blob: 9c7e57cc63a34f7231b77a7d8b395d3157a34ba6 [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>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090016#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#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
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070076 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080077
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070078 /*
79 * Intend to reclaim enough contenious memory rather than to reclaim
80 * enough amount memory. I.e, it's the mode for high order allocation.
81 */
82 bool lumpy_reclaim_mode;
83
Balbir Singh66e17072008-02-07 00:13:56 -080084 /* Which cgroup do we reclaim from */
85 struct mem_cgroup *mem_cgroup;
86
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -070087 /*
88 * Nodemask of nodes allowed by the caller. If NULL, all nodes
89 * are scanned.
90 */
91 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070092};
93
Linus Torvalds1da177e2005-04-16 15:20:36 -070094#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
95
96#ifdef ARCH_HAS_PREFETCH
97#define prefetch_prev_lru_page(_page, _base, _field) \
98 do { \
99 if ((_page)->lru.prev != _base) { \
100 struct page *prev; \
101 \
102 prev = lru_to_page(&(_page->lru)); \
103 prefetch(&prev->_field); \
104 } \
105 } while (0)
106#else
107#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
108#endif
109
110#ifdef ARCH_HAS_PREFETCHW
111#define prefetchw_prev_lru_page(_page, _base, _field) \
112 do { \
113 if ((_page)->lru.prev != _base) { \
114 struct page *prev; \
115 \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetchw(&prev->_field); \
118 } \
119 } while (0)
120#else
121#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
122#endif
123
124/*
125 * From 0 .. 100. Higher means more swappy.
126 */
127int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700128long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129
130static LIST_HEAD(shrinker_list);
131static DECLARE_RWSEM(shrinker_rwsem);
132
Balbir Singh00f0b822008-03-04 14:28:39 -0800133#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800134#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800135#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800136#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800137#endif
138
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800139static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
140 struct scan_control *sc)
141{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800142 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800143 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
144
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800145 return &zone->reclaim_stat;
146}
147
Vincent Li0b217672009-09-21 17:03:09 -0700148static unsigned long zone_nr_lru_pages(struct zone *zone,
149 struct scan_control *sc, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800150{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800151 if (!scanning_global_lru(sc))
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800152 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
153
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800154 return zone_page_state(zone, NR_LRU_BASE + lru);
155}
156
157
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158/*
159 * Add a shrinker callback to be called from the vm
160 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700161void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700163 shrinker->nr = 0;
164 down_write(&shrinker_rwsem);
165 list_add_tail(&shrinker->list, &shrinker_list);
166 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700168EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169
170/*
171 * Remove one
172 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700173void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174{
175 down_write(&shrinker_rwsem);
176 list_del(&shrinker->list);
177 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700179EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180
181#define SHRINK_BATCH 128
182/*
183 * Call the shrink functions to age shrinkable caches
184 *
185 * Here we assume it costs one seek to replace a lru page and that it also
186 * takes a seek to recreate a cache object. With this in mind we age equal
187 * percentages of the lru and ageable caches. This should balance the seeks
188 * generated by these structures.
189 *
Simon Arlott183ff222007-10-20 01:27:18 +0200190 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 * slab to avoid swapping.
192 *
193 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
194 *
195 * `lru_pages' represents the number of on-LRU pages in all the zones which
196 * are eligible for the caller's allocation attempt. It is used for balancing
197 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700198 *
199 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 */
Andrew Morton69e05942006-03-22 00:08:19 -0800201unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
202 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203{
204 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800205 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206
207 if (scanned == 0)
208 scanned = SWAP_CLUSTER_MAX;
209
210 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700211 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212
213 list_for_each_entry(shrinker, &shrinker_list, list) {
214 unsigned long long delta;
215 unsigned long total_scan;
Rusty Russell8e1f9362007-07-17 04:03:17 -0700216 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217
218 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800219 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220 do_div(delta, lru_pages + 1);
221 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800222 if (shrinker->nr < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700223 printk(KERN_ERR "shrink_slab: %pF negative objects to "
224 "delete nr=%ld\n",
225 shrinker->shrink, shrinker->nr);
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800226 shrinker->nr = max_pass;
227 }
228
229 /*
230 * Avoid risking looping forever due to too large nr value:
231 * never try to free more than twice the estimate number of
232 * freeable entries.
233 */
234 if (shrinker->nr > max_pass * 2)
235 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236
237 total_scan = shrinker->nr;
238 shrinker->nr = 0;
239
240 while (total_scan >= SHRINK_BATCH) {
241 long this_scan = SHRINK_BATCH;
242 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700243 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244
Rusty Russell8e1f9362007-07-17 04:03:17 -0700245 nr_before = (*shrinker->shrink)(0, gfp_mask);
246 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 if (shrink_ret == -1)
248 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700249 if (shrink_ret < nr_before)
250 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700251 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 total_scan -= this_scan;
253
254 cond_resched();
255 }
256
257 shrinker->nr += total_scan;
258 }
259 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700260 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261}
262
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263static inline int is_page_cache_freeable(struct page *page)
264{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700265 /*
266 * A freeable page cache page is referenced only by the caller
267 * that isolated the page, the page cache radix tree and
268 * optional buffer heads at page->private.
269 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700270 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700271}
272
273static int may_write_to_queue(struct backing_dev_info *bdi)
274{
Christoph Lameter930d9152006-01-08 01:00:47 -0800275 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700276 return 1;
277 if (!bdi_write_congested(bdi))
278 return 1;
279 if (bdi == current->backing_dev_info)
280 return 1;
281 return 0;
282}
283
284/*
285 * We detected a synchronous write error writing a page out. Probably
286 * -ENOSPC. We need to propagate that into the address_space for a subsequent
287 * fsync(), msync() or close().
288 *
289 * The tricky part is that after writepage we cannot touch the mapping: nothing
290 * prevents it from being freed up. But we have a ref on the page and once
291 * that page is locked, the mapping is pinned.
292 *
293 * We're allowed to run sleeping lock_page() here because we know the caller has
294 * __GFP_FS.
295 */
296static void handle_write_error(struct address_space *mapping,
297 struct page *page, int error)
298{
299 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700300 if (page_mapping(page) == mapping)
301 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700302 unlock_page(page);
303}
304
Andy Whitcroftc661b072007-08-22 14:01:26 -0700305/* Request for sync pageout. */
306enum pageout_io {
307 PAGEOUT_IO_ASYNC,
308 PAGEOUT_IO_SYNC,
309};
310
Christoph Lameter04e62a22006-06-23 02:03:38 -0700311/* possible outcome of pageout() */
312typedef enum {
313 /* failed to write page out, page is locked */
314 PAGE_KEEP,
315 /* move page to the active list, page is locked */
316 PAGE_ACTIVATE,
317 /* page has been sent to the disk successfully, page is unlocked */
318 PAGE_SUCCESS,
319 /* page is clean and locked */
320 PAGE_CLEAN,
321} pageout_t;
322
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323/*
Andrew Morton1742f192006-03-22 00:08:21 -0800324 * pageout is called by shrink_page_list() for each dirty page.
325 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700326 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700327static pageout_t pageout(struct page *page, struct address_space *mapping,
328 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329{
330 /*
331 * If the page is dirty, only perform writeback if that write
332 * will be non-blocking. To prevent this allocation from being
333 * stalled by pagecache activity. But note that there may be
334 * stalls if we need to run get_block(). We could test
335 * PagePrivate for that.
336 *
Vincent Li6aceb532009-12-14 17:58:49 -0800337 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700338 * this page's queue, we can perform writeback even if that
339 * will block.
340 *
341 * If the page is swapcache, write it back even if that would
342 * block, for some throttling. This happens by accident, because
343 * swap_backing_dev_info is bust: it doesn't reflect the
344 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345 */
346 if (!is_page_cache_freeable(page))
347 return PAGE_KEEP;
348 if (!mapping) {
349 /*
350 * Some data journaling orphaned pages can have
351 * page->mapping == NULL while being dirty with clean buffers.
352 */
David Howells266cf652009-04-03 16:42:36 +0100353 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700354 if (try_to_free_buffers(page)) {
355 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700356 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700357 return PAGE_CLEAN;
358 }
359 }
360 return PAGE_KEEP;
361 }
362 if (mapping->a_ops->writepage == NULL)
363 return PAGE_ACTIVATE;
364 if (!may_write_to_queue(mapping->backing_dev_info))
365 return PAGE_KEEP;
366
367 if (clear_page_dirty_for_io(page)) {
368 int res;
369 struct writeback_control wbc = {
370 .sync_mode = WB_SYNC_NONE,
371 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700372 .range_start = 0,
373 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 .nonblocking = 1,
375 .for_reclaim = 1,
376 };
377
378 SetPageReclaim(page);
379 res = mapping->a_ops->writepage(page, &wbc);
380 if (res < 0)
381 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800382 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700383 ClearPageReclaim(page);
384 return PAGE_ACTIVATE;
385 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700386
387 /*
388 * Wait on writeback if requested to. This happens when
389 * direct reclaiming a large contiguous area and the
390 * first attempt to free a range of pages fails.
391 */
392 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
393 wait_on_page_writeback(page);
394
Linus Torvalds1da177e2005-04-16 15:20:36 -0700395 if (!PageWriteback(page)) {
396 /* synchronous write or broken a_ops? */
397 ClearPageReclaim(page);
398 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700399 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 return PAGE_SUCCESS;
401 }
402
403 return PAGE_CLEAN;
404}
405
Andrew Mortona649fd92006-10-17 00:09:36 -0700406/*
Nick Piggine2867812008-07-25 19:45:30 -0700407 * Same as remove_mapping, but if the page is removed from the mapping, it
408 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700409 */
Nick Piggine2867812008-07-25 19:45:30 -0700410static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800411{
Nick Piggin28e4d962006-09-25 23:31:23 -0700412 BUG_ON(!PageLocked(page));
413 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800414
Nick Piggin19fd6232008-07-25 19:45:32 -0700415 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800416 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700417 * The non racy check for a busy page.
418 *
419 * Must be careful with the order of the tests. When someone has
420 * a ref to the page, it may be possible that they dirty it then
421 * drop the reference. So if PageDirty is tested before page_count
422 * here, then the following race may occur:
423 *
424 * get_user_pages(&page);
425 * [user mapping goes away]
426 * write_to(page);
427 * !PageDirty(page) [good]
428 * SetPageDirty(page);
429 * put_page(page);
430 * !page_count(page) [good, discard it]
431 *
432 * [oops, our write_to data is lost]
433 *
434 * Reversing the order of the tests ensures such a situation cannot
435 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
436 * load is not satisfied before that of page->_count.
437 *
438 * Note that if SetPageDirty is always performed via set_page_dirty,
439 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800440 */
Nick Piggine2867812008-07-25 19:45:30 -0700441 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800442 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700443 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
444 if (unlikely(PageDirty(page))) {
445 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800446 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700447 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800448
449 if (PageSwapCache(page)) {
450 swp_entry_t swap = { .val = page_private(page) };
451 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700452 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700453 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700454 } else {
455 __remove_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700456 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700457 mem_cgroup_uncharge_cache_page(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800458 }
459
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800460 return 1;
461
462cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700463 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800464 return 0;
465}
466
Linus Torvalds1da177e2005-04-16 15:20:36 -0700467/*
Nick Piggine2867812008-07-25 19:45:30 -0700468 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
469 * someone else has a ref on the page, abort and return 0. If it was
470 * successfully detached, return 1. Assumes the caller has a single ref on
471 * this page.
472 */
473int remove_mapping(struct address_space *mapping, struct page *page)
474{
475 if (__remove_mapping(mapping, page)) {
476 /*
477 * Unfreezing the refcount with 1 rather than 2 effectively
478 * drops the pagecache ref for us without requiring another
479 * atomic operation.
480 */
481 page_unfreeze_refs(page, 1);
482 return 1;
483 }
484 return 0;
485}
486
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700487/**
488 * putback_lru_page - put previously isolated page onto appropriate LRU list
489 * @page: page to be put back to appropriate lru list
490 *
491 * Add previously isolated @page to appropriate LRU list.
492 * Page may still be unevictable for other reasons.
493 *
494 * lru_lock must not be held, interrupts must be enabled.
495 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700496void putback_lru_page(struct page *page)
497{
498 int lru;
499 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700500 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700501
502 VM_BUG_ON(PageLRU(page));
503
504redo:
505 ClearPageUnevictable(page);
506
507 if (page_evictable(page, NULL)) {
508 /*
509 * For evictable pages, we can use the cache.
510 * In event of a race, worst case is we end up with an
511 * unevictable page on [in]active list.
512 * We know how to handle that.
513 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700514 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700515 lru_cache_add_lru(page, lru);
516 } else {
517 /*
518 * Put unevictable pages directly on zone's unevictable
519 * list.
520 */
521 lru = LRU_UNEVICTABLE;
522 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700523 /*
524 * When racing with an mlock clearing (page is
525 * unlocked), make sure that if the other thread does
526 * not observe our setting of PG_lru and fails
527 * isolation, we see PG_mlocked cleared below and move
528 * the page back to the evictable list.
529 *
530 * The other side is TestClearPageMlocked().
531 */
532 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700533 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700534
535 /*
536 * page's status can change while we move it among lru. If an evictable
537 * page is on unevictable list, it never be freed. To avoid that,
538 * check after we added it to the list, again.
539 */
540 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
541 if (!isolate_lru_page(page)) {
542 put_page(page);
543 goto redo;
544 }
545 /* This means someone else dropped this page from LRU
546 * So, it will be freed or putback to LRU again. There is
547 * nothing to do here.
548 */
549 }
550
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700551 if (was_unevictable && lru != LRU_UNEVICTABLE)
552 count_vm_event(UNEVICTABLE_PGRESCUED);
553 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
554 count_vm_event(UNEVICTABLE_PGCULLED);
555
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700556 put_page(page); /* drop ref from isolate */
557}
558
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800559enum page_references {
560 PAGEREF_RECLAIM,
561 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner64574742010-03-05 13:42:22 -0800562 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800563 PAGEREF_ACTIVATE,
564};
565
566static enum page_references page_check_references(struct page *page,
567 struct scan_control *sc)
568{
Johannes Weiner64574742010-03-05 13:42:22 -0800569 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800570 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800571
Johannes Weiner64574742010-03-05 13:42:22 -0800572 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
573 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800574
575 /* Lumpy reclaim - ignore references */
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700576 if (sc->lumpy_reclaim_mode)
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800577 return PAGEREF_RECLAIM;
578
579 /*
580 * Mlock lost the isolation race with us. Let try_to_unmap()
581 * move the page to the unevictable list.
582 */
583 if (vm_flags & VM_LOCKED)
584 return PAGEREF_RECLAIM;
585
Johannes Weiner64574742010-03-05 13:42:22 -0800586 if (referenced_ptes) {
587 if (PageAnon(page))
588 return PAGEREF_ACTIVATE;
589 /*
590 * All mapped pages start out with page table
591 * references from the instantiating fault, so we need
592 * to look twice if a mapped file page is used more
593 * than once.
594 *
595 * Mark it and spare it for another trip around the
596 * inactive list. Another page table reference will
597 * lead to its activation.
598 *
599 * Note: the mark is set for activated pages as well
600 * so that recently deactivated but used pages are
601 * quickly recovered.
602 */
603 SetPageReferenced(page);
604
605 if (referenced_page)
606 return PAGEREF_ACTIVATE;
607
608 return PAGEREF_KEEP;
609 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800610
611 /* Reclaim if clean, defer dirty pages to writeback */
Johannes Weiner64574742010-03-05 13:42:22 -0800612 if (referenced_page)
613 return PAGEREF_RECLAIM_CLEAN;
614
615 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800616}
617
Nick Piggine2867812008-07-25 19:45:30 -0700618/*
Andrew Morton1742f192006-03-22 00:08:21 -0800619 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700620 */
Andrew Morton1742f192006-03-22 00:08:21 -0800621static unsigned long shrink_page_list(struct list_head *page_list,
Andy Whitcroftc661b072007-08-22 14:01:26 -0700622 struct scan_control *sc,
623 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700624{
625 LIST_HEAD(ret_pages);
626 struct pagevec freed_pvec;
627 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800628 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700629
630 cond_resched();
631
632 pagevec_init(&freed_pvec, 1);
633 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800634 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 struct address_space *mapping;
636 struct page *page;
637 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700638
639 cond_resched();
640
641 page = lru_to_page(page_list);
642 list_del(&page->lru);
643
Nick Piggin529ae9a2008-08-02 12:01:03 +0200644 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 goto keep;
646
Nick Piggin725d7042006-09-25 23:30:55 -0700647 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700648
649 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800650
Nick Pigginb291f002008-10-18 20:26:44 -0700651 if (unlikely(!page_evictable(page, NULL)))
652 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700653
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700654 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800655 goto keep_locked;
656
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 /* Double the slab pressure for mapped and swapcache pages */
658 if (page_mapped(page) || PageSwapCache(page))
659 sc->nr_scanned++;
660
Andy Whitcroftc661b072007-08-22 14:01:26 -0700661 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
662 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
663
664 if (PageWriteback(page)) {
665 /*
666 * Synchronous reclaim is performed in two passes,
667 * first an asynchronous pass over the list to
668 * start parallel writeback, and a second synchronous
669 * pass to wait for the IO to complete. Wait here
670 * for any page for which writeback has already
671 * started.
672 */
673 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
674 wait_on_page_writeback(page);
Andrew Morton4dd4b922008-03-24 12:29:52 -0700675 else
Andy Whitcroftc661b072007-08-22 14:01:26 -0700676 goto keep_locked;
677 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700678
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800679 references = page_check_references(page, sc);
680 switch (references) {
681 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 goto activate_locked;
Johannes Weiner64574742010-03-05 13:42:22 -0800683 case PAGEREF_KEEP:
684 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800685 case PAGEREF_RECLAIM:
686 case PAGEREF_RECLAIM_CLEAN:
687 ; /* try to reclaim the page below */
688 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700689
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690 /*
691 * Anonymous process memory has backing store?
692 * Try to allocate it some swap space here.
693 */
Nick Pigginb291f002008-10-18 20:26:44 -0700694 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800695 if (!(sc->gfp_mask & __GFP_IO))
696 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800697 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800699 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700700 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701
702 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703
704 /*
705 * The page is mapped into the page tables of one or more
706 * processes. Try to unmap it here.
707 */
708 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200709 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710 case SWAP_FAIL:
711 goto activate_locked;
712 case SWAP_AGAIN:
713 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700714 case SWAP_MLOCK:
715 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700716 case SWAP_SUCCESS:
717 ; /* try to free the page below */
718 }
719 }
720
721 if (PageDirty(page)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800722 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700723 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700724 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800726 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 goto keep_locked;
728
729 /* Page is dirty, try to write it out here */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700730 switch (pageout(page, mapping, sync_writeback)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700731 case PAGE_KEEP:
732 goto keep_locked;
733 case PAGE_ACTIVATE:
734 goto activate_locked;
735 case PAGE_SUCCESS:
Andrew Morton4dd4b922008-03-24 12:29:52 -0700736 if (PageWriteback(page) || PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700737 goto keep;
738 /*
739 * A synchronous write - probably a ramdisk. Go
740 * ahead and try to reclaim the page.
741 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200742 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 goto keep;
744 if (PageDirty(page) || PageWriteback(page))
745 goto keep_locked;
746 mapping = page_mapping(page);
747 case PAGE_CLEAN:
748 ; /* try to free the page below */
749 }
750 }
751
752 /*
753 * If the page has buffers, try to free the buffer mappings
754 * associated with this page. If we succeed we try to free
755 * the page as well.
756 *
757 * We do this even if the page is PageDirty().
758 * try_to_release_page() does not perform I/O, but it is
759 * possible for a page to have PageDirty set, but it is actually
760 * clean (all its buffers are clean). This happens if the
761 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700762 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 * try_to_release_page() will discover that cleanness and will
764 * drop the buffers and mark the page clean - it can be freed.
765 *
766 * Rarely, pages can have buffers and no ->mapping. These are
767 * the pages which were not successfully invalidated in
768 * truncate_complete_page(). We try to drop those buffers here
769 * and if that worked, and the page is no longer mapped into
770 * process address space (page_count == 1) it can be freed.
771 * Otherwise, leave the page on the LRU so it is swappable.
772 */
David Howells266cf652009-04-03 16:42:36 +0100773 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700774 if (!try_to_release_page(page, sc->gfp_mask))
775 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700776 if (!mapping && page_count(page) == 1) {
777 unlock_page(page);
778 if (put_page_testzero(page))
779 goto free_it;
780 else {
781 /*
782 * rare race with speculative reference.
783 * the speculative reference will free
784 * this page shortly, so we may
785 * increment nr_reclaimed here (and
786 * leave it off the LRU).
787 */
788 nr_reclaimed++;
789 continue;
790 }
791 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 }
793
Nick Piggine2867812008-07-25 19:45:30 -0700794 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800795 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796
Nick Piggina978d6f2008-10-18 20:26:58 -0700797 /*
798 * At this point, we have no other references and there is
799 * no way to pick any more up (removed from LRU, removed
800 * from pagecache). Can use non-atomic bitops now (and
801 * we obviously don't have to worry about waking up a process
802 * waiting on the page lock, because there are no references.
803 */
804 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700805free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800806 nr_reclaimed++;
Nick Piggine2867812008-07-25 19:45:30 -0700807 if (!pagevec_add(&freed_pvec, page)) {
808 __pagevec_free(&freed_pvec);
809 pagevec_reinit(&freed_pvec);
810 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700811 continue;
812
Nick Pigginb291f002008-10-18 20:26:44 -0700813cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800814 if (PageSwapCache(page))
815 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700816 unlock_page(page);
817 putback_lru_page(page);
818 continue;
819
Linus Torvalds1da177e2005-04-16 15:20:36 -0700820activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700821 /* Not a candidate for swapping, so reclaim swap space. */
822 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800823 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700824 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700825 SetPageActive(page);
826 pgactivate++;
827keep_locked:
828 unlock_page(page);
829keep:
830 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700831 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700832 }
833 list_splice(&ret_pages, page_list);
834 if (pagevec_count(&freed_pvec))
Nick Piggine2867812008-07-25 19:45:30 -0700835 __pagevec_free(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700836 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800837 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700838}
839
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700840/*
841 * Attempt to remove the specified page from its LRU. Only take this page
842 * if it is of the appropriate PageActive status. Pages which are being
843 * freed elsewhere are also ignored.
844 *
845 * page: page to consider
846 * mode: one of the LRU isolation modes defined above
847 *
848 * returns 0 on success, -ve errno on failure.
849 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700850int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700851{
852 int ret = -EINVAL;
853
854 /* Only take pages on the LRU. */
855 if (!PageLRU(page))
856 return ret;
857
858 /*
859 * When checking the active state, we need to be sure we are
860 * dealing with comparible boolean values. Take the logical not
861 * of each.
862 */
863 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
864 return ret;
865
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700866 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700867 return ret;
868
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700869 /*
870 * When this function is being called for lumpy reclaim, we
871 * initially look into all LRU pages, active, inactive and
872 * unevictable; only give shrink_page_list evictable pages.
873 */
874 if (PageUnevictable(page))
875 return ret;
876
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700877 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800878
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700879 if (likely(get_page_unless_zero(page))) {
880 /*
881 * Be careful not to clear PageLRU until after we're
882 * sure the page is not being freed elsewhere -- the
883 * page release code relies on it.
884 */
885 ClearPageLRU(page);
886 ret = 0;
887 }
888
889 return ret;
890}
891
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800892/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700893 * zone->lru_lock is heavily contended. Some of the functions that
894 * shrink the lists perform better by taking out a batch of pages
895 * and working on them outside the LRU lock.
896 *
897 * For pagecache intensive workloads, this function is the hottest
898 * spot in the kernel (apart from copy_*_user functions).
899 *
900 * Appropriate locks must be held before calling this function.
901 *
902 * @nr_to_scan: The number of pages to look through on the list.
903 * @src: The LRU list to pull pages off.
904 * @dst: The temp list to put pages on to.
905 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700906 * @order: The caller's attempted allocation order
907 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700908 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 *
910 * returns how many pages were moved onto *@dst.
911 */
Andrew Morton69e05942006-03-22 00:08:19 -0800912static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
913 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700914 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915{
Andrew Morton69e05942006-03-22 00:08:19 -0800916 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800917 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700918
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800919 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700920 struct page *page;
921 unsigned long pfn;
922 unsigned long end_pfn;
923 unsigned long page_pfn;
924 int zone_id;
925
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926 page = lru_to_page(src);
927 prefetchw_prev_lru_page(page, src, flags);
928
Nick Piggin725d7042006-09-25 23:30:55 -0700929 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800930
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700931 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700932 case 0:
933 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700934 mem_cgroup_del_lru(page);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800935 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700936 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800937
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700938 case -EBUSY:
939 /* else it is being freed elsewhere */
940 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700941 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700942 continue;
943
944 default:
945 BUG();
946 }
947
948 if (!order)
949 continue;
950
951 /*
952 * Attempt to take all pages in the order aligned region
953 * surrounding the tag page. Only take those pages of
954 * the same active state as that tag page. We may safely
955 * round the target page pfn down to the requested order
956 * as the mem_map is guarenteed valid out to MAX_ORDER,
957 * where that page is in a different zone we will detect
958 * it from its zone id and abort this block scan.
959 */
960 zone_id = page_zone_id(page);
961 page_pfn = page_to_pfn(page);
962 pfn = page_pfn & ~((1 << order) - 1);
963 end_pfn = pfn + (1 << order);
964 for (; pfn < end_pfn; pfn++) {
965 struct page *cursor_page;
966
967 /* The target page is in the block, ignore it. */
968 if (unlikely(pfn == page_pfn))
969 continue;
970
971 /* Avoid holes within the zone. */
972 if (unlikely(!pfn_valid_within(pfn)))
973 break;
974
975 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700976
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700977 /* Check that we have not crossed a zone boundary. */
978 if (unlikely(page_zone_id(cursor_page) != zone_id))
979 continue;
Minchan Kimde2e7562009-09-21 17:01:43 -0700980
981 /*
982 * If we don't have enough swap space, reclaiming of
983 * anon page which don't already have a swap slot is
984 * pointless.
985 */
986 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
987 !PageSwapCache(cursor_page))
988 continue;
989
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -0700990 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700991 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +0900992 mem_cgroup_del_lru(cursor_page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700993 nr_taken++;
994 scan++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700995 }
996 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700997 }
998
999 *scanned = scan;
1000 return nr_taken;
1001}
1002
Balbir Singh66e17072008-02-07 00:13:56 -08001003static unsigned long isolate_pages_global(unsigned long nr,
1004 struct list_head *dst,
1005 unsigned long *scanned, int order,
1006 int mode, struct zone *z,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001007 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -08001008{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001009 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -08001010 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001011 lru += LRU_ACTIVE;
1012 if (file)
1013 lru += LRU_FILE;
1014 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001015 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -08001016}
1017
Linus Torvalds1da177e2005-04-16 15:20:36 -07001018/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001019 * clear_active_flags() is a helper for shrink_active_list(), clearing
1020 * any active bits from the pages in the list.
1021 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001022static unsigned long clear_active_flags(struct list_head *page_list,
1023 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001024{
1025 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001026 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001027 struct page *page;
1028
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001029 list_for_each_entry(page, page_list, lru) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07001030 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001031 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001032 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001033 ClearPageActive(page);
1034 nr_active++;
1035 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001036 count[lru]++;
1037 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001038
1039 return nr_active;
1040}
1041
Nick Piggin62695a82008-10-18 20:26:09 -07001042/**
1043 * isolate_lru_page - tries to isolate a page from its LRU list
1044 * @page: page to isolate from its LRU list
1045 *
1046 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1047 * vmstat statistic corresponding to whatever LRU list the page was on.
1048 *
1049 * Returns 0 if the page was removed from an LRU list.
1050 * Returns -EBUSY if the page was not on an LRU list.
1051 *
1052 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001053 * the active list, it will have PageActive set. If it was found on
1054 * the unevictable list, it will have the PageUnevictable bit set. That flag
1055 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001056 *
1057 * The vmstat statistic corresponding to the list on which the page was
1058 * found will be decremented.
1059 *
1060 * Restrictions:
1061 * (1) Must be called with an elevated refcount on the page. This is a
1062 * fundamentnal difference from isolate_lru_pages (which is called
1063 * without a stable reference).
1064 * (2) the lru_lock must not be held.
1065 * (3) interrupts must be enabled.
1066 */
1067int isolate_lru_page(struct page *page)
1068{
1069 int ret = -EBUSY;
1070
1071 if (PageLRU(page)) {
1072 struct zone *zone = page_zone(page);
1073
1074 spin_lock_irq(&zone->lru_lock);
1075 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001076 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001077 ret = 0;
1078 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001079
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001080 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001081 }
1082 spin_unlock_irq(&zone->lru_lock);
1083 }
1084 return ret;
1085}
1086
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001087/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001088 * Are there way too many processes in the direct reclaim path already?
1089 */
1090static int too_many_isolated(struct zone *zone, int file,
1091 struct scan_control *sc)
1092{
1093 unsigned long inactive, isolated;
1094
1095 if (current_is_kswapd())
1096 return 0;
1097
1098 if (!scanning_global_lru(sc))
1099 return 0;
1100
1101 if (file) {
1102 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1103 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1104 } else {
1105 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1106 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1107 }
1108
1109 return isolated > inactive;
1110}
1111
1112/*
Andrew Morton1742f192006-03-22 00:08:21 -08001113 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1114 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001115 */
Andrew Morton1742f192006-03-22 00:08:21 -08001116static unsigned long shrink_inactive_list(unsigned long max_scan,
Rik van Riel33c120e2008-10-18 20:26:36 -07001117 struct zone *zone, struct scan_control *sc,
1118 int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119{
1120 LIST_HEAD(page_list);
1121 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -08001122 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001123 unsigned long nr_reclaimed = 0;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001124 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001125
Rik van Riel35cd7812009-09-21 17:01:38 -07001126 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001127 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001128
1129 /* We are about to die and free our memory. Return now. */
1130 if (fatal_signal_pending(current))
1131 return SWAP_CLUSTER_MAX;
1132 }
1133
Linus Torvalds1da177e2005-04-16 15:20:36 -07001134
1135 pagevec_init(&pvec, 1);
1136
1137 lru_add_drain();
1138 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -08001139 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001140 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -08001141 unsigned long nr_taken;
1142 unsigned long nr_scan;
1143 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001144 unsigned long nr_active;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001145 unsigned int count[NR_LRU_LISTS] = { 0, };
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -07001146 int mode = sc->lumpy_reclaim_mode ? ISOLATE_BOTH : ISOLATE_INACTIVE;
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001147 unsigned long nr_anon;
1148 unsigned long nr_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001150 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001151 nr_taken = isolate_pages_global(SWAP_CLUSTER_MAX,
1152 &page_list, &nr_scan,
1153 sc->order, mode,
1154 zone, 0, file);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001155 zone->pages_scanned += nr_scan;
1156 if (current_is_kswapd())
1157 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1158 nr_scan);
1159 else
1160 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1161 nr_scan);
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001162 } else {
1163 nr_taken = mem_cgroup_isolate_pages(SWAP_CLUSTER_MAX,
1164 &page_list, &nr_scan,
1165 sc->order, mode,
1166 zone, sc->mem_cgroup,
1167 0, file);
1168 /*
1169 * mem_cgroup_isolate_pages() keeps track of
1170 * scanned pages on its own.
1171 */
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001172 }
1173
1174 if (nr_taken == 0)
1175 goto done;
1176
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001177 nr_active = clear_active_flags(&page_list, count);
Andy Whitcrofte9187bd2007-08-22 14:01:25 -07001178 __count_vm_events(PGDEACTIVATE, nr_active);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001179
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001180 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1181 -count[LRU_ACTIVE_FILE]);
1182 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1183 -count[LRU_INACTIVE_FILE]);
1184 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1185 -count[LRU_ACTIVE_ANON]);
1186 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1187 -count[LRU_INACTIVE_ANON]);
1188
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001189 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1190 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1191 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1192 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001193
Huang Shijie62c0c2f2009-12-14 17:59:48 -08001194 reclaim_stat->recent_scanned[0] += nr_anon;
1195 reclaim_stat->recent_scanned[1] += nr_file;
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001196
Linus Torvalds1da177e2005-04-16 15:20:36 -07001197 spin_unlock_irq(&zone->lru_lock);
1198
Andrew Morton69e05942006-03-22 00:08:19 -08001199 nr_scanned += nr_scan;
Andy Whitcroftc661b072007-08-22 14:01:26 -07001200 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1201
1202 /*
1203 * If we are direct reclaiming for contiguous pages and we do
1204 * not reclaim everything in the list, try again and wait
1205 * for IO to complete. This will stall high-order allocations
1206 * but that should be acceptable to the caller
1207 */
1208 if (nr_freed < nr_taken && !current_is_kswapd() &&
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -07001209 sc->lumpy_reclaim_mode) {
Jens Axboe8aa7e842009-07-09 14:52:32 +02001210 congestion_wait(BLK_RW_ASYNC, HZ/10);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001211
1212 /*
1213 * The attempt at page out may have made some
1214 * of the pages active, mark them inactive again.
1215 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001216 nr_active = clear_active_flags(&page_list, count);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001217 count_vm_events(PGDEACTIVATE, nr_active);
1218
1219 nr_freed += shrink_page_list(&page_list, sc,
1220 PAGEOUT_IO_SYNC);
1221 }
1222
Andrew Morton05ff5132006-03-22 00:08:20 -08001223 nr_reclaimed += nr_freed;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001224
Nick Piggina74609f2006-01-06 00:11:20 -08001225 local_irq_disable();
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001226 if (current_is_kswapd())
Christoph Lameterf8891e52006-06-30 01:55:45 -07001227 __count_vm_events(KSWAPD_STEAL, nr_freed);
Shantanu Goel918d3f92006-12-29 16:48:59 -08001228 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -08001229
1230 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231 /*
1232 * Put back any unfreeable pages.
1233 */
1234 while (!list_empty(&page_list)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001235 int lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -07001237 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238 list_del(&page->lru);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001239 if (unlikely(!page_evictable(page, NULL))) {
1240 spin_unlock_irq(&zone->lru_lock);
1241 putback_lru_page(page);
1242 spin_lock_irq(&zone->lru_lock);
1243 continue;
1244 }
1245 SetPageLRU(page);
1246 lru = page_lru(page);
1247 add_page_to_lru_list(zone, page, lru);
KOSAKI Motohiro74a1c482009-09-21 17:01:45 -07001248 if (is_active_lru(lru)) {
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001249 int file = is_file_lru(lru);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001250 reclaim_stat->recent_rotated[file]++;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001251 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001252 if (!pagevec_add(&pvec, page)) {
1253 spin_unlock_irq(&zone->lru_lock);
1254 __pagevec_release(&pvec);
1255 spin_lock_irq(&zone->lru_lock);
1256 }
1257 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001258 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1259 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1260
Andrew Morton69e05942006-03-22 00:08:19 -08001261 } while (nr_scanned < max_scan);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001262
Linus Torvalds1da177e2005-04-16 15:20:36 -07001263done:
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001264 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -08001266 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267}
1268
Martin Bligh3bb1a852006-10-28 10:38:24 -07001269/*
1270 * We are about to scan this zone at a certain priority level. If that priority
1271 * level is smaller (ie: more urgent) than the previous priority, then note
1272 * that priority level within the zone. This is done so that when the next
1273 * process comes in to scan this zone, it will immediately start out at this
1274 * priority level rather than having to build up its own scanning priority.
1275 * Here, this priority affects only the reclaim-mapped threshold.
1276 */
1277static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1278{
1279 if (priority < zone->prev_priority)
1280 zone->prev_priority = priority;
1281}
1282
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001283/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 * This moves pages from the active list to the inactive list.
1285 *
1286 * We move them the other way if the page is referenced by one or more
1287 * processes, from rmap.
1288 *
1289 * If the pages are mostly unmapped, the processing is fast and it is
1290 * appropriate to hold zone->lru_lock across the whole operation. But if
1291 * the pages are mapped, the processing is slow (page_referenced()) so we
1292 * should drop zone->lru_lock around each page. It's impossible to balance
1293 * this, so instead we remove the pages from the LRU while processing them.
1294 * It is safe to rely on PG_active against the non-LRU pages in here because
1295 * nobody will play with that bit on a non-LRU page.
1296 *
1297 * The downside is that we have to touch page->_count against each page.
1298 * But we had to alter page->flags anyway.
1299 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001300
Wu Fengguang3eb41402009-06-16 15:33:13 -07001301static void move_active_pages_to_lru(struct zone *zone,
1302 struct list_head *list,
1303 enum lru_list lru)
1304{
1305 unsigned long pgmoved = 0;
1306 struct pagevec pvec;
1307 struct page *page;
1308
1309 pagevec_init(&pvec, 1);
1310
1311 while (!list_empty(list)) {
1312 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001313
1314 VM_BUG_ON(PageLRU(page));
1315 SetPageLRU(page);
1316
Wu Fengguang3eb41402009-06-16 15:33:13 -07001317 list_move(&page->lru, &zone->lru[lru].list);
1318 mem_cgroup_add_lru_list(page, lru);
1319 pgmoved++;
1320
1321 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1322 spin_unlock_irq(&zone->lru_lock);
1323 if (buffer_heads_over_limit)
1324 pagevec_strip(&pvec);
1325 __pagevec_release(&pvec);
1326 spin_lock_irq(&zone->lru_lock);
1327 }
1328 }
1329 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1330 if (!is_active_lru(lru))
1331 __count_vm_events(PGDEACTIVATE, pgmoved);
1332}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001333
Andrew Morton1742f192006-03-22 00:08:21 -08001334static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001335 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001337 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001338 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001339 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001340 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001341 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001342 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001344 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001345 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001346
1347 lru_add_drain();
1348 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001349 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001350 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1351 &pgscanned, sc->order,
1352 ISOLATE_ACTIVE, zone,
1353 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001354 zone->pages_scanned += pgscanned;
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001355 } else {
1356 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1357 &pgscanned, sc->order,
1358 ISOLATE_ACTIVE, zone,
1359 sc->mem_cgroup, 1, file);
1360 /*
1361 * mem_cgroup_isolate_pages() keeps track of
1362 * scanned pages on its own.
1363 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001364 }
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001365
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001366 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001367
Wu Fengguang3eb41402009-06-16 15:33:13 -07001368 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001369 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001370 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001371 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001372 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001373 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374 spin_unlock_irq(&zone->lru_lock);
1375
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 while (!list_empty(&l_hold)) {
1377 cond_resched();
1378 page = lru_to_page(&l_hold);
1379 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001380
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001381 if (unlikely(!page_evictable(page, NULL))) {
1382 putback_lru_page(page);
1383 continue;
1384 }
1385
Johannes Weiner64574742010-03-05 13:42:22 -08001386 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001387 nr_rotated++;
Wu Fengguang8cab4752009-06-16 15:33:12 -07001388 /*
1389 * Identify referenced, file-backed active pages and
1390 * give them one more trip around the active list. So
1391 * that executable code get better chances to stay in
1392 * memory under moderate memory pressure. Anon pages
1393 * are not likely to be evicted by use-once streaming
1394 * IO, plus JVM can create lots of anon VM_EXEC pages,
1395 * so we ignore them here.
1396 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001397 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001398 list_add(&page->lru, &l_active);
1399 continue;
1400 }
1401 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001402
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001403 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404 list_add(&page->lru, &l_inactive);
1405 }
1406
Andrew Mortonb5557492009-01-06 14:40:13 -08001407 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001408 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001409 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001410 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001411 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001412 * Count referenced pages from currently used mappings as rotated,
1413 * even though only some of them are actually re-activated. This
1414 * helps balance scan pressure between file and anonymous pages in
1415 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001416 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001417 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001418
Wu Fengguang3eb41402009-06-16 15:33:13 -07001419 move_active_pages_to_lru(zone, &l_active,
1420 LRU_ACTIVE + file * LRU_FILE);
1421 move_active_pages_to_lru(zone, &l_inactive,
1422 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001423 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001424 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001425}
1426
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001427static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001428{
1429 unsigned long active, inactive;
1430
1431 active = zone_page_state(zone, NR_ACTIVE_ANON);
1432 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1433
1434 if (inactive * zone->inactive_ratio < active)
1435 return 1;
1436
1437 return 0;
1438}
1439
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001440/**
1441 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1442 * @zone: zone to check
1443 * @sc: scan control of this context
1444 *
1445 * Returns true if the zone does not have enough inactive anon pages,
1446 * meaning some active anon pages need to be deactivated.
1447 */
1448static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1449{
1450 int low;
1451
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001452 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001453 low = inactive_anon_is_low_global(zone);
1454 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001455 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001456 return low;
1457}
1458
Rik van Riel56e49d22009-06-16 15:32:28 -07001459static int inactive_file_is_low_global(struct zone *zone)
1460{
1461 unsigned long active, inactive;
1462
1463 active = zone_page_state(zone, NR_ACTIVE_FILE);
1464 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1465
1466 return (active > inactive);
1467}
1468
1469/**
1470 * inactive_file_is_low - check if file pages need to be deactivated
1471 * @zone: zone to check
1472 * @sc: scan control of this context
1473 *
1474 * When the system is doing streaming IO, memory pressure here
1475 * ensures that active file pages get deactivated, until more
1476 * than half of the file pages are on the inactive list.
1477 *
1478 * Once we get to that situation, protect the system's working
1479 * set from being evicted by disabling active file page aging.
1480 *
1481 * This uses a different ratio than the anonymous pages, because
1482 * the page cache uses a use-once replacement algorithm.
1483 */
1484static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1485{
1486 int low;
1487
1488 if (scanning_global_lru(sc))
1489 low = inactive_file_is_low_global(zone);
1490 else
1491 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1492 return low;
1493}
1494
Rik van Rielb39415b2009-12-14 17:59:48 -08001495static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1496 int file)
1497{
1498 if (file)
1499 return inactive_file_is_low(zone, sc);
1500 else
1501 return inactive_anon_is_low(zone, sc);
1502}
1503
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001504static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001505 struct zone *zone, struct scan_control *sc, int priority)
1506{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001507 int file = is_file_lru(lru);
1508
Rik van Rielb39415b2009-12-14 17:59:48 -08001509 if (is_active_lru(lru)) {
1510 if (inactive_list_is_low(zone, sc, file))
1511 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001512 return 0;
1513 }
1514
Rik van Riel33c120e2008-10-18 20:26:36 -07001515 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001516}
1517
Linus Torvalds1da177e2005-04-16 15:20:36 -07001518/*
Shaohua Li76a33fc2010-05-24 14:32:36 -07001519 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1520 * until we collected @swap_cluster_max pages to scan.
1521 */
1522static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1523 unsigned long *nr_saved_scan)
1524{
1525 unsigned long nr;
1526
1527 *nr_saved_scan += nr_to_scan;
1528 nr = *nr_saved_scan;
1529
1530 if (nr >= SWAP_CLUSTER_MAX)
1531 *nr_saved_scan = 0;
1532 else
1533 nr = 0;
1534
1535 return nr;
1536}
1537
1538/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001539 * Determine how aggressively the anon and file LRU lists should be
1540 * scanned. The relative value of each set of LRU lists is determined
1541 * by looking at the fraction of the pages scanned we did rotate back
1542 * onto the active list instead of evict.
1543 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001544 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001545 */
Shaohua Li76a33fc2010-05-24 14:32:36 -07001546static void get_scan_count(struct zone *zone, struct scan_control *sc,
1547 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001548{
1549 unsigned long anon, file, free;
1550 unsigned long anon_prio, file_prio;
1551 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001552 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001553 u64 fraction[2], denominator;
1554 enum lru_list l;
1555 int noswap = 0;
1556
1557 /* If we have no swap space, do not bother scanning anon pages. */
1558 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1559 noswap = 1;
1560 fraction[0] = 0;
1561 fraction[1] = 1;
1562 denominator = 1;
1563 goto out;
1564 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001565
Vincent Li0b217672009-09-21 17:03:09 -07001566 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1567 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1568 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1569 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001570
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001571 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001572 free = zone_page_state(zone, NR_FREE_PAGES);
1573 /* If we have very few page cache pages,
1574 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001575 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001576 fraction[0] = 1;
1577 fraction[1] = 0;
1578 denominator = 1;
1579 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001580 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001581 }
1582
1583 /*
1584 * OK, so we have swap space and a fair amount of page cache
1585 * pages. We use the recently rotated / recently scanned
1586 * ratios to determine how valuable each cache is.
1587 *
1588 * Because workloads change over time (and to avoid overflow)
1589 * we keep these statistics as a floating average, which ends
1590 * up weighing recent references more than old ones.
1591 *
1592 * anon in [0], file in [1]
1593 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001594 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001595 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001596 reclaim_stat->recent_scanned[0] /= 2;
1597 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001598 spin_unlock_irq(&zone->lru_lock);
1599 }
1600
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001601 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001602 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001603 reclaim_stat->recent_scanned[1] /= 2;
1604 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001605 spin_unlock_irq(&zone->lru_lock);
1606 }
1607
1608 /*
1609 * With swappiness at 100, anonymous and file have the same priority.
1610 * This scanning priority is essentially the inverse of IO cost.
1611 */
1612 anon_prio = sc->swappiness;
1613 file_prio = 200 - sc->swappiness;
1614
1615 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001616 * The amount of pressure on anon vs file pages is inversely
1617 * proportional to the fraction of recently scanned pages on
1618 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001619 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001620 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1621 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001622
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001623 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1624 fp /= reclaim_stat->recent_rotated[1] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001625
Shaohua Li76a33fc2010-05-24 14:32:36 -07001626 fraction[0] = ap;
1627 fraction[1] = fp;
1628 denominator = ap + fp + 1;
1629out:
1630 for_each_evictable_lru(l) {
1631 int file = is_file_lru(l);
1632 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001633
Shaohua Li76a33fc2010-05-24 14:32:36 -07001634 scan = zone_nr_lru_pages(zone, sc, l);
1635 if (priority || noswap) {
1636 scan >>= priority;
1637 scan = div64_u64(scan * fraction[file], denominator);
1638 }
1639 nr[l] = nr_scan_try_batch(scan,
1640 &reclaim_stat->nr_saved_scan[l]);
1641 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001642}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001643
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -07001644static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc)
1645{
1646 /*
1647 * If we need a large contiguous chunk of memory, or have
1648 * trouble getting a small set of contiguous pages, we
1649 * will reclaim both active and inactive pages.
1650 */
1651 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1652 sc->lumpy_reclaim_mode = 1;
1653 else if (sc->order && priority < DEF_PRIORITY - 2)
1654 sc->lumpy_reclaim_mode = 1;
1655 else
1656 sc->lumpy_reclaim_mode = 0;
1657}
1658
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001659/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1661 */
Rik van Riela79311c2009-01-06 14:40:01 -08001662static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001663 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001664{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001665 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001666 unsigned long nr_to_scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001667 enum lru_list l;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001668 unsigned long nr_reclaimed = sc->nr_reclaimed;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001669 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670
Shaohua Li76a33fc2010-05-24 14:32:36 -07001671 get_scan_count(zone, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001672
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -07001673 set_lumpy_reclaim_mode(priority, sc);
1674
Rik van Riel556adec2008-10-18 20:26:34 -07001675 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1676 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001677 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001678 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001679 nr_to_scan = min_t(unsigned long,
1680 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001681 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001682
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001683 nr_reclaimed += shrink_list(l, nr_to_scan,
1684 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001685 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001686 }
Rik van Riela79311c2009-01-06 14:40:01 -08001687 /*
1688 * On large memory systems, scan >> priority can become
1689 * really large. This is fine for the starting priority;
1690 * we want to put equal scanning pressure on each zone.
1691 * However, if the VM has a harder time of freeing pages,
1692 * with multiple processes reclaiming pages, the total
1693 * freeing target can get unreasonably large.
1694 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08001695 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001696 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 }
1698
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001699 sc->nr_reclaimed = nr_reclaimed;
1700
Rik van Riel556adec2008-10-18 20:26:34 -07001701 /*
1702 * Even if we did not try to evict anon pages at all, we want to
1703 * rebalance the anon lru active/inactive ratio.
1704 */
MinChan Kim69c85482009-06-16 15:32:44 -07001705 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
Rik van Riel556adec2008-10-18 20:26:34 -07001706 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1707
Andrew Morton232ea4d2007-02-28 20:13:21 -08001708 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709}
1710
1711/*
1712 * This is the direct reclaim path, for page-allocating processes. We only
1713 * try to reclaim pages from zones which will satisfy the caller's allocation
1714 * request.
1715 *
Mel Gorman41858962009-06-16 15:32:12 -07001716 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1717 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001718 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1719 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001720 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1721 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1722 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001724 * If a zone is deemed to be full of pinned pages then just give it a light
1725 * scan then give up on it.
1726 */
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001727static bool shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001728 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001729{
Mel Gorman54a6eb52008-04-28 02:12:16 -07001730 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Mel Gormandd1a2392008-04-28 02:12:17 -07001731 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001732 struct zone *zone;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001733 bool all_unreclaimable = true;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001734
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001735 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1736 sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001737 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001738 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001739 /*
1740 * Take care memory controller reclaiming has small influence
1741 * to global LRU.
1742 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001743 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001744 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1745 continue;
1746 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001747
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08001748 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001749 continue; /* Let kswapd poll it */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001750 } else {
1751 /*
1752 * Ignore cpuset limitation here. We just want to reduce
1753 * # of used pages by us regardless of memory shortage.
1754 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001755 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1756 priority);
1757 }
Nick Piggin408d8542006-09-25 23:31:27 -07001758
Rik van Riela79311c2009-01-06 14:40:01 -08001759 shrink_zone(priority, zone, sc);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001760 all_unreclaimable = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001762 return all_unreclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001764
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765/*
1766 * This is the main entry point to direct page reclaim.
1767 *
1768 * If a full scan of the inactive list fails to free enough memory then we
1769 * are "out of memory" and something needs to be killed.
1770 *
1771 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1772 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02001773 * caller can't do much about. We kick the writeback threads and take explicit
1774 * naps in the hope that some of these pages can be written. But if the
1775 * allocating task holds filesystem locks which prevent writeout this might not
1776 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001777 *
1778 * returns: 0, if no pages reclaimed
1779 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001780 */
Mel Gormandac1d272008-04-28 02:12:12 -07001781static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001782 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001783{
1784 int priority;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001785 bool all_unreclaimable;
Andrew Morton69e05942006-03-22 00:08:19 -08001786 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788 unsigned long lru_pages = 0;
Mel Gormandd1a2392008-04-28 02:12:17 -07001789 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001790 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -07001791 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001792 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001793
Miao Xiec0ff7452010-05-24 14:32:08 -07001794 get_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07001795 delayacct_freepages_start();
1796
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001797 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001798 count_vm_event(ALLOCSTALL);
1799 /*
1800 * mem_cgroup will not do shrink_slab.
1801 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001802 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001803 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001804
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001805 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1806 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807
Wu Fengguangadea02a2009-09-21 17:01:42 -07001808 lru_pages += zone_reclaimable_pages(zone);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001809 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810 }
1811
1812 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001813 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001814 if (!priority)
1815 disable_swap_token();
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001816 all_unreclaimable = shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001817 /*
1818 * Don't shrink slabs when reclaiming memory from
1819 * over limit cgroups
1820 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001821 if (scanning_global_lru(sc)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07001822 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001823 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08001824 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001825 reclaim_state->reclaimed_slab = 0;
1826 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 }
Balbir Singh66e17072008-02-07 00:13:56 -08001828 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001829 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001830 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001831
1832 /*
1833 * Try to write back as many pages as we just scanned. This
1834 * tends to cause slow streaming writers to write data to the
1835 * disk smoothly, at the dirtying rate, which is nice. But
1836 * that's undesirable in laptop mode, where we *want* lumpy
1837 * writeout. So in laptop mode, write out the whole world.
1838 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001839 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1840 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001841 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001842 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 }
1844
1845 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08001846 if (!sc->hibernation_mode && sc->nr_scanned &&
1847 priority < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02001848 congestion_wait(BLK_RW_ASYNC, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001849 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001850
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07001852 /*
1853 * Now that we've scanned all the zones at this priority level, note
1854 * that level within the zone so that the next thread which performs
1855 * scanning of this zone will immediately start out at this priority
1856 * level. This affects only the decision whether or not to bring
1857 * mapped pages onto the inactive list.
1858 */
1859 if (priority < 0)
1860 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001862 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001863 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001865 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1866 continue;
1867
1868 zone->prev_priority = priority;
1869 }
1870 } else
1871 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1872
Keika Kobayashi873b4772008-07-25 01:48:52 -07001873 delayacct_freepages_end();
Miao Xiec0ff7452010-05-24 14:32:08 -07001874 put_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07001875
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001876 if (sc->nr_reclaimed)
1877 return sc->nr_reclaimed;
1878
1879 /* top priority shrink_zones still had more to do? don't OOM, then */
1880 if (scanning_global_lru(sc) && !all_unreclaimable)
1881 return 1;
1882
1883 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001884}
1885
Mel Gormandac1d272008-04-28 02:12:12 -07001886unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001887 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08001888{
1889 struct scan_control sc = {
1890 .gfp_mask = gfp_mask,
1891 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001892 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001893 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001894 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08001895 .swappiness = vm_swappiness,
1896 .order = order,
1897 .mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001898 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08001899 };
1900
Mel Gormandd1a2392008-04-28 02:12:17 -07001901 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001902}
1903
Balbir Singh00f0b822008-03-04 14:28:39 -08001904#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08001905
Balbir Singh4e416952009-09-23 15:56:39 -07001906unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1907 gfp_t gfp_mask, bool noswap,
1908 unsigned int swappiness,
1909 struct zone *zone, int nid)
1910{
1911 struct scan_control sc = {
1912 .may_writepage = !laptop_mode,
1913 .may_unmap = 1,
1914 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07001915 .swappiness = swappiness,
1916 .order = 0,
1917 .mem_cgroup = mem,
Balbir Singh4e416952009-09-23 15:56:39 -07001918 };
1919 nodemask_t nm = nodemask_of_node(nid);
1920
1921 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1922 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1923 sc.nodemask = &nm;
1924 sc.nr_reclaimed = 0;
1925 sc.nr_scanned = 0;
1926 /*
1927 * NOTE: Although we can get the priority field, using it
1928 * here is not a good idea, since it limits the pages we can scan.
1929 * if we don't reclaim here, the shrink_zone from balance_pgdat
1930 * will pick up pages from other mem cgroup's as well. We hack
1931 * the priority and make it zero.
1932 */
1933 shrink_zone(0, zone, &sc);
1934 return sc.nr_reclaimed;
1935}
1936
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001937unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001938 gfp_t gfp_mask,
1939 bool noswap,
1940 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08001941{
Balbir Singh4e416952009-09-23 15:56:39 -07001942 struct zonelist *zonelist;
Balbir Singh66e17072008-02-07 00:13:56 -08001943 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08001944 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001945 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001946 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001947 .nr_to_reclaim = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001948 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08001949 .order = 0,
1950 .mem_cgroup = mem_cont,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001951 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08001952 };
Balbir Singh66e17072008-02-07 00:13:56 -08001953
Mel Gormandd1a2392008-04-28 02:12:17 -07001954 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1955 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1956 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1957 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001958}
1959#endif
1960
Mel Gormanf50de2d2009-12-14 17:58:53 -08001961/* is kswapd sleeping prematurely? */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001962static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
Mel Gormanf50de2d2009-12-14 17:58:53 -08001963{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001964 int i;
Mel Gormanf50de2d2009-12-14 17:58:53 -08001965
1966 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1967 if (remaining)
1968 return 1;
1969
1970 /* If after HZ/10, a zone is below the high mark, it's premature */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001971 for (i = 0; i < pgdat->nr_zones; i++) {
1972 struct zone *zone = pgdat->node_zones + i;
1973
1974 if (!populated_zone(zone))
1975 continue;
1976
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08001977 if (zone->all_unreclaimable)
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08001978 continue;
1979
Mel Gormanf50de2d2009-12-14 17:58:53 -08001980 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
1981 0, 0))
1982 return 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001983 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08001984
1985 return 0;
1986}
1987
Linus Torvalds1da177e2005-04-16 15:20:36 -07001988/*
1989 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07001990 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992 * Returns the number of pages which were actually freed.
1993 *
1994 * There is special handling here for zones which are full of pinned pages.
1995 * This can happen if the pages are all mlocked, or if they are all used by
1996 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1997 * What we do is to detect the case where all pages in the zone have been
1998 * scanned twice and there has been zero successful reclaim. Mark the zone as
1999 * dead and from now on, only perform a short scan. Basically we're polling
2000 * the zone for when the problem goes away.
2001 *
2002 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002003 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2004 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2005 * lower zones regardless of the number of free pages in the lower zones. This
2006 * interoperates with the page allocator fallback scheme to ensure that aging
2007 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002009static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002011 int all_zones_ok;
2012 int priority;
2013 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08002014 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08002016 struct scan_control sc = {
2017 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002018 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002019 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002020 /*
2021 * kswapd doesn't want to be bailed out while reclaim. because
2022 * we want to put equal scanning pressure on each zone.
2023 */
2024 .nr_to_reclaim = ULONG_MAX,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002025 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002026 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002027 .mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002028 };
Martin Bligh3bb1a852006-10-28 10:38:24 -07002029 /*
2030 * temp_priority is used to remember the scanning priority at which
Mel Gorman41858962009-06-16 15:32:12 -07002031 * this zone was successfully refilled to
2032 * free_pages == high_wmark_pages(zone).
Martin Bligh3bb1a852006-10-28 10:38:24 -07002033 */
2034 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002035
2036loop_again:
2037 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002038 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002039 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002040 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002041
Martin Bligh3bb1a852006-10-28 10:38:24 -07002042 for (i = 0; i < pgdat->nr_zones; i++)
2043 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044
2045 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2046 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2047 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002048 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002049
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002050 /* The swap token gets in the way of swapout... */
2051 if (!priority)
2052 disable_swap_token();
2053
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054 all_zones_ok = 1;
2055
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002056 /*
2057 * Scan in the highmem->dma direction for the highest
2058 * zone which needs scanning
2059 */
2060 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2061 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002062
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002063 if (!populated_zone(zone))
2064 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002065
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002066 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002067 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002068
Rik van Riel556adec2008-10-18 20:26:34 -07002069 /*
2070 * Do some background aging of the anon list, to give
2071 * pages a chance to be referenced before reclaiming.
2072 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002073 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002074 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2075 &sc, priority, 0);
2076
Mel Gorman41858962009-06-16 15:32:12 -07002077 if (!zone_watermark_ok(zone, order,
2078 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002079 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002080 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002081 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002082 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002083 if (i < 0)
2084 goto out;
2085
Linus Torvalds1da177e2005-04-16 15:20:36 -07002086 for (i = 0; i <= end_zone; i++) {
2087 struct zone *zone = pgdat->node_zones + i;
2088
Wu Fengguangadea02a2009-09-21 17:01:42 -07002089 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090 }
2091
2092 /*
2093 * Now scan the zone in the dma->highmem direction, stopping
2094 * at the last zone which needs scanning.
2095 *
2096 * We do this because the page allocator works in the opposite
2097 * direction. This prevents the page allocator from allocating
2098 * pages behind kswapd's direction of progress, which would
2099 * cause too much scanning of the lower zones.
2100 */
2101 for (i = 0; i <= end_zone; i++) {
2102 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002103 int nr_slab;
Balbir Singh4e416952009-09-23 15:56:39 -07002104 int nid, zid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002105
Con Kolivasf3fe6512006-01-06 00:11:15 -08002106 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 continue;
2108
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002109 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110 continue;
2111
Martin Bligh3bb1a852006-10-28 10:38:24 -07002112 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002113 sc.nr_scanned = 0;
Martin Bligh3bb1a852006-10-28 10:38:24 -07002114 note_zone_scanning_priority(zone, priority);
Balbir Singh4e416952009-09-23 15:56:39 -07002115
2116 nid = pgdat->node_id;
2117 zid = zone_idx(zone);
2118 /*
2119 * Call soft limit reclaim before calling shrink_zone.
2120 * For now we ignore the return value
2121 */
2122 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2123 nid, zid);
Rik van Riel32a43302007-10-16 01:24:50 -07002124 /*
2125 * We put equal pressure on every zone, unless one
2126 * zone has way too many pages free already.
2127 */
Mel Gorman41858962009-06-16 15:32:12 -07002128 if (!zone_watermark_ok(zone, order,
2129 8*high_wmark_pages(zone), end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002130 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002131 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002132 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2133 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002134 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002135 total_scanned += sc.nr_scanned;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002136 if (zone->all_unreclaimable)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002138 if (nr_slab == 0 &&
2139 zone->pages_scanned >= (zone_reclaimable_pages(zone) * 6))
2140 zone->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002141 /*
2142 * If we've done a decent amount of scanning and
2143 * the reclaim ratio is low, start doing writepage
2144 * even in laptop mode
2145 */
2146 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002147 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002148 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002149
Minchan Kim45973d72010-03-05 13:41:45 -08002150 if (!zone_watermark_ok(zone, order,
2151 high_wmark_pages(zone), end_zone, 0)) {
2152 all_zones_ok = 0;
2153 /*
2154 * We are still under min water mark. This
2155 * means that we have a GFP_ATOMIC allocation
2156 * failure risk. Hurry up!
2157 */
2158 if (!zone_watermark_ok(zone, order,
2159 min_wmark_pages(zone), end_zone, 0))
2160 has_under_min_watermark_zone = 1;
2161 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002162
Linus Torvalds1da177e2005-04-16 15:20:36 -07002163 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002164 if (all_zones_ok)
2165 break; /* kswapd: all done */
2166 /*
2167 * OK, kswapd is getting into trouble. Take a nap, then take
2168 * another pass across the zones.
2169 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002170 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2171 if (has_under_min_watermark_zone)
2172 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2173 else
2174 congestion_wait(BLK_RW_ASYNC, HZ/10);
2175 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176
2177 /*
2178 * We do this so kswapd doesn't build up large priorities for
2179 * example when it is freeing in parallel with allocators. It
2180 * matches the direct reclaim path behaviour in terms of impact
2181 * on zone->*_priority.
2182 */
Rik van Riela79311c2009-01-06 14:40:01 -08002183 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184 break;
2185 }
2186out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07002187 /*
2188 * Note within each zone the priority level at which this zone was
2189 * brought into a happy state. So that the next thread which scans this
2190 * zone will start out at that priority level.
2191 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192 for (i = 0; i < pgdat->nr_zones; i++) {
2193 struct zone *zone = pgdat->node_zones + i;
2194
Martin Bligh3bb1a852006-10-28 10:38:24 -07002195 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196 }
2197 if (!all_zones_ok) {
2198 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002199
2200 try_to_freeze();
2201
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002202 /*
2203 * Fragmentation may mean that the system cannot be
2204 * rebalanced for high-order allocations in all zones.
2205 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2206 * it means the zones have been fully scanned and are still
2207 * not balanced. For high-order allocations, there is
2208 * little point trying all over again as kswapd may
2209 * infinite loop.
2210 *
2211 * Instead, recheck all watermarks at order-0 as they
2212 * are the most important. If watermarks are ok, kswapd will go
2213 * back to sleep. High-order users can still perform direct
2214 * reclaim if they wish.
2215 */
2216 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2217 order = sc.order = 0;
2218
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 goto loop_again;
2220 }
2221
Rik van Riela79311c2009-01-06 14:40:01 -08002222 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002223}
2224
2225/*
2226 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002227 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002228 *
2229 * This basically trickles out pages so that we have _some_
2230 * free memory available even if there is no other activity
2231 * that frees anything up. This is needed for things like routing
2232 * etc, where we otherwise might have all activity going on in
2233 * asynchronous contexts that cannot page things out.
2234 *
2235 * If there are applications that are active memory-allocators
2236 * (most normal use), this basically shouldn't matter.
2237 */
2238static int kswapd(void *p)
2239{
2240 unsigned long order;
2241 pg_data_t *pgdat = (pg_data_t*)p;
2242 struct task_struct *tsk = current;
2243 DEFINE_WAIT(wait);
2244 struct reclaim_state reclaim_state = {
2245 .reclaimed_slab = 0,
2246 };
Rusty Russella70f7302009-03-13 14:49:46 +10302247 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002248
Nick Piggincf40bd12009-01-21 08:12:39 +01002249 lockdep_set_current_reclaim_state(GFP_KERNEL);
2250
Rusty Russell174596a2009-01-01 10:12:29 +10302251 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002252 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253 current->reclaim_state = &reclaim_state;
2254
2255 /*
2256 * Tell the memory management that we're a "memory allocator",
2257 * and that if we need more memory we should get access to it
2258 * regardless (see "__alloc_pages()"). "kswapd" should
2259 * never get caught in the normal page freeing logic.
2260 *
2261 * (Kswapd normally doesn't need memory anyway, but sometimes
2262 * you need a small amount of memory in order to be able to
2263 * page out something else, and this flag essentially protects
2264 * us from recursively trying to free more memory as we're
2265 * trying to free the first piece of memory in the first place).
2266 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002267 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002268 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269
2270 order = 0;
2271 for ( ; ; ) {
2272 unsigned long new_order;
David Rientjes8fe23e02009-12-14 17:58:33 -08002273 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002274
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2276 new_order = pgdat->kswapd_max_order;
2277 pgdat->kswapd_max_order = 0;
2278 if (order < new_order) {
2279 /*
2280 * Don't sleep if someone wants a larger 'order'
2281 * allocation
2282 */
2283 order = new_order;
2284 } else {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002285 if (!freezing(current) && !kthread_should_stop()) {
2286 long remaining = 0;
2287
2288 /* Try to sleep for a short interval */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002289 if (!sleeping_prematurely(pgdat, order, remaining)) {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002290 remaining = schedule_timeout(HZ/10);
2291 finish_wait(&pgdat->kswapd_wait, &wait);
2292 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2293 }
2294
2295 /*
2296 * After a short sleep, check if it was a
2297 * premature sleep. If not, then go fully
2298 * to sleep until explicitly woken up
2299 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002300 if (!sleeping_prematurely(pgdat, order, remaining))
Mel Gormanf50de2d2009-12-14 17:58:53 -08002301 schedule();
2302 else {
2303 if (remaining)
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002304 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002305 else
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002306 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002307 }
2308 }
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002309
Linus Torvalds1da177e2005-04-16 15:20:36 -07002310 order = pgdat->kswapd_max_order;
2311 }
2312 finish_wait(&pgdat->kswapd_wait, &wait);
2313
David Rientjes8fe23e02009-12-14 17:58:33 -08002314 ret = try_to_freeze();
2315 if (kthread_should_stop())
2316 break;
2317
2318 /*
2319 * We can speed up thawing tasks if we don't call balance_pgdat
2320 * after returning from the refrigerator
2321 */
2322 if (!ret)
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002323 balance_pgdat(pgdat, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002324 }
2325 return 0;
2326}
2327
2328/*
2329 * A zone is low on free memory, so wake its kswapd task to service it.
2330 */
2331void wakeup_kswapd(struct zone *zone, int order)
2332{
2333 pg_data_t *pgdat;
2334
Con Kolivasf3fe6512006-01-06 00:11:15 -08002335 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 return;
2337
2338 pgdat = zone->zone_pgdat;
Mel Gorman41858962009-06-16 15:32:12 -07002339 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340 return;
2341 if (pgdat->kswapd_max_order < order)
2342 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08002343 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002345 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002347 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348}
2349
Wu Fengguangadea02a2009-09-21 17:01:42 -07002350/*
2351 * The reclaimable count would be mostly accurate.
2352 * The less reclaimable pages may be
2353 * - mlocked pages, which will be moved to unevictable list when encountered
2354 * - mapped pages, which may require several travels to be reclaimed
2355 * - dirty pages, which is not "instantly" reclaimable
2356 */
2357unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002358{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002359 int nr;
2360
2361 nr = global_page_state(NR_ACTIVE_FILE) +
2362 global_page_state(NR_INACTIVE_FILE);
2363
2364 if (nr_swap_pages > 0)
2365 nr += global_page_state(NR_ACTIVE_ANON) +
2366 global_page_state(NR_INACTIVE_ANON);
2367
2368 return nr;
2369}
2370
2371unsigned long zone_reclaimable_pages(struct zone *zone)
2372{
2373 int nr;
2374
2375 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2376 zone_page_state(zone, NR_INACTIVE_FILE);
2377
2378 if (nr_swap_pages > 0)
2379 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2380 zone_page_state(zone, NR_INACTIVE_ANON);
2381
2382 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002383}
2384
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002385#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002386/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002387 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002388 * freed pages.
2389 *
2390 * Rather than trying to age LRUs the aim is to preserve the overall
2391 * LRU order by reclaiming preferentially
2392 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002394unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002396 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002397 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002398 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2399 .may_swap = 1,
2400 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002401 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002402 .nr_to_reclaim = nr_to_reclaim,
2403 .hibernation_mode = 1,
2404 .swappiness = vm_swappiness,
2405 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406 };
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002407 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2408 struct task_struct *p = current;
2409 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002410
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002411 p->flags |= PF_MEMALLOC;
2412 lockdep_set_current_reclaim_state(sc.gfp_mask);
2413 reclaim_state.reclaimed_slab = 0;
2414 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002415
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002416 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002417
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002418 p->reclaim_state = NULL;
2419 lockdep_clear_current_reclaim_state();
2420 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002421
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002422 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002423}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002424#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426/* It's optimal to keep kswapds on the same CPUs as their memory, but
2427 not required for correctness. So if the last cpu in a node goes
2428 away, we get changed to run anywhere: as the first one comes back,
2429 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002430static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002431 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002433 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002435 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002436 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002437 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302438 const struct cpumask *mask;
2439
2440 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002441
Rusty Russell3e597942009-01-01 10:12:24 +10302442 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002444 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 }
2446 }
2447 return NOTIFY_OK;
2448}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449
Yasunori Goto3218ae12006-06-27 02:53:33 -07002450/*
2451 * This kswapd start function will be called by init and node-hot-add.
2452 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2453 */
2454int kswapd_run(int nid)
2455{
2456 pg_data_t *pgdat = NODE_DATA(nid);
2457 int ret = 0;
2458
2459 if (pgdat->kswapd)
2460 return 0;
2461
2462 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2463 if (IS_ERR(pgdat->kswapd)) {
2464 /* failure at boot is fatal */
2465 BUG_ON(system_state == SYSTEM_BOOTING);
2466 printk("Failed to start kswapd on node %d\n",nid);
2467 ret = -1;
2468 }
2469 return ret;
2470}
2471
David Rientjes8fe23e02009-12-14 17:58:33 -08002472/*
2473 * Called by memory hotplug when all memory in a node is offlined.
2474 */
2475void kswapd_stop(int nid)
2476{
2477 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2478
2479 if (kswapd)
2480 kthread_stop(kswapd);
2481}
2482
Linus Torvalds1da177e2005-04-16 15:20:36 -07002483static int __init kswapd_init(void)
2484{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002485 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002486
Linus Torvalds1da177e2005-04-16 15:20:36 -07002487 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002488 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002489 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490 hotcpu_notifier(cpu_callback, 0);
2491 return 0;
2492}
2493
2494module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002495
2496#ifdef CONFIG_NUMA
2497/*
2498 * Zone reclaim mode
2499 *
2500 * If non-zero call zone_reclaim when the number of free pages falls below
2501 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002502 */
2503int zone_reclaim_mode __read_mostly;
2504
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002505#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002506#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002507#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2508#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2509
Christoph Lameter9eeff232006-01-18 17:42:31 -08002510/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002511 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2512 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2513 * a zone.
2514 */
2515#define ZONE_RECLAIM_PRIORITY 4
2516
Christoph Lameter9eeff232006-01-18 17:42:31 -08002517/*
Christoph Lameter96146342006-07-03 00:24:13 -07002518 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2519 * occur.
2520 */
2521int sysctl_min_unmapped_ratio = 1;
2522
2523/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002524 * If the number of slab pages in a zone grows beyond this percentage then
2525 * slab reclaim needs to occur.
2526 */
2527int sysctl_min_slab_ratio = 5;
2528
Mel Gorman90afa5d2009-06-16 15:33:20 -07002529static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2530{
2531 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2532 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2533 zone_page_state(zone, NR_ACTIVE_FILE);
2534
2535 /*
2536 * It's possible for there to be more file mapped pages than
2537 * accounted for by the pages on the file LRU lists because
2538 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2539 */
2540 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2541}
2542
2543/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2544static long zone_pagecache_reclaimable(struct zone *zone)
2545{
2546 long nr_pagecache_reclaimable;
2547 long delta = 0;
2548
2549 /*
2550 * If RECLAIM_SWAP is set, then all file pages are considered
2551 * potentially reclaimable. Otherwise, we have to worry about
2552 * pages like swapcache and zone_unmapped_file_pages() provides
2553 * a better estimate
2554 */
2555 if (zone_reclaim_mode & RECLAIM_SWAP)
2556 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2557 else
2558 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2559
2560 /* If we can't clean pages, remove dirty pages from consideration */
2561 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2562 delta += zone_page_state(zone, NR_FILE_DIRTY);
2563
2564 /* Watch for any possible underflows due to delta */
2565 if (unlikely(delta > nr_pagecache_reclaimable))
2566 delta = nr_pagecache_reclaimable;
2567
2568 return nr_pagecache_reclaimable - delta;
2569}
2570
Christoph Lameter0ff38492006-09-25 23:31:52 -07002571/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002572 * Try to free up some pages from this zone through reclaim.
2573 */
Andrew Morton179e9632006-03-22 00:08:18 -08002574static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002575{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002576 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002577 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002578 struct task_struct *p = current;
2579 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002580 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002581 struct scan_control sc = {
2582 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002583 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002584 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002585 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2586 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002587 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002588 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002589 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08002590 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07002591 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002592
2593 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08002594 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002595 /*
2596 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2597 * and we also need to be able to write out pages for RECLAIM_WRITE
2598 * and RECLAIM_SWAP.
2599 */
2600 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002601 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08002602 reclaim_state.reclaimed_slab = 0;
2603 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002604
Mel Gorman90afa5d2009-06-16 15:33:20 -07002605 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07002606 /*
2607 * Free memory by calling shrink zone with increasing
2608 * priorities until we have enough memory freed.
2609 */
2610 priority = ZONE_RECLAIM_PRIORITY;
2611 do {
Martin Bligh3bb1a852006-10-28 10:38:24 -07002612 note_zone_scanning_priority(zone, priority);
Rik van Riela79311c2009-01-06 14:40:01 -08002613 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002614 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002615 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002616 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002617
Christoph Lameter83e33a42006-09-25 23:31:53 -07002618 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2619 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002620 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002621 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002622 * many pages were freed in this zone. So we take the current
2623 * number of slab pages and shake the slab until it is reduced
2624 * by the same nr_pages that we used for reclaiming unmapped
2625 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002626 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002627 * Note that shrink_slab will free memory on all zones and may
2628 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002629 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07002630 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07002631 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2632 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07002633 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07002634
2635 /*
2636 * Update nr_reclaimed by the number of slab pages we
2637 * reclaimed from this zone.
2638 */
Rik van Riela79311c2009-01-06 14:40:01 -08002639 sc.nr_reclaimed += slab_reclaimable -
Christoph Lameter83e33a42006-09-25 23:31:53 -07002640 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002641 }
2642
Christoph Lameter9eeff232006-01-18 17:42:31 -08002643 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002644 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002645 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08002646 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002647}
Andrew Morton179e9632006-03-22 00:08:18 -08002648
2649int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2650{
Andrew Morton179e9632006-03-22 00:08:18 -08002651 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002652 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002653
2654 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002655 * Zone reclaim reclaims unmapped file backed pages and
2656 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002657 *
Christoph Lameter96146342006-07-03 00:24:13 -07002658 * A small portion of unmapped file backed pages is needed for
2659 * file I/O otherwise pages read by file I/O will be immediately
2660 * thrown out if the zone is overallocated. So we do not reclaim
2661 * if less than a specified percentage of the zone is used by
2662 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002663 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07002664 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2665 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002666 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08002667
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002668 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002669 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07002670
Andrew Morton179e9632006-03-22 00:08:18 -08002671 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002672 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002673 */
David Rientjesd773ed62007-10-16 23:26:01 -07002674 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002675 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08002676
2677 /*
2678 * Only run zone reclaim on the local zone or on zones that do not
2679 * have associated processors. This will favor the local processor
2680 * over remote processors and spread off node memory allocations
2681 * as wide as possible.
2682 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002683 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002684 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07002685 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07002686
2687 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002688 return ZONE_RECLAIM_NOSCAN;
2689
David Rientjesd773ed62007-10-16 23:26:01 -07002690 ret = __zone_reclaim(zone, gfp_mask, order);
2691 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2692
Mel Gorman24cf725182009-06-16 15:33:23 -07002693 if (!ret)
2694 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2695
David Rientjesd773ed62007-10-16 23:26:01 -07002696 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002697}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002698#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002699
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002700/*
2701 * page_evictable - test whether a page is evictable
2702 * @page: the page to test
2703 * @vma: the VMA in which the page is or will be mapped, may be NULL
2704 *
2705 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07002706 * lists vs unevictable list. The vma argument is !NULL when called from the
2707 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002708 *
2709 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002710 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07002711 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002712 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002713 */
2714int page_evictable(struct page *page, struct vm_area_struct *vma)
2715{
2716
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002717 if (mapping_unevictable(page_mapping(page)))
2718 return 0;
2719
Nick Pigginb291f002008-10-18 20:26:44 -07002720 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2721 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002722
2723 return 1;
2724}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002725
2726/**
2727 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2728 * @page: page to check evictability and move to appropriate lru list
2729 * @zone: zone page is in
2730 *
2731 * Checks a page for evictability and moves the page to the appropriate
2732 * zone lru list.
2733 *
2734 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2735 * have PageUnevictable set.
2736 */
2737static void check_move_unevictable_page(struct page *page, struct zone *zone)
2738{
2739 VM_BUG_ON(PageActive(page));
2740
2741retry:
2742 ClearPageUnevictable(page);
2743 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07002744 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002745
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002746 __dec_zone_state(zone, NR_UNEVICTABLE);
2747 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002748 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002749 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2750 __count_vm_event(UNEVICTABLE_PGRESCUED);
2751 } else {
2752 /*
2753 * rotate unevictable list
2754 */
2755 SetPageUnevictable(page);
2756 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002757 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002758 if (page_evictable(page, NULL))
2759 goto retry;
2760 }
2761}
2762
2763/**
2764 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2765 * @mapping: struct address_space to scan for evictable pages
2766 *
2767 * Scan all pages in mapping. Check unevictable pages for
2768 * evictability and move them to the appropriate zone lru list.
2769 */
2770void scan_mapping_unevictable_pages(struct address_space *mapping)
2771{
2772 pgoff_t next = 0;
2773 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2774 PAGE_CACHE_SHIFT;
2775 struct zone *zone;
2776 struct pagevec pvec;
2777
2778 if (mapping->nrpages == 0)
2779 return;
2780
2781 pagevec_init(&pvec, 0);
2782 while (next < end &&
2783 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2784 int i;
2785 int pg_scanned = 0;
2786
2787 zone = NULL;
2788
2789 for (i = 0; i < pagevec_count(&pvec); i++) {
2790 struct page *page = pvec.pages[i];
2791 pgoff_t page_index = page->index;
2792 struct zone *pagezone = page_zone(page);
2793
2794 pg_scanned++;
2795 if (page_index > next)
2796 next = page_index;
2797 next++;
2798
2799 if (pagezone != zone) {
2800 if (zone)
2801 spin_unlock_irq(&zone->lru_lock);
2802 zone = pagezone;
2803 spin_lock_irq(&zone->lru_lock);
2804 }
2805
2806 if (PageLRU(page) && PageUnevictable(page))
2807 check_move_unevictable_page(page, zone);
2808 }
2809 if (zone)
2810 spin_unlock_irq(&zone->lru_lock);
2811 pagevec_release(&pvec);
2812
2813 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2814 }
2815
2816}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002817
2818/**
2819 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2820 * @zone - zone of which to scan the unevictable list
2821 *
2822 * Scan @zone's unevictable LRU lists to check for pages that have become
2823 * evictable. Move those that have to @zone's inactive list where they
2824 * become candidates for reclaim, unless shrink_inactive_zone() decides
2825 * to reactivate them. Pages that are still unevictable are rotated
2826 * back onto @zone's unevictable list.
2827 */
2828#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08002829static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002830{
2831 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2832 unsigned long scan;
2833 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2834
2835 while (nr_to_scan > 0) {
2836 unsigned long batch_size = min(nr_to_scan,
2837 SCAN_UNEVICTABLE_BATCH_SIZE);
2838
2839 spin_lock_irq(&zone->lru_lock);
2840 for (scan = 0; scan < batch_size; scan++) {
2841 struct page *page = lru_to_page(l_unevictable);
2842
2843 if (!trylock_page(page))
2844 continue;
2845
2846 prefetchw_prev_lru_page(page, l_unevictable, flags);
2847
2848 if (likely(PageLRU(page) && PageUnevictable(page)))
2849 check_move_unevictable_page(page, zone);
2850
2851 unlock_page(page);
2852 }
2853 spin_unlock_irq(&zone->lru_lock);
2854
2855 nr_to_scan -= batch_size;
2856 }
2857}
2858
2859
2860/**
2861 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2862 *
2863 * A really big hammer: scan all zones' unevictable LRU lists to check for
2864 * pages that have become evictable. Move those back to the zones'
2865 * inactive list where they become candidates for reclaim.
2866 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2867 * and we add swap to the system. As such, it runs in the context of a task
2868 * that has possibly/probably made some previously unevictable pages
2869 * evictable.
2870 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08002871static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002872{
2873 struct zone *zone;
2874
2875 for_each_zone(zone) {
2876 scan_zone_unevictable_pages(zone);
2877 }
2878}
2879
2880/*
2881 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2882 * all nodes' unevictable lists for evictable pages
2883 */
2884unsigned long scan_unevictable_pages;
2885
2886int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002887 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002888 size_t *length, loff_t *ppos)
2889{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002890 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002891
2892 if (write && *(unsigned long *)table->data)
2893 scan_all_zones_unevictable_pages();
2894
2895 scan_unevictable_pages = 0;
2896 return 0;
2897}
2898
2899/*
2900 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2901 * a specified node's per zone unevictable lists for evictable pages.
2902 */
2903
2904static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2905 struct sysdev_attribute *attr,
2906 char *buf)
2907{
2908 return sprintf(buf, "0\n"); /* always zero; should fit... */
2909}
2910
2911static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2912 struct sysdev_attribute *attr,
2913 const char *buf, size_t count)
2914{
2915 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2916 struct zone *zone;
2917 unsigned long res;
2918 unsigned long req = strict_strtoul(buf, 10, &res);
2919
2920 if (!req)
2921 return 1; /* zero is no-op */
2922
2923 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2924 if (!populated_zone(zone))
2925 continue;
2926 scan_zone_unevictable_pages(zone);
2927 }
2928 return 1;
2929}
2930
2931
2932static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2933 read_scan_unevictable_node,
2934 write_scan_unevictable_node);
2935
2936int scan_unevictable_register_node(struct node *node)
2937{
2938 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2939}
2940
2941void scan_unevictable_unregister_node(struct node *node)
2942{
2943 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2944}
2945