blob: c26986c85ce01b1d15a8914065a829890054c210 [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
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080058 /* How many pages shrink_list() should reclaim */
59 unsigned long nr_to_reclaim;
60
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080061 unsigned long hibernation_mode;
62
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040064 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070065
66 int may_writepage;
67
Johannes Weinera6dc60f2009-03-31 15:19:30 -070068 /* Can mapped pages be reclaimed? */
69 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080070
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070071 /* Can pages be swapped as part of reclaim? */
72 int may_swap;
73
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070074 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070075
76 int all_unreclaimable;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070077
78 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080079
80 /* Which cgroup do we reclaim from */
81 struct mem_cgroup *mem_cgroup;
82
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -070083 /*
84 * Nodemask of nodes allowed by the caller. If NULL, all nodes
85 * are scanned.
86 */
87 nodemask_t *nodemask;
88
Balbir Singh66e17072008-02-07 00:13:56 -080089 /* Pluggable isolate pages callback */
90 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
91 unsigned long *scanned, int order, int mode,
92 struct zone *z, struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -070093 int active, int file);
Linus Torvalds1da177e2005-04-16 15:20:36 -070094};
95
Linus Torvalds1da177e2005-04-16 15:20:36 -070096#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
97
98#ifdef ARCH_HAS_PREFETCH
99#define prefetch_prev_lru_page(_page, _base, _field) \
100 do { \
101 if ((_page)->lru.prev != _base) { \
102 struct page *prev; \
103 \
104 prev = lru_to_page(&(_page->lru)); \
105 prefetch(&prev->_field); \
106 } \
107 } while (0)
108#else
109#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
110#endif
111
112#ifdef ARCH_HAS_PREFETCHW
113#define prefetchw_prev_lru_page(_page, _base, _field) \
114 do { \
115 if ((_page)->lru.prev != _base) { \
116 struct page *prev; \
117 \
118 prev = lru_to_page(&(_page->lru)); \
119 prefetchw(&prev->_field); \
120 } \
121 } while (0)
122#else
123#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
124#endif
125
126/*
127 * From 0 .. 100. Higher means more swappy.
128 */
129int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700130long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700131
132static LIST_HEAD(shrinker_list);
133static DECLARE_RWSEM(shrinker_rwsem);
134
Balbir Singh00f0b822008-03-04 14:28:39 -0800135#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800136#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800137#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800138#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800139#endif
140
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800141static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
142 struct scan_control *sc)
143{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800144 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800145 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
146
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800147 return &zone->reclaim_stat;
148}
149
Vincent Li0b217672009-09-21 17:03:09 -0700150static unsigned long zone_nr_lru_pages(struct zone *zone,
151 struct scan_control *sc, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800152{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800153 if (!scanning_global_lru(sc))
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800154 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
155
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800156 return zone_page_state(zone, NR_LRU_BASE + lru);
157}
158
159
Linus Torvalds1da177e2005-04-16 15:20:36 -0700160/*
161 * Add a shrinker callback to be called from the vm
162 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700163void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700164{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700165 shrinker->nr = 0;
166 down_write(&shrinker_rwsem);
167 list_add_tail(&shrinker->list, &shrinker_list);
168 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700170EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171
172/*
173 * Remove one
174 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700175void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176{
177 down_write(&shrinker_rwsem);
178 list_del(&shrinker->list);
179 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700181EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182
183#define SHRINK_BATCH 128
184/*
185 * Call the shrink functions to age shrinkable caches
186 *
187 * Here we assume it costs one seek to replace a lru page and that it also
188 * takes a seek to recreate a cache object. With this in mind we age equal
189 * percentages of the lru and ageable caches. This should balance the seeks
190 * generated by these structures.
191 *
Simon Arlott183ff222007-10-20 01:27:18 +0200192 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700193 * slab to avoid swapping.
194 *
195 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
196 *
197 * `lru_pages' represents the number of on-LRU pages in all the zones which
198 * are eligible for the caller's allocation attempt. It is used for balancing
199 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700200 *
201 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700202 */
Andrew Morton69e05942006-03-22 00:08:19 -0800203unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
204 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205{
206 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800207 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700208
209 if (scanned == 0)
210 scanned = SWAP_CLUSTER_MAX;
211
212 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700213 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214
215 list_for_each_entry(shrinker, &shrinker_list, list) {
216 unsigned long long delta;
217 unsigned long total_scan;
Rusty Russell8e1f9362007-07-17 04:03:17 -0700218 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219
220 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800221 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222 do_div(delta, lru_pages + 1);
223 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800224 if (shrinker->nr < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700225 printk(KERN_ERR "shrink_slab: %pF negative objects to "
226 "delete nr=%ld\n",
227 shrinker->shrink, shrinker->nr);
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800228 shrinker->nr = max_pass;
229 }
230
231 /*
232 * Avoid risking looping forever due to too large nr value:
233 * never try to free more than twice the estimate number of
234 * freeable entries.
235 */
236 if (shrinker->nr > max_pass * 2)
237 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700238
239 total_scan = shrinker->nr;
240 shrinker->nr = 0;
241
242 while (total_scan >= SHRINK_BATCH) {
243 long this_scan = SHRINK_BATCH;
244 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700245 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246
Rusty Russell8e1f9362007-07-17 04:03:17 -0700247 nr_before = (*shrinker->shrink)(0, gfp_mask);
248 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 if (shrink_ret == -1)
250 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700251 if (shrink_ret < nr_before)
252 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700253 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 total_scan -= this_scan;
255
256 cond_resched();
257 }
258
259 shrinker->nr += total_scan;
260 }
261 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700262 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263}
264
265/* Called without lock on whether page is mapped, so answer is unstable */
266static inline int page_mapping_inuse(struct page *page)
267{
268 struct address_space *mapping;
269
270 /* Page is in somebody's page tables. */
271 if (page_mapped(page))
272 return 1;
273
274 /* Be more reluctant to reclaim swapcache than pagecache */
275 if (PageSwapCache(page))
276 return 1;
277
278 mapping = page_mapping(page);
279 if (!mapping)
280 return 0;
281
282 /* File is mmap'd by somebody? */
283 return mapping_mapped(mapping);
284}
285
286static inline int is_page_cache_freeable(struct page *page)
287{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700288 /*
289 * A freeable page cache page is referenced only by the caller
290 * that isolated the page, the page cache radix tree and
291 * optional buffer heads at page->private.
292 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700293 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700294}
295
296static int may_write_to_queue(struct backing_dev_info *bdi)
297{
Christoph Lameter930d9152006-01-08 01:00:47 -0800298 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700299 return 1;
300 if (!bdi_write_congested(bdi))
301 return 1;
302 if (bdi == current->backing_dev_info)
303 return 1;
304 return 0;
305}
306
307/*
308 * We detected a synchronous write error writing a page out. Probably
309 * -ENOSPC. We need to propagate that into the address_space for a subsequent
310 * fsync(), msync() or close().
311 *
312 * The tricky part is that after writepage we cannot touch the mapping: nothing
313 * prevents it from being freed up. But we have a ref on the page and once
314 * that page is locked, the mapping is pinned.
315 *
316 * We're allowed to run sleeping lock_page() here because we know the caller has
317 * __GFP_FS.
318 */
319static void handle_write_error(struct address_space *mapping,
320 struct page *page, int error)
321{
322 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700323 if (page_mapping(page) == mapping)
324 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325 unlock_page(page);
326}
327
Andy Whitcroftc661b072007-08-22 14:01:26 -0700328/* Request for sync pageout. */
329enum pageout_io {
330 PAGEOUT_IO_ASYNC,
331 PAGEOUT_IO_SYNC,
332};
333
Christoph Lameter04e62a22006-06-23 02:03:38 -0700334/* possible outcome of pageout() */
335typedef enum {
336 /* failed to write page out, page is locked */
337 PAGE_KEEP,
338 /* move page to the active list, page is locked */
339 PAGE_ACTIVATE,
340 /* page has been sent to the disk successfully, page is unlocked */
341 PAGE_SUCCESS,
342 /* page is clean and locked */
343 PAGE_CLEAN,
344} pageout_t;
345
Linus Torvalds1da177e2005-04-16 15:20:36 -0700346/*
Andrew Morton1742f192006-03-22 00:08:21 -0800347 * pageout is called by shrink_page_list() for each dirty page.
348 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700350static pageout_t pageout(struct page *page, struct address_space *mapping,
351 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700352{
353 /*
354 * If the page is dirty, only perform writeback if that write
355 * will be non-blocking. To prevent this allocation from being
356 * stalled by pagecache activity. But note that there may be
357 * stalls if we need to run get_block(). We could test
358 * PagePrivate for that.
359 *
Vincent Li6aceb532009-12-14 17:58:49 -0800360 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 * this page's queue, we can perform writeback even if that
362 * will block.
363 *
364 * If the page is swapcache, write it back even if that would
365 * block, for some throttling. This happens by accident, because
366 * swap_backing_dev_info is bust: it doesn't reflect the
367 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700368 */
369 if (!is_page_cache_freeable(page))
370 return PAGE_KEEP;
371 if (!mapping) {
372 /*
373 * Some data journaling orphaned pages can have
374 * page->mapping == NULL while being dirty with clean buffers.
375 */
David Howells266cf652009-04-03 16:42:36 +0100376 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700377 if (try_to_free_buffers(page)) {
378 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700379 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700380 return PAGE_CLEAN;
381 }
382 }
383 return PAGE_KEEP;
384 }
385 if (mapping->a_ops->writepage == NULL)
386 return PAGE_ACTIVATE;
387 if (!may_write_to_queue(mapping->backing_dev_info))
388 return PAGE_KEEP;
389
390 if (clear_page_dirty_for_io(page)) {
391 int res;
392 struct writeback_control wbc = {
393 .sync_mode = WB_SYNC_NONE,
394 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700395 .range_start = 0,
396 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 .nonblocking = 1,
398 .for_reclaim = 1,
399 };
400
401 SetPageReclaim(page);
402 res = mapping->a_ops->writepage(page, &wbc);
403 if (res < 0)
404 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800405 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 ClearPageReclaim(page);
407 return PAGE_ACTIVATE;
408 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700409
410 /*
411 * Wait on writeback if requested to. This happens when
412 * direct reclaiming a large contiguous area and the
413 * first attempt to free a range of pages fails.
414 */
415 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
416 wait_on_page_writeback(page);
417
Linus Torvalds1da177e2005-04-16 15:20:36 -0700418 if (!PageWriteback(page)) {
419 /* synchronous write or broken a_ops? */
420 ClearPageReclaim(page);
421 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700422 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700423 return PAGE_SUCCESS;
424 }
425
426 return PAGE_CLEAN;
427}
428
Andrew Mortona649fd92006-10-17 00:09:36 -0700429/*
Nick Piggine2867812008-07-25 19:45:30 -0700430 * Same as remove_mapping, but if the page is removed from the mapping, it
431 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700432 */
Nick Piggine2867812008-07-25 19:45:30 -0700433static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800434{
Nick Piggin28e4d962006-09-25 23:31:23 -0700435 BUG_ON(!PageLocked(page));
436 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800437
Nick Piggin19fd6232008-07-25 19:45:32 -0700438 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800439 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700440 * The non racy check for a busy page.
441 *
442 * Must be careful with the order of the tests. When someone has
443 * a ref to the page, it may be possible that they dirty it then
444 * drop the reference. So if PageDirty is tested before page_count
445 * here, then the following race may occur:
446 *
447 * get_user_pages(&page);
448 * [user mapping goes away]
449 * write_to(page);
450 * !PageDirty(page) [good]
451 * SetPageDirty(page);
452 * put_page(page);
453 * !page_count(page) [good, discard it]
454 *
455 * [oops, our write_to data is lost]
456 *
457 * Reversing the order of the tests ensures such a situation cannot
458 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
459 * load is not satisfied before that of page->_count.
460 *
461 * Note that if SetPageDirty is always performed via set_page_dirty,
462 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800463 */
Nick Piggine2867812008-07-25 19:45:30 -0700464 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800465 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700466 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
467 if (unlikely(PageDirty(page))) {
468 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800469 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700470 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800471
472 if (PageSwapCache(page)) {
473 swp_entry_t swap = { .val = page_private(page) };
474 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700475 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700476 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700477 } else {
478 __remove_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700479 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700480 mem_cgroup_uncharge_cache_page(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800481 }
482
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800483 return 1;
484
485cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700486 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800487 return 0;
488}
489
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490/*
Nick Piggine2867812008-07-25 19:45:30 -0700491 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
492 * someone else has a ref on the page, abort and return 0. If it was
493 * successfully detached, return 1. Assumes the caller has a single ref on
494 * this page.
495 */
496int remove_mapping(struct address_space *mapping, struct page *page)
497{
498 if (__remove_mapping(mapping, page)) {
499 /*
500 * Unfreezing the refcount with 1 rather than 2 effectively
501 * drops the pagecache ref for us without requiring another
502 * atomic operation.
503 */
504 page_unfreeze_refs(page, 1);
505 return 1;
506 }
507 return 0;
508}
509
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700510/**
511 * putback_lru_page - put previously isolated page onto appropriate LRU list
512 * @page: page to be put back to appropriate lru list
513 *
514 * Add previously isolated @page to appropriate LRU list.
515 * Page may still be unevictable for other reasons.
516 *
517 * lru_lock must not be held, interrupts must be enabled.
518 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700519void putback_lru_page(struct page *page)
520{
521 int lru;
522 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700523 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700524
525 VM_BUG_ON(PageLRU(page));
526
527redo:
528 ClearPageUnevictable(page);
529
530 if (page_evictable(page, NULL)) {
531 /*
532 * For evictable pages, we can use the cache.
533 * In event of a race, worst case is we end up with an
534 * unevictable page on [in]active list.
535 * We know how to handle that.
536 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700537 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700538 lru_cache_add_lru(page, lru);
539 } else {
540 /*
541 * Put unevictable pages directly on zone's unevictable
542 * list.
543 */
544 lru = LRU_UNEVICTABLE;
545 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700546 /*
547 * When racing with an mlock clearing (page is
548 * unlocked), make sure that if the other thread does
549 * not observe our setting of PG_lru and fails
550 * isolation, we see PG_mlocked cleared below and move
551 * the page back to the evictable list.
552 *
553 * The other side is TestClearPageMlocked().
554 */
555 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700556 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700557
558 /*
559 * page's status can change while we move it among lru. If an evictable
560 * page is on unevictable list, it never be freed. To avoid that,
561 * check after we added it to the list, again.
562 */
563 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
564 if (!isolate_lru_page(page)) {
565 put_page(page);
566 goto redo;
567 }
568 /* This means someone else dropped this page from LRU
569 * So, it will be freed or putback to LRU again. There is
570 * nothing to do here.
571 */
572 }
573
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700574 if (was_unevictable && lru != LRU_UNEVICTABLE)
575 count_vm_event(UNEVICTABLE_PGRESCUED);
576 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
577 count_vm_event(UNEVICTABLE_PGCULLED);
578
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700579 put_page(page); /* drop ref from isolate */
580}
581
Nick Piggine2867812008-07-25 19:45:30 -0700582/*
Andrew Morton1742f192006-03-22 00:08:21 -0800583 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584 */
Andrew Morton1742f192006-03-22 00:08:21 -0800585static unsigned long shrink_page_list(struct list_head *page_list,
Andy Whitcroftc661b072007-08-22 14:01:26 -0700586 struct scan_control *sc,
587 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588{
589 LIST_HEAD(ret_pages);
590 struct pagevec freed_pvec;
591 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800592 unsigned long nr_reclaimed = 0;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700593 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594
595 cond_resched();
596
597 pagevec_init(&freed_pvec, 1);
598 while (!list_empty(page_list)) {
599 struct address_space *mapping;
600 struct page *page;
601 int may_enter_fs;
602 int referenced;
603
604 cond_resched();
605
606 page = lru_to_page(page_list);
607 list_del(&page->lru);
608
Nick Piggin529ae9a2008-08-02 12:01:03 +0200609 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 goto keep;
611
Nick Piggin725d7042006-09-25 23:30:55 -0700612 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613
614 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800615
Nick Pigginb291f002008-10-18 20:26:44 -0700616 if (unlikely(!page_evictable(page, NULL)))
617 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700618
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700619 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800620 goto keep_locked;
621
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 /* Double the slab pressure for mapped and swapcache pages */
623 if (page_mapped(page) || PageSwapCache(page))
624 sc->nr_scanned++;
625
Andy Whitcroftc661b072007-08-22 14:01:26 -0700626 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
627 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
628
629 if (PageWriteback(page)) {
630 /*
631 * Synchronous reclaim is performed in two passes,
632 * first an asynchronous pass over the list to
633 * start parallel writeback, and a second synchronous
634 * pass to wait for the IO to complete. Wait here
635 * for any page for which writeback has already
636 * started.
637 */
638 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
639 wait_on_page_writeback(page);
Andrew Morton4dd4b922008-03-24 12:29:52 -0700640 else
Andy Whitcroftc661b072007-08-22 14:01:26 -0700641 goto keep_locked;
642 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700643
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -0700644 referenced = page_referenced(page, 1,
645 sc->mem_cgroup, &vm_flags);
Minchan Kim03ef83a2009-08-26 14:29:23 -0700646 /*
647 * In active use or really unfreeable? Activate it.
648 * If page which have PG_mlocked lost isoltation race,
649 * try_to_unmap moves it to unevictable list
650 */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700651 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
Minchan Kim03ef83a2009-08-26 14:29:23 -0700652 referenced && page_mapping_inuse(page)
653 && !(vm_flags & VM_LOCKED))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700654 goto activate_locked;
655
Linus Torvalds1da177e2005-04-16 15:20:36 -0700656 /*
657 * Anonymous process memory has backing store?
658 * Try to allocate it some swap space here.
659 */
Nick Pigginb291f002008-10-18 20:26:44 -0700660 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800661 if (!(sc->gfp_mask & __GFP_IO))
662 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800663 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700664 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800665 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700666 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667
668 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700669
670 /*
671 * The page is mapped into the page tables of one or more
672 * processes. Try to unmap it here.
673 */
674 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200675 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 case SWAP_FAIL:
677 goto activate_locked;
678 case SWAP_AGAIN:
679 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700680 case SWAP_MLOCK:
681 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 case SWAP_SUCCESS:
683 ; /* try to free the page below */
684 }
685 }
686
687 if (PageDirty(page)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700688 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700690 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800692 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700693 goto keep_locked;
694
695 /* Page is dirty, try to write it out here */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700696 switch (pageout(page, mapping, sync_writeback)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697 case PAGE_KEEP:
698 goto keep_locked;
699 case PAGE_ACTIVATE:
700 goto activate_locked;
701 case PAGE_SUCCESS:
Andrew Morton4dd4b922008-03-24 12:29:52 -0700702 if (PageWriteback(page) || PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 goto keep;
704 /*
705 * A synchronous write - probably a ramdisk. Go
706 * ahead and try to reclaim the page.
707 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200708 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700709 goto keep;
710 if (PageDirty(page) || PageWriteback(page))
711 goto keep_locked;
712 mapping = page_mapping(page);
713 case PAGE_CLEAN:
714 ; /* try to free the page below */
715 }
716 }
717
718 /*
719 * If the page has buffers, try to free the buffer mappings
720 * associated with this page. If we succeed we try to free
721 * the page as well.
722 *
723 * We do this even if the page is PageDirty().
724 * try_to_release_page() does not perform I/O, but it is
725 * possible for a page to have PageDirty set, but it is actually
726 * clean (all its buffers are clean). This happens if the
727 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700728 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 * try_to_release_page() will discover that cleanness and will
730 * drop the buffers and mark the page clean - it can be freed.
731 *
732 * Rarely, pages can have buffers and no ->mapping. These are
733 * the pages which were not successfully invalidated in
734 * truncate_complete_page(). We try to drop those buffers here
735 * and if that worked, and the page is no longer mapped into
736 * process address space (page_count == 1) it can be freed.
737 * Otherwise, leave the page on the LRU so it is swappable.
738 */
David Howells266cf652009-04-03 16:42:36 +0100739 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700740 if (!try_to_release_page(page, sc->gfp_mask))
741 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700742 if (!mapping && page_count(page) == 1) {
743 unlock_page(page);
744 if (put_page_testzero(page))
745 goto free_it;
746 else {
747 /*
748 * rare race with speculative reference.
749 * the speculative reference will free
750 * this page shortly, so we may
751 * increment nr_reclaimed here (and
752 * leave it off the LRU).
753 */
754 nr_reclaimed++;
755 continue;
756 }
757 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700758 }
759
Nick Piggine2867812008-07-25 19:45:30 -0700760 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800761 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762
Nick Piggina978d6f2008-10-18 20:26:58 -0700763 /*
764 * At this point, we have no other references and there is
765 * no way to pick any more up (removed from LRU, removed
766 * from pagecache). Can use non-atomic bitops now (and
767 * we obviously don't have to worry about waking up a process
768 * waiting on the page lock, because there are no references.
769 */
770 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700771free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800772 nr_reclaimed++;
Nick Piggine2867812008-07-25 19:45:30 -0700773 if (!pagevec_add(&freed_pvec, page)) {
774 __pagevec_free(&freed_pvec);
775 pagevec_reinit(&freed_pvec);
776 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777 continue;
778
Nick Pigginb291f002008-10-18 20:26:44 -0700779cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800780 if (PageSwapCache(page))
781 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700782 unlock_page(page);
783 putback_lru_page(page);
784 continue;
785
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700787 /* Not a candidate for swapping, so reclaim swap space. */
788 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800789 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700790 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700791 SetPageActive(page);
792 pgactivate++;
793keep_locked:
794 unlock_page(page);
795keep:
796 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700797 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 }
799 list_splice(&ret_pages, page_list);
800 if (pagevec_count(&freed_pvec))
Nick Piggine2867812008-07-25 19:45:30 -0700801 __pagevec_free(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700802 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800803 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700804}
805
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700806/* LRU Isolation modes. */
807#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
808#define ISOLATE_ACTIVE 1 /* Isolate active pages. */
809#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
810
811/*
812 * Attempt to remove the specified page from its LRU. Only take this page
813 * if it is of the appropriate PageActive status. Pages which are being
814 * freed elsewhere are also ignored.
815 *
816 * page: page to consider
817 * mode: one of the LRU isolation modes defined above
818 *
819 * returns 0 on success, -ve errno on failure.
820 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700821int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700822{
823 int ret = -EINVAL;
824
825 /* Only take pages on the LRU. */
826 if (!PageLRU(page))
827 return ret;
828
829 /*
830 * When checking the active state, we need to be sure we are
831 * dealing with comparible boolean values. Take the logical not
832 * of each.
833 */
834 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
835 return ret;
836
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700837 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700838 return ret;
839
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700840 /*
841 * When this function is being called for lumpy reclaim, we
842 * initially look into all LRU pages, active, inactive and
843 * unevictable; only give shrink_page_list evictable pages.
844 */
845 if (PageUnevictable(page))
846 return ret;
847
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700848 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800849
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700850 if (likely(get_page_unless_zero(page))) {
851 /*
852 * Be careful not to clear PageLRU until after we're
853 * sure the page is not being freed elsewhere -- the
854 * page release code relies on it.
855 */
856 ClearPageLRU(page);
857 ret = 0;
858 }
859
860 return ret;
861}
862
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800863/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700864 * zone->lru_lock is heavily contended. Some of the functions that
865 * shrink the lists perform better by taking out a batch of pages
866 * and working on them outside the LRU lock.
867 *
868 * For pagecache intensive workloads, this function is the hottest
869 * spot in the kernel (apart from copy_*_user functions).
870 *
871 * Appropriate locks must be held before calling this function.
872 *
873 * @nr_to_scan: The number of pages to look through on the list.
874 * @src: The LRU list to pull pages off.
875 * @dst: The temp list to put pages on to.
876 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700877 * @order: The caller's attempted allocation order
878 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700879 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700880 *
881 * returns how many pages were moved onto *@dst.
882 */
Andrew Morton69e05942006-03-22 00:08:19 -0800883static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
884 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700885 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700886{
Andrew Morton69e05942006-03-22 00:08:19 -0800887 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800888 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800890 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700891 struct page *page;
892 unsigned long pfn;
893 unsigned long end_pfn;
894 unsigned long page_pfn;
895 int zone_id;
896
Linus Torvalds1da177e2005-04-16 15:20:36 -0700897 page = lru_to_page(src);
898 prefetchw_prev_lru_page(page, src, flags);
899
Nick Piggin725d7042006-09-25 23:30:55 -0700900 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800901
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700902 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700903 case 0:
904 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700905 mem_cgroup_del_lru(page);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800906 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700907 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800908
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700909 case -EBUSY:
910 /* else it is being freed elsewhere */
911 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700912 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700913 continue;
914
915 default:
916 BUG();
917 }
918
919 if (!order)
920 continue;
921
922 /*
923 * Attempt to take all pages in the order aligned region
924 * surrounding the tag page. Only take those pages of
925 * the same active state as that tag page. We may safely
926 * round the target page pfn down to the requested order
927 * as the mem_map is guarenteed valid out to MAX_ORDER,
928 * where that page is in a different zone we will detect
929 * it from its zone id and abort this block scan.
930 */
931 zone_id = page_zone_id(page);
932 page_pfn = page_to_pfn(page);
933 pfn = page_pfn & ~((1 << order) - 1);
934 end_pfn = pfn + (1 << order);
935 for (; pfn < end_pfn; pfn++) {
936 struct page *cursor_page;
937
938 /* The target page is in the block, ignore it. */
939 if (unlikely(pfn == page_pfn))
940 continue;
941
942 /* Avoid holes within the zone. */
943 if (unlikely(!pfn_valid_within(pfn)))
944 break;
945
946 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700947
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700948 /* Check that we have not crossed a zone boundary. */
949 if (unlikely(page_zone_id(cursor_page) != zone_id))
950 continue;
Minchan Kimde2e7562009-09-21 17:01:43 -0700951
952 /*
953 * If we don't have enough swap space, reclaiming of
954 * anon page which don't already have a swap slot is
955 * pointless.
956 */
957 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
958 !PageSwapCache(cursor_page))
959 continue;
960
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -0700961 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700962 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +0900963 mem_cgroup_del_lru(cursor_page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700964 nr_taken++;
965 scan++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700966 }
967 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 }
969
970 *scanned = scan;
971 return nr_taken;
972}
973
Balbir Singh66e17072008-02-07 00:13:56 -0800974static unsigned long isolate_pages_global(unsigned long nr,
975 struct list_head *dst,
976 unsigned long *scanned, int order,
977 int mode, struct zone *z,
978 struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700979 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -0800980{
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700981 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -0800982 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700983 lru += LRU_ACTIVE;
984 if (file)
985 lru += LRU_FILE;
986 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -0700987 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -0800988}
989
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700991 * clear_active_flags() is a helper for shrink_active_list(), clearing
992 * any active bits from the pages in the list.
993 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700994static unsigned long clear_active_flags(struct list_head *page_list,
995 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700996{
997 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700998 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700999 struct page *page;
1000
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001001 list_for_each_entry(page, page_list, lru) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07001002 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001003 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001004 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001005 ClearPageActive(page);
1006 nr_active++;
1007 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001008 count[lru]++;
1009 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001010
1011 return nr_active;
1012}
1013
Nick Piggin62695a82008-10-18 20:26:09 -07001014/**
1015 * isolate_lru_page - tries to isolate a page from its LRU list
1016 * @page: page to isolate from its LRU list
1017 *
1018 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1019 * vmstat statistic corresponding to whatever LRU list the page was on.
1020 *
1021 * Returns 0 if the page was removed from an LRU list.
1022 * Returns -EBUSY if the page was not on an LRU list.
1023 *
1024 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001025 * the active list, it will have PageActive set. If it was found on
1026 * the unevictable list, it will have the PageUnevictable bit set. That flag
1027 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001028 *
1029 * The vmstat statistic corresponding to the list on which the page was
1030 * found will be decremented.
1031 *
1032 * Restrictions:
1033 * (1) Must be called with an elevated refcount on the page. This is a
1034 * fundamentnal difference from isolate_lru_pages (which is called
1035 * without a stable reference).
1036 * (2) the lru_lock must not be held.
1037 * (3) interrupts must be enabled.
1038 */
1039int isolate_lru_page(struct page *page)
1040{
1041 int ret = -EBUSY;
1042
1043 if (PageLRU(page)) {
1044 struct zone *zone = page_zone(page);
1045
1046 spin_lock_irq(&zone->lru_lock);
1047 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001048 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001049 ret = 0;
1050 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001051
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001052 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001053 }
1054 spin_unlock_irq(&zone->lru_lock);
1055 }
1056 return ret;
1057}
1058
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001059/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001060 * Are there way too many processes in the direct reclaim path already?
1061 */
1062static int too_many_isolated(struct zone *zone, int file,
1063 struct scan_control *sc)
1064{
1065 unsigned long inactive, isolated;
1066
1067 if (current_is_kswapd())
1068 return 0;
1069
1070 if (!scanning_global_lru(sc))
1071 return 0;
1072
1073 if (file) {
1074 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1075 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1076 } else {
1077 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1078 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1079 }
1080
1081 return isolated > inactive;
1082}
1083
1084/*
Andrew Morton1742f192006-03-22 00:08:21 -08001085 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1086 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 */
Andrew Morton1742f192006-03-22 00:08:21 -08001088static unsigned long shrink_inactive_list(unsigned long max_scan,
Rik van Riel33c120e2008-10-18 20:26:36 -07001089 struct zone *zone, struct scan_control *sc,
1090 int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091{
1092 LIST_HEAD(page_list);
1093 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -08001094 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001095 unsigned long nr_reclaimed = 0;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001096 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001097 int lumpy_reclaim = 0;
1098
Rik van Riel35cd7812009-09-21 17:01:38 -07001099 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001100 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001101
1102 /* We are about to die and free our memory. Return now. */
1103 if (fatal_signal_pending(current))
1104 return SWAP_CLUSTER_MAX;
1105 }
1106
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001107 /*
1108 * If we need a large contiguous chunk of memory, or have
1109 * trouble getting a small set of contiguous pages, we
1110 * will reclaim both active and inactive pages.
1111 *
1112 * We use the same threshold as pageout congestion_wait below.
1113 */
1114 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1115 lumpy_reclaim = 1;
1116 else if (sc->order && priority < DEF_PRIORITY - 2)
1117 lumpy_reclaim = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001118
1119 pagevec_init(&pvec, 1);
1120
1121 lru_add_drain();
1122 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -08001123 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001124 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -08001125 unsigned long nr_taken;
1126 unsigned long nr_scan;
1127 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001128 unsigned long nr_active;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001129 unsigned int count[NR_LRU_LISTS] = { 0, };
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001130 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001131 unsigned long nr_anon;
1132 unsigned long nr_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001133
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001134 nr_taken = sc->isolate_pages(SWAP_CLUSTER_MAX,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001135 &page_list, &nr_scan, sc->order, mode,
1136 zone, sc->mem_cgroup, 0, file);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001137
1138 if (scanning_global_lru(sc)) {
1139 zone->pages_scanned += nr_scan;
1140 if (current_is_kswapd())
1141 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1142 nr_scan);
1143 else
1144 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1145 nr_scan);
1146 }
1147
1148 if (nr_taken == 0)
1149 goto done;
1150
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001151 nr_active = clear_active_flags(&page_list, count);
Andy Whitcrofte9187bd2007-08-22 14:01:25 -07001152 __count_vm_events(PGDEACTIVATE, nr_active);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001153
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001154 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1155 -count[LRU_ACTIVE_FILE]);
1156 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1157 -count[LRU_INACTIVE_FILE]);
1158 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1159 -count[LRU_ACTIVE_ANON]);
1160 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1161 -count[LRU_INACTIVE_ANON]);
1162
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001163 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1164 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1165 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1166 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001167
Huang Shijie62c0c2f2009-12-14 17:59:48 -08001168 reclaim_stat->recent_scanned[0] += nr_anon;
1169 reclaim_stat->recent_scanned[1] += nr_file;
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001170
Linus Torvalds1da177e2005-04-16 15:20:36 -07001171 spin_unlock_irq(&zone->lru_lock);
1172
Andrew Morton69e05942006-03-22 00:08:19 -08001173 nr_scanned += nr_scan;
Andy Whitcroftc661b072007-08-22 14:01:26 -07001174 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1175
1176 /*
1177 * If we are direct reclaiming for contiguous pages and we do
1178 * not reclaim everything in the list, try again and wait
1179 * for IO to complete. This will stall high-order allocations
1180 * but that should be acceptable to the caller
1181 */
1182 if (nr_freed < nr_taken && !current_is_kswapd() &&
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001183 lumpy_reclaim) {
Jens Axboe8aa7e842009-07-09 14:52:32 +02001184 congestion_wait(BLK_RW_ASYNC, HZ/10);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001185
1186 /*
1187 * The attempt at page out may have made some
1188 * of the pages active, mark them inactive again.
1189 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001190 nr_active = clear_active_flags(&page_list, count);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001191 count_vm_events(PGDEACTIVATE, nr_active);
1192
1193 nr_freed += shrink_page_list(&page_list, sc,
1194 PAGEOUT_IO_SYNC);
1195 }
1196
Andrew Morton05ff5132006-03-22 00:08:20 -08001197 nr_reclaimed += nr_freed;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001198
Nick Piggina74609f2006-01-06 00:11:20 -08001199 local_irq_disable();
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001200 if (current_is_kswapd())
Christoph Lameterf8891e52006-06-30 01:55:45 -07001201 __count_vm_events(KSWAPD_STEAL, nr_freed);
Shantanu Goel918d3f92006-12-29 16:48:59 -08001202 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -08001203
1204 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001205 /*
1206 * Put back any unfreeable pages.
1207 */
1208 while (!list_empty(&page_list)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001209 int lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001210 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -07001211 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212 list_del(&page->lru);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001213 if (unlikely(!page_evictable(page, NULL))) {
1214 spin_unlock_irq(&zone->lru_lock);
1215 putback_lru_page(page);
1216 spin_lock_irq(&zone->lru_lock);
1217 continue;
1218 }
1219 SetPageLRU(page);
1220 lru = page_lru(page);
1221 add_page_to_lru_list(zone, page, lru);
KOSAKI Motohiro74a1c482009-09-21 17:01:45 -07001222 if (is_active_lru(lru)) {
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001223 int file = is_file_lru(lru);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001224 reclaim_stat->recent_rotated[file]++;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001225 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 if (!pagevec_add(&pvec, page)) {
1227 spin_unlock_irq(&zone->lru_lock);
1228 __pagevec_release(&pvec);
1229 spin_lock_irq(&zone->lru_lock);
1230 }
1231 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001232 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1233 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1234
Andrew Morton69e05942006-03-22 00:08:19 -08001235 } while (nr_scanned < max_scan);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001236
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237done:
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001238 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001239 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -08001240 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241}
1242
Martin Bligh3bb1a852006-10-28 10:38:24 -07001243/*
1244 * We are about to scan this zone at a certain priority level. If that priority
1245 * level is smaller (ie: more urgent) than the previous priority, then note
1246 * that priority level within the zone. This is done so that when the next
1247 * process comes in to scan this zone, it will immediately start out at this
1248 * priority level rather than having to build up its own scanning priority.
1249 * Here, this priority affects only the reclaim-mapped threshold.
1250 */
1251static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1252{
1253 if (priority < zone->prev_priority)
1254 zone->prev_priority = priority;
1255}
1256
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001257/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 * This moves pages from the active list to the inactive list.
1259 *
1260 * We move them the other way if the page is referenced by one or more
1261 * processes, from rmap.
1262 *
1263 * If the pages are mostly unmapped, the processing is fast and it is
1264 * appropriate to hold zone->lru_lock across the whole operation. But if
1265 * the pages are mapped, the processing is slow (page_referenced()) so we
1266 * should drop zone->lru_lock around each page. It's impossible to balance
1267 * this, so instead we remove the pages from the LRU while processing them.
1268 * It is safe to rely on PG_active against the non-LRU pages in here because
1269 * nobody will play with that bit on a non-LRU page.
1270 *
1271 * The downside is that we have to touch page->_count against each page.
1272 * But we had to alter page->flags anyway.
1273 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001274
Wu Fengguang3eb41402009-06-16 15:33:13 -07001275static void move_active_pages_to_lru(struct zone *zone,
1276 struct list_head *list,
1277 enum lru_list lru)
1278{
1279 unsigned long pgmoved = 0;
1280 struct pagevec pvec;
1281 struct page *page;
1282
1283 pagevec_init(&pvec, 1);
1284
1285 while (!list_empty(list)) {
1286 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001287
1288 VM_BUG_ON(PageLRU(page));
1289 SetPageLRU(page);
1290
Wu Fengguang3eb41402009-06-16 15:33:13 -07001291 list_move(&page->lru, &zone->lru[lru].list);
1292 mem_cgroup_add_lru_list(page, lru);
1293 pgmoved++;
1294
1295 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1296 spin_unlock_irq(&zone->lru_lock);
1297 if (buffer_heads_over_limit)
1298 pagevec_strip(&pvec);
1299 __pagevec_release(&pvec);
1300 spin_lock_irq(&zone->lru_lock);
1301 }
1302 }
1303 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1304 if (!is_active_lru(lru))
1305 __count_vm_events(PGDEACTIVATE, pgmoved);
1306}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001307
Andrew Morton1742f192006-03-22 00:08:21 -08001308static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001309 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001310{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001311 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001312 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001313 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001315 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001316 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001317 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001318 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001319 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001320
1321 lru_add_drain();
1322 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001323 nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
Balbir Singh66e17072008-02-07 00:13:56 -08001324 ISOLATE_ACTIVE, zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001325 sc->mem_cgroup, 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001326 /*
1327 * zone->pages_scanned is used for detect zone's oom
1328 * mem_cgroup remembers nr_scan by itself.
1329 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001330 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001331 zone->pages_scanned += pgscanned;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001332 }
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001333 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001334
Wu Fengguang3eb41402009-06-16 15:33:13 -07001335 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001336 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001337 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001338 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001339 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001340 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001341 spin_unlock_irq(&zone->lru_lock);
1342
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 while (!list_empty(&l_hold)) {
1344 cond_resched();
1345 page = lru_to_page(&l_hold);
1346 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001347
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001348 if (unlikely(!page_evictable(page, NULL))) {
1349 putback_lru_page(page);
1350 continue;
1351 }
1352
Rik van Riel7e9cd482008-10-18 20:26:35 -07001353 /* page_referenced clears PageReferenced */
1354 if (page_mapping_inuse(page) &&
Wu Fengguang8cab4752009-06-16 15:33:12 -07001355 page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001356 nr_rotated++;
Wu Fengguang8cab4752009-06-16 15:33:12 -07001357 /*
1358 * Identify referenced, file-backed active pages and
1359 * give them one more trip around the active list. So
1360 * that executable code get better chances to stay in
1361 * memory under moderate memory pressure. Anon pages
1362 * are not likely to be evicted by use-once streaming
1363 * IO, plus JVM can create lots of anon VM_EXEC pages,
1364 * so we ignore them here.
1365 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001366 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001367 list_add(&page->lru, &l_active);
1368 continue;
1369 }
1370 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001371
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001372 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001373 list_add(&page->lru, &l_inactive);
1374 }
1375
Andrew Mortonb5557492009-01-06 14:40:13 -08001376 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001377 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001378 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001379 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001380 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001381 * Count referenced pages from currently used mappings as rotated,
1382 * even though only some of them are actually re-activated. This
1383 * helps balance scan pressure between file and anonymous pages in
1384 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001385 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001386 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001387
Wu Fengguang3eb41402009-06-16 15:33:13 -07001388 move_active_pages_to_lru(zone, &l_active,
1389 LRU_ACTIVE + file * LRU_FILE);
1390 move_active_pages_to_lru(zone, &l_inactive,
1391 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001392 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001393 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394}
1395
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001396static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001397{
1398 unsigned long active, inactive;
1399
1400 active = zone_page_state(zone, NR_ACTIVE_ANON);
1401 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1402
1403 if (inactive * zone->inactive_ratio < active)
1404 return 1;
1405
1406 return 0;
1407}
1408
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001409/**
1410 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1411 * @zone: zone to check
1412 * @sc: scan control of this context
1413 *
1414 * Returns true if the zone does not have enough inactive anon pages,
1415 * meaning some active anon pages need to be deactivated.
1416 */
1417static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1418{
1419 int low;
1420
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001421 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001422 low = inactive_anon_is_low_global(zone);
1423 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001424 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001425 return low;
1426}
1427
Rik van Riel56e49d22009-06-16 15:32:28 -07001428static int inactive_file_is_low_global(struct zone *zone)
1429{
1430 unsigned long active, inactive;
1431
1432 active = zone_page_state(zone, NR_ACTIVE_FILE);
1433 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1434
1435 return (active > inactive);
1436}
1437
1438/**
1439 * inactive_file_is_low - check if file pages need to be deactivated
1440 * @zone: zone to check
1441 * @sc: scan control of this context
1442 *
1443 * When the system is doing streaming IO, memory pressure here
1444 * ensures that active file pages get deactivated, until more
1445 * than half of the file pages are on the inactive list.
1446 *
1447 * Once we get to that situation, protect the system's working
1448 * set from being evicted by disabling active file page aging.
1449 *
1450 * This uses a different ratio than the anonymous pages, because
1451 * the page cache uses a use-once replacement algorithm.
1452 */
1453static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1454{
1455 int low;
1456
1457 if (scanning_global_lru(sc))
1458 low = inactive_file_is_low_global(zone);
1459 else
1460 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1461 return low;
1462}
1463
Rik van Rielb39415b2009-12-14 17:59:48 -08001464static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1465 int file)
1466{
1467 if (file)
1468 return inactive_file_is_low(zone, sc);
1469 else
1470 return inactive_anon_is_low(zone, sc);
1471}
1472
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001473static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001474 struct zone *zone, struct scan_control *sc, int priority)
1475{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001476 int file = is_file_lru(lru);
1477
Rik van Rielb39415b2009-12-14 17:59:48 -08001478 if (is_active_lru(lru)) {
1479 if (inactive_list_is_low(zone, sc, file))
1480 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001481 return 0;
1482 }
1483
Rik van Riel33c120e2008-10-18 20:26:36 -07001484 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001485}
1486
Linus Torvalds1da177e2005-04-16 15:20:36 -07001487/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001488 * Determine how aggressively the anon and file LRU lists should be
1489 * scanned. The relative value of each set of LRU lists is determined
1490 * by looking at the fraction of the pages scanned we did rotate back
1491 * onto the active list instead of evict.
1492 *
1493 * percent[0] specifies how much pressure to put on ram/swap backed
1494 * memory, while percent[1] determines pressure on the file LRUs.
1495 */
1496static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1497 unsigned long *percent)
1498{
1499 unsigned long anon, file, free;
1500 unsigned long anon_prio, file_prio;
1501 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001502 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001503
Vincent Li0b217672009-09-21 17:03:09 -07001504 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1505 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1506 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1507 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001508
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001509 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001510 free = zone_page_state(zone, NR_FREE_PAGES);
1511 /* If we have very few page cache pages,
1512 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001513 if (unlikely(file + free <= high_wmark_pages(zone))) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001514 percent[0] = 100;
1515 percent[1] = 0;
1516 return;
1517 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001518 }
1519
1520 /*
1521 * OK, so we have swap space and a fair amount of page cache
1522 * pages. We use the recently rotated / recently scanned
1523 * ratios to determine how valuable each cache is.
1524 *
1525 * Because workloads change over time (and to avoid overflow)
1526 * we keep these statistics as a floating average, which ends
1527 * up weighing recent references more than old ones.
1528 *
1529 * anon in [0], file in [1]
1530 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001531 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001532 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001533 reclaim_stat->recent_scanned[0] /= 2;
1534 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001535 spin_unlock_irq(&zone->lru_lock);
1536 }
1537
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001538 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001539 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001540 reclaim_stat->recent_scanned[1] /= 2;
1541 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001542 spin_unlock_irq(&zone->lru_lock);
1543 }
1544
1545 /*
1546 * With swappiness at 100, anonymous and file have the same priority.
1547 * This scanning priority is essentially the inverse of IO cost.
1548 */
1549 anon_prio = sc->swappiness;
1550 file_prio = 200 - sc->swappiness;
1551
1552 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001553 * The amount of pressure on anon vs file pages is inversely
1554 * proportional to the fraction of recently scanned pages on
1555 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001556 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001557 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1558 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001559
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001560 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1561 fp /= reclaim_stat->recent_rotated[1] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001562
1563 /* Normalize to percentages */
1564 percent[0] = 100 * ap / (ap + fp + 1);
1565 percent[1] = 100 - percent[0];
1566}
1567
Wu Fengguang6e08a362009-06-16 15:32:29 -07001568/*
1569 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1570 * until we collected @swap_cluster_max pages to scan.
1571 */
1572static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001573 unsigned long *nr_saved_scan)
Wu Fengguang6e08a362009-06-16 15:32:29 -07001574{
1575 unsigned long nr;
1576
1577 *nr_saved_scan += nr_to_scan;
1578 nr = *nr_saved_scan;
1579
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001580 if (nr >= SWAP_CLUSTER_MAX)
Wu Fengguang6e08a362009-06-16 15:32:29 -07001581 *nr_saved_scan = 0;
1582 else
1583 nr = 0;
1584
1585 return nr;
1586}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001587
1588/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001589 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1590 */
Rik van Riela79311c2009-01-06 14:40:01 -08001591static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001592 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001594 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001595 unsigned long nr_to_scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001596 unsigned long percent[2]; /* anon @ 0; file @ 1 */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001597 enum lru_list l;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001598 unsigned long nr_reclaimed = sc->nr_reclaimed;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001599 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Wu Fengguangf8629632009-09-21 17:03:11 -07001600 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001601 int noswap = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001602
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001603 /* If we have no swap space, do not bother scanning anon pages. */
1604 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1605 noswap = 1;
1606 percent[0] = 0;
1607 percent[1] = 100;
1608 } else
1609 get_scan_ratio(zone, sc, percent);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001610
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001611 for_each_evictable_lru(l) {
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001612 int file = is_file_lru(l);
Andrew Morton8713e012009-04-30 15:08:55 -07001613 unsigned long scan;
Johannes Weinere0f79b82008-10-18 20:26:55 -07001614
Vincent Li0b217672009-09-21 17:03:09 -07001615 scan = zone_nr_lru_pages(zone, sc, l);
Daisuke Nishimura9198e962009-06-16 15:33:15 -07001616 if (priority || noswap) {
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001617 scan >>= priority;
1618 scan = (scan * percent[file]) / 100;
1619 }
Wu Fengguangf8629632009-09-21 17:03:11 -07001620 nr[l] = nr_scan_try_batch(scan,
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001621 &reclaim_stat->nr_saved_scan[l]);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001622 }
1623
Rik van Riel556adec2008-10-18 20:26:34 -07001624 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1625 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001626 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001627 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001628 nr_to_scan = min_t(unsigned long,
1629 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001630 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001631
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001632 nr_reclaimed += shrink_list(l, nr_to_scan,
1633 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001634 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001635 }
Rik van Riela79311c2009-01-06 14:40:01 -08001636 /*
1637 * On large memory systems, scan >> priority can become
1638 * really large. This is fine for the starting priority;
1639 * we want to put equal scanning pressure on each zone.
1640 * However, if the VM has a harder time of freeing pages,
1641 * with multiple processes reclaiming pages, the total
1642 * freeing target can get unreasonably large.
1643 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08001644 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001645 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001646 }
1647
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001648 sc->nr_reclaimed = nr_reclaimed;
1649
Rik van Riel556adec2008-10-18 20:26:34 -07001650 /*
1651 * Even if we did not try to evict anon pages at all, we want to
1652 * rebalance the anon lru active/inactive ratio.
1653 */
MinChan Kim69c85482009-06-16 15:32:44 -07001654 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
Rik van Riel556adec2008-10-18 20:26:34 -07001655 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1656
Andrew Morton232ea4d2007-02-28 20:13:21 -08001657 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658}
1659
1660/*
1661 * This is the direct reclaim path, for page-allocating processes. We only
1662 * try to reclaim pages from zones which will satisfy the caller's allocation
1663 * request.
1664 *
Mel Gorman41858962009-06-16 15:32:12 -07001665 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1666 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1668 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001669 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1670 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1671 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001673 * If a zone is deemed to be full of pinned pages then just give it a light
1674 * scan then give up on it.
1675 */
Rik van Riela79311c2009-01-06 14:40:01 -08001676static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001677 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678{
Mel Gorman54a6eb52008-04-28 02:12:16 -07001679 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Mel Gormandd1a2392008-04-28 02:12:17 -07001680 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001681 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001682
Nick Piggin408d8542006-09-25 23:31:27 -07001683 sc->all_unreclaimable = 1;
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001684 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1685 sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001686 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001687 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001688 /*
1689 * Take care memory controller reclaiming has small influence
1690 * to global LRU.
1691 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001692 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001693 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1694 continue;
1695 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001696
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001697 if (zone_is_all_unreclaimable(zone) &&
1698 priority != DEF_PRIORITY)
1699 continue; /* Let kswapd poll it */
1700 sc->all_unreclaimable = 0;
1701 } else {
1702 /*
1703 * Ignore cpuset limitation here. We just want to reduce
1704 * # of used pages by us regardless of memory shortage.
1705 */
1706 sc->all_unreclaimable = 0;
1707 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1708 priority);
1709 }
Nick Piggin408d8542006-09-25 23:31:27 -07001710
Rik van Riela79311c2009-01-06 14:40:01 -08001711 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 }
1713}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001714
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715/*
1716 * This is the main entry point to direct page reclaim.
1717 *
1718 * If a full scan of the inactive list fails to free enough memory then we
1719 * are "out of memory" and something needs to be killed.
1720 *
1721 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1722 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02001723 * caller can't do much about. We kick the writeback threads and take explicit
1724 * naps in the hope that some of these pages can be written. But if the
1725 * allocating task holds filesystem locks which prevent writeout this might not
1726 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001727 *
1728 * returns: 0, if no pages reclaimed
1729 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730 */
Mel Gormandac1d272008-04-28 02:12:12 -07001731static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001732 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001733{
1734 int priority;
kosaki.motohiro@jp.fujitsu.comc700be32008-06-12 15:21:27 -07001735 unsigned long ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001736 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001737 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738 unsigned long lru_pages = 0;
Mel Gormandd1a2392008-04-28 02:12:17 -07001739 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001740 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -07001741 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001742 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743
Keika Kobayashi873b4772008-07-25 01:48:52 -07001744 delayacct_freepages_start();
1745
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001746 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001747 count_vm_event(ALLOCSTALL);
1748 /*
1749 * mem_cgroup will not do shrink_slab.
1750 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001751 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001752 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001754 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1755 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756
Wu Fengguangadea02a2009-09-21 17:01:42 -07001757 lru_pages += zone_reclaimable_pages(zone);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001758 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759 }
1760
1761 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001762 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001763 if (!priority)
1764 disable_swap_token();
Rik van Riela79311c2009-01-06 14:40:01 -08001765 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001766 /*
1767 * Don't shrink slabs when reclaiming memory from
1768 * over limit cgroups
1769 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001770 if (scanning_global_lru(sc)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07001771 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001772 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08001773 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001774 reclaim_state->reclaimed_slab = 0;
1775 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776 }
Balbir Singh66e17072008-02-07 00:13:56 -08001777 total_scanned += sc->nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001778 if (sc->nr_reclaimed >= sc->nr_to_reclaim) {
Rik van Riela79311c2009-01-06 14:40:01 -08001779 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 goto out;
1781 }
1782
1783 /*
1784 * Try to write back as many pages as we just scanned. This
1785 * tends to cause slow streaming writers to write data to the
1786 * disk smoothly, at the dirtying rate, which is nice. But
1787 * that's undesirable in laptop mode, where we *want* lumpy
1788 * writeout. So in laptop mode, write out the whole world.
1789 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001790 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1791 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001792 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001793 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794 }
1795
1796 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08001797 if (!sc->hibernation_mode && sc->nr_scanned &&
1798 priority < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02001799 congestion_wait(BLK_RW_ASYNC, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 }
Fernando Luis Vazquez Cao87547ee2008-07-29 22:33:42 -07001801 /* top priority shrink_zones still had more to do? don't OOM, then */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001802 if (!sc->all_unreclaimable && scanning_global_lru(sc))
Rik van Riela79311c2009-01-06 14:40:01 -08001803 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07001805 /*
1806 * Now that we've scanned all the zones at this priority level, note
1807 * that level within the zone so that the next thread which performs
1808 * scanning of this zone will immediately start out at this priority
1809 * level. This affects only the decision whether or not to bring
1810 * mapped pages onto the inactive list.
1811 */
1812 if (priority < 0)
1813 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001815 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001816 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001817
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001818 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1819 continue;
1820
1821 zone->prev_priority = priority;
1822 }
1823 } else
1824 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1825
Keika Kobayashi873b4772008-07-25 01:48:52 -07001826 delayacct_freepages_end();
1827
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 return ret;
1829}
1830
Mel Gormandac1d272008-04-28 02:12:12 -07001831unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001832 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08001833{
1834 struct scan_control sc = {
1835 .gfp_mask = gfp_mask,
1836 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001837 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001838 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001839 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08001840 .swappiness = vm_swappiness,
1841 .order = order,
1842 .mem_cgroup = NULL,
1843 .isolate_pages = isolate_pages_global,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001844 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08001845 };
1846
Mel Gormandd1a2392008-04-28 02:12:17 -07001847 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001848}
1849
Balbir Singh00f0b822008-03-04 14:28:39 -08001850#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08001851
Balbir Singh4e416952009-09-23 15:56:39 -07001852unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1853 gfp_t gfp_mask, bool noswap,
1854 unsigned int swappiness,
1855 struct zone *zone, int nid)
1856{
1857 struct scan_control sc = {
1858 .may_writepage = !laptop_mode,
1859 .may_unmap = 1,
1860 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07001861 .swappiness = swappiness,
1862 .order = 0,
1863 .mem_cgroup = mem,
1864 .isolate_pages = mem_cgroup_isolate_pages,
1865 };
1866 nodemask_t nm = nodemask_of_node(nid);
1867
1868 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1869 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1870 sc.nodemask = &nm;
1871 sc.nr_reclaimed = 0;
1872 sc.nr_scanned = 0;
1873 /*
1874 * NOTE: Although we can get the priority field, using it
1875 * here is not a good idea, since it limits the pages we can scan.
1876 * if we don't reclaim here, the shrink_zone from balance_pgdat
1877 * will pick up pages from other mem cgroup's as well. We hack
1878 * the priority and make it zero.
1879 */
1880 shrink_zone(0, zone, &sc);
1881 return sc.nr_reclaimed;
1882}
1883
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001884unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001885 gfp_t gfp_mask,
1886 bool noswap,
1887 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08001888{
Balbir Singh4e416952009-09-23 15:56:39 -07001889 struct zonelist *zonelist;
Balbir Singh66e17072008-02-07 00:13:56 -08001890 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08001891 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001892 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001893 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001894 .nr_to_reclaim = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001895 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08001896 .order = 0,
1897 .mem_cgroup = mem_cont,
1898 .isolate_pages = mem_cgroup_isolate_pages,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001899 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08001900 };
Balbir Singh66e17072008-02-07 00:13:56 -08001901
Mel Gormandd1a2392008-04-28 02:12:17 -07001902 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1903 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1904 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1905 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001906}
1907#endif
1908
Mel Gormanf50de2d2009-12-14 17:58:53 -08001909/* is kswapd sleeping prematurely? */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001910static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
Mel Gormanf50de2d2009-12-14 17:58:53 -08001911{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001912 int i;
Mel Gormanf50de2d2009-12-14 17:58:53 -08001913
1914 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1915 if (remaining)
1916 return 1;
1917
1918 /* If after HZ/10, a zone is below the high mark, it's premature */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001919 for (i = 0; i < pgdat->nr_zones; i++) {
1920 struct zone *zone = pgdat->node_zones + i;
1921
1922 if (!populated_zone(zone))
1923 continue;
1924
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08001925 if (zone_is_all_unreclaimable(zone))
1926 continue;
1927
Mel Gormanf50de2d2009-12-14 17:58:53 -08001928 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
1929 0, 0))
1930 return 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001931 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08001932
1933 return 0;
1934}
1935
Linus Torvalds1da177e2005-04-16 15:20:36 -07001936/*
1937 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07001938 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001940 * Returns the number of pages which were actually freed.
1941 *
1942 * There is special handling here for zones which are full of pinned pages.
1943 * This can happen if the pages are all mlocked, or if they are all used by
1944 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1945 * What we do is to detect the case where all pages in the zone have been
1946 * scanned twice and there has been zero successful reclaim. Mark the zone as
1947 * dead and from now on, only perform a short scan. Basically we're polling
1948 * the zone for when the problem goes away.
1949 *
1950 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07001951 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1952 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1953 * lower zones regardless of the number of free pages in the lower zones. This
1954 * interoperates with the page allocator fallback scheme to ensure that aging
1955 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001956 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001957static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001958{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959 int all_zones_ok;
1960 int priority;
1961 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08001962 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001963 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08001964 struct scan_control sc = {
1965 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001966 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001967 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001968 /*
1969 * kswapd doesn't want to be bailed out while reclaim. because
1970 * we want to put equal scanning pressure on each zone.
1971 */
1972 .nr_to_reclaim = ULONG_MAX,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001973 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001974 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08001975 .mem_cgroup = NULL,
1976 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08001977 };
Martin Bligh3bb1a852006-10-28 10:38:24 -07001978 /*
1979 * temp_priority is used to remember the scanning priority at which
Mel Gorman41858962009-06-16 15:32:12 -07001980 * this zone was successfully refilled to
1981 * free_pages == high_wmark_pages(zone).
Martin Bligh3bb1a852006-10-28 10:38:24 -07001982 */
1983 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001984
1985loop_again:
1986 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08001987 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07001988 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001989 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990
Martin Bligh3bb1a852006-10-28 10:38:24 -07001991 for (i = 0; i < pgdat->nr_zones; i++)
1992 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993
1994 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1995 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1996 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001997 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001999 /* The swap token gets in the way of swapout... */
2000 if (!priority)
2001 disable_swap_token();
2002
Linus Torvalds1da177e2005-04-16 15:20:36 -07002003 all_zones_ok = 1;
2004
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002005 /*
2006 * Scan in the highmem->dma direction for the highest
2007 * zone which needs scanning
2008 */
2009 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2010 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002012 if (!populated_zone(zone))
2013 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002014
David Rientjese815af92007-10-16 23:25:54 -07002015 if (zone_is_all_unreclaimable(zone) &&
2016 priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002017 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018
Rik van Riel556adec2008-10-18 20:26:34 -07002019 /*
2020 * Do some background aging of the anon list, to give
2021 * pages a chance to be referenced before reclaiming.
2022 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002023 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002024 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2025 &sc, priority, 0);
2026
Mel Gorman41858962009-06-16 15:32:12 -07002027 if (!zone_watermark_ok(zone, order,
2028 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002029 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002030 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002033 if (i < 0)
2034 goto out;
2035
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 for (i = 0; i <= end_zone; i++) {
2037 struct zone *zone = pgdat->node_zones + i;
2038
Wu Fengguangadea02a2009-09-21 17:01:42 -07002039 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 }
2041
2042 /*
2043 * Now scan the zone in the dma->highmem direction, stopping
2044 * at the last zone which needs scanning.
2045 *
2046 * We do this because the page allocator works in the opposite
2047 * direction. This prevents the page allocator from allocating
2048 * pages behind kswapd's direction of progress, which would
2049 * cause too much scanning of the lower zones.
2050 */
2051 for (i = 0; i <= end_zone; i++) {
2052 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002053 int nr_slab;
Balbir Singh4e416952009-09-23 15:56:39 -07002054 int nid, zid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055
Con Kolivasf3fe6512006-01-06 00:11:15 -08002056 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002057 continue;
2058
David Rientjese815af92007-10-16 23:25:54 -07002059 if (zone_is_all_unreclaimable(zone) &&
2060 priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002061 continue;
2062
Mel Gorman41858962009-06-16 15:32:12 -07002063 if (!zone_watermark_ok(zone, order,
2064 high_wmark_pages(zone), end_zone, 0))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002065 all_zones_ok = 0;
Martin Bligh3bb1a852006-10-28 10:38:24 -07002066 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067 sc.nr_scanned = 0;
Martin Bligh3bb1a852006-10-28 10:38:24 -07002068 note_zone_scanning_priority(zone, priority);
Balbir Singh4e416952009-09-23 15:56:39 -07002069
2070 nid = pgdat->node_id;
2071 zid = zone_idx(zone);
2072 /*
2073 * Call soft limit reclaim before calling shrink_zone.
2074 * For now we ignore the return value
2075 */
2076 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2077 nid, zid);
Rik van Riel32a43302007-10-16 01:24:50 -07002078 /*
2079 * We put equal pressure on every zone, unless one
2080 * zone has way too many pages free already.
2081 */
Mel Gorman41858962009-06-16 15:32:12 -07002082 if (!zone_watermark_ok(zone, order,
2083 8*high_wmark_pages(zone), end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002084 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002086 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2087 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002088 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002089 total_scanned += sc.nr_scanned;
David Rientjese815af92007-10-16 23:25:54 -07002090 if (zone_is_all_unreclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002091 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002092 if (nr_slab == 0 && zone->pages_scanned >=
Wu Fengguangadea02a2009-09-21 17:01:42 -07002093 (zone_reclaimable_pages(zone) * 6))
David Rientjese815af92007-10-16 23:25:54 -07002094 zone_set_flag(zone,
2095 ZONE_ALL_UNRECLAIMABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 /*
2097 * If we've done a decent amount of scanning and
2098 * the reclaim ratio is low, start doing writepage
2099 * even in laptop mode
2100 */
2101 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002102 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002104
2105 /*
2106 * We are still under min water mark. it mean we have
2107 * GFP_ATOMIC allocation failure risk. Hurry up!
2108 */
2109 if (!zone_watermark_ok(zone, order, min_wmark_pages(zone),
2110 end_zone, 0))
2111 has_under_min_watermark_zone = 1;
2112
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002114 if (all_zones_ok)
2115 break; /* kswapd: all done */
2116 /*
2117 * OK, kswapd is getting into trouble. Take a nap, then take
2118 * another pass across the zones.
2119 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002120 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2121 if (has_under_min_watermark_zone)
2122 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2123 else
2124 congestion_wait(BLK_RW_ASYNC, HZ/10);
2125 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002126
2127 /*
2128 * We do this so kswapd doesn't build up large priorities for
2129 * example when it is freeing in parallel with allocators. It
2130 * matches the direct reclaim path behaviour in terms of impact
2131 * on zone->*_priority.
2132 */
Rik van Riela79311c2009-01-06 14:40:01 -08002133 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002134 break;
2135 }
2136out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07002137 /*
2138 * Note within each zone the priority level at which this zone was
2139 * brought into a happy state. So that the next thread which scans this
2140 * zone will start out at that priority level.
2141 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002142 for (i = 0; i < pgdat->nr_zones; i++) {
2143 struct zone *zone = pgdat->node_zones + i;
2144
Martin Bligh3bb1a852006-10-28 10:38:24 -07002145 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002146 }
2147 if (!all_zones_ok) {
2148 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002149
2150 try_to_freeze();
2151
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002152 /*
2153 * Fragmentation may mean that the system cannot be
2154 * rebalanced for high-order allocations in all zones.
2155 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2156 * it means the zones have been fully scanned and are still
2157 * not balanced. For high-order allocations, there is
2158 * little point trying all over again as kswapd may
2159 * infinite loop.
2160 *
2161 * Instead, recheck all watermarks at order-0 as they
2162 * are the most important. If watermarks are ok, kswapd will go
2163 * back to sleep. High-order users can still perform direct
2164 * reclaim if they wish.
2165 */
2166 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2167 order = sc.order = 0;
2168
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169 goto loop_again;
2170 }
2171
Rik van Riela79311c2009-01-06 14:40:01 -08002172 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002173}
2174
2175/*
2176 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002177 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178 *
2179 * This basically trickles out pages so that we have _some_
2180 * free memory available even if there is no other activity
2181 * that frees anything up. This is needed for things like routing
2182 * etc, where we otherwise might have all activity going on in
2183 * asynchronous contexts that cannot page things out.
2184 *
2185 * If there are applications that are active memory-allocators
2186 * (most normal use), this basically shouldn't matter.
2187 */
2188static int kswapd(void *p)
2189{
2190 unsigned long order;
2191 pg_data_t *pgdat = (pg_data_t*)p;
2192 struct task_struct *tsk = current;
2193 DEFINE_WAIT(wait);
2194 struct reclaim_state reclaim_state = {
2195 .reclaimed_slab = 0,
2196 };
Rusty Russella70f7302009-03-13 14:49:46 +10302197 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198
Nick Piggincf40bd12009-01-21 08:12:39 +01002199 lockdep_set_current_reclaim_state(GFP_KERNEL);
2200
Rusty Russell174596a2009-01-01 10:12:29 +10302201 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002202 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 current->reclaim_state = &reclaim_state;
2204
2205 /*
2206 * Tell the memory management that we're a "memory allocator",
2207 * and that if we need more memory we should get access to it
2208 * regardless (see "__alloc_pages()"). "kswapd" should
2209 * never get caught in the normal page freeing logic.
2210 *
2211 * (Kswapd normally doesn't need memory anyway, but sometimes
2212 * you need a small amount of memory in order to be able to
2213 * page out something else, and this flag essentially protects
2214 * us from recursively trying to free more memory as we're
2215 * trying to free the first piece of memory in the first place).
2216 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002217 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002218 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219
2220 order = 0;
2221 for ( ; ; ) {
2222 unsigned long new_order;
David Rientjes8fe23e02009-12-14 17:58:33 -08002223 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002224
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2226 new_order = pgdat->kswapd_max_order;
2227 pgdat->kswapd_max_order = 0;
2228 if (order < new_order) {
2229 /*
2230 * Don't sleep if someone wants a larger 'order'
2231 * allocation
2232 */
2233 order = new_order;
2234 } else {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002235 if (!freezing(current) && !kthread_should_stop()) {
2236 long remaining = 0;
2237
2238 /* Try to sleep for a short interval */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002239 if (!sleeping_prematurely(pgdat, order, remaining)) {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002240 remaining = schedule_timeout(HZ/10);
2241 finish_wait(&pgdat->kswapd_wait, &wait);
2242 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2243 }
2244
2245 /*
2246 * After a short sleep, check if it was a
2247 * premature sleep. If not, then go fully
2248 * to sleep until explicitly woken up
2249 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002250 if (!sleeping_prematurely(pgdat, order, remaining))
Mel Gormanf50de2d2009-12-14 17:58:53 -08002251 schedule();
2252 else {
2253 if (remaining)
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002254 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002255 else
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002256 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002257 }
2258 }
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002259
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260 order = pgdat->kswapd_max_order;
2261 }
2262 finish_wait(&pgdat->kswapd_wait, &wait);
2263
David Rientjes8fe23e02009-12-14 17:58:33 -08002264 ret = try_to_freeze();
2265 if (kthread_should_stop())
2266 break;
2267
2268 /*
2269 * We can speed up thawing tasks if we don't call balance_pgdat
2270 * after returning from the refrigerator
2271 */
2272 if (!ret)
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002273 balance_pgdat(pgdat, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 }
2275 return 0;
2276}
2277
2278/*
2279 * A zone is low on free memory, so wake its kswapd task to service it.
2280 */
2281void wakeup_kswapd(struct zone *zone, int order)
2282{
2283 pg_data_t *pgdat;
2284
Con Kolivasf3fe6512006-01-06 00:11:15 -08002285 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 return;
2287
2288 pgdat = zone->zone_pgdat;
Mel Gorman41858962009-06-16 15:32:12 -07002289 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290 return;
2291 if (pgdat->kswapd_max_order < order)
2292 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08002293 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002294 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002295 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002296 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002297 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298}
2299
Wu Fengguangadea02a2009-09-21 17:01:42 -07002300/*
2301 * The reclaimable count would be mostly accurate.
2302 * The less reclaimable pages may be
2303 * - mlocked pages, which will be moved to unevictable list when encountered
2304 * - mapped pages, which may require several travels to be reclaimed
2305 * - dirty pages, which is not "instantly" reclaimable
2306 */
2307unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002308{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002309 int nr;
2310
2311 nr = global_page_state(NR_ACTIVE_FILE) +
2312 global_page_state(NR_INACTIVE_FILE);
2313
2314 if (nr_swap_pages > 0)
2315 nr += global_page_state(NR_ACTIVE_ANON) +
2316 global_page_state(NR_INACTIVE_ANON);
2317
2318 return nr;
2319}
2320
2321unsigned long zone_reclaimable_pages(struct zone *zone)
2322{
2323 int nr;
2324
2325 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2326 zone_page_state(zone, NR_INACTIVE_FILE);
2327
2328 if (nr_swap_pages > 0)
2329 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2330 zone_page_state(zone, NR_INACTIVE_ANON);
2331
2332 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002333}
2334
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002335#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002337 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002338 * freed pages.
2339 *
2340 * Rather than trying to age LRUs the aim is to preserve the overall
2341 * LRU order by reclaiming preferentially
2342 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002344unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002345{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002346 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002347 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002348 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2349 .may_swap = 1,
2350 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002351 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002352 .nr_to_reclaim = nr_to_reclaim,
2353 .hibernation_mode = 1,
2354 .swappiness = vm_swappiness,
2355 .order = 0,
Balbir Singh66e17072008-02-07 00:13:56 -08002356 .isolate_pages = isolate_pages_global,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002357 };
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002358 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2359 struct task_struct *p = current;
2360 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002362 p->flags |= PF_MEMALLOC;
2363 lockdep_set_current_reclaim_state(sc.gfp_mask);
2364 reclaim_state.reclaimed_slab = 0;
2365 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002366
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002367 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002368
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002369 p->reclaim_state = NULL;
2370 lockdep_clear_current_reclaim_state();
2371 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002372
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002373 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002374}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002375#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002376
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377/* It's optimal to keep kswapds on the same CPUs as their memory, but
2378 not required for correctness. So if the last cpu in a node goes
2379 away, we get changed to run anywhere: as the first one comes back,
2380 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002381static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002382 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002384 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002385
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002386 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002387 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002388 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302389 const struct cpumask *mask;
2390
2391 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002392
Rusty Russell3e597942009-01-01 10:12:24 +10302393 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002395 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002396 }
2397 }
2398 return NOTIFY_OK;
2399}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400
Yasunori Goto3218ae12006-06-27 02:53:33 -07002401/*
2402 * This kswapd start function will be called by init and node-hot-add.
2403 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2404 */
2405int kswapd_run(int nid)
2406{
2407 pg_data_t *pgdat = NODE_DATA(nid);
2408 int ret = 0;
2409
2410 if (pgdat->kswapd)
2411 return 0;
2412
2413 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2414 if (IS_ERR(pgdat->kswapd)) {
2415 /* failure at boot is fatal */
2416 BUG_ON(system_state == SYSTEM_BOOTING);
2417 printk("Failed to start kswapd on node %d\n",nid);
2418 ret = -1;
2419 }
2420 return ret;
2421}
2422
David Rientjes8fe23e02009-12-14 17:58:33 -08002423/*
2424 * Called by memory hotplug when all memory in a node is offlined.
2425 */
2426void kswapd_stop(int nid)
2427{
2428 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2429
2430 if (kswapd)
2431 kthread_stop(kswapd);
2432}
2433
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434static int __init kswapd_init(void)
2435{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002436 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002437
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002439 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002440 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 hotcpu_notifier(cpu_callback, 0);
2442 return 0;
2443}
2444
2445module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002446
2447#ifdef CONFIG_NUMA
2448/*
2449 * Zone reclaim mode
2450 *
2451 * If non-zero call zone_reclaim when the number of free pages falls below
2452 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002453 */
2454int zone_reclaim_mode __read_mostly;
2455
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002456#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002457#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002458#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2459#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2460
Christoph Lameter9eeff232006-01-18 17:42:31 -08002461/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002462 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2463 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2464 * a zone.
2465 */
2466#define ZONE_RECLAIM_PRIORITY 4
2467
Christoph Lameter9eeff232006-01-18 17:42:31 -08002468/*
Christoph Lameter96146342006-07-03 00:24:13 -07002469 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2470 * occur.
2471 */
2472int sysctl_min_unmapped_ratio = 1;
2473
2474/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002475 * If the number of slab pages in a zone grows beyond this percentage then
2476 * slab reclaim needs to occur.
2477 */
2478int sysctl_min_slab_ratio = 5;
2479
Mel Gorman90afa5d2009-06-16 15:33:20 -07002480static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2481{
2482 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2483 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2484 zone_page_state(zone, NR_ACTIVE_FILE);
2485
2486 /*
2487 * It's possible for there to be more file mapped pages than
2488 * accounted for by the pages on the file LRU lists because
2489 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2490 */
2491 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2492}
2493
2494/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2495static long zone_pagecache_reclaimable(struct zone *zone)
2496{
2497 long nr_pagecache_reclaimable;
2498 long delta = 0;
2499
2500 /*
2501 * If RECLAIM_SWAP is set, then all file pages are considered
2502 * potentially reclaimable. Otherwise, we have to worry about
2503 * pages like swapcache and zone_unmapped_file_pages() provides
2504 * a better estimate
2505 */
2506 if (zone_reclaim_mode & RECLAIM_SWAP)
2507 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2508 else
2509 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2510
2511 /* If we can't clean pages, remove dirty pages from consideration */
2512 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2513 delta += zone_page_state(zone, NR_FILE_DIRTY);
2514
2515 /* Watch for any possible underflows due to delta */
2516 if (unlikely(delta > nr_pagecache_reclaimable))
2517 delta = nr_pagecache_reclaimable;
2518
2519 return nr_pagecache_reclaimable - delta;
2520}
2521
Christoph Lameter0ff38492006-09-25 23:31:52 -07002522/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002523 * Try to free up some pages from this zone through reclaim.
2524 */
Andrew Morton179e9632006-03-22 00:08:18 -08002525static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002526{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002527 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002528 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002529 struct task_struct *p = current;
2530 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002531 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002532 struct scan_control sc = {
2533 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002534 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002535 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002536 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2537 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002538 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002539 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002540 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002541 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08002542 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07002543 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002544
2545 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08002546 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002547 /*
2548 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2549 * and we also need to be able to write out pages for RECLAIM_WRITE
2550 * and RECLAIM_SWAP.
2551 */
2552 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002553 reclaim_state.reclaimed_slab = 0;
2554 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002555
Mel Gorman90afa5d2009-06-16 15:33:20 -07002556 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07002557 /*
2558 * Free memory by calling shrink zone with increasing
2559 * priorities until we have enough memory freed.
2560 */
2561 priority = ZONE_RECLAIM_PRIORITY;
2562 do {
Martin Bligh3bb1a852006-10-28 10:38:24 -07002563 note_zone_scanning_priority(zone, priority);
Rik van Riela79311c2009-01-06 14:40:01 -08002564 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002565 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002566 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002567 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002568
Christoph Lameter83e33a42006-09-25 23:31:53 -07002569 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2570 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002571 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002572 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002573 * many pages were freed in this zone. So we take the current
2574 * number of slab pages and shake the slab until it is reduced
2575 * by the same nr_pages that we used for reclaiming unmapped
2576 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002577 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002578 * Note that shrink_slab will free memory on all zones and may
2579 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002580 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07002581 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07002582 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2583 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07002584 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07002585
2586 /*
2587 * Update nr_reclaimed by the number of slab pages we
2588 * reclaimed from this zone.
2589 */
Rik van Riela79311c2009-01-06 14:40:01 -08002590 sc.nr_reclaimed += slab_reclaimable -
Christoph Lameter83e33a42006-09-25 23:31:53 -07002591 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002592 }
2593
Christoph Lameter9eeff232006-01-18 17:42:31 -08002594 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002595 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
Rik van Riela79311c2009-01-06 14:40:01 -08002596 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002597}
Andrew Morton179e9632006-03-22 00:08:18 -08002598
2599int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2600{
Andrew Morton179e9632006-03-22 00:08:18 -08002601 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002602 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002603
2604 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002605 * Zone reclaim reclaims unmapped file backed pages and
2606 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002607 *
Christoph Lameter96146342006-07-03 00:24:13 -07002608 * A small portion of unmapped file backed pages is needed for
2609 * file I/O otherwise pages read by file I/O will be immediately
2610 * thrown out if the zone is overallocated. So we do not reclaim
2611 * if less than a specified percentage of the zone is used by
2612 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002613 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07002614 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2615 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002616 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08002617
David Rientjesd773ed62007-10-16 23:26:01 -07002618 if (zone_is_all_unreclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002619 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07002620
Andrew Morton179e9632006-03-22 00:08:18 -08002621 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002622 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002623 */
David Rientjesd773ed62007-10-16 23:26:01 -07002624 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002625 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08002626
2627 /*
2628 * Only run zone reclaim on the local zone or on zones that do not
2629 * have associated processors. This will favor the local processor
2630 * over remote processors and spread off node memory allocations
2631 * as wide as possible.
2632 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002633 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002634 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07002635 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07002636
2637 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002638 return ZONE_RECLAIM_NOSCAN;
2639
David Rientjesd773ed62007-10-16 23:26:01 -07002640 ret = __zone_reclaim(zone, gfp_mask, order);
2641 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2642
Mel Gorman24cf725182009-06-16 15:33:23 -07002643 if (!ret)
2644 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2645
David Rientjesd773ed62007-10-16 23:26:01 -07002646 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002647}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002648#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002649
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002650/*
2651 * page_evictable - test whether a page is evictable
2652 * @page: the page to test
2653 * @vma: the VMA in which the page is or will be mapped, may be NULL
2654 *
2655 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07002656 * lists vs unevictable list. The vma argument is !NULL when called from the
2657 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002658 *
2659 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002660 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07002661 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002662 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002663 */
2664int page_evictable(struct page *page, struct vm_area_struct *vma)
2665{
2666
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002667 if (mapping_unevictable(page_mapping(page)))
2668 return 0;
2669
Nick Pigginb291f002008-10-18 20:26:44 -07002670 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2671 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002672
2673 return 1;
2674}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002675
2676/**
2677 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2678 * @page: page to check evictability and move to appropriate lru list
2679 * @zone: zone page is in
2680 *
2681 * Checks a page for evictability and moves the page to the appropriate
2682 * zone lru list.
2683 *
2684 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2685 * have PageUnevictable set.
2686 */
2687static void check_move_unevictable_page(struct page *page, struct zone *zone)
2688{
2689 VM_BUG_ON(PageActive(page));
2690
2691retry:
2692 ClearPageUnevictable(page);
2693 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07002694 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002695
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002696 __dec_zone_state(zone, NR_UNEVICTABLE);
2697 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002698 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002699 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2700 __count_vm_event(UNEVICTABLE_PGRESCUED);
2701 } else {
2702 /*
2703 * rotate unevictable list
2704 */
2705 SetPageUnevictable(page);
2706 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002707 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002708 if (page_evictable(page, NULL))
2709 goto retry;
2710 }
2711}
2712
2713/**
2714 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2715 * @mapping: struct address_space to scan for evictable pages
2716 *
2717 * Scan all pages in mapping. Check unevictable pages for
2718 * evictability and move them to the appropriate zone lru list.
2719 */
2720void scan_mapping_unevictable_pages(struct address_space *mapping)
2721{
2722 pgoff_t next = 0;
2723 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2724 PAGE_CACHE_SHIFT;
2725 struct zone *zone;
2726 struct pagevec pvec;
2727
2728 if (mapping->nrpages == 0)
2729 return;
2730
2731 pagevec_init(&pvec, 0);
2732 while (next < end &&
2733 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2734 int i;
2735 int pg_scanned = 0;
2736
2737 zone = NULL;
2738
2739 for (i = 0; i < pagevec_count(&pvec); i++) {
2740 struct page *page = pvec.pages[i];
2741 pgoff_t page_index = page->index;
2742 struct zone *pagezone = page_zone(page);
2743
2744 pg_scanned++;
2745 if (page_index > next)
2746 next = page_index;
2747 next++;
2748
2749 if (pagezone != zone) {
2750 if (zone)
2751 spin_unlock_irq(&zone->lru_lock);
2752 zone = pagezone;
2753 spin_lock_irq(&zone->lru_lock);
2754 }
2755
2756 if (PageLRU(page) && PageUnevictable(page))
2757 check_move_unevictable_page(page, zone);
2758 }
2759 if (zone)
2760 spin_unlock_irq(&zone->lru_lock);
2761 pagevec_release(&pvec);
2762
2763 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2764 }
2765
2766}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002767
2768/**
2769 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2770 * @zone - zone of which to scan the unevictable list
2771 *
2772 * Scan @zone's unevictable LRU lists to check for pages that have become
2773 * evictable. Move those that have to @zone's inactive list where they
2774 * become candidates for reclaim, unless shrink_inactive_zone() decides
2775 * to reactivate them. Pages that are still unevictable are rotated
2776 * back onto @zone's unevictable list.
2777 */
2778#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08002779static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002780{
2781 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2782 unsigned long scan;
2783 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2784
2785 while (nr_to_scan > 0) {
2786 unsigned long batch_size = min(nr_to_scan,
2787 SCAN_UNEVICTABLE_BATCH_SIZE);
2788
2789 spin_lock_irq(&zone->lru_lock);
2790 for (scan = 0; scan < batch_size; scan++) {
2791 struct page *page = lru_to_page(l_unevictable);
2792
2793 if (!trylock_page(page))
2794 continue;
2795
2796 prefetchw_prev_lru_page(page, l_unevictable, flags);
2797
2798 if (likely(PageLRU(page) && PageUnevictable(page)))
2799 check_move_unevictable_page(page, zone);
2800
2801 unlock_page(page);
2802 }
2803 spin_unlock_irq(&zone->lru_lock);
2804
2805 nr_to_scan -= batch_size;
2806 }
2807}
2808
2809
2810/**
2811 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2812 *
2813 * A really big hammer: scan all zones' unevictable LRU lists to check for
2814 * pages that have become evictable. Move those back to the zones'
2815 * inactive list where they become candidates for reclaim.
2816 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2817 * and we add swap to the system. As such, it runs in the context of a task
2818 * that has possibly/probably made some previously unevictable pages
2819 * evictable.
2820 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08002821static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002822{
2823 struct zone *zone;
2824
2825 for_each_zone(zone) {
2826 scan_zone_unevictable_pages(zone);
2827 }
2828}
2829
2830/*
2831 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2832 * all nodes' unevictable lists for evictable pages
2833 */
2834unsigned long scan_unevictable_pages;
2835
2836int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002837 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002838 size_t *length, loff_t *ppos)
2839{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002840 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002841
2842 if (write && *(unsigned long *)table->data)
2843 scan_all_zones_unevictable_pages();
2844
2845 scan_unevictable_pages = 0;
2846 return 0;
2847}
2848
2849/*
2850 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2851 * a specified node's per zone unevictable lists for evictable pages.
2852 */
2853
2854static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2855 struct sysdev_attribute *attr,
2856 char *buf)
2857{
2858 return sprintf(buf, "0\n"); /* always zero; should fit... */
2859}
2860
2861static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2862 struct sysdev_attribute *attr,
2863 const char *buf, size_t count)
2864{
2865 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2866 struct zone *zone;
2867 unsigned long res;
2868 unsigned long req = strict_strtoul(buf, 10, &res);
2869
2870 if (!req)
2871 return 1; /* zero is no-op */
2872
2873 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2874 if (!populated_zone(zone))
2875 continue;
2876 scan_zone_unevictable_pages(zone);
2877 }
2878 return 1;
2879}
2880
2881
2882static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2883 read_scan_unevictable_node,
2884 write_scan_unevictable_node);
2885
2886int scan_unevictable_register_node(struct node *node)
2887{
2888 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2889}
2890
2891void scan_unevictable_unregister_node(struct node *node)
2892{
2893 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2894}
2895