blob: 79c809895fba777d6069ae3778046def74a29b81 [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
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265static inline int is_page_cache_freeable(struct page *page)
266{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700267 /*
268 * A freeable page cache page is referenced only by the caller
269 * that isolated the page, the page cache radix tree and
270 * optional buffer heads at page->private.
271 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700272 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273}
274
275static int may_write_to_queue(struct backing_dev_info *bdi)
276{
Christoph Lameter930d9152006-01-08 01:00:47 -0800277 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 return 1;
279 if (!bdi_write_congested(bdi))
280 return 1;
281 if (bdi == current->backing_dev_info)
282 return 1;
283 return 0;
284}
285
286/*
287 * We detected a synchronous write error writing a page out. Probably
288 * -ENOSPC. We need to propagate that into the address_space for a subsequent
289 * fsync(), msync() or close().
290 *
291 * The tricky part is that after writepage we cannot touch the mapping: nothing
292 * prevents it from being freed up. But we have a ref on the page and once
293 * that page is locked, the mapping is pinned.
294 *
295 * We're allowed to run sleeping lock_page() here because we know the caller has
296 * __GFP_FS.
297 */
298static void handle_write_error(struct address_space *mapping,
299 struct page *page, int error)
300{
301 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700302 if (page_mapping(page) == mapping)
303 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 unlock_page(page);
305}
306
Andy Whitcroftc661b072007-08-22 14:01:26 -0700307/* Request for sync pageout. */
308enum pageout_io {
309 PAGEOUT_IO_ASYNC,
310 PAGEOUT_IO_SYNC,
311};
312
Christoph Lameter04e62a22006-06-23 02:03:38 -0700313/* possible outcome of pageout() */
314typedef enum {
315 /* failed to write page out, page is locked */
316 PAGE_KEEP,
317 /* move page to the active list, page is locked */
318 PAGE_ACTIVATE,
319 /* page has been sent to the disk successfully, page is unlocked */
320 PAGE_SUCCESS,
321 /* page is clean and locked */
322 PAGE_CLEAN,
323} pageout_t;
324
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325/*
Andrew Morton1742f192006-03-22 00:08:21 -0800326 * pageout is called by shrink_page_list() for each dirty page.
327 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700329static pageout_t pageout(struct page *page, struct address_space *mapping,
330 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331{
332 /*
333 * If the page is dirty, only perform writeback if that write
334 * will be non-blocking. To prevent this allocation from being
335 * stalled by pagecache activity. But note that there may be
336 * stalls if we need to run get_block(). We could test
337 * PagePrivate for that.
338 *
Vincent Li6aceb532009-12-14 17:58:49 -0800339 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 * this page's queue, we can perform writeback even if that
341 * will block.
342 *
343 * If the page is swapcache, write it back even if that would
344 * block, for some throttling. This happens by accident, because
345 * swap_backing_dev_info is bust: it doesn't reflect the
346 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 */
348 if (!is_page_cache_freeable(page))
349 return PAGE_KEEP;
350 if (!mapping) {
351 /*
352 * Some data journaling orphaned pages can have
353 * page->mapping == NULL while being dirty with clean buffers.
354 */
David Howells266cf652009-04-03 16:42:36 +0100355 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356 if (try_to_free_buffers(page)) {
357 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700358 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 return PAGE_CLEAN;
360 }
361 }
362 return PAGE_KEEP;
363 }
364 if (mapping->a_ops->writepage == NULL)
365 return PAGE_ACTIVATE;
366 if (!may_write_to_queue(mapping->backing_dev_info))
367 return PAGE_KEEP;
368
369 if (clear_page_dirty_for_io(page)) {
370 int res;
371 struct writeback_control wbc = {
372 .sync_mode = WB_SYNC_NONE,
373 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700374 .range_start = 0,
375 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 .nonblocking = 1,
377 .for_reclaim = 1,
378 };
379
380 SetPageReclaim(page);
381 res = mapping->a_ops->writepage(page, &wbc);
382 if (res < 0)
383 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800384 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 ClearPageReclaim(page);
386 return PAGE_ACTIVATE;
387 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700388
389 /*
390 * Wait on writeback if requested to. This happens when
391 * direct reclaiming a large contiguous area and the
392 * first attempt to free a range of pages fails.
393 */
394 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
395 wait_on_page_writeback(page);
396
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 if (!PageWriteback(page)) {
398 /* synchronous write or broken a_ops? */
399 ClearPageReclaim(page);
400 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700401 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402 return PAGE_SUCCESS;
403 }
404
405 return PAGE_CLEAN;
406}
407
Andrew Mortona649fd92006-10-17 00:09:36 -0700408/*
Nick Piggine2867812008-07-25 19:45:30 -0700409 * Same as remove_mapping, but if the page is removed from the mapping, it
410 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700411 */
Nick Piggine2867812008-07-25 19:45:30 -0700412static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800413{
Nick Piggin28e4d962006-09-25 23:31:23 -0700414 BUG_ON(!PageLocked(page));
415 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800416
Nick Piggin19fd6232008-07-25 19:45:32 -0700417 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800418 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700419 * The non racy check for a busy page.
420 *
421 * Must be careful with the order of the tests. When someone has
422 * a ref to the page, it may be possible that they dirty it then
423 * drop the reference. So if PageDirty is tested before page_count
424 * here, then the following race may occur:
425 *
426 * get_user_pages(&page);
427 * [user mapping goes away]
428 * write_to(page);
429 * !PageDirty(page) [good]
430 * SetPageDirty(page);
431 * put_page(page);
432 * !page_count(page) [good, discard it]
433 *
434 * [oops, our write_to data is lost]
435 *
436 * Reversing the order of the tests ensures such a situation cannot
437 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
438 * load is not satisfied before that of page->_count.
439 *
440 * Note that if SetPageDirty is always performed via set_page_dirty,
441 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800442 */
Nick Piggine2867812008-07-25 19:45:30 -0700443 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800444 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700445 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
446 if (unlikely(PageDirty(page))) {
447 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800448 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700449 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800450
451 if (PageSwapCache(page)) {
452 swp_entry_t swap = { .val = page_private(page) };
453 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700454 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700455 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700456 } else {
457 __remove_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700458 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700459 mem_cgroup_uncharge_cache_page(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800460 }
461
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800462 return 1;
463
464cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700465 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800466 return 0;
467}
468
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469/*
Nick Piggine2867812008-07-25 19:45:30 -0700470 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
471 * someone else has a ref on the page, abort and return 0. If it was
472 * successfully detached, return 1. Assumes the caller has a single ref on
473 * this page.
474 */
475int remove_mapping(struct address_space *mapping, struct page *page)
476{
477 if (__remove_mapping(mapping, page)) {
478 /*
479 * Unfreezing the refcount with 1 rather than 2 effectively
480 * drops the pagecache ref for us without requiring another
481 * atomic operation.
482 */
483 page_unfreeze_refs(page, 1);
484 return 1;
485 }
486 return 0;
487}
488
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700489/**
490 * putback_lru_page - put previously isolated page onto appropriate LRU list
491 * @page: page to be put back to appropriate lru list
492 *
493 * Add previously isolated @page to appropriate LRU list.
494 * Page may still be unevictable for other reasons.
495 *
496 * lru_lock must not be held, interrupts must be enabled.
497 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700498void putback_lru_page(struct page *page)
499{
500 int lru;
501 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700502 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700503
504 VM_BUG_ON(PageLRU(page));
505
506redo:
507 ClearPageUnevictable(page);
508
509 if (page_evictable(page, NULL)) {
510 /*
511 * For evictable pages, we can use the cache.
512 * In event of a race, worst case is we end up with an
513 * unevictable page on [in]active list.
514 * We know how to handle that.
515 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700516 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700517 lru_cache_add_lru(page, lru);
518 } else {
519 /*
520 * Put unevictable pages directly on zone's unevictable
521 * list.
522 */
523 lru = LRU_UNEVICTABLE;
524 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700525 /*
526 * When racing with an mlock clearing (page is
527 * unlocked), make sure that if the other thread does
528 * not observe our setting of PG_lru and fails
529 * isolation, we see PG_mlocked cleared below and move
530 * the page back to the evictable list.
531 *
532 * The other side is TestClearPageMlocked().
533 */
534 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700535 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700536
537 /*
538 * page's status can change while we move it among lru. If an evictable
539 * page is on unevictable list, it never be freed. To avoid that,
540 * check after we added it to the list, again.
541 */
542 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
543 if (!isolate_lru_page(page)) {
544 put_page(page);
545 goto redo;
546 }
547 /* This means someone else dropped this page from LRU
548 * So, it will be freed or putback to LRU again. There is
549 * nothing to do here.
550 */
551 }
552
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700553 if (was_unevictable && lru != LRU_UNEVICTABLE)
554 count_vm_event(UNEVICTABLE_PGRESCUED);
555 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
556 count_vm_event(UNEVICTABLE_PGCULLED);
557
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700558 put_page(page); /* drop ref from isolate */
559}
560
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800561enum page_references {
562 PAGEREF_RECLAIM,
563 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner64574742010-03-05 13:42:22 -0800564 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800565 PAGEREF_ACTIVATE,
566};
567
568static enum page_references page_check_references(struct page *page,
569 struct scan_control *sc)
570{
Johannes Weiner64574742010-03-05 13:42:22 -0800571 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800572 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800573
Johannes Weiner64574742010-03-05 13:42:22 -0800574 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
575 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800576
577 /* Lumpy reclaim - ignore references */
578 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
579 return PAGEREF_RECLAIM;
580
581 /*
582 * Mlock lost the isolation race with us. Let try_to_unmap()
583 * move the page to the unevictable list.
584 */
585 if (vm_flags & VM_LOCKED)
586 return PAGEREF_RECLAIM;
587
Johannes Weiner64574742010-03-05 13:42:22 -0800588 if (referenced_ptes) {
589 if (PageAnon(page))
590 return PAGEREF_ACTIVATE;
591 /*
592 * All mapped pages start out with page table
593 * references from the instantiating fault, so we need
594 * to look twice if a mapped file page is used more
595 * than once.
596 *
597 * Mark it and spare it for another trip around the
598 * inactive list. Another page table reference will
599 * lead to its activation.
600 *
601 * Note: the mark is set for activated pages as well
602 * so that recently deactivated but used pages are
603 * quickly recovered.
604 */
605 SetPageReferenced(page);
606
607 if (referenced_page)
608 return PAGEREF_ACTIVATE;
609
610 return PAGEREF_KEEP;
611 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800612
613 /* Reclaim if clean, defer dirty pages to writeback */
Johannes Weiner64574742010-03-05 13:42:22 -0800614 if (referenced_page)
615 return PAGEREF_RECLAIM_CLEAN;
616
617 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800618}
619
Nick Piggine2867812008-07-25 19:45:30 -0700620/*
Andrew Morton1742f192006-03-22 00:08:21 -0800621 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 */
Andrew Morton1742f192006-03-22 00:08:21 -0800623static unsigned long shrink_page_list(struct list_head *page_list,
Andy Whitcroftc661b072007-08-22 14:01:26 -0700624 struct scan_control *sc,
625 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626{
627 LIST_HEAD(ret_pages);
628 struct pagevec freed_pvec;
629 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800630 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631
632 cond_resched();
633
634 pagevec_init(&freed_pvec, 1);
635 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800636 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 struct address_space *mapping;
638 struct page *page;
639 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640
641 cond_resched();
642
643 page = lru_to_page(page_list);
644 list_del(&page->lru);
645
Nick Piggin529ae9a2008-08-02 12:01:03 +0200646 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 goto keep;
648
Nick Piggin725d7042006-09-25 23:30:55 -0700649 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650
651 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800652
Nick Pigginb291f002008-10-18 20:26:44 -0700653 if (unlikely(!page_evictable(page, NULL)))
654 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700655
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700656 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800657 goto keep_locked;
658
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 /* Double the slab pressure for mapped and swapcache pages */
660 if (page_mapped(page) || PageSwapCache(page))
661 sc->nr_scanned++;
662
Andy Whitcroftc661b072007-08-22 14:01:26 -0700663 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
664 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
665
666 if (PageWriteback(page)) {
667 /*
668 * Synchronous reclaim is performed in two passes,
669 * first an asynchronous pass over the list to
670 * start parallel writeback, and a second synchronous
671 * pass to wait for the IO to complete. Wait here
672 * for any page for which writeback has already
673 * started.
674 */
675 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
676 wait_on_page_writeback(page);
Andrew Morton4dd4b922008-03-24 12:29:52 -0700677 else
Andy Whitcroftc661b072007-08-22 14:01:26 -0700678 goto keep_locked;
679 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800681 references = page_check_references(page, sc);
682 switch (references) {
683 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 goto activate_locked;
Johannes Weiner64574742010-03-05 13:42:22 -0800685 case PAGEREF_KEEP:
686 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800687 case PAGEREF_RECLAIM:
688 case PAGEREF_RECLAIM_CLEAN:
689 ; /* try to reclaim the page below */
690 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 /*
693 * Anonymous process memory has backing store?
694 * Try to allocate it some swap space here.
695 */
Nick Pigginb291f002008-10-18 20:26:44 -0700696 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800697 if (!(sc->gfp_mask & __GFP_IO))
698 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800699 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800701 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700702 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703
704 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705
706 /*
707 * The page is mapped into the page tables of one or more
708 * processes. Try to unmap it here.
709 */
710 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200711 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 case SWAP_FAIL:
713 goto activate_locked;
714 case SWAP_AGAIN:
715 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700716 case SWAP_MLOCK:
717 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718 case SWAP_SUCCESS:
719 ; /* try to free the page below */
720 }
721 }
722
723 if (PageDirty(page)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800724 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700726 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800728 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 goto keep_locked;
730
731 /* Page is dirty, try to write it out here */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700732 switch (pageout(page, mapping, sync_writeback)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 case PAGE_KEEP:
734 goto keep_locked;
735 case PAGE_ACTIVATE:
736 goto activate_locked;
737 case PAGE_SUCCESS:
Andrew Morton4dd4b922008-03-24 12:29:52 -0700738 if (PageWriteback(page) || PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 goto keep;
740 /*
741 * A synchronous write - probably a ramdisk. Go
742 * ahead and try to reclaim the page.
743 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200744 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 goto keep;
746 if (PageDirty(page) || PageWriteback(page))
747 goto keep_locked;
748 mapping = page_mapping(page);
749 case PAGE_CLEAN:
750 ; /* try to free the page below */
751 }
752 }
753
754 /*
755 * If the page has buffers, try to free the buffer mappings
756 * associated with this page. If we succeed we try to free
757 * the page as well.
758 *
759 * We do this even if the page is PageDirty().
760 * try_to_release_page() does not perform I/O, but it is
761 * possible for a page to have PageDirty set, but it is actually
762 * clean (all its buffers are clean). This happens if the
763 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700764 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 * try_to_release_page() will discover that cleanness and will
766 * drop the buffers and mark the page clean - it can be freed.
767 *
768 * Rarely, pages can have buffers and no ->mapping. These are
769 * the pages which were not successfully invalidated in
770 * truncate_complete_page(). We try to drop those buffers here
771 * and if that worked, and the page is no longer mapped into
772 * process address space (page_count == 1) it can be freed.
773 * Otherwise, leave the page on the LRU so it is swappable.
774 */
David Howells266cf652009-04-03 16:42:36 +0100775 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 if (!try_to_release_page(page, sc->gfp_mask))
777 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700778 if (!mapping && page_count(page) == 1) {
779 unlock_page(page);
780 if (put_page_testzero(page))
781 goto free_it;
782 else {
783 /*
784 * rare race with speculative reference.
785 * the speculative reference will free
786 * this page shortly, so we may
787 * increment nr_reclaimed here (and
788 * leave it off the LRU).
789 */
790 nr_reclaimed++;
791 continue;
792 }
793 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 }
795
Nick Piggine2867812008-07-25 19:45:30 -0700796 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800797 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798
Nick Piggina978d6f2008-10-18 20:26:58 -0700799 /*
800 * At this point, we have no other references and there is
801 * no way to pick any more up (removed from LRU, removed
802 * from pagecache). Can use non-atomic bitops now (and
803 * we obviously don't have to worry about waking up a process
804 * waiting on the page lock, because there are no references.
805 */
806 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700807free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800808 nr_reclaimed++;
Nick Piggine2867812008-07-25 19:45:30 -0700809 if (!pagevec_add(&freed_pvec, page)) {
810 __pagevec_free(&freed_pvec);
811 pagevec_reinit(&freed_pvec);
812 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 continue;
814
Nick Pigginb291f002008-10-18 20:26:44 -0700815cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800816 if (PageSwapCache(page))
817 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700818 unlock_page(page);
819 putback_lru_page(page);
820 continue;
821
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700823 /* Not a candidate for swapping, so reclaim swap space. */
824 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800825 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700826 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 SetPageActive(page);
828 pgactivate++;
829keep_locked:
830 unlock_page(page);
831keep:
832 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700833 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 }
835 list_splice(&ret_pages, page_list);
836 if (pagevec_count(&freed_pvec))
Nick Piggine2867812008-07-25 19:45:30 -0700837 __pagevec_free(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700838 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800839 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840}
841
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700842/* LRU Isolation modes. */
843#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
844#define ISOLATE_ACTIVE 1 /* Isolate active pages. */
845#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
846
847/*
848 * Attempt to remove the specified page from its LRU. Only take this page
849 * if it is of the appropriate PageActive status. Pages which are being
850 * freed elsewhere are also ignored.
851 *
852 * page: page to consider
853 * mode: one of the LRU isolation modes defined above
854 *
855 * returns 0 on success, -ve errno on failure.
856 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700857int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700858{
859 int ret = -EINVAL;
860
861 /* Only take pages on the LRU. */
862 if (!PageLRU(page))
863 return ret;
864
865 /*
866 * When checking the active state, we need to be sure we are
867 * dealing with comparible boolean values. Take the logical not
868 * of each.
869 */
870 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
871 return ret;
872
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700873 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700874 return ret;
875
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700876 /*
877 * When this function is being called for lumpy reclaim, we
878 * initially look into all LRU pages, active, inactive and
879 * unevictable; only give shrink_page_list evictable pages.
880 */
881 if (PageUnevictable(page))
882 return ret;
883
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700884 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800885
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700886 if (likely(get_page_unless_zero(page))) {
887 /*
888 * Be careful not to clear PageLRU until after we're
889 * sure the page is not being freed elsewhere -- the
890 * page release code relies on it.
891 */
892 ClearPageLRU(page);
893 ret = 0;
894 }
895
896 return ret;
897}
898
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800899/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700900 * zone->lru_lock is heavily contended. Some of the functions that
901 * shrink the lists perform better by taking out a batch of pages
902 * and working on them outside the LRU lock.
903 *
904 * For pagecache intensive workloads, this function is the hottest
905 * spot in the kernel (apart from copy_*_user functions).
906 *
907 * Appropriate locks must be held before calling this function.
908 *
909 * @nr_to_scan: The number of pages to look through on the list.
910 * @src: The LRU list to pull pages off.
911 * @dst: The temp list to put pages on to.
912 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700913 * @order: The caller's attempted allocation order
914 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700915 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700916 *
917 * returns how many pages were moved onto *@dst.
918 */
Andrew Morton69e05942006-03-22 00:08:19 -0800919static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
920 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700921 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922{
Andrew Morton69e05942006-03-22 00:08:19 -0800923 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800924 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700925
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800926 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700927 struct page *page;
928 unsigned long pfn;
929 unsigned long end_pfn;
930 unsigned long page_pfn;
931 int zone_id;
932
Linus Torvalds1da177e2005-04-16 15:20:36 -0700933 page = lru_to_page(src);
934 prefetchw_prev_lru_page(page, src, flags);
935
Nick Piggin725d7042006-09-25 23:30:55 -0700936 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800937
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700938 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700939 case 0:
940 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700941 mem_cgroup_del_lru(page);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800942 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700943 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800944
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700945 case -EBUSY:
946 /* else it is being freed elsewhere */
947 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700948 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700949 continue;
950
951 default:
952 BUG();
953 }
954
955 if (!order)
956 continue;
957
958 /*
959 * Attempt to take all pages in the order aligned region
960 * surrounding the tag page. Only take those pages of
961 * the same active state as that tag page. We may safely
962 * round the target page pfn down to the requested order
963 * as the mem_map is guarenteed valid out to MAX_ORDER,
964 * where that page is in a different zone we will detect
965 * it from its zone id and abort this block scan.
966 */
967 zone_id = page_zone_id(page);
968 page_pfn = page_to_pfn(page);
969 pfn = page_pfn & ~((1 << order) - 1);
970 end_pfn = pfn + (1 << order);
971 for (; pfn < end_pfn; pfn++) {
972 struct page *cursor_page;
973
974 /* The target page is in the block, ignore it. */
975 if (unlikely(pfn == page_pfn))
976 continue;
977
978 /* Avoid holes within the zone. */
979 if (unlikely(!pfn_valid_within(pfn)))
980 break;
981
982 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700983
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700984 /* Check that we have not crossed a zone boundary. */
985 if (unlikely(page_zone_id(cursor_page) != zone_id))
986 continue;
Minchan Kimde2e7562009-09-21 17:01:43 -0700987
988 /*
989 * If we don't have enough swap space, reclaiming of
990 * anon page which don't already have a swap slot is
991 * pointless.
992 */
993 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
994 !PageSwapCache(cursor_page))
995 continue;
996
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -0700997 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700998 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +0900999 mem_cgroup_del_lru(cursor_page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001000 nr_taken++;
1001 scan++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001002 }
1003 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 }
1005
1006 *scanned = scan;
1007 return nr_taken;
1008}
1009
Balbir Singh66e17072008-02-07 00:13:56 -08001010static unsigned long isolate_pages_global(unsigned long nr,
1011 struct list_head *dst,
1012 unsigned long *scanned, int order,
1013 int mode, struct zone *z,
1014 struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001015 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -08001016{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001017 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -08001018 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001019 lru += LRU_ACTIVE;
1020 if (file)
1021 lru += LRU_FILE;
1022 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001023 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -08001024}
1025
Linus Torvalds1da177e2005-04-16 15:20:36 -07001026/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001027 * clear_active_flags() is a helper for shrink_active_list(), clearing
1028 * any active bits from the pages in the list.
1029 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001030static unsigned long clear_active_flags(struct list_head *page_list,
1031 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001032{
1033 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001034 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001035 struct page *page;
1036
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001037 list_for_each_entry(page, page_list, lru) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07001038 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001039 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001040 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001041 ClearPageActive(page);
1042 nr_active++;
1043 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001044 count[lru]++;
1045 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001046
1047 return nr_active;
1048}
1049
Nick Piggin62695a82008-10-18 20:26:09 -07001050/**
1051 * isolate_lru_page - tries to isolate a page from its LRU list
1052 * @page: page to isolate from its LRU list
1053 *
1054 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1055 * vmstat statistic corresponding to whatever LRU list the page was on.
1056 *
1057 * Returns 0 if the page was removed from an LRU list.
1058 * Returns -EBUSY if the page was not on an LRU list.
1059 *
1060 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001061 * the active list, it will have PageActive set. If it was found on
1062 * the unevictable list, it will have the PageUnevictable bit set. That flag
1063 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001064 *
1065 * The vmstat statistic corresponding to the list on which the page was
1066 * found will be decremented.
1067 *
1068 * Restrictions:
1069 * (1) Must be called with an elevated refcount on the page. This is a
1070 * fundamentnal difference from isolate_lru_pages (which is called
1071 * without a stable reference).
1072 * (2) the lru_lock must not be held.
1073 * (3) interrupts must be enabled.
1074 */
1075int isolate_lru_page(struct page *page)
1076{
1077 int ret = -EBUSY;
1078
1079 if (PageLRU(page)) {
1080 struct zone *zone = page_zone(page);
1081
1082 spin_lock_irq(&zone->lru_lock);
1083 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001084 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001085 ret = 0;
1086 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001087
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001088 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001089 }
1090 spin_unlock_irq(&zone->lru_lock);
1091 }
1092 return ret;
1093}
1094
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001095/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001096 * Are there way too many processes in the direct reclaim path already?
1097 */
1098static int too_many_isolated(struct zone *zone, int file,
1099 struct scan_control *sc)
1100{
1101 unsigned long inactive, isolated;
1102
1103 if (current_is_kswapd())
1104 return 0;
1105
1106 if (!scanning_global_lru(sc))
1107 return 0;
1108
1109 if (file) {
1110 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1111 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1112 } else {
1113 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1114 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1115 }
1116
1117 return isolated > inactive;
1118}
1119
1120/*
Andrew Morton1742f192006-03-22 00:08:21 -08001121 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1122 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001123 */
Andrew Morton1742f192006-03-22 00:08:21 -08001124static unsigned long shrink_inactive_list(unsigned long max_scan,
Rik van Riel33c120e2008-10-18 20:26:36 -07001125 struct zone *zone, struct scan_control *sc,
1126 int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001127{
1128 LIST_HEAD(page_list);
1129 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -08001130 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001131 unsigned long nr_reclaimed = 0;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001132 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001133 int lumpy_reclaim = 0;
1134
Rik van Riel35cd7812009-09-21 17:01:38 -07001135 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001136 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001137
1138 /* We are about to die and free our memory. Return now. */
1139 if (fatal_signal_pending(current))
1140 return SWAP_CLUSTER_MAX;
1141 }
1142
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001143 /*
1144 * If we need a large contiguous chunk of memory, or have
1145 * trouble getting a small set of contiguous pages, we
1146 * will reclaim both active and inactive pages.
1147 *
1148 * We use the same threshold as pageout congestion_wait below.
1149 */
1150 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1151 lumpy_reclaim = 1;
1152 else if (sc->order && priority < DEF_PRIORITY - 2)
1153 lumpy_reclaim = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001154
1155 pagevec_init(&pvec, 1);
1156
1157 lru_add_drain();
1158 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -08001159 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001160 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -08001161 unsigned long nr_taken;
1162 unsigned long nr_scan;
1163 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001164 unsigned long nr_active;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001165 unsigned int count[NR_LRU_LISTS] = { 0, };
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001166 int mode = lumpy_reclaim ? ISOLATE_BOTH : ISOLATE_INACTIVE;
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001167 unsigned long nr_anon;
1168 unsigned long nr_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001170 nr_taken = sc->isolate_pages(SWAP_CLUSTER_MAX,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001171 &page_list, &nr_scan, sc->order, mode,
1172 zone, sc->mem_cgroup, 0, file);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001173
1174 if (scanning_global_lru(sc)) {
1175 zone->pages_scanned += nr_scan;
1176 if (current_is_kswapd())
1177 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1178 nr_scan);
1179 else
1180 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1181 nr_scan);
1182 }
1183
1184 if (nr_taken == 0)
1185 goto done;
1186
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001187 nr_active = clear_active_flags(&page_list, count);
Andy Whitcrofte9187bd2007-08-22 14:01:25 -07001188 __count_vm_events(PGDEACTIVATE, nr_active);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001189
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001190 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1191 -count[LRU_ACTIVE_FILE]);
1192 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1193 -count[LRU_INACTIVE_FILE]);
1194 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1195 -count[LRU_ACTIVE_ANON]);
1196 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1197 -count[LRU_INACTIVE_ANON]);
1198
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001199 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1200 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1201 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1202 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001203
Huang Shijie62c0c2f2009-12-14 17:59:48 -08001204 reclaim_stat->recent_scanned[0] += nr_anon;
1205 reclaim_stat->recent_scanned[1] += nr_file;
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001206
Linus Torvalds1da177e2005-04-16 15:20:36 -07001207 spin_unlock_irq(&zone->lru_lock);
1208
Andrew Morton69e05942006-03-22 00:08:19 -08001209 nr_scanned += nr_scan;
Andy Whitcroftc661b072007-08-22 14:01:26 -07001210 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1211
1212 /*
1213 * If we are direct reclaiming for contiguous pages and we do
1214 * not reclaim everything in the list, try again and wait
1215 * for IO to complete. This will stall high-order allocations
1216 * but that should be acceptable to the caller
1217 */
1218 if (nr_freed < nr_taken && !current_is_kswapd() &&
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001219 lumpy_reclaim) {
Jens Axboe8aa7e842009-07-09 14:52:32 +02001220 congestion_wait(BLK_RW_ASYNC, HZ/10);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001221
1222 /*
1223 * The attempt at page out may have made some
1224 * of the pages active, mark them inactive again.
1225 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001226 nr_active = clear_active_flags(&page_list, count);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001227 count_vm_events(PGDEACTIVATE, nr_active);
1228
1229 nr_freed += shrink_page_list(&page_list, sc,
1230 PAGEOUT_IO_SYNC);
1231 }
1232
Andrew Morton05ff5132006-03-22 00:08:20 -08001233 nr_reclaimed += nr_freed;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001234
Nick Piggina74609f2006-01-06 00:11:20 -08001235 local_irq_disable();
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001236 if (current_is_kswapd())
Christoph Lameterf8891e52006-06-30 01:55:45 -07001237 __count_vm_events(KSWAPD_STEAL, nr_freed);
Shantanu Goel918d3f92006-12-29 16:48:59 -08001238 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -08001239
1240 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001241 /*
1242 * Put back any unfreeable pages.
1243 */
1244 while (!list_empty(&page_list)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001245 int lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001246 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -07001247 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001248 list_del(&page->lru);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001249 if (unlikely(!page_evictable(page, NULL))) {
1250 spin_unlock_irq(&zone->lru_lock);
1251 putback_lru_page(page);
1252 spin_lock_irq(&zone->lru_lock);
1253 continue;
1254 }
1255 SetPageLRU(page);
1256 lru = page_lru(page);
1257 add_page_to_lru_list(zone, page, lru);
KOSAKI Motohiro74a1c482009-09-21 17:01:45 -07001258 if (is_active_lru(lru)) {
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001259 int file = is_file_lru(lru);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001260 reclaim_stat->recent_rotated[file]++;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001261 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001262 if (!pagevec_add(&pvec, page)) {
1263 spin_unlock_irq(&zone->lru_lock);
1264 __pagevec_release(&pvec);
1265 spin_lock_irq(&zone->lru_lock);
1266 }
1267 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001268 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1269 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1270
Andrew Morton69e05942006-03-22 00:08:19 -08001271 } while (nr_scanned < max_scan);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001272
Linus Torvalds1da177e2005-04-16 15:20:36 -07001273done:
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001274 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001275 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -08001276 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001277}
1278
Martin Bligh3bb1a852006-10-28 10:38:24 -07001279/*
1280 * We are about to scan this zone at a certain priority level. If that priority
1281 * level is smaller (ie: more urgent) than the previous priority, then note
1282 * that priority level within the zone. This is done so that when the next
1283 * process comes in to scan this zone, it will immediately start out at this
1284 * priority level rather than having to build up its own scanning priority.
1285 * Here, this priority affects only the reclaim-mapped threshold.
1286 */
1287static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1288{
1289 if (priority < zone->prev_priority)
1290 zone->prev_priority = priority;
1291}
1292
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001293/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 * This moves pages from the active list to the inactive list.
1295 *
1296 * We move them the other way if the page is referenced by one or more
1297 * processes, from rmap.
1298 *
1299 * If the pages are mostly unmapped, the processing is fast and it is
1300 * appropriate to hold zone->lru_lock across the whole operation. But if
1301 * the pages are mapped, the processing is slow (page_referenced()) so we
1302 * should drop zone->lru_lock around each page. It's impossible to balance
1303 * this, so instead we remove the pages from the LRU while processing them.
1304 * It is safe to rely on PG_active against the non-LRU pages in here because
1305 * nobody will play with that bit on a non-LRU page.
1306 *
1307 * The downside is that we have to touch page->_count against each page.
1308 * But we had to alter page->flags anyway.
1309 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001310
Wu Fengguang3eb41402009-06-16 15:33:13 -07001311static void move_active_pages_to_lru(struct zone *zone,
1312 struct list_head *list,
1313 enum lru_list lru)
1314{
1315 unsigned long pgmoved = 0;
1316 struct pagevec pvec;
1317 struct page *page;
1318
1319 pagevec_init(&pvec, 1);
1320
1321 while (!list_empty(list)) {
1322 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001323
1324 VM_BUG_ON(PageLRU(page));
1325 SetPageLRU(page);
1326
Wu Fengguang3eb41402009-06-16 15:33:13 -07001327 list_move(&page->lru, &zone->lru[lru].list);
1328 mem_cgroup_add_lru_list(page, lru);
1329 pgmoved++;
1330
1331 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1332 spin_unlock_irq(&zone->lru_lock);
1333 if (buffer_heads_over_limit)
1334 pagevec_strip(&pvec);
1335 __pagevec_release(&pvec);
1336 spin_lock_irq(&zone->lru_lock);
1337 }
1338 }
1339 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1340 if (!is_active_lru(lru))
1341 __count_vm_events(PGDEACTIVATE, pgmoved);
1342}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001343
Andrew Morton1742f192006-03-22 00:08:21 -08001344static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001345 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001347 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001348 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001349 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001351 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001352 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001353 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001354 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001355 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001356
1357 lru_add_drain();
1358 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001359 nr_taken = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
Balbir Singh66e17072008-02-07 00:13:56 -08001360 ISOLATE_ACTIVE, zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001361 sc->mem_cgroup, 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001362 /*
1363 * zone->pages_scanned is used for detect zone's oom
1364 * mem_cgroup remembers nr_scan by itself.
1365 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001366 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001367 zone->pages_scanned += pgscanned;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001368 }
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001369 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001370
Wu Fengguang3eb41402009-06-16 15:33:13 -07001371 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001372 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001373 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001374 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001375 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001376 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001377 spin_unlock_irq(&zone->lru_lock);
1378
Linus Torvalds1da177e2005-04-16 15:20:36 -07001379 while (!list_empty(&l_hold)) {
1380 cond_resched();
1381 page = lru_to_page(&l_hold);
1382 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001383
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001384 if (unlikely(!page_evictable(page, NULL))) {
1385 putback_lru_page(page);
1386 continue;
1387 }
1388
Johannes Weiner64574742010-03-05 13:42:22 -08001389 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001390 nr_rotated++;
Wu Fengguang8cab4752009-06-16 15:33:12 -07001391 /*
1392 * Identify referenced, file-backed active pages and
1393 * give them one more trip around the active list. So
1394 * that executable code get better chances to stay in
1395 * memory under moderate memory pressure. Anon pages
1396 * are not likely to be evicted by use-once streaming
1397 * IO, plus JVM can create lots of anon VM_EXEC pages,
1398 * so we ignore them here.
1399 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001400 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001401 list_add(&page->lru, &l_active);
1402 continue;
1403 }
1404 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001405
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001406 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407 list_add(&page->lru, &l_inactive);
1408 }
1409
Andrew Mortonb5557492009-01-06 14:40:13 -08001410 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001411 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001412 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001413 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001414 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001415 * Count referenced pages from currently used mappings as rotated,
1416 * even though only some of them are actually re-activated. This
1417 * helps balance scan pressure between file and anonymous pages in
1418 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001419 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001420 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001421
Wu Fengguang3eb41402009-06-16 15:33:13 -07001422 move_active_pages_to_lru(zone, &l_active,
1423 LRU_ACTIVE + file * LRU_FILE);
1424 move_active_pages_to_lru(zone, &l_inactive,
1425 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001426 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001427 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428}
1429
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001430static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001431{
1432 unsigned long active, inactive;
1433
1434 active = zone_page_state(zone, NR_ACTIVE_ANON);
1435 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1436
1437 if (inactive * zone->inactive_ratio < active)
1438 return 1;
1439
1440 return 0;
1441}
1442
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001443/**
1444 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1445 * @zone: zone to check
1446 * @sc: scan control of this context
1447 *
1448 * Returns true if the zone does not have enough inactive anon pages,
1449 * meaning some active anon pages need to be deactivated.
1450 */
1451static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1452{
1453 int low;
1454
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001455 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001456 low = inactive_anon_is_low_global(zone);
1457 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001458 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001459 return low;
1460}
1461
Rik van Riel56e49d22009-06-16 15:32:28 -07001462static int inactive_file_is_low_global(struct zone *zone)
1463{
1464 unsigned long active, inactive;
1465
1466 active = zone_page_state(zone, NR_ACTIVE_FILE);
1467 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1468
1469 return (active > inactive);
1470}
1471
1472/**
1473 * inactive_file_is_low - check if file pages need to be deactivated
1474 * @zone: zone to check
1475 * @sc: scan control of this context
1476 *
1477 * When the system is doing streaming IO, memory pressure here
1478 * ensures that active file pages get deactivated, until more
1479 * than half of the file pages are on the inactive list.
1480 *
1481 * Once we get to that situation, protect the system's working
1482 * set from being evicted by disabling active file page aging.
1483 *
1484 * This uses a different ratio than the anonymous pages, because
1485 * the page cache uses a use-once replacement algorithm.
1486 */
1487static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1488{
1489 int low;
1490
1491 if (scanning_global_lru(sc))
1492 low = inactive_file_is_low_global(zone);
1493 else
1494 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1495 return low;
1496}
1497
Rik van Rielb39415b2009-12-14 17:59:48 -08001498static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1499 int file)
1500{
1501 if (file)
1502 return inactive_file_is_low(zone, sc);
1503 else
1504 return inactive_anon_is_low(zone, sc);
1505}
1506
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001507static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001508 struct zone *zone, struct scan_control *sc, int priority)
1509{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001510 int file = is_file_lru(lru);
1511
Rik van Rielb39415b2009-12-14 17:59:48 -08001512 if (is_active_lru(lru)) {
1513 if (inactive_list_is_low(zone, sc, file))
1514 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001515 return 0;
1516 }
1517
Rik van Riel33c120e2008-10-18 20:26:36 -07001518 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001519}
1520
Linus Torvalds1da177e2005-04-16 15:20:36 -07001521/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001522 * Determine how aggressively the anon and file LRU lists should be
1523 * scanned. The relative value of each set of LRU lists is determined
1524 * by looking at the fraction of the pages scanned we did rotate back
1525 * onto the active list instead of evict.
1526 *
1527 * percent[0] specifies how much pressure to put on ram/swap backed
1528 * memory, while percent[1] determines pressure on the file LRUs.
1529 */
1530static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1531 unsigned long *percent)
1532{
1533 unsigned long anon, file, free;
1534 unsigned long anon_prio, file_prio;
1535 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001536 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001537
KOSAKI Motohiro84b18492010-03-05 13:41:47 -08001538 /* If we have no swap space, do not bother scanning anon pages. */
1539 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1540 percent[0] = 0;
1541 percent[1] = 100;
1542 return;
1543 }
1544
Vincent Li0b217672009-09-21 17:03:09 -07001545 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1546 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1547 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1548 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001549
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001550 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001551 free = zone_page_state(zone, NR_FREE_PAGES);
1552 /* If we have very few page cache pages,
1553 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001554 if (unlikely(file + free <= high_wmark_pages(zone))) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001555 percent[0] = 100;
1556 percent[1] = 0;
1557 return;
1558 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001559 }
1560
1561 /*
1562 * OK, so we have swap space and a fair amount of page cache
1563 * pages. We use the recently rotated / recently scanned
1564 * ratios to determine how valuable each cache is.
1565 *
1566 * Because workloads change over time (and to avoid overflow)
1567 * we keep these statistics as a floating average, which ends
1568 * up weighing recent references more than old ones.
1569 *
1570 * anon in [0], file in [1]
1571 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001572 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001573 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001574 reclaim_stat->recent_scanned[0] /= 2;
1575 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001576 spin_unlock_irq(&zone->lru_lock);
1577 }
1578
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001579 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001580 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001581 reclaim_stat->recent_scanned[1] /= 2;
1582 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001583 spin_unlock_irq(&zone->lru_lock);
1584 }
1585
1586 /*
1587 * With swappiness at 100, anonymous and file have the same priority.
1588 * This scanning priority is essentially the inverse of IO cost.
1589 */
1590 anon_prio = sc->swappiness;
1591 file_prio = 200 - sc->swappiness;
1592
1593 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001594 * The amount of pressure on anon vs file pages is inversely
1595 * proportional to the fraction of recently scanned pages on
1596 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001597 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001598 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1599 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001600
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001601 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1602 fp /= reclaim_stat->recent_rotated[1] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001603
1604 /* Normalize to percentages */
1605 percent[0] = 100 * ap / (ap + fp + 1);
1606 percent[1] = 100 - percent[0];
1607}
1608
Wu Fengguang6e08a362009-06-16 15:32:29 -07001609/*
1610 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1611 * until we collected @swap_cluster_max pages to scan.
1612 */
1613static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001614 unsigned long *nr_saved_scan)
Wu Fengguang6e08a362009-06-16 15:32:29 -07001615{
1616 unsigned long nr;
1617
1618 *nr_saved_scan += nr_to_scan;
1619 nr = *nr_saved_scan;
1620
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001621 if (nr >= SWAP_CLUSTER_MAX)
Wu Fengguang6e08a362009-06-16 15:32:29 -07001622 *nr_saved_scan = 0;
1623 else
1624 nr = 0;
1625
1626 return nr;
1627}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001628
1629/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001630 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1631 */
Rik van Riela79311c2009-01-06 14:40:01 -08001632static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001633 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001634{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001635 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001636 unsigned long nr_to_scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001637 unsigned long percent[2]; /* anon @ 0; file @ 1 */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001638 enum lru_list l;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001639 unsigned long nr_reclaimed = sc->nr_reclaimed;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001640 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Wu Fengguangf8629632009-09-21 17:03:11 -07001641 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642
KOSAKI Motohiro84b18492010-03-05 13:41:47 -08001643 get_scan_ratio(zone, sc, percent);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001644
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001645 for_each_evictable_lru(l) {
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001646 int file = is_file_lru(l);
Andrew Morton8713e012009-04-30 15:08:55 -07001647 unsigned long scan;
Johannes Weinere0f79b82008-10-18 20:26:55 -07001648
KOSAKI Motohiro84b18492010-03-05 13:41:47 -08001649 if (percent[file] == 0) {
1650 nr[l] = 0;
1651 continue;
1652 }
1653
Vincent Li0b217672009-09-21 17:03:09 -07001654 scan = zone_nr_lru_pages(zone, sc, l);
KOSAKI Motohiro84b18492010-03-05 13:41:47 -08001655 if (priority) {
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001656 scan >>= priority;
1657 scan = (scan * percent[file]) / 100;
1658 }
Wu Fengguangf8629632009-09-21 17:03:11 -07001659 nr[l] = nr_scan_try_batch(scan,
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001660 &reclaim_stat->nr_saved_scan[l]);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001661 }
1662
Rik van Riel556adec2008-10-18 20:26:34 -07001663 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1664 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001665 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001666 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001667 nr_to_scan = min_t(unsigned long,
1668 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001669 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001671 nr_reclaimed += shrink_list(l, nr_to_scan,
1672 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001673 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001674 }
Rik van Riela79311c2009-01-06 14:40:01 -08001675 /*
1676 * On large memory systems, scan >> priority can become
1677 * really large. This is fine for the starting priority;
1678 * we want to put equal scanning pressure on each zone.
1679 * However, if the VM has a harder time of freeing pages,
1680 * with multiple processes reclaiming pages, the total
1681 * freeing target can get unreasonably large.
1682 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08001683 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001684 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 }
1686
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001687 sc->nr_reclaimed = nr_reclaimed;
1688
Rik van Riel556adec2008-10-18 20:26:34 -07001689 /*
1690 * Even if we did not try to evict anon pages at all, we want to
1691 * rebalance the anon lru active/inactive ratio.
1692 */
MinChan Kim69c85482009-06-16 15:32:44 -07001693 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
Rik van Riel556adec2008-10-18 20:26:34 -07001694 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1695
Andrew Morton232ea4d2007-02-28 20:13:21 -08001696 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697}
1698
1699/*
1700 * This is the direct reclaim path, for page-allocating processes. We only
1701 * try to reclaim pages from zones which will satisfy the caller's allocation
1702 * request.
1703 *
Mel Gorman41858962009-06-16 15:32:12 -07001704 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1705 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001706 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1707 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001708 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1709 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1710 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712 * If a zone is deemed to be full of pinned pages then just give it a light
1713 * scan then give up on it.
1714 */
Rik van Riela79311c2009-01-06 14:40:01 -08001715static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001716 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001717{
Mel Gorman54a6eb52008-04-28 02:12:16 -07001718 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Mel Gormandd1a2392008-04-28 02:12:17 -07001719 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001720 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001721
Nick Piggin408d8542006-09-25 23:31:27 -07001722 sc->all_unreclaimable = 1;
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001723 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1724 sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001725 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001727 /*
1728 * Take care memory controller reclaiming has small influence
1729 * to global LRU.
1730 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001731 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001732 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1733 continue;
1734 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08001736 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001737 continue; /* Let kswapd poll it */
1738 sc->all_unreclaimable = 0;
1739 } else {
1740 /*
1741 * Ignore cpuset limitation here. We just want to reduce
1742 * # of used pages by us regardless of memory shortage.
1743 */
1744 sc->all_unreclaimable = 0;
1745 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1746 priority);
1747 }
Nick Piggin408d8542006-09-25 23:31:27 -07001748
Rik van Riela79311c2009-01-06 14:40:01 -08001749 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001750 }
1751}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001752
Linus Torvalds1da177e2005-04-16 15:20:36 -07001753/*
1754 * This is the main entry point to direct page reclaim.
1755 *
1756 * If a full scan of the inactive list fails to free enough memory then we
1757 * are "out of memory" and something needs to be killed.
1758 *
1759 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1760 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02001761 * caller can't do much about. We kick the writeback threads and take explicit
1762 * naps in the hope that some of these pages can be written. But if the
1763 * allocating task holds filesystem locks which prevent writeout this might not
1764 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001765 *
1766 * returns: 0, if no pages reclaimed
1767 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001768 */
Mel Gormandac1d272008-04-28 02:12:12 -07001769static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001770 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001771{
1772 int priority;
kosaki.motohiro@jp.fujitsu.comc700be32008-06-12 15:21:27 -07001773 unsigned long ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001774 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001775 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001776 unsigned long lru_pages = 0;
Mel Gormandd1a2392008-04-28 02:12:17 -07001777 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001778 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -07001779 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001780 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781
Keika Kobayashi873b4772008-07-25 01:48:52 -07001782 delayacct_freepages_start();
1783
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001784 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001785 count_vm_event(ALLOCSTALL);
1786 /*
1787 * mem_cgroup will not do shrink_slab.
1788 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001789 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001790 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001792 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1793 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001794
Wu Fengguangadea02a2009-09-21 17:01:42 -07001795 lru_pages += zone_reclaimable_pages(zone);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001796 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001797 }
1798
1799 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001800 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001801 if (!priority)
1802 disable_swap_token();
Rik van Riela79311c2009-01-06 14:40:01 -08001803 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001804 /*
1805 * Don't shrink slabs when reclaiming memory from
1806 * over limit cgroups
1807 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001808 if (scanning_global_lru(sc)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07001809 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001810 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08001811 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001812 reclaim_state->reclaimed_slab = 0;
1813 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814 }
Balbir Singh66e17072008-02-07 00:13:56 -08001815 total_scanned += sc->nr_scanned;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001816 if (sc->nr_reclaimed >= sc->nr_to_reclaim) {
Rik van Riela79311c2009-01-06 14:40:01 -08001817 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818 goto out;
1819 }
1820
1821 /*
1822 * Try to write back as many pages as we just scanned. This
1823 * tends to cause slow streaming writers to write data to the
1824 * disk smoothly, at the dirtying rate, which is nice. But
1825 * that's undesirable in laptop mode, where we *want* lumpy
1826 * writeout. So in laptop mode, write out the whole world.
1827 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001828 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1829 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001830 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001831 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 }
1833
1834 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08001835 if (!sc->hibernation_mode && sc->nr_scanned &&
1836 priority < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02001837 congestion_wait(BLK_RW_ASYNC, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001838 }
Fernando Luis Vazquez Cao87547ee2008-07-29 22:33:42 -07001839 /* top priority shrink_zones still had more to do? don't OOM, then */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001840 if (!sc->all_unreclaimable && scanning_global_lru(sc))
Rik van Riela79311c2009-01-06 14:40:01 -08001841 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001842out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07001843 /*
1844 * Now that we've scanned all the zones at this priority level, note
1845 * that level within the zone so that the next thread which performs
1846 * scanning of this zone will immediately start out at this priority
1847 * level. This affects only the decision whether or not to bring
1848 * mapped pages onto the inactive list.
1849 */
1850 if (priority < 0)
1851 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001853 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001854 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001856 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1857 continue;
1858
1859 zone->prev_priority = priority;
1860 }
1861 } else
1862 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1863
Keika Kobayashi873b4772008-07-25 01:48:52 -07001864 delayacct_freepages_end();
1865
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866 return ret;
1867}
1868
Mel Gormandac1d272008-04-28 02:12:12 -07001869unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001870 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08001871{
1872 struct scan_control sc = {
1873 .gfp_mask = gfp_mask,
1874 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001875 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001876 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001877 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08001878 .swappiness = vm_swappiness,
1879 .order = order,
1880 .mem_cgroup = NULL,
1881 .isolate_pages = isolate_pages_global,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001882 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08001883 };
1884
Mel Gormandd1a2392008-04-28 02:12:17 -07001885 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001886}
1887
Balbir Singh00f0b822008-03-04 14:28:39 -08001888#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08001889
Balbir Singh4e416952009-09-23 15:56:39 -07001890unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1891 gfp_t gfp_mask, bool noswap,
1892 unsigned int swappiness,
1893 struct zone *zone, int nid)
1894{
1895 struct scan_control sc = {
1896 .may_writepage = !laptop_mode,
1897 .may_unmap = 1,
1898 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07001899 .swappiness = swappiness,
1900 .order = 0,
1901 .mem_cgroup = mem,
1902 .isolate_pages = mem_cgroup_isolate_pages,
1903 };
1904 nodemask_t nm = nodemask_of_node(nid);
1905
1906 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1907 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1908 sc.nodemask = &nm;
1909 sc.nr_reclaimed = 0;
1910 sc.nr_scanned = 0;
1911 /*
1912 * NOTE: Although we can get the priority field, using it
1913 * here is not a good idea, since it limits the pages we can scan.
1914 * if we don't reclaim here, the shrink_zone from balance_pgdat
1915 * will pick up pages from other mem cgroup's as well. We hack
1916 * the priority and make it zero.
1917 */
1918 shrink_zone(0, zone, &sc);
1919 return sc.nr_reclaimed;
1920}
1921
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001922unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001923 gfp_t gfp_mask,
1924 bool noswap,
1925 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08001926{
Balbir Singh4e416952009-09-23 15:56:39 -07001927 struct zonelist *zonelist;
Balbir Singh66e17072008-02-07 00:13:56 -08001928 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08001929 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001930 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001931 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001932 .nr_to_reclaim = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001933 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08001934 .order = 0,
1935 .mem_cgroup = mem_cont,
1936 .isolate_pages = mem_cgroup_isolate_pages,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001937 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08001938 };
Balbir Singh66e17072008-02-07 00:13:56 -08001939
Mel Gormandd1a2392008-04-28 02:12:17 -07001940 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1941 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1942 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1943 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001944}
1945#endif
1946
Mel Gormanf50de2d2009-12-14 17:58:53 -08001947/* is kswapd sleeping prematurely? */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001948static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
Mel Gormanf50de2d2009-12-14 17:58:53 -08001949{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001950 int i;
Mel Gormanf50de2d2009-12-14 17:58:53 -08001951
1952 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1953 if (remaining)
1954 return 1;
1955
1956 /* If after HZ/10, a zone is below the high mark, it's premature */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001957 for (i = 0; i < pgdat->nr_zones; i++) {
1958 struct zone *zone = pgdat->node_zones + i;
1959
1960 if (!populated_zone(zone))
1961 continue;
1962
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08001963 if (zone->all_unreclaimable)
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08001964 continue;
1965
Mel Gormanf50de2d2009-12-14 17:58:53 -08001966 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
1967 0, 0))
1968 return 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001969 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08001970
1971 return 0;
1972}
1973
Linus Torvalds1da177e2005-04-16 15:20:36 -07001974/*
1975 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07001976 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001977 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001978 * Returns the number of pages which were actually freed.
1979 *
1980 * There is special handling here for zones which are full of pinned pages.
1981 * This can happen if the pages are all mlocked, or if they are all used by
1982 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1983 * What we do is to detect the case where all pages in the zone have been
1984 * scanned twice and there has been zero successful reclaim. Mark the zone as
1985 * dead and from now on, only perform a short scan. Basically we're polling
1986 * the zone for when the problem goes away.
1987 *
1988 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07001989 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1990 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
1991 * lower zones regardless of the number of free pages in the lower zones. This
1992 * interoperates with the page allocator fallback scheme to ensure that aging
1993 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001995static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997 int all_zones_ok;
1998 int priority;
1999 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08002000 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002001 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08002002 struct scan_control sc = {
2003 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002004 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002005 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002006 /*
2007 * kswapd doesn't want to be bailed out while reclaim. because
2008 * we want to put equal scanning pressure on each zone.
2009 */
2010 .nr_to_reclaim = ULONG_MAX,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002011 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002012 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002013 .mem_cgroup = NULL,
2014 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08002015 };
Martin Bligh3bb1a852006-10-28 10:38:24 -07002016 /*
2017 * temp_priority is used to remember the scanning priority at which
Mel Gorman41858962009-06-16 15:32:12 -07002018 * this zone was successfully refilled to
2019 * free_pages == high_wmark_pages(zone).
Martin Bligh3bb1a852006-10-28 10:38:24 -07002020 */
2021 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022
2023loop_again:
2024 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002025 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002026 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002027 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002028
Martin Bligh3bb1a852006-10-28 10:38:24 -07002029 for (i = 0; i < pgdat->nr_zones; i++)
2030 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002031
2032 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2033 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2034 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002035 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002037 /* The swap token gets in the way of swapout... */
2038 if (!priority)
2039 disable_swap_token();
2040
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041 all_zones_ok = 1;
2042
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002043 /*
2044 * Scan in the highmem->dma direction for the highest
2045 * zone which needs scanning
2046 */
2047 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2048 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002050 if (!populated_zone(zone))
2051 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002052
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002053 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002054 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002055
Rik van Riel556adec2008-10-18 20:26:34 -07002056 /*
2057 * Do some background aging of the anon list, to give
2058 * pages a chance to be referenced before reclaiming.
2059 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002060 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002061 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2062 &sc, priority, 0);
2063
Mel Gorman41858962009-06-16 15:32:12 -07002064 if (!zone_watermark_ok(zone, order,
2065 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002066 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002067 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002070 if (i < 0)
2071 goto out;
2072
Linus Torvalds1da177e2005-04-16 15:20:36 -07002073 for (i = 0; i <= end_zone; i++) {
2074 struct zone *zone = pgdat->node_zones + i;
2075
Wu Fengguangadea02a2009-09-21 17:01:42 -07002076 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002077 }
2078
2079 /*
2080 * Now scan the zone in the dma->highmem direction, stopping
2081 * at the last zone which needs scanning.
2082 *
2083 * We do this because the page allocator works in the opposite
2084 * direction. This prevents the page allocator from allocating
2085 * pages behind kswapd's direction of progress, which would
2086 * cause too much scanning of the lower zones.
2087 */
2088 for (i = 0; i <= end_zone; i++) {
2089 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002090 int nr_slab;
Balbir Singh4e416952009-09-23 15:56:39 -07002091 int nid, zid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092
Con Kolivasf3fe6512006-01-06 00:11:15 -08002093 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002094 continue;
2095
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002096 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097 continue;
2098
Martin Bligh3bb1a852006-10-28 10:38:24 -07002099 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002100 sc.nr_scanned = 0;
Martin Bligh3bb1a852006-10-28 10:38:24 -07002101 note_zone_scanning_priority(zone, priority);
Balbir Singh4e416952009-09-23 15:56:39 -07002102
2103 nid = pgdat->node_id;
2104 zid = zone_idx(zone);
2105 /*
2106 * Call soft limit reclaim before calling shrink_zone.
2107 * For now we ignore the return value
2108 */
2109 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2110 nid, zid);
Rik van Riel32a43302007-10-16 01:24:50 -07002111 /*
2112 * We put equal pressure on every zone, unless one
2113 * zone has way too many pages free already.
2114 */
Mel Gorman41858962009-06-16 15:32:12 -07002115 if (!zone_watermark_ok(zone, order,
2116 8*high_wmark_pages(zone), end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002117 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002119 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2120 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002121 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002122 total_scanned += sc.nr_scanned;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002123 if (zone->all_unreclaimable)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002125 if (nr_slab == 0 &&
2126 zone->pages_scanned >= (zone_reclaimable_pages(zone) * 6))
2127 zone->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 /*
2129 * If we've done a decent amount of scanning and
2130 * the reclaim ratio is low, start doing writepage
2131 * even in laptop mode
2132 */
2133 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002134 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002136
Minchan Kim45973d72010-03-05 13:41:45 -08002137 if (!zone_watermark_ok(zone, order,
2138 high_wmark_pages(zone), end_zone, 0)) {
2139 all_zones_ok = 0;
2140 /*
2141 * We are still under min water mark. This
2142 * means that we have a GFP_ATOMIC allocation
2143 * failure risk. Hurry up!
2144 */
2145 if (!zone_watermark_ok(zone, order,
2146 min_wmark_pages(zone), end_zone, 0))
2147 has_under_min_watermark_zone = 1;
2148 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002149
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151 if (all_zones_ok)
2152 break; /* kswapd: all done */
2153 /*
2154 * OK, kswapd is getting into trouble. Take a nap, then take
2155 * another pass across the zones.
2156 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002157 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2158 if (has_under_min_watermark_zone)
2159 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2160 else
2161 congestion_wait(BLK_RW_ASYNC, HZ/10);
2162 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163
2164 /*
2165 * We do this so kswapd doesn't build up large priorities for
2166 * example when it is freeing in parallel with allocators. It
2167 * matches the direct reclaim path behaviour in terms of impact
2168 * on zone->*_priority.
2169 */
Rik van Riela79311c2009-01-06 14:40:01 -08002170 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171 break;
2172 }
2173out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07002174 /*
2175 * Note within each zone the priority level at which this zone was
2176 * brought into a happy state. So that the next thread which scans this
2177 * zone will start out at that priority level.
2178 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002179 for (i = 0; i < pgdat->nr_zones; i++) {
2180 struct zone *zone = pgdat->node_zones + i;
2181
Martin Bligh3bb1a852006-10-28 10:38:24 -07002182 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002183 }
2184 if (!all_zones_ok) {
2185 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002186
2187 try_to_freeze();
2188
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002189 /*
2190 * Fragmentation may mean that the system cannot be
2191 * rebalanced for high-order allocations in all zones.
2192 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2193 * it means the zones have been fully scanned and are still
2194 * not balanced. For high-order allocations, there is
2195 * little point trying all over again as kswapd may
2196 * infinite loop.
2197 *
2198 * Instead, recheck all watermarks at order-0 as they
2199 * are the most important. If watermarks are ok, kswapd will go
2200 * back to sleep. High-order users can still perform direct
2201 * reclaim if they wish.
2202 */
2203 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2204 order = sc.order = 0;
2205
Linus Torvalds1da177e2005-04-16 15:20:36 -07002206 goto loop_again;
2207 }
2208
Rik van Riela79311c2009-01-06 14:40:01 -08002209 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210}
2211
2212/*
2213 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002214 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002215 *
2216 * This basically trickles out pages so that we have _some_
2217 * free memory available even if there is no other activity
2218 * that frees anything up. This is needed for things like routing
2219 * etc, where we otherwise might have all activity going on in
2220 * asynchronous contexts that cannot page things out.
2221 *
2222 * If there are applications that are active memory-allocators
2223 * (most normal use), this basically shouldn't matter.
2224 */
2225static int kswapd(void *p)
2226{
2227 unsigned long order;
2228 pg_data_t *pgdat = (pg_data_t*)p;
2229 struct task_struct *tsk = current;
2230 DEFINE_WAIT(wait);
2231 struct reclaim_state reclaim_state = {
2232 .reclaimed_slab = 0,
2233 };
Rusty Russella70f7302009-03-13 14:49:46 +10302234 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235
Nick Piggincf40bd12009-01-21 08:12:39 +01002236 lockdep_set_current_reclaim_state(GFP_KERNEL);
2237
Rusty Russell174596a2009-01-01 10:12:29 +10302238 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002239 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002240 current->reclaim_state = &reclaim_state;
2241
2242 /*
2243 * Tell the memory management that we're a "memory allocator",
2244 * and that if we need more memory we should get access to it
2245 * regardless (see "__alloc_pages()"). "kswapd" should
2246 * never get caught in the normal page freeing logic.
2247 *
2248 * (Kswapd normally doesn't need memory anyway, but sometimes
2249 * you need a small amount of memory in order to be able to
2250 * page out something else, and this flag essentially protects
2251 * us from recursively trying to free more memory as we're
2252 * trying to free the first piece of memory in the first place).
2253 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002254 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002255 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002256
2257 order = 0;
2258 for ( ; ; ) {
2259 unsigned long new_order;
David Rientjes8fe23e02009-12-14 17:58:33 -08002260 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002261
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2263 new_order = pgdat->kswapd_max_order;
2264 pgdat->kswapd_max_order = 0;
2265 if (order < new_order) {
2266 /*
2267 * Don't sleep if someone wants a larger 'order'
2268 * allocation
2269 */
2270 order = new_order;
2271 } else {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002272 if (!freezing(current) && !kthread_should_stop()) {
2273 long remaining = 0;
2274
2275 /* Try to sleep for a short interval */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002276 if (!sleeping_prematurely(pgdat, order, remaining)) {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002277 remaining = schedule_timeout(HZ/10);
2278 finish_wait(&pgdat->kswapd_wait, &wait);
2279 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2280 }
2281
2282 /*
2283 * After a short sleep, check if it was a
2284 * premature sleep. If not, then go fully
2285 * to sleep until explicitly woken up
2286 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002287 if (!sleeping_prematurely(pgdat, order, remaining))
Mel Gormanf50de2d2009-12-14 17:58:53 -08002288 schedule();
2289 else {
2290 if (remaining)
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002291 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002292 else
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002293 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002294 }
2295 }
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002296
Linus Torvalds1da177e2005-04-16 15:20:36 -07002297 order = pgdat->kswapd_max_order;
2298 }
2299 finish_wait(&pgdat->kswapd_wait, &wait);
2300
David Rientjes8fe23e02009-12-14 17:58:33 -08002301 ret = try_to_freeze();
2302 if (kthread_should_stop())
2303 break;
2304
2305 /*
2306 * We can speed up thawing tasks if we don't call balance_pgdat
2307 * after returning from the refrigerator
2308 */
2309 if (!ret)
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002310 balance_pgdat(pgdat, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002311 }
2312 return 0;
2313}
2314
2315/*
2316 * A zone is low on free memory, so wake its kswapd task to service it.
2317 */
2318void wakeup_kswapd(struct zone *zone, int order)
2319{
2320 pg_data_t *pgdat;
2321
Con Kolivasf3fe6512006-01-06 00:11:15 -08002322 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 return;
2324
2325 pgdat = zone->zone_pgdat;
Mel Gorman41858962009-06-16 15:32:12 -07002326 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002327 return;
2328 if (pgdat->kswapd_max_order < order)
2329 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08002330 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002331 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002332 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002334 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335}
2336
Wu Fengguangadea02a2009-09-21 17:01:42 -07002337/*
2338 * The reclaimable count would be mostly accurate.
2339 * The less reclaimable pages may be
2340 * - mlocked pages, which will be moved to unevictable list when encountered
2341 * - mapped pages, which may require several travels to be reclaimed
2342 * - dirty pages, which is not "instantly" reclaimable
2343 */
2344unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002345{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002346 int nr;
2347
2348 nr = global_page_state(NR_ACTIVE_FILE) +
2349 global_page_state(NR_INACTIVE_FILE);
2350
2351 if (nr_swap_pages > 0)
2352 nr += global_page_state(NR_ACTIVE_ANON) +
2353 global_page_state(NR_INACTIVE_ANON);
2354
2355 return nr;
2356}
2357
2358unsigned long zone_reclaimable_pages(struct zone *zone)
2359{
2360 int nr;
2361
2362 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2363 zone_page_state(zone, NR_INACTIVE_FILE);
2364
2365 if (nr_swap_pages > 0)
2366 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2367 zone_page_state(zone, NR_INACTIVE_ANON);
2368
2369 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002370}
2371
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002372#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002373/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002374 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002375 * freed pages.
2376 *
2377 * Rather than trying to age LRUs the aim is to preserve the overall
2378 * LRU order by reclaiming preferentially
2379 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002381unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002382{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002383 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002384 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002385 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2386 .may_swap = 1,
2387 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002388 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002389 .nr_to_reclaim = nr_to_reclaim,
2390 .hibernation_mode = 1,
2391 .swappiness = vm_swappiness,
2392 .order = 0,
Balbir Singh66e17072008-02-07 00:13:56 -08002393 .isolate_pages = isolate_pages_global,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002394 };
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002395 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2396 struct task_struct *p = current;
2397 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002398
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002399 p->flags |= PF_MEMALLOC;
2400 lockdep_set_current_reclaim_state(sc.gfp_mask);
2401 reclaim_state.reclaimed_slab = 0;
2402 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002403
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002404 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002405
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002406 p->reclaim_state = NULL;
2407 lockdep_clear_current_reclaim_state();
2408 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002409
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002410 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002412#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002413
Linus Torvalds1da177e2005-04-16 15:20:36 -07002414/* It's optimal to keep kswapds on the same CPUs as their memory, but
2415 not required for correctness. So if the last cpu in a node goes
2416 away, we get changed to run anywhere: as the first one comes back,
2417 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002418static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002419 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002421 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002423 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002424 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002425 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302426 const struct cpumask *mask;
2427
2428 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002429
Rusty Russell3e597942009-01-01 10:12:24 +10302430 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002432 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002433 }
2434 }
2435 return NOTIFY_OK;
2436}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437
Yasunori Goto3218ae12006-06-27 02:53:33 -07002438/*
2439 * This kswapd start function will be called by init and node-hot-add.
2440 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2441 */
2442int kswapd_run(int nid)
2443{
2444 pg_data_t *pgdat = NODE_DATA(nid);
2445 int ret = 0;
2446
2447 if (pgdat->kswapd)
2448 return 0;
2449
2450 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2451 if (IS_ERR(pgdat->kswapd)) {
2452 /* failure at boot is fatal */
2453 BUG_ON(system_state == SYSTEM_BOOTING);
2454 printk("Failed to start kswapd on node %d\n",nid);
2455 ret = -1;
2456 }
2457 return ret;
2458}
2459
David Rientjes8fe23e02009-12-14 17:58:33 -08002460/*
2461 * Called by memory hotplug when all memory in a node is offlined.
2462 */
2463void kswapd_stop(int nid)
2464{
2465 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2466
2467 if (kswapd)
2468 kthread_stop(kswapd);
2469}
2470
Linus Torvalds1da177e2005-04-16 15:20:36 -07002471static int __init kswapd_init(void)
2472{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002473 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002474
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002476 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002477 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478 hotcpu_notifier(cpu_callback, 0);
2479 return 0;
2480}
2481
2482module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002483
2484#ifdef CONFIG_NUMA
2485/*
2486 * Zone reclaim mode
2487 *
2488 * If non-zero call zone_reclaim when the number of free pages falls below
2489 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002490 */
2491int zone_reclaim_mode __read_mostly;
2492
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002493#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002494#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002495#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2496#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2497
Christoph Lameter9eeff232006-01-18 17:42:31 -08002498/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002499 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2500 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2501 * a zone.
2502 */
2503#define ZONE_RECLAIM_PRIORITY 4
2504
Christoph Lameter9eeff232006-01-18 17:42:31 -08002505/*
Christoph Lameter96146342006-07-03 00:24:13 -07002506 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2507 * occur.
2508 */
2509int sysctl_min_unmapped_ratio = 1;
2510
2511/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002512 * If the number of slab pages in a zone grows beyond this percentage then
2513 * slab reclaim needs to occur.
2514 */
2515int sysctl_min_slab_ratio = 5;
2516
Mel Gorman90afa5d2009-06-16 15:33:20 -07002517static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2518{
2519 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2520 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2521 zone_page_state(zone, NR_ACTIVE_FILE);
2522
2523 /*
2524 * It's possible for there to be more file mapped pages than
2525 * accounted for by the pages on the file LRU lists because
2526 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2527 */
2528 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2529}
2530
2531/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2532static long zone_pagecache_reclaimable(struct zone *zone)
2533{
2534 long nr_pagecache_reclaimable;
2535 long delta = 0;
2536
2537 /*
2538 * If RECLAIM_SWAP is set, then all file pages are considered
2539 * potentially reclaimable. Otherwise, we have to worry about
2540 * pages like swapcache and zone_unmapped_file_pages() provides
2541 * a better estimate
2542 */
2543 if (zone_reclaim_mode & RECLAIM_SWAP)
2544 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2545 else
2546 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2547
2548 /* If we can't clean pages, remove dirty pages from consideration */
2549 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2550 delta += zone_page_state(zone, NR_FILE_DIRTY);
2551
2552 /* Watch for any possible underflows due to delta */
2553 if (unlikely(delta > nr_pagecache_reclaimable))
2554 delta = nr_pagecache_reclaimable;
2555
2556 return nr_pagecache_reclaimable - delta;
2557}
2558
Christoph Lameter0ff38492006-09-25 23:31:52 -07002559/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002560 * Try to free up some pages from this zone through reclaim.
2561 */
Andrew Morton179e9632006-03-22 00:08:18 -08002562static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002563{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002564 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002565 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002566 struct task_struct *p = current;
2567 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002568 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002569 struct scan_control sc = {
2570 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002571 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002572 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002573 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2574 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002575 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002576 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002577 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002578 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08002579 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07002580 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002581
2582 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08002583 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002584 /*
2585 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2586 * and we also need to be able to write out pages for RECLAIM_WRITE
2587 * and RECLAIM_SWAP.
2588 */
2589 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002590 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08002591 reclaim_state.reclaimed_slab = 0;
2592 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002593
Mel Gorman90afa5d2009-06-16 15:33:20 -07002594 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07002595 /*
2596 * Free memory by calling shrink zone with increasing
2597 * priorities until we have enough memory freed.
2598 */
2599 priority = ZONE_RECLAIM_PRIORITY;
2600 do {
Martin Bligh3bb1a852006-10-28 10:38:24 -07002601 note_zone_scanning_priority(zone, priority);
Rik van Riela79311c2009-01-06 14:40:01 -08002602 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002603 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002604 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002605 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002606
Christoph Lameter83e33a42006-09-25 23:31:53 -07002607 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2608 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002609 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002610 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002611 * many pages were freed in this zone. So we take the current
2612 * number of slab pages and shake the slab until it is reduced
2613 * by the same nr_pages that we used for reclaiming unmapped
2614 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002615 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002616 * Note that shrink_slab will free memory on all zones and may
2617 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002618 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07002619 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07002620 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2621 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07002622 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07002623
2624 /*
2625 * Update nr_reclaimed by the number of slab pages we
2626 * reclaimed from this zone.
2627 */
Rik van Riela79311c2009-01-06 14:40:01 -08002628 sc.nr_reclaimed += slab_reclaimable -
Christoph Lameter83e33a42006-09-25 23:31:53 -07002629 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002630 }
2631
Christoph Lameter9eeff232006-01-18 17:42:31 -08002632 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002633 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002634 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08002635 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002636}
Andrew Morton179e9632006-03-22 00:08:18 -08002637
2638int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2639{
Andrew Morton179e9632006-03-22 00:08:18 -08002640 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002641 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002642
2643 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002644 * Zone reclaim reclaims unmapped file backed pages and
2645 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002646 *
Christoph Lameter96146342006-07-03 00:24:13 -07002647 * A small portion of unmapped file backed pages is needed for
2648 * file I/O otherwise pages read by file I/O will be immediately
2649 * thrown out if the zone is overallocated. So we do not reclaim
2650 * if less than a specified percentage of the zone is used by
2651 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002652 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07002653 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2654 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002655 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08002656
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002657 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002658 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07002659
Andrew Morton179e9632006-03-22 00:08:18 -08002660 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002661 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002662 */
David Rientjesd773ed62007-10-16 23:26:01 -07002663 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002664 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08002665
2666 /*
2667 * Only run zone reclaim on the local zone or on zones that do not
2668 * have associated processors. This will favor the local processor
2669 * over remote processors and spread off node memory allocations
2670 * as wide as possible.
2671 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002672 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002673 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07002674 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07002675
2676 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002677 return ZONE_RECLAIM_NOSCAN;
2678
David Rientjesd773ed62007-10-16 23:26:01 -07002679 ret = __zone_reclaim(zone, gfp_mask, order);
2680 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2681
Mel Gorman24cf725182009-06-16 15:33:23 -07002682 if (!ret)
2683 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2684
David Rientjesd773ed62007-10-16 23:26:01 -07002685 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002686}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002687#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002688
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002689/*
2690 * page_evictable - test whether a page is evictable
2691 * @page: the page to test
2692 * @vma: the VMA in which the page is or will be mapped, may be NULL
2693 *
2694 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07002695 * lists vs unevictable list. The vma argument is !NULL when called from the
2696 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002697 *
2698 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002699 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07002700 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002701 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002702 */
2703int page_evictable(struct page *page, struct vm_area_struct *vma)
2704{
2705
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002706 if (mapping_unevictable(page_mapping(page)))
2707 return 0;
2708
Nick Pigginb291f002008-10-18 20:26:44 -07002709 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2710 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002711
2712 return 1;
2713}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002714
2715/**
2716 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2717 * @page: page to check evictability and move to appropriate lru list
2718 * @zone: zone page is in
2719 *
2720 * Checks a page for evictability and moves the page to the appropriate
2721 * zone lru list.
2722 *
2723 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2724 * have PageUnevictable set.
2725 */
2726static void check_move_unevictable_page(struct page *page, struct zone *zone)
2727{
2728 VM_BUG_ON(PageActive(page));
2729
2730retry:
2731 ClearPageUnevictable(page);
2732 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07002733 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002734
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002735 __dec_zone_state(zone, NR_UNEVICTABLE);
2736 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002737 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002738 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2739 __count_vm_event(UNEVICTABLE_PGRESCUED);
2740 } else {
2741 /*
2742 * rotate unevictable list
2743 */
2744 SetPageUnevictable(page);
2745 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002746 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002747 if (page_evictable(page, NULL))
2748 goto retry;
2749 }
2750}
2751
2752/**
2753 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2754 * @mapping: struct address_space to scan for evictable pages
2755 *
2756 * Scan all pages in mapping. Check unevictable pages for
2757 * evictability and move them to the appropriate zone lru list.
2758 */
2759void scan_mapping_unevictable_pages(struct address_space *mapping)
2760{
2761 pgoff_t next = 0;
2762 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2763 PAGE_CACHE_SHIFT;
2764 struct zone *zone;
2765 struct pagevec pvec;
2766
2767 if (mapping->nrpages == 0)
2768 return;
2769
2770 pagevec_init(&pvec, 0);
2771 while (next < end &&
2772 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2773 int i;
2774 int pg_scanned = 0;
2775
2776 zone = NULL;
2777
2778 for (i = 0; i < pagevec_count(&pvec); i++) {
2779 struct page *page = pvec.pages[i];
2780 pgoff_t page_index = page->index;
2781 struct zone *pagezone = page_zone(page);
2782
2783 pg_scanned++;
2784 if (page_index > next)
2785 next = page_index;
2786 next++;
2787
2788 if (pagezone != zone) {
2789 if (zone)
2790 spin_unlock_irq(&zone->lru_lock);
2791 zone = pagezone;
2792 spin_lock_irq(&zone->lru_lock);
2793 }
2794
2795 if (PageLRU(page) && PageUnevictable(page))
2796 check_move_unevictable_page(page, zone);
2797 }
2798 if (zone)
2799 spin_unlock_irq(&zone->lru_lock);
2800 pagevec_release(&pvec);
2801
2802 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2803 }
2804
2805}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002806
2807/**
2808 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2809 * @zone - zone of which to scan the unevictable list
2810 *
2811 * Scan @zone's unevictable LRU lists to check for pages that have become
2812 * evictable. Move those that have to @zone's inactive list where they
2813 * become candidates for reclaim, unless shrink_inactive_zone() decides
2814 * to reactivate them. Pages that are still unevictable are rotated
2815 * back onto @zone's unevictable list.
2816 */
2817#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08002818static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002819{
2820 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2821 unsigned long scan;
2822 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2823
2824 while (nr_to_scan > 0) {
2825 unsigned long batch_size = min(nr_to_scan,
2826 SCAN_UNEVICTABLE_BATCH_SIZE);
2827
2828 spin_lock_irq(&zone->lru_lock);
2829 for (scan = 0; scan < batch_size; scan++) {
2830 struct page *page = lru_to_page(l_unevictable);
2831
2832 if (!trylock_page(page))
2833 continue;
2834
2835 prefetchw_prev_lru_page(page, l_unevictable, flags);
2836
2837 if (likely(PageLRU(page) && PageUnevictable(page)))
2838 check_move_unevictable_page(page, zone);
2839
2840 unlock_page(page);
2841 }
2842 spin_unlock_irq(&zone->lru_lock);
2843
2844 nr_to_scan -= batch_size;
2845 }
2846}
2847
2848
2849/**
2850 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2851 *
2852 * A really big hammer: scan all zones' unevictable LRU lists to check for
2853 * pages that have become evictable. Move those back to the zones'
2854 * inactive list where they become candidates for reclaim.
2855 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2856 * and we add swap to the system. As such, it runs in the context of a task
2857 * that has possibly/probably made some previously unevictable pages
2858 * evictable.
2859 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08002860static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002861{
2862 struct zone *zone;
2863
2864 for_each_zone(zone) {
2865 scan_zone_unevictable_pages(zone);
2866 }
2867}
2868
2869/*
2870 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2871 * all nodes' unevictable lists for evictable pages
2872 */
2873unsigned long scan_unevictable_pages;
2874
2875int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002876 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002877 size_t *length, loff_t *ppos)
2878{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002879 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002880
2881 if (write && *(unsigned long *)table->data)
2882 scan_all_zones_unevictable_pages();
2883
2884 scan_unevictable_pages = 0;
2885 return 0;
2886}
2887
2888/*
2889 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2890 * a specified node's per zone unevictable lists for evictable pages.
2891 */
2892
2893static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2894 struct sysdev_attribute *attr,
2895 char *buf)
2896{
2897 return sprintf(buf, "0\n"); /* always zero; should fit... */
2898}
2899
2900static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2901 struct sysdev_attribute *attr,
2902 const char *buf, size_t count)
2903{
2904 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2905 struct zone *zone;
2906 unsigned long res;
2907 unsigned long req = strict_strtoul(buf, 10, &res);
2908
2909 if (!req)
2910 return 1; /* zero is no-op */
2911
2912 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2913 if (!populated_zone(zone))
2914 continue;
2915 scan_zone_unevictable_pages(zone);
2916 }
2917 return 1;
2918}
2919
2920
2921static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2922 read_scan_unevictable_node,
2923 write_scan_unevictable_node);
2924
2925int scan_unevictable_register_node(struct node *node)
2926{
2927 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2928}
2929
2930void scan_unevictable_unregister_node(struct node *node)
2931{
2932 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2933}
2934