blob: b07c48b09a93224f27b54a490c5beb857c744564 [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>
Balbir Singh66e17072008-02-07 00:13:56 -080040#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070041#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070042#include <linux/sysctl.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44#include <asm/tlbflush.h>
45#include <asm/div64.h>
46
47#include <linux/swapops.h>
48
Nick Piggin0f8053a2006-03-22 00:08:33 -080049#include "internal.h"
50
Linus Torvalds1da177e2005-04-16 15:20:36 -070051struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 /* Incremented by the number of inactive pages that were scanned */
53 unsigned long nr_scanned;
54
Rik van Riela79311c2009-01-06 14:40:01 -080055 /* Number of pages freed so far during a call to shrink_zones() */
56 unsigned long nr_reclaimed;
57
Linus Torvalds1da177e2005-04-16 15:20:36 -070058 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040059 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
61 int may_writepage;
62
Christoph Lameterf1fd1062006-01-18 17:42:30 -080063 /* Can pages be swapped as part of reclaim? */
64 int may_swap;
65
Linus Torvalds1da177e2005-04-16 15:20:36 -070066 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
67 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
68 * In this context, it doesn't matter that we scan the
69 * whole list at once. */
70 int swap_cluster_max;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070071
72 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070073
74 int all_unreclaimable;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070075
76 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080077
78 /* Which cgroup do we reclaim from */
79 struct mem_cgroup *mem_cgroup;
80
81 /* Pluggable isolate pages callback */
82 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
83 unsigned long *scanned, int order, int mode,
84 struct zone *z, struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -070085 int active, int file);
Linus Torvalds1da177e2005-04-16 15:20:36 -070086};
87
Linus Torvalds1da177e2005-04-16 15:20:36 -070088#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
89
90#ifdef ARCH_HAS_PREFETCH
91#define prefetch_prev_lru_page(_page, _base, _field) \
92 do { \
93 if ((_page)->lru.prev != _base) { \
94 struct page *prev; \
95 \
96 prev = lru_to_page(&(_page->lru)); \
97 prefetch(&prev->_field); \
98 } \
99 } while (0)
100#else
101#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
102#endif
103
104#ifdef ARCH_HAS_PREFETCHW
105#define prefetchw_prev_lru_page(_page, _base, _field) \
106 do { \
107 if ((_page)->lru.prev != _base) { \
108 struct page *prev; \
109 \
110 prev = lru_to_page(&(_page->lru)); \
111 prefetchw(&prev->_field); \
112 } \
113 } while (0)
114#else
115#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
116#endif
117
118/*
119 * From 0 .. 100. Higher means more swappy.
120 */
121int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700122long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700123
124static LIST_HEAD(shrinker_list);
125static DECLARE_RWSEM(shrinker_rwsem);
126
Balbir Singh00f0b822008-03-04 14:28:39 -0800127#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800128#define scan_global_lru(sc) (!(sc)->mem_cgroup)
129#else
130#define scan_global_lru(sc) (1)
131#endif
132
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133/*
134 * Add a shrinker callback to be called from the vm
135 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700136void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700138 shrinker->nr = 0;
139 down_write(&shrinker_rwsem);
140 list_add_tail(&shrinker->list, &shrinker_list);
141 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700142}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700143EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144
145/*
146 * Remove one
147 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700148void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149{
150 down_write(&shrinker_rwsem);
151 list_del(&shrinker->list);
152 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700153}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700154EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700155
156#define SHRINK_BATCH 128
157/*
158 * Call the shrink functions to age shrinkable caches
159 *
160 * Here we assume it costs one seek to replace a lru page and that it also
161 * takes a seek to recreate a cache object. With this in mind we age equal
162 * percentages of the lru and ageable caches. This should balance the seeks
163 * generated by these structures.
164 *
Simon Arlott183ff222007-10-20 01:27:18 +0200165 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166 * slab to avoid swapping.
167 *
168 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
169 *
170 * `lru_pages' represents the number of on-LRU pages in all the zones which
171 * are eligible for the caller's allocation attempt. It is used for balancing
172 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700173 *
174 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700175 */
Andrew Morton69e05942006-03-22 00:08:19 -0800176unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
177 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178{
179 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800180 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181
182 if (scanned == 0)
183 scanned = SWAP_CLUSTER_MAX;
184
185 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700186 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187
188 list_for_each_entry(shrinker, &shrinker_list, list) {
189 unsigned long long delta;
190 unsigned long total_scan;
Rusty Russell8e1f9362007-07-17 04:03:17 -0700191 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700192
193 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800194 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700195 do_div(delta, lru_pages + 1);
196 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800197 if (shrinker->nr < 0) {
198 printk(KERN_ERR "%s: nr=%ld\n",
Harvey Harrisond40cee22008-04-30 00:55:07 -0700199 __func__, shrinker->nr);
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800200 shrinker->nr = max_pass;
201 }
202
203 /*
204 * Avoid risking looping forever due to too large nr value:
205 * never try to free more than twice the estimate number of
206 * freeable entries.
207 */
208 if (shrinker->nr > max_pass * 2)
209 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210
211 total_scan = shrinker->nr;
212 shrinker->nr = 0;
213
214 while (total_scan >= SHRINK_BATCH) {
215 long this_scan = SHRINK_BATCH;
216 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700217 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700218
Rusty Russell8e1f9362007-07-17 04:03:17 -0700219 nr_before = (*shrinker->shrink)(0, gfp_mask);
220 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221 if (shrink_ret == -1)
222 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700223 if (shrink_ret < nr_before)
224 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700225 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226 total_scan -= this_scan;
227
228 cond_resched();
229 }
230
231 shrinker->nr += total_scan;
232 }
233 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700234 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700235}
236
237/* Called without lock on whether page is mapped, so answer is unstable */
238static inline int page_mapping_inuse(struct page *page)
239{
240 struct address_space *mapping;
241
242 /* Page is in somebody's page tables. */
243 if (page_mapped(page))
244 return 1;
245
246 /* Be more reluctant to reclaim swapcache than pagecache */
247 if (PageSwapCache(page))
248 return 1;
249
250 mapping = page_mapping(page);
251 if (!mapping)
252 return 0;
253
254 /* File is mmap'd by somebody? */
255 return mapping_mapped(mapping);
256}
257
258static inline int is_page_cache_freeable(struct page *page)
259{
260 return page_count(page) - !!PagePrivate(page) == 2;
261}
262
263static int may_write_to_queue(struct backing_dev_info *bdi)
264{
Christoph Lameter930d9152006-01-08 01:00:47 -0800265 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700266 return 1;
267 if (!bdi_write_congested(bdi))
268 return 1;
269 if (bdi == current->backing_dev_info)
270 return 1;
271 return 0;
272}
273
274/*
275 * We detected a synchronous write error writing a page out. Probably
276 * -ENOSPC. We need to propagate that into the address_space for a subsequent
277 * fsync(), msync() or close().
278 *
279 * The tricky part is that after writepage we cannot touch the mapping: nothing
280 * prevents it from being freed up. But we have a ref on the page and once
281 * that page is locked, the mapping is pinned.
282 *
283 * We're allowed to run sleeping lock_page() here because we know the caller has
284 * __GFP_FS.
285 */
286static void handle_write_error(struct address_space *mapping,
287 struct page *page, int error)
288{
289 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700290 if (page_mapping(page) == mapping)
291 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 unlock_page(page);
293}
294
Andy Whitcroftc661b072007-08-22 14:01:26 -0700295/* Request for sync pageout. */
296enum pageout_io {
297 PAGEOUT_IO_ASYNC,
298 PAGEOUT_IO_SYNC,
299};
300
Christoph Lameter04e62a22006-06-23 02:03:38 -0700301/* possible outcome of pageout() */
302typedef enum {
303 /* failed to write page out, page is locked */
304 PAGE_KEEP,
305 /* move page to the active list, page is locked */
306 PAGE_ACTIVATE,
307 /* page has been sent to the disk successfully, page is unlocked */
308 PAGE_SUCCESS,
309 /* page is clean and locked */
310 PAGE_CLEAN,
311} pageout_t;
312
Linus Torvalds1da177e2005-04-16 15:20:36 -0700313/*
Andrew Morton1742f192006-03-22 00:08:21 -0800314 * pageout is called by shrink_page_list() for each dirty page.
315 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700317static pageout_t pageout(struct page *page, struct address_space *mapping,
318 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700319{
320 /*
321 * If the page is dirty, only perform writeback if that write
322 * will be non-blocking. To prevent this allocation from being
323 * stalled by pagecache activity. But note that there may be
324 * stalls if we need to run get_block(). We could test
325 * PagePrivate for that.
326 *
327 * If this process is currently in generic_file_write() against
328 * this page's queue, we can perform writeback even if that
329 * will block.
330 *
331 * If the page is swapcache, write it back even if that would
332 * block, for some throttling. This happens by accident, because
333 * swap_backing_dev_info is bust: it doesn't reflect the
334 * congestion state of the swapdevs. Easy to fix, if needed.
335 * See swapfile.c:page_queue_congested().
336 */
337 if (!is_page_cache_freeable(page))
338 return PAGE_KEEP;
339 if (!mapping) {
340 /*
341 * Some data journaling orphaned pages can have
342 * page->mapping == NULL while being dirty with clean buffers.
343 */
akpm@osdl.org323aca62005-04-16 15:24:06 -0700344 if (PagePrivate(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 if (try_to_free_buffers(page)) {
346 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700347 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 return PAGE_CLEAN;
349 }
350 }
351 return PAGE_KEEP;
352 }
353 if (mapping->a_ops->writepage == NULL)
354 return PAGE_ACTIVATE;
355 if (!may_write_to_queue(mapping->backing_dev_info))
356 return PAGE_KEEP;
357
358 if (clear_page_dirty_for_io(page)) {
359 int res;
360 struct writeback_control wbc = {
361 .sync_mode = WB_SYNC_NONE,
362 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700363 .range_start = 0,
364 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365 .nonblocking = 1,
366 .for_reclaim = 1,
367 };
368
369 SetPageReclaim(page);
370 res = mapping->a_ops->writepage(page, &wbc);
371 if (res < 0)
372 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800373 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 ClearPageReclaim(page);
375 return PAGE_ACTIVATE;
376 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700377
378 /*
379 * Wait on writeback if requested to. This happens when
380 * direct reclaiming a large contiguous area and the
381 * first attempt to free a range of pages fails.
382 */
383 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
384 wait_on_page_writeback(page);
385
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386 if (!PageWriteback(page)) {
387 /* synchronous write or broken a_ops? */
388 ClearPageReclaim(page);
389 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700390 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 return PAGE_SUCCESS;
392 }
393
394 return PAGE_CLEAN;
395}
396
Andrew Mortona649fd92006-10-17 00:09:36 -0700397/*
Nick Piggine2867812008-07-25 19:45:30 -0700398 * Same as remove_mapping, but if the page is removed from the mapping, it
399 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700400 */
Nick Piggine2867812008-07-25 19:45:30 -0700401static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800402{
Nick Piggin28e4d962006-09-25 23:31:23 -0700403 BUG_ON(!PageLocked(page));
404 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800405
Nick Piggin19fd6232008-07-25 19:45:32 -0700406 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800407 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700408 * The non racy check for a busy page.
409 *
410 * Must be careful with the order of the tests. When someone has
411 * a ref to the page, it may be possible that they dirty it then
412 * drop the reference. So if PageDirty is tested before page_count
413 * here, then the following race may occur:
414 *
415 * get_user_pages(&page);
416 * [user mapping goes away]
417 * write_to(page);
418 * !PageDirty(page) [good]
419 * SetPageDirty(page);
420 * put_page(page);
421 * !page_count(page) [good, discard it]
422 *
423 * [oops, our write_to data is lost]
424 *
425 * Reversing the order of the tests ensures such a situation cannot
426 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
427 * load is not satisfied before that of page->_count.
428 *
429 * Note that if SetPageDirty is always performed via set_page_dirty,
430 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800431 */
Nick Piggine2867812008-07-25 19:45:30 -0700432 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800433 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700434 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
435 if (unlikely(PageDirty(page))) {
436 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800437 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700438 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800439
440 if (PageSwapCache(page)) {
441 swp_entry_t swap = { .val = page_private(page) };
442 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700443 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800444 swap_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700445 } else {
446 __remove_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700447 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800448 }
449
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800450 return 1;
451
452cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700453 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800454 return 0;
455}
456
Linus Torvalds1da177e2005-04-16 15:20:36 -0700457/*
Nick Piggine2867812008-07-25 19:45:30 -0700458 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
459 * someone else has a ref on the page, abort and return 0. If it was
460 * successfully detached, return 1. Assumes the caller has a single ref on
461 * this page.
462 */
463int remove_mapping(struct address_space *mapping, struct page *page)
464{
465 if (__remove_mapping(mapping, page)) {
466 /*
467 * Unfreezing the refcount with 1 rather than 2 effectively
468 * drops the pagecache ref for us without requiring another
469 * atomic operation.
470 */
471 page_unfreeze_refs(page, 1);
472 return 1;
473 }
474 return 0;
475}
476
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700477/**
478 * putback_lru_page - put previously isolated page onto appropriate LRU list
479 * @page: page to be put back to appropriate lru list
480 *
481 * Add previously isolated @page to appropriate LRU list.
482 * Page may still be unevictable for other reasons.
483 *
484 * lru_lock must not be held, interrupts must be enabled.
485 */
486#ifdef CONFIG_UNEVICTABLE_LRU
487void putback_lru_page(struct page *page)
488{
489 int lru;
490 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700491 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700492
493 VM_BUG_ON(PageLRU(page));
494
495redo:
496 ClearPageUnevictable(page);
497
498 if (page_evictable(page, NULL)) {
499 /*
500 * For evictable pages, we can use the cache.
501 * In event of a race, worst case is we end up with an
502 * unevictable page on [in]active list.
503 * We know how to handle that.
504 */
505 lru = active + page_is_file_cache(page);
506 lru_cache_add_lru(page, lru);
507 } else {
508 /*
509 * Put unevictable pages directly on zone's unevictable
510 * list.
511 */
512 lru = LRU_UNEVICTABLE;
513 add_page_to_unevictable_list(page);
514 }
515 mem_cgroup_move_lists(page, lru);
516
517 /*
518 * page's status can change while we move it among lru. If an evictable
519 * page is on unevictable list, it never be freed. To avoid that,
520 * check after we added it to the list, again.
521 */
522 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
523 if (!isolate_lru_page(page)) {
524 put_page(page);
525 goto redo;
526 }
527 /* This means someone else dropped this page from LRU
528 * So, it will be freed or putback to LRU again. There is
529 * nothing to do here.
530 */
531 }
532
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700533 if (was_unevictable && lru != LRU_UNEVICTABLE)
534 count_vm_event(UNEVICTABLE_PGRESCUED);
535 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
536 count_vm_event(UNEVICTABLE_PGCULLED);
537
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700538 put_page(page); /* drop ref from isolate */
539}
540
541#else /* CONFIG_UNEVICTABLE_LRU */
542
543void putback_lru_page(struct page *page)
544{
545 int lru;
546 VM_BUG_ON(PageLRU(page));
547
548 lru = !!TestClearPageActive(page) + page_is_file_cache(page);
549 lru_cache_add_lru(page, lru);
550 mem_cgroup_move_lists(page, lru);
551 put_page(page);
552}
553#endif /* CONFIG_UNEVICTABLE_LRU */
554
555
Nick Piggine2867812008-07-25 19:45:30 -0700556/*
Andrew Morton1742f192006-03-22 00:08:21 -0800557 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558 */
Andrew Morton1742f192006-03-22 00:08:21 -0800559static unsigned long shrink_page_list(struct list_head *page_list,
Andy Whitcroftc661b072007-08-22 14:01:26 -0700560 struct scan_control *sc,
561 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700562{
563 LIST_HEAD(ret_pages);
564 struct pagevec freed_pvec;
565 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800566 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567
568 cond_resched();
569
570 pagevec_init(&freed_pvec, 1);
571 while (!list_empty(page_list)) {
572 struct address_space *mapping;
573 struct page *page;
574 int may_enter_fs;
575 int referenced;
576
577 cond_resched();
578
579 page = lru_to_page(page_list);
580 list_del(&page->lru);
581
Nick Piggin529ae9a2008-08-02 12:01:03 +0200582 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700583 goto keep;
584
Nick Piggin725d7042006-09-25 23:30:55 -0700585 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586
587 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800588
Nick Pigginb291f002008-10-18 20:26:44 -0700589 if (unlikely(!page_evictable(page, NULL)))
590 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700591
Christoph Lameter80e43422006-02-11 17:55:53 -0800592 if (!sc->may_swap && page_mapped(page))
593 goto keep_locked;
594
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595 /* Double the slab pressure for mapped and swapcache pages */
596 if (page_mapped(page) || PageSwapCache(page))
597 sc->nr_scanned++;
598
Andy Whitcroftc661b072007-08-22 14:01:26 -0700599 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
600 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
601
602 if (PageWriteback(page)) {
603 /*
604 * Synchronous reclaim is performed in two passes,
605 * first an asynchronous pass over the list to
606 * start parallel writeback, and a second synchronous
607 * pass to wait for the IO to complete. Wait here
608 * for any page for which writeback has already
609 * started.
610 */
611 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
612 wait_on_page_writeback(page);
Andrew Morton4dd4b922008-03-24 12:29:52 -0700613 else
Andy Whitcroftc661b072007-08-22 14:01:26 -0700614 goto keep_locked;
615 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616
Balbir Singhbed71612008-02-07 00:14:01 -0800617 referenced = page_referenced(page, 1, sc->mem_cgroup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700618 /* In active use or really unfreeable? Activate it. */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700619 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
620 referenced && page_mapping_inuse(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621 goto activate_locked;
622
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623 /*
624 * Anonymous process memory has backing store?
625 * Try to allocate it some swap space here.
626 */
Nick Pigginb291f002008-10-18 20:26:44 -0700627 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800628 if (!(sc->gfp_mask & __GFP_IO))
629 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800630 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800632 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700633 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634
635 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700636
637 /*
638 * The page is mapped into the page tables of one or more
639 * processes. Try to unmap it here.
640 */
641 if (page_mapped(page) && mapping) {
Christoph Lametera48d07a2006-02-01 03:05:38 -0800642 switch (try_to_unmap(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 case SWAP_FAIL:
644 goto activate_locked;
645 case SWAP_AGAIN:
646 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700647 case SWAP_MLOCK:
648 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700649 case SWAP_SUCCESS:
650 ; /* try to free the page below */
651 }
652 }
653
654 if (PageDirty(page)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700655 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700657 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800659 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660 goto keep_locked;
661
662 /* Page is dirty, try to write it out here */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700663 switch (pageout(page, mapping, sync_writeback)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 case PAGE_KEEP:
665 goto keep_locked;
666 case PAGE_ACTIVATE:
667 goto activate_locked;
668 case PAGE_SUCCESS:
Andrew Morton4dd4b922008-03-24 12:29:52 -0700669 if (PageWriteback(page) || PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670 goto keep;
671 /*
672 * A synchronous write - probably a ramdisk. Go
673 * ahead and try to reclaim the page.
674 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200675 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 goto keep;
677 if (PageDirty(page) || PageWriteback(page))
678 goto keep_locked;
679 mapping = page_mapping(page);
680 case PAGE_CLEAN:
681 ; /* try to free the page below */
682 }
683 }
684
685 /*
686 * If the page has buffers, try to free the buffer mappings
687 * associated with this page. If we succeed we try to free
688 * the page as well.
689 *
690 * We do this even if the page is PageDirty().
691 * try_to_release_page() does not perform I/O, but it is
692 * possible for a page to have PageDirty set, but it is actually
693 * clean (all its buffers are clean). This happens if the
694 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700695 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 * try_to_release_page() will discover that cleanness and will
697 * drop the buffers and mark the page clean - it can be freed.
698 *
699 * Rarely, pages can have buffers and no ->mapping. These are
700 * the pages which were not successfully invalidated in
701 * truncate_complete_page(). We try to drop those buffers here
702 * and if that worked, and the page is no longer mapped into
703 * process address space (page_count == 1) it can be freed.
704 * Otherwise, leave the page on the LRU so it is swappable.
705 */
706 if (PagePrivate(page)) {
707 if (!try_to_release_page(page, sc->gfp_mask))
708 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700709 if (!mapping && page_count(page) == 1) {
710 unlock_page(page);
711 if (put_page_testzero(page))
712 goto free_it;
713 else {
714 /*
715 * rare race with speculative reference.
716 * the speculative reference will free
717 * this page shortly, so we may
718 * increment nr_reclaimed here (and
719 * leave it off the LRU).
720 */
721 nr_reclaimed++;
722 continue;
723 }
724 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 }
726
Nick Piggine2867812008-07-25 19:45:30 -0700727 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800728 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729
Nick Piggina978d6f2008-10-18 20:26:58 -0700730 /*
731 * At this point, we have no other references and there is
732 * no way to pick any more up (removed from LRU, removed
733 * from pagecache). Can use non-atomic bitops now (and
734 * we obviously don't have to worry about waking up a process
735 * waiting on the page lock, because there are no references.
736 */
737 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700738free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800739 nr_reclaimed++;
Nick Piggine2867812008-07-25 19:45:30 -0700740 if (!pagevec_add(&freed_pvec, page)) {
741 __pagevec_free(&freed_pvec);
742 pagevec_reinit(&freed_pvec);
743 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700744 continue;
745
Nick Pigginb291f002008-10-18 20:26:44 -0700746cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800747 if (PageSwapCache(page))
748 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700749 unlock_page(page);
750 putback_lru_page(page);
751 continue;
752
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700754 /* Not a candidate for swapping, so reclaim swap space. */
755 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800756 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700757 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 SetPageActive(page);
759 pgactivate++;
760keep_locked:
761 unlock_page(page);
762keep:
763 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700764 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 }
766 list_splice(&ret_pages, page_list);
767 if (pagevec_count(&freed_pvec))
Nick Piggine2867812008-07-25 19:45:30 -0700768 __pagevec_free(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700769 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800770 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771}
772
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700773/* LRU Isolation modes. */
774#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
775#define ISOLATE_ACTIVE 1 /* Isolate active pages. */
776#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
777
778/*
779 * Attempt to remove the specified page from its LRU. Only take this page
780 * if it is of the appropriate PageActive status. Pages which are being
781 * freed elsewhere are also ignored.
782 *
783 * page: page to consider
784 * mode: one of the LRU isolation modes defined above
785 *
786 * returns 0 on success, -ve errno on failure.
787 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700788int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700789{
790 int ret = -EINVAL;
791
792 /* Only take pages on the LRU. */
793 if (!PageLRU(page))
794 return ret;
795
796 /*
797 * When checking the active state, we need to be sure we are
798 * dealing with comparible boolean values. Take the logical not
799 * of each.
800 */
801 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
802 return ret;
803
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700804 if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
805 return ret;
806
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700807 /*
808 * When this function is being called for lumpy reclaim, we
809 * initially look into all LRU pages, active, inactive and
810 * unevictable; only give shrink_page_list evictable pages.
811 */
812 if (PageUnevictable(page))
813 return ret;
814
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700815 ret = -EBUSY;
816 if (likely(get_page_unless_zero(page))) {
817 /*
818 * Be careful not to clear PageLRU until after we're
819 * sure the page is not being freed elsewhere -- the
820 * page release code relies on it.
821 */
822 ClearPageLRU(page);
823 ret = 0;
824 }
825
826 return ret;
827}
828
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800829/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700830 * zone->lru_lock is heavily contended. Some of the functions that
831 * shrink the lists perform better by taking out a batch of pages
832 * and working on them outside the LRU lock.
833 *
834 * For pagecache intensive workloads, this function is the hottest
835 * spot in the kernel (apart from copy_*_user functions).
836 *
837 * Appropriate locks must be held before calling this function.
838 *
839 * @nr_to_scan: The number of pages to look through on the list.
840 * @src: The LRU list to pull pages off.
841 * @dst: The temp list to put pages on to.
842 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700843 * @order: The caller's attempted allocation order
844 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700845 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700846 *
847 * returns how many pages were moved onto *@dst.
848 */
Andrew Morton69e05942006-03-22 00:08:19 -0800849static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
850 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700851 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852{
Andrew Morton69e05942006-03-22 00:08:19 -0800853 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800854 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800856 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700857 struct page *page;
858 unsigned long pfn;
859 unsigned long end_pfn;
860 unsigned long page_pfn;
861 int zone_id;
862
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 page = lru_to_page(src);
864 prefetchw_prev_lru_page(page, src, flags);
865
Nick Piggin725d7042006-09-25 23:30:55 -0700866 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800867
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700868 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700869 case 0:
870 list_move(&page->lru, dst);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800871 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700872 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800873
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700874 case -EBUSY:
875 /* else it is being freed elsewhere */
876 list_move(&page->lru, src);
877 continue;
878
879 default:
880 BUG();
881 }
882
883 if (!order)
884 continue;
885
886 /*
887 * Attempt to take all pages in the order aligned region
888 * surrounding the tag page. Only take those pages of
889 * the same active state as that tag page. We may safely
890 * round the target page pfn down to the requested order
891 * as the mem_map is guarenteed valid out to MAX_ORDER,
892 * where that page is in a different zone we will detect
893 * it from its zone id and abort this block scan.
894 */
895 zone_id = page_zone_id(page);
896 page_pfn = page_to_pfn(page);
897 pfn = page_pfn & ~((1 << order) - 1);
898 end_pfn = pfn + (1 << order);
899 for (; pfn < end_pfn; pfn++) {
900 struct page *cursor_page;
901
902 /* The target page is in the block, ignore it. */
903 if (unlikely(pfn == page_pfn))
904 continue;
905
906 /* Avoid holes within the zone. */
907 if (unlikely(!pfn_valid_within(pfn)))
908 break;
909
910 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700911
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700912 /* Check that we have not crossed a zone boundary. */
913 if (unlikely(page_zone_id(cursor_page) != zone_id))
914 continue;
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700915 switch (__isolate_lru_page(cursor_page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700916 case 0:
917 list_move(&cursor_page->lru, dst);
918 nr_taken++;
919 scan++;
920 break;
921
922 case -EBUSY:
923 /* else it is being freed elsewhere */
924 list_move(&cursor_page->lru, src);
925 default:
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700926 break; /* ! on LRU or wrong list */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700927 }
928 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700929 }
930
931 *scanned = scan;
932 return nr_taken;
933}
934
Balbir Singh66e17072008-02-07 00:13:56 -0800935static unsigned long isolate_pages_global(unsigned long nr,
936 struct list_head *dst,
937 unsigned long *scanned, int order,
938 int mode, struct zone *z,
939 struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700940 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -0800941{
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700942 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -0800943 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700944 lru += LRU_ACTIVE;
945 if (file)
946 lru += LRU_FILE;
947 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
948 mode, !!file);
Balbir Singh66e17072008-02-07 00:13:56 -0800949}
950
Linus Torvalds1da177e2005-04-16 15:20:36 -0700951/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700952 * clear_active_flags() is a helper for shrink_active_list(), clearing
953 * any active bits from the pages in the list.
954 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700955static unsigned long clear_active_flags(struct list_head *page_list,
956 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700957{
958 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700959 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700960 struct page *page;
961
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700962 list_for_each_entry(page, page_list, lru) {
963 lru = page_is_file_cache(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700964 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700965 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700966 ClearPageActive(page);
967 nr_active++;
968 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700969 count[lru]++;
970 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700971
972 return nr_active;
973}
974
Nick Piggin62695a82008-10-18 20:26:09 -0700975/**
976 * isolate_lru_page - tries to isolate a page from its LRU list
977 * @page: page to isolate from its LRU list
978 *
979 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
980 * vmstat statistic corresponding to whatever LRU list the page was on.
981 *
982 * Returns 0 if the page was removed from an LRU list.
983 * Returns -EBUSY if the page was not on an LRU list.
984 *
985 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700986 * the active list, it will have PageActive set. If it was found on
987 * the unevictable list, it will have the PageUnevictable bit set. That flag
988 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -0700989 *
990 * The vmstat statistic corresponding to the list on which the page was
991 * found will be decremented.
992 *
993 * Restrictions:
994 * (1) Must be called with an elevated refcount on the page. This is a
995 * fundamentnal difference from isolate_lru_pages (which is called
996 * without a stable reference).
997 * (2) the lru_lock must not be held.
998 * (3) interrupts must be enabled.
999 */
1000int isolate_lru_page(struct page *page)
1001{
1002 int ret = -EBUSY;
1003
1004 if (PageLRU(page)) {
1005 struct zone *zone = page_zone(page);
1006
1007 spin_lock_irq(&zone->lru_lock);
1008 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001009 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001010 ret = 0;
1011 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001012
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001013 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001014 }
1015 spin_unlock_irq(&zone->lru_lock);
1016 }
1017 return ret;
1018}
1019
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001020/*
Andrew Morton1742f192006-03-22 00:08:21 -08001021 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1022 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001023 */
Andrew Morton1742f192006-03-22 00:08:21 -08001024static unsigned long shrink_inactive_list(unsigned long max_scan,
Rik van Riel33c120e2008-10-18 20:26:36 -07001025 struct zone *zone, struct scan_control *sc,
1026 int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027{
1028 LIST_HEAD(page_list);
1029 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -08001030 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001031 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001032
1033 pagevec_init(&pvec, 1);
1034
1035 lru_add_drain();
1036 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -08001037 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001038 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -08001039 unsigned long nr_taken;
1040 unsigned long nr_scan;
1041 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001042 unsigned long nr_active;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001043 unsigned int count[NR_LRU_LISTS] = { 0, };
Rik van Riel33c120e2008-10-18 20:26:36 -07001044 int mode = ISOLATE_INACTIVE;
1045
1046 /*
1047 * If we need a large contiguous chunk of memory, or have
1048 * trouble getting a small set of contiguous pages, we
1049 * will reclaim both active and inactive pages.
1050 *
1051 * We use the same threshold as pageout congestion_wait below.
1052 */
1053 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1054 mode = ISOLATE_BOTH;
1055 else if (sc->order && priority < DEF_PRIORITY - 2)
1056 mode = ISOLATE_BOTH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001057
Balbir Singh66e17072008-02-07 00:13:56 -08001058 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001059 &page_list, &nr_scan, sc->order, mode,
1060 zone, sc->mem_cgroup, 0, file);
1061 nr_active = clear_active_flags(&page_list, count);
Andy Whitcrofte9187bd2007-08-22 14:01:25 -07001062 __count_vm_events(PGDEACTIVATE, nr_active);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001063
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001064 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1065 -count[LRU_ACTIVE_FILE]);
1066 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1067 -count[LRU_INACTIVE_FILE]);
1068 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1069 -count[LRU_ACTIVE_ANON]);
1070 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1071 -count[LRU_INACTIVE_ANON]);
1072
1073 if (scan_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001074 zone->pages_scanned += nr_scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001075 zone->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1076 zone->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1077 zone->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1078 zone->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1079 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001080 spin_unlock_irq(&zone->lru_lock);
1081
Andrew Morton69e05942006-03-22 00:08:19 -08001082 nr_scanned += nr_scan;
Andy Whitcroftc661b072007-08-22 14:01:26 -07001083 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1084
1085 /*
1086 * If we are direct reclaiming for contiguous pages and we do
1087 * not reclaim everything in the list, try again and wait
1088 * for IO to complete. This will stall high-order allocations
1089 * but that should be acceptable to the caller
1090 */
1091 if (nr_freed < nr_taken && !current_is_kswapd() &&
1092 sc->order > PAGE_ALLOC_COSTLY_ORDER) {
1093 congestion_wait(WRITE, HZ/10);
1094
1095 /*
1096 * The attempt at page out may have made some
1097 * of the pages active, mark them inactive again.
1098 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001099 nr_active = clear_active_flags(&page_list, count);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001100 count_vm_events(PGDEACTIVATE, nr_active);
1101
1102 nr_freed += shrink_page_list(&page_list, sc,
1103 PAGEOUT_IO_SYNC);
1104 }
1105
Andrew Morton05ff5132006-03-22 00:08:20 -08001106 nr_reclaimed += nr_freed;
Nick Piggina74609f2006-01-06 00:11:20 -08001107 local_irq_disable();
1108 if (current_is_kswapd()) {
Christoph Lameterf8891e52006-06-30 01:55:45 -07001109 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1110 __count_vm_events(KSWAPD_STEAL, nr_freed);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001111 } else if (scan_global_lru(sc))
Christoph Lameterf8891e52006-06-30 01:55:45 -07001112 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001113
Shantanu Goel918d3f92006-12-29 16:48:59 -08001114 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -08001115
Wu Fengguangfb8d14e2006-03-22 00:08:28 -08001116 if (nr_taken == 0)
1117 goto done;
1118
Nick Piggina74609f2006-01-06 00:11:20 -08001119 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120 /*
1121 * Put back any unfreeable pages.
1122 */
1123 while (!list_empty(&page_list)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001124 int lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -07001126 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127 list_del(&page->lru);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001128 if (unlikely(!page_evictable(page, NULL))) {
1129 spin_unlock_irq(&zone->lru_lock);
1130 putback_lru_page(page);
1131 spin_lock_irq(&zone->lru_lock);
1132 continue;
1133 }
1134 SetPageLRU(page);
1135 lru = page_lru(page);
1136 add_page_to_lru_list(zone, page, lru);
1137 mem_cgroup_move_lists(page, lru);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001138 if (PageActive(page) && scan_global_lru(sc)) {
1139 int file = !!page_is_file_cache(page);
1140 zone->recent_rotated[file]++;
1141 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 if (!pagevec_add(&pvec, page)) {
1143 spin_unlock_irq(&zone->lru_lock);
1144 __pagevec_release(&pvec);
1145 spin_lock_irq(&zone->lru_lock);
1146 }
1147 }
Andrew Morton69e05942006-03-22 00:08:19 -08001148 } while (nr_scanned < max_scan);
Wu Fengguangfb8d14e2006-03-22 00:08:28 -08001149 spin_unlock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001150done:
Wu Fengguangfb8d14e2006-03-22 00:08:28 -08001151 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -08001153 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154}
1155
Martin Bligh3bb1a852006-10-28 10:38:24 -07001156/*
1157 * We are about to scan this zone at a certain priority level. If that priority
1158 * level is smaller (ie: more urgent) than the previous priority, then note
1159 * that priority level within the zone. This is done so that when the next
1160 * process comes in to scan this zone, it will immediately start out at this
1161 * priority level rather than having to build up its own scanning priority.
1162 * Here, this priority affects only the reclaim-mapped threshold.
1163 */
1164static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1165{
1166 if (priority < zone->prev_priority)
1167 zone->prev_priority = priority;
1168}
1169
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001170/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 * This moves pages from the active list to the inactive list.
1172 *
1173 * We move them the other way if the page is referenced by one or more
1174 * processes, from rmap.
1175 *
1176 * If the pages are mostly unmapped, the processing is fast and it is
1177 * appropriate to hold zone->lru_lock across the whole operation. But if
1178 * the pages are mapped, the processing is slow (page_referenced()) so we
1179 * should drop zone->lru_lock around each page. It's impossible to balance
1180 * this, so instead we remove the pages from the LRU while processing them.
1181 * It is safe to rely on PG_active against the non-LRU pages in here because
1182 * nobody will play with that bit on a non-LRU page.
1183 *
1184 * The downside is that we have to touch page->_count against each page.
1185 * But we had to alter page->flags anyway.
1186 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001187
1188
Andrew Morton1742f192006-03-22 00:08:21 -08001189static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001190 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001191{
Andrew Morton69e05942006-03-22 00:08:19 -08001192 unsigned long pgmoved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193 int pgdeactivate = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001194 unsigned long pgscanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001195 LIST_HEAD(l_hold); /* The pages which were snipped off */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001196 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 struct page *page;
1198 struct pagevec pvec;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001199 enum lru_list lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001200
1201 lru_add_drain();
1202 spin_lock_irq(&zone->lru_lock);
Balbir Singh66e17072008-02-07 00:13:56 -08001203 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1204 ISOLATE_ACTIVE, zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001205 sc->mem_cgroup, 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001206 /*
1207 * zone->pages_scanned is used for detect zone's oom
1208 * mem_cgroup remembers nr_scan by itself.
1209 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001210 if (scan_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001211 zone->pages_scanned += pgscanned;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001212 zone->recent_scanned[!!file] += pgmoved;
1213 }
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001214
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001215 if (file)
1216 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1217 else
1218 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 spin_unlock_irq(&zone->lru_lock);
1220
Rik van Riel556adec2008-10-18 20:26:34 -07001221 pgmoved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222 while (!list_empty(&l_hold)) {
1223 cond_resched();
1224 page = lru_to_page(&l_hold);
1225 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001226
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001227 if (unlikely(!page_evictable(page, NULL))) {
1228 putback_lru_page(page);
1229 continue;
1230 }
1231
Rik van Riel7e9cd482008-10-18 20:26:35 -07001232 /* page_referenced clears PageReferenced */
1233 if (page_mapping_inuse(page) &&
1234 page_referenced(page, 0, sc->mem_cgroup))
1235 pgmoved++;
1236
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 list_add(&page->lru, &l_inactive);
1238 }
1239
Andrew Mortonb5557492009-01-06 14:40:13 -08001240 /*
1241 * Move the pages to the [file or anon] inactive list.
1242 */
1243 pagevec_init(&pvec, 1);
1244 pgmoved = 0;
1245 lru = LRU_BASE + file * LRU_FILE;
1246
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001247 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001248 /*
Rik van Riel7e9cd482008-10-18 20:26:35 -07001249 * Count referenced pages from currently used mappings as
1250 * rotated, even though they are moved to the inactive list.
1251 * This helps balance scan pressure between file and anonymous
1252 * pages in get_scan_ratio.
1253 */
KOSAKI Motohiro077cbc52009-01-06 14:39:42 -08001254 if (scan_global_lru(sc))
1255 zone->recent_rotated[!!file] += pgmoved;
Rik van Riel556adec2008-10-18 20:26:34 -07001256
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 while (!list_empty(&l_inactive)) {
1258 page = lru_to_page(&l_inactive);
1259 prefetchw_prev_lru_page(page, &l_inactive, flags);
Nick Piggin725d7042006-09-25 23:30:55 -07001260 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001261 SetPageLRU(page);
Nick Piggin725d7042006-09-25 23:30:55 -07001262 VM_BUG_ON(!PageActive(page));
Nick Piggin4c84cac2006-03-22 00:08:00 -08001263 ClearPageActive(page);
1264
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001265 list_move(&page->lru, &zone->lru[lru].list);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001266 mem_cgroup_move_lists(page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 pgmoved++;
1268 if (!pagevec_add(&pvec, page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001269 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270 spin_unlock_irq(&zone->lru_lock);
1271 pgdeactivate += pgmoved;
1272 pgmoved = 0;
1273 if (buffer_heads_over_limit)
1274 pagevec_strip(&pvec);
1275 __pagevec_release(&pvec);
1276 spin_lock_irq(&zone->lru_lock);
1277 }
1278 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001279 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001280 pgdeactivate += pgmoved;
1281 if (buffer_heads_over_limit) {
1282 spin_unlock_irq(&zone->lru_lock);
1283 pagevec_strip(&pvec);
1284 spin_lock_irq(&zone->lru_lock);
1285 }
Christoph Lameterf8891e52006-06-30 01:55:45 -07001286 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1287 __count_vm_events(PGDEACTIVATE, pgdeactivate);
1288 spin_unlock_irq(&zone->lru_lock);
Rik van Riel68a223942008-10-18 20:26:23 -07001289 if (vm_swap_full())
1290 pagevec_swap_free(&pvec);
Nick Piggina74609f2006-01-06 00:11:20 -08001291
1292 pagevec_release(&pvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001293}
1294
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001295static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001296 struct zone *zone, struct scan_control *sc, int priority)
1297{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001298 int file = is_file_lru(lru);
1299
Rik van Riel556adec2008-10-18 20:26:34 -07001300 if (lru == LRU_ACTIVE_FILE) {
1301 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1302 return 0;
1303 }
1304
1305 if (lru == LRU_ACTIVE_ANON &&
1306 (!scan_global_lru(sc) || inactive_anon_is_low(zone))) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001307 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001308 return 0;
1309 }
Rik van Riel33c120e2008-10-18 20:26:36 -07001310 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001311}
1312
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001314 * Determine how aggressively the anon and file LRU lists should be
1315 * scanned. The relative value of each set of LRU lists is determined
1316 * by looking at the fraction of the pages scanned we did rotate back
1317 * onto the active list instead of evict.
1318 *
1319 * percent[0] specifies how much pressure to put on ram/swap backed
1320 * memory, while percent[1] determines pressure on the file LRUs.
1321 */
1322static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1323 unsigned long *percent)
1324{
1325 unsigned long anon, file, free;
1326 unsigned long anon_prio, file_prio;
1327 unsigned long ap, fp;
1328
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001329 /* If we have no swap space, do not bother scanning anon pages. */
1330 if (nr_swap_pages <= 0) {
1331 percent[0] = 0;
1332 percent[1] = 100;
1333 return;
1334 }
1335
Hugh Dickinsb9627162009-01-06 14:39:41 -08001336 anon = zone_page_state(zone, NR_ACTIVE_ANON) +
1337 zone_page_state(zone, NR_INACTIVE_ANON);
1338 file = zone_page_state(zone, NR_ACTIVE_FILE) +
1339 zone_page_state(zone, NR_INACTIVE_FILE);
1340 free = zone_page_state(zone, NR_FREE_PAGES);
1341
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001342 /* If we have very few page cache pages, force-scan anon pages. */
1343 if (unlikely(file + free <= zone->pages_high)) {
1344 percent[0] = 100;
1345 percent[1] = 0;
1346 return;
1347 }
1348
1349 /*
1350 * OK, so we have swap space and a fair amount of page cache
1351 * pages. We use the recently rotated / recently scanned
1352 * ratios to determine how valuable each cache is.
1353 *
1354 * Because workloads change over time (and to avoid overflow)
1355 * we keep these statistics as a floating average, which ends
1356 * up weighing recent references more than old ones.
1357 *
1358 * anon in [0], file in [1]
1359 */
1360 if (unlikely(zone->recent_scanned[0] > anon / 4)) {
1361 spin_lock_irq(&zone->lru_lock);
1362 zone->recent_scanned[0] /= 2;
1363 zone->recent_rotated[0] /= 2;
1364 spin_unlock_irq(&zone->lru_lock);
1365 }
1366
1367 if (unlikely(zone->recent_scanned[1] > file / 4)) {
1368 spin_lock_irq(&zone->lru_lock);
1369 zone->recent_scanned[1] /= 2;
1370 zone->recent_rotated[1] /= 2;
1371 spin_unlock_irq(&zone->lru_lock);
1372 }
1373
1374 /*
1375 * With swappiness at 100, anonymous and file have the same priority.
1376 * This scanning priority is essentially the inverse of IO cost.
1377 */
1378 anon_prio = sc->swappiness;
1379 file_prio = 200 - sc->swappiness;
1380
1381 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001382 * The amount of pressure on anon vs file pages is inversely
1383 * proportional to the fraction of recently scanned pages on
1384 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001385 */
1386 ap = (anon_prio + 1) * (zone->recent_scanned[0] + 1);
1387 ap /= zone->recent_rotated[0] + 1;
1388
1389 fp = (file_prio + 1) * (zone->recent_scanned[1] + 1);
1390 fp /= zone->recent_rotated[1] + 1;
1391
1392 /* Normalize to percentages */
1393 percent[0] = 100 * ap / (ap + fp + 1);
1394 percent[1] = 100 - percent[0];
1395}
1396
1397
1398/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001399 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1400 */
Rik van Riela79311c2009-01-06 14:40:01 -08001401static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001402 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001403{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001404 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001405 unsigned long nr_to_scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001406 unsigned long percent[2]; /* anon @ 0; file @ 1 */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001407 enum lru_list l;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001408 unsigned long nr_reclaimed = sc->nr_reclaimed;
1409 unsigned long swap_cluster_max = sc->swap_cluster_max;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001410
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001411 get_scan_ratio(zone, sc, percent);
1412
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001413 for_each_evictable_lru(l) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001414 if (scan_global_lru(sc)) {
1415 int file = is_file_lru(l);
1416 int scan;
Johannes Weinere0f79b82008-10-18 20:26:55 -07001417
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001418 scan = zone_page_state(zone, NR_LRU_BASE + l);
1419 if (priority) {
1420 scan >>= priority;
1421 scan = (scan * percent[file]) / 100;
1422 }
Johannes Weinere0f79b82008-10-18 20:26:55 -07001423 zone->lru[l].nr_scan += scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001424 nr[l] = zone->lru[l].nr_scan;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001425 if (nr[l] >= swap_cluster_max)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001426 zone->lru[l].nr_scan = 0;
1427 else
1428 nr[l] = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001429 } else {
1430 /*
1431 * This reclaim occurs not because zone memory shortage
1432 * but because memory controller hits its limit.
1433 * Don't modify zone reclaim related data.
1434 */
1435 nr[l] = mem_cgroup_calc_reclaim(sc->mem_cgroup, zone,
1436 priority, l);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001437 }
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001438 }
1439
Rik van Riel556adec2008-10-18 20:26:34 -07001440 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1441 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001442 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001443 if (nr[l]) {
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001444 nr_to_scan = min(nr[l], swap_cluster_max);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001445 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001447 nr_reclaimed += shrink_list(l, nr_to_scan,
1448 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001449 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001450 }
Rik van Riela79311c2009-01-06 14:40:01 -08001451 /*
1452 * On large memory systems, scan >> priority can become
1453 * really large. This is fine for the starting priority;
1454 * we want to put equal scanning pressure on each zone.
1455 * However, if the VM has a harder time of freeing pages,
1456 * with multiple processes reclaiming pages, the total
1457 * freeing target can get unreasonably large.
1458 */
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001459 if (nr_reclaimed > swap_cluster_max &&
Rik van Riela79311c2009-01-06 14:40:01 -08001460 priority < DEF_PRIORITY && !current_is_kswapd())
1461 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001462 }
1463
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001464 sc->nr_reclaimed = nr_reclaimed;
1465
Rik van Riel556adec2008-10-18 20:26:34 -07001466 /*
1467 * Even if we did not try to evict anon pages at all, we want to
1468 * rebalance the anon lru active/inactive ratio.
1469 */
1470 if (!scan_global_lru(sc) || inactive_anon_is_low(zone))
1471 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1472 else if (!scan_global_lru(sc))
1473 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1474
Andrew Morton232ea4d2007-02-28 20:13:21 -08001475 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001476}
1477
1478/*
1479 * This is the direct reclaim path, for page-allocating processes. We only
1480 * try to reclaim pages from zones which will satisfy the caller's allocation
1481 * request.
1482 *
1483 * We reclaim from a zone even if that zone is over pages_high. Because:
1484 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1485 * allocation or
1486 * b) The zones may be over pages_high but they must go *over* pages_high to
1487 * satisfy the `incremental min' zone defense algorithm.
1488 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489 * If a zone is deemed to be full of pinned pages then just give it a light
1490 * scan then give up on it.
1491 */
Rik van Riela79311c2009-01-06 14:40:01 -08001492static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001493 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001494{
Mel Gorman54a6eb52008-04-28 02:12:16 -07001495 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Mel Gormandd1a2392008-04-28 02:12:17 -07001496 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001497 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001498
Nick Piggin408d8542006-09-25 23:31:27 -07001499 sc->all_unreclaimable = 1;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001500 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001501 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001503 /*
1504 * Take care memory controller reclaiming has small influence
1505 * to global LRU.
1506 */
1507 if (scan_global_lru(sc)) {
1508 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1509 continue;
1510 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001512 if (zone_is_all_unreclaimable(zone) &&
1513 priority != DEF_PRIORITY)
1514 continue; /* Let kswapd poll it */
1515 sc->all_unreclaimable = 0;
1516 } else {
1517 /*
1518 * Ignore cpuset limitation here. We just want to reduce
1519 * # of used pages by us regardless of memory shortage.
1520 */
1521 sc->all_unreclaimable = 0;
1522 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1523 priority);
1524 }
Nick Piggin408d8542006-09-25 23:31:27 -07001525
Rik van Riela79311c2009-01-06 14:40:01 -08001526 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527 }
1528}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001529
Linus Torvalds1da177e2005-04-16 15:20:36 -07001530/*
1531 * This is the main entry point to direct page reclaim.
1532 *
1533 * If a full scan of the inactive list fails to free enough memory then we
1534 * are "out of memory" and something needs to be killed.
1535 *
1536 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1537 * high - the zone may be full of dirty or under-writeback pages, which this
1538 * caller can't do much about. We kick pdflush and take explicit naps in the
1539 * hope that some of these pages can be written. But if the allocating task
1540 * holds filesystem locks which prevent writeout this might not work, and the
1541 * allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001542 *
1543 * returns: 0, if no pages reclaimed
1544 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 */
Mel Gormandac1d272008-04-28 02:12:12 -07001546static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001547 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548{
1549 int priority;
kosaki.motohiro@jp.fujitsu.comc700be32008-06-12 15:21:27 -07001550 unsigned long ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001551 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001553 unsigned long lru_pages = 0;
Mel Gormandd1a2392008-04-28 02:12:17 -07001554 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001555 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -07001556 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001557
Keika Kobayashi873b4772008-07-25 01:48:52 -07001558 delayacct_freepages_start();
1559
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001560 if (scan_global_lru(sc))
1561 count_vm_event(ALLOCSTALL);
1562 /*
1563 * mem_cgroup will not do shrink_slab.
1564 */
1565 if (scan_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001566 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001568 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1569 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001570
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001571 lru_pages += zone_lru_pages(zone);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001572 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573 }
1574
1575 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001576 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001577 if (!priority)
1578 disable_swap_token();
Rik van Riela79311c2009-01-06 14:40:01 -08001579 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001580 /*
1581 * Don't shrink slabs when reclaiming memory from
1582 * over limit cgroups
1583 */
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001584 if (scan_global_lru(sc)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07001585 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001586 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08001587 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001588 reclaim_state->reclaimed_slab = 0;
1589 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590 }
Balbir Singh66e17072008-02-07 00:13:56 -08001591 total_scanned += sc->nr_scanned;
Rik van Riela79311c2009-01-06 14:40:01 -08001592 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1593 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594 goto out;
1595 }
1596
1597 /*
1598 * Try to write back as many pages as we just scanned. This
1599 * tends to cause slow streaming writers to write data to the
1600 * disk smoothly, at the dirtying rate, which is nice. But
1601 * that's undesirable in laptop mode, where we *want* lumpy
1602 * writeout. So in laptop mode, write out the whole world.
1603 */
Balbir Singh66e17072008-02-07 00:13:56 -08001604 if (total_scanned > sc->swap_cluster_max +
1605 sc->swap_cluster_max / 2) {
Pekka J Enberg687a21c2005-06-28 20:44:55 -07001606 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001607 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 }
1609
1610 /* Take a nap, wait for some writeback to complete */
Andrew Morton4dd4b922008-03-24 12:29:52 -07001611 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001612 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 }
Fernando Luis Vazquez Cao87547ee2008-07-29 22:33:42 -07001614 /* top priority shrink_zones still had more to do? don't OOM, then */
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001615 if (!sc->all_unreclaimable && scan_global_lru(sc))
Rik van Riela79311c2009-01-06 14:40:01 -08001616 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001617out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07001618 /*
1619 * Now that we've scanned all the zones at this priority level, note
1620 * that level within the zone so that the next thread which performs
1621 * scanning of this zone will immediately start out at this priority
1622 * level. This affects only the decision whether or not to bring
1623 * mapped pages onto the inactive list.
1624 */
1625 if (priority < 0)
1626 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001628 if (scan_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001629 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001631 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1632 continue;
1633
1634 zone->prev_priority = priority;
1635 }
1636 } else
1637 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1638
Keika Kobayashi873b4772008-07-25 01:48:52 -07001639 delayacct_freepages_end();
1640
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 return ret;
1642}
1643
Mel Gormandac1d272008-04-28 02:12:12 -07001644unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1645 gfp_t gfp_mask)
Balbir Singh66e17072008-02-07 00:13:56 -08001646{
1647 struct scan_control sc = {
1648 .gfp_mask = gfp_mask,
1649 .may_writepage = !laptop_mode,
1650 .swap_cluster_max = SWAP_CLUSTER_MAX,
1651 .may_swap = 1,
1652 .swappiness = vm_swappiness,
1653 .order = order,
1654 .mem_cgroup = NULL,
1655 .isolate_pages = isolate_pages_global,
1656 };
1657
Mel Gormandd1a2392008-04-28 02:12:17 -07001658 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001659}
1660
Balbir Singh00f0b822008-03-04 14:28:39 -08001661#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08001662
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001663unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1664 gfp_t gfp_mask)
Balbir Singh66e17072008-02-07 00:13:56 -08001665{
1666 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08001667 .may_writepage = !laptop_mode,
1668 .may_swap = 1,
1669 .swap_cluster_max = SWAP_CLUSTER_MAX,
1670 .swappiness = vm_swappiness,
1671 .order = 0,
1672 .mem_cgroup = mem_cont,
1673 .isolate_pages = mem_cgroup_isolate_pages,
1674 };
Mel Gormandac1d272008-04-28 02:12:12 -07001675 struct zonelist *zonelist;
Balbir Singh66e17072008-02-07 00:13:56 -08001676
Mel Gormandd1a2392008-04-28 02:12:17 -07001677 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1678 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1679 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1680 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001681}
1682#endif
1683
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684/*
1685 * For kswapd, balance_pgdat() will work across all this node's zones until
1686 * they are all at pages_high.
1687 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 * Returns the number of pages which were actually freed.
1689 *
1690 * There is special handling here for zones which are full of pinned pages.
1691 * This can happen if the pages are all mlocked, or if they are all used by
1692 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1693 * What we do is to detect the case where all pages in the zone have been
1694 * scanned twice and there has been zero successful reclaim. Mark the zone as
1695 * dead and from now on, only perform a short scan. Basically we're polling
1696 * the zone for when the problem goes away.
1697 *
1698 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1699 * zones which have free_pages > pages_high, but once a zone is found to have
1700 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1701 * of the number of free pages in the lower zones. This interoperates with
1702 * the page allocator fallback scheme to ensure that aging of pages is balanced
1703 * across the zones.
1704 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001705static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707 int all_zones_ok;
1708 int priority;
1709 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08001710 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08001712 struct scan_control sc = {
1713 .gfp_mask = GFP_KERNEL,
1714 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001715 .swap_cluster_max = SWAP_CLUSTER_MAX,
1716 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001717 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08001718 .mem_cgroup = NULL,
1719 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08001720 };
Martin Bligh3bb1a852006-10-28 10:38:24 -07001721 /*
1722 * temp_priority is used to remember the scanning priority at which
1723 * this zone was successfully refilled to free_pages == pages_high.
1724 */
1725 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726
1727loop_again:
1728 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08001729 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07001730 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001731 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732
Martin Bligh3bb1a852006-10-28 10:38:24 -07001733 for (i = 0; i < pgdat->nr_zones; i++)
1734 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735
1736 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1737 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1738 unsigned long lru_pages = 0;
1739
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001740 /* The swap token gets in the way of swapout... */
1741 if (!priority)
1742 disable_swap_token();
1743
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744 all_zones_ok = 1;
1745
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001746 /*
1747 * Scan in the highmem->dma direction for the highest
1748 * zone which needs scanning
1749 */
1750 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1751 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001753 if (!populated_zone(zone))
1754 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001755
David Rientjese815af92007-10-16 23:25:54 -07001756 if (zone_is_all_unreclaimable(zone) &&
1757 priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001758 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759
Rik van Riel556adec2008-10-18 20:26:34 -07001760 /*
1761 * Do some background aging of the anon list, to give
1762 * pages a chance to be referenced before reclaiming.
1763 */
1764 if (inactive_anon_is_low(zone))
1765 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1766 &sc, priority, 0);
1767
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001768 if (!zone_watermark_ok(zone, order, zone->pages_high,
1769 0, 0)) {
1770 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001771 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001772 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001773 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001774 if (i < 0)
1775 goto out;
1776
Linus Torvalds1da177e2005-04-16 15:20:36 -07001777 for (i = 0; i <= end_zone; i++) {
1778 struct zone *zone = pgdat->node_zones + i;
1779
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001780 lru_pages += zone_lru_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781 }
1782
1783 /*
1784 * Now scan the zone in the dma->highmem direction, stopping
1785 * at the last zone which needs scanning.
1786 *
1787 * We do this because the page allocator works in the opposite
1788 * direction. This prevents the page allocator from allocating
1789 * pages behind kswapd's direction of progress, which would
1790 * cause too much scanning of the lower zones.
1791 */
1792 for (i = 0; i <= end_zone; i++) {
1793 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001794 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795
Con Kolivasf3fe6512006-01-06 00:11:15 -08001796 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 continue;
1798
David Rientjese815af92007-10-16 23:25:54 -07001799 if (zone_is_all_unreclaimable(zone) &&
1800 priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801 continue;
1802
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001803 if (!zone_watermark_ok(zone, order, zone->pages_high,
1804 end_zone, 0))
1805 all_zones_ok = 0;
Martin Bligh3bb1a852006-10-28 10:38:24 -07001806 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807 sc.nr_scanned = 0;
Martin Bligh3bb1a852006-10-28 10:38:24 -07001808 note_zone_scanning_priority(zone, priority);
Rik van Riel32a43302007-10-16 01:24:50 -07001809 /*
1810 * We put equal pressure on every zone, unless one
1811 * zone has way too many pages free already.
1812 */
1813 if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1814 end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08001815 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001816 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001817 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1818 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08001819 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001820 total_scanned += sc.nr_scanned;
David Rientjese815af92007-10-16 23:25:54 -07001821 if (zone_is_all_unreclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001822 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001823 if (nr_slab == 0 && zone->pages_scanned >=
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001824 (zone_lru_pages(zone) * 6))
David Rientjese815af92007-10-16 23:25:54 -07001825 zone_set_flag(zone,
1826 ZONE_ALL_UNRECLAIMABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 /*
1828 * If we've done a decent amount of scanning and
1829 * the reclaim ratio is low, start doing writepage
1830 * even in laptop mode
1831 */
1832 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08001833 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001834 sc.may_writepage = 1;
1835 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 if (all_zones_ok)
1837 break; /* kswapd: all done */
1838 /*
1839 * OK, kswapd is getting into trouble. Take a nap, then take
1840 * another pass across the zones.
1841 */
Andrew Morton4dd4b922008-03-24 12:29:52 -07001842 if (total_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001843 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001844
1845 /*
1846 * We do this so kswapd doesn't build up large priorities for
1847 * example when it is freeing in parallel with allocators. It
1848 * matches the direct reclaim path behaviour in terms of impact
1849 * on zone->*_priority.
1850 */
Rik van Riela79311c2009-01-06 14:40:01 -08001851 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 break;
1853 }
1854out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07001855 /*
1856 * Note within each zone the priority level at which this zone was
1857 * brought into a happy state. So that the next thread which scans this
1858 * zone will start out at that priority level.
1859 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860 for (i = 0; i < pgdat->nr_zones; i++) {
1861 struct zone *zone = pgdat->node_zones + i;
1862
Martin Bligh3bb1a852006-10-28 10:38:24 -07001863 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 }
1865 if (!all_zones_ok) {
1866 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08001867
1868 try_to_freeze();
1869
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08001870 /*
1871 * Fragmentation may mean that the system cannot be
1872 * rebalanced for high-order allocations in all zones.
1873 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1874 * it means the zones have been fully scanned and are still
1875 * not balanced. For high-order allocations, there is
1876 * little point trying all over again as kswapd may
1877 * infinite loop.
1878 *
1879 * Instead, recheck all watermarks at order-0 as they
1880 * are the most important. If watermarks are ok, kswapd will go
1881 * back to sleep. High-order users can still perform direct
1882 * reclaim if they wish.
1883 */
1884 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
1885 order = sc.order = 0;
1886
Linus Torvalds1da177e2005-04-16 15:20:36 -07001887 goto loop_again;
1888 }
1889
Rik van Riela79311c2009-01-06 14:40:01 -08001890 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891}
1892
1893/*
1894 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001895 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 *
1897 * This basically trickles out pages so that we have _some_
1898 * free memory available even if there is no other activity
1899 * that frees anything up. This is needed for things like routing
1900 * etc, where we otherwise might have all activity going on in
1901 * asynchronous contexts that cannot page things out.
1902 *
1903 * If there are applications that are active memory-allocators
1904 * (most normal use), this basically shouldn't matter.
1905 */
1906static int kswapd(void *p)
1907{
1908 unsigned long order;
1909 pg_data_t *pgdat = (pg_data_t*)p;
1910 struct task_struct *tsk = current;
1911 DEFINE_WAIT(wait);
1912 struct reclaim_state reclaim_state = {
1913 .reclaimed_slab = 0,
1914 };
Mike Travisc5f59f02008-04-04 18:11:10 -07001915 node_to_cpumask_ptr(cpumask, pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916
Rusty Russell174596a2009-01-01 10:12:29 +10301917 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07001918 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 current->reclaim_state = &reclaim_state;
1920
1921 /*
1922 * Tell the memory management that we're a "memory allocator",
1923 * and that if we need more memory we should get access to it
1924 * regardless (see "__alloc_pages()"). "kswapd" should
1925 * never get caught in the normal page freeing logic.
1926 *
1927 * (Kswapd normally doesn't need memory anyway, but sometimes
1928 * you need a small amount of memory in order to be able to
1929 * page out something else, and this flag essentially protects
1930 * us from recursively trying to free more memory as we're
1931 * trying to free the first piece of memory in the first place).
1932 */
Christoph Lameter930d9152006-01-08 01:00:47 -08001933 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07001934 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001935
1936 order = 0;
1937 for ( ; ; ) {
1938 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07001939
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1941 new_order = pgdat->kswapd_max_order;
1942 pgdat->kswapd_max_order = 0;
1943 if (order < new_order) {
1944 /*
1945 * Don't sleep if someone wants a larger 'order'
1946 * allocation
1947 */
1948 order = new_order;
1949 } else {
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07001950 if (!freezing(current))
1951 schedule();
1952
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953 order = pgdat->kswapd_max_order;
1954 }
1955 finish_wait(&pgdat->kswapd_wait, &wait);
1956
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07001957 if (!try_to_freeze()) {
1958 /* We can speed up thawing tasks if we don't call
1959 * balance_pgdat after returning from the refrigerator
1960 */
1961 balance_pgdat(pgdat, order);
1962 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963 }
1964 return 0;
1965}
1966
1967/*
1968 * A zone is low on free memory, so wake its kswapd task to service it.
1969 */
1970void wakeup_kswapd(struct zone *zone, int order)
1971{
1972 pg_data_t *pgdat;
1973
Con Kolivasf3fe6512006-01-06 00:11:15 -08001974 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975 return;
1976
1977 pgdat = zone->zone_pgdat;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001978 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 return;
1980 if (pgdat->kswapd_max_order < order)
1981 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08001982 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001984 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001986 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987}
1988
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001989unsigned long global_lru_pages(void)
1990{
1991 return global_page_state(NR_ACTIVE_ANON)
1992 + global_page_state(NR_ACTIVE_FILE)
1993 + global_page_state(NR_INACTIVE_ANON)
1994 + global_page_state(NR_INACTIVE_FILE);
1995}
1996
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997#ifdef CONFIG_PM
1998/*
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001999 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
2000 * from LRU lists system-wide, for given pass and priority, and returns the
2001 * number of reclaimed pages
2002 *
2003 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2004 */
Nigel Cunninghame07aa052006-12-22 01:07:21 -08002005static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
2006 int pass, struct scan_control *sc)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002007{
2008 struct zone *zone;
2009 unsigned long nr_to_scan, ret = 0;
Christoph Lameterb69408e2008-10-18 20:26:14 -07002010 enum lru_list l;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002011
2012 for_each_zone(zone) {
2013
2014 if (!populated_zone(zone))
2015 continue;
2016
David Rientjese815af92007-10-16 23:25:54 -07002017 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002018 continue;
2019
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002020 for_each_evictable_lru(l) {
2021 /* For pass = 0, we don't shrink the active list */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002022 if (pass == 0 &&
2023 (l == LRU_ACTIVE || l == LRU_ACTIVE_FILE))
Christoph Lameterb69408e2008-10-18 20:26:14 -07002024 continue;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002025
Christoph Lameterb69408e2008-10-18 20:26:14 -07002026 zone->lru[l].nr_scan +=
2027 (zone_page_state(zone, NR_LRU_BASE + l)
2028 >> prio) + 1;
2029 if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
2030 zone->lru[l].nr_scan = 0;
2031 nr_to_scan = min(nr_pages,
2032 zone_page_state(zone,
2033 NR_LRU_BASE + l));
2034 ret += shrink_list(l, nr_to_scan, zone,
2035 sc, prio);
2036 if (ret >= nr_pages)
2037 return ret;
2038 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002039 }
2040 }
2041
2042 return ret;
2043}
2044
2045/*
2046 * Try to free `nr_pages' of memory, system-wide, and return the number of
2047 * freed pages.
2048 *
2049 * Rather than trying to age LRUs the aim is to preserve the overall
2050 * LRU order by reclaiming preferentially
2051 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 */
Andrew Morton69e05942006-03-22 00:08:19 -08002053unsigned long shrink_all_memory(unsigned long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002055 unsigned long lru_pages, nr_slab;
Andrew Morton69e05942006-03-22 00:08:19 -08002056 unsigned long ret = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002057 int pass;
2058 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002059 struct scan_control sc = {
2060 .gfp_mask = GFP_KERNEL,
2061 .may_swap = 0,
2062 .swap_cluster_max = nr_pages,
2063 .may_writepage = 1,
2064 .swappiness = vm_swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08002065 .isolate_pages = isolate_pages_global,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002066 };
2067
2068 current->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002069
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002070 lru_pages = global_lru_pages();
Christoph Lameter972d1a72006-09-25 23:31:51 -07002071 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002072 /* If slab caches are huge, it's better to hit them first */
2073 while (nr_slab >= lru_pages) {
2074 reclaim_state.reclaimed_slab = 0;
2075 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2076 if (!reclaim_state.reclaimed_slab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 break;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002078
2079 ret += reclaim_state.reclaimed_slab;
2080 if (ret >= nr_pages)
2081 goto out;
2082
2083 nr_slab -= reclaim_state.reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002085
2086 /*
2087 * We try to shrink LRUs in 5 passes:
2088 * 0 = Reclaim from inactive_list only
2089 * 1 = Reclaim from active list but don't reclaim mapped
2090 * 2 = 2nd pass of type 1
2091 * 3 = Reclaim mapped (normal reclaim)
2092 * 4 = 2nd pass of type 3
2093 */
2094 for (pass = 0; pass < 5; pass++) {
2095 int prio;
2096
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002097 /* Force reclaiming mapped pages in the passes #3 and #4 */
2098 if (pass > 2) {
2099 sc.may_swap = 1;
2100 sc.swappiness = 100;
2101 }
2102
2103 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
2104 unsigned long nr_to_scan = nr_pages - ret;
2105
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002106 sc.nr_scanned = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002107 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
2108 if (ret >= nr_pages)
2109 goto out;
2110
2111 reclaim_state.reclaimed_slab = 0;
Andrew Morton76395d32007-01-05 16:37:05 -08002112 shrink_slab(sc.nr_scanned, sc.gfp_mask,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002113 global_lru_pages());
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002114 ret += reclaim_state.reclaimed_slab;
2115 if (ret >= nr_pages)
2116 goto out;
2117
2118 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07002119 congestion_wait(WRITE, HZ / 10);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002120 }
Rafael J. Wysocki248a0302006-03-22 00:09:04 -08002121 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002122
2123 /*
2124 * If ret = 0, we could not shrink LRUs, but there may be something
2125 * in slab caches
2126 */
Andrew Morton76395d32007-01-05 16:37:05 -08002127 if (!ret) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002128 do {
2129 reclaim_state.reclaimed_slab = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002130 shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002131 ret += reclaim_state.reclaimed_slab;
2132 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
Andrew Morton76395d32007-01-05 16:37:05 -08002133 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002134
2135out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136 current->reclaim_state = NULL;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002137
Linus Torvalds1da177e2005-04-16 15:20:36 -07002138 return ret;
2139}
2140#endif
2141
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142/* It's optimal to keep kswapds on the same CPUs as their memory, but
2143 not required for correctness. So if the last cpu in a node goes
2144 away, we get changed to run anywhere: as the first one comes back,
2145 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002146static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002147 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002149 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002151 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002152 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002153 pg_data_t *pgdat = NODE_DATA(nid);
2154 node_to_cpumask_ptr(mask, pgdat->node_id);
2155
Rusty Russell3e597942009-01-01 10:12:24 +10302156 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002158 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002159 }
2160 }
2161 return NOTIFY_OK;
2162}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163
Yasunori Goto3218ae12006-06-27 02:53:33 -07002164/*
2165 * This kswapd start function will be called by init and node-hot-add.
2166 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2167 */
2168int kswapd_run(int nid)
2169{
2170 pg_data_t *pgdat = NODE_DATA(nid);
2171 int ret = 0;
2172
2173 if (pgdat->kswapd)
2174 return 0;
2175
2176 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2177 if (IS_ERR(pgdat->kswapd)) {
2178 /* failure at boot is fatal */
2179 BUG_ON(system_state == SYSTEM_BOOTING);
2180 printk("Failed to start kswapd on node %d\n",nid);
2181 ret = -1;
2182 }
2183 return ret;
2184}
2185
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186static int __init kswapd_init(void)
2187{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002188 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002189
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002191 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002192 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193 hotcpu_notifier(cpu_callback, 0);
2194 return 0;
2195}
2196
2197module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002198
2199#ifdef CONFIG_NUMA
2200/*
2201 * Zone reclaim mode
2202 *
2203 * If non-zero call zone_reclaim when the number of free pages falls below
2204 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002205 */
2206int zone_reclaim_mode __read_mostly;
2207
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002208#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002209#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002210#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2211#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2212
Christoph Lameter9eeff232006-01-18 17:42:31 -08002213/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002214 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2215 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2216 * a zone.
2217 */
2218#define ZONE_RECLAIM_PRIORITY 4
2219
Christoph Lameter9eeff232006-01-18 17:42:31 -08002220/*
Christoph Lameter96146342006-07-03 00:24:13 -07002221 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2222 * occur.
2223 */
2224int sysctl_min_unmapped_ratio = 1;
2225
2226/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002227 * If the number of slab pages in a zone grows beyond this percentage then
2228 * slab reclaim needs to occur.
2229 */
2230int sysctl_min_slab_ratio = 5;
2231
2232/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002233 * Try to free up some pages from this zone through reclaim.
2234 */
Andrew Morton179e9632006-03-22 00:08:18 -08002235static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002236{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002237 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002238 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002239 struct task_struct *p = current;
2240 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002241 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002242 struct scan_control sc = {
2243 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2244 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
Andrew Morton69e05942006-03-22 00:08:19 -08002245 .swap_cluster_max = max_t(unsigned long, nr_pages,
2246 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002247 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002248 .swappiness = vm_swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08002249 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08002250 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07002251 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002252
2253 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08002254 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002255 /*
2256 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2257 * and we also need to be able to write out pages for RECLAIM_WRITE
2258 * and RECLAIM_SWAP.
2259 */
2260 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002261 reclaim_state.reclaimed_slab = 0;
2262 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002263
Christoph Lameter0ff38492006-09-25 23:31:52 -07002264 if (zone_page_state(zone, NR_FILE_PAGES) -
2265 zone_page_state(zone, NR_FILE_MAPPED) >
2266 zone->min_unmapped_pages) {
2267 /*
2268 * Free memory by calling shrink zone with increasing
2269 * priorities until we have enough memory freed.
2270 */
2271 priority = ZONE_RECLAIM_PRIORITY;
2272 do {
Martin Bligh3bb1a852006-10-28 10:38:24 -07002273 note_zone_scanning_priority(zone, priority);
Rik van Riela79311c2009-01-06 14:40:01 -08002274 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002275 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002276 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002277 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002278
Christoph Lameter83e33a42006-09-25 23:31:53 -07002279 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2280 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002281 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002282 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002283 * many pages were freed in this zone. So we take the current
2284 * number of slab pages and shake the slab until it is reduced
2285 * by the same nr_pages that we used for reclaiming unmapped
2286 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002287 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002288 * Note that shrink_slab will free memory on all zones and may
2289 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002290 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07002291 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07002292 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2293 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07002294 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07002295
2296 /*
2297 * Update nr_reclaimed by the number of slab pages we
2298 * reclaimed from this zone.
2299 */
Rik van Riela79311c2009-01-06 14:40:01 -08002300 sc.nr_reclaimed += slab_reclaimable -
Christoph Lameter83e33a42006-09-25 23:31:53 -07002301 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002302 }
2303
Christoph Lameter9eeff232006-01-18 17:42:31 -08002304 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002305 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
Rik van Riela79311c2009-01-06 14:40:01 -08002306 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002307}
Andrew Morton179e9632006-03-22 00:08:18 -08002308
2309int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2310{
Andrew Morton179e9632006-03-22 00:08:18 -08002311 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002312 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002313
2314 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002315 * Zone reclaim reclaims unmapped file backed pages and
2316 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002317 *
Christoph Lameter96146342006-07-03 00:24:13 -07002318 * A small portion of unmapped file backed pages is needed for
2319 * file I/O otherwise pages read by file I/O will be immediately
2320 * thrown out if the zone is overallocated. So we do not reclaim
2321 * if less than a specified percentage of the zone is used by
2322 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002323 */
Christoph Lameter34aa1332006-06-30 01:55:37 -07002324 if (zone_page_state(zone, NR_FILE_PAGES) -
Christoph Lameter0ff38492006-09-25 23:31:52 -07002325 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2326 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2327 <= zone->min_slab_pages)
Christoph Lameter96146342006-07-03 00:24:13 -07002328 return 0;
Andrew Morton179e9632006-03-22 00:08:18 -08002329
David Rientjesd773ed62007-10-16 23:26:01 -07002330 if (zone_is_all_unreclaimable(zone))
2331 return 0;
2332
Andrew Morton179e9632006-03-22 00:08:18 -08002333 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002334 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002335 */
David Rientjesd773ed62007-10-16 23:26:01 -07002336 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Andrew Morton179e9632006-03-22 00:08:18 -08002337 return 0;
2338
2339 /*
2340 * Only run zone reclaim on the local zone or on zones that do not
2341 * have associated processors. This will favor the local processor
2342 * over remote processors and spread off node memory allocations
2343 * as wide as possible.
2344 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002345 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002346 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Andrew Morton179e9632006-03-22 00:08:18 -08002347 return 0;
David Rientjesd773ed62007-10-16 23:26:01 -07002348
2349 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2350 return 0;
2351 ret = __zone_reclaim(zone, gfp_mask, order);
2352 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2353
2354 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002355}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002356#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002357
2358#ifdef CONFIG_UNEVICTABLE_LRU
2359/*
2360 * page_evictable - test whether a page is evictable
2361 * @page: the page to test
2362 * @vma: the VMA in which the page is or will be mapped, may be NULL
2363 *
2364 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07002365 * lists vs unevictable list. The vma argument is !NULL when called from the
2366 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002367 *
2368 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002369 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07002370 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002371 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002372 */
2373int page_evictable(struct page *page, struct vm_area_struct *vma)
2374{
2375
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002376 if (mapping_unevictable(page_mapping(page)))
2377 return 0;
2378
Nick Pigginb291f002008-10-18 20:26:44 -07002379 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2380 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002381
2382 return 1;
2383}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002384
2385/**
2386 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2387 * @page: page to check evictability and move to appropriate lru list
2388 * @zone: zone page is in
2389 *
2390 * Checks a page for evictability and moves the page to the appropriate
2391 * zone lru list.
2392 *
2393 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2394 * have PageUnevictable set.
2395 */
2396static void check_move_unevictable_page(struct page *page, struct zone *zone)
2397{
2398 VM_BUG_ON(PageActive(page));
2399
2400retry:
2401 ClearPageUnevictable(page);
2402 if (page_evictable(page, NULL)) {
2403 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002404
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002405 __dec_zone_state(zone, NR_UNEVICTABLE);
2406 list_move(&page->lru, &zone->lru[l].list);
2407 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2408 __count_vm_event(UNEVICTABLE_PGRESCUED);
2409 } else {
2410 /*
2411 * rotate unevictable list
2412 */
2413 SetPageUnevictable(page);
2414 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
2415 if (page_evictable(page, NULL))
2416 goto retry;
2417 }
2418}
2419
2420/**
2421 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2422 * @mapping: struct address_space to scan for evictable pages
2423 *
2424 * Scan all pages in mapping. Check unevictable pages for
2425 * evictability and move them to the appropriate zone lru list.
2426 */
2427void scan_mapping_unevictable_pages(struct address_space *mapping)
2428{
2429 pgoff_t next = 0;
2430 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2431 PAGE_CACHE_SHIFT;
2432 struct zone *zone;
2433 struct pagevec pvec;
2434
2435 if (mapping->nrpages == 0)
2436 return;
2437
2438 pagevec_init(&pvec, 0);
2439 while (next < end &&
2440 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2441 int i;
2442 int pg_scanned = 0;
2443
2444 zone = NULL;
2445
2446 for (i = 0; i < pagevec_count(&pvec); i++) {
2447 struct page *page = pvec.pages[i];
2448 pgoff_t page_index = page->index;
2449 struct zone *pagezone = page_zone(page);
2450
2451 pg_scanned++;
2452 if (page_index > next)
2453 next = page_index;
2454 next++;
2455
2456 if (pagezone != zone) {
2457 if (zone)
2458 spin_unlock_irq(&zone->lru_lock);
2459 zone = pagezone;
2460 spin_lock_irq(&zone->lru_lock);
2461 }
2462
2463 if (PageLRU(page) && PageUnevictable(page))
2464 check_move_unevictable_page(page, zone);
2465 }
2466 if (zone)
2467 spin_unlock_irq(&zone->lru_lock);
2468 pagevec_release(&pvec);
2469
2470 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2471 }
2472
2473}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002474
2475/**
2476 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2477 * @zone - zone of which to scan the unevictable list
2478 *
2479 * Scan @zone's unevictable LRU lists to check for pages that have become
2480 * evictable. Move those that have to @zone's inactive list where they
2481 * become candidates for reclaim, unless shrink_inactive_zone() decides
2482 * to reactivate them. Pages that are still unevictable are rotated
2483 * back onto @zone's unevictable list.
2484 */
2485#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08002486static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002487{
2488 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2489 unsigned long scan;
2490 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2491
2492 while (nr_to_scan > 0) {
2493 unsigned long batch_size = min(nr_to_scan,
2494 SCAN_UNEVICTABLE_BATCH_SIZE);
2495
2496 spin_lock_irq(&zone->lru_lock);
2497 for (scan = 0; scan < batch_size; scan++) {
2498 struct page *page = lru_to_page(l_unevictable);
2499
2500 if (!trylock_page(page))
2501 continue;
2502
2503 prefetchw_prev_lru_page(page, l_unevictable, flags);
2504
2505 if (likely(PageLRU(page) && PageUnevictable(page)))
2506 check_move_unevictable_page(page, zone);
2507
2508 unlock_page(page);
2509 }
2510 spin_unlock_irq(&zone->lru_lock);
2511
2512 nr_to_scan -= batch_size;
2513 }
2514}
2515
2516
2517/**
2518 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2519 *
2520 * A really big hammer: scan all zones' unevictable LRU lists to check for
2521 * pages that have become evictable. Move those back to the zones'
2522 * inactive list where they become candidates for reclaim.
2523 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2524 * and we add swap to the system. As such, it runs in the context of a task
2525 * that has possibly/probably made some previously unevictable pages
2526 * evictable.
2527 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08002528static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002529{
2530 struct zone *zone;
2531
2532 for_each_zone(zone) {
2533 scan_zone_unevictable_pages(zone);
2534 }
2535}
2536
2537/*
2538 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2539 * all nodes' unevictable lists for evictable pages
2540 */
2541unsigned long scan_unevictable_pages;
2542
2543int scan_unevictable_handler(struct ctl_table *table, int write,
2544 struct file *file, void __user *buffer,
2545 size_t *length, loff_t *ppos)
2546{
2547 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2548
2549 if (write && *(unsigned long *)table->data)
2550 scan_all_zones_unevictable_pages();
2551
2552 scan_unevictable_pages = 0;
2553 return 0;
2554}
2555
2556/*
2557 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2558 * a specified node's per zone unevictable lists for evictable pages.
2559 */
2560
2561static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2562 struct sysdev_attribute *attr,
2563 char *buf)
2564{
2565 return sprintf(buf, "0\n"); /* always zero; should fit... */
2566}
2567
2568static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2569 struct sysdev_attribute *attr,
2570 const char *buf, size_t count)
2571{
2572 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2573 struct zone *zone;
2574 unsigned long res;
2575 unsigned long req = strict_strtoul(buf, 10, &res);
2576
2577 if (!req)
2578 return 1; /* zero is no-op */
2579
2580 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2581 if (!populated_zone(zone))
2582 continue;
2583 scan_zone_unevictable_pages(zone);
2584 }
2585 return 1;
2586}
2587
2588
2589static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2590 read_scan_unevictable_node,
2591 write_scan_unevictable_node);
2592
2593int scan_unevictable_register_node(struct node *node)
2594{
2595 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2596}
2597
2598void scan_unevictable_unregister_node(struct node *node)
2599{
2600 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2601}
2602
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002603#endif