blob: 777af57fd8c8c80971d7abaf4edb974e360288c9 [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 Weinera6dc60f2009-03-31 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
Vincent Li0b217672009-09-21 17:03:09 -0700151static unsigned long zone_nr_lru_pages(struct zone *zone,
152 struct scan_control *sc, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800153{
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{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700289 /*
290 * A freeable page cache page is referenced only by the caller
291 * that isolated the page, the page cache radix tree and
292 * optional buffer heads at page->private.
293 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700294 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700295}
296
297static int may_write_to_queue(struct backing_dev_info *bdi)
298{
Christoph Lameter930d9152006-01-08 01:00:47 -0800299 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700300 return 1;
301 if (!bdi_write_congested(bdi))
302 return 1;
303 if (bdi == current->backing_dev_info)
304 return 1;
305 return 0;
306}
307
308/*
309 * We detected a synchronous write error writing a page out. Probably
310 * -ENOSPC. We need to propagate that into the address_space for a subsequent
311 * fsync(), msync() or close().
312 *
313 * The tricky part is that after writepage we cannot touch the mapping: nothing
314 * prevents it from being freed up. But we have a ref on the page and once
315 * that page is locked, the mapping is pinned.
316 *
317 * We're allowed to run sleeping lock_page() here because we know the caller has
318 * __GFP_FS.
319 */
320static void handle_write_error(struct address_space *mapping,
321 struct page *page, int error)
322{
323 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700324 if (page_mapping(page) == mapping)
325 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326 unlock_page(page);
327}
328
Andy Whitcroftc661b072007-08-22 14:01:26 -0700329/* Request for sync pageout. */
330enum pageout_io {
331 PAGEOUT_IO_ASYNC,
332 PAGEOUT_IO_SYNC,
333};
334
Christoph Lameter04e62a22006-06-23 02:03:38 -0700335/* possible outcome of pageout() */
336typedef enum {
337 /* failed to write page out, page is locked */
338 PAGE_KEEP,
339 /* move page to the active list, page is locked */
340 PAGE_ACTIVATE,
341 /* page has been sent to the disk successfully, page is unlocked */
342 PAGE_SUCCESS,
343 /* page is clean and locked */
344 PAGE_CLEAN,
345} pageout_t;
346
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347/*
Andrew Morton1742f192006-03-22 00:08:21 -0800348 * pageout is called by shrink_page_list() for each dirty page.
349 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700350 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700351static pageout_t pageout(struct page *page, struct address_space *mapping,
352 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353{
354 /*
355 * If the page is dirty, only perform writeback if that write
356 * will be non-blocking. To prevent this allocation from being
357 * stalled by pagecache activity. But note that there may be
358 * stalls if we need to run get_block(). We could test
359 * PagePrivate for that.
360 *
361 * If this process is currently in generic_file_write() against
362 * this page's queue, we can perform writeback even if that
363 * will block.
364 *
365 * If the page is swapcache, write it back even if that would
366 * block, for some throttling. This happens by accident, because
367 * swap_backing_dev_info is bust: it doesn't reflect the
368 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 */
370 if (!is_page_cache_freeable(page))
371 return PAGE_KEEP;
372 if (!mapping) {
373 /*
374 * Some data journaling orphaned pages can have
375 * page->mapping == NULL while being dirty with clean buffers.
376 */
David Howells266cf652009-04-03 16:42:36 +0100377 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 if (try_to_free_buffers(page)) {
379 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700380 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 return PAGE_CLEAN;
382 }
383 }
384 return PAGE_KEEP;
385 }
386 if (mapping->a_ops->writepage == NULL)
387 return PAGE_ACTIVATE;
388 if (!may_write_to_queue(mapping->backing_dev_info))
389 return PAGE_KEEP;
390
391 if (clear_page_dirty_for_io(page)) {
392 int res;
393 struct writeback_control wbc = {
394 .sync_mode = WB_SYNC_NONE,
395 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700396 .range_start = 0,
397 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700398 .nonblocking = 1,
399 .for_reclaim = 1,
400 };
401
402 SetPageReclaim(page);
403 res = mapping->a_ops->writepage(page, &wbc);
404 if (res < 0)
405 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800406 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700407 ClearPageReclaim(page);
408 return PAGE_ACTIVATE;
409 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700410
411 /*
412 * Wait on writeback if requested to. This happens when
413 * direct reclaiming a large contiguous area and the
414 * first attempt to free a range of pages fails.
415 */
416 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
417 wait_on_page_writeback(page);
418
Linus Torvalds1da177e2005-04-16 15:20:36 -0700419 if (!PageWriteback(page)) {
420 /* synchronous write or broken a_ops? */
421 ClearPageReclaim(page);
422 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700423 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700424 return PAGE_SUCCESS;
425 }
426
427 return PAGE_CLEAN;
428}
429
Andrew Mortona649fd92006-10-17 00:09:36 -0700430/*
Nick Piggine2867812008-07-25 19:45:30 -0700431 * Same as remove_mapping, but if the page is removed from the mapping, it
432 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700433 */
Nick Piggine2867812008-07-25 19:45:30 -0700434static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800435{
Nick Piggin28e4d962006-09-25 23:31:23 -0700436 BUG_ON(!PageLocked(page));
437 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800438
Nick Piggin19fd6232008-07-25 19:45:32 -0700439 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800440 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700441 * The non racy check for a busy page.
442 *
443 * Must be careful with the order of the tests. When someone has
444 * a ref to the page, it may be possible that they dirty it then
445 * drop the reference. So if PageDirty is tested before page_count
446 * here, then the following race may occur:
447 *
448 * get_user_pages(&page);
449 * [user mapping goes away]
450 * write_to(page);
451 * !PageDirty(page) [good]
452 * SetPageDirty(page);
453 * put_page(page);
454 * !page_count(page) [good, discard it]
455 *
456 * [oops, our write_to data is lost]
457 *
458 * Reversing the order of the tests ensures such a situation cannot
459 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
460 * load is not satisfied before that of page->_count.
461 *
462 * Note that if SetPageDirty is always performed via set_page_dirty,
463 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800464 */
Nick Piggine2867812008-07-25 19:45:30 -0700465 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800466 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700467 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
468 if (unlikely(PageDirty(page))) {
469 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800470 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700471 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800472
473 if (PageSwapCache(page)) {
474 swp_entry_t swap = { .val = page_private(page) };
475 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700476 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700477 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700478 } else {
479 __remove_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700480 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700481 mem_cgroup_uncharge_cache_page(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800482 }
483
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800484 return 1;
485
486cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700487 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800488 return 0;
489}
490
Linus Torvalds1da177e2005-04-16 15:20:36 -0700491/*
Nick Piggine2867812008-07-25 19:45:30 -0700492 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
493 * someone else has a ref on the page, abort and return 0. If it was
494 * successfully detached, return 1. Assumes the caller has a single ref on
495 * this page.
496 */
497int remove_mapping(struct address_space *mapping, struct page *page)
498{
499 if (__remove_mapping(mapping, page)) {
500 /*
501 * Unfreezing the refcount with 1 rather than 2 effectively
502 * drops the pagecache ref for us without requiring another
503 * atomic operation.
504 */
505 page_unfreeze_refs(page, 1);
506 return 1;
507 }
508 return 0;
509}
510
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700511/**
512 * putback_lru_page - put previously isolated page onto appropriate LRU list
513 * @page: page to be put back to appropriate lru list
514 *
515 * Add previously isolated @page to appropriate LRU list.
516 * Page may still be unevictable for other reasons.
517 *
518 * lru_lock must not be held, interrupts must be enabled.
519 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700520void putback_lru_page(struct page *page)
521{
522 int lru;
523 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700524 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700525
526 VM_BUG_ON(PageLRU(page));
527
528redo:
529 ClearPageUnevictable(page);
530
531 if (page_evictable(page, NULL)) {
532 /*
533 * For evictable pages, we can use the cache.
534 * In event of a race, worst case is we end up with an
535 * unevictable page on [in]active list.
536 * We know how to handle that.
537 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700538 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700539 lru_cache_add_lru(page, lru);
540 } else {
541 /*
542 * Put unevictable pages directly on zone's unevictable
543 * list.
544 */
545 lru = LRU_UNEVICTABLE;
546 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700547 /*
548 * When racing with an mlock clearing (page is
549 * unlocked), make sure that if the other thread does
550 * not observe our setting of PG_lru and fails
551 * isolation, we see PG_mlocked cleared below and move
552 * the page back to the evictable list.
553 *
554 * The other side is TestClearPageMlocked().
555 */
556 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700557 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700558
559 /*
560 * page's status can change while we move it among lru. If an evictable
561 * page is on unevictable list, it never be freed. To avoid that,
562 * check after we added it to the list, again.
563 */
564 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
565 if (!isolate_lru_page(page)) {
566 put_page(page);
567 goto redo;
568 }
569 /* This means someone else dropped this page from LRU
570 * So, it will be freed or putback to LRU again. There is
571 * nothing to do here.
572 */
573 }
574
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700575 if (was_unevictable && lru != LRU_UNEVICTABLE)
576 count_vm_event(UNEVICTABLE_PGRESCUED);
577 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
578 count_vm_event(UNEVICTABLE_PGCULLED);
579
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700580 put_page(page); /* drop ref from isolate */
581}
582
Nick Piggine2867812008-07-25 19:45:30 -0700583/*
Andrew Morton1742f192006-03-22 00:08:21 -0800584 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585 */
Andrew Morton1742f192006-03-22 00:08:21 -0800586static unsigned long shrink_page_list(struct list_head *page_list,
Andy Whitcroftc661b072007-08-22 14:01:26 -0700587 struct scan_control *sc,
588 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700589{
590 LIST_HEAD(ret_pages);
591 struct pagevec freed_pvec;
592 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800593 unsigned long nr_reclaimed = 0;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700594 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595
596 cond_resched();
597
598 pagevec_init(&freed_pvec, 1);
599 while (!list_empty(page_list)) {
600 struct address_space *mapping;
601 struct page *page;
602 int may_enter_fs;
603 int referenced;
604
605 cond_resched();
606
607 page = lru_to_page(page_list);
608 list_del(&page->lru);
609
Nick Piggin529ae9a2008-08-02 12:01:03 +0200610 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611 goto keep;
612
Nick Piggin725d7042006-09-25 23:30:55 -0700613 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614
615 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800616
Nick Pigginb291f002008-10-18 20:26:44 -0700617 if (unlikely(!page_evictable(page, NULL)))
618 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700619
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700620 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800621 goto keep_locked;
622
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623 /* Double the slab pressure for mapped and swapcache pages */
624 if (page_mapped(page) || PageSwapCache(page))
625 sc->nr_scanned++;
626
Andy Whitcroftc661b072007-08-22 14:01:26 -0700627 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
628 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
629
630 if (PageWriteback(page)) {
631 /*
632 * Synchronous reclaim is performed in two passes,
633 * first an asynchronous pass over the list to
634 * start parallel writeback, and a second synchronous
635 * pass to wait for the IO to complete. Wait here
636 * for any page for which writeback has already
637 * started.
638 */
639 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
640 wait_on_page_writeback(page);
Andrew Morton4dd4b922008-03-24 12:29:52 -0700641 else
Andy Whitcroftc661b072007-08-22 14:01:26 -0700642 goto keep_locked;
643 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700644
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700645 referenced = page_referenced(page, 1,
646 sc->mem_cgroup, &vm_flags);
Minchan Kim03ef83a2009-08-26 14:29:23 -0700647 /*
648 * In active use or really unfreeable? Activate it.
649 * If page which have PG_mlocked lost isoltation race,
650 * try_to_unmap moves it to unevictable list
651 */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700652 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
Minchan Kim03ef83a2009-08-26 14:29:23 -0700653 referenced && page_mapping_inuse(page)
654 && !(vm_flags & VM_LOCKED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 goto activate_locked;
656
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 /*
658 * Anonymous process memory has backing store?
659 * Try to allocate it some swap space here.
660 */
Nick Pigginb291f002008-10-18 20:26:44 -0700661 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800662 if (!(sc->gfp_mask & __GFP_IO))
663 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800664 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700665 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800666 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700667 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700668
669 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700670
671 /*
672 * The page is mapped into the page tables of one or more
673 * processes. Try to unmap it here.
674 */
675 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200676 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700677 case SWAP_FAIL:
678 goto activate_locked;
679 case SWAP_AGAIN:
680 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700681 case SWAP_MLOCK:
682 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700683 case SWAP_SUCCESS:
684 ; /* try to free the page below */
685 }
686 }
687
688 if (PageDirty(page)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700689 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700691 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800693 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 goto keep_locked;
695
696 /* Page is dirty, try to write it out here */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700697 switch (pageout(page, mapping, sync_writeback)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 case PAGE_KEEP:
699 goto keep_locked;
700 case PAGE_ACTIVATE:
701 goto activate_locked;
702 case PAGE_SUCCESS:
Andrew Morton4dd4b922008-03-24 12:29:52 -0700703 if (PageWriteback(page) || PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700704 goto keep;
705 /*
706 * A synchronous write - probably a ramdisk. Go
707 * ahead and try to reclaim the page.
708 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200709 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 goto keep;
711 if (PageDirty(page) || PageWriteback(page))
712 goto keep_locked;
713 mapping = page_mapping(page);
714 case PAGE_CLEAN:
715 ; /* try to free the page below */
716 }
717 }
718
719 /*
720 * If the page has buffers, try to free the buffer mappings
721 * associated with this page. If we succeed we try to free
722 * the page as well.
723 *
724 * We do this even if the page is PageDirty().
725 * try_to_release_page() does not perform I/O, but it is
726 * possible for a page to have PageDirty set, but it is actually
727 * clean (all its buffers are clean). This happens if the
728 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700729 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700730 * try_to_release_page() will discover that cleanness and will
731 * drop the buffers and mark the page clean - it can be freed.
732 *
733 * Rarely, pages can have buffers and no ->mapping. These are
734 * the pages which were not successfully invalidated in
735 * truncate_complete_page(). We try to drop those buffers here
736 * and if that worked, and the page is no longer mapped into
737 * process address space (page_count == 1) it can be freed.
738 * Otherwise, leave the page on the LRU so it is swappable.
739 */
David Howells266cf652009-04-03 16:42:36 +0100740 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700741 if (!try_to_release_page(page, sc->gfp_mask))
742 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700743 if (!mapping && page_count(page) == 1) {
744 unlock_page(page);
745 if (put_page_testzero(page))
746 goto free_it;
747 else {
748 /*
749 * rare race with speculative reference.
750 * the speculative reference will free
751 * this page shortly, so we may
752 * increment nr_reclaimed here (and
753 * leave it off the LRU).
754 */
755 nr_reclaimed++;
756 continue;
757 }
758 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700759 }
760
Nick Piggine2867812008-07-25 19:45:30 -0700761 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800762 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763
Nick Piggina978d6f2008-10-18 20:26:58 -0700764 /*
765 * At this point, we have no other references and there is
766 * no way to pick any more up (removed from LRU, removed
767 * from pagecache). Can use non-atomic bitops now (and
768 * we obviously don't have to worry about waking up a process
769 * waiting on the page lock, because there are no references.
770 */
771 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700772free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800773 nr_reclaimed++;
Nick Piggine2867812008-07-25 19:45:30 -0700774 if (!pagevec_add(&freed_pvec, page)) {
775 __pagevec_free(&freed_pvec);
776 pagevec_reinit(&freed_pvec);
777 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778 continue;
779
Nick Pigginb291f002008-10-18 20:26:44 -0700780cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800781 if (PageSwapCache(page))
782 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700783 unlock_page(page);
784 putback_lru_page(page);
785 continue;
786
Linus Torvalds1da177e2005-04-16 15:20:36 -0700787activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700788 /* Not a candidate for swapping, so reclaim swap space. */
789 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800790 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700791 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 SetPageActive(page);
793 pgactivate++;
794keep_locked:
795 unlock_page(page);
796keep:
797 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700798 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 }
800 list_splice(&ret_pages, page_list);
801 if (pagevec_count(&freed_pvec))
Nick Piggine2867812008-07-25 19:45:30 -0700802 __pagevec_free(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700803 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800804 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805}
806
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700807/* LRU Isolation modes. */
808#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
809#define ISOLATE_ACTIVE 1 /* Isolate active pages. */
810#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
811
812/*
813 * Attempt to remove the specified page from its LRU. Only take this page
814 * if it is of the appropriate PageActive status. Pages which are being
815 * freed elsewhere are also ignored.
816 *
817 * page: page to consider
818 * mode: one of the LRU isolation modes defined above
819 *
820 * returns 0 on success, -ve errno on failure.
821 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700822int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700823{
824 int ret = -EINVAL;
825
826 /* Only take pages on the LRU. */
827 if (!PageLRU(page))
828 return ret;
829
830 /*
831 * When checking the active state, we need to be sure we are
832 * dealing with comparible boolean values. Take the logical not
833 * of each.
834 */
835 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
836 return ret;
837
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700838 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700839 return ret;
840
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700841 /*
842 * When this function is being called for lumpy reclaim, we
843 * initially look into all LRU pages, active, inactive and
844 * unevictable; only give shrink_page_list evictable pages.
845 */
846 if (PageUnevictable(page))
847 return ret;
848
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700849 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800850
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700851 if (likely(get_page_unless_zero(page))) {
852 /*
853 * Be careful not to clear PageLRU until after we're
854 * sure the page is not being freed elsewhere -- the
855 * page release code relies on it.
856 */
857 ClearPageLRU(page);
858 ret = 0;
859 }
860
861 return ret;
862}
863
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800864/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700865 * zone->lru_lock is heavily contended. Some of the functions that
866 * shrink the lists perform better by taking out a batch of pages
867 * and working on them outside the LRU lock.
868 *
869 * For pagecache intensive workloads, this function is the hottest
870 * spot in the kernel (apart from copy_*_user functions).
871 *
872 * Appropriate locks must be held before calling this function.
873 *
874 * @nr_to_scan: The number of pages to look through on the list.
875 * @src: The LRU list to pull pages off.
876 * @dst: The temp list to put pages on to.
877 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700878 * @order: The caller's attempted allocation order
879 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700880 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881 *
882 * returns how many pages were moved onto *@dst.
883 */
Andrew Morton69e05942006-03-22 00:08:19 -0800884static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
885 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700886 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700887{
Andrew Morton69e05942006-03-22 00:08:19 -0800888 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800889 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700890
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800891 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700892 struct page *page;
893 unsigned long pfn;
894 unsigned long end_pfn;
895 unsigned long page_pfn;
896 int zone_id;
897
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898 page = lru_to_page(src);
899 prefetchw_prev_lru_page(page, src, flags);
900
Nick Piggin725d7042006-09-25 23:30:55 -0700901 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800902
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700903 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700904 case 0:
905 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700906 mem_cgroup_del_lru(page);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800907 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700908 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800909
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700910 case -EBUSY:
911 /* else it is being freed elsewhere */
912 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700913 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700914 continue;
915
916 default:
917 BUG();
918 }
919
920 if (!order)
921 continue;
922
923 /*
924 * Attempt to take all pages in the order aligned region
925 * surrounding the tag page. Only take those pages of
926 * the same active state as that tag page. We may safely
927 * round the target page pfn down to the requested order
928 * as the mem_map is guarenteed valid out to MAX_ORDER,
929 * where that page is in a different zone we will detect
930 * it from its zone id and abort this block scan.
931 */
932 zone_id = page_zone_id(page);
933 page_pfn = page_to_pfn(page);
934 pfn = page_pfn & ~((1 << order) - 1);
935 end_pfn = pfn + (1 << order);
936 for (; pfn < end_pfn; pfn++) {
937 struct page *cursor_page;
938
939 /* The target page is in the block, ignore it. */
940 if (unlikely(pfn == page_pfn))
941 continue;
942
943 /* Avoid holes within the zone. */
944 if (unlikely(!pfn_valid_within(pfn)))
945 break;
946
947 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700948
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700949 /* Check that we have not crossed a zone boundary. */
950 if (unlikely(page_zone_id(cursor_page) != zone_id))
951 continue;
Minchan Kimde2e7562009-09-21 17:01:43 -0700952
953 /*
954 * If we don't have enough swap space, reclaiming of
955 * anon page which don't already have a swap slot is
956 * pointless.
957 */
958 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
959 !PageSwapCache(cursor_page))
960 continue;
961
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -0700962 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700963 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +0900964 mem_cgroup_del_lru(cursor_page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700965 nr_taken++;
966 scan++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700967 }
968 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 }
970
971 *scanned = scan;
972 return nr_taken;
973}
974
Balbir Singh66e17072008-02-07 00:13:56 -0800975static unsigned long isolate_pages_global(unsigned long nr,
976 struct list_head *dst,
977 unsigned long *scanned, int order,
978 int mode, struct zone *z,
979 struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700980 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -0800981{
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700982 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -0800983 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700984 lru += LRU_ACTIVE;
985 if (file)
986 lru += LRU_FILE;
987 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -0700988 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -0800989}
990
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700992 * clear_active_flags() is a helper for shrink_active_list(), clearing
993 * any active bits from the pages in the list.
994 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700995static unsigned long clear_active_flags(struct list_head *page_list,
996 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700997{
998 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700999 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001000 struct page *page;
1001
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001002 list_for_each_entry(page, page_list, lru) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07001003 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001004 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001005 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001006 ClearPageActive(page);
1007 nr_active++;
1008 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001009 count[lru]++;
1010 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001011
1012 return nr_active;
1013}
1014
Nick Piggin62695a82008-10-18 20:26:09 -07001015/**
1016 * isolate_lru_page - tries to isolate a page from its LRU list
1017 * @page: page to isolate from its LRU list
1018 *
1019 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1020 * vmstat statistic corresponding to whatever LRU list the page was on.
1021 *
1022 * Returns 0 if the page was removed from an LRU list.
1023 * Returns -EBUSY if the page was not on an LRU list.
1024 *
1025 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001026 * the active list, it will have PageActive set. If it was found on
1027 * the unevictable list, it will have the PageUnevictable bit set. That flag
1028 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001029 *
1030 * The vmstat statistic corresponding to the list on which the page was
1031 * found will be decremented.
1032 *
1033 * Restrictions:
1034 * (1) Must be called with an elevated refcount on the page. This is a
1035 * fundamentnal difference from isolate_lru_pages (which is called
1036 * without a stable reference).
1037 * (2) the lru_lock must not be held.
1038 * (3) interrupts must be enabled.
1039 */
1040int isolate_lru_page(struct page *page)
1041{
1042 int ret = -EBUSY;
1043
1044 if (PageLRU(page)) {
1045 struct zone *zone = page_zone(page);
1046
1047 spin_lock_irq(&zone->lru_lock);
1048 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001049 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001050 ret = 0;
1051 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001052
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001053 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001054 }
1055 spin_unlock_irq(&zone->lru_lock);
1056 }
1057 return ret;
1058}
1059
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001060/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001061 * Are there way too many processes in the direct reclaim path already?
1062 */
1063static int too_many_isolated(struct zone *zone, int file,
1064 struct scan_control *sc)
1065{
1066 unsigned long inactive, isolated;
1067
1068 if (current_is_kswapd())
1069 return 0;
1070
1071 if (!scanning_global_lru(sc))
1072 return 0;
1073
1074 if (file) {
1075 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1076 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1077 } else {
1078 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1079 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1080 }
1081
1082 return isolated > inactive;
1083}
1084
1085/*
Andrew Morton1742f192006-03-22 00:08:21 -08001086 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1087 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001088 */
Andrew Morton1742f192006-03-22 00:08:21 -08001089static unsigned long shrink_inactive_list(unsigned long max_scan,
Rik van Riel33c120e2008-10-18 20:26:36 -07001090 struct zone *zone, struct scan_control *sc,
1091 int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001092{
1093 LIST_HEAD(page_list);
1094 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -08001095 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001096 unsigned long nr_reclaimed = 0;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001097 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001098 int lumpy_reclaim = 0;
1099
Rik van Riel35cd7812009-09-21 17:01:38 -07001100 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001101 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001102
1103 /* We are about to die and free our memory. Return now. */
1104 if (fatal_signal_pending(current))
1105 return SWAP_CLUSTER_MAX;
1106 }
1107
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001108 /*
1109 * If we need a large contiguous chunk of memory, or have
1110 * trouble getting a small set of contiguous pages, we
1111 * will reclaim both active and inactive pages.
1112 *
1113 * We use the same threshold as pageout congestion_wait below.
1114 */
1115 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1116 lumpy_reclaim = 1;
1117 else if (sc->order && priority < DEF_PRIORITY - 2)
1118 lumpy_reclaim = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119
1120 pagevec_init(&pvec, 1);
1121
1122 lru_add_drain();
1123 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -08001124 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -08001126 unsigned long nr_taken;
1127 unsigned long nr_scan;
1128 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001129 unsigned long nr_active;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001130 unsigned int count[NR_LRU_LISTS] = { 0, };
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001131 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001132 unsigned long nr_anon;
1133 unsigned long nr_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134
Balbir Singh66e17072008-02-07 00:13:56 -08001135 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001136 &page_list, &nr_scan, sc->order, mode,
1137 zone, sc->mem_cgroup, 0, file);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001138
1139 if (scanning_global_lru(sc)) {
1140 zone->pages_scanned += nr_scan;
1141 if (current_is_kswapd())
1142 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1143 nr_scan);
1144 else
1145 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1146 nr_scan);
1147 }
1148
1149 if (nr_taken == 0)
1150 goto done;
1151
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001152 nr_active = clear_active_flags(&page_list, count);
Andy Whitcrofte9187bd2007-08-22 14:01:25 -07001153 __count_vm_events(PGDEACTIVATE, nr_active);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001154
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001155 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1156 -count[LRU_ACTIVE_FILE]);
1157 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1158 -count[LRU_INACTIVE_FILE]);
1159 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1160 -count[LRU_ACTIVE_ANON]);
1161 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1162 -count[LRU_INACTIVE_ANON]);
1163
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001164 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1165 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1166 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1167 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001168
1169 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1170 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1171 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1172 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1173
Linus Torvalds1da177e2005-04-16 15:20:36 -07001174 spin_unlock_irq(&zone->lru_lock);
1175
Andrew Morton69e05942006-03-22 00:08:19 -08001176 nr_scanned += nr_scan;
Andy Whitcroftc661b072007-08-22 14:01:26 -07001177 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1178
1179 /*
1180 * If we are direct reclaiming for contiguous pages and we do
1181 * not reclaim everything in the list, try again and wait
1182 * for IO to complete. This will stall high-order allocations
1183 * but that should be acceptable to the caller
1184 */
1185 if (nr_freed < nr_taken && !current_is_kswapd() &&
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001186 lumpy_reclaim) {
Jens Axboe8aa7e842009-07-09 14:52:32 +02001187 congestion_wait(BLK_RW_ASYNC, HZ/10);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001188
1189 /*
1190 * The attempt at page out may have made some
1191 * of the pages active, mark them inactive again.
1192 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001193 nr_active = clear_active_flags(&page_list, count);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001194 count_vm_events(PGDEACTIVATE, nr_active);
1195
1196 nr_freed += shrink_page_list(&page_list, sc,
1197 PAGEOUT_IO_SYNC);
1198 }
1199
Andrew Morton05ff5132006-03-22 00:08:20 -08001200 nr_reclaimed += nr_freed;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001201
Nick Piggina74609f2006-01-06 00:11:20 -08001202 local_irq_disable();
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001203 if (current_is_kswapd())
Christoph Lameterf8891e52006-06-30 01:55:45 -07001204 __count_vm_events(KSWAPD_STEAL, nr_freed);
Shantanu Goel918d3f92006-12-29 16:48:59 -08001205 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -08001206
1207 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 /*
1209 * Put back any unfreeable pages.
1210 */
1211 while (!list_empty(&page_list)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001212 int lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001213 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -07001214 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 list_del(&page->lru);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001216 if (unlikely(!page_evictable(page, NULL))) {
1217 spin_unlock_irq(&zone->lru_lock);
1218 putback_lru_page(page);
1219 spin_lock_irq(&zone->lru_lock);
1220 continue;
1221 }
1222 SetPageLRU(page);
1223 lru = page_lru(page);
1224 add_page_to_lru_list(zone, page, lru);
KOSAKI Motohiro74a1c482009-09-21 17:01:45 -07001225 if (is_active_lru(lru)) {
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001226 int file = is_file_lru(lru);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001227 reclaim_stat->recent_rotated[file]++;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001228 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 if (!pagevec_add(&pvec, page)) {
1230 spin_unlock_irq(&zone->lru_lock);
1231 __pagevec_release(&pvec);
1232 spin_lock_irq(&zone->lru_lock);
1233 }
1234 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001235 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1236 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1237
Andrew Morton69e05942006-03-22 00:08:19 -08001238 } while (nr_scanned < max_scan);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001239
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240done:
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001241 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -08001243 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001244}
1245
Martin Bligh3bb1a852006-10-28 10:38:24 -07001246/*
1247 * We are about to scan this zone at a certain priority level. If that priority
1248 * level is smaller (ie: more urgent) than the previous priority, then note
1249 * that priority level within the zone. This is done so that when the next
1250 * process comes in to scan this zone, it will immediately start out at this
1251 * priority level rather than having to build up its own scanning priority.
1252 * Here, this priority affects only the reclaim-mapped threshold.
1253 */
1254static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1255{
1256 if (priority < zone->prev_priority)
1257 zone->prev_priority = priority;
1258}
1259
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001260/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 * This moves pages from the active list to the inactive list.
1262 *
1263 * We move them the other way if the page is referenced by one or more
1264 * processes, from rmap.
1265 *
1266 * If the pages are mostly unmapped, the processing is fast and it is
1267 * appropriate to hold zone->lru_lock across the whole operation. But if
1268 * the pages are mapped, the processing is slow (page_referenced()) so we
1269 * should drop zone->lru_lock around each page. It's impossible to balance
1270 * this, so instead we remove the pages from the LRU while processing them.
1271 * It is safe to rely on PG_active against the non-LRU pages in here because
1272 * nobody will play with that bit on a non-LRU page.
1273 *
1274 * The downside is that we have to touch page->_count against each page.
1275 * But we had to alter page->flags anyway.
1276 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001277
Wu Fengguang3eb41402009-06-16 15:33:13 -07001278static void move_active_pages_to_lru(struct zone *zone,
1279 struct list_head *list,
1280 enum lru_list lru)
1281{
1282 unsigned long pgmoved = 0;
1283 struct pagevec pvec;
1284 struct page *page;
1285
1286 pagevec_init(&pvec, 1);
1287
1288 while (!list_empty(list)) {
1289 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001290
1291 VM_BUG_ON(PageLRU(page));
1292 SetPageLRU(page);
1293
Wu Fengguang3eb41402009-06-16 15:33:13 -07001294 list_move(&page->lru, &zone->lru[lru].list);
1295 mem_cgroup_add_lru_list(page, lru);
1296 pgmoved++;
1297
1298 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1299 spin_unlock_irq(&zone->lru_lock);
1300 if (buffer_heads_over_limit)
1301 pagevec_strip(&pvec);
1302 __pagevec_release(&pvec);
1303 spin_lock_irq(&zone->lru_lock);
1304 }
1305 }
1306 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1307 if (!is_active_lru(lru))
1308 __count_vm_events(PGDEACTIVATE, pgmoved);
1309}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001310
Andrew Morton1742f192006-03-22 00:08:21 -08001311static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001312 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001314 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001315 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001316 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001318 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001319 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001321 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001322 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323
1324 lru_add_drain();
1325 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001326 nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
Balbir Singh66e17072008-02-07 00:13:56 -08001327 ISOLATE_ACTIVE, zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001328 sc->mem_cgroup, 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001329 /*
1330 * zone->pages_scanned is used for detect zone's oom
1331 * mem_cgroup remembers nr_scan by itself.
1332 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001333 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001334 zone->pages_scanned += pgscanned;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001335 }
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001336 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001337
Wu Fengguang3eb41402009-06-16 15:33:13 -07001338 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001339 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001340 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001341 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001342 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001343 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 spin_unlock_irq(&zone->lru_lock);
1345
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346 while (!list_empty(&l_hold)) {
1347 cond_resched();
1348 page = lru_to_page(&l_hold);
1349 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001350
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001351 if (unlikely(!page_evictable(page, NULL))) {
1352 putback_lru_page(page);
1353 continue;
1354 }
1355
Rik van Riel7e9cd482008-10-18 20:26:35 -07001356 /* page_referenced clears PageReferenced */
1357 if (page_mapping_inuse(page) &&
Wu Fengguang8cab4752009-06-16 15:33:12 -07001358 page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001359 nr_rotated++;
Wu Fengguang8cab4752009-06-16 15:33:12 -07001360 /*
1361 * Identify referenced, file-backed active pages and
1362 * give them one more trip around the active list. So
1363 * that executable code get better chances to stay in
1364 * memory under moderate memory pressure. Anon pages
1365 * are not likely to be evicted by use-once streaming
1366 * IO, plus JVM can create lots of anon VM_EXEC pages,
1367 * so we ignore them here.
1368 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001369 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001370 list_add(&page->lru, &l_active);
1371 continue;
1372 }
1373 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001374
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001375 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 list_add(&page->lru, &l_inactive);
1377 }
1378
Andrew Mortonb5557492009-01-06 14:40:13 -08001379 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001380 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001381 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001382 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001383 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001384 * Count referenced pages from currently used mappings as rotated,
1385 * even though only some of them are actually re-activated. This
1386 * helps balance scan pressure between file and anonymous pages in
1387 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001388 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001389 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001390
Wu Fengguang3eb41402009-06-16 15:33:13 -07001391 move_active_pages_to_lru(zone, &l_active,
1392 LRU_ACTIVE + file * LRU_FILE);
1393 move_active_pages_to_lru(zone, &l_inactive,
1394 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001395 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001396 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397}
1398
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001399static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001400{
1401 unsigned long active, inactive;
1402
1403 active = zone_page_state(zone, NR_ACTIVE_ANON);
1404 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1405
1406 if (inactive * zone->inactive_ratio < active)
1407 return 1;
1408
1409 return 0;
1410}
1411
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001412/**
1413 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1414 * @zone: zone to check
1415 * @sc: scan control of this context
1416 *
1417 * Returns true if the zone does not have enough inactive anon pages,
1418 * meaning some active anon pages need to be deactivated.
1419 */
1420static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1421{
1422 int low;
1423
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001424 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001425 low = inactive_anon_is_low_global(zone);
1426 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001427 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001428 return low;
1429}
1430
Rik van Riel56e49d22009-06-16 15:32:28 -07001431static int inactive_file_is_low_global(struct zone *zone)
1432{
1433 unsigned long active, inactive;
1434
1435 active = zone_page_state(zone, NR_ACTIVE_FILE);
1436 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1437
1438 return (active > inactive);
1439}
1440
1441/**
1442 * inactive_file_is_low - check if file pages need to be deactivated
1443 * @zone: zone to check
1444 * @sc: scan control of this context
1445 *
1446 * When the system is doing streaming IO, memory pressure here
1447 * ensures that active file pages get deactivated, until more
1448 * than half of the file pages are on the inactive list.
1449 *
1450 * Once we get to that situation, protect the system's working
1451 * set from being evicted by disabling active file page aging.
1452 *
1453 * This uses a different ratio than the anonymous pages, because
1454 * the page cache uses a use-once replacement algorithm.
1455 */
1456static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1457{
1458 int low;
1459
1460 if (scanning_global_lru(sc))
1461 low = inactive_file_is_low_global(zone);
1462 else
1463 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1464 return low;
1465}
1466
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001467static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001468 struct zone *zone, struct scan_control *sc, int priority)
1469{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001470 int file = is_file_lru(lru);
1471
Rik van Riel56e49d22009-06-16 15:32:28 -07001472 if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
Rik van Riel556adec2008-10-18 20:26:34 -07001473 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1474 return 0;
1475 }
1476
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001477 if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001478 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001479 return 0;
1480 }
Rik van Riel33c120e2008-10-18 20:26:36 -07001481 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001482}
1483
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001485 * Determine how aggressively the anon and file LRU lists should be
1486 * scanned. The relative value of each set of LRU lists is determined
1487 * by looking at the fraction of the pages scanned we did rotate back
1488 * onto the active list instead of evict.
1489 *
1490 * percent[0] specifies how much pressure to put on ram/swap backed
1491 * memory, while percent[1] determines pressure on the file LRUs.
1492 */
1493static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1494 unsigned long *percent)
1495{
1496 unsigned long anon, file, free;
1497 unsigned long anon_prio, file_prio;
1498 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001499 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001500
Vincent Li0b217672009-09-21 17:03:09 -07001501 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1502 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1503 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1504 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001505
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001506 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001507 free = zone_page_state(zone, NR_FREE_PAGES);
1508 /* If we have very few page cache pages,
1509 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001510 if (unlikely(file + free <= high_wmark_pages(zone))) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001511 percent[0] = 100;
1512 percent[1] = 0;
1513 return;
1514 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001515 }
1516
1517 /*
1518 * OK, so we have swap space and a fair amount of page cache
1519 * pages. We use the recently rotated / recently scanned
1520 * ratios to determine how valuable each cache is.
1521 *
1522 * Because workloads change over time (and to avoid overflow)
1523 * we keep these statistics as a floating average, which ends
1524 * up weighing recent references more than old ones.
1525 *
1526 * anon in [0], file in [1]
1527 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001528 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001529 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001530 reclaim_stat->recent_scanned[0] /= 2;
1531 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001532 spin_unlock_irq(&zone->lru_lock);
1533 }
1534
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001535 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001536 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001537 reclaim_stat->recent_scanned[1] /= 2;
1538 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001539 spin_unlock_irq(&zone->lru_lock);
1540 }
1541
1542 /*
1543 * With swappiness at 100, anonymous and file have the same priority.
1544 * This scanning priority is essentially the inverse of IO cost.
1545 */
1546 anon_prio = sc->swappiness;
1547 file_prio = 200 - sc->swappiness;
1548
1549 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001550 * The amount of pressure on anon vs file pages is inversely
1551 * proportional to the fraction of recently scanned pages on
1552 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001553 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001554 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1555 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001556
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001557 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1558 fp /= reclaim_stat->recent_rotated[1] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001559
1560 /* Normalize to percentages */
1561 percent[0] = 100 * ap / (ap + fp + 1);
1562 percent[1] = 100 - percent[0];
1563}
1564
Wu Fengguang6e08a362009-06-16 15:32:29 -07001565/*
1566 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1567 * until we collected @swap_cluster_max pages to scan.
1568 */
1569static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1570 unsigned long *nr_saved_scan,
1571 unsigned long swap_cluster_max)
1572{
1573 unsigned long nr;
1574
1575 *nr_saved_scan += nr_to_scan;
1576 nr = *nr_saved_scan;
1577
1578 if (nr >= swap_cluster_max)
1579 *nr_saved_scan = 0;
1580 else
1581 nr = 0;
1582
1583 return nr;
1584}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001585
1586/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001587 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1588 */
Rik van Riela79311c2009-01-06 14:40:01 -08001589static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001590 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001592 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001593 unsigned long nr_to_scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001594 unsigned long percent[2]; /* anon @ 0; file @ 1 */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001595 enum lru_list l;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001596 unsigned long nr_reclaimed = sc->nr_reclaimed;
1597 unsigned long swap_cluster_max = sc->swap_cluster_max;
Wu Fengguangf8629632009-09-21 17:03:11 -07001598 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001599 int noswap = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001601 /* If we have no swap space, do not bother scanning anon pages. */
1602 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1603 noswap = 1;
1604 percent[0] = 0;
1605 percent[1] = 100;
1606 } else
1607 get_scan_ratio(zone, sc, percent);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001608
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001609 for_each_evictable_lru(l) {
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001610 int file = is_file_lru(l);
Andrew Morton8713e012009-04-30 15:08:55 -07001611 unsigned long scan;
Johannes Weinere0f79b82008-10-18 20:26:55 -07001612
Vincent Li0b217672009-09-21 17:03:09 -07001613 scan = zone_nr_lru_pages(zone, sc, l);
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001614 if (priority || noswap) {
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001615 scan >>= priority;
1616 scan = (scan * percent[file]) / 100;
1617 }
Wu Fengguangf8629632009-09-21 17:03:11 -07001618 nr[l] = nr_scan_try_batch(scan,
1619 &reclaim_stat->nr_saved_scan[l],
1620 swap_cluster_max);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001621 }
1622
Rik van Riel556adec2008-10-18 20:26:34 -07001623 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1624 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001625 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001626 if (nr[l]) {
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001627 nr_to_scan = min(nr[l], swap_cluster_max);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001628 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001630 nr_reclaimed += shrink_list(l, nr_to_scan,
1631 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001632 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633 }
Rik van Riela79311c2009-01-06 14:40:01 -08001634 /*
1635 * On large memory systems, scan >> priority can become
1636 * really large. This is fine for the starting priority;
1637 * we want to put equal scanning pressure on each zone.
1638 * However, if the VM has a harder time of freeing pages,
1639 * with multiple processes reclaiming pages, the total
1640 * freeing target can get unreasonably large.
1641 */
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001642 if (nr_reclaimed > swap_cluster_max &&
Rik van Riela79311c2009-01-06 14:40:01 -08001643 priority < DEF_PRIORITY && !current_is_kswapd())
1644 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001645 }
1646
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001647 sc->nr_reclaimed = nr_reclaimed;
1648
Rik van Riel556adec2008-10-18 20:26:34 -07001649 /*
1650 * Even if we did not try to evict anon pages at all, we want to
1651 * rebalance the anon lru active/inactive ratio.
1652 */
MinChan Kim69c85482009-06-16 15:32:44 -07001653 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
Rik van Riel556adec2008-10-18 20:26:34 -07001654 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1655
Andrew Morton232ea4d2007-02-28 20:13:21 -08001656 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001657}
1658
1659/*
1660 * This is the direct reclaim path, for page-allocating processes. We only
1661 * try to reclaim pages from zones which will satisfy the caller's allocation
1662 * request.
1663 *
Mel Gorman41858962009-06-16 15:32:12 -07001664 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1665 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1667 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001668 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1669 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1670 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 * If a zone is deemed to be full of pinned pages then just give it a light
1673 * scan then give up on it.
1674 */
Rik van Riela79311c2009-01-06 14:40:01 -08001675static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001676 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677{
Mel Gorman54a6eb52008-04-28 02:12:16 -07001678 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Mel Gormandd1a2392008-04-28 02:12:17 -07001679 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001680 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001681
Nick Piggin408d8542006-09-25 23:31:27 -07001682 sc->all_unreclaimable = 1;
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001683 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1684 sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001685 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001687 /*
1688 * Take care memory controller reclaiming has small influence
1689 * to global LRU.
1690 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001691 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001692 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1693 continue;
1694 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001695
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001696 if (zone_is_all_unreclaimable(zone) &&
1697 priority != DEF_PRIORITY)
1698 continue; /* Let kswapd poll it */
1699 sc->all_unreclaimable = 0;
1700 } else {
1701 /*
1702 * Ignore cpuset limitation here. We just want to reduce
1703 * # of used pages by us regardless of memory shortage.
1704 */
1705 sc->all_unreclaimable = 0;
1706 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1707 priority);
1708 }
Nick Piggin408d8542006-09-25 23:31:27 -07001709
Rik van Riela79311c2009-01-06 14:40:01 -08001710 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 }
1712}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001713
Linus Torvalds1da177e2005-04-16 15:20:36 -07001714/*
1715 * This is the main entry point to direct page reclaim.
1716 *
1717 * If a full scan of the inactive list fails to free enough memory then we
1718 * are "out of memory" and something needs to be killed.
1719 *
1720 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1721 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02001722 * caller can't do much about. We kick the writeback threads and take explicit
1723 * naps in the hope that some of these pages can be written. But if the
1724 * allocating task holds filesystem locks which prevent writeout this might not
1725 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001726 *
1727 * returns: 0, if no pages reclaimed
1728 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729 */
Mel Gormandac1d272008-04-28 02:12:12 -07001730static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001731 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732{
1733 int priority;
kosaki.motohiro@jp.fujitsu.comc700be32008-06-12 15:21:27 -07001734 unsigned long ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001735 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001736 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 unsigned long lru_pages = 0;
Mel Gormandd1a2392008-04-28 02:12:17 -07001738 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001739 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -07001740 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741
Keika Kobayashi873b4772008-07-25 01:48:52 -07001742 delayacct_freepages_start();
1743
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001744 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001745 count_vm_event(ALLOCSTALL);
1746 /*
1747 * mem_cgroup will not do shrink_slab.
1748 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001749 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001750 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001752 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1753 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754
Wu Fengguangadea02a2009-09-21 17:01:42 -07001755 lru_pages += zone_reclaimable_pages(zone);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001756 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001757 }
1758
1759 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001760 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001761 if (!priority)
1762 disable_swap_token();
Rik van Riela79311c2009-01-06 14:40:01 -08001763 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001764 /*
1765 * Don't shrink slabs when reclaiming memory from
1766 * over limit cgroups
1767 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001768 if (scanning_global_lru(sc)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07001769 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001770 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08001771 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001772 reclaim_state->reclaimed_slab = 0;
1773 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001774 }
Balbir Singh66e17072008-02-07 00:13:56 -08001775 total_scanned += sc->nr_scanned;
Rik van Riela79311c2009-01-06 14:40:01 -08001776 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1777 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001778 goto out;
1779 }
1780
1781 /*
1782 * Try to write back as many pages as we just scanned. This
1783 * tends to cause slow streaming writers to write data to the
1784 * disk smoothly, at the dirtying rate, which is nice. But
1785 * that's undesirable in laptop mode, where we *want* lumpy
1786 * writeout. So in laptop mode, write out the whole world.
1787 */
Balbir Singh66e17072008-02-07 00:13:56 -08001788 if (total_scanned > sc->swap_cluster_max +
1789 sc->swap_cluster_max / 2) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001790 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001791 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001792 }
1793
1794 /* Take a nap, wait for some writeback to complete */
Andrew Morton4dd4b922008-03-24 12:29:52 -07001795 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02001796 congestion_wait(BLK_RW_ASYNC, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 }
Fernando Luis Vazquez Cao87547ee2008-07-29 22:33:42 -07001798 /* top priority shrink_zones still had more to do? don't OOM, then */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001799 if (!sc->all_unreclaimable && scanning_global_lru(sc))
Rik van Riela79311c2009-01-06 14:40:01 -08001800 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07001802 /*
1803 * Now that we've scanned all the zones at this priority level, note
1804 * that level within the zone so that the next thread which performs
1805 * scanning of this zone will immediately start out at this priority
1806 * level. This affects only the decision whether or not to bring
1807 * mapped pages onto the inactive list.
1808 */
1809 if (priority < 0)
1810 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001812 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001813 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001815 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1816 continue;
1817
1818 zone->prev_priority = priority;
1819 }
1820 } else
1821 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1822
Keika Kobayashi873b4772008-07-25 01:48:52 -07001823 delayacct_freepages_end();
1824
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 return ret;
1826}
1827
Mel Gormandac1d272008-04-28 02:12:12 -07001828unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001829 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08001830{
1831 struct scan_control sc = {
1832 .gfp_mask = gfp_mask,
1833 .may_writepage = !laptop_mode,
1834 .swap_cluster_max = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001835 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001836 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08001837 .swappiness = vm_swappiness,
1838 .order = order,
1839 .mem_cgroup = NULL,
1840 .isolate_pages = isolate_pages_global,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001841 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08001842 };
1843
Mel Gormandd1a2392008-04-28 02:12:17 -07001844 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001845}
1846
Balbir Singh00f0b822008-03-04 14:28:39 -08001847#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08001848
Balbir Singh4e416952009-09-23 15:56:39 -07001849unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1850 gfp_t gfp_mask, bool noswap,
1851 unsigned int swappiness,
1852 struct zone *zone, int nid)
1853{
1854 struct scan_control sc = {
1855 .may_writepage = !laptop_mode,
1856 .may_unmap = 1,
1857 .may_swap = !noswap,
1858 .swap_cluster_max = SWAP_CLUSTER_MAX,
1859 .swappiness = swappiness,
1860 .order = 0,
1861 .mem_cgroup = mem,
1862 .isolate_pages = mem_cgroup_isolate_pages,
1863 };
1864 nodemask_t nm = nodemask_of_node(nid);
1865
1866 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1867 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1868 sc.nodemask = &nm;
1869 sc.nr_reclaimed = 0;
1870 sc.nr_scanned = 0;
1871 /*
1872 * NOTE: Although we can get the priority field, using it
1873 * here is not a good idea, since it limits the pages we can scan.
1874 * if we don't reclaim here, the shrink_zone from balance_pgdat
1875 * will pick up pages from other mem cgroup's as well. We hack
1876 * the priority and make it zero.
1877 */
1878 shrink_zone(0, zone, &sc);
1879 return sc.nr_reclaimed;
1880}
1881
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001882unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001883 gfp_t gfp_mask,
1884 bool noswap,
1885 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08001886{
Balbir Singh4e416952009-09-23 15:56:39 -07001887 struct zonelist *zonelist;
Balbir Singh66e17072008-02-07 00:13:56 -08001888 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08001889 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001890 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001891 .may_swap = !noswap,
Balbir Singh66e17072008-02-07 00:13:56 -08001892 .swap_cluster_max = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001893 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08001894 .order = 0,
1895 .mem_cgroup = mem_cont,
1896 .isolate_pages = mem_cgroup_isolate_pages,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001897 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08001898 };
Balbir Singh66e17072008-02-07 00:13:56 -08001899
Mel Gormandd1a2392008-04-28 02:12:17 -07001900 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1901 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1902 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1903 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001904}
1905#endif
1906
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907/*
1908 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07001909 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001910 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911 * Returns the number of pages which were actually freed.
1912 *
1913 * There is special handling here for zones which are full of pinned pages.
1914 * This can happen if the pages are all mlocked, or if they are all used by
1915 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1916 * What we do is to detect the case where all pages in the zone have been
1917 * scanned twice and there has been zero successful reclaim. Mark the zone as
1918 * dead and from now on, only perform a short scan. Basically we're polling
1919 * the zone for when the problem goes away.
1920 *
1921 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07001922 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1923 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1924 * lower zones regardless of the number of free pages in the lower zones. This
1925 * interoperates with the page allocator fallback scheme to ensure that aging
1926 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001927 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001928static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001929{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930 int all_zones_ok;
1931 int priority;
1932 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08001933 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001934 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08001935 struct scan_control sc = {
1936 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001937 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001938 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001939 .swap_cluster_max = SWAP_CLUSTER_MAX,
1940 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001941 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08001942 .mem_cgroup = NULL,
1943 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08001944 };
Martin Bligh3bb1a852006-10-28 10:38:24 -07001945 /*
1946 * temp_priority is used to remember the scanning priority at which
Mel Gorman41858962009-06-16 15:32:12 -07001947 * this zone was successfully refilled to
1948 * free_pages == high_wmark_pages(zone).
Martin Bligh3bb1a852006-10-28 10:38:24 -07001949 */
1950 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951
1952loop_again:
1953 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08001954 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07001955 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001956 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001957
Martin Bligh3bb1a852006-10-28 10:38:24 -07001958 for (i = 0; i < pgdat->nr_zones; i++)
1959 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960
1961 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1962 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1963 unsigned long lru_pages = 0;
1964
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001965 /* The swap token gets in the way of swapout... */
1966 if (!priority)
1967 disable_swap_token();
1968
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 all_zones_ok = 1;
1970
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001971 /*
1972 * Scan in the highmem->dma direction for the highest
1973 * zone which needs scanning
1974 */
1975 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1976 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001978 if (!populated_zone(zone))
1979 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001980
David Rientjese815af92007-10-16 23:25:54 -07001981 if (zone_is_all_unreclaimable(zone) &&
1982 priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001983 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984
Rik van Riel556adec2008-10-18 20:26:34 -07001985 /*
1986 * Do some background aging of the anon list, to give
1987 * pages a chance to be referenced before reclaiming.
1988 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001989 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07001990 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1991 &sc, priority, 0);
1992
Mel Gorman41858962009-06-16 15:32:12 -07001993 if (!zone_watermark_ok(zone, order,
1994 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001995 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001996 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001999 if (i < 0)
2000 goto out;
2001
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 for (i = 0; i <= end_zone; i++) {
2003 struct zone *zone = pgdat->node_zones + i;
2004
Wu Fengguangadea02a2009-09-21 17:01:42 -07002005 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 }
2007
2008 /*
2009 * Now scan the zone in the dma->highmem direction, stopping
2010 * at the last zone which needs scanning.
2011 *
2012 * We do this because the page allocator works in the opposite
2013 * direction. This prevents the page allocator from allocating
2014 * pages behind kswapd's direction of progress, which would
2015 * cause too much scanning of the lower zones.
2016 */
2017 for (i = 0; i <= end_zone; i++) {
2018 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002019 int nr_slab;
Balbir Singh4e416952009-09-23 15:56:39 -07002020 int nid, zid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021
Con Kolivasf3fe6512006-01-06 00:11:15 -08002022 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 continue;
2024
David Rientjese815af92007-10-16 23:25:54 -07002025 if (zone_is_all_unreclaimable(zone) &&
2026 priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027 continue;
2028
Mel Gorman41858962009-06-16 15:32:12 -07002029 if (!zone_watermark_ok(zone, order,
2030 high_wmark_pages(zone), end_zone, 0))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002031 all_zones_ok = 0;
Martin Bligh3bb1a852006-10-28 10:38:24 -07002032 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002033 sc.nr_scanned = 0;
Martin Bligh3bb1a852006-10-28 10:38:24 -07002034 note_zone_scanning_priority(zone, priority);
Balbir Singh4e416952009-09-23 15:56:39 -07002035
2036 nid = pgdat->node_id;
2037 zid = zone_idx(zone);
2038 /*
2039 * Call soft limit reclaim before calling shrink_zone.
2040 * For now we ignore the return value
2041 */
2042 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2043 nid, zid);
Rik van Riel32a43302007-10-16 01:24:50 -07002044 /*
2045 * We put equal pressure on every zone, unless one
2046 * zone has way too many pages free already.
2047 */
Mel Gorman41858962009-06-16 15:32:12 -07002048 if (!zone_watermark_ok(zone, order,
2049 8*high_wmark_pages(zone), end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002050 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002052 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2053 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002054 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055 total_scanned += sc.nr_scanned;
David Rientjese815af92007-10-16 23:25:54 -07002056 if (zone_is_all_unreclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002058 if (nr_slab == 0 && zone->pages_scanned >=
Wu Fengguangadea02a2009-09-21 17:01:42 -07002059 (zone_reclaimable_pages(zone) * 6))
David Rientjese815af92007-10-16 23:25:54 -07002060 zone_set_flag(zone,
2061 ZONE_ALL_UNRECLAIMABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062 /*
2063 * If we've done a decent amount of scanning and
2064 * the reclaim ratio is low, start doing writepage
2065 * even in laptop mode
2066 */
2067 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002068 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 sc.may_writepage = 1;
2070 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002071 if (all_zones_ok)
2072 break; /* kswapd: all done */
2073 /*
2074 * OK, kswapd is getting into trouble. Take a nap, then take
2075 * another pass across the zones.
2076 */
Andrew Morton4dd4b922008-03-24 12:29:52 -07002077 if (total_scanned && priority < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02002078 congestion_wait(BLK_RW_ASYNC, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002079
2080 /*
2081 * We do this so kswapd doesn't build up large priorities for
2082 * example when it is freeing in parallel with allocators. It
2083 * matches the direct reclaim path behaviour in terms of impact
2084 * on zone->*_priority.
2085 */
Rik van Riela79311c2009-01-06 14:40:01 -08002086 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002087 break;
2088 }
2089out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07002090 /*
2091 * Note within each zone the priority level at which this zone was
2092 * brought into a happy state. So that the next thread which scans this
2093 * zone will start out at that priority level.
2094 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002095 for (i = 0; i < pgdat->nr_zones; i++) {
2096 struct zone *zone = pgdat->node_zones + i;
2097
Martin Bligh3bb1a852006-10-28 10:38:24 -07002098 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099 }
2100 if (!all_zones_ok) {
2101 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002102
2103 try_to_freeze();
2104
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002105 /*
2106 * Fragmentation may mean that the system cannot be
2107 * rebalanced for high-order allocations in all zones.
2108 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2109 * it means the zones have been fully scanned and are still
2110 * not balanced. For high-order allocations, there is
2111 * little point trying all over again as kswapd may
2112 * infinite loop.
2113 *
2114 * Instead, recheck all watermarks at order-0 as they
2115 * are the most important. If watermarks are ok, kswapd will go
2116 * back to sleep. High-order users can still perform direct
2117 * reclaim if they wish.
2118 */
2119 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2120 order = sc.order = 0;
2121
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 goto loop_again;
2123 }
2124
Rik van Riela79311c2009-01-06 14:40:01 -08002125 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126}
2127
2128/*
2129 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002130 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 *
2132 * This basically trickles out pages so that we have _some_
2133 * free memory available even if there is no other activity
2134 * that frees anything up. This is needed for things like routing
2135 * etc, where we otherwise might have all activity going on in
2136 * asynchronous contexts that cannot page things out.
2137 *
2138 * If there are applications that are active memory-allocators
2139 * (most normal use), this basically shouldn't matter.
2140 */
2141static int kswapd(void *p)
2142{
2143 unsigned long order;
2144 pg_data_t *pgdat = (pg_data_t*)p;
2145 struct task_struct *tsk = current;
2146 DEFINE_WAIT(wait);
2147 struct reclaim_state reclaim_state = {
2148 .reclaimed_slab = 0,
2149 };
Rusty Russella70f7302009-03-13 14:49:46 +10302150 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151
Nick Piggincf40bd12009-01-21 08:12:39 +01002152 lockdep_set_current_reclaim_state(GFP_KERNEL);
2153
Rusty Russell174596a2009-01-01 10:12:29 +10302154 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002155 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002156 current->reclaim_state = &reclaim_state;
2157
2158 /*
2159 * Tell the memory management that we're a "memory allocator",
2160 * and that if we need more memory we should get access to it
2161 * regardless (see "__alloc_pages()"). "kswapd" should
2162 * never get caught in the normal page freeing logic.
2163 *
2164 * (Kswapd normally doesn't need memory anyway, but sometimes
2165 * you need a small amount of memory in order to be able to
2166 * page out something else, and this flag essentially protects
2167 * us from recursively trying to free more memory as we're
2168 * trying to free the first piece of memory in the first place).
2169 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002170 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002171 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002172
2173 order = 0;
2174 for ( ; ; ) {
2175 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002176
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2178 new_order = pgdat->kswapd_max_order;
2179 pgdat->kswapd_max_order = 0;
2180 if (order < new_order) {
2181 /*
2182 * Don't sleep if someone wants a larger 'order'
2183 * allocation
2184 */
2185 order = new_order;
2186 } else {
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002187 if (!freezing(current))
2188 schedule();
2189
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190 order = pgdat->kswapd_max_order;
2191 }
2192 finish_wait(&pgdat->kswapd_wait, &wait);
2193
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002194 if (!try_to_freeze()) {
2195 /* We can speed up thawing tasks if we don't call
2196 * balance_pgdat after returning from the refrigerator
2197 */
2198 balance_pgdat(pgdat, order);
2199 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200 }
2201 return 0;
2202}
2203
2204/*
2205 * A zone is low on free memory, so wake its kswapd task to service it.
2206 */
2207void wakeup_kswapd(struct zone *zone, int order)
2208{
2209 pg_data_t *pgdat;
2210
Con Kolivasf3fe6512006-01-06 00:11:15 -08002211 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 return;
2213
2214 pgdat = zone->zone_pgdat;
Mel Gorman41858962009-06-16 15:32:12 -07002215 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216 return;
2217 if (pgdat->kswapd_max_order < order)
2218 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08002219 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002220 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002221 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002223 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224}
2225
Wu Fengguangadea02a2009-09-21 17:01:42 -07002226/*
2227 * The reclaimable count would be mostly accurate.
2228 * The less reclaimable pages may be
2229 * - mlocked pages, which will be moved to unevictable list when encountered
2230 * - mapped pages, which may require several travels to be reclaimed
2231 * - dirty pages, which is not "instantly" reclaimable
2232 */
2233unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002234{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002235 int nr;
2236
2237 nr = global_page_state(NR_ACTIVE_FILE) +
2238 global_page_state(NR_INACTIVE_FILE);
2239
2240 if (nr_swap_pages > 0)
2241 nr += global_page_state(NR_ACTIVE_ANON) +
2242 global_page_state(NR_INACTIVE_ANON);
2243
2244 return nr;
2245}
2246
2247unsigned long zone_reclaimable_pages(struct zone *zone)
2248{
2249 int nr;
2250
2251 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2252 zone_page_state(zone, NR_INACTIVE_FILE);
2253
2254 if (nr_swap_pages > 0)
2255 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2256 zone_page_state(zone, NR_INACTIVE_ANON);
2257
2258 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002259}
2260
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002261#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262/*
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002263 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
MinChan Kimd9796772009-03-31 15:19:34 -07002264 * from LRU lists system-wide, for given pass and priority.
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002265 *
2266 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2267 */
MinChan Kimd9796772009-03-31 15:19:34 -07002268static void shrink_all_zones(unsigned long nr_pages, int prio,
Nigel Cunninghame07aa052006-12-22 01:07:21 -08002269 int pass, struct scan_control *sc)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002270{
2271 struct zone *zone;
MinChan Kimd9796772009-03-31 15:19:34 -07002272 unsigned long nr_reclaimed = 0;
Wu Fengguangf8629632009-09-21 17:03:11 -07002273 struct zone_reclaim_stat *reclaim_stat;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002274
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002275 for_each_populated_zone(zone) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002276 enum lru_list l;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002277
David Rientjese815af92007-10-16 23:25:54 -07002278 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002279 continue;
2280
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002281 for_each_evictable_lru(l) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002282 enum zone_stat_item ls = NR_LRU_BASE + l;
2283 unsigned long lru_pages = zone_page_state(zone, ls);
2284
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002285 /* For pass = 0, we don't shrink the active list */
Johannes Weiner0cb57252009-02-14 02:04:10 +01002286 if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2287 l == LRU_ACTIVE_FILE))
Christoph Lameterb69408e2008-10-18 20:26:14 -07002288 continue;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002289
Wu Fengguangf8629632009-09-21 17:03:11 -07002290 reclaim_stat = get_reclaim_stat(zone, sc);
2291 reclaim_stat->nr_saved_scan[l] +=
2292 (lru_pages >> prio) + 1;
2293 if (reclaim_stat->nr_saved_scan[l]
2294 >= nr_pages || pass > 3) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002295 unsigned long nr_to_scan;
2296
Wu Fengguangf8629632009-09-21 17:03:11 -07002297 reclaim_stat->nr_saved_scan[l] = 0;
Johannes Weiner0cb57252009-02-14 02:04:10 +01002298 nr_to_scan = min(nr_pages, lru_pages);
MinChan Kimd9796772009-03-31 15:19:34 -07002299 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
Christoph Lameterb69408e2008-10-18 20:26:14 -07002300 sc, prio);
MinChan Kimd9796772009-03-31 15:19:34 -07002301 if (nr_reclaimed >= nr_pages) {
Rafael J. Wysockia21e2552009-04-18 17:23:41 +02002302 sc->nr_reclaimed += nr_reclaimed;
MinChan Kimd9796772009-03-31 15:19:34 -07002303 return;
2304 }
Christoph Lameterb69408e2008-10-18 20:26:14 -07002305 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002306 }
2307 }
Rafael J. Wysockia21e2552009-04-18 17:23:41 +02002308 sc->nr_reclaimed += nr_reclaimed;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002309}
2310
2311/*
2312 * Try to free `nr_pages' of memory, system-wide, and return the number of
2313 * freed pages.
2314 *
2315 * Rather than trying to age LRUs the aim is to preserve the overall
2316 * LRU order by reclaiming preferentially
2317 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002318 */
Andrew Morton69e05942006-03-22 00:08:19 -08002319unsigned long shrink_all_memory(unsigned long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002321 unsigned long lru_pages, nr_slab;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002322 int pass;
2323 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002324 struct scan_control sc = {
2325 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002326 .may_unmap = 0,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002327 .may_writepage = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002328 .isolate_pages = isolate_pages_global,
Rafael J. Wysockia21e2552009-04-18 17:23:41 +02002329 .nr_reclaimed = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002330 };
2331
2332 current->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002333
Wu Fengguangadea02a2009-09-21 17:01:42 -07002334 lru_pages = global_reclaimable_pages();
Christoph Lameter972d1a72006-09-25 23:31:51 -07002335 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002336 /* If slab caches are huge, it's better to hit them first */
2337 while (nr_slab >= lru_pages) {
2338 reclaim_state.reclaimed_slab = 0;
2339 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2340 if (!reclaim_state.reclaimed_slab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002341 break;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002342
MinChan Kimd9796772009-03-31 15:19:34 -07002343 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2344 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002345 goto out;
2346
2347 nr_slab -= reclaim_state.reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002349
2350 /*
2351 * We try to shrink LRUs in 5 passes:
2352 * 0 = Reclaim from inactive_list only
2353 * 1 = Reclaim from active list but don't reclaim mapped
2354 * 2 = 2nd pass of type 1
2355 * 3 = Reclaim mapped (normal reclaim)
2356 * 4 = 2nd pass of type 3
2357 */
2358 for (pass = 0; pass < 5; pass++) {
2359 int prio;
2360
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002361 /* Force reclaiming mapped pages in the passes #3 and #4 */
Johannes Weiner30491032009-02-14 02:03:08 +01002362 if (pass > 2)
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002363 sc.may_unmap = 1;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002364
2365 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
MinChan Kimd9796772009-03-31 15:19:34 -07002366 unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002367
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002368 sc.nr_scanned = 0;
Johannes Weiner9786bf82009-03-31 15:19:35 -07002369 sc.swap_cluster_max = nr_to_scan;
MinChan Kimd9796772009-03-31 15:19:34 -07002370 shrink_all_zones(nr_to_scan, prio, pass, &sc);
2371 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002372 goto out;
2373
2374 reclaim_state.reclaimed_slab = 0;
Andrew Morton76395d32007-01-05 16:37:05 -08002375 shrink_slab(sc.nr_scanned, sc.gfp_mask,
Wu Fengguangadea02a2009-09-21 17:01:42 -07002376 global_reclaimable_pages());
MinChan Kimd9796772009-03-31 15:19:34 -07002377 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2378 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002379 goto out;
2380
2381 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02002382 congestion_wait(BLK_RW_ASYNC, HZ / 10);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002383 }
Rafael J. Wysocki248a0302006-03-22 00:09:04 -08002384 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002385
2386 /*
MinChan Kimd9796772009-03-31 15:19:34 -07002387 * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2388 * something in slab caches
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002389 */
MinChan Kimd9796772009-03-31 15:19:34 -07002390 if (!sc.nr_reclaimed) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002391 do {
2392 reclaim_state.reclaimed_slab = 0;
Wu Fengguangadea02a2009-09-21 17:01:42 -07002393 shrink_slab(nr_pages, sc.gfp_mask,
2394 global_reclaimable_pages());
MinChan Kimd9796772009-03-31 15:19:34 -07002395 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2396 } while (sc.nr_reclaimed < nr_pages &&
2397 reclaim_state.reclaimed_slab > 0);
Andrew Morton76395d32007-01-05 16:37:05 -08002398 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002399
MinChan Kimd9796772009-03-31 15:19:34 -07002400
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002401out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 current->reclaim_state = NULL;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002403
MinChan Kimd9796772009-03-31 15:19:34 -07002404 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002405}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002406#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408/* It's optimal to keep kswapds on the same CPUs as their memory, but
2409 not required for correctness. So if the last cpu in a node goes
2410 away, we get changed to run anywhere: as the first one comes back,
2411 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002412static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002413 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002415 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002416
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002417 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002418 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002419 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302420 const struct cpumask *mask;
2421
2422 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002423
Rusty Russell3e597942009-01-01 10:12:24 +10302424 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002426 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427 }
2428 }
2429 return NOTIFY_OK;
2430}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431
Yasunori Goto3218ae12006-06-27 02:53:33 -07002432/*
2433 * This kswapd start function will be called by init and node-hot-add.
2434 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2435 */
2436int kswapd_run(int nid)
2437{
2438 pg_data_t *pgdat = NODE_DATA(nid);
2439 int ret = 0;
2440
2441 if (pgdat->kswapd)
2442 return 0;
2443
2444 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2445 if (IS_ERR(pgdat->kswapd)) {
2446 /* failure at boot is fatal */
2447 BUG_ON(system_state == SYSTEM_BOOTING);
2448 printk("Failed to start kswapd on node %d\n",nid);
2449 ret = -1;
2450 }
2451 return ret;
2452}
2453
Linus Torvalds1da177e2005-04-16 15:20:36 -07002454static int __init kswapd_init(void)
2455{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002456 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002457
Linus Torvalds1da177e2005-04-16 15:20:36 -07002458 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002459 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002460 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461 hotcpu_notifier(cpu_callback, 0);
2462 return 0;
2463}
2464
2465module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002466
2467#ifdef CONFIG_NUMA
2468/*
2469 * Zone reclaim mode
2470 *
2471 * If non-zero call zone_reclaim when the number of free pages falls below
2472 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002473 */
2474int zone_reclaim_mode __read_mostly;
2475
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002476#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002477#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002478#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2479#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2480
Christoph Lameter9eeff232006-01-18 17:42:31 -08002481/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002482 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2483 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2484 * a zone.
2485 */
2486#define ZONE_RECLAIM_PRIORITY 4
2487
Christoph Lameter9eeff232006-01-18 17:42:31 -08002488/*
Christoph Lameter96146342006-07-03 00:24:13 -07002489 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2490 * occur.
2491 */
2492int sysctl_min_unmapped_ratio = 1;
2493
2494/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002495 * If the number of slab pages in a zone grows beyond this percentage then
2496 * slab reclaim needs to occur.
2497 */
2498int sysctl_min_slab_ratio = 5;
2499
Mel Gorman90afa5d2009-06-16 15:33:20 -07002500static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2501{
2502 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2503 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2504 zone_page_state(zone, NR_ACTIVE_FILE);
2505
2506 /*
2507 * It's possible for there to be more file mapped pages than
2508 * accounted for by the pages on the file LRU lists because
2509 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2510 */
2511 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2512}
2513
2514/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2515static long zone_pagecache_reclaimable(struct zone *zone)
2516{
2517 long nr_pagecache_reclaimable;
2518 long delta = 0;
2519
2520 /*
2521 * If RECLAIM_SWAP is set, then all file pages are considered
2522 * potentially reclaimable. Otherwise, we have to worry about
2523 * pages like swapcache and zone_unmapped_file_pages() provides
2524 * a better estimate
2525 */
2526 if (zone_reclaim_mode & RECLAIM_SWAP)
2527 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2528 else
2529 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2530
2531 /* If we can't clean pages, remove dirty pages from consideration */
2532 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2533 delta += zone_page_state(zone, NR_FILE_DIRTY);
2534
2535 /* Watch for any possible underflows due to delta */
2536 if (unlikely(delta > nr_pagecache_reclaimable))
2537 delta = nr_pagecache_reclaimable;
2538
2539 return nr_pagecache_reclaimable - delta;
2540}
2541
Christoph Lameter0ff38492006-09-25 23:31:52 -07002542/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002543 * Try to free up some pages from this zone through reclaim.
2544 */
Andrew Morton179e9632006-03-22 00:08:18 -08002545static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002546{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002547 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002548 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002549 struct task_struct *p = current;
2550 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002551 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002552 struct scan_control sc = {
2553 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002554 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002555 .may_swap = 1,
Andrew Morton69e05942006-03-22 00:08:19 -08002556 .swap_cluster_max = max_t(unsigned long, nr_pages,
2557 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002558 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002559 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002560 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002561 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08002562 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07002563 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002564
2565 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08002566 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002567 /*
2568 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2569 * and we also need to be able to write out pages for RECLAIM_WRITE
2570 * and RECLAIM_SWAP.
2571 */
2572 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002573 reclaim_state.reclaimed_slab = 0;
2574 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002575
Mel Gorman90afa5d2009-06-16 15:33:20 -07002576 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07002577 /*
2578 * Free memory by calling shrink zone with increasing
2579 * priorities until we have enough memory freed.
2580 */
2581 priority = ZONE_RECLAIM_PRIORITY;
2582 do {
Martin Bligh3bb1a852006-10-28 10:38:24 -07002583 note_zone_scanning_priority(zone, priority);
Rik van Riela79311c2009-01-06 14:40:01 -08002584 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002585 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002586 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002587 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002588
Christoph Lameter83e33a42006-09-25 23:31:53 -07002589 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2590 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002591 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002592 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002593 * many pages were freed in this zone. So we take the current
2594 * number of slab pages and shake the slab until it is reduced
2595 * by the same nr_pages that we used for reclaiming unmapped
2596 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002597 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002598 * Note that shrink_slab will free memory on all zones and may
2599 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002600 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07002601 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07002602 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2603 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07002604 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07002605
2606 /*
2607 * Update nr_reclaimed by the number of slab pages we
2608 * reclaimed from this zone.
2609 */
Rik van Riela79311c2009-01-06 14:40:01 -08002610 sc.nr_reclaimed += slab_reclaimable -
Christoph Lameter83e33a42006-09-25 23:31:53 -07002611 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002612 }
2613
Christoph Lameter9eeff232006-01-18 17:42:31 -08002614 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002615 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
Rik van Riela79311c2009-01-06 14:40:01 -08002616 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002617}
Andrew Morton179e9632006-03-22 00:08:18 -08002618
2619int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2620{
Andrew Morton179e9632006-03-22 00:08:18 -08002621 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002622 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002623
2624 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002625 * Zone reclaim reclaims unmapped file backed pages and
2626 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002627 *
Christoph Lameter96146342006-07-03 00:24:13 -07002628 * A small portion of unmapped file backed pages is needed for
2629 * file I/O otherwise pages read by file I/O will be immediately
2630 * thrown out if the zone is overallocated. So we do not reclaim
2631 * if less than a specified percentage of the zone is used by
2632 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002633 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07002634 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2635 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002636 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08002637
David Rientjesd773ed62007-10-16 23:26:01 -07002638 if (zone_is_all_unreclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002639 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07002640
Andrew Morton179e9632006-03-22 00:08:18 -08002641 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002642 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002643 */
David Rientjesd773ed62007-10-16 23:26:01 -07002644 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002645 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08002646
2647 /*
2648 * Only run zone reclaim on the local zone or on zones that do not
2649 * have associated processors. This will favor the local processor
2650 * over remote processors and spread off node memory allocations
2651 * as wide as possible.
2652 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002653 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002654 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07002655 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07002656
2657 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002658 return ZONE_RECLAIM_NOSCAN;
2659
David Rientjesd773ed62007-10-16 23:26:01 -07002660 ret = __zone_reclaim(zone, gfp_mask, order);
2661 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2662
Mel Gorman24cf725182009-06-16 15:33:23 -07002663 if (!ret)
2664 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2665
David Rientjesd773ed62007-10-16 23:26:01 -07002666 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002667}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002668#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002669
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002670/*
2671 * page_evictable - test whether a page is evictable
2672 * @page: the page to test
2673 * @vma: the VMA in which the page is or will be mapped, may be NULL
2674 *
2675 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07002676 * lists vs unevictable list. The vma argument is !NULL when called from the
2677 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002678 *
2679 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002680 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07002681 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002682 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002683 */
2684int page_evictable(struct page *page, struct vm_area_struct *vma)
2685{
2686
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002687 if (mapping_unevictable(page_mapping(page)))
2688 return 0;
2689
Nick Pigginb291f002008-10-18 20:26:44 -07002690 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2691 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002692
2693 return 1;
2694}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002695
2696/**
2697 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2698 * @page: page to check evictability and move to appropriate lru list
2699 * @zone: zone page is in
2700 *
2701 * Checks a page for evictability and moves the page to the appropriate
2702 * zone lru list.
2703 *
2704 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2705 * have PageUnevictable set.
2706 */
2707static void check_move_unevictable_page(struct page *page, struct zone *zone)
2708{
2709 VM_BUG_ON(PageActive(page));
2710
2711retry:
2712 ClearPageUnevictable(page);
2713 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07002714 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002715
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002716 __dec_zone_state(zone, NR_UNEVICTABLE);
2717 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002718 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002719 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2720 __count_vm_event(UNEVICTABLE_PGRESCUED);
2721 } else {
2722 /*
2723 * rotate unevictable list
2724 */
2725 SetPageUnevictable(page);
2726 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002727 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002728 if (page_evictable(page, NULL))
2729 goto retry;
2730 }
2731}
2732
2733/**
2734 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2735 * @mapping: struct address_space to scan for evictable pages
2736 *
2737 * Scan all pages in mapping. Check unevictable pages for
2738 * evictability and move them to the appropriate zone lru list.
2739 */
2740void scan_mapping_unevictable_pages(struct address_space *mapping)
2741{
2742 pgoff_t next = 0;
2743 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2744 PAGE_CACHE_SHIFT;
2745 struct zone *zone;
2746 struct pagevec pvec;
2747
2748 if (mapping->nrpages == 0)
2749 return;
2750
2751 pagevec_init(&pvec, 0);
2752 while (next < end &&
2753 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2754 int i;
2755 int pg_scanned = 0;
2756
2757 zone = NULL;
2758
2759 for (i = 0; i < pagevec_count(&pvec); i++) {
2760 struct page *page = pvec.pages[i];
2761 pgoff_t page_index = page->index;
2762 struct zone *pagezone = page_zone(page);
2763
2764 pg_scanned++;
2765 if (page_index > next)
2766 next = page_index;
2767 next++;
2768
2769 if (pagezone != zone) {
2770 if (zone)
2771 spin_unlock_irq(&zone->lru_lock);
2772 zone = pagezone;
2773 spin_lock_irq(&zone->lru_lock);
2774 }
2775
2776 if (PageLRU(page) && PageUnevictable(page))
2777 check_move_unevictable_page(page, zone);
2778 }
2779 if (zone)
2780 spin_unlock_irq(&zone->lru_lock);
2781 pagevec_release(&pvec);
2782
2783 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2784 }
2785
2786}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002787
2788/**
2789 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2790 * @zone - zone of which to scan the unevictable list
2791 *
2792 * Scan @zone's unevictable LRU lists to check for pages that have become
2793 * evictable. Move those that have to @zone's inactive list where they
2794 * become candidates for reclaim, unless shrink_inactive_zone() decides
2795 * to reactivate them. Pages that are still unevictable are rotated
2796 * back onto @zone's unevictable list.
2797 */
2798#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08002799static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002800{
2801 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2802 unsigned long scan;
2803 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2804
2805 while (nr_to_scan > 0) {
2806 unsigned long batch_size = min(nr_to_scan,
2807 SCAN_UNEVICTABLE_BATCH_SIZE);
2808
2809 spin_lock_irq(&zone->lru_lock);
2810 for (scan = 0; scan < batch_size; scan++) {
2811 struct page *page = lru_to_page(l_unevictable);
2812
2813 if (!trylock_page(page))
2814 continue;
2815
2816 prefetchw_prev_lru_page(page, l_unevictable, flags);
2817
2818 if (likely(PageLRU(page) && PageUnevictable(page)))
2819 check_move_unevictable_page(page, zone);
2820
2821 unlock_page(page);
2822 }
2823 spin_unlock_irq(&zone->lru_lock);
2824
2825 nr_to_scan -= batch_size;
2826 }
2827}
2828
2829
2830/**
2831 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2832 *
2833 * A really big hammer: scan all zones' unevictable LRU lists to check for
2834 * pages that have become evictable. Move those back to the zones'
2835 * inactive list where they become candidates for reclaim.
2836 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2837 * and we add swap to the system. As such, it runs in the context of a task
2838 * that has possibly/probably made some previously unevictable pages
2839 * evictable.
2840 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08002841static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002842{
2843 struct zone *zone;
2844
2845 for_each_zone(zone) {
2846 scan_zone_unevictable_pages(zone);
2847 }
2848}
2849
2850/*
2851 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2852 * all nodes' unevictable lists for evictable pages
2853 */
2854unsigned long scan_unevictable_pages;
2855
2856int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002857 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002858 size_t *length, loff_t *ppos)
2859{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002860 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002861
2862 if (write && *(unsigned long *)table->data)
2863 scan_all_zones_unevictable_pages();
2864
2865 scan_unevictable_pages = 0;
2866 return 0;
2867}
2868
2869/*
2870 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2871 * a specified node's per zone unevictable lists for evictable pages.
2872 */
2873
2874static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2875 struct sysdev_attribute *attr,
2876 char *buf)
2877{
2878 return sprintf(buf, "0\n"); /* always zero; should fit... */
2879}
2880
2881static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2882 struct sysdev_attribute *attr,
2883 const char *buf, size_t count)
2884{
2885 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2886 struct zone *zone;
2887 unsigned long res;
2888 unsigned long req = strict_strtoul(buf, 10, &res);
2889
2890 if (!req)
2891 return 1; /* zero is no-op */
2892
2893 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2894 if (!populated_zone(zone))
2895 continue;
2896 scan_zone_unevictable_pages(zone);
2897 }
2898 return 1;
2899}
2900
2901
2902static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2903 read_scan_unevictable_node,
2904 write_scan_unevictable_node);
2905
2906int scan_unevictable_register_node(struct node *node)
2907{
2908 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2909}
2910
2911void scan_unevictable_unregister_node(struct node *node)
2912{
2913 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2914}
2915