blob: 9ca587c692748adbc715b440ff5aa8c89357c511 [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
Mel Gorman33906bc2010-08-09 17:19:16 -070051#define CREATE_TRACE_POINTS
52#include <trace/events/vmscan.h>
53
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -070054enum lumpy_mode {
55 LUMPY_MODE_NONE,
56 LUMPY_MODE_ASYNC,
57 LUMPY_MODE_SYNC,
58};
59
Linus Torvalds1da177e2005-04-16 15:20:36 -070060struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070061 /* Incremented by the number of inactive pages that were scanned */
62 unsigned long nr_scanned;
63
Rik van Riela79311c2009-01-06 14:40:01 -080064 /* Number of pages freed so far during a call to shrink_zones() */
65 unsigned long nr_reclaimed;
66
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080067 /* How many pages shrink_list() should reclaim */
68 unsigned long nr_to_reclaim;
69
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080070 unsigned long hibernation_mode;
71
Linus Torvalds1da177e2005-04-16 15:20:36 -070072 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040073 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
75 int may_writepage;
76
Johannes Weinera6dc60f2009-03-31 15:19:30 -070077 /* Can mapped pages be reclaimed? */
78 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080079
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070080 /* Can pages be swapped as part of reclaim? */
81 int may_swap;
82
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070083 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070084
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070085 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080086
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070087 /*
Nikanth Karthikesan415b54e2010-08-17 15:39:09 +053088 * Intend to reclaim enough continuous memory rather than reclaim
89 * enough amount of memory. i.e, mode for high order allocation.
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070090 */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -070091 enum lumpy_mode lumpy_reclaim_mode;
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070092
Balbir Singh66e17072008-02-07 00:13:56 -080093 /* Which cgroup do we reclaim from */
94 struct mem_cgroup *mem_cgroup;
95
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -070096 /*
97 * Nodemask of nodes allowed by the caller. If NULL, all nodes
98 * are scanned.
99 */
100 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700101};
102
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
104
105#ifdef ARCH_HAS_PREFETCH
106#define prefetch_prev_lru_page(_page, _base, _field) \
107 do { \
108 if ((_page)->lru.prev != _base) { \
109 struct page *prev; \
110 \
111 prev = lru_to_page(&(_page->lru)); \
112 prefetch(&prev->_field); \
113 } \
114 } while (0)
115#else
116#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
117#endif
118
119#ifdef ARCH_HAS_PREFETCHW
120#define prefetchw_prev_lru_page(_page, _base, _field) \
121 do { \
122 if ((_page)->lru.prev != _base) { \
123 struct page *prev; \
124 \
125 prev = lru_to_page(&(_page->lru)); \
126 prefetchw(&prev->_field); \
127 } \
128 } while (0)
129#else
130#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
131#endif
132
133/*
134 * From 0 .. 100. Higher means more swappy.
135 */
136int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700137long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138
139static LIST_HEAD(shrinker_list);
140static DECLARE_RWSEM(shrinker_rwsem);
141
Balbir Singh00f0b822008-03-04 14:28:39 -0800142#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800143#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800144#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800145#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800146#endif
147
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800148static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
149 struct scan_control *sc)
150{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800151 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800152 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
153
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800154 return &zone->reclaim_stat;
155}
156
Vincent Li0b217672009-09-21 17:03:09 -0700157static unsigned long zone_nr_lru_pages(struct zone *zone,
158 struct scan_control *sc, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800159{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800160 if (!scanning_global_lru(sc))
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800161 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
162
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800163 return zone_page_state(zone, NR_LRU_BASE + lru);
164}
165
166
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167/*
168 * Add a shrinker callback to be called from the vm
169 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700170void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700172 shrinker->nr = 0;
173 down_write(&shrinker_rwsem);
174 list_add_tail(&shrinker->list, &shrinker_list);
175 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700177EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178
179/*
180 * Remove one
181 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700182void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183{
184 down_write(&shrinker_rwsem);
185 list_del(&shrinker->list);
186 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700188EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700189
190#define SHRINK_BATCH 128
191/*
192 * Call the shrink functions to age shrinkable caches
193 *
194 * Here we assume it costs one seek to replace a lru page and that it also
195 * takes a seek to recreate a cache object. With this in mind we age equal
196 * percentages of the lru and ageable caches. This should balance the seeks
197 * generated by these structures.
198 *
Simon Arlott183ff222007-10-20 01:27:18 +0200199 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 * slab to avoid swapping.
201 *
202 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
203 *
204 * `lru_pages' represents the number of on-LRU pages in all the zones which
205 * are eligible for the caller's allocation attempt. It is used for balancing
206 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700207 *
208 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700209 */
Andrew Morton69e05942006-03-22 00:08:19 -0800210unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
211 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212{
213 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800214 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215
216 if (scanned == 0)
217 scanned = SWAP_CLUSTER_MAX;
218
219 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700220 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221
222 list_for_each_entry(shrinker, &shrinker_list, list) {
223 unsigned long long delta;
224 unsigned long total_scan;
Dave Chinner7f8275d2010-07-19 14:56:17 +1000225 unsigned long max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700226
Dave Chinner7f8275d2010-07-19 14:56:17 +1000227 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700228 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800229 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230 do_div(delta, lru_pages + 1);
231 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800232 if (shrinker->nr < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700233 printk(KERN_ERR "shrink_slab: %pF negative objects to "
234 "delete nr=%ld\n",
235 shrinker->shrink, shrinker->nr);
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800236 shrinker->nr = max_pass;
237 }
238
239 /*
240 * Avoid risking looping forever due to too large nr value:
241 * never try to free more than twice the estimate number of
242 * freeable entries.
243 */
244 if (shrinker->nr > max_pass * 2)
245 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700246
247 total_scan = shrinker->nr;
248 shrinker->nr = 0;
249
250 while (total_scan >= SHRINK_BATCH) {
251 long this_scan = SHRINK_BATCH;
252 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700253 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254
Dave Chinner7f8275d2010-07-19 14:56:17 +1000255 nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
256 shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
257 gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700258 if (shrink_ret == -1)
259 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700260 if (shrink_ret < nr_before)
261 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700262 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263 total_scan -= this_scan;
264
265 cond_resched();
266 }
267
268 shrinker->nr += total_scan;
269 }
270 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700271 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700272}
273
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700274static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc,
275 bool sync)
276{
277 enum lumpy_mode mode = sync ? LUMPY_MODE_SYNC : LUMPY_MODE_ASYNC;
278
279 /*
280 * Some reclaim have alredy been failed. No worth to try synchronous
281 * lumpy reclaim.
282 */
283 if (sync && sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
284 return;
285
286 /*
287 * If we need a large contiguous chunk of memory, or have
288 * trouble getting a small set of contiguous pages, we
289 * will reclaim both active and inactive pages.
290 */
291 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
292 sc->lumpy_reclaim_mode = mode;
293 else if (sc->order && priority < DEF_PRIORITY - 2)
294 sc->lumpy_reclaim_mode = mode;
295 else
296 sc->lumpy_reclaim_mode = LUMPY_MODE_NONE;
297}
298
299static void disable_lumpy_reclaim_mode(struct scan_control *sc)
300{
301 sc->lumpy_reclaim_mode = LUMPY_MODE_NONE;
302}
303
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304static inline int is_page_cache_freeable(struct page *page)
305{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700306 /*
307 * A freeable page cache page is referenced only by the caller
308 * that isolated the page, the page cache radix tree and
309 * optional buffer heads at page->private.
310 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700311 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312}
313
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700314static int may_write_to_queue(struct backing_dev_info *bdi,
315 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700316{
Christoph Lameter930d9152006-01-08 01:00:47 -0800317 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 return 1;
319 if (!bdi_write_congested(bdi))
320 return 1;
321 if (bdi == current->backing_dev_info)
322 return 1;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700323
324 /* lumpy reclaim for hugepage often need a lot of write */
325 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
326 return 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700327 return 0;
328}
329
330/*
331 * We detected a synchronous write error writing a page out. Probably
332 * -ENOSPC. We need to propagate that into the address_space for a subsequent
333 * fsync(), msync() or close().
334 *
335 * The tricky part is that after writepage we cannot touch the mapping: nothing
336 * prevents it from being freed up. But we have a ref on the page and once
337 * that page is locked, the mapping is pinned.
338 *
339 * We're allowed to run sleeping lock_page() here because we know the caller has
340 * __GFP_FS.
341 */
342static void handle_write_error(struct address_space *mapping,
343 struct page *page, int error)
344{
Nick Piggina6aa62a2010-07-20 13:24:25 -0700345 lock_page_nosync(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700346 if (page_mapping(page) == mapping)
347 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700348 unlock_page(page);
349}
350
Christoph Lameter04e62a22006-06-23 02:03:38 -0700351/* possible outcome of pageout() */
352typedef enum {
353 /* failed to write page out, page is locked */
354 PAGE_KEEP,
355 /* move page to the active list, page is locked */
356 PAGE_ACTIVATE,
357 /* page has been sent to the disk successfully, page is unlocked */
358 PAGE_SUCCESS,
359 /* page is clean and locked */
360 PAGE_CLEAN,
361} pageout_t;
362
Linus Torvalds1da177e2005-04-16 15:20:36 -0700363/*
Andrew Morton1742f192006-03-22 00:08:21 -0800364 * pageout is called by shrink_page_list() for each dirty page.
365 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700367static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700368 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369{
370 /*
371 * If the page is dirty, only perform writeback if that write
372 * will be non-blocking. To prevent this allocation from being
373 * stalled by pagecache activity. But note that there may be
374 * stalls if we need to run get_block(). We could test
375 * PagePrivate for that.
376 *
Vincent Li6aceb532009-12-14 17:58:49 -0800377 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700378 * this page's queue, we can perform writeback even if that
379 * will block.
380 *
381 * If the page is swapcache, write it back even if that would
382 * block, for some throttling. This happens by accident, because
383 * swap_backing_dev_info is bust: it doesn't reflect the
384 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 */
386 if (!is_page_cache_freeable(page))
387 return PAGE_KEEP;
388 if (!mapping) {
389 /*
390 * Some data journaling orphaned pages can have
391 * page->mapping == NULL while being dirty with clean buffers.
392 */
David Howells266cf652009-04-03 16:42:36 +0100393 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700394 if (try_to_free_buffers(page)) {
395 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700396 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 return PAGE_CLEAN;
398 }
399 }
400 return PAGE_KEEP;
401 }
402 if (mapping->a_ops->writepage == NULL)
403 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700404 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700405 return PAGE_KEEP;
406
407 if (clear_page_dirty_for_io(page)) {
408 int res;
409 struct writeback_control wbc = {
410 .sync_mode = WB_SYNC_NONE,
411 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700412 .range_start = 0,
413 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 .for_reclaim = 1,
415 };
416
417 SetPageReclaim(page);
418 res = mapping->a_ops->writepage(page, &wbc);
419 if (res < 0)
420 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800421 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700422 ClearPageReclaim(page);
423 return PAGE_ACTIVATE;
424 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700425
426 /*
427 * Wait on writeback if requested to. This happens when
428 * direct reclaiming a large contiguous area and the
429 * first attempt to free a range of pages fails.
430 */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700431 if (PageWriteback(page) &&
432 sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC)
Andy Whitcroftc661b072007-08-22 14:01:26 -0700433 wait_on_page_writeback(page);
434
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435 if (!PageWriteback(page)) {
436 /* synchronous write or broken a_ops? */
437 ClearPageReclaim(page);
438 }
Mel Gorman755f0222010-08-09 17:19:18 -0700439 trace_mm_vmscan_writepage(page,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700440 trace_reclaim_flags(page, sc->lumpy_reclaim_mode));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700441 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700442 return PAGE_SUCCESS;
443 }
444
445 return PAGE_CLEAN;
446}
447
Andrew Mortona649fd92006-10-17 00:09:36 -0700448/*
Nick Piggine2867812008-07-25 19:45:30 -0700449 * Same as remove_mapping, but if the page is removed from the mapping, it
450 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700451 */
Nick Piggine2867812008-07-25 19:45:30 -0700452static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800453{
Nick Piggin28e4d962006-09-25 23:31:23 -0700454 BUG_ON(!PageLocked(page));
455 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800456
Nick Piggin19fd6232008-07-25 19:45:32 -0700457 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800458 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700459 * The non racy check for a busy page.
460 *
461 * Must be careful with the order of the tests. When someone has
462 * a ref to the page, it may be possible that they dirty it then
463 * drop the reference. So if PageDirty is tested before page_count
464 * here, then the following race may occur:
465 *
466 * get_user_pages(&page);
467 * [user mapping goes away]
468 * write_to(page);
469 * !PageDirty(page) [good]
470 * SetPageDirty(page);
471 * put_page(page);
472 * !page_count(page) [good, discard it]
473 *
474 * [oops, our write_to data is lost]
475 *
476 * Reversing the order of the tests ensures such a situation cannot
477 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
478 * load is not satisfied before that of page->_count.
479 *
480 * Note that if SetPageDirty is always performed via set_page_dirty,
481 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800482 */
Nick Piggine2867812008-07-25 19:45:30 -0700483 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800484 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700485 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
486 if (unlikely(PageDirty(page))) {
487 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800488 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700489 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800490
491 if (PageSwapCache(page)) {
492 swp_entry_t swap = { .val = page_private(page) };
493 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700494 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700495 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700496 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500497 void (*freepage)(struct page *);
498
499 freepage = mapping->a_ops->freepage;
500
Nick Piggine2867812008-07-25 19:45:30 -0700501 __remove_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700502 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700503 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500504
505 if (freepage != NULL)
506 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800507 }
508
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800509 return 1;
510
511cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700512 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800513 return 0;
514}
515
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516/*
Nick Piggine2867812008-07-25 19:45:30 -0700517 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
518 * someone else has a ref on the page, abort and return 0. If it was
519 * successfully detached, return 1. Assumes the caller has a single ref on
520 * this page.
521 */
522int remove_mapping(struct address_space *mapping, struct page *page)
523{
524 if (__remove_mapping(mapping, page)) {
525 /*
526 * Unfreezing the refcount with 1 rather than 2 effectively
527 * drops the pagecache ref for us without requiring another
528 * atomic operation.
529 */
530 page_unfreeze_refs(page, 1);
531 return 1;
532 }
533 return 0;
534}
535
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700536/**
537 * putback_lru_page - put previously isolated page onto appropriate LRU list
538 * @page: page to be put back to appropriate lru list
539 *
540 * Add previously isolated @page to appropriate LRU list.
541 * Page may still be unevictable for other reasons.
542 *
543 * lru_lock must not be held, interrupts must be enabled.
544 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700545void putback_lru_page(struct page *page)
546{
547 int lru;
548 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700549 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700550
551 VM_BUG_ON(PageLRU(page));
552
553redo:
554 ClearPageUnevictable(page);
555
556 if (page_evictable(page, NULL)) {
557 /*
558 * For evictable pages, we can use the cache.
559 * In event of a race, worst case is we end up with an
560 * unevictable page on [in]active list.
561 * We know how to handle that.
562 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700563 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700564 lru_cache_add_lru(page, lru);
565 } else {
566 /*
567 * Put unevictable pages directly on zone's unevictable
568 * list.
569 */
570 lru = LRU_UNEVICTABLE;
571 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700572 /*
573 * When racing with an mlock clearing (page is
574 * unlocked), make sure that if the other thread does
575 * not observe our setting of PG_lru and fails
576 * isolation, we see PG_mlocked cleared below and move
577 * the page back to the evictable list.
578 *
579 * The other side is TestClearPageMlocked().
580 */
581 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700582 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700583
584 /*
585 * page's status can change while we move it among lru. If an evictable
586 * page is on unevictable list, it never be freed. To avoid that,
587 * check after we added it to the list, again.
588 */
589 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
590 if (!isolate_lru_page(page)) {
591 put_page(page);
592 goto redo;
593 }
594 /* This means someone else dropped this page from LRU
595 * So, it will be freed or putback to LRU again. There is
596 * nothing to do here.
597 */
598 }
599
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700600 if (was_unevictable && lru != LRU_UNEVICTABLE)
601 count_vm_event(UNEVICTABLE_PGRESCUED);
602 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
603 count_vm_event(UNEVICTABLE_PGCULLED);
604
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700605 put_page(page); /* drop ref from isolate */
606}
607
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800608enum page_references {
609 PAGEREF_RECLAIM,
610 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner64574742010-03-05 13:42:22 -0800611 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800612 PAGEREF_ACTIVATE,
613};
614
615static enum page_references page_check_references(struct page *page,
616 struct scan_control *sc)
617{
Johannes Weiner64574742010-03-05 13:42:22 -0800618 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800619 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800620
Johannes Weiner64574742010-03-05 13:42:22 -0800621 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
622 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800623
624 /* Lumpy reclaim - ignore references */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700625 if (sc->lumpy_reclaim_mode != LUMPY_MODE_NONE)
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800626 return PAGEREF_RECLAIM;
627
628 /*
629 * Mlock lost the isolation race with us. Let try_to_unmap()
630 * move the page to the unevictable list.
631 */
632 if (vm_flags & VM_LOCKED)
633 return PAGEREF_RECLAIM;
634
Johannes Weiner64574742010-03-05 13:42:22 -0800635 if (referenced_ptes) {
636 if (PageAnon(page))
637 return PAGEREF_ACTIVATE;
638 /*
639 * All mapped pages start out with page table
640 * references from the instantiating fault, so we need
641 * to look twice if a mapped file page is used more
642 * than once.
643 *
644 * Mark it and spare it for another trip around the
645 * inactive list. Another page table reference will
646 * lead to its activation.
647 *
648 * Note: the mark is set for activated pages as well
649 * so that recently deactivated but used pages are
650 * quickly recovered.
651 */
652 SetPageReferenced(page);
653
654 if (referenced_page)
655 return PAGEREF_ACTIVATE;
656
657 return PAGEREF_KEEP;
658 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800659
660 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700661 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner64574742010-03-05 13:42:22 -0800662 return PAGEREF_RECLAIM_CLEAN;
663
664 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800665}
666
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700667static noinline_for_stack void free_page_list(struct list_head *free_pages)
668{
669 struct pagevec freed_pvec;
670 struct page *page, *tmp;
671
672 pagevec_init(&freed_pvec, 1);
673
674 list_for_each_entry_safe(page, tmp, free_pages, lru) {
675 list_del(&page->lru);
676 if (!pagevec_add(&freed_pvec, page)) {
677 __pagevec_free(&freed_pvec);
678 pagevec_reinit(&freed_pvec);
679 }
680 }
681
682 pagevec_free(&freed_pvec);
683}
684
Nick Piggine2867812008-07-25 19:45:30 -0700685/*
Andrew Morton1742f192006-03-22 00:08:21 -0800686 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 */
Andrew Morton1742f192006-03-22 00:08:21 -0800688static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman0e093d992010-10-26 14:21:45 -0700689 struct zone *zone,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700690 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691{
692 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700693 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700695 unsigned long nr_dirty = 0;
696 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800697 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700698
699 cond_resched();
700
Linus Torvalds1da177e2005-04-16 15:20:36 -0700701 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800702 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703 struct address_space *mapping;
704 struct page *page;
705 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706
707 cond_resched();
708
709 page = lru_to_page(page_list);
710 list_del(&page->lru);
711
Nick Piggin529ae9a2008-08-02 12:01:03 +0200712 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700713 goto keep;
714
Nick Piggin725d7042006-09-25 23:30:55 -0700715 VM_BUG_ON(PageActive(page));
Mel Gorman0e093d992010-10-26 14:21:45 -0700716 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700717
718 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800719
Nick Pigginb291f002008-10-18 20:26:44 -0700720 if (unlikely(!page_evictable(page, NULL)))
721 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700722
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700723 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800724 goto keep_locked;
725
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 /* Double the slab pressure for mapped and swapcache pages */
727 if (page_mapped(page) || PageSwapCache(page))
728 sc->nr_scanned++;
729
Andy Whitcroftc661b072007-08-22 14:01:26 -0700730 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
731 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
732
733 if (PageWriteback(page)) {
734 /*
735 * Synchronous reclaim is performed in two passes,
736 * first an asynchronous pass over the list to
737 * start parallel writeback, and a second synchronous
738 * pass to wait for the IO to complete. Wait here
739 * for any page for which writeback has already
740 * started.
741 */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700742 if (sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC &&
743 may_enter_fs)
Andy Whitcroftc661b072007-08-22 14:01:26 -0700744 wait_on_page_writeback(page);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700745 else {
746 unlock_page(page);
747 goto keep_lumpy;
748 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700749 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700750
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800751 references = page_check_references(page, sc);
752 switch (references) {
753 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754 goto activate_locked;
Johannes Weiner64574742010-03-05 13:42:22 -0800755 case PAGEREF_KEEP:
756 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800757 case PAGEREF_RECLAIM:
758 case PAGEREF_RECLAIM_CLEAN:
759 ; /* try to reclaim the page below */
760 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700761
Linus Torvalds1da177e2005-04-16 15:20:36 -0700762 /*
763 * Anonymous process memory has backing store?
764 * Try to allocate it some swap space here.
765 */
Nick Pigginb291f002008-10-18 20:26:44 -0700766 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800767 if (!(sc->gfp_mask & __GFP_IO))
768 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800769 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700770 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800771 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700772 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773
774 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775
776 /*
777 * The page is mapped into the page tables of one or more
778 * processes. Try to unmap it here.
779 */
780 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200781 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 case SWAP_FAIL:
783 goto activate_locked;
784 case SWAP_AGAIN:
785 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700786 case SWAP_MLOCK:
787 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788 case SWAP_SUCCESS:
789 ; /* try to free the page below */
790 }
791 }
792
793 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700794 nr_dirty++;
795
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800796 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700797 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700798 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800800 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700801 goto keep_locked;
802
803 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700804 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700805 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700806 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 goto keep_locked;
808 case PAGE_ACTIVATE:
809 goto activate_locked;
810 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700811 if (PageWriteback(page))
812 goto keep_lumpy;
813 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700815
Linus Torvalds1da177e2005-04-16 15:20:36 -0700816 /*
817 * A synchronous write - probably a ramdisk. Go
818 * ahead and try to reclaim the page.
819 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200820 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700821 goto keep;
822 if (PageDirty(page) || PageWriteback(page))
823 goto keep_locked;
824 mapping = page_mapping(page);
825 case PAGE_CLEAN:
826 ; /* try to free the page below */
827 }
828 }
829
830 /*
831 * If the page has buffers, try to free the buffer mappings
832 * associated with this page. If we succeed we try to free
833 * the page as well.
834 *
835 * We do this even if the page is PageDirty().
836 * try_to_release_page() does not perform I/O, but it is
837 * possible for a page to have PageDirty set, but it is actually
838 * clean (all its buffers are clean). This happens if the
839 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700840 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700841 * try_to_release_page() will discover that cleanness and will
842 * drop the buffers and mark the page clean - it can be freed.
843 *
844 * Rarely, pages can have buffers and no ->mapping. These are
845 * the pages which were not successfully invalidated in
846 * truncate_complete_page(). We try to drop those buffers here
847 * and if that worked, and the page is no longer mapped into
848 * process address space (page_count == 1) it can be freed.
849 * Otherwise, leave the page on the LRU so it is swappable.
850 */
David Howells266cf652009-04-03 16:42:36 +0100851 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700852 if (!try_to_release_page(page, sc->gfp_mask))
853 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700854 if (!mapping && page_count(page) == 1) {
855 unlock_page(page);
856 if (put_page_testzero(page))
857 goto free_it;
858 else {
859 /*
860 * rare race with speculative reference.
861 * the speculative reference will free
862 * this page shortly, so we may
863 * increment nr_reclaimed here (and
864 * leave it off the LRU).
865 */
866 nr_reclaimed++;
867 continue;
868 }
869 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700870 }
871
Nick Piggine2867812008-07-25 19:45:30 -0700872 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800873 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700874
Nick Piggina978d6f2008-10-18 20:26:58 -0700875 /*
876 * At this point, we have no other references and there is
877 * no way to pick any more up (removed from LRU, removed
878 * from pagecache). Can use non-atomic bitops now (and
879 * we obviously don't have to worry about waking up a process
880 * waiting on the page lock, because there are no references.
881 */
882 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700883free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800884 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700885
886 /*
887 * Is there need to periodically free_page_list? It would
888 * appear not as the counts should be low
889 */
890 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891 continue;
892
Nick Pigginb291f002008-10-18 20:26:44 -0700893cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800894 if (PageSwapCache(page))
895 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700896 unlock_page(page);
897 putback_lru_page(page);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700898 disable_lumpy_reclaim_mode(sc);
Nick Pigginb291f002008-10-18 20:26:44 -0700899 continue;
900
Linus Torvalds1da177e2005-04-16 15:20:36 -0700901activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700902 /* Not a candidate for swapping, so reclaim swap space. */
903 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800904 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700905 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 SetPageActive(page);
907 pgactivate++;
908keep_locked:
909 unlock_page(page);
910keep:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700911 disable_lumpy_reclaim_mode(sc);
912keep_lumpy:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700913 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700914 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700915 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700916
Mel Gorman0e093d992010-10-26 14:21:45 -0700917 /*
918 * Tag a zone as congested if all the dirty pages encountered were
919 * backed by a congested BDI. In this case, reclaimers should just
920 * back off and wait for congestion to clear because further reclaim
921 * will encounter the same problem
922 */
Shaohua Li1dce0712010-11-11 14:05:17 -0800923 if (nr_dirty == nr_congested && nr_dirty != 0)
Mel Gorman0e093d992010-10-26 14:21:45 -0700924 zone_set_flag(zone, ZONE_CONGESTED);
925
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700926 free_page_list(&free_pages);
927
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700929 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800930 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700931}
932
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700933/*
934 * Attempt to remove the specified page from its LRU. Only take this page
935 * if it is of the appropriate PageActive status. Pages which are being
936 * freed elsewhere are also ignored.
937 *
938 * page: page to consider
939 * mode: one of the LRU isolation modes defined above
940 *
941 * returns 0 on success, -ve errno on failure.
942 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700943int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700944{
945 int ret = -EINVAL;
946
947 /* Only take pages on the LRU. */
948 if (!PageLRU(page))
949 return ret;
950
951 /*
952 * When checking the active state, we need to be sure we are
953 * dealing with comparible boolean values. Take the logical not
954 * of each.
955 */
956 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
957 return ret;
958
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700959 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700960 return ret;
961
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700962 /*
963 * When this function is being called for lumpy reclaim, we
964 * initially look into all LRU pages, active, inactive and
965 * unevictable; only give shrink_page_list evictable pages.
966 */
967 if (PageUnevictable(page))
968 return ret;
969
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700970 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800971
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700972 if (likely(get_page_unless_zero(page))) {
973 /*
974 * Be careful not to clear PageLRU until after we're
975 * sure the page is not being freed elsewhere -- the
976 * page release code relies on it.
977 */
978 ClearPageLRU(page);
979 ret = 0;
980 }
981
982 return ret;
983}
984
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800985/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 * zone->lru_lock is heavily contended. Some of the functions that
987 * shrink the lists perform better by taking out a batch of pages
988 * and working on them outside the LRU lock.
989 *
990 * For pagecache intensive workloads, this function is the hottest
991 * spot in the kernel (apart from copy_*_user functions).
992 *
993 * Appropriate locks must be held before calling this function.
994 *
995 * @nr_to_scan: The number of pages to look through on the list.
996 * @src: The LRU list to pull pages off.
997 * @dst: The temp list to put pages on to.
998 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700999 * @order: The caller's attempted allocation order
1000 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001001 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 *
1003 * returns how many pages were moved onto *@dst.
1004 */
Andrew Morton69e05942006-03-22 00:08:19 -08001005static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
1006 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001007 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001008{
Andrew Morton69e05942006-03-22 00:08:19 -08001009 unsigned long nr_taken = 0;
Mel Gormana8a94d12010-08-09 17:19:17 -07001010 unsigned long nr_lumpy_taken = 0;
1011 unsigned long nr_lumpy_dirty = 0;
1012 unsigned long nr_lumpy_failed = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001013 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001014
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001015 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001016 struct page *page;
1017 unsigned long pfn;
1018 unsigned long end_pfn;
1019 unsigned long page_pfn;
1020 int zone_id;
1021
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 page = lru_to_page(src);
1023 prefetchw_prev_lru_page(page, src, flags);
1024
Nick Piggin725d7042006-09-25 23:30:55 -07001025 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001026
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001027 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001028 case 0:
1029 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001030 mem_cgroup_del_lru(page);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -08001031 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001032 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001033
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001034 case -EBUSY:
1035 /* else it is being freed elsewhere */
1036 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001037 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001038 continue;
1039
1040 default:
1041 BUG();
1042 }
1043
1044 if (!order)
1045 continue;
1046
1047 /*
1048 * Attempt to take all pages in the order aligned region
1049 * surrounding the tag page. Only take those pages of
1050 * the same active state as that tag page. We may safely
1051 * round the target page pfn down to the requested order
1052 * as the mem_map is guarenteed valid out to MAX_ORDER,
1053 * where that page is in a different zone we will detect
1054 * it from its zone id and abort this block scan.
1055 */
1056 zone_id = page_zone_id(page);
1057 page_pfn = page_to_pfn(page);
1058 pfn = page_pfn & ~((1 << order) - 1);
1059 end_pfn = pfn + (1 << order);
1060 for (; pfn < end_pfn; pfn++) {
1061 struct page *cursor_page;
1062
1063 /* The target page is in the block, ignore it. */
1064 if (unlikely(pfn == page_pfn))
1065 continue;
1066
1067 /* Avoid holes within the zone. */
1068 if (unlikely(!pfn_valid_within(pfn)))
1069 break;
1070
1071 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001072
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001073 /* Check that we have not crossed a zone boundary. */
1074 if (unlikely(page_zone_id(cursor_page) != zone_id))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001075 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001076
1077 /*
1078 * If we don't have enough swap space, reclaiming of
1079 * anon page which don't already have a swap slot is
1080 * pointless.
1081 */
1082 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001083 !PageSwapCache(cursor_page))
1084 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001085
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -07001086 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001087 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +09001088 mem_cgroup_del_lru(cursor_page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001089 nr_taken++;
Mel Gormana8a94d12010-08-09 17:19:17 -07001090 nr_lumpy_taken++;
1091 if (PageDirty(cursor_page))
1092 nr_lumpy_dirty++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001093 scan++;
Mel Gormana8a94d12010-08-09 17:19:17 -07001094 } else {
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001095 /* the page is freed already. */
1096 if (!page_count(cursor_page))
1097 continue;
1098 break;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001099 }
1100 }
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001101
1102 /* If we break out of the loop above, lumpy reclaim failed */
1103 if (pfn < end_pfn)
1104 nr_lumpy_failed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001105 }
1106
1107 *scanned = scan;
Mel Gormana8a94d12010-08-09 17:19:17 -07001108
1109 trace_mm_vmscan_lru_isolate(order,
1110 nr_to_scan, scan,
1111 nr_taken,
1112 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1113 mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001114 return nr_taken;
1115}
1116
Balbir Singh66e17072008-02-07 00:13:56 -08001117static unsigned long isolate_pages_global(unsigned long nr,
1118 struct list_head *dst,
1119 unsigned long *scanned, int order,
1120 int mode, struct zone *z,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001121 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -08001122{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001123 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -08001124 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001125 lru += LRU_ACTIVE;
1126 if (file)
1127 lru += LRU_FILE;
1128 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001129 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -08001130}
1131
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001133 * clear_active_flags() is a helper for shrink_active_list(), clearing
1134 * any active bits from the pages in the list.
1135 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001136static unsigned long clear_active_flags(struct list_head *page_list,
1137 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001138{
1139 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001140 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001141 struct page *page;
1142
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001143 list_for_each_entry(page, page_list, lru) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07001144 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001145 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001146 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001147 ClearPageActive(page);
1148 nr_active++;
1149 }
Mel Gorman1489fa12010-08-09 17:19:33 -07001150 if (count)
1151 count[lru]++;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001152 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001153
1154 return nr_active;
1155}
1156
Nick Piggin62695a82008-10-18 20:26:09 -07001157/**
1158 * isolate_lru_page - tries to isolate a page from its LRU list
1159 * @page: page to isolate from its LRU list
1160 *
1161 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1162 * vmstat statistic corresponding to whatever LRU list the page was on.
1163 *
1164 * Returns 0 if the page was removed from an LRU list.
1165 * Returns -EBUSY if the page was not on an LRU list.
1166 *
1167 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001168 * the active list, it will have PageActive set. If it was found on
1169 * the unevictable list, it will have the PageUnevictable bit set. That flag
1170 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001171 *
1172 * The vmstat statistic corresponding to the list on which the page was
1173 * found will be decremented.
1174 *
1175 * Restrictions:
1176 * (1) Must be called with an elevated refcount on the page. This is a
1177 * fundamentnal difference from isolate_lru_pages (which is called
1178 * without a stable reference).
1179 * (2) the lru_lock must not be held.
1180 * (3) interrupts must be enabled.
1181 */
1182int isolate_lru_page(struct page *page)
1183{
1184 int ret = -EBUSY;
1185
1186 if (PageLRU(page)) {
1187 struct zone *zone = page_zone(page);
1188
1189 spin_lock_irq(&zone->lru_lock);
1190 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001191 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001192 ret = 0;
1193 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001194
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001195 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001196 }
1197 spin_unlock_irq(&zone->lru_lock);
1198 }
1199 return ret;
1200}
1201
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001202/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001203 * Are there way too many processes in the direct reclaim path already?
1204 */
1205static int too_many_isolated(struct zone *zone, int file,
1206 struct scan_control *sc)
1207{
1208 unsigned long inactive, isolated;
1209
1210 if (current_is_kswapd())
1211 return 0;
1212
1213 if (!scanning_global_lru(sc))
1214 return 0;
1215
1216 if (file) {
1217 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1218 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1219 } else {
1220 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1221 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1222 }
1223
1224 return isolated > inactive;
1225}
1226
1227/*
Mel Gorman66635622010-08-09 17:19:30 -07001228 * TODO: Try merging with migrations version of putback_lru_pages
1229 */
1230static noinline_for_stack void
Mel Gorman1489fa12010-08-09 17:19:33 -07001231putback_lru_pages(struct zone *zone, struct scan_control *sc,
Mel Gorman66635622010-08-09 17:19:30 -07001232 unsigned long nr_anon, unsigned long nr_file,
1233 struct list_head *page_list)
1234{
1235 struct page *page;
1236 struct pagevec pvec;
Mel Gorman1489fa12010-08-09 17:19:33 -07001237 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Mel Gorman66635622010-08-09 17:19:30 -07001238
1239 pagevec_init(&pvec, 1);
1240
1241 /*
1242 * Put back any unfreeable pages.
1243 */
1244 spin_lock(&zone->lru_lock);
1245 while (!list_empty(page_list)) {
1246 int lru;
1247 page = lru_to_page(page_list);
1248 VM_BUG_ON(PageLRU(page));
1249 list_del(&page->lru);
1250 if (unlikely(!page_evictable(page, NULL))) {
1251 spin_unlock_irq(&zone->lru_lock);
1252 putback_lru_page(page);
1253 spin_lock_irq(&zone->lru_lock);
1254 continue;
1255 }
1256 SetPageLRU(page);
1257 lru = page_lru(page);
1258 add_page_to_lru_list(zone, page, lru);
1259 if (is_active_lru(lru)) {
1260 int file = is_file_lru(lru);
1261 reclaim_stat->recent_rotated[file]++;
1262 }
1263 if (!pagevec_add(&pvec, page)) {
1264 spin_unlock_irq(&zone->lru_lock);
1265 __pagevec_release(&pvec);
1266 spin_lock_irq(&zone->lru_lock);
1267 }
1268 }
1269 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1270 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1271
1272 spin_unlock_irq(&zone->lru_lock);
1273 pagevec_release(&pvec);
1274}
1275
Mel Gorman1489fa12010-08-09 17:19:33 -07001276static noinline_for_stack void update_isolated_counts(struct zone *zone,
1277 struct scan_control *sc,
1278 unsigned long *nr_anon,
1279 unsigned long *nr_file,
1280 struct list_head *isolated_list)
1281{
1282 unsigned long nr_active;
1283 unsigned int count[NR_LRU_LISTS] = { 0, };
1284 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1285
1286 nr_active = clear_active_flags(isolated_list, count);
1287 __count_vm_events(PGDEACTIVATE, nr_active);
1288
1289 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1290 -count[LRU_ACTIVE_FILE]);
1291 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1292 -count[LRU_INACTIVE_FILE]);
1293 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1294 -count[LRU_ACTIVE_ANON]);
1295 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1296 -count[LRU_INACTIVE_ANON]);
1297
1298 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1299 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1300 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1301 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1302
1303 reclaim_stat->recent_scanned[0] += *nr_anon;
1304 reclaim_stat->recent_scanned[1] += *nr_file;
1305}
1306
Mel Gorman66635622010-08-09 17:19:30 -07001307/*
Wu Fengguange31f36982010-08-09 17:20:01 -07001308 * Returns true if the caller should wait to clean dirty/writeback pages.
1309 *
1310 * If we are direct reclaiming for contiguous pages and we do not reclaim
1311 * everything in the list, try again and wait for writeback IO to complete.
1312 * This will stall high-order allocations noticeably. Only do that when really
1313 * need to free the pages under high memory pressure.
1314 */
1315static inline bool should_reclaim_stall(unsigned long nr_taken,
1316 unsigned long nr_freed,
1317 int priority,
1318 struct scan_control *sc)
1319{
1320 int lumpy_stall_priority;
1321
1322 /* kswapd should not stall on sync IO */
1323 if (current_is_kswapd())
1324 return false;
1325
1326 /* Only stall on lumpy reclaim */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001327 if (sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
Wu Fengguange31f36982010-08-09 17:20:01 -07001328 return false;
1329
1330 /* If we have relaimed everything on the isolated list, no stall */
1331 if (nr_freed == nr_taken)
1332 return false;
1333
1334 /*
1335 * For high-order allocations, there are two stall thresholds.
1336 * High-cost allocations stall immediately where as lower
1337 * order allocations such as stacks require the scanning
1338 * priority to be much higher before stalling.
1339 */
1340 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1341 lumpy_stall_priority = DEF_PRIORITY;
1342 else
1343 lumpy_stall_priority = DEF_PRIORITY / 3;
1344
1345 return priority <= lumpy_stall_priority;
1346}
1347
1348/*
Andrew Morton1742f192006-03-22 00:08:21 -08001349 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1350 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001351 */
Mel Gorman66635622010-08-09 17:19:30 -07001352static noinline_for_stack unsigned long
1353shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1354 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001355{
1356 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001357 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001358 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001359 unsigned long nr_taken;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001360 unsigned long nr_anon;
1361 unsigned long nr_file;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001362
Rik van Riel35cd7812009-09-21 17:01:38 -07001363 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001364 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001365
1366 /* We are about to die and free our memory. Return now. */
1367 if (fatal_signal_pending(current))
1368 return SWAP_CLUSTER_MAX;
1369 }
1370
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001371 set_lumpy_reclaim_mode(priority, sc, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001372 lru_add_drain();
1373 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001375 if (scanning_global_lru(sc)) {
1376 nr_taken = isolate_pages_global(nr_to_scan,
1377 &page_list, &nr_scanned, sc->order,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001378 sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1379 ISOLATE_INACTIVE : ISOLATE_BOTH,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001380 zone, 0, file);
1381 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001382 if (current_is_kswapd())
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001383 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1384 nr_scanned);
1385 else
1386 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1387 nr_scanned);
1388 } else {
1389 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1390 &page_list, &nr_scanned, sc->order,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001391 sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1392 ISOLATE_INACTIVE : ISOLATE_BOTH,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001393 zone, sc->mem_cgroup,
1394 0, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001395 /*
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001396 * mem_cgroup_isolate_pages() keeps track of
1397 * scanned pages on its own.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001398 */
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001399 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001400
Mel Gorman66635622010-08-09 17:19:30 -07001401 if (nr_taken == 0) {
1402 spin_unlock_irq(&zone->lru_lock);
1403 return 0;
1404 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001405
Mel Gorman1489fa12010-08-09 17:19:33 -07001406 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001407
1408 spin_unlock_irq(&zone->lru_lock);
1409
Mel Gorman0e093d992010-10-26 14:21:45 -07001410 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001411
Wu Fengguange31f36982010-08-09 17:20:01 -07001412 /* Check if we should syncronously wait for writeback */
1413 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001414 set_lumpy_reclaim_mode(priority, sc, true);
Mel Gorman0e093d992010-10-26 14:21:45 -07001415 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001416 }
1417
1418 local_irq_disable();
1419 if (current_is_kswapd())
1420 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1421 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1422
Mel Gorman1489fa12010-08-09 17:19:33 -07001423 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07001424
1425 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1426 zone_idx(zone),
1427 nr_scanned, nr_reclaimed,
1428 priority,
1429 trace_shrink_flags(file, sc->lumpy_reclaim_mode));
Andrew Morton05ff5132006-03-22 00:08:20 -08001430 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001431}
1432
Martin Bligh3bb1a852006-10-28 10:38:24 -07001433/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434 * This moves pages from the active list to the inactive list.
1435 *
1436 * We move them the other way if the page is referenced by one or more
1437 * processes, from rmap.
1438 *
1439 * If the pages are mostly unmapped, the processing is fast and it is
1440 * appropriate to hold zone->lru_lock across the whole operation. But if
1441 * the pages are mapped, the processing is slow (page_referenced()) so we
1442 * should drop zone->lru_lock around each page. It's impossible to balance
1443 * this, so instead we remove the pages from the LRU while processing them.
1444 * It is safe to rely on PG_active against the non-LRU pages in here because
1445 * nobody will play with that bit on a non-LRU page.
1446 *
1447 * The downside is that we have to touch page->_count against each page.
1448 * But we had to alter page->flags anyway.
1449 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001450
Wu Fengguang3eb41402009-06-16 15:33:13 -07001451static void move_active_pages_to_lru(struct zone *zone,
1452 struct list_head *list,
1453 enum lru_list lru)
1454{
1455 unsigned long pgmoved = 0;
1456 struct pagevec pvec;
1457 struct page *page;
1458
1459 pagevec_init(&pvec, 1);
1460
1461 while (!list_empty(list)) {
1462 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001463
1464 VM_BUG_ON(PageLRU(page));
1465 SetPageLRU(page);
1466
Wu Fengguang3eb41402009-06-16 15:33:13 -07001467 list_move(&page->lru, &zone->lru[lru].list);
1468 mem_cgroup_add_lru_list(page, lru);
1469 pgmoved++;
1470
1471 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1472 spin_unlock_irq(&zone->lru_lock);
1473 if (buffer_heads_over_limit)
1474 pagevec_strip(&pvec);
1475 __pagevec_release(&pvec);
1476 spin_lock_irq(&zone->lru_lock);
1477 }
1478 }
1479 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1480 if (!is_active_lru(lru))
1481 __count_vm_events(PGDEACTIVATE, pgmoved);
1482}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001483
Andrew Morton1742f192006-03-22 00:08:21 -08001484static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001485 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001487 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001488 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001489 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001491 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001492 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001494 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001495 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001496
1497 lru_add_drain();
1498 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001499 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001500 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1501 &pgscanned, sc->order,
1502 ISOLATE_ACTIVE, zone,
1503 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001504 zone->pages_scanned += pgscanned;
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001505 } else {
1506 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1507 &pgscanned, sc->order,
1508 ISOLATE_ACTIVE, zone,
1509 sc->mem_cgroup, 1, file);
1510 /*
1511 * mem_cgroup_isolate_pages() keeps track of
1512 * scanned pages on its own.
1513 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001514 }
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001515
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001516 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001517
Wu Fengguang3eb41402009-06-16 15:33:13 -07001518 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001519 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001520 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001521 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001522 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001523 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001524 spin_unlock_irq(&zone->lru_lock);
1525
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526 while (!list_empty(&l_hold)) {
1527 cond_resched();
1528 page = lru_to_page(&l_hold);
1529 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001530
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001531 if (unlikely(!page_evictable(page, NULL))) {
1532 putback_lru_page(page);
1533 continue;
1534 }
1535
Johannes Weiner64574742010-03-05 13:42:22 -08001536 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001537 nr_rotated++;
Wu Fengguang8cab4752009-06-16 15:33:12 -07001538 /*
1539 * Identify referenced, file-backed active pages and
1540 * give them one more trip around the active list. So
1541 * that executable code get better chances to stay in
1542 * memory under moderate memory pressure. Anon pages
1543 * are not likely to be evicted by use-once streaming
1544 * IO, plus JVM can create lots of anon VM_EXEC pages,
1545 * so we ignore them here.
1546 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001547 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001548 list_add(&page->lru, &l_active);
1549 continue;
1550 }
1551 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001552
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001553 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001554 list_add(&page->lru, &l_inactive);
1555 }
1556
Andrew Mortonb5557492009-01-06 14:40:13 -08001557 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001558 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001559 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001560 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001561 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001562 * Count referenced pages from currently used mappings as rotated,
1563 * even though only some of them are actually re-activated. This
1564 * helps balance scan pressure between file and anonymous pages in
1565 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001566 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001567 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001568
Wu Fengguang3eb41402009-06-16 15:33:13 -07001569 move_active_pages_to_lru(zone, &l_active,
1570 LRU_ACTIVE + file * LRU_FILE);
1571 move_active_pages_to_lru(zone, &l_inactive,
1572 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001573 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001574 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575}
1576
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001577#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001578static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001579{
1580 unsigned long active, inactive;
1581
1582 active = zone_page_state(zone, NR_ACTIVE_ANON);
1583 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1584
1585 if (inactive * zone->inactive_ratio < active)
1586 return 1;
1587
1588 return 0;
1589}
1590
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001591/**
1592 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1593 * @zone: zone to check
1594 * @sc: scan control of this context
1595 *
1596 * Returns true if the zone does not have enough inactive anon pages,
1597 * meaning some active anon pages need to be deactivated.
1598 */
1599static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1600{
1601 int low;
1602
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001603 /*
1604 * If we don't have swap space, anonymous page deactivation
1605 * is pointless.
1606 */
1607 if (!total_swap_pages)
1608 return 0;
1609
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001610 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001611 low = inactive_anon_is_low_global(zone);
1612 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001613 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001614 return low;
1615}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001616#else
1617static inline int inactive_anon_is_low(struct zone *zone,
1618 struct scan_control *sc)
1619{
1620 return 0;
1621}
1622#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001623
Rik van Riel56e49d22009-06-16 15:32:28 -07001624static int inactive_file_is_low_global(struct zone *zone)
1625{
1626 unsigned long active, inactive;
1627
1628 active = zone_page_state(zone, NR_ACTIVE_FILE);
1629 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1630
1631 return (active > inactive);
1632}
1633
1634/**
1635 * inactive_file_is_low - check if file pages need to be deactivated
1636 * @zone: zone to check
1637 * @sc: scan control of this context
1638 *
1639 * When the system is doing streaming IO, memory pressure here
1640 * ensures that active file pages get deactivated, until more
1641 * than half of the file pages are on the inactive list.
1642 *
1643 * Once we get to that situation, protect the system's working
1644 * set from being evicted by disabling active file page aging.
1645 *
1646 * This uses a different ratio than the anonymous pages, because
1647 * the page cache uses a use-once replacement algorithm.
1648 */
1649static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1650{
1651 int low;
1652
1653 if (scanning_global_lru(sc))
1654 low = inactive_file_is_low_global(zone);
1655 else
1656 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1657 return low;
1658}
1659
Rik van Rielb39415b2009-12-14 17:59:48 -08001660static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1661 int file)
1662{
1663 if (file)
1664 return inactive_file_is_low(zone, sc);
1665 else
1666 return inactive_anon_is_low(zone, sc);
1667}
1668
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001669static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001670 struct zone *zone, struct scan_control *sc, int priority)
1671{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001672 int file = is_file_lru(lru);
1673
Rik van Rielb39415b2009-12-14 17:59:48 -08001674 if (is_active_lru(lru)) {
1675 if (inactive_list_is_low(zone, sc, file))
1676 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001677 return 0;
1678 }
1679
Rik van Riel33c120e2008-10-18 20:26:36 -07001680 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001681}
1682
Linus Torvalds1da177e2005-04-16 15:20:36 -07001683/*
Shaohua Li76a33fc2010-05-24 14:32:36 -07001684 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1685 * until we collected @swap_cluster_max pages to scan.
1686 */
1687static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1688 unsigned long *nr_saved_scan)
1689{
1690 unsigned long nr;
1691
1692 *nr_saved_scan += nr_to_scan;
1693 nr = *nr_saved_scan;
1694
1695 if (nr >= SWAP_CLUSTER_MAX)
1696 *nr_saved_scan = 0;
1697 else
1698 nr = 0;
1699
1700 return nr;
1701}
1702
1703/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001704 * Determine how aggressively the anon and file LRU lists should be
1705 * scanned. The relative value of each set of LRU lists is determined
1706 * by looking at the fraction of the pages scanned we did rotate back
1707 * onto the active list instead of evict.
1708 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001709 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001710 */
Shaohua Li76a33fc2010-05-24 14:32:36 -07001711static void get_scan_count(struct zone *zone, struct scan_control *sc,
1712 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001713{
1714 unsigned long anon, file, free;
1715 unsigned long anon_prio, file_prio;
1716 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001717 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001718 u64 fraction[2], denominator;
1719 enum lru_list l;
1720 int noswap = 0;
1721
1722 /* If we have no swap space, do not bother scanning anon pages. */
1723 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1724 noswap = 1;
1725 fraction[0] = 0;
1726 fraction[1] = 1;
1727 denominator = 1;
1728 goto out;
1729 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001730
Vincent Li0b217672009-09-21 17:03:09 -07001731 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1732 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1733 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1734 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001735
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001736 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001737 free = zone_page_state(zone, NR_FREE_PAGES);
1738 /* If we have very few page cache pages,
1739 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001740 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001741 fraction[0] = 1;
1742 fraction[1] = 0;
1743 denominator = 1;
1744 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001745 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001746 }
1747
1748 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001749 * With swappiness at 100, anonymous and file have the same priority.
1750 * This scanning priority is essentially the inverse of IO cost.
1751 */
1752 anon_prio = sc->swappiness;
1753 file_prio = 200 - sc->swappiness;
1754
1755 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001756 * OK, so we have swap space and a fair amount of page cache
1757 * pages. We use the recently rotated / recently scanned
1758 * ratios to determine how valuable each cache is.
1759 *
1760 * Because workloads change over time (and to avoid overflow)
1761 * we keep these statistics as a floating average, which ends
1762 * up weighing recent references more than old ones.
1763 *
1764 * anon in [0], file in [1]
1765 */
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001766 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001767 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001768 reclaim_stat->recent_scanned[0] /= 2;
1769 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001770 }
1771
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001772 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001773 reclaim_stat->recent_scanned[1] /= 2;
1774 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001775 }
1776
1777 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001778 * The amount of pressure on anon vs file pages is inversely
1779 * proportional to the fraction of recently scanned pages on
1780 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001781 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001782 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1783 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001784
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001785 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1786 fp /= reclaim_stat->recent_rotated[1] + 1;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001787 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001788
Shaohua Li76a33fc2010-05-24 14:32:36 -07001789 fraction[0] = ap;
1790 fraction[1] = fp;
1791 denominator = ap + fp + 1;
1792out:
1793 for_each_evictable_lru(l) {
1794 int file = is_file_lru(l);
1795 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001796
Shaohua Li76a33fc2010-05-24 14:32:36 -07001797 scan = zone_nr_lru_pages(zone, sc, l);
1798 if (priority || noswap) {
1799 scan >>= priority;
1800 scan = div64_u64(scan * fraction[file], denominator);
1801 }
1802 nr[l] = nr_scan_try_batch(scan,
1803 &reclaim_stat->nr_saved_scan[l]);
1804 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001805}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001806
1807/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001808 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1809 */
Rik van Riela79311c2009-01-06 14:40:01 -08001810static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001811 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001813 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001814 unsigned long nr_to_scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001815 enum lru_list l;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001816 unsigned long nr_reclaimed = sc->nr_reclaimed;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001817 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818
Shaohua Li76a33fc2010-05-24 14:32:36 -07001819 get_scan_count(zone, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001820
Rik van Riel556adec2008-10-18 20:26:34 -07001821 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1822 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001823 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001824 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001825 nr_to_scan = min_t(unsigned long,
1826 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001827 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001829 nr_reclaimed += shrink_list(l, nr_to_scan,
1830 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001831 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 }
Rik van Riela79311c2009-01-06 14:40:01 -08001833 /*
1834 * On large memory systems, scan >> priority can become
1835 * really large. This is fine for the starting priority;
1836 * we want to put equal scanning pressure on each zone.
1837 * However, if the VM has a harder time of freeing pages,
1838 * with multiple processes reclaiming pages, the total
1839 * freeing target can get unreasonably large.
1840 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08001841 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001842 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001843 }
1844
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001845 sc->nr_reclaimed = nr_reclaimed;
1846
Rik van Riel556adec2008-10-18 20:26:34 -07001847 /*
1848 * Even if we did not try to evict anon pages at all, we want to
1849 * rebalance the anon lru active/inactive ratio.
1850 */
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001851 if (inactive_anon_is_low(zone, sc))
Rik van Riel556adec2008-10-18 20:26:34 -07001852 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1853
Andrew Morton232ea4d2007-02-28 20:13:21 -08001854 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855}
1856
1857/*
1858 * This is the direct reclaim path, for page-allocating processes. We only
1859 * try to reclaim pages from zones which will satisfy the caller's allocation
1860 * request.
1861 *
Mel Gorman41858962009-06-16 15:32:12 -07001862 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1863 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1865 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001866 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1867 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1868 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870 * If a zone is deemed to be full of pinned pages then just give it a light
1871 * scan then give up on it.
1872 */
Minchan Kimd1908362010-09-22 13:05:01 -07001873static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001874 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875{
Mel Gormandd1a2392008-04-28 02:12:17 -07001876 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001877 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001878
Mel Gormand4debc62010-08-09 17:19:29 -07001879 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1880 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001881 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001883 /*
1884 * Take care memory controller reclaiming has small influence
1885 * to global LRU.
1886 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001887 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001888 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1889 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08001890 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001891 continue; /* Let kswapd poll it */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001892 }
Nick Piggin408d8542006-09-25 23:31:27 -07001893
Rik van Riela79311c2009-01-06 14:40:01 -08001894 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001895 }
Minchan Kimd1908362010-09-22 13:05:01 -07001896}
1897
1898static bool zone_reclaimable(struct zone *zone)
1899{
1900 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1901}
1902
1903/*
1904 * As hibernation is going on, kswapd is freezed so that it can't mark
1905 * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1906 * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1907 */
1908static bool all_unreclaimable(struct zonelist *zonelist,
1909 struct scan_control *sc)
1910{
1911 struct zoneref *z;
1912 struct zone *zone;
1913 bool all_unreclaimable = true;
1914
1915 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1916 gfp_zone(sc->gfp_mask), sc->nodemask) {
1917 if (!populated_zone(zone))
1918 continue;
1919 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1920 continue;
1921 if (zone_reclaimable(zone)) {
1922 all_unreclaimable = false;
1923 break;
1924 }
1925 }
1926
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001927 return all_unreclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001929
Linus Torvalds1da177e2005-04-16 15:20:36 -07001930/*
1931 * This is the main entry point to direct page reclaim.
1932 *
1933 * If a full scan of the inactive list fails to free enough memory then we
1934 * are "out of memory" and something needs to be killed.
1935 *
1936 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1937 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02001938 * caller can't do much about. We kick the writeback threads and take explicit
1939 * naps in the hope that some of these pages can be written. But if the
1940 * allocating task holds filesystem locks which prevent writeout this might not
1941 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001942 *
1943 * returns: 0, if no pages reclaimed
1944 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001945 */
Mel Gormandac1d272008-04-28 02:12:12 -07001946static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001947 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948{
1949 int priority;
Andrew Morton69e05942006-03-22 00:08:19 -08001950 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07001952 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001953 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001954 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001955
Miao Xiec0ff7452010-05-24 14:32:08 -07001956 get_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07001957 delayacct_freepages_start();
1958
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001959 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001960 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961
1962 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001963 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001964 if (!priority)
1965 disable_swap_token();
Minchan Kimd1908362010-09-22 13:05:01 -07001966 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001967 /*
1968 * Don't shrink slabs when reclaiming memory from
1969 * over limit cgroups
1970 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001971 if (scanning_global_lru(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07001972 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07001973 for_each_zone_zonelist(zone, z, zonelist,
1974 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07001975 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1976 continue;
1977
1978 lru_pages += zone_reclaimable_pages(zone);
1979 }
1980
Mel Gormandd1a2392008-04-28 02:12:17 -07001981 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001982 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08001983 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001984 reclaim_state->reclaimed_slab = 0;
1985 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001986 }
Balbir Singh66e17072008-02-07 00:13:56 -08001987 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001988 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990
1991 /*
1992 * Try to write back as many pages as we just scanned. This
1993 * tends to cause slow streaming writers to write data to the
1994 * disk smoothly, at the dirtying rate, which is nice. But
1995 * that's undesirable in laptop mode, where we *want* lumpy
1996 * writeout. So in laptop mode, write out the whole world.
1997 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001998 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1999 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02002000 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08002001 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 }
2003
2004 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002005 if (!sc->hibernation_mode && sc->nr_scanned &&
Mel Gorman0e093d992010-10-26 14:21:45 -07002006 priority < DEF_PRIORITY - 2) {
2007 struct zone *preferred_zone;
2008
2009 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2010 NULL, &preferred_zone);
2011 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2012 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002014
Linus Torvalds1da177e2005-04-16 15:20:36 -07002015out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002016 delayacct_freepages_end();
Miao Xiec0ff7452010-05-24 14:32:08 -07002017 put_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002018
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002019 if (sc->nr_reclaimed)
2020 return sc->nr_reclaimed;
2021
2022 /* top priority shrink_zones still had more to do? don't OOM, then */
Minchan Kimd1908362010-09-22 13:05:01 -07002023 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002024 return 1;
2025
2026 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002027}
2028
Mel Gormandac1d272008-04-28 02:12:12 -07002029unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002030 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002031{
Mel Gorman33906bc2010-08-09 17:19:16 -07002032 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002033 struct scan_control sc = {
2034 .gfp_mask = gfp_mask,
2035 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002036 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002037 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002038 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002039 .swappiness = vm_swappiness,
2040 .order = order,
2041 .mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002042 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002043 };
2044
Mel Gorman33906bc2010-08-09 17:19:16 -07002045 trace_mm_vmscan_direct_reclaim_begin(order,
2046 sc.may_writepage,
2047 gfp_mask);
2048
2049 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2050
2051 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2052
2053 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002054}
2055
Balbir Singh00f0b822008-03-04 14:28:39 -08002056#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08002057
Balbir Singh4e416952009-09-23 15:56:39 -07002058unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2059 gfp_t gfp_mask, bool noswap,
2060 unsigned int swappiness,
KOSAKI Motohiro14fec792010-08-10 18:03:05 -07002061 struct zone *zone)
Balbir Singh4e416952009-09-23 15:56:39 -07002062{
2063 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002064 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002065 .may_writepage = !laptop_mode,
2066 .may_unmap = 1,
2067 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002068 .swappiness = swappiness,
2069 .order = 0,
2070 .mem_cgroup = mem,
Balbir Singh4e416952009-09-23 15:56:39 -07002071 };
Balbir Singh4e416952009-09-23 15:56:39 -07002072 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2073 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002074
2075 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2076 sc.may_writepage,
2077 sc.gfp_mask);
2078
Balbir Singh4e416952009-09-23 15:56:39 -07002079 /*
2080 * NOTE: Although we can get the priority field, using it
2081 * here is not a good idea, since it limits the pages we can scan.
2082 * if we don't reclaim here, the shrink_zone from balance_pgdat
2083 * will pick up pages from other mem cgroup's as well. We hack
2084 * the priority and make it zero.
2085 */
2086 shrink_zone(0, zone, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002087
2088 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2089
Balbir Singh4e416952009-09-23 15:56:39 -07002090 return sc.nr_reclaimed;
2091}
2092
Balbir Singhe1a1cd52008-02-07 00:14:02 -08002093unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002094 gfp_t gfp_mask,
2095 bool noswap,
2096 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08002097{
Balbir Singh4e416952009-09-23 15:56:39 -07002098 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002099 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002100 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002101 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002102 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002103 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002104 .nr_to_reclaim = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002105 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08002106 .order = 0,
2107 .mem_cgroup = mem_cont,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002108 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08002109 };
Balbir Singh66e17072008-02-07 00:13:56 -08002110
Mel Gormandd1a2392008-04-28 02:12:17 -07002111 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2112 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2113 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002114
2115 trace_mm_vmscan_memcg_reclaim_begin(0,
2116 sc.may_writepage,
2117 sc.gfp_mask);
2118
2119 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2120
2121 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2122
2123 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002124}
2125#endif
2126
Mel Gormanf50de2d2009-12-14 17:58:53 -08002127/* is kswapd sleeping prematurely? */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002128static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002129{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002130 int i;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002131
2132 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2133 if (remaining)
2134 return 1;
2135
2136 /* If after HZ/10, a zone is below the high mark, it's premature */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002137 for (i = 0; i < pgdat->nr_zones; i++) {
2138 struct zone *zone = pgdat->node_zones + i;
2139
2140 if (!populated_zone(zone))
2141 continue;
2142
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002143 if (zone->all_unreclaimable)
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002144 continue;
2145
Mel Gormanf50de2d2009-12-14 17:58:53 -08002146 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
2147 0, 0))
2148 return 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002149 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002150
2151 return 0;
2152}
2153
Linus Torvalds1da177e2005-04-16 15:20:36 -07002154/*
2155 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002156 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002157 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002158 * Returns the number of pages which were actually freed.
2159 *
2160 * There is special handling here for zones which are full of pinned pages.
2161 * This can happen if the pages are all mlocked, or if they are all used by
2162 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2163 * What we do is to detect the case where all pages in the zone have been
2164 * scanned twice and there has been zero successful reclaim. Mark the zone as
2165 * dead and from now on, only perform a short scan. Basically we're polling
2166 * the zone for when the problem goes away.
2167 *
2168 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002169 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2170 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2171 * lower zones regardless of the number of free pages in the lower zones. This
2172 * interoperates with the page allocator fallback scheme to ensure that aging
2173 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002175static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002176{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002177 int all_zones_ok;
2178 int priority;
2179 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08002180 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002181 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08002182 struct scan_control sc = {
2183 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002184 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002185 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002186 /*
2187 * kswapd doesn't want to be bailed out while reclaim. because
2188 * we want to put equal scanning pressure on each zone.
2189 */
2190 .nr_to_reclaim = ULONG_MAX,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002191 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002192 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002193 .mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002194 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195loop_again:
2196 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002197 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002198 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002199 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002200
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2202 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2203 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002204 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002205
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002206 /* The swap token gets in the way of swapout... */
2207 if (!priority)
2208 disable_swap_token();
2209
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 all_zones_ok = 1;
2211
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002212 /*
2213 * Scan in the highmem->dma direction for the highest
2214 * zone which needs scanning
2215 */
2216 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2217 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002218
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002219 if (!populated_zone(zone))
2220 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002222 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002223 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002224
Rik van Riel556adec2008-10-18 20:26:34 -07002225 /*
2226 * Do some background aging of the anon list, to give
2227 * pages a chance to be referenced before reclaiming.
2228 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002229 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002230 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2231 &sc, priority, 0);
2232
Mel Gorman41858962009-06-16 15:32:12 -07002233 if (!zone_watermark_ok(zone, order,
2234 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002235 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002236 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002237 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002238 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002239 if (i < 0)
2240 goto out;
2241
Linus Torvalds1da177e2005-04-16 15:20:36 -07002242 for (i = 0; i <= end_zone; i++) {
2243 struct zone *zone = pgdat->node_zones + i;
2244
Wu Fengguangadea02a2009-09-21 17:01:42 -07002245 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 }
2247
2248 /*
2249 * Now scan the zone in the dma->highmem direction, stopping
2250 * at the last zone which needs scanning.
2251 *
2252 * We do this because the page allocator works in the opposite
2253 * direction. This prevents the page allocator from allocating
2254 * pages behind kswapd's direction of progress, which would
2255 * cause too much scanning of the lower zones.
2256 */
2257 for (i = 0; i <= end_zone; i++) {
2258 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002259 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260
Con Kolivasf3fe6512006-01-06 00:11:15 -08002261 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 continue;
2263
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002264 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002265 continue;
2266
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002268
Balbir Singh4e416952009-09-23 15:56:39 -07002269 /*
2270 * Call soft limit reclaim before calling shrink_zone.
2271 * For now we ignore the return value
2272 */
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002273 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2274
Rik van Riel32a43302007-10-16 01:24:50 -07002275 /*
2276 * We put equal pressure on every zone, unless one
2277 * zone has way too many pages free already.
2278 */
Mel Gorman41858962009-06-16 15:32:12 -07002279 if (!zone_watermark_ok(zone, order,
2280 8*high_wmark_pages(zone), end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002281 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002282 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002283 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2284 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002285 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 total_scanned += sc.nr_scanned;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002287 if (zone->all_unreclaimable)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002288 continue;
Minchan Kimd1908362010-09-22 13:05:01 -07002289 if (nr_slab == 0 && !zone_reclaimable(zone))
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002290 zone->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291 /*
2292 * If we've done a decent amount of scanning and
2293 * the reclaim ratio is low, start doing writepage
2294 * even in laptop mode
2295 */
2296 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002297 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002298 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002299
Minchan Kim45973d72010-03-05 13:41:45 -08002300 if (!zone_watermark_ok(zone, order,
2301 high_wmark_pages(zone), end_zone, 0)) {
2302 all_zones_ok = 0;
2303 /*
2304 * We are still under min water mark. This
2305 * means that we have a GFP_ATOMIC allocation
2306 * failure risk. Hurry up!
2307 */
2308 if (!zone_watermark_ok(zone, order,
2309 min_wmark_pages(zone), end_zone, 0))
2310 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002311 } else {
2312 /*
2313 * If a zone reaches its high watermark,
2314 * consider it to be no longer congested. It's
2315 * possible there are dirty pages backed by
2316 * congested BDIs but as pressure is relieved,
2317 * spectulatively avoid congestion waits
2318 */
2319 zone_clear_flag(zone, ZONE_CONGESTED);
Minchan Kim45973d72010-03-05 13:41:45 -08002320 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002321
Linus Torvalds1da177e2005-04-16 15:20:36 -07002322 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323 if (all_zones_ok)
2324 break; /* kswapd: all done */
2325 /*
2326 * OK, kswapd is getting into trouble. Take a nap, then take
2327 * another pass across the zones.
2328 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002329 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2330 if (has_under_min_watermark_zone)
2331 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2332 else
2333 congestion_wait(BLK_RW_ASYNC, HZ/10);
2334 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002335
2336 /*
2337 * We do this so kswapd doesn't build up large priorities for
2338 * example when it is freeing in parallel with allocators. It
2339 * matches the direct reclaim path behaviour in terms of impact
2340 * on zone->*_priority.
2341 */
Rik van Riela79311c2009-01-06 14:40:01 -08002342 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002343 break;
2344 }
2345out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 if (!all_zones_ok) {
2347 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002348
2349 try_to_freeze();
2350
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002351 /*
2352 * Fragmentation may mean that the system cannot be
2353 * rebalanced for high-order allocations in all zones.
2354 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2355 * it means the zones have been fully scanned and are still
2356 * not balanced. For high-order allocations, there is
2357 * little point trying all over again as kswapd may
2358 * infinite loop.
2359 *
2360 * Instead, recheck all watermarks at order-0 as they
2361 * are the most important. If watermarks are ok, kswapd will go
2362 * back to sleep. High-order users can still perform direct
2363 * reclaim if they wish.
2364 */
2365 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2366 order = sc.order = 0;
2367
Linus Torvalds1da177e2005-04-16 15:20:36 -07002368 goto loop_again;
2369 }
2370
Rik van Riela79311c2009-01-06 14:40:01 -08002371 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002372}
2373
2374/*
2375 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002376 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377 *
2378 * This basically trickles out pages so that we have _some_
2379 * free memory available even if there is no other activity
2380 * that frees anything up. This is needed for things like routing
2381 * etc, where we otherwise might have all activity going on in
2382 * asynchronous contexts that cannot page things out.
2383 *
2384 * If there are applications that are active memory-allocators
2385 * (most normal use), this basically shouldn't matter.
2386 */
2387static int kswapd(void *p)
2388{
2389 unsigned long order;
2390 pg_data_t *pgdat = (pg_data_t*)p;
2391 struct task_struct *tsk = current;
2392 DEFINE_WAIT(wait);
2393 struct reclaim_state reclaim_state = {
2394 .reclaimed_slab = 0,
2395 };
Rusty Russella70f7302009-03-13 14:49:46 +10302396 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397
Nick Piggincf40bd12009-01-21 08:12:39 +01002398 lockdep_set_current_reclaim_state(GFP_KERNEL);
2399
Rusty Russell174596a2009-01-01 10:12:29 +10302400 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002401 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002402 current->reclaim_state = &reclaim_state;
2403
2404 /*
2405 * Tell the memory management that we're a "memory allocator",
2406 * and that if we need more memory we should get access to it
2407 * regardless (see "__alloc_pages()"). "kswapd" should
2408 * never get caught in the normal page freeing logic.
2409 *
2410 * (Kswapd normally doesn't need memory anyway, but sometimes
2411 * you need a small amount of memory in order to be able to
2412 * page out something else, and this flag essentially protects
2413 * us from recursively trying to free more memory as we're
2414 * trying to free the first piece of memory in the first place).
2415 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002416 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002417 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002418
2419 order = 0;
2420 for ( ; ; ) {
2421 unsigned long new_order;
David Rientjes8fe23e02009-12-14 17:58:33 -08002422 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002423
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2425 new_order = pgdat->kswapd_max_order;
2426 pgdat->kswapd_max_order = 0;
2427 if (order < new_order) {
2428 /*
2429 * Don't sleep if someone wants a larger 'order'
2430 * allocation
2431 */
2432 order = new_order;
2433 } else {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002434 if (!freezing(current) && !kthread_should_stop()) {
2435 long remaining = 0;
2436
2437 /* Try to sleep for a short interval */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002438 if (!sleeping_prematurely(pgdat, order, remaining)) {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002439 remaining = schedule_timeout(HZ/10);
2440 finish_wait(&pgdat->kswapd_wait, &wait);
2441 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2442 }
2443
2444 /*
2445 * After a short sleep, check if it was a
2446 * premature sleep. If not, then go fully
2447 * to sleep until explicitly woken up
2448 */
Mel Gorman33906bc2010-08-09 17:19:16 -07002449 if (!sleeping_prematurely(pgdat, order, remaining)) {
2450 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002451 schedule();
Mel Gorman33906bc2010-08-09 17:19:16 -07002452 } else {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002453 if (remaining)
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002454 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002455 else
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002456 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002457 }
2458 }
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002459
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460 order = pgdat->kswapd_max_order;
2461 }
2462 finish_wait(&pgdat->kswapd_wait, &wait);
2463
David Rientjes8fe23e02009-12-14 17:58:33 -08002464 ret = try_to_freeze();
2465 if (kthread_should_stop())
2466 break;
2467
2468 /*
2469 * We can speed up thawing tasks if we don't call balance_pgdat
2470 * after returning from the refrigerator
2471 */
Mel Gorman33906bc2010-08-09 17:19:16 -07002472 if (!ret) {
2473 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002474 balance_pgdat(pgdat, order);
Mel Gorman33906bc2010-08-09 17:19:16 -07002475 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476 }
2477 return 0;
2478}
2479
2480/*
2481 * A zone is low on free memory, so wake its kswapd task to service it.
2482 */
2483void wakeup_kswapd(struct zone *zone, int order)
2484{
2485 pg_data_t *pgdat;
2486
Con Kolivasf3fe6512006-01-06 00:11:15 -08002487 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002488 return;
2489
2490 pgdat = zone->zone_pgdat;
Mel Gorman41858962009-06-16 15:32:12 -07002491 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 return;
2493 if (pgdat->kswapd_max_order < order)
2494 pgdat->kswapd_max_order = order;
Mel Gorman33906bc2010-08-09 17:19:16 -07002495 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Paul Jackson02a0e532006-12-13 00:34:25 -08002496 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002497 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002498 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002499 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002500 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501}
2502
Wu Fengguangadea02a2009-09-21 17:01:42 -07002503/*
2504 * The reclaimable count would be mostly accurate.
2505 * The less reclaimable pages may be
2506 * - mlocked pages, which will be moved to unevictable list when encountered
2507 * - mapped pages, which may require several travels to be reclaimed
2508 * - dirty pages, which is not "instantly" reclaimable
2509 */
2510unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002511{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002512 int nr;
2513
2514 nr = global_page_state(NR_ACTIVE_FILE) +
2515 global_page_state(NR_INACTIVE_FILE);
2516
2517 if (nr_swap_pages > 0)
2518 nr += global_page_state(NR_ACTIVE_ANON) +
2519 global_page_state(NR_INACTIVE_ANON);
2520
2521 return nr;
2522}
2523
2524unsigned long zone_reclaimable_pages(struct zone *zone)
2525{
2526 int nr;
2527
2528 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2529 zone_page_state(zone, NR_INACTIVE_FILE);
2530
2531 if (nr_swap_pages > 0)
2532 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2533 zone_page_state(zone, NR_INACTIVE_ANON);
2534
2535 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002536}
2537
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002538#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002540 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002541 * freed pages.
2542 *
2543 * Rather than trying to age LRUs the aim is to preserve the overall
2544 * LRU order by reclaiming preferentially
2545 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002547unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002549 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002550 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002551 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2552 .may_swap = 1,
2553 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002554 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002555 .nr_to_reclaim = nr_to_reclaim,
2556 .hibernation_mode = 1,
2557 .swappiness = vm_swappiness,
2558 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002559 };
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002560 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2561 struct task_struct *p = current;
2562 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002564 p->flags |= PF_MEMALLOC;
2565 lockdep_set_current_reclaim_state(sc.gfp_mask);
2566 reclaim_state.reclaimed_slab = 0;
2567 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002568
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002569 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002570
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002571 p->reclaim_state = NULL;
2572 lockdep_clear_current_reclaim_state();
2573 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002574
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002575 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002576}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002577#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579/* It's optimal to keep kswapds on the same CPUs as their memory, but
2580 not required for correctness. So if the last cpu in a node goes
2581 away, we get changed to run anywhere: as the first one comes back,
2582 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002583static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002584 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002585{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002586 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002587
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002588 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002589 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002590 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302591 const struct cpumask *mask;
2592
2593 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002594
Rusty Russell3e597942009-01-01 10:12:24 +10302595 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002596 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002597 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002598 }
2599 }
2600 return NOTIFY_OK;
2601}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002602
Yasunori Goto3218ae12006-06-27 02:53:33 -07002603/*
2604 * This kswapd start function will be called by init and node-hot-add.
2605 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2606 */
2607int kswapd_run(int nid)
2608{
2609 pg_data_t *pgdat = NODE_DATA(nid);
2610 int ret = 0;
2611
2612 if (pgdat->kswapd)
2613 return 0;
2614
2615 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2616 if (IS_ERR(pgdat->kswapd)) {
2617 /* failure at boot is fatal */
2618 BUG_ON(system_state == SYSTEM_BOOTING);
2619 printk("Failed to start kswapd on node %d\n",nid);
2620 ret = -1;
2621 }
2622 return ret;
2623}
2624
David Rientjes8fe23e02009-12-14 17:58:33 -08002625/*
2626 * Called by memory hotplug when all memory in a node is offlined.
2627 */
2628void kswapd_stop(int nid)
2629{
2630 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2631
2632 if (kswapd)
2633 kthread_stop(kswapd);
2634}
2635
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636static int __init kswapd_init(void)
2637{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002638 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002639
Linus Torvalds1da177e2005-04-16 15:20:36 -07002640 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002641 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002642 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002643 hotcpu_notifier(cpu_callback, 0);
2644 return 0;
2645}
2646
2647module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002648
2649#ifdef CONFIG_NUMA
2650/*
2651 * Zone reclaim mode
2652 *
2653 * If non-zero call zone_reclaim when the number of free pages falls below
2654 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002655 */
2656int zone_reclaim_mode __read_mostly;
2657
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002658#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002659#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002660#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2661#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2662
Christoph Lameter9eeff232006-01-18 17:42:31 -08002663/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002664 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2665 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2666 * a zone.
2667 */
2668#define ZONE_RECLAIM_PRIORITY 4
2669
Christoph Lameter9eeff232006-01-18 17:42:31 -08002670/*
Christoph Lameter96146342006-07-03 00:24:13 -07002671 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2672 * occur.
2673 */
2674int sysctl_min_unmapped_ratio = 1;
2675
2676/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002677 * If the number of slab pages in a zone grows beyond this percentage then
2678 * slab reclaim needs to occur.
2679 */
2680int sysctl_min_slab_ratio = 5;
2681
Mel Gorman90afa5d2009-06-16 15:33:20 -07002682static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2683{
2684 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2685 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2686 zone_page_state(zone, NR_ACTIVE_FILE);
2687
2688 /*
2689 * It's possible for there to be more file mapped pages than
2690 * accounted for by the pages on the file LRU lists because
2691 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2692 */
2693 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2694}
2695
2696/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2697static long zone_pagecache_reclaimable(struct zone *zone)
2698{
2699 long nr_pagecache_reclaimable;
2700 long delta = 0;
2701
2702 /*
2703 * If RECLAIM_SWAP is set, then all file pages are considered
2704 * potentially reclaimable. Otherwise, we have to worry about
2705 * pages like swapcache and zone_unmapped_file_pages() provides
2706 * a better estimate
2707 */
2708 if (zone_reclaim_mode & RECLAIM_SWAP)
2709 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2710 else
2711 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2712
2713 /* If we can't clean pages, remove dirty pages from consideration */
2714 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2715 delta += zone_page_state(zone, NR_FILE_DIRTY);
2716
2717 /* Watch for any possible underflows due to delta */
2718 if (unlikely(delta > nr_pagecache_reclaimable))
2719 delta = nr_pagecache_reclaimable;
2720
2721 return nr_pagecache_reclaimable - delta;
2722}
2723
Christoph Lameter0ff38492006-09-25 23:31:52 -07002724/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002725 * Try to free up some pages from this zone through reclaim.
2726 */
Andrew Morton179e9632006-03-22 00:08:18 -08002727static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002728{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002729 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002730 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002731 struct task_struct *p = current;
2732 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002733 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002734 struct scan_control sc = {
2735 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002736 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002737 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002738 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2739 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002740 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002741 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002742 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08002743 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07002744 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002745
Christoph Lameter9eeff232006-01-18 17:42:31 -08002746 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002747 /*
2748 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2749 * and we also need to be able to write out pages for RECLAIM_WRITE
2750 * and RECLAIM_SWAP.
2751 */
2752 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002753 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08002754 reclaim_state.reclaimed_slab = 0;
2755 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002756
Mel Gorman90afa5d2009-06-16 15:33:20 -07002757 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07002758 /*
2759 * Free memory by calling shrink zone with increasing
2760 * priorities until we have enough memory freed.
2761 */
2762 priority = ZONE_RECLAIM_PRIORITY;
2763 do {
Rik van Riela79311c2009-01-06 14:40:01 -08002764 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002765 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002766 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002767 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002768
KOSAKI Motohiro15748042010-08-09 17:19:50 -07002769 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2770 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002771 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002772 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002773 * many pages were freed in this zone. So we take the current
2774 * number of slab pages and shake the slab until it is reduced
2775 * by the same nr_pages that we used for reclaiming unmapped
2776 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002777 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002778 * Note that shrink_slab will free memory on all zones and may
2779 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002780 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07002781 for (;;) {
2782 unsigned long lru_pages = zone_reclaimable_pages(zone);
2783
2784 /* No reclaimable slab or very low memory pressure */
2785 if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2786 break;
2787
2788 /* Freed enough memory */
2789 nr_slab_pages1 = zone_page_state(zone,
2790 NR_SLAB_RECLAIMABLE);
2791 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2792 break;
2793 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07002794
2795 /*
2796 * Update nr_reclaimed by the number of slab pages we
2797 * reclaimed from this zone.
2798 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07002799 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2800 if (nr_slab_pages1 < nr_slab_pages0)
2801 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002802 }
2803
Christoph Lameter9eeff232006-01-18 17:42:31 -08002804 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002805 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002806 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08002807 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002808}
Andrew Morton179e9632006-03-22 00:08:18 -08002809
2810int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2811{
Andrew Morton179e9632006-03-22 00:08:18 -08002812 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002813 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002814
2815 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002816 * Zone reclaim reclaims unmapped file backed pages and
2817 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002818 *
Christoph Lameter96146342006-07-03 00:24:13 -07002819 * A small portion of unmapped file backed pages is needed for
2820 * file I/O otherwise pages read by file I/O will be immediately
2821 * thrown out if the zone is overallocated. So we do not reclaim
2822 * if less than a specified percentage of the zone is used by
2823 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002824 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07002825 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2826 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002827 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08002828
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002829 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002830 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07002831
Andrew Morton179e9632006-03-22 00:08:18 -08002832 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002833 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002834 */
David Rientjesd773ed62007-10-16 23:26:01 -07002835 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002836 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08002837
2838 /*
2839 * Only run zone reclaim on the local zone or on zones that do not
2840 * have associated processors. This will favor the local processor
2841 * over remote processors and spread off node memory allocations
2842 * as wide as possible.
2843 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002844 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002845 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07002846 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07002847
2848 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002849 return ZONE_RECLAIM_NOSCAN;
2850
David Rientjesd773ed62007-10-16 23:26:01 -07002851 ret = __zone_reclaim(zone, gfp_mask, order);
2852 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2853
Mel Gorman24cf725182009-06-16 15:33:23 -07002854 if (!ret)
2855 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2856
David Rientjesd773ed62007-10-16 23:26:01 -07002857 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002858}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002859#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002860
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002861/*
2862 * page_evictable - test whether a page is evictable
2863 * @page: the page to test
2864 * @vma: the VMA in which the page is or will be mapped, may be NULL
2865 *
2866 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07002867 * lists vs unevictable list. The vma argument is !NULL when called from the
2868 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002869 *
2870 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002871 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07002872 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002873 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002874 */
2875int page_evictable(struct page *page, struct vm_area_struct *vma)
2876{
2877
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002878 if (mapping_unevictable(page_mapping(page)))
2879 return 0;
2880
Nick Pigginb291f002008-10-18 20:26:44 -07002881 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2882 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002883
2884 return 1;
2885}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002886
2887/**
2888 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2889 * @page: page to check evictability and move to appropriate lru list
2890 * @zone: zone page is in
2891 *
2892 * Checks a page for evictability and moves the page to the appropriate
2893 * zone lru list.
2894 *
2895 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2896 * have PageUnevictable set.
2897 */
2898static void check_move_unevictable_page(struct page *page, struct zone *zone)
2899{
2900 VM_BUG_ON(PageActive(page));
2901
2902retry:
2903 ClearPageUnevictable(page);
2904 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07002905 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002906
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002907 __dec_zone_state(zone, NR_UNEVICTABLE);
2908 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002909 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002910 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2911 __count_vm_event(UNEVICTABLE_PGRESCUED);
2912 } else {
2913 /*
2914 * rotate unevictable list
2915 */
2916 SetPageUnevictable(page);
2917 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002918 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002919 if (page_evictable(page, NULL))
2920 goto retry;
2921 }
2922}
2923
2924/**
2925 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2926 * @mapping: struct address_space to scan for evictable pages
2927 *
2928 * Scan all pages in mapping. Check unevictable pages for
2929 * evictability and move them to the appropriate zone lru list.
2930 */
2931void scan_mapping_unevictable_pages(struct address_space *mapping)
2932{
2933 pgoff_t next = 0;
2934 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2935 PAGE_CACHE_SHIFT;
2936 struct zone *zone;
2937 struct pagevec pvec;
2938
2939 if (mapping->nrpages == 0)
2940 return;
2941
2942 pagevec_init(&pvec, 0);
2943 while (next < end &&
2944 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2945 int i;
2946 int pg_scanned = 0;
2947
2948 zone = NULL;
2949
2950 for (i = 0; i < pagevec_count(&pvec); i++) {
2951 struct page *page = pvec.pages[i];
2952 pgoff_t page_index = page->index;
2953 struct zone *pagezone = page_zone(page);
2954
2955 pg_scanned++;
2956 if (page_index > next)
2957 next = page_index;
2958 next++;
2959
2960 if (pagezone != zone) {
2961 if (zone)
2962 spin_unlock_irq(&zone->lru_lock);
2963 zone = pagezone;
2964 spin_lock_irq(&zone->lru_lock);
2965 }
2966
2967 if (PageLRU(page) && PageUnevictable(page))
2968 check_move_unevictable_page(page, zone);
2969 }
2970 if (zone)
2971 spin_unlock_irq(&zone->lru_lock);
2972 pagevec_release(&pvec);
2973
2974 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2975 }
2976
2977}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002978
2979/**
2980 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2981 * @zone - zone of which to scan the unevictable list
2982 *
2983 * Scan @zone's unevictable LRU lists to check for pages that have become
2984 * evictable. Move those that have to @zone's inactive list where they
2985 * become candidates for reclaim, unless shrink_inactive_zone() decides
2986 * to reactivate them. Pages that are still unevictable are rotated
2987 * back onto @zone's unevictable list.
2988 */
2989#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08002990static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002991{
2992 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2993 unsigned long scan;
2994 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2995
2996 while (nr_to_scan > 0) {
2997 unsigned long batch_size = min(nr_to_scan,
2998 SCAN_UNEVICTABLE_BATCH_SIZE);
2999
3000 spin_lock_irq(&zone->lru_lock);
3001 for (scan = 0; scan < batch_size; scan++) {
3002 struct page *page = lru_to_page(l_unevictable);
3003
3004 if (!trylock_page(page))
3005 continue;
3006
3007 prefetchw_prev_lru_page(page, l_unevictable, flags);
3008
3009 if (likely(PageLRU(page) && PageUnevictable(page)))
3010 check_move_unevictable_page(page, zone);
3011
3012 unlock_page(page);
3013 }
3014 spin_unlock_irq(&zone->lru_lock);
3015
3016 nr_to_scan -= batch_size;
3017 }
3018}
3019
3020
3021/**
3022 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3023 *
3024 * A really big hammer: scan all zones' unevictable LRU lists to check for
3025 * pages that have become evictable. Move those back to the zones'
3026 * inactive list where they become candidates for reclaim.
3027 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3028 * and we add swap to the system. As such, it runs in the context of a task
3029 * that has possibly/probably made some previously unevictable pages
3030 * evictable.
3031 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08003032static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003033{
3034 struct zone *zone;
3035
3036 for_each_zone(zone) {
3037 scan_zone_unevictable_pages(zone);
3038 }
3039}
3040
3041/*
3042 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3043 * all nodes' unevictable lists for evictable pages
3044 */
3045unsigned long scan_unevictable_pages;
3046
3047int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003048 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003049 size_t *length, loff_t *ppos)
3050{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003051 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003052
3053 if (write && *(unsigned long *)table->data)
3054 scan_all_zones_unevictable_pages();
3055
3056 scan_unevictable_pages = 0;
3057 return 0;
3058}
3059
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003060#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003061/*
3062 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3063 * a specified node's per zone unevictable lists for evictable pages.
3064 */
3065
3066static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3067 struct sysdev_attribute *attr,
3068 char *buf)
3069{
3070 return sprintf(buf, "0\n"); /* always zero; should fit... */
3071}
3072
3073static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3074 struct sysdev_attribute *attr,
3075 const char *buf, size_t count)
3076{
3077 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3078 struct zone *zone;
3079 unsigned long res;
3080 unsigned long req = strict_strtoul(buf, 10, &res);
3081
3082 if (!req)
3083 return 1; /* zero is no-op */
3084
3085 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3086 if (!populated_zone(zone))
3087 continue;
3088 scan_zone_unevictable_pages(zone);
3089 }
3090 return 1;
3091}
3092
3093
3094static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3095 read_scan_unevictable_node,
3096 write_scan_unevictable_node);
3097
3098int scan_unevictable_register_node(struct node *node)
3099{
3100 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3101}
3102
3103void scan_unevictable_unregister_node(struct node *node)
3104{
3105 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3106}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003107#endif