blob: ece2ecb081025dd5f1ebf74ed24c0ce70bd5a4bf [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
Johannes Weinera6dc60f82009-03-31 15:19:30 -070063 /* Can mapped pages be reclaimed? */
64 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080065
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070066 /* Can pages be swapped as part of reclaim? */
67 int may_swap;
68
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
70 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
71 * In this context, it doesn't matter that we scan the
72 * whole list at once. */
73 int swap_cluster_max;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070074
75 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070076
77 int all_unreclaimable;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070078
79 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080080
81 /* Which cgroup do we reclaim from */
82 struct mem_cgroup *mem_cgroup;
83
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -070084 /*
85 * Nodemask of nodes allowed by the caller. If NULL, all nodes
86 * are scanned.
87 */
88 nodemask_t *nodemask;
89
Balbir Singh66e17072008-02-07 00:13:56 -080090 /* Pluggable isolate pages callback */
91 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
92 unsigned long *scanned, int order, int mode,
93 struct zone *z, struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -070094 int active, int file);
Linus Torvalds1da177e2005-04-16 15:20:36 -070095};
96
Linus Torvalds1da177e2005-04-16 15:20:36 -070097#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99#ifdef ARCH_HAS_PREFETCH
100#define prefetch_prev_lru_page(_page, _base, _field) \
101 do { \
102 if ((_page)->lru.prev != _base) { \
103 struct page *prev; \
104 \
105 prev = lru_to_page(&(_page->lru)); \
106 prefetch(&prev->_field); \
107 } \
108 } while (0)
109#else
110#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111#endif
112
113#ifdef ARCH_HAS_PREFETCHW
114#define prefetchw_prev_lru_page(_page, _base, _field) \
115 do { \
116 if ((_page)->lru.prev != _base) { \
117 struct page *prev; \
118 \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetchw(&prev->_field); \
121 } \
122 } while (0)
123#else
124#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125#endif
126
127/*
128 * From 0 .. 100. Higher means more swappy.
129 */
130int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700131long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132
133static LIST_HEAD(shrinker_list);
134static DECLARE_RWSEM(shrinker_rwsem);
135
Balbir Singh00f0b822008-03-04 14:28:39 -0800136#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800137#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800138#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800139#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800140#endif
141
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800142static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
143 struct scan_control *sc)
144{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800145 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800146 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
147
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800148 return &zone->reclaim_stat;
149}
150
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800151static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc,
152 enum lru_list lru)
153{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800154 if (!scanning_global_lru(sc))
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800155 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
156
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800157 return zone_page_state(zone, NR_LRU_BASE + lru);
158}
159
160
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161/*
162 * Add a shrinker callback to be called from the vm
163 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700164void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700165{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700166 shrinker->nr = 0;
167 down_write(&shrinker_rwsem);
168 list_add_tail(&shrinker->list, &shrinker_list);
169 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700171EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700172
173/*
174 * Remove one
175 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700176void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177{
178 down_write(&shrinker_rwsem);
179 list_del(&shrinker->list);
180 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700182EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183
184#define SHRINK_BATCH 128
185/*
186 * Call the shrink functions to age shrinkable caches
187 *
188 * Here we assume it costs one seek to replace a lru page and that it also
189 * takes a seek to recreate a cache object. With this in mind we age equal
190 * percentages of the lru and ageable caches. This should balance the seeks
191 * generated by these structures.
192 *
Simon Arlott183ff222007-10-20 01:27:18 +0200193 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700194 * slab to avoid swapping.
195 *
196 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
197 *
198 * `lru_pages' represents the number of on-LRU pages in all the zones which
199 * are eligible for the caller's allocation attempt. It is used for balancing
200 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700201 *
202 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203 */
Andrew Morton69e05942006-03-22 00:08:19 -0800204unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
205 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206{
207 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800208 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209
210 if (scanned == 0)
211 scanned = SWAP_CLUSTER_MAX;
212
213 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700214 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215
216 list_for_each_entry(shrinker, &shrinker_list, list) {
217 unsigned long long delta;
218 unsigned long total_scan;
Rusty Russell8e1f9362007-07-17 04:03:17 -0700219 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220
221 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800222 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223 do_div(delta, lru_pages + 1);
224 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800225 if (shrinker->nr < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700226 printk(KERN_ERR "shrink_slab: %pF negative objects to "
227 "delete nr=%ld\n",
228 shrinker->shrink, shrinker->nr);
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800229 shrinker->nr = max_pass;
230 }
231
232 /*
233 * Avoid risking looping forever due to too large nr value:
234 * never try to free more than twice the estimate number of
235 * freeable entries.
236 */
237 if (shrinker->nr > max_pass * 2)
238 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700239
240 total_scan = shrinker->nr;
241 shrinker->nr = 0;
242
243 while (total_scan >= SHRINK_BATCH) {
244 long this_scan = SHRINK_BATCH;
245 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700246 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247
Rusty Russell8e1f9362007-07-17 04:03:17 -0700248 nr_before = (*shrinker->shrink)(0, gfp_mask);
249 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700250 if (shrink_ret == -1)
251 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700252 if (shrink_ret < nr_before)
253 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700254 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700255 total_scan -= this_scan;
256
257 cond_resched();
258 }
259
260 shrinker->nr += total_scan;
261 }
262 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700263 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700264}
265
266/* Called without lock on whether page is mapped, so answer is unstable */
267static inline int page_mapping_inuse(struct page *page)
268{
269 struct address_space *mapping;
270
271 /* Page is in somebody's page tables. */
272 if (page_mapped(page))
273 return 1;
274
275 /* Be more reluctant to reclaim swapcache than pagecache */
276 if (PageSwapCache(page))
277 return 1;
278
279 mapping = page_mapping(page);
280 if (!mapping)
281 return 0;
282
283 /* File is mmap'd by somebody? */
284 return mapping_mapped(mapping);
285}
286
287static inline int is_page_cache_freeable(struct page *page)
288{
David Howells266cf652009-04-03 16:42:36 +0100289 return page_count(page) - !!page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700290}
291
292static int may_write_to_queue(struct backing_dev_info *bdi)
293{
Christoph Lameter930d9152006-01-08 01:00:47 -0800294 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295 return 1;
296 if (!bdi_write_congested(bdi))
297 return 1;
298 if (bdi == current->backing_dev_info)
299 return 1;
300 return 0;
301}
302
303/*
304 * We detected a synchronous write error writing a page out. Probably
305 * -ENOSPC. We need to propagate that into the address_space for a subsequent
306 * fsync(), msync() or close().
307 *
308 * The tricky part is that after writepage we cannot touch the mapping: nothing
309 * prevents it from being freed up. But we have a ref on the page and once
310 * that page is locked, the mapping is pinned.
311 *
312 * We're allowed to run sleeping lock_page() here because we know the caller has
313 * __GFP_FS.
314 */
315static void handle_write_error(struct address_space *mapping,
316 struct page *page, int error)
317{
318 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700319 if (page_mapping(page) == mapping)
320 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700321 unlock_page(page);
322}
323
Andy Whitcroftc661b072007-08-22 14:01:26 -0700324/* Request for sync pageout. */
325enum pageout_io {
326 PAGEOUT_IO_ASYNC,
327 PAGEOUT_IO_SYNC,
328};
329
Christoph Lameter04e62a22006-06-23 02:03:38 -0700330/* possible outcome of pageout() */
331typedef enum {
332 /* failed to write page out, page is locked */
333 PAGE_KEEP,
334 /* move page to the active list, page is locked */
335 PAGE_ACTIVATE,
336 /* page has been sent to the disk successfully, page is unlocked */
337 PAGE_SUCCESS,
338 /* page is clean and locked */
339 PAGE_CLEAN,
340} pageout_t;
341
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342/*
Andrew Morton1742f192006-03-22 00:08:21 -0800343 * pageout is called by shrink_page_list() for each dirty page.
344 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700346static pageout_t pageout(struct page *page, struct address_space *mapping,
347 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348{
349 /*
350 * If the page is dirty, only perform writeback if that write
351 * will be non-blocking. To prevent this allocation from being
352 * stalled by pagecache activity. But note that there may be
353 * stalls if we need to run get_block(). We could test
354 * PagePrivate for that.
355 *
356 * If this process is currently in generic_file_write() against
357 * this page's queue, we can perform writeback even if that
358 * will block.
359 *
360 * If the page is swapcache, write it back even if that would
361 * block, for some throttling. This happens by accident, because
362 * swap_backing_dev_info is bust: it doesn't reflect the
363 * congestion state of the swapdevs. Easy to fix, if needed.
364 * See swapfile.c:page_queue_congested().
365 */
366 if (!is_page_cache_freeable(page))
367 return PAGE_KEEP;
368 if (!mapping) {
369 /*
370 * Some data journaling orphaned pages can have
371 * page->mapping == NULL while being dirty with clean buffers.
372 */
David Howells266cf652009-04-03 16:42:36 +0100373 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 if (try_to_free_buffers(page)) {
375 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700376 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 return PAGE_CLEAN;
378 }
379 }
380 return PAGE_KEEP;
381 }
382 if (mapping->a_ops->writepage == NULL)
383 return PAGE_ACTIVATE;
384 if (!may_write_to_queue(mapping->backing_dev_info))
385 return PAGE_KEEP;
386
387 if (clear_page_dirty_for_io(page)) {
388 int res;
389 struct writeback_control wbc = {
390 .sync_mode = WB_SYNC_NONE,
391 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700392 .range_start = 0,
393 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 .nonblocking = 1,
395 .for_reclaim = 1,
396 };
397
398 SetPageReclaim(page);
399 res = mapping->a_ops->writepage(page, &wbc);
400 if (res < 0)
401 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800402 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700403 ClearPageReclaim(page);
404 return PAGE_ACTIVATE;
405 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700406
407 /*
408 * Wait on writeback if requested to. This happens when
409 * direct reclaiming a large contiguous area and the
410 * first attempt to free a range of pages fails.
411 */
412 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
413 wait_on_page_writeback(page);
414
Linus Torvalds1da177e2005-04-16 15:20:36 -0700415 if (!PageWriteback(page)) {
416 /* synchronous write or broken a_ops? */
417 ClearPageReclaim(page);
418 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700419 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700420 return PAGE_SUCCESS;
421 }
422
423 return PAGE_CLEAN;
424}
425
Andrew Mortona649fd92006-10-17 00:09:36 -0700426/*
Nick Piggine2867812008-07-25 19:45:30 -0700427 * Same as remove_mapping, but if the page is removed from the mapping, it
428 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700429 */
Nick Piggine2867812008-07-25 19:45:30 -0700430static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800431{
Nick Piggin28e4d962006-09-25 23:31:23 -0700432 BUG_ON(!PageLocked(page));
433 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800434
Nick Piggin19fd6232008-07-25 19:45:32 -0700435 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800436 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700437 * The non racy check for a busy page.
438 *
439 * Must be careful with the order of the tests. When someone has
440 * a ref to the page, it may be possible that they dirty it then
441 * drop the reference. So if PageDirty is tested before page_count
442 * here, then the following race may occur:
443 *
444 * get_user_pages(&page);
445 * [user mapping goes away]
446 * write_to(page);
447 * !PageDirty(page) [good]
448 * SetPageDirty(page);
449 * put_page(page);
450 * !page_count(page) [good, discard it]
451 *
452 * [oops, our write_to data is lost]
453 *
454 * Reversing the order of the tests ensures such a situation cannot
455 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
456 * load is not satisfied before that of page->_count.
457 *
458 * Note that if SetPageDirty is always performed via set_page_dirty,
459 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800460 */
Nick Piggine2867812008-07-25 19:45:30 -0700461 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800462 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700463 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
464 if (unlikely(PageDirty(page))) {
465 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800466 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700467 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800468
469 if (PageSwapCache(page)) {
470 swp_entry_t swap = { .val = page_private(page) };
471 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700472 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700473 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700474 } else {
475 __remove_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700476 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700477 mem_cgroup_uncharge_cache_page(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800478 }
479
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800480 return 1;
481
482cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700483 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800484 return 0;
485}
486
Linus Torvalds1da177e2005-04-16 15:20:36 -0700487/*
Nick Piggine2867812008-07-25 19:45:30 -0700488 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
489 * someone else has a ref on the page, abort and return 0. If it was
490 * successfully detached, return 1. Assumes the caller has a single ref on
491 * this page.
492 */
493int remove_mapping(struct address_space *mapping, struct page *page)
494{
495 if (__remove_mapping(mapping, page)) {
496 /*
497 * Unfreezing the refcount with 1 rather than 2 effectively
498 * drops the pagecache ref for us without requiring another
499 * atomic operation.
500 */
501 page_unfreeze_refs(page, 1);
502 return 1;
503 }
504 return 0;
505}
506
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700507/**
508 * putback_lru_page - put previously isolated page onto appropriate LRU list
509 * @page: page to be put back to appropriate lru list
510 *
511 * Add previously isolated @page to appropriate LRU list.
512 * Page may still be unevictable for other reasons.
513 *
514 * lru_lock must not be held, interrupts must be enabled.
515 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700516void putback_lru_page(struct page *page)
517{
518 int lru;
519 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700520 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700521
522 VM_BUG_ON(PageLRU(page));
523
524redo:
525 ClearPageUnevictable(page);
526
527 if (page_evictable(page, NULL)) {
528 /*
529 * For evictable pages, we can use the cache.
530 * In event of a race, worst case is we end up with an
531 * unevictable page on [in]active list.
532 * We know how to handle that.
533 */
534 lru = active + page_is_file_cache(page);
535 lru_cache_add_lru(page, lru);
536 } else {
537 /*
538 * Put unevictable pages directly on zone's unevictable
539 * list.
540 */
541 lru = LRU_UNEVICTABLE;
542 add_page_to_unevictable_list(page);
543 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700544
545 /*
546 * page's status can change while we move it among lru. If an evictable
547 * page is on unevictable list, it never be freed. To avoid that,
548 * check after we added it to the list, again.
549 */
550 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
551 if (!isolate_lru_page(page)) {
552 put_page(page);
553 goto redo;
554 }
555 /* This means someone else dropped this page from LRU
556 * So, it will be freed or putback to LRU again. There is
557 * nothing to do here.
558 */
559 }
560
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700561 if (was_unevictable && lru != LRU_UNEVICTABLE)
562 count_vm_event(UNEVICTABLE_PGRESCUED);
563 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
564 count_vm_event(UNEVICTABLE_PGCULLED);
565
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700566 put_page(page); /* drop ref from isolate */
567}
568
Nick Piggine2867812008-07-25 19:45:30 -0700569/*
Andrew Morton1742f192006-03-22 00:08:21 -0800570 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571 */
Andrew Morton1742f192006-03-22 00:08:21 -0800572static unsigned long shrink_page_list(struct list_head *page_list,
Andy Whitcroftc661b072007-08-22 14:01:26 -0700573 struct scan_control *sc,
574 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575{
576 LIST_HEAD(ret_pages);
577 struct pagevec freed_pvec;
578 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800579 unsigned long nr_reclaimed = 0;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700580 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700581
582 cond_resched();
583
584 pagevec_init(&freed_pvec, 1);
585 while (!list_empty(page_list)) {
586 struct address_space *mapping;
587 struct page *page;
588 int may_enter_fs;
589 int referenced;
590
591 cond_resched();
592
593 page = lru_to_page(page_list);
594 list_del(&page->lru);
595
Nick Piggin529ae9a2008-08-02 12:01:03 +0200596 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700597 goto keep;
598
Nick Piggin725d7042006-09-25 23:30:55 -0700599 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600
601 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800602
Nick Pigginb291f002008-10-18 20:26:44 -0700603 if (unlikely(!page_evictable(page, NULL)))
604 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700605
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700606 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800607 goto keep_locked;
608
Linus Torvalds1da177e2005-04-16 15:20:36 -0700609 /* Double the slab pressure for mapped and swapcache pages */
610 if (page_mapped(page) || PageSwapCache(page))
611 sc->nr_scanned++;
612
Andy Whitcroftc661b072007-08-22 14:01:26 -0700613 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
614 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
615
616 if (PageWriteback(page)) {
617 /*
618 * Synchronous reclaim is performed in two passes,
619 * first an asynchronous pass over the list to
620 * start parallel writeback, and a second synchronous
621 * pass to wait for the IO to complete. Wait here
622 * for any page for which writeback has already
623 * started.
624 */
625 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
626 wait_on_page_writeback(page);
Andrew Morton4dd4b922008-03-24 12:29:52 -0700627 else
Andy Whitcroftc661b072007-08-22 14:01:26 -0700628 goto keep_locked;
629 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700630
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700631 referenced = page_referenced(page, 1,
632 sc->mem_cgroup, &vm_flags);
Minchan Kim03ef83a2009-08-26 14:29:23 -0700633 /*
634 * In active use or really unfreeable? Activate it.
635 * If page which have PG_mlocked lost isoltation race,
636 * try_to_unmap moves it to unevictable list
637 */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700638 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
Minchan Kim03ef83a2009-08-26 14:29:23 -0700639 referenced && page_mapping_inuse(page)
640 && !(vm_flags & VM_LOCKED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700641 goto activate_locked;
642
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643 /*
644 * Anonymous process memory has backing store?
645 * Try to allocate it some swap space here.
646 */
Nick Pigginb291f002008-10-18 20:26:44 -0700647 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800648 if (!(sc->gfp_mask & __GFP_IO))
649 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800650 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700651 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800652 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700653 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654
655 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656
657 /*
658 * The page is mapped into the page tables of one or more
659 * processes. Try to unmap it here.
660 */
661 if (page_mapped(page) && mapping) {
Christoph Lametera48d07a2006-02-01 03:05:38 -0800662 switch (try_to_unmap(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663 case SWAP_FAIL:
664 goto activate_locked;
665 case SWAP_AGAIN:
666 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700667 case SWAP_MLOCK:
668 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669 case SWAP_SUCCESS:
670 ; /* try to free the page below */
671 }
672 }
673
674 if (PageDirty(page)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700675 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700677 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800679 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 goto keep_locked;
681
682 /* Page is dirty, try to write it out here */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700683 switch (pageout(page, mapping, sync_writeback)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 case PAGE_KEEP:
685 goto keep_locked;
686 case PAGE_ACTIVATE:
687 goto activate_locked;
688 case PAGE_SUCCESS:
Andrew Morton4dd4b922008-03-24 12:29:52 -0700689 if (PageWriteback(page) || PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 goto keep;
691 /*
692 * A synchronous write - probably a ramdisk. Go
693 * ahead and try to reclaim the page.
694 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200695 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 goto keep;
697 if (PageDirty(page) || PageWriteback(page))
698 goto keep_locked;
699 mapping = page_mapping(page);
700 case PAGE_CLEAN:
701 ; /* try to free the page below */
702 }
703 }
704
705 /*
706 * If the page has buffers, try to free the buffer mappings
707 * associated with this page. If we succeed we try to free
708 * the page as well.
709 *
710 * We do this even if the page is PageDirty().
711 * try_to_release_page() does not perform I/O, but it is
712 * possible for a page to have PageDirty set, but it is actually
713 * clean (all its buffers are clean). This happens if the
714 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700715 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 * try_to_release_page() will discover that cleanness and will
717 * drop the buffers and mark the page clean - it can be freed.
718 *
719 * Rarely, pages can have buffers and no ->mapping. These are
720 * the pages which were not successfully invalidated in
721 * truncate_complete_page(). We try to drop those buffers here
722 * and if that worked, and the page is no longer mapped into
723 * process address space (page_count == 1) it can be freed.
724 * Otherwise, leave the page on the LRU so it is swappable.
725 */
David Howells266cf652009-04-03 16:42:36 +0100726 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 if (!try_to_release_page(page, sc->gfp_mask))
728 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700729 if (!mapping && page_count(page) == 1) {
730 unlock_page(page);
731 if (put_page_testzero(page))
732 goto free_it;
733 else {
734 /*
735 * rare race with speculative reference.
736 * the speculative reference will free
737 * this page shortly, so we may
738 * increment nr_reclaimed here (and
739 * leave it off the LRU).
740 */
741 nr_reclaimed++;
742 continue;
743 }
744 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 }
746
Nick Piggine2867812008-07-25 19:45:30 -0700747 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800748 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749
Nick Piggina978d6f2008-10-18 20:26:58 -0700750 /*
751 * At this point, we have no other references and there is
752 * no way to pick any more up (removed from LRU, removed
753 * from pagecache). Can use non-atomic bitops now (and
754 * we obviously don't have to worry about waking up a process
755 * waiting on the page lock, because there are no references.
756 */
757 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700758free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800759 nr_reclaimed++;
Nick Piggine2867812008-07-25 19:45:30 -0700760 if (!pagevec_add(&freed_pvec, page)) {
761 __pagevec_free(&freed_pvec);
762 pagevec_reinit(&freed_pvec);
763 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700764 continue;
765
Nick Pigginb291f002008-10-18 20:26:44 -0700766cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800767 if (PageSwapCache(page))
768 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700769 unlock_page(page);
770 putback_lru_page(page);
771 continue;
772
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700774 /* Not a candidate for swapping, so reclaim swap space. */
775 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800776 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700777 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 SetPageActive(page);
779 pgactivate++;
780keep_locked:
781 unlock_page(page);
782keep:
783 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700784 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700785 }
786 list_splice(&ret_pages, page_list);
787 if (pagevec_count(&freed_pvec))
Nick Piggine2867812008-07-25 19:45:30 -0700788 __pagevec_free(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700789 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800790 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791}
792
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700793/* LRU Isolation modes. */
794#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
795#define ISOLATE_ACTIVE 1 /* Isolate active pages. */
796#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
797
798/*
799 * Attempt to remove the specified page from its LRU. Only take this page
800 * if it is of the appropriate PageActive status. Pages which are being
801 * freed elsewhere are also ignored.
802 *
803 * page: page to consider
804 * mode: one of the LRU isolation modes defined above
805 *
806 * returns 0 on success, -ve errno on failure.
807 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700808int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700809{
810 int ret = -EINVAL;
811
812 /* Only take pages on the LRU. */
813 if (!PageLRU(page))
814 return ret;
815
816 /*
817 * When checking the active state, we need to be sure we are
818 * dealing with comparible boolean values. Take the logical not
819 * of each.
820 */
821 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
822 return ret;
823
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700824 if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
825 return ret;
826
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700827 /*
828 * When this function is being called for lumpy reclaim, we
829 * initially look into all LRU pages, active, inactive and
830 * unevictable; only give shrink_page_list evictable pages.
831 */
832 if (PageUnevictable(page))
833 return ret;
834
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700835 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800836
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700837 if (likely(get_page_unless_zero(page))) {
838 /*
839 * Be careful not to clear PageLRU until after we're
840 * sure the page is not being freed elsewhere -- the
841 * page release code relies on it.
842 */
843 ClearPageLRU(page);
844 ret = 0;
845 }
846
847 return ret;
848}
849
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800850/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700851 * zone->lru_lock is heavily contended. Some of the functions that
852 * shrink the lists perform better by taking out a batch of pages
853 * and working on them outside the LRU lock.
854 *
855 * For pagecache intensive workloads, this function is the hottest
856 * spot in the kernel (apart from copy_*_user functions).
857 *
858 * Appropriate locks must be held before calling this function.
859 *
860 * @nr_to_scan: The number of pages to look through on the list.
861 * @src: The LRU list to pull pages off.
862 * @dst: The temp list to put pages on to.
863 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700864 * @order: The caller's attempted allocation order
865 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700866 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867 *
868 * returns how many pages were moved onto *@dst.
869 */
Andrew Morton69e05942006-03-22 00:08:19 -0800870static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
871 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700872 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700873{
Andrew Morton69e05942006-03-22 00:08:19 -0800874 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800875 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700876
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800877 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700878 struct page *page;
879 unsigned long pfn;
880 unsigned long end_pfn;
881 unsigned long page_pfn;
882 int zone_id;
883
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 page = lru_to_page(src);
885 prefetchw_prev_lru_page(page, src, flags);
886
Nick Piggin725d7042006-09-25 23:30:55 -0700887 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800888
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700889 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700890 case 0:
891 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700892 mem_cgroup_del_lru(page);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800893 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700894 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800895
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700896 case -EBUSY:
897 /* else it is being freed elsewhere */
898 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700899 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700900 continue;
901
902 default:
903 BUG();
904 }
905
906 if (!order)
907 continue;
908
909 /*
910 * Attempt to take all pages in the order aligned region
911 * surrounding the tag page. Only take those pages of
912 * the same active state as that tag page. We may safely
913 * round the target page pfn down to the requested order
914 * as the mem_map is guarenteed valid out to MAX_ORDER,
915 * where that page is in a different zone we will detect
916 * it from its zone id and abort this block scan.
917 */
918 zone_id = page_zone_id(page);
919 page_pfn = page_to_pfn(page);
920 pfn = page_pfn & ~((1 << order) - 1);
921 end_pfn = pfn + (1 << order);
922 for (; pfn < end_pfn; pfn++) {
923 struct page *cursor_page;
924
925 /* The target page is in the block, ignore it. */
926 if (unlikely(pfn == page_pfn))
927 continue;
928
929 /* Avoid holes within the zone. */
930 if (unlikely(!pfn_valid_within(pfn)))
931 break;
932
933 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700934
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700935 /* Check that we have not crossed a zone boundary. */
936 if (unlikely(page_zone_id(cursor_page) != zone_id))
937 continue;
Minchan Kimde2e7562009-09-21 17:01:43 -0700938
939 /*
940 * If we don't have enough swap space, reclaiming of
941 * anon page which don't already have a swap slot is
942 * pointless.
943 */
944 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
945 !PageSwapCache(cursor_page))
946 continue;
947
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -0700948 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700949 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +0900950 mem_cgroup_del_lru(cursor_page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700951 nr_taken++;
952 scan++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700953 }
954 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 }
956
957 *scanned = scan;
958 return nr_taken;
959}
960
Balbir Singh66e17072008-02-07 00:13:56 -0800961static unsigned long isolate_pages_global(unsigned long nr,
962 struct list_head *dst,
963 unsigned long *scanned, int order,
964 int mode, struct zone *z,
965 struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700966 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -0800967{
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700968 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -0800969 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700970 lru += LRU_ACTIVE;
971 if (file)
972 lru += LRU_FILE;
973 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
974 mode, !!file);
Balbir Singh66e17072008-02-07 00:13:56 -0800975}
976
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700978 * clear_active_flags() is a helper for shrink_active_list(), clearing
979 * any active bits from the pages in the list.
980 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700981static unsigned long clear_active_flags(struct list_head *page_list,
982 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700983{
984 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700985 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700986 struct page *page;
987
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700988 list_for_each_entry(page, page_list, lru) {
989 lru = page_is_file_cache(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700990 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700991 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700992 ClearPageActive(page);
993 nr_active++;
994 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700995 count[lru]++;
996 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700997
998 return nr_active;
999}
1000
Nick Piggin62695a82008-10-18 20:26:09 -07001001/**
1002 * isolate_lru_page - tries to isolate a page from its LRU list
1003 * @page: page to isolate from its LRU list
1004 *
1005 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1006 * vmstat statistic corresponding to whatever LRU list the page was on.
1007 *
1008 * Returns 0 if the page was removed from an LRU list.
1009 * Returns -EBUSY if the page was not on an LRU list.
1010 *
1011 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001012 * the active list, it will have PageActive set. If it was found on
1013 * the unevictable list, it will have the PageUnevictable bit set. That flag
1014 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001015 *
1016 * The vmstat statistic corresponding to the list on which the page was
1017 * found will be decremented.
1018 *
1019 * Restrictions:
1020 * (1) Must be called with an elevated refcount on the page. This is a
1021 * fundamentnal difference from isolate_lru_pages (which is called
1022 * without a stable reference).
1023 * (2) the lru_lock must not be held.
1024 * (3) interrupts must be enabled.
1025 */
1026int isolate_lru_page(struct page *page)
1027{
1028 int ret = -EBUSY;
1029
1030 if (PageLRU(page)) {
1031 struct zone *zone = page_zone(page);
1032
1033 spin_lock_irq(&zone->lru_lock);
1034 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001035 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001036 ret = 0;
1037 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001038
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001039 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001040 }
1041 spin_unlock_irq(&zone->lru_lock);
1042 }
1043 return ret;
1044}
1045
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001046/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001047 * Are there way too many processes in the direct reclaim path already?
1048 */
1049static int too_many_isolated(struct zone *zone, int file,
1050 struct scan_control *sc)
1051{
1052 unsigned long inactive, isolated;
1053
1054 if (current_is_kswapd())
1055 return 0;
1056
1057 if (!scanning_global_lru(sc))
1058 return 0;
1059
1060 if (file) {
1061 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1062 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1063 } else {
1064 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1065 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1066 }
1067
1068 return isolated > inactive;
1069}
1070
1071/*
Andrew Morton1742f192006-03-22 00:08:21 -08001072 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1073 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001074 */
Andrew Morton1742f192006-03-22 00:08:21 -08001075static unsigned long shrink_inactive_list(unsigned long max_scan,
Rik van Riel33c120e2008-10-18 20:26:36 -07001076 struct zone *zone, struct scan_control *sc,
1077 int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078{
1079 LIST_HEAD(page_list);
1080 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -08001081 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001082 unsigned long nr_reclaimed = 0;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001083 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001084 int lumpy_reclaim = 0;
1085
Rik van Riel35cd7812009-09-21 17:01:38 -07001086 while (unlikely(too_many_isolated(zone, file, sc))) {
1087 congestion_wait(WRITE, HZ/10);
1088
1089 /* We are about to die and free our memory. Return now. */
1090 if (fatal_signal_pending(current))
1091 return SWAP_CLUSTER_MAX;
1092 }
1093
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001094 /*
1095 * If we need a large contiguous chunk of memory, or have
1096 * trouble getting a small set of contiguous pages, we
1097 * will reclaim both active and inactive pages.
1098 *
1099 * We use the same threshold as pageout congestion_wait below.
1100 */
1101 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1102 lumpy_reclaim = 1;
1103 else if (sc->order && priority < DEF_PRIORITY - 2)
1104 lumpy_reclaim = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105
1106 pagevec_init(&pvec, 1);
1107
1108 lru_add_drain();
1109 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -08001110 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001111 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -08001112 unsigned long nr_taken;
1113 unsigned long nr_scan;
1114 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001115 unsigned long nr_active;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001116 unsigned int count[NR_LRU_LISTS] = { 0, };
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001117 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001118 unsigned long nr_anon;
1119 unsigned long nr_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001120
Balbir Singh66e17072008-02-07 00:13:56 -08001121 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001122 &page_list, &nr_scan, sc->order, mode,
1123 zone, sc->mem_cgroup, 0, file);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001124
1125 if (scanning_global_lru(sc)) {
1126 zone->pages_scanned += nr_scan;
1127 if (current_is_kswapd())
1128 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1129 nr_scan);
1130 else
1131 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1132 nr_scan);
1133 }
1134
1135 if (nr_taken == 0)
1136 goto done;
1137
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001138 nr_active = clear_active_flags(&page_list, count);
Andy Whitcrofte9187bd2007-08-22 14:01:25 -07001139 __count_vm_events(PGDEACTIVATE, nr_active);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001140
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001141 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1142 -count[LRU_ACTIVE_FILE]);
1143 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1144 -count[LRU_INACTIVE_FILE]);
1145 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1146 -count[LRU_ACTIVE_ANON]);
1147 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1148 -count[LRU_INACTIVE_ANON]);
1149
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001150 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1151 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1152 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1153 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001154
1155 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1156 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1157 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1158 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1159
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160 spin_unlock_irq(&zone->lru_lock);
1161
Andrew Morton69e05942006-03-22 00:08:19 -08001162 nr_scanned += nr_scan;
Andy Whitcroftc661b072007-08-22 14:01:26 -07001163 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1164
1165 /*
1166 * If we are direct reclaiming for contiguous pages and we do
1167 * not reclaim everything in the list, try again and wait
1168 * for IO to complete. This will stall high-order allocations
1169 * but that should be acceptable to the caller
1170 */
1171 if (nr_freed < nr_taken && !current_is_kswapd() &&
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001172 lumpy_reclaim) {
Jens Axboe8aa7e842009-07-09 14:52:32 +02001173 congestion_wait(BLK_RW_ASYNC, HZ/10);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001174
1175 /*
1176 * The attempt at page out may have made some
1177 * of the pages active, mark them inactive again.
1178 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001179 nr_active = clear_active_flags(&page_list, count);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001180 count_vm_events(PGDEACTIVATE, nr_active);
1181
1182 nr_freed += shrink_page_list(&page_list, sc,
1183 PAGEOUT_IO_SYNC);
1184 }
1185
Andrew Morton05ff5132006-03-22 00:08:20 -08001186 nr_reclaimed += nr_freed;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001187
Nick Piggina74609f2006-01-06 00:11:20 -08001188 local_irq_disable();
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001189 if (current_is_kswapd())
Christoph Lameterf8891e52006-06-30 01:55:45 -07001190 __count_vm_events(KSWAPD_STEAL, nr_freed);
Shantanu Goel918d3f92006-12-29 16:48:59 -08001191 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -08001192
1193 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001194 /*
1195 * Put back any unfreeable pages.
1196 */
1197 while (!list_empty(&page_list)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001198 int lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -07001200 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001201 list_del(&page->lru);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001202 if (unlikely(!page_evictable(page, NULL))) {
1203 spin_unlock_irq(&zone->lru_lock);
1204 putback_lru_page(page);
1205 spin_lock_irq(&zone->lru_lock);
1206 continue;
1207 }
1208 SetPageLRU(page);
1209 lru = page_lru(page);
1210 add_page_to_lru_list(zone, page, lru);
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001211 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001212 int file = !!page_is_file_cache(page);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001213 reclaim_stat->recent_rotated[file]++;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001214 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 if (!pagevec_add(&pvec, page)) {
1216 spin_unlock_irq(&zone->lru_lock);
1217 __pagevec_release(&pvec);
1218 spin_lock_irq(&zone->lru_lock);
1219 }
1220 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001221 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1222 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1223
Andrew Morton69e05942006-03-22 00:08:19 -08001224 } while (nr_scanned < max_scan);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001225
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226done:
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001227 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -08001229 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230}
1231
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001232/*
1233 * We are about to scan this zone at a certain priority level. If that priority
1234 * level is smaller (ie: more urgent) than the previous priority, then note
1235 * that priority level within the zone. This is done so that when the next
1236 * process comes in to scan this zone, it will immediately start out at this
1237 * priority level rather than having to build up its own scanning priority.
1238 * Here, this priority affects only the reclaim-mapped threshold.
1239 */
1240static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1241{
1242 if (priority < zone->prev_priority)
1243 zone->prev_priority = priority;
1244}
1245
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001246/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247 * This moves pages from the active list to the inactive list.
1248 *
1249 * We move them the other way if the page is referenced by one or more
1250 * processes, from rmap.
1251 *
1252 * If the pages are mostly unmapped, the processing is fast and it is
1253 * appropriate to hold zone->lru_lock across the whole operation. But if
1254 * the pages are mapped, the processing is slow (page_referenced()) so we
1255 * should drop zone->lru_lock around each page. It's impossible to balance
1256 * this, so instead we remove the pages from the LRU while processing them.
1257 * It is safe to rely on PG_active against the non-LRU pages in here because
1258 * nobody will play with that bit on a non-LRU page.
1259 *
1260 * The downside is that we have to touch page->_count against each page.
1261 * But we had to alter page->flags anyway.
1262 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001263
Wu Fengguang3eb41402009-06-16 15:33:13 -07001264static void move_active_pages_to_lru(struct zone *zone,
1265 struct list_head *list,
1266 enum lru_list lru)
1267{
1268 unsigned long pgmoved = 0;
1269 struct pagevec pvec;
1270 struct page *page;
1271
1272 pagevec_init(&pvec, 1);
1273
1274 while (!list_empty(list)) {
1275 page = lru_to_page(list);
1276 prefetchw_prev_lru_page(page, list, flags);
1277
1278 VM_BUG_ON(PageLRU(page));
1279 SetPageLRU(page);
1280
1281 VM_BUG_ON(!PageActive(page));
1282 if (!is_active_lru(lru))
1283 ClearPageActive(page); /* we are de-activating */
1284
1285 list_move(&page->lru, &zone->lru[lru].list);
1286 mem_cgroup_add_lru_list(page, lru);
1287 pgmoved++;
1288
1289 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1290 spin_unlock_irq(&zone->lru_lock);
1291 if (buffer_heads_over_limit)
1292 pagevec_strip(&pvec);
1293 __pagevec_release(&pvec);
1294 spin_lock_irq(&zone->lru_lock);
1295 }
1296 }
1297 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1298 if (!is_active_lru(lru))
1299 __count_vm_events(PGDEACTIVATE, pgmoved);
1300}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001301
Andrew Morton1742f192006-03-22 00:08:21 -08001302static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001303 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001304{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001305 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001306 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001307 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001308 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001309 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001310 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001312 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001313 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314
1315 lru_add_drain();
1316 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001317 nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
Balbir Singh66e17072008-02-07 00:13:56 -08001318 ISOLATE_ACTIVE, zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001319 sc->mem_cgroup, 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001320 /*
1321 * zone->pages_scanned is used for detect zone's oom
1322 * mem_cgroup remembers nr_scan by itself.
1323 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001324 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001325 zone->pages_scanned += pgscanned;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001326 }
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001327 reclaim_stat->recent_scanned[!!file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001328
Wu Fengguang3eb41402009-06-16 15:33:13 -07001329 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001330 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001331 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001332 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001333 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001334 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001335 spin_unlock_irq(&zone->lru_lock);
1336
Linus Torvalds1da177e2005-04-16 15:20:36 -07001337 while (!list_empty(&l_hold)) {
1338 cond_resched();
1339 page = lru_to_page(&l_hold);
1340 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001341
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001342 if (unlikely(!page_evictable(page, NULL))) {
1343 putback_lru_page(page);
1344 continue;
1345 }
1346
Rik van Riel7e9cd482008-10-18 20:26:35 -07001347 /* page_referenced clears PageReferenced */
1348 if (page_mapping_inuse(page) &&
Wu Fengguang8cab4752009-06-16 15:33:12 -07001349 page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001350 nr_rotated++;
Wu Fengguang8cab4752009-06-16 15:33:12 -07001351 /*
1352 * Identify referenced, file-backed active pages and
1353 * give them one more trip around the active list. So
1354 * that executable code get better chances to stay in
1355 * memory under moderate memory pressure. Anon pages
1356 * are not likely to be evicted by use-once streaming
1357 * IO, plus JVM can create lots of anon VM_EXEC pages,
1358 * so we ignore them here.
1359 */
1360 if ((vm_flags & VM_EXEC) && !PageAnon(page)) {
1361 list_add(&page->lru, &l_active);
1362 continue;
1363 }
1364 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001365
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366 list_add(&page->lru, &l_inactive);
1367 }
1368
Andrew Mortonb5557492009-01-06 14:40:13 -08001369 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001370 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001371 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001372 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001373 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001374 * Count referenced pages from currently used mappings as rotated,
1375 * even though only some of them are actually re-activated. This
1376 * helps balance scan pressure between file and anonymous pages in
1377 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001378 */
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001379 reclaim_stat->recent_rotated[!!file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001380
Wu Fengguang3eb41402009-06-16 15:33:13 -07001381 move_active_pages_to_lru(zone, &l_active,
1382 LRU_ACTIVE + file * LRU_FILE);
1383 move_active_pages_to_lru(zone, &l_inactive,
1384 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001385 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001386 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387}
1388
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001389static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001390{
1391 unsigned long active, inactive;
1392
1393 active = zone_page_state(zone, NR_ACTIVE_ANON);
1394 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1395
1396 if (inactive * zone->inactive_ratio < active)
1397 return 1;
1398
1399 return 0;
1400}
1401
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001402/**
1403 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1404 * @zone: zone to check
1405 * @sc: scan control of this context
1406 *
1407 * Returns true if the zone does not have enough inactive anon pages,
1408 * meaning some active anon pages need to be deactivated.
1409 */
1410static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1411{
1412 int low;
1413
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001414 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001415 low = inactive_anon_is_low_global(zone);
1416 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001417 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001418 return low;
1419}
1420
Rik van Riel56e49d22009-06-16 15:32:28 -07001421static int inactive_file_is_low_global(struct zone *zone)
1422{
1423 unsigned long active, inactive;
1424
1425 active = zone_page_state(zone, NR_ACTIVE_FILE);
1426 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1427
1428 return (active > inactive);
1429}
1430
1431/**
1432 * inactive_file_is_low - check if file pages need to be deactivated
1433 * @zone: zone to check
1434 * @sc: scan control of this context
1435 *
1436 * When the system is doing streaming IO, memory pressure here
1437 * ensures that active file pages get deactivated, until more
1438 * than half of the file pages are on the inactive list.
1439 *
1440 * Once we get to that situation, protect the system's working
1441 * set from being evicted by disabling active file page aging.
1442 *
1443 * This uses a different ratio than the anonymous pages, because
1444 * the page cache uses a use-once replacement algorithm.
1445 */
1446static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1447{
1448 int low;
1449
1450 if (scanning_global_lru(sc))
1451 low = inactive_file_is_low_global(zone);
1452 else
1453 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1454 return low;
1455}
1456
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001457static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001458 struct zone *zone, struct scan_control *sc, int priority)
1459{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001460 int file = is_file_lru(lru);
1461
Rik van Riel56e49d22009-06-16 15:32:28 -07001462 if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
Rik van Riel556adec2008-10-18 20:26:34 -07001463 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1464 return 0;
1465 }
1466
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001467 if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001468 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001469 return 0;
1470 }
Rik van Riel33c120e2008-10-18 20:26:36 -07001471 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001472}
1473
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001475 * Determine how aggressively the anon and file LRU lists should be
1476 * scanned. The relative value of each set of LRU lists is determined
1477 * by looking at the fraction of the pages scanned we did rotate back
1478 * onto the active list instead of evict.
1479 *
1480 * percent[0] specifies how much pressure to put on ram/swap backed
1481 * memory, while percent[1] determines pressure on the file LRUs.
1482 */
1483static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1484 unsigned long *percent)
1485{
1486 unsigned long anon, file, free;
1487 unsigned long anon_prio, file_prio;
1488 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001489 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001490
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -08001491 anon = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1492 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1493 file = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1494 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001495
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001496 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001497 free = zone_page_state(zone, NR_FREE_PAGES);
1498 /* If we have very few page cache pages,
1499 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001500 if (unlikely(file + free <= high_wmark_pages(zone))) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001501 percent[0] = 100;
1502 percent[1] = 0;
1503 return;
1504 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001505 }
1506
1507 /*
1508 * OK, so we have swap space and a fair amount of page cache
1509 * pages. We use the recently rotated / recently scanned
1510 * ratios to determine how valuable each cache is.
1511 *
1512 * Because workloads change over time (and to avoid overflow)
1513 * we keep these statistics as a floating average, which ends
1514 * up weighing recent references more than old ones.
1515 *
1516 * anon in [0], file in [1]
1517 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001518 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001519 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001520 reclaim_stat->recent_scanned[0] /= 2;
1521 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001522 spin_unlock_irq(&zone->lru_lock);
1523 }
1524
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001525 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001526 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001527 reclaim_stat->recent_scanned[1] /= 2;
1528 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001529 spin_unlock_irq(&zone->lru_lock);
1530 }
1531
1532 /*
1533 * With swappiness at 100, anonymous and file have the same priority.
1534 * This scanning priority is essentially the inverse of IO cost.
1535 */
1536 anon_prio = sc->swappiness;
1537 file_prio = 200 - sc->swappiness;
1538
1539 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001540 * The amount of pressure on anon vs file pages is inversely
1541 * proportional to the fraction of recently scanned pages on
1542 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001543 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001544 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1545 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001546
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001547 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1548 fp /= reclaim_stat->recent_rotated[1] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001549
1550 /* Normalize to percentages */
1551 percent[0] = 100 * ap / (ap + fp + 1);
1552 percent[1] = 100 - percent[0];
1553}
1554
Wu Fengguang6e08a362009-06-16 15:32:29 -07001555/*
1556 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1557 * until we collected @swap_cluster_max pages to scan.
1558 */
1559static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1560 unsigned long *nr_saved_scan,
1561 unsigned long swap_cluster_max)
1562{
1563 unsigned long nr;
1564
1565 *nr_saved_scan += nr_to_scan;
1566 nr = *nr_saved_scan;
1567
1568 if (nr >= swap_cluster_max)
1569 *nr_saved_scan = 0;
1570 else
1571 nr = 0;
1572
1573 return nr;
1574}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001575
1576/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1578 */
Rik van Riela79311c2009-01-06 14:40:01 -08001579static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001580 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001582 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001583 unsigned long nr_to_scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001584 unsigned long percent[2]; /* anon @ 0; file @ 1 */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001585 enum lru_list l;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001586 unsigned long nr_reclaimed = sc->nr_reclaimed;
1587 unsigned long swap_cluster_max = sc->swap_cluster_max;
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001588 int noswap = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001590 /* If we have no swap space, do not bother scanning anon pages. */
1591 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1592 noswap = 1;
1593 percent[0] = 0;
1594 percent[1] = 100;
1595 } else
1596 get_scan_ratio(zone, sc, percent);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001597
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001598 for_each_evictable_lru(l) {
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001599 int file = is_file_lru(l);
Andrew Morton8713e012009-04-30 15:08:55 -07001600 unsigned long scan;
Johannes Weinere0f79b82008-10-18 20:26:55 -07001601
KOSAKI Motohirof272b7b2009-03-12 14:31:36 -07001602 scan = zone_nr_pages(zone, sc, l);
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001603 if (priority || noswap) {
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001604 scan >>= priority;
1605 scan = (scan * percent[file]) / 100;
1606 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001607 if (scanning_global_lru(sc))
1608 nr[l] = nr_scan_try_batch(scan,
1609 &zone->lru[l].nr_saved_scan,
1610 swap_cluster_max);
1611 else
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001612 nr[l] = scan;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001613 }
1614
Rik van Riel556adec2008-10-18 20:26:34 -07001615 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1616 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001617 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001618 if (nr[l]) {
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001619 nr_to_scan = min(nr[l], swap_cluster_max);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001620 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001622 nr_reclaimed += shrink_list(l, nr_to_scan,
1623 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001624 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001625 }
Rik van Riela79311c2009-01-06 14:40:01 -08001626 /*
1627 * On large memory systems, scan >> priority can become
1628 * really large. This is fine for the starting priority;
1629 * we want to put equal scanning pressure on each zone.
1630 * However, if the VM has a harder time of freeing pages,
1631 * with multiple processes reclaiming pages, the total
1632 * freeing target can get unreasonably large.
1633 */
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001634 if (nr_reclaimed > swap_cluster_max &&
Rik van Riela79311c2009-01-06 14:40:01 -08001635 priority < DEF_PRIORITY && !current_is_kswapd())
1636 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637 }
1638
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001639 sc->nr_reclaimed = nr_reclaimed;
1640
Rik van Riel556adec2008-10-18 20:26:34 -07001641 /*
1642 * Even if we did not try to evict anon pages at all, we want to
1643 * rebalance the anon lru active/inactive ratio.
1644 */
MinChan Kim69c85482009-06-16 15:32:44 -07001645 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
Rik van Riel556adec2008-10-18 20:26:34 -07001646 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1647
Andrew Morton232ea4d2007-02-28 20:13:21 -08001648 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001649}
1650
1651/*
1652 * This is the direct reclaim path, for page-allocating processes. We only
1653 * try to reclaim pages from zones which will satisfy the caller's allocation
1654 * request.
1655 *
Mel Gorman41858962009-06-16 15:32:12 -07001656 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1657 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1659 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001660 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1661 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1662 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001663 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664 * If a zone is deemed to be full of pinned pages then just give it a light
1665 * scan then give up on it.
1666 */
Rik van Riela79311c2009-01-06 14:40:01 -08001667static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001668 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669{
Mel Gorman54a6eb52008-04-28 02:12:16 -07001670 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Mel Gormandd1a2392008-04-28 02:12:17 -07001671 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001672 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001673
Nick Piggin408d8542006-09-25 23:31:27 -07001674 sc->all_unreclaimable = 1;
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001675 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1676 sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001677 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001679 /*
1680 * Take care memory controller reclaiming has small influence
1681 * to global LRU.
1682 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001683 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001684 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1685 continue;
1686 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001688 if (zone_is_all_unreclaimable(zone) &&
1689 priority != DEF_PRIORITY)
1690 continue; /* Let kswapd poll it */
1691 sc->all_unreclaimable = 0;
1692 } else {
1693 /*
1694 * Ignore cpuset limitation here. We just want to reduce
1695 * # of used pages by us regardless of memory shortage.
1696 */
1697 sc->all_unreclaimable = 0;
1698 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1699 priority);
1700 }
Nick Piggin408d8542006-09-25 23:31:27 -07001701
Rik van Riela79311c2009-01-06 14:40:01 -08001702 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 }
1704}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001705
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706/*
1707 * This is the main entry point to direct page reclaim.
1708 *
1709 * If a full scan of the inactive list fails to free enough memory then we
1710 * are "out of memory" and something needs to be killed.
1711 *
1712 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1713 * high - the zone may be full of dirty or under-writeback pages, which this
1714 * caller can't do much about. We kick pdflush and take explicit naps in the
1715 * hope that some of these pages can be written. But if the allocating task
1716 * holds filesystem locks which prevent writeout this might not work, and the
1717 * allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001718 *
1719 * returns: 0, if no pages reclaimed
1720 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721 */
Mel Gormandac1d272008-04-28 02:12:12 -07001722static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001723 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724{
1725 int priority;
kosaki.motohiro@jp.fujitsu.comc700be32008-06-12 15:21:27 -07001726 unsigned long ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001727 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001728 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 unsigned long lru_pages = 0;
Mel Gormandd1a2392008-04-28 02:12:17 -07001730 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001731 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -07001732 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733
Keika Kobayashi873b4772008-07-25 01:48:52 -07001734 delayacct_freepages_start();
1735
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001736 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001737 count_vm_event(ALLOCSTALL);
1738 /*
1739 * mem_cgroup will not do shrink_slab.
1740 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001741 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001742 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001744 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1745 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746
Wu Fengguangadea02a2009-09-21 17:01:42 -07001747 lru_pages += zone_reclaimable_pages(zone);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001748 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749 }
1750
1751 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001752 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001753 if (!priority)
1754 disable_swap_token();
Rik van Riela79311c2009-01-06 14:40:01 -08001755 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001756 /*
1757 * Don't shrink slabs when reclaiming memory from
1758 * over limit cgroups
1759 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001760 if (scanning_global_lru(sc)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07001761 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001762 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08001763 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001764 reclaim_state->reclaimed_slab = 0;
1765 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766 }
Balbir Singh66e17072008-02-07 00:13:56 -08001767 total_scanned += sc->nr_scanned;
Rik van Riela79311c2009-01-06 14:40:01 -08001768 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1769 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001770 goto out;
1771 }
1772
1773 /*
1774 * Try to write back as many pages as we just scanned. This
1775 * tends to cause slow streaming writers to write data to the
1776 * disk smoothly, at the dirtying rate, which is nice. But
1777 * that's undesirable in laptop mode, where we *want* lumpy
1778 * writeout. So in laptop mode, write out the whole world.
1779 */
Balbir Singh66e17072008-02-07 00:13:56 -08001780 if (total_scanned > sc->swap_cluster_max +
1781 sc->swap_cluster_max / 2) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001782 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001783 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001784 }
1785
1786 /* Take a nap, wait for some writeback to complete */
Andrew Morton4dd4b922008-03-24 12:29:52 -07001787 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02001788 congestion_wait(BLK_RW_ASYNC, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 }
Fernando Luis Vazquez Cao87547ee2008-07-29 22:33:42 -07001790 /* top priority shrink_zones still had more to do? don't OOM, then */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001791 if (!sc->all_unreclaimable && scanning_global_lru(sc))
Rik van Riela79311c2009-01-06 14:40:01 -08001792 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001794 /*
1795 * Now that we've scanned all the zones at this priority level, note
1796 * that level within the zone so that the next thread which performs
1797 * scanning of this zone will immediately start out at this priority
1798 * level. This affects only the decision whether or not to bring
1799 * mapped pages onto the inactive list.
1800 */
1801 if (priority < 0)
1802 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001804 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001805 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001807 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1808 continue;
1809
1810 zone->prev_priority = priority;
1811 }
1812 } else
1813 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1814
Keika Kobayashi873b4772008-07-25 01:48:52 -07001815 delayacct_freepages_end();
1816
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817 return ret;
1818}
1819
Mel Gormandac1d272008-04-28 02:12:12 -07001820unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001821 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08001822{
1823 struct scan_control sc = {
1824 .gfp_mask = gfp_mask,
1825 .may_writepage = !laptop_mode,
1826 .swap_cluster_max = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07001827 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001828 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08001829 .swappiness = vm_swappiness,
1830 .order = order,
1831 .mem_cgroup = NULL,
1832 .isolate_pages = isolate_pages_global,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001833 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08001834 };
1835
Mel Gormandd1a2392008-04-28 02:12:17 -07001836 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001837}
1838
Balbir Singh00f0b822008-03-04 14:28:39 -08001839#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08001840
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001841unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001842 gfp_t gfp_mask,
1843 bool noswap,
1844 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08001845{
1846 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08001847 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07001848 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001849 .may_swap = !noswap,
Balbir Singh66e17072008-02-07 00:13:56 -08001850 .swap_cluster_max = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001851 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08001852 .order = 0,
1853 .mem_cgroup = mem_cont,
1854 .isolate_pages = mem_cgroup_isolate_pages,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001855 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08001856 };
Mel Gormandac1d272008-04-28 02:12:12 -07001857 struct zonelist *zonelist;
Balbir Singh66e17072008-02-07 00:13:56 -08001858
Mel Gormandd1a2392008-04-28 02:12:17 -07001859 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1860 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1861 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1862 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001863}
1864#endif
1865
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866/*
1867 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07001868 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870 * Returns the number of pages which were actually freed.
1871 *
1872 * There is special handling here for zones which are full of pinned pages.
1873 * This can happen if the pages are all mlocked, or if they are all used by
1874 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1875 * What we do is to detect the case where all pages in the zone have been
1876 * scanned twice and there has been zero successful reclaim. Mark the zone as
1877 * dead and from now on, only perform a short scan. Basically we're polling
1878 * the zone for when the problem goes away.
1879 *
1880 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07001881 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1882 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1883 * lower zones regardless of the number of free pages in the lower zones. This
1884 * interoperates with the page allocator fallback scheme to ensure that aging
1885 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001887static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 int all_zones_ok;
1890 int priority;
1891 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08001892 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001893 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08001894 struct scan_control sc = {
1895 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07001896 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001897 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001898 .swap_cluster_max = SWAP_CLUSTER_MAX,
1899 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001900 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08001901 .mem_cgroup = NULL,
1902 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08001903 };
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001904 /*
1905 * temp_priority is used to remember the scanning priority at which
Mel Gorman41858962009-06-16 15:32:12 -07001906 * this zone was successfully refilled to
1907 * free_pages == high_wmark_pages(zone).
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001908 */
1909 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910
1911loop_again:
1912 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08001913 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07001914 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001915 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001917 for (i = 0; i < pgdat->nr_zones; i++)
1918 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919
1920 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1921 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1922 unsigned long lru_pages = 0;
1923
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001924 /* The swap token gets in the way of swapout... */
1925 if (!priority)
1926 disable_swap_token();
1927
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928 all_zones_ok = 1;
1929
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001930 /*
1931 * Scan in the highmem->dma direction for the highest
1932 * zone which needs scanning
1933 */
1934 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1935 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001937 if (!populated_zone(zone))
1938 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939
David Rientjese815af92007-10-16 23:25:54 -07001940 if (zone_is_all_unreclaimable(zone) &&
1941 priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001942 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001943
Rik van Riel556adec2008-10-18 20:26:34 -07001944 /*
1945 * Do some background aging of the anon list, to give
1946 * pages a chance to be referenced before reclaiming.
1947 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001948 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07001949 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1950 &sc, priority, 0);
1951
Mel Gorman41858962009-06-16 15:32:12 -07001952 if (!zone_watermark_ok(zone, order,
1953 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001954 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001955 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001958 if (i < 0)
1959 goto out;
1960
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961 for (i = 0; i <= end_zone; i++) {
1962 struct zone *zone = pgdat->node_zones + i;
1963
Wu Fengguangadea02a2009-09-21 17:01:42 -07001964 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001965 }
1966
1967 /*
1968 * Now scan the zone in the dma->highmem direction, stopping
1969 * at the last zone which needs scanning.
1970 *
1971 * We do this because the page allocator works in the opposite
1972 * direction. This prevents the page allocator from allocating
1973 * pages behind kswapd's direction of progress, which would
1974 * cause too much scanning of the lower zones.
1975 */
1976 for (i = 0; i <= end_zone; i++) {
1977 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001978 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979
Con Kolivasf3fe6512006-01-06 00:11:15 -08001980 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001981 continue;
1982
David Rientjese815af92007-10-16 23:25:54 -07001983 if (zone_is_all_unreclaimable(zone) &&
1984 priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001985 continue;
1986
Mel Gorman41858962009-06-16 15:32:12 -07001987 if (!zone_watermark_ok(zone, order,
1988 high_wmark_pages(zone), end_zone, 0))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001989 all_zones_ok = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001990 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 sc.nr_scanned = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001992 note_zone_scanning_priority(zone, priority);
Rik van Riel32a43302007-10-16 01:24:50 -07001993 /*
1994 * We put equal pressure on every zone, unless one
1995 * zone has way too many pages free already.
1996 */
Mel Gorman41858962009-06-16 15:32:12 -07001997 if (!zone_watermark_ok(zone, order,
1998 8*high_wmark_pages(zone), end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08001999 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002000 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002001 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2002 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002003 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002004 total_scanned += sc.nr_scanned;
David Rientjese815af92007-10-16 23:25:54 -07002005 if (zone_is_all_unreclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002007 if (nr_slab == 0 && zone->pages_scanned >=
Wu Fengguangadea02a2009-09-21 17:01:42 -07002008 (zone_reclaimable_pages(zone) * 6))
David Rientjese815af92007-10-16 23:25:54 -07002009 zone_set_flag(zone,
2010 ZONE_ALL_UNRECLAIMABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011 /*
2012 * If we've done a decent amount of scanning and
2013 * the reclaim ratio is low, start doing writepage
2014 * even in laptop mode
2015 */
2016 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002017 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018 sc.may_writepage = 1;
2019 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020 if (all_zones_ok)
2021 break; /* kswapd: all done */
2022 /*
2023 * OK, kswapd is getting into trouble. Take a nap, then take
2024 * another pass across the zones.
2025 */
Andrew Morton4dd4b922008-03-24 12:29:52 -07002026 if (total_scanned && priority < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02002027 congestion_wait(BLK_RW_ASYNC, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028
2029 /*
2030 * We do this so kswapd doesn't build up large priorities for
2031 * example when it is freeing in parallel with allocators. It
2032 * matches the direct reclaim path behaviour in terms of impact
2033 * on zone->*_priority.
2034 */
Rik van Riela79311c2009-01-06 14:40:01 -08002035 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 break;
2037 }
2038out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07002039 /*
2040 * Note within each zone the priority level at which this zone was
2041 * brought into a happy state. So that the next thread which scans this
2042 * zone will start out at that priority level.
2043 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044 for (i = 0; i < pgdat->nr_zones; i++) {
2045 struct zone *zone = pgdat->node_zones + i;
2046
Martin Bligh3bb1a8522006-10-28 10:38:24 -07002047 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002048 }
2049 if (!all_zones_ok) {
2050 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002051
2052 try_to_freeze();
2053
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002054 /*
2055 * Fragmentation may mean that the system cannot be
2056 * rebalanced for high-order allocations in all zones.
2057 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2058 * it means the zones have been fully scanned and are still
2059 * not balanced. For high-order allocations, there is
2060 * little point trying all over again as kswapd may
2061 * infinite loop.
2062 *
2063 * Instead, recheck all watermarks at order-0 as they
2064 * are the most important. If watermarks are ok, kswapd will go
2065 * back to sleep. High-order users can still perform direct
2066 * reclaim if they wish.
2067 */
2068 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2069 order = sc.order = 0;
2070
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071 goto loop_again;
2072 }
2073
Rik van Riela79311c2009-01-06 14:40:01 -08002074 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002075}
2076
2077/*
2078 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002079 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002080 *
2081 * This basically trickles out pages so that we have _some_
2082 * free memory available even if there is no other activity
2083 * that frees anything up. This is needed for things like routing
2084 * etc, where we otherwise might have all activity going on in
2085 * asynchronous contexts that cannot page things out.
2086 *
2087 * If there are applications that are active memory-allocators
2088 * (most normal use), this basically shouldn't matter.
2089 */
2090static int kswapd(void *p)
2091{
2092 unsigned long order;
2093 pg_data_t *pgdat = (pg_data_t*)p;
2094 struct task_struct *tsk = current;
2095 DEFINE_WAIT(wait);
2096 struct reclaim_state reclaim_state = {
2097 .reclaimed_slab = 0,
2098 };
Rusty Russella70f7302009-03-13 14:49:46 +10302099 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100
Nick Piggincf40bd12009-01-21 08:12:39 +01002101 lockdep_set_current_reclaim_state(GFP_KERNEL);
2102
Rusty Russell174596a2009-01-01 10:12:29 +10302103 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002104 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105 current->reclaim_state = &reclaim_state;
2106
2107 /*
2108 * Tell the memory management that we're a "memory allocator",
2109 * and that if we need more memory we should get access to it
2110 * regardless (see "__alloc_pages()"). "kswapd" should
2111 * never get caught in the normal page freeing logic.
2112 *
2113 * (Kswapd normally doesn't need memory anyway, but sometimes
2114 * you need a small amount of memory in order to be able to
2115 * page out something else, and this flag essentially protects
2116 * us from recursively trying to free more memory as we're
2117 * trying to free the first piece of memory in the first place).
2118 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002119 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002120 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121
2122 order = 0;
2123 for ( ; ; ) {
2124 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002125
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2127 new_order = pgdat->kswapd_max_order;
2128 pgdat->kswapd_max_order = 0;
2129 if (order < new_order) {
2130 /*
2131 * Don't sleep if someone wants a larger 'order'
2132 * allocation
2133 */
2134 order = new_order;
2135 } else {
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002136 if (!freezing(current))
2137 schedule();
2138
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 order = pgdat->kswapd_max_order;
2140 }
2141 finish_wait(&pgdat->kswapd_wait, &wait);
2142
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002143 if (!try_to_freeze()) {
2144 /* We can speed up thawing tasks if we don't call
2145 * balance_pgdat after returning from the refrigerator
2146 */
2147 balance_pgdat(pgdat, order);
2148 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002149 }
2150 return 0;
2151}
2152
2153/*
2154 * A zone is low on free memory, so wake its kswapd task to service it.
2155 */
2156void wakeup_kswapd(struct zone *zone, int order)
2157{
2158 pg_data_t *pgdat;
2159
Con Kolivasf3fe6512006-01-06 00:11:15 -08002160 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002161 return;
2162
2163 pgdat = zone->zone_pgdat;
Mel Gorman41858962009-06-16 15:32:12 -07002164 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165 return;
2166 if (pgdat->kswapd_max_order < order)
2167 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08002168 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002170 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002172 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173}
2174
Wu Fengguangadea02a2009-09-21 17:01:42 -07002175/*
2176 * The reclaimable count would be mostly accurate.
2177 * The less reclaimable pages may be
2178 * - mlocked pages, which will be moved to unevictable list when encountered
2179 * - mapped pages, which may require several travels to be reclaimed
2180 * - dirty pages, which is not "instantly" reclaimable
2181 */
2182unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002183{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002184 int nr;
2185
2186 nr = global_page_state(NR_ACTIVE_FILE) +
2187 global_page_state(NR_INACTIVE_FILE);
2188
2189 if (nr_swap_pages > 0)
2190 nr += global_page_state(NR_ACTIVE_ANON) +
2191 global_page_state(NR_INACTIVE_ANON);
2192
2193 return nr;
2194}
2195
2196unsigned long zone_reclaimable_pages(struct zone *zone)
2197{
2198 int nr;
2199
2200 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2201 zone_page_state(zone, NR_INACTIVE_FILE);
2202
2203 if (nr_swap_pages > 0)
2204 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2205 zone_page_state(zone, NR_INACTIVE_ANON);
2206
2207 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002208}
2209
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002210#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211/*
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002212 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
MinChan Kimd9796772009-03-31 15:19:34 -07002213 * from LRU lists system-wide, for given pass and priority.
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002214 *
2215 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2216 */
MinChan Kimd9796772009-03-31 15:19:34 -07002217static void shrink_all_zones(unsigned long nr_pages, int prio,
Nigel Cunninghame07aa052006-12-22 01:07:21 -08002218 int pass, struct scan_control *sc)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002219{
2220 struct zone *zone;
MinChan Kimd9796772009-03-31 15:19:34 -07002221 unsigned long nr_reclaimed = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002222
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002223 for_each_populated_zone(zone) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002224 enum lru_list l;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002225
David Rientjese815af92007-10-16 23:25:54 -07002226 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002227 continue;
2228
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002229 for_each_evictable_lru(l) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002230 enum zone_stat_item ls = NR_LRU_BASE + l;
2231 unsigned long lru_pages = zone_page_state(zone, ls);
2232
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002233 /* For pass = 0, we don't shrink the active list */
Johannes Weiner0cb57252009-02-14 02:04:10 +01002234 if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2235 l == LRU_ACTIVE_FILE))
Christoph Lameterb69408e2008-10-18 20:26:14 -07002236 continue;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002237
Wu Fengguang6e08a362009-06-16 15:32:29 -07002238 zone->lru[l].nr_saved_scan += (lru_pages >> prio) + 1;
2239 if (zone->lru[l].nr_saved_scan >= nr_pages || pass > 3) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002240 unsigned long nr_to_scan;
2241
Wu Fengguang6e08a362009-06-16 15:32:29 -07002242 zone->lru[l].nr_saved_scan = 0;
Johannes Weiner0cb57252009-02-14 02:04:10 +01002243 nr_to_scan = min(nr_pages, lru_pages);
MinChan Kimd9796772009-03-31 15:19:34 -07002244 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
Christoph Lameterb69408e2008-10-18 20:26:14 -07002245 sc, prio);
MinChan Kimd9796772009-03-31 15:19:34 -07002246 if (nr_reclaimed >= nr_pages) {
Rafael J. Wysockia21e2552009-04-18 17:23:41 +02002247 sc->nr_reclaimed += nr_reclaimed;
MinChan Kimd9796772009-03-31 15:19:34 -07002248 return;
2249 }
Christoph Lameterb69408e2008-10-18 20:26:14 -07002250 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002251 }
2252 }
Rafael J. Wysockia21e2552009-04-18 17:23:41 +02002253 sc->nr_reclaimed += nr_reclaimed;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002254}
2255
2256/*
2257 * Try to free `nr_pages' of memory, system-wide, and return the number of
2258 * freed pages.
2259 *
2260 * Rather than trying to age LRUs the aim is to preserve the overall
2261 * LRU order by reclaiming preferentially
2262 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002263 */
Andrew Morton69e05942006-03-22 00:08:19 -08002264unsigned long shrink_all_memory(unsigned long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002266 unsigned long lru_pages, nr_slab;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002267 int pass;
2268 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002269 struct scan_control sc = {
2270 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002271 .may_unmap = 0,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002272 .may_writepage = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002273 .isolate_pages = isolate_pages_global,
Rafael J. Wysockia21e2552009-04-18 17:23:41 +02002274 .nr_reclaimed = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 };
2276
2277 current->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002278
Wu Fengguangadea02a2009-09-21 17:01:42 -07002279 lru_pages = global_reclaimable_pages();
Christoph Lameter972d1a72006-09-25 23:31:51 -07002280 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002281 /* If slab caches are huge, it's better to hit them first */
2282 while (nr_slab >= lru_pages) {
2283 reclaim_state.reclaimed_slab = 0;
2284 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2285 if (!reclaim_state.reclaimed_slab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 break;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002287
MinChan Kimd9796772009-03-31 15:19:34 -07002288 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2289 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002290 goto out;
2291
2292 nr_slab -= reclaim_state.reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002294
2295 /*
2296 * We try to shrink LRUs in 5 passes:
2297 * 0 = Reclaim from inactive_list only
2298 * 1 = Reclaim from active list but don't reclaim mapped
2299 * 2 = 2nd pass of type 1
2300 * 3 = Reclaim mapped (normal reclaim)
2301 * 4 = 2nd pass of type 3
2302 */
2303 for (pass = 0; pass < 5; pass++) {
2304 int prio;
2305
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002306 /* Force reclaiming mapped pages in the passes #3 and #4 */
Johannes Weiner30491032009-02-14 02:03:08 +01002307 if (pass > 2)
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002308 sc.may_unmap = 1;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002309
2310 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
MinChan Kimd9796772009-03-31 15:19:34 -07002311 unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002312
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002313 sc.nr_scanned = 0;
Johannes Weiner9786bf82009-03-31 15:19:35 -07002314 sc.swap_cluster_max = nr_to_scan;
MinChan Kimd9796772009-03-31 15:19:34 -07002315 shrink_all_zones(nr_to_scan, prio, pass, &sc);
2316 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002317 goto out;
2318
2319 reclaim_state.reclaimed_slab = 0;
Andrew Morton76395d32007-01-05 16:37:05 -08002320 shrink_slab(sc.nr_scanned, sc.gfp_mask,
Wu Fengguangadea02a2009-09-21 17:01:42 -07002321 global_reclaimable_pages());
MinChan Kimd9796772009-03-31 15:19:34 -07002322 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2323 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002324 goto out;
2325
2326 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02002327 congestion_wait(BLK_RW_ASYNC, HZ / 10);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002328 }
Rafael J. Wysocki248a0302006-03-22 00:09:04 -08002329 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002330
2331 /*
MinChan Kimd9796772009-03-31 15:19:34 -07002332 * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2333 * something in slab caches
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002334 */
MinChan Kimd9796772009-03-31 15:19:34 -07002335 if (!sc.nr_reclaimed) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002336 do {
2337 reclaim_state.reclaimed_slab = 0;
Wu Fengguangadea02a2009-09-21 17:01:42 -07002338 shrink_slab(nr_pages, sc.gfp_mask,
2339 global_reclaimable_pages());
MinChan Kimd9796772009-03-31 15:19:34 -07002340 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2341 } while (sc.nr_reclaimed < nr_pages &&
2342 reclaim_state.reclaimed_slab > 0);
Andrew Morton76395d32007-01-05 16:37:05 -08002343 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002344
MinChan Kimd9796772009-03-31 15:19:34 -07002345
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002346out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002347 current->reclaim_state = NULL;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002348
MinChan Kimd9796772009-03-31 15:19:34 -07002349 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002351#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002352
Linus Torvalds1da177e2005-04-16 15:20:36 -07002353/* It's optimal to keep kswapds on the same CPUs as their memory, but
2354 not required for correctness. So if the last cpu in a node goes
2355 away, we get changed to run anywhere: as the first one comes back,
2356 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002357static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002358 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002360 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002362 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002363 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002364 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302365 const struct cpumask *mask;
2366
2367 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002368
Rusty Russell3e597942009-01-01 10:12:24 +10302369 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002371 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372 }
2373 }
2374 return NOTIFY_OK;
2375}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376
Yasunori Goto3218ae12006-06-27 02:53:33 -07002377/*
2378 * This kswapd start function will be called by init and node-hot-add.
2379 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2380 */
2381int kswapd_run(int nid)
2382{
2383 pg_data_t *pgdat = NODE_DATA(nid);
2384 int ret = 0;
2385
2386 if (pgdat->kswapd)
2387 return 0;
2388
2389 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2390 if (IS_ERR(pgdat->kswapd)) {
2391 /* failure at boot is fatal */
2392 BUG_ON(system_state == SYSTEM_BOOTING);
2393 printk("Failed to start kswapd on node %d\n",nid);
2394 ret = -1;
2395 }
2396 return ret;
2397}
2398
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399static int __init kswapd_init(void)
2400{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002401 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002402
Linus Torvalds1da177e2005-04-16 15:20:36 -07002403 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002404 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002405 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406 hotcpu_notifier(cpu_callback, 0);
2407 return 0;
2408}
2409
2410module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002411
2412#ifdef CONFIG_NUMA
2413/*
2414 * Zone reclaim mode
2415 *
2416 * If non-zero call zone_reclaim when the number of free pages falls below
2417 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002418 */
2419int zone_reclaim_mode __read_mostly;
2420
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002421#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002422#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002423#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2424#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2425
Christoph Lameter9eeff232006-01-18 17:42:31 -08002426/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002427 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2428 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2429 * a zone.
2430 */
2431#define ZONE_RECLAIM_PRIORITY 4
2432
Christoph Lameter9eeff232006-01-18 17:42:31 -08002433/*
Christoph Lameter96146342006-07-03 00:24:13 -07002434 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2435 * occur.
2436 */
2437int sysctl_min_unmapped_ratio = 1;
2438
2439/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002440 * If the number of slab pages in a zone grows beyond this percentage then
2441 * slab reclaim needs to occur.
2442 */
2443int sysctl_min_slab_ratio = 5;
2444
Mel Gorman90afa5d2009-06-16 15:33:20 -07002445static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2446{
2447 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2448 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2449 zone_page_state(zone, NR_ACTIVE_FILE);
2450
2451 /*
2452 * It's possible for there to be more file mapped pages than
2453 * accounted for by the pages on the file LRU lists because
2454 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2455 */
2456 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2457}
2458
2459/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2460static long zone_pagecache_reclaimable(struct zone *zone)
2461{
2462 long nr_pagecache_reclaimable;
2463 long delta = 0;
2464
2465 /*
2466 * If RECLAIM_SWAP is set, then all file pages are considered
2467 * potentially reclaimable. Otherwise, we have to worry about
2468 * pages like swapcache and zone_unmapped_file_pages() provides
2469 * a better estimate
2470 */
2471 if (zone_reclaim_mode & RECLAIM_SWAP)
2472 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2473 else
2474 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2475
2476 /* If we can't clean pages, remove dirty pages from consideration */
2477 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2478 delta += zone_page_state(zone, NR_FILE_DIRTY);
2479
2480 /* Watch for any possible underflows due to delta */
2481 if (unlikely(delta > nr_pagecache_reclaimable))
2482 delta = nr_pagecache_reclaimable;
2483
2484 return nr_pagecache_reclaimable - delta;
2485}
2486
Christoph Lameter0ff38492006-09-25 23:31:52 -07002487/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002488 * Try to free up some pages from this zone through reclaim.
2489 */
Andrew Morton179e9632006-03-22 00:08:18 -08002490static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002491{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002492 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002493 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002494 struct task_struct *p = current;
2495 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002496 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002497 struct scan_control sc = {
2498 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002499 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002500 .may_swap = 1,
Andrew Morton69e05942006-03-22 00:08:19 -08002501 .swap_cluster_max = max_t(unsigned long, nr_pages,
2502 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002503 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002504 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002505 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002506 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08002507 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07002508 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002509
2510 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08002511 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002512 /*
2513 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2514 * and we also need to be able to write out pages for RECLAIM_WRITE
2515 * and RECLAIM_SWAP.
2516 */
2517 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002518 reclaim_state.reclaimed_slab = 0;
2519 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002520
Mel Gorman90afa5d2009-06-16 15:33:20 -07002521 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07002522 /*
2523 * Free memory by calling shrink zone with increasing
2524 * priorities until we have enough memory freed.
2525 */
2526 priority = ZONE_RECLAIM_PRIORITY;
2527 do {
Martin Bligh3bb1a8522006-10-28 10:38:24 -07002528 note_zone_scanning_priority(zone, priority);
Rik van Riela79311c2009-01-06 14:40:01 -08002529 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002530 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002531 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002532 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002533
Christoph Lameter83e33a42006-09-25 23:31:53 -07002534 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2535 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002536 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002537 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002538 * many pages were freed in this zone. So we take the current
2539 * number of slab pages and shake the slab until it is reduced
2540 * by the same nr_pages that we used for reclaiming unmapped
2541 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002542 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002543 * Note that shrink_slab will free memory on all zones and may
2544 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002545 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07002546 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07002547 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2548 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07002549 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07002550
2551 /*
2552 * Update nr_reclaimed by the number of slab pages we
2553 * reclaimed from this zone.
2554 */
Rik van Riela79311c2009-01-06 14:40:01 -08002555 sc.nr_reclaimed += slab_reclaimable -
Christoph Lameter83e33a42006-09-25 23:31:53 -07002556 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002557 }
2558
Christoph Lameter9eeff232006-01-18 17:42:31 -08002559 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002560 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
Rik van Riela79311c2009-01-06 14:40:01 -08002561 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002562}
Andrew Morton179e9632006-03-22 00:08:18 -08002563
2564int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2565{
Andrew Morton179e9632006-03-22 00:08:18 -08002566 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002567 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002568
2569 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002570 * Zone reclaim reclaims unmapped file backed pages and
2571 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002572 *
Christoph Lameter96146342006-07-03 00:24:13 -07002573 * A small portion of unmapped file backed pages is needed for
2574 * file I/O otherwise pages read by file I/O will be immediately
2575 * thrown out if the zone is overallocated. So we do not reclaim
2576 * if less than a specified percentage of the zone is used by
2577 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002578 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07002579 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2580 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002581 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08002582
David Rientjesd773ed62007-10-16 23:26:01 -07002583 if (zone_is_all_unreclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002584 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07002585
Andrew Morton179e9632006-03-22 00:08:18 -08002586 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002587 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002588 */
David Rientjesd773ed62007-10-16 23:26:01 -07002589 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002590 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08002591
2592 /*
2593 * Only run zone reclaim on the local zone or on zones that do not
2594 * have associated processors. This will favor the local processor
2595 * over remote processors and spread off node memory allocations
2596 * as wide as possible.
2597 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002598 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002599 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07002600 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07002601
2602 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002603 return ZONE_RECLAIM_NOSCAN;
2604
David Rientjesd773ed62007-10-16 23:26:01 -07002605 ret = __zone_reclaim(zone, gfp_mask, order);
2606 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2607
Mel Gorman24cf725182009-06-16 15:33:23 -07002608 if (!ret)
2609 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2610
David Rientjesd773ed62007-10-16 23:26:01 -07002611 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002612}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002613#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002614
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002615/*
2616 * page_evictable - test whether a page is evictable
2617 * @page: the page to test
2618 * @vma: the VMA in which the page is or will be mapped, may be NULL
2619 *
2620 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07002621 * lists vs unevictable list. The vma argument is !NULL when called from the
2622 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002623 *
2624 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002625 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07002626 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002627 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002628 */
2629int page_evictable(struct page *page, struct vm_area_struct *vma)
2630{
2631
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002632 if (mapping_unevictable(page_mapping(page)))
2633 return 0;
2634
Nick Pigginb291f002008-10-18 20:26:44 -07002635 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2636 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002637
2638 return 1;
2639}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002640
2641/**
2642 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2643 * @page: page to check evictability and move to appropriate lru list
2644 * @zone: zone page is in
2645 *
2646 * Checks a page for evictability and moves the page to the appropriate
2647 * zone lru list.
2648 *
2649 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2650 * have PageUnevictable set.
2651 */
2652static void check_move_unevictable_page(struct page *page, struct zone *zone)
2653{
2654 VM_BUG_ON(PageActive(page));
2655
2656retry:
2657 ClearPageUnevictable(page);
2658 if (page_evictable(page, NULL)) {
2659 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002660
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002661 __dec_zone_state(zone, NR_UNEVICTABLE);
2662 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002663 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002664 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2665 __count_vm_event(UNEVICTABLE_PGRESCUED);
2666 } else {
2667 /*
2668 * rotate unevictable list
2669 */
2670 SetPageUnevictable(page);
2671 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002672 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002673 if (page_evictable(page, NULL))
2674 goto retry;
2675 }
2676}
2677
2678/**
2679 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2680 * @mapping: struct address_space to scan for evictable pages
2681 *
2682 * Scan all pages in mapping. Check unevictable pages for
2683 * evictability and move them to the appropriate zone lru list.
2684 */
2685void scan_mapping_unevictable_pages(struct address_space *mapping)
2686{
2687 pgoff_t next = 0;
2688 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2689 PAGE_CACHE_SHIFT;
2690 struct zone *zone;
2691 struct pagevec pvec;
2692
2693 if (mapping->nrpages == 0)
2694 return;
2695
2696 pagevec_init(&pvec, 0);
2697 while (next < end &&
2698 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2699 int i;
2700 int pg_scanned = 0;
2701
2702 zone = NULL;
2703
2704 for (i = 0; i < pagevec_count(&pvec); i++) {
2705 struct page *page = pvec.pages[i];
2706 pgoff_t page_index = page->index;
2707 struct zone *pagezone = page_zone(page);
2708
2709 pg_scanned++;
2710 if (page_index > next)
2711 next = page_index;
2712 next++;
2713
2714 if (pagezone != zone) {
2715 if (zone)
2716 spin_unlock_irq(&zone->lru_lock);
2717 zone = pagezone;
2718 spin_lock_irq(&zone->lru_lock);
2719 }
2720
2721 if (PageLRU(page) && PageUnevictable(page))
2722 check_move_unevictable_page(page, zone);
2723 }
2724 if (zone)
2725 spin_unlock_irq(&zone->lru_lock);
2726 pagevec_release(&pvec);
2727
2728 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2729 }
2730
2731}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002732
2733/**
2734 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2735 * @zone - zone of which to scan the unevictable list
2736 *
2737 * Scan @zone's unevictable LRU lists to check for pages that have become
2738 * evictable. Move those that have to @zone's inactive list where they
2739 * become candidates for reclaim, unless shrink_inactive_zone() decides
2740 * to reactivate them. Pages that are still unevictable are rotated
2741 * back onto @zone's unevictable list.
2742 */
2743#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08002744static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002745{
2746 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2747 unsigned long scan;
2748 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2749
2750 while (nr_to_scan > 0) {
2751 unsigned long batch_size = min(nr_to_scan,
2752 SCAN_UNEVICTABLE_BATCH_SIZE);
2753
2754 spin_lock_irq(&zone->lru_lock);
2755 for (scan = 0; scan < batch_size; scan++) {
2756 struct page *page = lru_to_page(l_unevictable);
2757
2758 if (!trylock_page(page))
2759 continue;
2760
2761 prefetchw_prev_lru_page(page, l_unevictable, flags);
2762
2763 if (likely(PageLRU(page) && PageUnevictable(page)))
2764 check_move_unevictable_page(page, zone);
2765
2766 unlock_page(page);
2767 }
2768 spin_unlock_irq(&zone->lru_lock);
2769
2770 nr_to_scan -= batch_size;
2771 }
2772}
2773
2774
2775/**
2776 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2777 *
2778 * A really big hammer: scan all zones' unevictable LRU lists to check for
2779 * pages that have become evictable. Move those back to the zones'
2780 * inactive list where they become candidates for reclaim.
2781 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2782 * and we add swap to the system. As such, it runs in the context of a task
2783 * that has possibly/probably made some previously unevictable pages
2784 * evictable.
2785 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08002786static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002787{
2788 struct zone *zone;
2789
2790 for_each_zone(zone) {
2791 scan_zone_unevictable_pages(zone);
2792 }
2793}
2794
2795/*
2796 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2797 * all nodes' unevictable lists for evictable pages
2798 */
2799unsigned long scan_unevictable_pages;
2800
2801int scan_unevictable_handler(struct ctl_table *table, int write,
2802 struct file *file, void __user *buffer,
2803 size_t *length, loff_t *ppos)
2804{
2805 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2806
2807 if (write && *(unsigned long *)table->data)
2808 scan_all_zones_unevictable_pages();
2809
2810 scan_unevictable_pages = 0;
2811 return 0;
2812}
2813
2814/*
2815 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2816 * a specified node's per zone unevictable lists for evictable pages.
2817 */
2818
2819static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2820 struct sysdev_attribute *attr,
2821 char *buf)
2822{
2823 return sprintf(buf, "0\n"); /* always zero; should fit... */
2824}
2825
2826static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2827 struct sysdev_attribute *attr,
2828 const char *buf, size_t count)
2829{
2830 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2831 struct zone *zone;
2832 unsigned long res;
2833 unsigned long req = strict_strtoul(buf, 10, &res);
2834
2835 if (!req)
2836 return 1; /* zero is no-op */
2837
2838 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2839 if (!populated_zone(zone))
2840 continue;
2841 scan_zone_unevictable_pages(zone);
2842 }
2843 return 1;
2844}
2845
2846
2847static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2848 read_scan_unevictable_node,
2849 write_scan_unevictable_node);
2850
2851int scan_unevictable_register_node(struct node *node)
2852{
2853 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2854}
2855
2856void scan_unevictable_unregister_node(struct node *node)
2857{
2858 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2859}
2860