blob: 64e438898832371277e8baeb89fd5ebb1f72a14a [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
16#include <linux/slab.h>
17#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070022#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/file.h>
24#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28#include <linux/mm_inline.h>
29#include <linux/pagevec.h>
30#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
35#include <linux/notifier.h>
36#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080037#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070038#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080039#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080040#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070041#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070042#include <linux/sysctl.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44#include <asm/tlbflush.h>
45#include <asm/div64.h>
46
47#include <linux/swapops.h>
48
Nick Piggin0f8053a2006-03-22 00:08:33 -080049#include "internal.h"
50
Linus Torvalds1da177e2005-04-16 15:20:36 -070051struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 /* Incremented by the number of inactive pages that were scanned */
53 unsigned long nr_scanned;
54
Rik van Riela79311c2009-01-06 14:40:01 -080055 /* Number of pages freed so far during a call to shrink_zones() */
56 unsigned long nr_reclaimed;
57
Linus Torvalds1da177e2005-04-16 15:20:36 -070058 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040059 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
61 int may_writepage;
62
Johannes Weinera6dc60f82009-03-31 15:19:30 -070063 /* Can mapped pages be reclaimed? */
64 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080065
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070066 /* Can pages be swapped as part of reclaim? */
67 int may_swap;
68
Linus Torvalds1da177e2005-04-16 15:20:36 -070069 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
70 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
71 * In this context, it doesn't matter that we scan the
72 * whole list at once. */
73 int swap_cluster_max;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070074
75 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070076
77 int all_unreclaimable;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070078
79 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080080
81 /* Which cgroup do we reclaim from */
82 struct mem_cgroup *mem_cgroup;
83
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -070084 /*
85 * Nodemask of nodes allowed by the caller. If NULL, all nodes
86 * are scanned.
87 */
88 nodemask_t *nodemask;
89
Balbir Singh66e17072008-02-07 00:13:56 -080090 /* Pluggable isolate pages callback */
91 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
92 unsigned long *scanned, int order, int mode,
93 struct zone *z, struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -070094 int active, int file);
Linus Torvalds1da177e2005-04-16 15:20:36 -070095};
96
Linus Torvalds1da177e2005-04-16 15:20:36 -070097#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
98
99#ifdef ARCH_HAS_PREFETCH
100#define prefetch_prev_lru_page(_page, _base, _field) \
101 do { \
102 if ((_page)->lru.prev != _base) { \
103 struct page *prev; \
104 \
105 prev = lru_to_page(&(_page->lru)); \
106 prefetch(&prev->_field); \
107 } \
108 } while (0)
109#else
110#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
111#endif
112
113#ifdef ARCH_HAS_PREFETCHW
114#define prefetchw_prev_lru_page(_page, _base, _field) \
115 do { \
116 if ((_page)->lru.prev != _base) { \
117 struct page *prev; \
118 \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetchw(&prev->_field); \
121 } \
122 } while (0)
123#else
124#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
125#endif
126
127/*
128 * From 0 .. 100. Higher means more swappy.
129 */
130int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700131long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132
133static LIST_HEAD(shrinker_list);
134static DECLARE_RWSEM(shrinker_rwsem);
135
Balbir Singh00f0b822008-03-04 14:28:39 -0800136#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800137#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800138#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800139#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800140#endif
141
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800142static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
143 struct scan_control *sc)
144{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800145 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800146 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
147
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800148 return &zone->reclaim_stat;
149}
150
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);
547 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700548
549 /*
550 * page's status can change while we move it among lru. If an evictable
551 * page is on unevictable list, it never be freed. To avoid that,
552 * check after we added it to the list, again.
553 */
554 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
555 if (!isolate_lru_page(page)) {
556 put_page(page);
557 goto redo;
558 }
559 /* This means someone else dropped this page from LRU
560 * So, it will be freed or putback to LRU again. There is
561 * nothing to do here.
562 */
563 }
564
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700565 if (was_unevictable && lru != LRU_UNEVICTABLE)
566 count_vm_event(UNEVICTABLE_PGRESCUED);
567 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
568 count_vm_event(UNEVICTABLE_PGCULLED);
569
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700570 put_page(page); /* drop ref from isolate */
571}
572
Nick Piggine2867812008-07-25 19:45:30 -0700573/*
Andrew Morton1742f192006-03-22 00:08:21 -0800574 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 */
Andrew Morton1742f192006-03-22 00:08:21 -0800576static unsigned long shrink_page_list(struct list_head *page_list,
Andy Whitcroftc661b072007-08-22 14:01:26 -0700577 struct scan_control *sc,
578 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579{
580 LIST_HEAD(ret_pages);
581 struct pagevec freed_pvec;
582 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800583 unsigned long nr_reclaimed = 0;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700584 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700585
586 cond_resched();
587
588 pagevec_init(&freed_pvec, 1);
589 while (!list_empty(page_list)) {
590 struct address_space *mapping;
591 struct page *page;
592 int may_enter_fs;
593 int referenced;
594
595 cond_resched();
596
597 page = lru_to_page(page_list);
598 list_del(&page->lru);
599
Nick Piggin529ae9a2008-08-02 12:01:03 +0200600 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700601 goto keep;
602
Nick Piggin725d7042006-09-25 23:30:55 -0700603 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700604
605 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800606
Nick Pigginb291f002008-10-18 20:26:44 -0700607 if (unlikely(!page_evictable(page, NULL)))
608 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700609
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700610 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800611 goto keep_locked;
612
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613 /* Double the slab pressure for mapped and swapcache pages */
614 if (page_mapped(page) || PageSwapCache(page))
615 sc->nr_scanned++;
616
Andy Whitcroftc661b072007-08-22 14:01:26 -0700617 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
618 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
619
620 if (PageWriteback(page)) {
621 /*
622 * Synchronous reclaim is performed in two passes,
623 * first an asynchronous pass over the list to
624 * start parallel writeback, and a second synchronous
625 * pass to wait for the IO to complete. Wait here
626 * for any page for which writeback has already
627 * started.
628 */
629 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
630 wait_on_page_writeback(page);
Andrew Morton4dd4b922008-03-24 12:29:52 -0700631 else
Andy Whitcroftc661b072007-08-22 14:01:26 -0700632 goto keep_locked;
633 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700634
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700635 referenced = page_referenced(page, 1,
636 sc->mem_cgroup, &vm_flags);
Minchan Kim03ef83a2009-08-26 14:29:23 -0700637 /*
638 * In active use or really unfreeable? Activate it.
639 * If page which have PG_mlocked lost isoltation race,
640 * try_to_unmap moves it to unevictable list
641 */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700642 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
Minchan Kim03ef83a2009-08-26 14:29:23 -0700643 referenced && page_mapping_inuse(page)
644 && !(vm_flags & VM_LOCKED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 goto activate_locked;
646
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 /*
648 * Anonymous process memory has backing store?
649 * Try to allocate it some swap space here.
650 */
Nick Pigginb291f002008-10-18 20:26:44 -0700651 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800652 if (!(sc->gfp_mask & __GFP_IO))
653 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800654 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800656 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700657 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658
659 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660
661 /*
662 * The page is mapped into the page tables of one or more
663 * processes. Try to unmap it here.
664 */
665 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200666 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 case SWAP_FAIL:
668 goto activate_locked;
669 case SWAP_AGAIN:
670 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700671 case SWAP_MLOCK:
672 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 case SWAP_SUCCESS:
674 ; /* try to free the page below */
675 }
676 }
677
678 if (PageDirty(page)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700679 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700681 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800683 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 goto keep_locked;
685
686 /* Page is dirty, try to write it out here */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700687 switch (pageout(page, mapping, sync_writeback)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 case PAGE_KEEP:
689 goto keep_locked;
690 case PAGE_ACTIVATE:
691 goto activate_locked;
692 case PAGE_SUCCESS:
Andrew Morton4dd4b922008-03-24 12:29:52 -0700693 if (PageWriteback(page) || PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 goto keep;
695 /*
696 * A synchronous write - probably a ramdisk. Go
697 * ahead and try to reclaim the page.
698 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200699 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 goto keep;
701 if (PageDirty(page) || PageWriteback(page))
702 goto keep_locked;
703 mapping = page_mapping(page);
704 case PAGE_CLEAN:
705 ; /* try to free the page below */
706 }
707 }
708
709 /*
710 * If the page has buffers, try to free the buffer mappings
711 * associated with this page. If we succeed we try to free
712 * the page as well.
713 *
714 * We do this even if the page is PageDirty().
715 * try_to_release_page() does not perform I/O, but it is
716 * possible for a page to have PageDirty set, but it is actually
717 * clean (all its buffers are clean). This happens if the
718 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700719 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 * try_to_release_page() will discover that cleanness and will
721 * drop the buffers and mark the page clean - it can be freed.
722 *
723 * Rarely, pages can have buffers and no ->mapping. These are
724 * the pages which were not successfully invalidated in
725 * truncate_complete_page(). We try to drop those buffers here
726 * and if that worked, and the page is no longer mapped into
727 * process address space (page_count == 1) it can be freed.
728 * Otherwise, leave the page on the LRU so it is swappable.
729 */
David Howells266cf652009-04-03 16:42:36 +0100730 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731 if (!try_to_release_page(page, sc->gfp_mask))
732 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700733 if (!mapping && page_count(page) == 1) {
734 unlock_page(page);
735 if (put_page_testzero(page))
736 goto free_it;
737 else {
738 /*
739 * rare race with speculative reference.
740 * the speculative reference will free
741 * this page shortly, so we may
742 * increment nr_reclaimed here (and
743 * leave it off the LRU).
744 */
745 nr_reclaimed++;
746 continue;
747 }
748 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 }
750
Nick Piggine2867812008-07-25 19:45:30 -0700751 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800752 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753
Nick Piggina978d6f2008-10-18 20:26:58 -0700754 /*
755 * At this point, we have no other references and there is
756 * no way to pick any more up (removed from LRU, removed
757 * from pagecache). Can use non-atomic bitops now (and
758 * we obviously don't have to worry about waking up a process
759 * waiting on the page lock, because there are no references.
760 */
761 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700762free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800763 nr_reclaimed++;
Nick Piggine2867812008-07-25 19:45:30 -0700764 if (!pagevec_add(&freed_pvec, page)) {
765 __pagevec_free(&freed_pvec);
766 pagevec_reinit(&freed_pvec);
767 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 continue;
769
Nick Pigginb291f002008-10-18 20:26:44 -0700770cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800771 if (PageSwapCache(page))
772 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700773 unlock_page(page);
774 putback_lru_page(page);
775 continue;
776
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700778 /* Not a candidate for swapping, so reclaim swap space. */
779 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800780 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700781 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 SetPageActive(page);
783 pgactivate++;
784keep_locked:
785 unlock_page(page);
786keep:
787 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700788 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 }
790 list_splice(&ret_pages, page_list);
791 if (pagevec_count(&freed_pvec))
Nick Piggine2867812008-07-25 19:45:30 -0700792 __pagevec_free(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700793 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800794 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795}
796
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700797/* LRU Isolation modes. */
798#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
799#define ISOLATE_ACTIVE 1 /* Isolate active pages. */
800#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
801
802/*
803 * Attempt to remove the specified page from its LRU. Only take this page
804 * if it is of the appropriate PageActive status. Pages which are being
805 * freed elsewhere are also ignored.
806 *
807 * page: page to consider
808 * mode: one of the LRU isolation modes defined above
809 *
810 * returns 0 on success, -ve errno on failure.
811 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700812int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700813{
814 int ret = -EINVAL;
815
816 /* Only take pages on the LRU. */
817 if (!PageLRU(page))
818 return ret;
819
820 /*
821 * When checking the active state, we need to be sure we are
822 * dealing with comparible boolean values. Take the logical not
823 * of each.
824 */
825 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
826 return ret;
827
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700828 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700829 return ret;
830
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700831 /*
832 * When this function is being called for lumpy reclaim, we
833 * initially look into all LRU pages, active, inactive and
834 * unevictable; only give shrink_page_list evictable pages.
835 */
836 if (PageUnevictable(page))
837 return ret;
838
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700839 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800840
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700841 if (likely(get_page_unless_zero(page))) {
842 /*
843 * Be careful not to clear PageLRU until after we're
844 * sure the page is not being freed elsewhere -- the
845 * page release code relies on it.
846 */
847 ClearPageLRU(page);
848 ret = 0;
849 }
850
851 return ret;
852}
853
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800854/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700855 * zone->lru_lock is heavily contended. Some of the functions that
856 * shrink the lists perform better by taking out a batch of pages
857 * and working on them outside the LRU lock.
858 *
859 * For pagecache intensive workloads, this function is the hottest
860 * spot in the kernel (apart from copy_*_user functions).
861 *
862 * Appropriate locks must be held before calling this function.
863 *
864 * @nr_to_scan: The number of pages to look through on the list.
865 * @src: The LRU list to pull pages off.
866 * @dst: The temp list to put pages on to.
867 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700868 * @order: The caller's attempted allocation order
869 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700870 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700871 *
872 * returns how many pages were moved onto *@dst.
873 */
Andrew Morton69e05942006-03-22 00:08:19 -0800874static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
875 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700876 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700877{
Andrew Morton69e05942006-03-22 00:08:19 -0800878 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800879 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800881 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700882 struct page *page;
883 unsigned long pfn;
884 unsigned long end_pfn;
885 unsigned long page_pfn;
886 int zone_id;
887
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 page = lru_to_page(src);
889 prefetchw_prev_lru_page(page, src, flags);
890
Nick Piggin725d7042006-09-25 23:30:55 -0700891 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800892
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700893 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700894 case 0:
895 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700896 mem_cgroup_del_lru(page);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800897 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700898 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800899
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700900 case -EBUSY:
901 /* else it is being freed elsewhere */
902 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700903 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700904 continue;
905
906 default:
907 BUG();
908 }
909
910 if (!order)
911 continue;
912
913 /*
914 * Attempt to take all pages in the order aligned region
915 * surrounding the tag page. Only take those pages of
916 * the same active state as that tag page. We may safely
917 * round the target page pfn down to the requested order
918 * as the mem_map is guarenteed valid out to MAX_ORDER,
919 * where that page is in a different zone we will detect
920 * it from its zone id and abort this block scan.
921 */
922 zone_id = page_zone_id(page);
923 page_pfn = page_to_pfn(page);
924 pfn = page_pfn & ~((1 << order) - 1);
925 end_pfn = pfn + (1 << order);
926 for (; pfn < end_pfn; pfn++) {
927 struct page *cursor_page;
928
929 /* The target page is in the block, ignore it. */
930 if (unlikely(pfn == page_pfn))
931 continue;
932
933 /* Avoid holes within the zone. */
934 if (unlikely(!pfn_valid_within(pfn)))
935 break;
936
937 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700938
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700939 /* Check that we have not crossed a zone boundary. */
940 if (unlikely(page_zone_id(cursor_page) != zone_id))
941 continue;
Minchan Kimde2e7562009-09-21 17:01:43 -0700942
943 /*
944 * If we don't have enough swap space, reclaiming of
945 * anon page which don't already have a swap slot is
946 * pointless.
947 */
948 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
949 !PageSwapCache(cursor_page))
950 continue;
951
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -0700952 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700953 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +0900954 mem_cgroup_del_lru(cursor_page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700955 nr_taken++;
956 scan++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700957 }
958 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 }
960
961 *scanned = scan;
962 return nr_taken;
963}
964
Balbir Singh66e17072008-02-07 00:13:56 -0800965static unsigned long isolate_pages_global(unsigned long nr,
966 struct list_head *dst,
967 unsigned long *scanned, int order,
968 int mode, struct zone *z,
969 struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700970 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -0800971{
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700972 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -0800973 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700974 lru += LRU_ACTIVE;
975 if (file)
976 lru += LRU_FILE;
977 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -0700978 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -0800979}
980
Linus Torvalds1da177e2005-04-16 15:20:36 -0700981/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700982 * clear_active_flags() is a helper for shrink_active_list(), clearing
983 * any active bits from the pages in the list.
984 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700985static unsigned long clear_active_flags(struct list_head *page_list,
986 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700987{
988 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700989 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700990 struct page *page;
991
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700992 list_for_each_entry(page, page_list, lru) {
Johannes Weiner401a8e12009-09-21 17:02:58 -0700993 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700994 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700995 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700996 ClearPageActive(page);
997 nr_active++;
998 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700999 count[lru]++;
1000 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001001
1002 return nr_active;
1003}
1004
Nick Piggin62695a82008-10-18 20:26:09 -07001005/**
1006 * isolate_lru_page - tries to isolate a page from its LRU list
1007 * @page: page to isolate from its LRU list
1008 *
1009 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1010 * vmstat statistic corresponding to whatever LRU list the page was on.
1011 *
1012 * Returns 0 if the page was removed from an LRU list.
1013 * Returns -EBUSY if the page was not on an LRU list.
1014 *
1015 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001016 * the active list, it will have PageActive set. If it was found on
1017 * the unevictable list, it will have the PageUnevictable bit set. That flag
1018 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001019 *
1020 * The vmstat statistic corresponding to the list on which the page was
1021 * found will be decremented.
1022 *
1023 * Restrictions:
1024 * (1) Must be called with an elevated refcount on the page. This is a
1025 * fundamentnal difference from isolate_lru_pages (which is called
1026 * without a stable reference).
1027 * (2) the lru_lock must not be held.
1028 * (3) interrupts must be enabled.
1029 */
1030int isolate_lru_page(struct page *page)
1031{
1032 int ret = -EBUSY;
1033
1034 if (PageLRU(page)) {
1035 struct zone *zone = page_zone(page);
1036
1037 spin_lock_irq(&zone->lru_lock);
1038 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001039 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001040 ret = 0;
1041 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001042
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001043 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001044 }
1045 spin_unlock_irq(&zone->lru_lock);
1046 }
1047 return ret;
1048}
1049
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001050/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001051 * Are there way too many processes in the direct reclaim path already?
1052 */
1053static int too_many_isolated(struct zone *zone, int file,
1054 struct scan_control *sc)
1055{
1056 unsigned long inactive, isolated;
1057
1058 if (current_is_kswapd())
1059 return 0;
1060
1061 if (!scanning_global_lru(sc))
1062 return 0;
1063
1064 if (file) {
1065 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1066 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1067 } else {
1068 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1069 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1070 }
1071
1072 return isolated > inactive;
1073}
1074
1075/*
Andrew Morton1742f192006-03-22 00:08:21 -08001076 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1077 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078 */
Andrew Morton1742f192006-03-22 00:08:21 -08001079static unsigned long shrink_inactive_list(unsigned long max_scan,
Rik van Riel33c120e2008-10-18 20:26:36 -07001080 struct zone *zone, struct scan_control *sc,
1081 int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001082{
1083 LIST_HEAD(page_list);
1084 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -08001085 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001086 unsigned long nr_reclaimed = 0;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001087 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001088 int lumpy_reclaim = 0;
1089
Rik van Riel35cd7812009-09-21 17:01:38 -07001090 while (unlikely(too_many_isolated(zone, file, sc))) {
1091 congestion_wait(WRITE, HZ/10);
1092
1093 /* We are about to die and free our memory. Return now. */
1094 if (fatal_signal_pending(current))
1095 return SWAP_CLUSTER_MAX;
1096 }
1097
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001098 /*
1099 * If we need a large contiguous chunk of memory, or have
1100 * trouble getting a small set of contiguous pages, we
1101 * will reclaim both active and inactive pages.
1102 *
1103 * We use the same threshold as pageout congestion_wait below.
1104 */
1105 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1106 lumpy_reclaim = 1;
1107 else if (sc->order && priority < DEF_PRIORITY - 2)
1108 lumpy_reclaim = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001109
1110 pagevec_init(&pvec, 1);
1111
1112 lru_add_drain();
1113 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -08001114 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -08001116 unsigned long nr_taken;
1117 unsigned long nr_scan;
1118 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001119 unsigned long nr_active;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001120 unsigned int count[NR_LRU_LISTS] = { 0, };
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001121 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001122 unsigned long nr_anon;
1123 unsigned long nr_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124
Balbir Singh66e17072008-02-07 00:13:56 -08001125 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001126 &page_list, &nr_scan, sc->order, mode,
1127 zone, sc->mem_cgroup, 0, file);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001128
1129 if (scanning_global_lru(sc)) {
1130 zone->pages_scanned += nr_scan;
1131 if (current_is_kswapd())
1132 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1133 nr_scan);
1134 else
1135 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1136 nr_scan);
1137 }
1138
1139 if (nr_taken == 0)
1140 goto done;
1141
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001142 nr_active = clear_active_flags(&page_list, count);
Andy Whitcrofte9187bd2007-08-22 14:01:25 -07001143 __count_vm_events(PGDEACTIVATE, nr_active);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001144
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001145 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1146 -count[LRU_ACTIVE_FILE]);
1147 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1148 -count[LRU_INACTIVE_FILE]);
1149 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1150 -count[LRU_ACTIVE_ANON]);
1151 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1152 -count[LRU_INACTIVE_ANON]);
1153
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001154 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1155 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1156 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1157 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001158
1159 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1160 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1161 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1162 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1163
Linus Torvalds1da177e2005-04-16 15:20:36 -07001164 spin_unlock_irq(&zone->lru_lock);
1165
Andrew Morton69e05942006-03-22 00:08:19 -08001166 nr_scanned += nr_scan;
Andy Whitcroftc661b072007-08-22 14:01:26 -07001167 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1168
1169 /*
1170 * If we are direct reclaiming for contiguous pages and we do
1171 * not reclaim everything in the list, try again and wait
1172 * for IO to complete. This will stall high-order allocations
1173 * but that should be acceptable to the caller
1174 */
1175 if (nr_freed < nr_taken && !current_is_kswapd() &&
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001176 lumpy_reclaim) {
Jens Axboe8aa7e842009-07-09 14:52:32 +02001177 congestion_wait(BLK_RW_ASYNC, HZ/10);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001178
1179 /*
1180 * The attempt at page out may have made some
1181 * of the pages active, mark them inactive again.
1182 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001183 nr_active = clear_active_flags(&page_list, count);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001184 count_vm_events(PGDEACTIVATE, nr_active);
1185
1186 nr_freed += shrink_page_list(&page_list, sc,
1187 PAGEOUT_IO_SYNC);
1188 }
1189
Andrew Morton05ff5132006-03-22 00:08:20 -08001190 nr_reclaimed += nr_freed;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001191
Nick Piggina74609f2006-01-06 00:11:20 -08001192 local_irq_disable();
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001193 if (current_is_kswapd())
Christoph Lameterf8891e52006-06-30 01:55:45 -07001194 __count_vm_events(KSWAPD_STEAL, nr_freed);
Shantanu Goel918d3f92006-12-29 16:48:59 -08001195 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -08001196
1197 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 /*
1199 * Put back any unfreeable pages.
1200 */
1201 while (!list_empty(&page_list)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001202 int lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001203 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -07001204 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 list_del(&page->lru);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001206 if (unlikely(!page_evictable(page, NULL))) {
1207 spin_unlock_irq(&zone->lru_lock);
1208 putback_lru_page(page);
1209 spin_lock_irq(&zone->lru_lock);
1210 continue;
1211 }
1212 SetPageLRU(page);
1213 lru = page_lru(page);
1214 add_page_to_lru_list(zone, page, lru);
KOSAKI Motohiro74a1c482009-09-21 17:01:45 -07001215 if (is_active_lru(lru)) {
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001216 int file = is_file_lru(lru);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001217 reclaim_stat->recent_rotated[file]++;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001218 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 if (!pagevec_add(&pvec, page)) {
1220 spin_unlock_irq(&zone->lru_lock);
1221 __pagevec_release(&pvec);
1222 spin_lock_irq(&zone->lru_lock);
1223 }
1224 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001225 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1226 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1227
Andrew Morton69e05942006-03-22 00:08:19 -08001228 } while (nr_scanned < max_scan);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001229
Linus Torvalds1da177e2005-04-16 15:20:36 -07001230done:
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001231 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001232 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -08001233 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001234}
1235
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001236/*
1237 * We are about to scan this zone at a certain priority level. If that priority
1238 * level is smaller (ie: more urgent) than the previous priority, then note
1239 * that priority level within the zone. This is done so that when the next
1240 * process comes in to scan this zone, it will immediately start out at this
1241 * priority level rather than having to build up its own scanning priority.
1242 * Here, this priority affects only the reclaim-mapped threshold.
1243 */
1244static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1245{
1246 if (priority < zone->prev_priority)
1247 zone->prev_priority = priority;
1248}
1249
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001250/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251 * This moves pages from the active list to the inactive list.
1252 *
1253 * We move them the other way if the page is referenced by one or more
1254 * processes, from rmap.
1255 *
1256 * If the pages are mostly unmapped, the processing is fast and it is
1257 * appropriate to hold zone->lru_lock across the whole operation. But if
1258 * the pages are mapped, the processing is slow (page_referenced()) so we
1259 * should drop zone->lru_lock around each page. It's impossible to balance
1260 * this, so instead we remove the pages from the LRU while processing them.
1261 * It is safe to rely on PG_active against the non-LRU pages in here because
1262 * nobody will play with that bit on a non-LRU page.
1263 *
1264 * The downside is that we have to touch page->_count against each page.
1265 * But we had to alter page->flags anyway.
1266 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001267
Wu Fengguang3eb41402009-06-16 15:33:13 -07001268static void move_active_pages_to_lru(struct zone *zone,
1269 struct list_head *list,
1270 enum lru_list lru)
1271{
1272 unsigned long pgmoved = 0;
1273 struct pagevec pvec;
1274 struct page *page;
1275
1276 pagevec_init(&pvec, 1);
1277
1278 while (!list_empty(list)) {
1279 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001280
1281 VM_BUG_ON(PageLRU(page));
1282 SetPageLRU(page);
1283
Wu Fengguang3eb41402009-06-16 15:33:13 -07001284 list_move(&page->lru, &zone->lru[lru].list);
1285 mem_cgroup_add_lru_list(page, lru);
1286 pgmoved++;
1287
1288 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1289 spin_unlock_irq(&zone->lru_lock);
1290 if (buffer_heads_over_limit)
1291 pagevec_strip(&pvec);
1292 __pagevec_release(&pvec);
1293 spin_lock_irq(&zone->lru_lock);
1294 }
1295 }
1296 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1297 if (!is_active_lru(lru))
1298 __count_vm_events(PGDEACTIVATE, pgmoved);
1299}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001300
Andrew Morton1742f192006-03-22 00:08:21 -08001301static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001302 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001303{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001304 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001305 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001306 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001308 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001309 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001311 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001312 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001313
1314 lru_add_drain();
1315 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001316 nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
Balbir Singh66e17072008-02-07 00:13:56 -08001317 ISOLATE_ACTIVE, zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001318 sc->mem_cgroup, 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001319 /*
1320 * zone->pages_scanned is used for detect zone's oom
1321 * mem_cgroup remembers nr_scan by itself.
1322 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001323 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001324 zone->pages_scanned += pgscanned;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001325 }
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001326 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001327
Wu Fengguang3eb41402009-06-16 15:33:13 -07001328 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001329 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001330 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001331 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001332 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001333 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001334 spin_unlock_irq(&zone->lru_lock);
1335
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 while (!list_empty(&l_hold)) {
1337 cond_resched();
1338 page = lru_to_page(&l_hold);
1339 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001340
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001341 if (unlikely(!page_evictable(page, NULL))) {
1342 putback_lru_page(page);
1343 continue;
1344 }
1345
Rik van Riel7e9cd482008-10-18 20:26:35 -07001346 /* page_referenced clears PageReferenced */
1347 if (page_mapping_inuse(page) &&
Wu Fengguang8cab4752009-06-16 15:33:12 -07001348 page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001349 nr_rotated++;
Wu Fengguang8cab4752009-06-16 15:33:12 -07001350 /*
1351 * Identify referenced, file-backed active pages and
1352 * give them one more trip around the active list. So
1353 * that executable code get better chances to stay in
1354 * memory under moderate memory pressure. Anon pages
1355 * are not likely to be evicted by use-once streaming
1356 * IO, plus JVM can create lots of anon VM_EXEC pages,
1357 * so we ignore them here.
1358 */
1359 if ((vm_flags & VM_EXEC) && !PageAnon(page)) {
1360 list_add(&page->lru, &l_active);
1361 continue;
1362 }
1363 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001364
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001365 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001366 list_add(&page->lru, &l_inactive);
1367 }
1368
Andrew Mortonb5557492009-01-06 14:40:13 -08001369 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001370 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001371 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001372 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001373 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001374 * Count referenced pages from currently used mappings as rotated,
1375 * even though only some of them are actually re-activated. This
1376 * helps balance scan pressure between file and anonymous pages in
1377 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001378 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001379 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001380
Wu Fengguang3eb41402009-06-16 15:33:13 -07001381 move_active_pages_to_lru(zone, &l_active,
1382 LRU_ACTIVE + file * LRU_FILE);
1383 move_active_pages_to_lru(zone, &l_inactive,
1384 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001385 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001386 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001387}
1388
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001389static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001390{
1391 unsigned long active, inactive;
1392
1393 active = zone_page_state(zone, NR_ACTIVE_ANON);
1394 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1395
1396 if (inactive * zone->inactive_ratio < active)
1397 return 1;
1398
1399 return 0;
1400}
1401
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001402/**
1403 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1404 * @zone: zone to check
1405 * @sc: scan control of this context
1406 *
1407 * Returns true if the zone does not have enough inactive anon pages,
1408 * meaning some active anon pages need to be deactivated.
1409 */
1410static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1411{
1412 int low;
1413
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001414 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001415 low = inactive_anon_is_low_global(zone);
1416 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001417 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001418 return low;
1419}
1420
Rik van Riel56e49d22009-06-16 15:32:28 -07001421static int inactive_file_is_low_global(struct zone *zone)
1422{
1423 unsigned long active, inactive;
1424
1425 active = zone_page_state(zone, NR_ACTIVE_FILE);
1426 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1427
1428 return (active > inactive);
1429}
1430
1431/**
1432 * inactive_file_is_low - check if file pages need to be deactivated
1433 * @zone: zone to check
1434 * @sc: scan control of this context
1435 *
1436 * When the system is doing streaming IO, memory pressure here
1437 * ensures that active file pages get deactivated, until more
1438 * than half of the file pages are on the inactive list.
1439 *
1440 * Once we get to that situation, protect the system's working
1441 * set from being evicted by disabling active file page aging.
1442 *
1443 * This uses a different ratio than the anonymous pages, because
1444 * the page cache uses a use-once replacement algorithm.
1445 */
1446static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1447{
1448 int low;
1449
1450 if (scanning_global_lru(sc))
1451 low = inactive_file_is_low_global(zone);
1452 else
1453 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1454 return low;
1455}
1456
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001457static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001458 struct zone *zone, struct scan_control *sc, int priority)
1459{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001460 int file = is_file_lru(lru);
1461
Rik van Riel56e49d22009-06-16 15:32:28 -07001462 if (lru == LRU_ACTIVE_FILE && inactive_file_is_low(zone, sc)) {
Rik van Riel556adec2008-10-18 20:26:34 -07001463 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1464 return 0;
1465 }
1466
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001467 if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001468 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001469 return 0;
1470 }
Rik van Riel33c120e2008-10-18 20:26:36 -07001471 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001472}
1473
Linus Torvalds1da177e2005-04-16 15:20:36 -07001474/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001475 * Determine how aggressively the anon and file LRU lists should be
1476 * scanned. The relative value of each set of LRU lists is determined
1477 * by looking at the fraction of the pages scanned we did rotate back
1478 * onto the active list instead of evict.
1479 *
1480 * percent[0] specifies how much pressure to put on ram/swap backed
1481 * memory, while percent[1] determines pressure on the file LRUs.
1482 */
1483static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1484 unsigned long *percent)
1485{
1486 unsigned long anon, file, free;
1487 unsigned long anon_prio, file_prio;
1488 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001489 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001490
Vincent Li0b217672009-09-21 17:03:09 -07001491 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1492 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1493 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1494 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001495
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001496 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001497 free = zone_page_state(zone, NR_FREE_PAGES);
1498 /* If we have very few page cache pages,
1499 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001500 if (unlikely(file + free <= high_wmark_pages(zone))) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001501 percent[0] = 100;
1502 percent[1] = 0;
1503 return;
1504 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001505 }
1506
1507 /*
1508 * OK, so we have swap space and a fair amount of page cache
1509 * pages. We use the recently rotated / recently scanned
1510 * ratios to determine how valuable each cache is.
1511 *
1512 * Because workloads change over time (and to avoid overflow)
1513 * we keep these statistics as a floating average, which ends
1514 * up weighing recent references more than old ones.
1515 *
1516 * anon in [0], file in [1]
1517 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001518 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001519 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001520 reclaim_stat->recent_scanned[0] /= 2;
1521 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001522 spin_unlock_irq(&zone->lru_lock);
1523 }
1524
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001525 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001526 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001527 reclaim_stat->recent_scanned[1] /= 2;
1528 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001529 spin_unlock_irq(&zone->lru_lock);
1530 }
1531
1532 /*
1533 * With swappiness at 100, anonymous and file have the same priority.
1534 * This scanning priority is essentially the inverse of IO cost.
1535 */
1536 anon_prio = sc->swappiness;
1537 file_prio = 200 - sc->swappiness;
1538
1539 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001540 * The amount of pressure on anon vs file pages is inversely
1541 * proportional to the fraction of recently scanned pages on
1542 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001543 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001544 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1545 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001546
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001547 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1548 fp /= reclaim_stat->recent_rotated[1] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001549
1550 /* Normalize to percentages */
1551 percent[0] = 100 * ap / (ap + fp + 1);
1552 percent[1] = 100 - percent[0];
1553}
1554
Wu Fengguang6e08a362009-06-16 15:32:29 -07001555/*
1556 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1557 * until we collected @swap_cluster_max pages to scan.
1558 */
1559static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1560 unsigned long *nr_saved_scan,
1561 unsigned long swap_cluster_max)
1562{
1563 unsigned long nr;
1564
1565 *nr_saved_scan += nr_to_scan;
1566 nr = *nr_saved_scan;
1567
1568 if (nr >= swap_cluster_max)
1569 *nr_saved_scan = 0;
1570 else
1571 nr = 0;
1572
1573 return nr;
1574}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001575
1576/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001577 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1578 */
Rik van Riela79311c2009-01-06 14:40:01 -08001579static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001580 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001581{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001582 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001583 unsigned long nr_to_scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001584 unsigned long percent[2]; /* anon @ 0; file @ 1 */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001585 enum lru_list l;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001586 unsigned long nr_reclaimed = sc->nr_reclaimed;
1587 unsigned long swap_cluster_max = sc->swap_cluster_max;
Wu Fengguangf8629632009-09-21 17:03:11 -07001588 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001589 int noswap = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001590
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001591 /* If we have no swap space, do not bother scanning anon pages. */
1592 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1593 noswap = 1;
1594 percent[0] = 0;
1595 percent[1] = 100;
1596 } else
1597 get_scan_ratio(zone, sc, percent);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001598
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001599 for_each_evictable_lru(l) {
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001600 int file = is_file_lru(l);
Andrew Morton8713e012009-04-30 15:08:55 -07001601 unsigned long scan;
Johannes Weinere0f79b82008-10-18 20:26:55 -07001602
Vincent Li0b217672009-09-21 17:03:09 -07001603 scan = zone_nr_lru_pages(zone, sc, l);
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001604 if (priority || noswap) {
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001605 scan >>= priority;
1606 scan = (scan * percent[file]) / 100;
1607 }
Wu Fengguangf8629632009-09-21 17:03:11 -07001608 nr[l] = nr_scan_try_batch(scan,
1609 &reclaim_stat->nr_saved_scan[l],
1610 swap_cluster_max);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001611 }
1612
Rik van Riel556adec2008-10-18 20:26:34 -07001613 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1614 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001615 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001616 if (nr[l]) {
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001617 nr_to_scan = min(nr[l], swap_cluster_max);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001618 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001620 nr_reclaimed += shrink_list(l, nr_to_scan,
1621 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001622 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623 }
Rik van Riela79311c2009-01-06 14:40:01 -08001624 /*
1625 * On large memory systems, scan >> priority can become
1626 * really large. This is fine for the starting priority;
1627 * we want to put equal scanning pressure on each zone.
1628 * However, if the VM has a harder time of freeing pages,
1629 * with multiple processes reclaiming pages, the total
1630 * freeing target can get unreasonably large.
1631 */
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001632 if (nr_reclaimed > swap_cluster_max &&
Rik van Riela79311c2009-01-06 14:40:01 -08001633 priority < DEF_PRIORITY && !current_is_kswapd())
1634 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 }
1636
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001637 sc->nr_reclaimed = nr_reclaimed;
1638
Rik van Riel556adec2008-10-18 20:26:34 -07001639 /*
1640 * Even if we did not try to evict anon pages at all, we want to
1641 * rebalance the anon lru active/inactive ratio.
1642 */
MinChan Kim69c85482009-06-16 15:32:44 -07001643 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
Rik van Riel556adec2008-10-18 20:26:34 -07001644 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1645
Andrew Morton232ea4d2007-02-28 20:13:21 -08001646 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001647}
1648
1649/*
1650 * This is the direct reclaim path, for page-allocating processes. We only
1651 * try to reclaim pages from zones which will satisfy the caller's allocation
1652 * request.
1653 *
Mel Gorman41858962009-06-16 15:32:12 -07001654 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1655 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1657 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001658 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1659 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1660 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662 * If a zone is deemed to be full of pinned pages then just give it a light
1663 * scan then give up on it.
1664 */
Rik van Riela79311c2009-01-06 14:40:01 -08001665static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001666 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667{
Mel Gorman54a6eb52008-04-28 02:12:16 -07001668 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Mel Gormandd1a2392008-04-28 02:12:17 -07001669 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001670 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001671
Nick Piggin408d8542006-09-25 23:31:27 -07001672 sc->all_unreclaimable = 1;
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001673 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1674 sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001675 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001677 /*
1678 * Take care memory controller reclaiming has small influence
1679 * to global LRU.
1680 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001681 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001682 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1683 continue;
1684 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001686 if (zone_is_all_unreclaimable(zone) &&
1687 priority != DEF_PRIORITY)
1688 continue; /* Let kswapd poll it */
1689 sc->all_unreclaimable = 0;
1690 } else {
1691 /*
1692 * Ignore cpuset limitation here. We just want to reduce
1693 * # of used pages by us regardless of memory shortage.
1694 */
1695 sc->all_unreclaimable = 0;
1696 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1697 priority);
1698 }
Nick Piggin408d8542006-09-25 23:31:27 -07001699
Rik van Riela79311c2009-01-06 14:40:01 -08001700 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001701 }
1702}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001703
Linus Torvalds1da177e2005-04-16 15:20:36 -07001704/*
1705 * This is the main entry point to direct page reclaim.
1706 *
1707 * If a full scan of the inactive list fails to free enough memory then we
1708 * are "out of memory" and something needs to be killed.
1709 *
1710 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1711 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02001712 * caller can't do much about. We kick the writeback threads and take explicit
1713 * naps in the hope that some of these pages can be written. But if the
1714 * allocating task holds filesystem locks which prevent writeout this might not
1715 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001716 *
1717 * returns: 0, if no pages reclaimed
1718 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 */
Mel Gormandac1d272008-04-28 02:12:12 -07001720static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001721 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001722{
1723 int priority;
kosaki.motohiro@jp.fujitsu.comc700be32008-06-12 15:21:27 -07001724 unsigned long ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001725 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001727 unsigned long lru_pages = 0;
Mel Gormandd1a2392008-04-28 02:12:17 -07001728 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001729 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -07001730 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731
Keika Kobayashi873b4772008-07-25 01:48:52 -07001732 delayacct_freepages_start();
1733
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001734 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001735 count_vm_event(ALLOCSTALL);
1736 /*
1737 * mem_cgroup will not do shrink_slab.
1738 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001739 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001740 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001741
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001742 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1743 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001744
Wu Fengguangadea02a2009-09-21 17:01:42 -07001745 lru_pages += zone_reclaimable_pages(zone);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001746 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747 }
1748
1749 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001750 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001751 if (!priority)
1752 disable_swap_token();
Rik van Riela79311c2009-01-06 14:40:01 -08001753 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001754 /*
1755 * Don't shrink slabs when reclaiming memory from
1756 * over limit cgroups
1757 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001758 if (scanning_global_lru(sc)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07001759 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001760 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08001761 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001762 reclaim_state->reclaimed_slab = 0;
1763 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001764 }
Balbir Singh66e17072008-02-07 00:13:56 -08001765 total_scanned += sc->nr_scanned;
Rik van Riela79311c2009-01-06 14:40:01 -08001766 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1767 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 goto out;
1769 }
1770
1771 /*
1772 * Try to write back as many pages as we just scanned. This
1773 * tends to cause slow streaming writers to write data to the
1774 * disk smoothly, at the dirtying rate, which is nice. But
1775 * that's undesirable in laptop mode, where we *want* lumpy
1776 * writeout. So in laptop mode, write out the whole world.
1777 */
Balbir Singh66e17072008-02-07 00:13:56 -08001778 if (total_scanned > sc->swap_cluster_max +
1779 sc->swap_cluster_max / 2) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001780 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001781 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782 }
1783
1784 /* Take a nap, wait for some writeback to complete */
Andrew Morton4dd4b922008-03-24 12:29:52 -07001785 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02001786 congestion_wait(BLK_RW_ASYNC, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 }
Fernando Luis Vazquez Cao87547ee2008-07-29 22:33:42 -07001788 /* top priority shrink_zones still had more to do? don't OOM, then */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001789 if (!sc->all_unreclaimable && scanning_global_lru(sc))
Rik van Riela79311c2009-01-06 14:40:01 -08001790 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001792 /*
1793 * Now that we've scanned all the zones at this priority level, note
1794 * that level within the zone so that the next thread which performs
1795 * scanning of this zone will immediately start out at this priority
1796 * level. This affects only the decision whether or not to bring
1797 * mapped pages onto the inactive list.
1798 */
1799 if (priority < 0)
1800 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001802 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001803 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001805 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1806 continue;
1807
1808 zone->prev_priority = priority;
1809 }
1810 } else
1811 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1812
Keika Kobayashi873b4772008-07-25 01:48:52 -07001813 delayacct_freepages_end();
1814
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815 return ret;
1816}
1817
Mel Gormandac1d272008-04-28 02:12:12 -07001818unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001819 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08001820{
1821 struct scan_control sc = {
1822 .gfp_mask = gfp_mask,
1823 .may_writepage = !laptop_mode,
1824 .swap_cluster_max = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07001825 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001826 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08001827 .swappiness = vm_swappiness,
1828 .order = order,
1829 .mem_cgroup = NULL,
1830 .isolate_pages = isolate_pages_global,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001831 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08001832 };
1833
Mel Gormandd1a2392008-04-28 02:12:17 -07001834 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001835}
1836
Balbir Singh00f0b822008-03-04 14:28:39 -08001837#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08001838
Balbir Singh4e416952009-09-23 15:56:39 -07001839unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1840 gfp_t gfp_mask, bool noswap,
1841 unsigned int swappiness,
1842 struct zone *zone, int nid)
1843{
1844 struct scan_control sc = {
1845 .may_writepage = !laptop_mode,
1846 .may_unmap = 1,
1847 .may_swap = !noswap,
1848 .swap_cluster_max = SWAP_CLUSTER_MAX,
1849 .swappiness = swappiness,
1850 .order = 0,
1851 .mem_cgroup = mem,
1852 .isolate_pages = mem_cgroup_isolate_pages,
1853 };
1854 nodemask_t nm = nodemask_of_node(nid);
1855
1856 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1857 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1858 sc.nodemask = &nm;
1859 sc.nr_reclaimed = 0;
1860 sc.nr_scanned = 0;
1861 /*
1862 * NOTE: Although we can get the priority field, using it
1863 * here is not a good idea, since it limits the pages we can scan.
1864 * if we don't reclaim here, the shrink_zone from balance_pgdat
1865 * will pick up pages from other mem cgroup's as well. We hack
1866 * the priority and make it zero.
1867 */
1868 shrink_zone(0, zone, &sc);
1869 return sc.nr_reclaimed;
1870}
1871
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001872unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001873 gfp_t gfp_mask,
1874 bool noswap,
1875 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08001876{
Balbir Singh4e416952009-09-23 15:56:39 -07001877 struct zonelist *zonelist;
Balbir Singh66e17072008-02-07 00:13:56 -08001878 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08001879 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07001880 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001881 .may_swap = !noswap,
Balbir Singh66e17072008-02-07 00:13:56 -08001882 .swap_cluster_max = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001883 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08001884 .order = 0,
1885 .mem_cgroup = mem_cont,
1886 .isolate_pages = mem_cgroup_isolate_pages,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001887 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08001888 };
Balbir Singh66e17072008-02-07 00:13:56 -08001889
Mel Gormandd1a2392008-04-28 02:12:17 -07001890 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1891 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1892 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1893 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001894}
1895#endif
1896
Linus Torvalds1da177e2005-04-16 15:20:36 -07001897/*
1898 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07001899 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001900 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001901 * Returns the number of pages which were actually freed.
1902 *
1903 * There is special handling here for zones which are full of pinned pages.
1904 * This can happen if the pages are all mlocked, or if they are all used by
1905 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1906 * What we do is to detect the case where all pages in the zone have been
1907 * scanned twice and there has been zero successful reclaim. Mark the zone as
1908 * dead and from now on, only perform a short scan. Basically we're polling
1909 * the zone for when the problem goes away.
1910 *
1911 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07001912 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1913 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1914 * lower zones regardless of the number of free pages in the lower zones. This
1915 * interoperates with the page allocator fallback scheme to ensure that aging
1916 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001917 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001918static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001920 int all_zones_ok;
1921 int priority;
1922 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08001923 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001924 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08001925 struct scan_control sc = {
1926 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07001927 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001928 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001929 .swap_cluster_max = SWAP_CLUSTER_MAX,
1930 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001931 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08001932 .mem_cgroup = NULL,
1933 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08001934 };
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001935 /*
1936 * temp_priority is used to remember the scanning priority at which
Mel Gorman41858962009-06-16 15:32:12 -07001937 * this zone was successfully refilled to
1938 * free_pages == high_wmark_pages(zone).
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001939 */
1940 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941
1942loop_again:
1943 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08001944 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07001945 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001946 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001948 for (i = 0; i < pgdat->nr_zones; i++)
1949 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950
1951 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1952 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1953 unsigned long lru_pages = 0;
1954
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001955 /* The swap token gets in the way of swapout... */
1956 if (!priority)
1957 disable_swap_token();
1958
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 all_zones_ok = 1;
1960
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001961 /*
1962 * Scan in the highmem->dma direction for the highest
1963 * zone which needs scanning
1964 */
1965 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1966 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001967
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001968 if (!populated_zone(zone))
1969 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001970
David Rientjese815af92007-10-16 23:25:54 -07001971 if (zone_is_all_unreclaimable(zone) &&
1972 priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001973 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974
Rik van Riel556adec2008-10-18 20:26:34 -07001975 /*
1976 * Do some background aging of the anon list, to give
1977 * pages a chance to be referenced before reclaiming.
1978 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001979 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07001980 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1981 &sc, priority, 0);
1982
Mel Gorman41858962009-06-16 15:32:12 -07001983 if (!zone_watermark_ok(zone, order,
1984 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001985 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001986 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001987 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001989 if (i < 0)
1990 goto out;
1991
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992 for (i = 0; i <= end_zone; i++) {
1993 struct zone *zone = pgdat->node_zones + i;
1994
Wu Fengguangadea02a2009-09-21 17:01:42 -07001995 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 }
1997
1998 /*
1999 * Now scan the zone in the dma->highmem direction, stopping
2000 * at the last zone which needs scanning.
2001 *
2002 * We do this because the page allocator works in the opposite
2003 * direction. This prevents the page allocator from allocating
2004 * pages behind kswapd's direction of progress, which would
2005 * cause too much scanning of the lower zones.
2006 */
2007 for (i = 0; i <= end_zone; i++) {
2008 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002009 int nr_slab;
Balbir Singh4e416952009-09-23 15:56:39 -07002010 int nid, zid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011
Con Kolivasf3fe6512006-01-06 00:11:15 -08002012 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013 continue;
2014
David Rientjese815af92007-10-16 23:25:54 -07002015 if (zone_is_all_unreclaimable(zone) &&
2016 priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002017 continue;
2018
Mel Gorman41858962009-06-16 15:32:12 -07002019 if (!zone_watermark_ok(zone, order,
2020 high_wmark_pages(zone), end_zone, 0))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002021 all_zones_ok = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07002022 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002023 sc.nr_scanned = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07002024 note_zone_scanning_priority(zone, priority);
Balbir Singh4e416952009-09-23 15:56:39 -07002025
2026 nid = pgdat->node_id;
2027 zid = zone_idx(zone);
2028 /*
2029 * Call soft limit reclaim before calling shrink_zone.
2030 * For now we ignore the return value
2031 */
2032 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2033 nid, zid);
Rik van Riel32a43302007-10-16 01:24:50 -07002034 /*
2035 * We put equal pressure on every zone, unless one
2036 * zone has way too many pages free already.
2037 */
Mel Gorman41858962009-06-16 15:32:12 -07002038 if (!zone_watermark_ok(zone, order,
2039 8*high_wmark_pages(zone), end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002040 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002042 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2043 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002044 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002045 total_scanned += sc.nr_scanned;
David Rientjese815af92007-10-16 23:25:54 -07002046 if (zone_is_all_unreclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002047 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002048 if (nr_slab == 0 && zone->pages_scanned >=
Wu Fengguangadea02a2009-09-21 17:01:42 -07002049 (zone_reclaimable_pages(zone) * 6))
David Rientjese815af92007-10-16 23:25:54 -07002050 zone_set_flag(zone,
2051 ZONE_ALL_UNRECLAIMABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052 /*
2053 * If we've done a decent amount of scanning and
2054 * the reclaim ratio is low, start doing writepage
2055 * even in laptop mode
2056 */
2057 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002058 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002059 sc.may_writepage = 1;
2060 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061 if (all_zones_ok)
2062 break; /* kswapd: all done */
2063 /*
2064 * OK, kswapd is getting into trouble. Take a nap, then take
2065 * another pass across the zones.
2066 */
Andrew Morton4dd4b922008-03-24 12:29:52 -07002067 if (total_scanned && priority < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02002068 congestion_wait(BLK_RW_ASYNC, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069
2070 /*
2071 * We do this so kswapd doesn't build up large priorities for
2072 * example when it is freeing in parallel with allocators. It
2073 * matches the direct reclaim path behaviour in terms of impact
2074 * on zone->*_priority.
2075 */
Rik van Riela79311c2009-01-06 14:40:01 -08002076 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 break;
2078 }
2079out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07002080 /*
2081 * Note within each zone the priority level at which this zone was
2082 * brought into a happy state. So that the next thread which scans this
2083 * zone will start out at that priority level.
2084 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 for (i = 0; i < pgdat->nr_zones; i++) {
2086 struct zone *zone = pgdat->node_zones + i;
2087
Martin Bligh3bb1a8522006-10-28 10:38:24 -07002088 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089 }
2090 if (!all_zones_ok) {
2091 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002092
2093 try_to_freeze();
2094
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002095 /*
2096 * Fragmentation may mean that the system cannot be
2097 * rebalanced for high-order allocations in all zones.
2098 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2099 * it means the zones have been fully scanned and are still
2100 * not balanced. For high-order allocations, there is
2101 * little point trying all over again as kswapd may
2102 * infinite loop.
2103 *
2104 * Instead, recheck all watermarks at order-0 as they
2105 * are the most important. If watermarks are ok, kswapd will go
2106 * back to sleep. High-order users can still perform direct
2107 * reclaim if they wish.
2108 */
2109 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2110 order = sc.order = 0;
2111
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112 goto loop_again;
2113 }
2114
Rik van Riela79311c2009-01-06 14:40:01 -08002115 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002116}
2117
2118/*
2119 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002120 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002121 *
2122 * This basically trickles out pages so that we have _some_
2123 * free memory available even if there is no other activity
2124 * that frees anything up. This is needed for things like routing
2125 * etc, where we otherwise might have all activity going on in
2126 * asynchronous contexts that cannot page things out.
2127 *
2128 * If there are applications that are active memory-allocators
2129 * (most normal use), this basically shouldn't matter.
2130 */
2131static int kswapd(void *p)
2132{
2133 unsigned long order;
2134 pg_data_t *pgdat = (pg_data_t*)p;
2135 struct task_struct *tsk = current;
2136 DEFINE_WAIT(wait);
2137 struct reclaim_state reclaim_state = {
2138 .reclaimed_slab = 0,
2139 };
Rusty Russella70f7302009-03-13 14:49:46 +10302140 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141
Nick Piggincf40bd12009-01-21 08:12:39 +01002142 lockdep_set_current_reclaim_state(GFP_KERNEL);
2143
Rusty Russell174596a2009-01-01 10:12:29 +10302144 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002145 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 current->reclaim_state = &reclaim_state;
2147
2148 /*
2149 * Tell the memory management that we're a "memory allocator",
2150 * and that if we need more memory we should get access to it
2151 * regardless (see "__alloc_pages()"). "kswapd" should
2152 * never get caught in the normal page freeing logic.
2153 *
2154 * (Kswapd normally doesn't need memory anyway, but sometimes
2155 * you need a small amount of memory in order to be able to
2156 * page out something else, and this flag essentially protects
2157 * us from recursively trying to free more memory as we're
2158 * trying to free the first piece of memory in the first place).
2159 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002160 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002161 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002162
2163 order = 0;
2164 for ( ; ; ) {
2165 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002166
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2168 new_order = pgdat->kswapd_max_order;
2169 pgdat->kswapd_max_order = 0;
2170 if (order < new_order) {
2171 /*
2172 * Don't sleep if someone wants a larger 'order'
2173 * allocation
2174 */
2175 order = new_order;
2176 } else {
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002177 if (!freezing(current))
2178 schedule();
2179
Linus Torvalds1da177e2005-04-16 15:20:36 -07002180 order = pgdat->kswapd_max_order;
2181 }
2182 finish_wait(&pgdat->kswapd_wait, &wait);
2183
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002184 if (!try_to_freeze()) {
2185 /* We can speed up thawing tasks if we don't call
2186 * balance_pgdat after returning from the refrigerator
2187 */
2188 balance_pgdat(pgdat, order);
2189 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002190 }
2191 return 0;
2192}
2193
2194/*
2195 * A zone is low on free memory, so wake its kswapd task to service it.
2196 */
2197void wakeup_kswapd(struct zone *zone, int order)
2198{
2199 pg_data_t *pgdat;
2200
Con Kolivasf3fe6512006-01-06 00:11:15 -08002201 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202 return;
2203
2204 pgdat = zone->zone_pgdat;
Mel Gorman41858962009-06-16 15:32:12 -07002205 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 return;
2207 if (pgdat->kswapd_max_order < order)
2208 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08002209 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002211 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002213 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214}
2215
Wu Fengguangadea02a2009-09-21 17:01:42 -07002216/*
2217 * The reclaimable count would be mostly accurate.
2218 * The less reclaimable pages may be
2219 * - mlocked pages, which will be moved to unevictable list when encountered
2220 * - mapped pages, which may require several travels to be reclaimed
2221 * - dirty pages, which is not "instantly" reclaimable
2222 */
2223unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002224{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002225 int nr;
2226
2227 nr = global_page_state(NR_ACTIVE_FILE) +
2228 global_page_state(NR_INACTIVE_FILE);
2229
2230 if (nr_swap_pages > 0)
2231 nr += global_page_state(NR_ACTIVE_ANON) +
2232 global_page_state(NR_INACTIVE_ANON);
2233
2234 return nr;
2235}
2236
2237unsigned long zone_reclaimable_pages(struct zone *zone)
2238{
2239 int nr;
2240
2241 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2242 zone_page_state(zone, NR_INACTIVE_FILE);
2243
2244 if (nr_swap_pages > 0)
2245 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2246 zone_page_state(zone, NR_INACTIVE_ANON);
2247
2248 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002249}
2250
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002251#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002252/*
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002253 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
MinChan Kimd9796772009-03-31 15:19:34 -07002254 * from LRU lists system-wide, for given pass and priority.
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002255 *
2256 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2257 */
MinChan Kimd9796772009-03-31 15:19:34 -07002258static void shrink_all_zones(unsigned long nr_pages, int prio,
Nigel Cunninghame07aa052006-12-22 01:07:21 -08002259 int pass, struct scan_control *sc)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002260{
2261 struct zone *zone;
MinChan Kimd9796772009-03-31 15:19:34 -07002262 unsigned long nr_reclaimed = 0;
Wu Fengguangf8629632009-09-21 17:03:11 -07002263 struct zone_reclaim_stat *reclaim_stat;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002264
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002265 for_each_populated_zone(zone) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002266 enum lru_list l;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002267
David Rientjese815af92007-10-16 23:25:54 -07002268 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002269 continue;
2270
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002271 for_each_evictable_lru(l) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002272 enum zone_stat_item ls = NR_LRU_BASE + l;
2273 unsigned long lru_pages = zone_page_state(zone, ls);
2274
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002275 /* For pass = 0, we don't shrink the active list */
Johannes Weiner0cb57252009-02-14 02:04:10 +01002276 if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2277 l == LRU_ACTIVE_FILE))
Christoph Lameterb69408e2008-10-18 20:26:14 -07002278 continue;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002279
Wu Fengguangf8629632009-09-21 17:03:11 -07002280 reclaim_stat = get_reclaim_stat(zone, sc);
2281 reclaim_stat->nr_saved_scan[l] +=
2282 (lru_pages >> prio) + 1;
2283 if (reclaim_stat->nr_saved_scan[l]
2284 >= nr_pages || pass > 3) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002285 unsigned long nr_to_scan;
2286
Wu Fengguangf8629632009-09-21 17:03:11 -07002287 reclaim_stat->nr_saved_scan[l] = 0;
Johannes Weiner0cb57252009-02-14 02:04:10 +01002288 nr_to_scan = min(nr_pages, lru_pages);
MinChan Kimd9796772009-03-31 15:19:34 -07002289 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
Christoph Lameterb69408e2008-10-18 20:26:14 -07002290 sc, prio);
MinChan Kimd9796772009-03-31 15:19:34 -07002291 if (nr_reclaimed >= nr_pages) {
Rafael J. Wysockia21e2552009-04-18 17:23:41 +02002292 sc->nr_reclaimed += nr_reclaimed;
MinChan Kimd9796772009-03-31 15:19:34 -07002293 return;
2294 }
Christoph Lameterb69408e2008-10-18 20:26:14 -07002295 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002296 }
2297 }
Rafael J. Wysockia21e2552009-04-18 17:23:41 +02002298 sc->nr_reclaimed += nr_reclaimed;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002299}
2300
2301/*
2302 * Try to free `nr_pages' of memory, system-wide, and return the number of
2303 * freed pages.
2304 *
2305 * Rather than trying to age LRUs the aim is to preserve the overall
2306 * LRU order by reclaiming preferentially
2307 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002308 */
Andrew Morton69e05942006-03-22 00:08:19 -08002309unsigned long shrink_all_memory(unsigned long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002311 unsigned long lru_pages, nr_slab;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002312 int pass;
2313 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002314 struct scan_control sc = {
2315 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002316 .may_unmap = 0,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002317 .may_writepage = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002318 .isolate_pages = isolate_pages_global,
Rafael J. Wysockia21e2552009-04-18 17:23:41 +02002319 .nr_reclaimed = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002320 };
2321
2322 current->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002323
Wu Fengguangadea02a2009-09-21 17:01:42 -07002324 lru_pages = global_reclaimable_pages();
Christoph Lameter972d1a72006-09-25 23:31:51 -07002325 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002326 /* If slab caches are huge, it's better to hit them first */
2327 while (nr_slab >= lru_pages) {
2328 reclaim_state.reclaimed_slab = 0;
2329 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2330 if (!reclaim_state.reclaimed_slab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331 break;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002332
MinChan Kimd9796772009-03-31 15:19:34 -07002333 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2334 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002335 goto out;
2336
2337 nr_slab -= reclaim_state.reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002339
2340 /*
2341 * We try to shrink LRUs in 5 passes:
2342 * 0 = Reclaim from inactive_list only
2343 * 1 = Reclaim from active list but don't reclaim mapped
2344 * 2 = 2nd pass of type 1
2345 * 3 = Reclaim mapped (normal reclaim)
2346 * 4 = 2nd pass of type 3
2347 */
2348 for (pass = 0; pass < 5; pass++) {
2349 int prio;
2350
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002351 /* Force reclaiming mapped pages in the passes #3 and #4 */
Johannes Weiner30491032009-02-14 02:03:08 +01002352 if (pass > 2)
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002353 sc.may_unmap = 1;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002354
2355 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
MinChan Kimd9796772009-03-31 15:19:34 -07002356 unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002357
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002358 sc.nr_scanned = 0;
Johannes Weiner9786bf82009-03-31 15:19:35 -07002359 sc.swap_cluster_max = nr_to_scan;
MinChan Kimd9796772009-03-31 15:19:34 -07002360 shrink_all_zones(nr_to_scan, prio, pass, &sc);
2361 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002362 goto out;
2363
2364 reclaim_state.reclaimed_slab = 0;
Andrew Morton76395d32007-01-05 16:37:05 -08002365 shrink_slab(sc.nr_scanned, sc.gfp_mask,
Wu Fengguangadea02a2009-09-21 17:01:42 -07002366 global_reclaimable_pages());
MinChan Kimd9796772009-03-31 15:19:34 -07002367 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2368 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002369 goto out;
2370
2371 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02002372 congestion_wait(BLK_RW_ASYNC, HZ / 10);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002373 }
Rafael J. Wysocki248a0302006-03-22 00:09:04 -08002374 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002375
2376 /*
MinChan Kimd9796772009-03-31 15:19:34 -07002377 * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2378 * something in slab caches
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002379 */
MinChan Kimd9796772009-03-31 15:19:34 -07002380 if (!sc.nr_reclaimed) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002381 do {
2382 reclaim_state.reclaimed_slab = 0;
Wu Fengguangadea02a2009-09-21 17:01:42 -07002383 shrink_slab(nr_pages, sc.gfp_mask,
2384 global_reclaimable_pages());
MinChan Kimd9796772009-03-31 15:19:34 -07002385 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2386 } while (sc.nr_reclaimed < nr_pages &&
2387 reclaim_state.reclaimed_slab > 0);
Andrew Morton76395d32007-01-05 16:37:05 -08002388 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002389
MinChan Kimd9796772009-03-31 15:19:34 -07002390
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002391out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002392 current->reclaim_state = NULL;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002393
MinChan Kimd9796772009-03-31 15:19:34 -07002394 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002396#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398/* It's optimal to keep kswapds on the same CPUs as their memory, but
2399 not required for correctness. So if the last cpu in a node goes
2400 away, we get changed to run anywhere: as the first one comes back,
2401 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002402static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002403 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002405 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002407 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002408 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002409 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302410 const struct cpumask *mask;
2411
2412 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002413
Rusty Russell3e597942009-01-01 10:12:24 +10302414 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002416 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 }
2418 }
2419 return NOTIFY_OK;
2420}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002421
Yasunori Goto3218ae12006-06-27 02:53:33 -07002422/*
2423 * This kswapd start function will be called by init and node-hot-add.
2424 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2425 */
2426int kswapd_run(int nid)
2427{
2428 pg_data_t *pgdat = NODE_DATA(nid);
2429 int ret = 0;
2430
2431 if (pgdat->kswapd)
2432 return 0;
2433
2434 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2435 if (IS_ERR(pgdat->kswapd)) {
2436 /* failure at boot is fatal */
2437 BUG_ON(system_state == SYSTEM_BOOTING);
2438 printk("Failed to start kswapd on node %d\n",nid);
2439 ret = -1;
2440 }
2441 return ret;
2442}
2443
Linus Torvalds1da177e2005-04-16 15:20:36 -07002444static int __init kswapd_init(void)
2445{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002446 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002447
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002449 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002450 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451 hotcpu_notifier(cpu_callback, 0);
2452 return 0;
2453}
2454
2455module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002456
2457#ifdef CONFIG_NUMA
2458/*
2459 * Zone reclaim mode
2460 *
2461 * If non-zero call zone_reclaim when the number of free pages falls below
2462 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002463 */
2464int zone_reclaim_mode __read_mostly;
2465
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002466#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002467#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002468#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2469#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2470
Christoph Lameter9eeff232006-01-18 17:42:31 -08002471/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002472 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2473 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2474 * a zone.
2475 */
2476#define ZONE_RECLAIM_PRIORITY 4
2477
Christoph Lameter9eeff232006-01-18 17:42:31 -08002478/*
Christoph Lameter96146342006-07-03 00:24:13 -07002479 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2480 * occur.
2481 */
2482int sysctl_min_unmapped_ratio = 1;
2483
2484/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002485 * If the number of slab pages in a zone grows beyond this percentage then
2486 * slab reclaim needs to occur.
2487 */
2488int sysctl_min_slab_ratio = 5;
2489
Mel Gorman90afa5d2009-06-16 15:33:20 -07002490static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2491{
2492 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2493 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2494 zone_page_state(zone, NR_ACTIVE_FILE);
2495
2496 /*
2497 * It's possible for there to be more file mapped pages than
2498 * accounted for by the pages on the file LRU lists because
2499 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2500 */
2501 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2502}
2503
2504/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2505static long zone_pagecache_reclaimable(struct zone *zone)
2506{
2507 long nr_pagecache_reclaimable;
2508 long delta = 0;
2509
2510 /*
2511 * If RECLAIM_SWAP is set, then all file pages are considered
2512 * potentially reclaimable. Otherwise, we have to worry about
2513 * pages like swapcache and zone_unmapped_file_pages() provides
2514 * a better estimate
2515 */
2516 if (zone_reclaim_mode & RECLAIM_SWAP)
2517 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2518 else
2519 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2520
2521 /* If we can't clean pages, remove dirty pages from consideration */
2522 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2523 delta += zone_page_state(zone, NR_FILE_DIRTY);
2524
2525 /* Watch for any possible underflows due to delta */
2526 if (unlikely(delta > nr_pagecache_reclaimable))
2527 delta = nr_pagecache_reclaimable;
2528
2529 return nr_pagecache_reclaimable - delta;
2530}
2531
Christoph Lameter0ff38492006-09-25 23:31:52 -07002532/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002533 * Try to free up some pages from this zone through reclaim.
2534 */
Andrew Morton179e9632006-03-22 00:08:18 -08002535static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002536{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002537 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002538 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002539 struct task_struct *p = current;
2540 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002541 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002542 struct scan_control sc = {
2543 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002544 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002545 .may_swap = 1,
Andrew Morton69e05942006-03-22 00:08:19 -08002546 .swap_cluster_max = max_t(unsigned long, nr_pages,
2547 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002548 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002549 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002550 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002551 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08002552 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07002553 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002554
2555 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08002556 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002557 /*
2558 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2559 * and we also need to be able to write out pages for RECLAIM_WRITE
2560 * and RECLAIM_SWAP.
2561 */
2562 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002563 reclaim_state.reclaimed_slab = 0;
2564 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002565
Mel Gorman90afa5d2009-06-16 15:33:20 -07002566 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07002567 /*
2568 * Free memory by calling shrink zone with increasing
2569 * priorities until we have enough memory freed.
2570 */
2571 priority = ZONE_RECLAIM_PRIORITY;
2572 do {
Martin Bligh3bb1a8522006-10-28 10:38:24 -07002573 note_zone_scanning_priority(zone, priority);
Rik van Riela79311c2009-01-06 14:40:01 -08002574 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002575 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002576 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002577 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002578
Christoph Lameter83e33a42006-09-25 23:31:53 -07002579 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2580 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002581 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002582 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002583 * many pages were freed in this zone. So we take the current
2584 * number of slab pages and shake the slab until it is reduced
2585 * by the same nr_pages that we used for reclaiming unmapped
2586 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002587 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002588 * Note that shrink_slab will free memory on all zones and may
2589 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002590 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07002591 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07002592 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2593 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07002594 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07002595
2596 /*
2597 * Update nr_reclaimed by the number of slab pages we
2598 * reclaimed from this zone.
2599 */
Rik van Riela79311c2009-01-06 14:40:01 -08002600 sc.nr_reclaimed += slab_reclaimable -
Christoph Lameter83e33a42006-09-25 23:31:53 -07002601 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002602 }
2603
Christoph Lameter9eeff232006-01-18 17:42:31 -08002604 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002605 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
Rik van Riela79311c2009-01-06 14:40:01 -08002606 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002607}
Andrew Morton179e9632006-03-22 00:08:18 -08002608
2609int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2610{
Andrew Morton179e9632006-03-22 00:08:18 -08002611 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002612 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002613
2614 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002615 * Zone reclaim reclaims unmapped file backed pages and
2616 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002617 *
Christoph Lameter96146342006-07-03 00:24:13 -07002618 * A small portion of unmapped file backed pages is needed for
2619 * file I/O otherwise pages read by file I/O will be immediately
2620 * thrown out if the zone is overallocated. So we do not reclaim
2621 * if less than a specified percentage of the zone is used by
2622 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002623 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07002624 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2625 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002626 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08002627
David Rientjesd773ed62007-10-16 23:26:01 -07002628 if (zone_is_all_unreclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002629 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07002630
Andrew Morton179e9632006-03-22 00:08:18 -08002631 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002632 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002633 */
David Rientjesd773ed62007-10-16 23:26:01 -07002634 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002635 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08002636
2637 /*
2638 * Only run zone reclaim on the local zone or on zones that do not
2639 * have associated processors. This will favor the local processor
2640 * over remote processors and spread off node memory allocations
2641 * as wide as possible.
2642 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002643 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002644 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07002645 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07002646
2647 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002648 return ZONE_RECLAIM_NOSCAN;
2649
David Rientjesd773ed62007-10-16 23:26:01 -07002650 ret = __zone_reclaim(zone, gfp_mask, order);
2651 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2652
Mel Gorman24cf725182009-06-16 15:33:23 -07002653 if (!ret)
2654 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2655
David Rientjesd773ed62007-10-16 23:26:01 -07002656 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002657}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002658#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002659
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002660/*
2661 * page_evictable - test whether a page is evictable
2662 * @page: the page to test
2663 * @vma: the VMA in which the page is or will be mapped, may be NULL
2664 *
2665 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07002666 * lists vs unevictable list. The vma argument is !NULL when called from the
2667 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002668 *
2669 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002670 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07002671 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002672 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002673 */
2674int page_evictable(struct page *page, struct vm_area_struct *vma)
2675{
2676
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002677 if (mapping_unevictable(page_mapping(page)))
2678 return 0;
2679
Nick Pigginb291f002008-10-18 20:26:44 -07002680 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2681 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002682
2683 return 1;
2684}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002685
2686/**
2687 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2688 * @page: page to check evictability and move to appropriate lru list
2689 * @zone: zone page is in
2690 *
2691 * Checks a page for evictability and moves the page to the appropriate
2692 * zone lru list.
2693 *
2694 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2695 * have PageUnevictable set.
2696 */
2697static void check_move_unevictable_page(struct page *page, struct zone *zone)
2698{
2699 VM_BUG_ON(PageActive(page));
2700
2701retry:
2702 ClearPageUnevictable(page);
2703 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07002704 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002705
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002706 __dec_zone_state(zone, NR_UNEVICTABLE);
2707 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002708 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002709 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2710 __count_vm_event(UNEVICTABLE_PGRESCUED);
2711 } else {
2712 /*
2713 * rotate unevictable list
2714 */
2715 SetPageUnevictable(page);
2716 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002717 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002718 if (page_evictable(page, NULL))
2719 goto retry;
2720 }
2721}
2722
2723/**
2724 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2725 * @mapping: struct address_space to scan for evictable pages
2726 *
2727 * Scan all pages in mapping. Check unevictable pages for
2728 * evictability and move them to the appropriate zone lru list.
2729 */
2730void scan_mapping_unevictable_pages(struct address_space *mapping)
2731{
2732 pgoff_t next = 0;
2733 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2734 PAGE_CACHE_SHIFT;
2735 struct zone *zone;
2736 struct pagevec pvec;
2737
2738 if (mapping->nrpages == 0)
2739 return;
2740
2741 pagevec_init(&pvec, 0);
2742 while (next < end &&
2743 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2744 int i;
2745 int pg_scanned = 0;
2746
2747 zone = NULL;
2748
2749 for (i = 0; i < pagevec_count(&pvec); i++) {
2750 struct page *page = pvec.pages[i];
2751 pgoff_t page_index = page->index;
2752 struct zone *pagezone = page_zone(page);
2753
2754 pg_scanned++;
2755 if (page_index > next)
2756 next = page_index;
2757 next++;
2758
2759 if (pagezone != zone) {
2760 if (zone)
2761 spin_unlock_irq(&zone->lru_lock);
2762 zone = pagezone;
2763 spin_lock_irq(&zone->lru_lock);
2764 }
2765
2766 if (PageLRU(page) && PageUnevictable(page))
2767 check_move_unevictable_page(page, zone);
2768 }
2769 if (zone)
2770 spin_unlock_irq(&zone->lru_lock);
2771 pagevec_release(&pvec);
2772
2773 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2774 }
2775
2776}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002777
2778/**
2779 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2780 * @zone - zone of which to scan the unevictable list
2781 *
2782 * Scan @zone's unevictable LRU lists to check for pages that have become
2783 * evictable. Move those that have to @zone's inactive list where they
2784 * become candidates for reclaim, unless shrink_inactive_zone() decides
2785 * to reactivate them. Pages that are still unevictable are rotated
2786 * back onto @zone's unevictable list.
2787 */
2788#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08002789static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002790{
2791 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2792 unsigned long scan;
2793 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2794
2795 while (nr_to_scan > 0) {
2796 unsigned long batch_size = min(nr_to_scan,
2797 SCAN_UNEVICTABLE_BATCH_SIZE);
2798
2799 spin_lock_irq(&zone->lru_lock);
2800 for (scan = 0; scan < batch_size; scan++) {
2801 struct page *page = lru_to_page(l_unevictable);
2802
2803 if (!trylock_page(page))
2804 continue;
2805
2806 prefetchw_prev_lru_page(page, l_unevictable, flags);
2807
2808 if (likely(PageLRU(page) && PageUnevictable(page)))
2809 check_move_unevictable_page(page, zone);
2810
2811 unlock_page(page);
2812 }
2813 spin_unlock_irq(&zone->lru_lock);
2814
2815 nr_to_scan -= batch_size;
2816 }
2817}
2818
2819
2820/**
2821 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2822 *
2823 * A really big hammer: scan all zones' unevictable LRU lists to check for
2824 * pages that have become evictable. Move those back to the zones'
2825 * inactive list where they become candidates for reclaim.
2826 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2827 * and we add swap to the system. As such, it runs in the context of a task
2828 * that has possibly/probably made some previously unevictable pages
2829 * evictable.
2830 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08002831static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002832{
2833 struct zone *zone;
2834
2835 for_each_zone(zone) {
2836 scan_zone_unevictable_pages(zone);
2837 }
2838}
2839
2840/*
2841 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2842 * all nodes' unevictable lists for evictable pages
2843 */
2844unsigned long scan_unevictable_pages;
2845
2846int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002847 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002848 size_t *length, loff_t *ppos)
2849{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002850 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002851
2852 if (write && *(unsigned long *)table->data)
2853 scan_all_zones_unevictable_pages();
2854
2855 scan_unevictable_pages = 0;
2856 return 0;
2857}
2858
2859/*
2860 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2861 * a specified node's per zone unevictable lists for evictable pages.
2862 */
2863
2864static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2865 struct sysdev_attribute *attr,
2866 char *buf)
2867{
2868 return sprintf(buf, "0\n"); /* always zero; should fit... */
2869}
2870
2871static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2872 struct sysdev_attribute *attr,
2873 const char *buf, size_t count)
2874{
2875 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2876 struct zone *zone;
2877 unsigned long res;
2878 unsigned long req = strict_strtoul(buf, 10, &res);
2879
2880 if (!req)
2881 return 1; /* zero is no-op */
2882
2883 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2884 if (!populated_zone(zone))
2885 continue;
2886 scan_zone_unevictable_pages(zone);
2887 }
2888 return 1;
2889}
2890
2891
2892static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2893 read_scan_unevictable_node,
2894 write_scan_unevictable_node);
2895
2896int scan_unevictable_register_node(struct node *node)
2897{
2898 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2899}
2900
2901void scan_unevictable_unregister_node(struct node *node)
2902{
2903 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2904}
2905