blob: b8a6fdc2131258db15321021375117cae27cec0d [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 {
497 __remove_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700498 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700499 mem_cgroup_uncharge_cache_page(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800500 }
501
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800502 return 1;
503
504cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700505 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800506 return 0;
507}
508
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509/*
Nick Piggine2867812008-07-25 19:45:30 -0700510 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
511 * someone else has a ref on the page, abort and return 0. If it was
512 * successfully detached, return 1. Assumes the caller has a single ref on
513 * this page.
514 */
515int remove_mapping(struct address_space *mapping, struct page *page)
516{
517 if (__remove_mapping(mapping, page)) {
518 /*
519 * Unfreezing the refcount with 1 rather than 2 effectively
520 * drops the pagecache ref for us without requiring another
521 * atomic operation.
522 */
523 page_unfreeze_refs(page, 1);
524 return 1;
525 }
526 return 0;
527}
528
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700529/**
530 * putback_lru_page - put previously isolated page onto appropriate LRU list
531 * @page: page to be put back to appropriate lru list
532 *
533 * Add previously isolated @page to appropriate LRU list.
534 * Page may still be unevictable for other reasons.
535 *
536 * lru_lock must not be held, interrupts must be enabled.
537 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700538void putback_lru_page(struct page *page)
539{
540 int lru;
541 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700542 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700543
544 VM_BUG_ON(PageLRU(page));
545
546redo:
547 ClearPageUnevictable(page);
548
549 if (page_evictable(page, NULL)) {
550 /*
551 * For evictable pages, we can use the cache.
552 * In event of a race, worst case is we end up with an
553 * unevictable page on [in]active list.
554 * We know how to handle that.
555 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700556 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700557 lru_cache_add_lru(page, lru);
558 } else {
559 /*
560 * Put unevictable pages directly on zone's unevictable
561 * list.
562 */
563 lru = LRU_UNEVICTABLE;
564 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700565 /*
566 * When racing with an mlock clearing (page is
567 * unlocked), make sure that if the other thread does
568 * not observe our setting of PG_lru and fails
569 * isolation, we see PG_mlocked cleared below and move
570 * the page back to the evictable list.
571 *
572 * The other side is TestClearPageMlocked().
573 */
574 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700575 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700576
577 /*
578 * page's status can change while we move it among lru. If an evictable
579 * page is on unevictable list, it never be freed. To avoid that,
580 * check after we added it to the list, again.
581 */
582 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
583 if (!isolate_lru_page(page)) {
584 put_page(page);
585 goto redo;
586 }
587 /* This means someone else dropped this page from LRU
588 * So, it will be freed or putback to LRU again. There is
589 * nothing to do here.
590 */
591 }
592
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700593 if (was_unevictable && lru != LRU_UNEVICTABLE)
594 count_vm_event(UNEVICTABLE_PGRESCUED);
595 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
596 count_vm_event(UNEVICTABLE_PGCULLED);
597
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700598 put_page(page); /* drop ref from isolate */
599}
600
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800601enum page_references {
602 PAGEREF_RECLAIM,
603 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner64574742010-03-05 13:42:22 -0800604 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800605 PAGEREF_ACTIVATE,
606};
607
608static enum page_references page_check_references(struct page *page,
609 struct scan_control *sc)
610{
Johannes Weiner64574742010-03-05 13:42:22 -0800611 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800612 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800613
Johannes Weiner64574742010-03-05 13:42:22 -0800614 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
615 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800616
617 /* Lumpy reclaim - ignore references */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700618 if (sc->lumpy_reclaim_mode != LUMPY_MODE_NONE)
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800619 return PAGEREF_RECLAIM;
620
621 /*
622 * Mlock lost the isolation race with us. Let try_to_unmap()
623 * move the page to the unevictable list.
624 */
625 if (vm_flags & VM_LOCKED)
626 return PAGEREF_RECLAIM;
627
Johannes Weiner64574742010-03-05 13:42:22 -0800628 if (referenced_ptes) {
629 if (PageAnon(page))
630 return PAGEREF_ACTIVATE;
631 /*
632 * All mapped pages start out with page table
633 * references from the instantiating fault, so we need
634 * to look twice if a mapped file page is used more
635 * than once.
636 *
637 * Mark it and spare it for another trip around the
638 * inactive list. Another page table reference will
639 * lead to its activation.
640 *
641 * Note: the mark is set for activated pages as well
642 * so that recently deactivated but used pages are
643 * quickly recovered.
644 */
645 SetPageReferenced(page);
646
647 if (referenced_page)
648 return PAGEREF_ACTIVATE;
649
650 return PAGEREF_KEEP;
651 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800652
653 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700654 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner64574742010-03-05 13:42:22 -0800655 return PAGEREF_RECLAIM_CLEAN;
656
657 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800658}
659
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700660static noinline_for_stack void free_page_list(struct list_head *free_pages)
661{
662 struct pagevec freed_pvec;
663 struct page *page, *tmp;
664
665 pagevec_init(&freed_pvec, 1);
666
667 list_for_each_entry_safe(page, tmp, free_pages, lru) {
668 list_del(&page->lru);
669 if (!pagevec_add(&freed_pvec, page)) {
670 __pagevec_free(&freed_pvec);
671 pagevec_reinit(&freed_pvec);
672 }
673 }
674
675 pagevec_free(&freed_pvec);
676}
677
Nick Piggine2867812008-07-25 19:45:30 -0700678/*
Andrew Morton1742f192006-03-22 00:08:21 -0800679 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 */
Andrew Morton1742f192006-03-22 00:08:21 -0800681static unsigned long shrink_page_list(struct list_head *page_list,
Mel Gorman0e093d992010-10-26 14:21:45 -0700682 struct zone *zone,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700683 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684{
685 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700686 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700687 int pgactivate = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700688 unsigned long nr_dirty = 0;
689 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800690 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691
692 cond_resched();
693
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800695 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700696 struct address_space *mapping;
697 struct page *page;
698 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700699
700 cond_resched();
701
702 page = lru_to_page(page_list);
703 list_del(&page->lru);
704
Nick Piggin529ae9a2008-08-02 12:01:03 +0200705 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700706 goto keep;
707
Nick Piggin725d7042006-09-25 23:30:55 -0700708 VM_BUG_ON(PageActive(page));
Mel Gorman0e093d992010-10-26 14:21:45 -0700709 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700710
711 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800712
Nick Pigginb291f002008-10-18 20:26:44 -0700713 if (unlikely(!page_evictable(page, NULL)))
714 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700715
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700716 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800717 goto keep_locked;
718
Linus Torvalds1da177e2005-04-16 15:20:36 -0700719 /* Double the slab pressure for mapped and swapcache pages */
720 if (page_mapped(page) || PageSwapCache(page))
721 sc->nr_scanned++;
722
Andy Whitcroftc661b072007-08-22 14:01:26 -0700723 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
724 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
725
726 if (PageWriteback(page)) {
727 /*
728 * Synchronous reclaim is performed in two passes,
729 * first an asynchronous pass over the list to
730 * start parallel writeback, and a second synchronous
731 * pass to wait for the IO to complete. Wait here
732 * for any page for which writeback has already
733 * started.
734 */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700735 if (sc->lumpy_reclaim_mode == LUMPY_MODE_SYNC &&
736 may_enter_fs)
Andy Whitcroftc661b072007-08-22 14:01:26 -0700737 wait_on_page_writeback(page);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700738 else {
739 unlock_page(page);
740 goto keep_lumpy;
741 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700742 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800744 references = page_check_references(page, sc);
745 switch (references) {
746 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700747 goto activate_locked;
Johannes Weiner64574742010-03-05 13:42:22 -0800748 case PAGEREF_KEEP:
749 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800750 case PAGEREF_RECLAIM:
751 case PAGEREF_RECLAIM_CLEAN:
752 ; /* try to reclaim the page below */
753 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700754
Linus Torvalds1da177e2005-04-16 15:20:36 -0700755 /*
756 * Anonymous process memory has backing store?
757 * Try to allocate it some swap space here.
758 */
Nick Pigginb291f002008-10-18 20:26:44 -0700759 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800760 if (!(sc->gfp_mask & __GFP_IO))
761 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800762 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700763 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800764 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700765 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700766
767 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768
769 /*
770 * The page is mapped into the page tables of one or more
771 * processes. Try to unmap it here.
772 */
773 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200774 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 case SWAP_FAIL:
776 goto activate_locked;
777 case SWAP_AGAIN:
778 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700779 case SWAP_MLOCK:
780 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700781 case SWAP_SUCCESS:
782 ; /* try to free the page below */
783 }
784 }
785
786 if (PageDirty(page)) {
Mel Gorman0e093d992010-10-26 14:21:45 -0700787 nr_dirty++;
788
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800789 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700790 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700791 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700792 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800793 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 goto keep_locked;
795
796 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700797 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798 case PAGE_KEEP:
Mel Gorman0e093d992010-10-26 14:21:45 -0700799 nr_congested++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 goto keep_locked;
801 case PAGE_ACTIVATE:
802 goto activate_locked;
803 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700804 if (PageWriteback(page))
805 goto keep_lumpy;
806 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700807 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700808
Linus Torvalds1da177e2005-04-16 15:20:36 -0700809 /*
810 * A synchronous write - probably a ramdisk. Go
811 * ahead and try to reclaim the page.
812 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200813 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700814 goto keep;
815 if (PageDirty(page) || PageWriteback(page))
816 goto keep_locked;
817 mapping = page_mapping(page);
818 case PAGE_CLEAN:
819 ; /* try to free the page below */
820 }
821 }
822
823 /*
824 * If the page has buffers, try to free the buffer mappings
825 * associated with this page. If we succeed we try to free
826 * the page as well.
827 *
828 * We do this even if the page is PageDirty().
829 * try_to_release_page() does not perform I/O, but it is
830 * possible for a page to have PageDirty set, but it is actually
831 * clean (all its buffers are clean). This happens if the
832 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700833 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 * try_to_release_page() will discover that cleanness and will
835 * drop the buffers and mark the page clean - it can be freed.
836 *
837 * Rarely, pages can have buffers and no ->mapping. These are
838 * the pages which were not successfully invalidated in
839 * truncate_complete_page(). We try to drop those buffers here
840 * and if that worked, and the page is no longer mapped into
841 * process address space (page_count == 1) it can be freed.
842 * Otherwise, leave the page on the LRU so it is swappable.
843 */
David Howells266cf652009-04-03 16:42:36 +0100844 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700845 if (!try_to_release_page(page, sc->gfp_mask))
846 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700847 if (!mapping && page_count(page) == 1) {
848 unlock_page(page);
849 if (put_page_testzero(page))
850 goto free_it;
851 else {
852 /*
853 * rare race with speculative reference.
854 * the speculative reference will free
855 * this page shortly, so we may
856 * increment nr_reclaimed here (and
857 * leave it off the LRU).
858 */
859 nr_reclaimed++;
860 continue;
861 }
862 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700863 }
864
Nick Piggine2867812008-07-25 19:45:30 -0700865 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800866 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700867
Nick Piggina978d6f2008-10-18 20:26:58 -0700868 /*
869 * At this point, we have no other references and there is
870 * no way to pick any more up (removed from LRU, removed
871 * from pagecache). Can use non-atomic bitops now (and
872 * we obviously don't have to worry about waking up a process
873 * waiting on the page lock, because there are no references.
874 */
875 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700876free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800877 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700878
879 /*
880 * Is there need to periodically free_page_list? It would
881 * appear not as the counts should be low
882 */
883 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700884 continue;
885
Nick Pigginb291f002008-10-18 20:26:44 -0700886cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800887 if (PageSwapCache(page))
888 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700889 unlock_page(page);
890 putback_lru_page(page);
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700891 disable_lumpy_reclaim_mode(sc);
Nick Pigginb291f002008-10-18 20:26:44 -0700892 continue;
893
Linus Torvalds1da177e2005-04-16 15:20:36 -0700894activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700895 /* Not a candidate for swapping, so reclaim swap space. */
896 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800897 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700898 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700899 SetPageActive(page);
900 pgactivate++;
901keep_locked:
902 unlock_page(page);
903keep:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700904 disable_lumpy_reclaim_mode(sc);
905keep_lumpy:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700906 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700907 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700909
Mel Gorman0e093d992010-10-26 14:21:45 -0700910 /*
911 * Tag a zone as congested if all the dirty pages encountered were
912 * backed by a congested BDI. In this case, reclaimers should just
913 * back off and wait for congestion to clear because further reclaim
914 * will encounter the same problem
915 */
916 if (nr_dirty == nr_congested)
917 zone_set_flag(zone, ZONE_CONGESTED);
918
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700919 free_page_list(&free_pages);
920
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700922 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800923 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700924}
925
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700926/*
927 * Attempt to remove the specified page from its LRU. Only take this page
928 * if it is of the appropriate PageActive status. Pages which are being
929 * freed elsewhere are also ignored.
930 *
931 * page: page to consider
932 * mode: one of the LRU isolation modes defined above
933 *
934 * returns 0 on success, -ve errno on failure.
935 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700936int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700937{
938 int ret = -EINVAL;
939
940 /* Only take pages on the LRU. */
941 if (!PageLRU(page))
942 return ret;
943
944 /*
945 * When checking the active state, we need to be sure we are
946 * dealing with comparible boolean values. Take the logical not
947 * of each.
948 */
949 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
950 return ret;
951
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700952 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700953 return ret;
954
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700955 /*
956 * When this function is being called for lumpy reclaim, we
957 * initially look into all LRU pages, active, inactive and
958 * unevictable; only give shrink_page_list evictable pages.
959 */
960 if (PageUnevictable(page))
961 return ret;
962
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700963 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800964
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700965 if (likely(get_page_unless_zero(page))) {
966 /*
967 * Be careful not to clear PageLRU until after we're
968 * sure the page is not being freed elsewhere -- the
969 * page release code relies on it.
970 */
971 ClearPageLRU(page);
972 ret = 0;
973 }
974
975 return ret;
976}
977
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800978/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979 * zone->lru_lock is heavily contended. Some of the functions that
980 * shrink the lists perform better by taking out a batch of pages
981 * and working on them outside the LRU lock.
982 *
983 * For pagecache intensive workloads, this function is the hottest
984 * spot in the kernel (apart from copy_*_user functions).
985 *
986 * Appropriate locks must be held before calling this function.
987 *
988 * @nr_to_scan: The number of pages to look through on the list.
989 * @src: The LRU list to pull pages off.
990 * @dst: The temp list to put pages on to.
991 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700992 * @order: The caller's attempted allocation order
993 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700994 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700995 *
996 * returns how many pages were moved onto *@dst.
997 */
Andrew Morton69e05942006-03-22 00:08:19 -0800998static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
999 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001000 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001001{
Andrew Morton69e05942006-03-22 00:08:19 -08001002 unsigned long nr_taken = 0;
Mel Gormana8a94d12010-08-09 17:19:17 -07001003 unsigned long nr_lumpy_taken = 0;
1004 unsigned long nr_lumpy_dirty = 0;
1005 unsigned long nr_lumpy_failed = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001006 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001007
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001008 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001009 struct page *page;
1010 unsigned long pfn;
1011 unsigned long end_pfn;
1012 unsigned long page_pfn;
1013 int zone_id;
1014
Linus Torvalds1da177e2005-04-16 15:20:36 -07001015 page = lru_to_page(src);
1016 prefetchw_prev_lru_page(page, src, flags);
1017
Nick Piggin725d7042006-09-25 23:30:55 -07001018 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001019
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001020 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001021 case 0:
1022 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001023 mem_cgroup_del_lru(page);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -08001024 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001025 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001026
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001027 case -EBUSY:
1028 /* else it is being freed elsewhere */
1029 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -07001030 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001031 continue;
1032
1033 default:
1034 BUG();
1035 }
1036
1037 if (!order)
1038 continue;
1039
1040 /*
1041 * Attempt to take all pages in the order aligned region
1042 * surrounding the tag page. Only take those pages of
1043 * the same active state as that tag page. We may safely
1044 * round the target page pfn down to the requested order
1045 * as the mem_map is guarenteed valid out to MAX_ORDER,
1046 * where that page is in a different zone we will detect
1047 * it from its zone id and abort this block scan.
1048 */
1049 zone_id = page_zone_id(page);
1050 page_pfn = page_to_pfn(page);
1051 pfn = page_pfn & ~((1 << order) - 1);
1052 end_pfn = pfn + (1 << order);
1053 for (; pfn < end_pfn; pfn++) {
1054 struct page *cursor_page;
1055
1056 /* The target page is in the block, ignore it. */
1057 if (unlikely(pfn == page_pfn))
1058 continue;
1059
1060 /* Avoid holes within the zone. */
1061 if (unlikely(!pfn_valid_within(pfn)))
1062 break;
1063
1064 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001065
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001066 /* Check that we have not crossed a zone boundary. */
1067 if (unlikely(page_zone_id(cursor_page) != zone_id))
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001068 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001069
1070 /*
1071 * If we don't have enough swap space, reclaiming of
1072 * anon page which don't already have a swap slot is
1073 * pointless.
1074 */
1075 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001076 !PageSwapCache(cursor_page))
1077 break;
Minchan Kimde2e7562009-09-21 17:01:43 -07001078
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -07001079 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001080 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +09001081 mem_cgroup_del_lru(cursor_page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001082 nr_taken++;
Mel Gormana8a94d12010-08-09 17:19:17 -07001083 nr_lumpy_taken++;
1084 if (PageDirty(cursor_page))
1085 nr_lumpy_dirty++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001086 scan++;
Mel Gormana8a94d12010-08-09 17:19:17 -07001087 } else {
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001088 /* the page is freed already. */
1089 if (!page_count(cursor_page))
1090 continue;
1091 break;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001092 }
1093 }
KOSAKI Motohiro08fc4682010-10-26 14:21:44 -07001094
1095 /* If we break out of the loop above, lumpy reclaim failed */
1096 if (pfn < end_pfn)
1097 nr_lumpy_failed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001098 }
1099
1100 *scanned = scan;
Mel Gormana8a94d12010-08-09 17:19:17 -07001101
1102 trace_mm_vmscan_lru_isolate(order,
1103 nr_to_scan, scan,
1104 nr_taken,
1105 nr_lumpy_taken, nr_lumpy_dirty, nr_lumpy_failed,
1106 mode);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001107 return nr_taken;
1108}
1109
Balbir Singh66e17072008-02-07 00:13:56 -08001110static unsigned long isolate_pages_global(unsigned long nr,
1111 struct list_head *dst,
1112 unsigned long *scanned, int order,
1113 int mode, struct zone *z,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001114 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -08001115{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001116 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -08001117 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001118 lru += LRU_ACTIVE;
1119 if (file)
1120 lru += LRU_FILE;
1121 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001122 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -08001123}
1124
Linus Torvalds1da177e2005-04-16 15:20:36 -07001125/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001126 * clear_active_flags() is a helper for shrink_active_list(), clearing
1127 * any active bits from the pages in the list.
1128 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001129static unsigned long clear_active_flags(struct list_head *page_list,
1130 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001131{
1132 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001133 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001134 struct page *page;
1135
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001136 list_for_each_entry(page, page_list, lru) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07001137 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001138 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001139 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001140 ClearPageActive(page);
1141 nr_active++;
1142 }
Mel Gorman1489fa12010-08-09 17:19:33 -07001143 if (count)
1144 count[lru]++;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001145 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001146
1147 return nr_active;
1148}
1149
Nick Piggin62695a82008-10-18 20:26:09 -07001150/**
1151 * isolate_lru_page - tries to isolate a page from its LRU list
1152 * @page: page to isolate from its LRU list
1153 *
1154 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1155 * vmstat statistic corresponding to whatever LRU list the page was on.
1156 *
1157 * Returns 0 if the page was removed from an LRU list.
1158 * Returns -EBUSY if the page was not on an LRU list.
1159 *
1160 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001161 * the active list, it will have PageActive set. If it was found on
1162 * the unevictable list, it will have the PageUnevictable bit set. That flag
1163 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001164 *
1165 * The vmstat statistic corresponding to the list on which the page was
1166 * found will be decremented.
1167 *
1168 * Restrictions:
1169 * (1) Must be called with an elevated refcount on the page. This is a
1170 * fundamentnal difference from isolate_lru_pages (which is called
1171 * without a stable reference).
1172 * (2) the lru_lock must not be held.
1173 * (3) interrupts must be enabled.
1174 */
1175int isolate_lru_page(struct page *page)
1176{
1177 int ret = -EBUSY;
1178
1179 if (PageLRU(page)) {
1180 struct zone *zone = page_zone(page);
1181
1182 spin_lock_irq(&zone->lru_lock);
1183 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001184 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001185 ret = 0;
1186 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001187
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001188 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001189 }
1190 spin_unlock_irq(&zone->lru_lock);
1191 }
1192 return ret;
1193}
1194
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001195/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001196 * Are there way too many processes in the direct reclaim path already?
1197 */
1198static int too_many_isolated(struct zone *zone, int file,
1199 struct scan_control *sc)
1200{
1201 unsigned long inactive, isolated;
1202
1203 if (current_is_kswapd())
1204 return 0;
1205
1206 if (!scanning_global_lru(sc))
1207 return 0;
1208
1209 if (file) {
1210 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1211 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1212 } else {
1213 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1214 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1215 }
1216
1217 return isolated > inactive;
1218}
1219
1220/*
Mel Gorman66635622010-08-09 17:19:30 -07001221 * TODO: Try merging with migrations version of putback_lru_pages
1222 */
1223static noinline_for_stack void
Mel Gorman1489fa12010-08-09 17:19:33 -07001224putback_lru_pages(struct zone *zone, struct scan_control *sc,
Mel Gorman66635622010-08-09 17:19:30 -07001225 unsigned long nr_anon, unsigned long nr_file,
1226 struct list_head *page_list)
1227{
1228 struct page *page;
1229 struct pagevec pvec;
Mel Gorman1489fa12010-08-09 17:19:33 -07001230 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Mel Gorman66635622010-08-09 17:19:30 -07001231
1232 pagevec_init(&pvec, 1);
1233
1234 /*
1235 * Put back any unfreeable pages.
1236 */
1237 spin_lock(&zone->lru_lock);
1238 while (!list_empty(page_list)) {
1239 int lru;
1240 page = lru_to_page(page_list);
1241 VM_BUG_ON(PageLRU(page));
1242 list_del(&page->lru);
1243 if (unlikely(!page_evictable(page, NULL))) {
1244 spin_unlock_irq(&zone->lru_lock);
1245 putback_lru_page(page);
1246 spin_lock_irq(&zone->lru_lock);
1247 continue;
1248 }
1249 SetPageLRU(page);
1250 lru = page_lru(page);
1251 add_page_to_lru_list(zone, page, lru);
1252 if (is_active_lru(lru)) {
1253 int file = is_file_lru(lru);
1254 reclaim_stat->recent_rotated[file]++;
1255 }
1256 if (!pagevec_add(&pvec, page)) {
1257 spin_unlock_irq(&zone->lru_lock);
1258 __pagevec_release(&pvec);
1259 spin_lock_irq(&zone->lru_lock);
1260 }
1261 }
1262 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1263 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1264
1265 spin_unlock_irq(&zone->lru_lock);
1266 pagevec_release(&pvec);
1267}
1268
Mel Gorman1489fa12010-08-09 17:19:33 -07001269static noinline_for_stack void update_isolated_counts(struct zone *zone,
1270 struct scan_control *sc,
1271 unsigned long *nr_anon,
1272 unsigned long *nr_file,
1273 struct list_head *isolated_list)
1274{
1275 unsigned long nr_active;
1276 unsigned int count[NR_LRU_LISTS] = { 0, };
1277 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
1278
1279 nr_active = clear_active_flags(isolated_list, count);
1280 __count_vm_events(PGDEACTIVATE, nr_active);
1281
1282 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1283 -count[LRU_ACTIVE_FILE]);
1284 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1285 -count[LRU_INACTIVE_FILE]);
1286 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1287 -count[LRU_ACTIVE_ANON]);
1288 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1289 -count[LRU_INACTIVE_ANON]);
1290
1291 *nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1292 *nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1293 __mod_zone_page_state(zone, NR_ISOLATED_ANON, *nr_anon);
1294 __mod_zone_page_state(zone, NR_ISOLATED_FILE, *nr_file);
1295
1296 reclaim_stat->recent_scanned[0] += *nr_anon;
1297 reclaim_stat->recent_scanned[1] += *nr_file;
1298}
1299
Mel Gorman66635622010-08-09 17:19:30 -07001300/*
Wu Fengguange31f36982010-08-09 17:20:01 -07001301 * Returns true if the caller should wait to clean dirty/writeback pages.
1302 *
1303 * If we are direct reclaiming for contiguous pages and we do not reclaim
1304 * everything in the list, try again and wait for writeback IO to complete.
1305 * This will stall high-order allocations noticeably. Only do that when really
1306 * need to free the pages under high memory pressure.
1307 */
1308static inline bool should_reclaim_stall(unsigned long nr_taken,
1309 unsigned long nr_freed,
1310 int priority,
1311 struct scan_control *sc)
1312{
1313 int lumpy_stall_priority;
1314
1315 /* kswapd should not stall on sync IO */
1316 if (current_is_kswapd())
1317 return false;
1318
1319 /* Only stall on lumpy reclaim */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001320 if (sc->lumpy_reclaim_mode == LUMPY_MODE_NONE)
Wu Fengguange31f36982010-08-09 17:20:01 -07001321 return false;
1322
1323 /* If we have relaimed everything on the isolated list, no stall */
1324 if (nr_freed == nr_taken)
1325 return false;
1326
1327 /*
1328 * For high-order allocations, there are two stall thresholds.
1329 * High-cost allocations stall immediately where as lower
1330 * order allocations such as stacks require the scanning
1331 * priority to be much higher before stalling.
1332 */
1333 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1334 lumpy_stall_priority = DEF_PRIORITY;
1335 else
1336 lumpy_stall_priority = DEF_PRIORITY / 3;
1337
1338 return priority <= lumpy_stall_priority;
1339}
1340
1341/*
Andrew Morton1742f192006-03-22 00:08:21 -08001342 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1343 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001344 */
Mel Gorman66635622010-08-09 17:19:30 -07001345static noinline_for_stack unsigned long
1346shrink_inactive_list(unsigned long nr_to_scan, struct zone *zone,
1347 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348{
1349 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001350 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001351 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001352 unsigned long nr_taken;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001353 unsigned long nr_anon;
1354 unsigned long nr_file;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001355
Rik van Riel35cd7812009-09-21 17:01:38 -07001356 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001357 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001358
1359 /* We are about to die and free our memory. Return now. */
1360 if (fatal_signal_pending(current))
1361 return SWAP_CLUSTER_MAX;
1362 }
1363
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001364 set_lumpy_reclaim_mode(priority, sc, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365 lru_add_drain();
1366 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001367
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001368 if (scanning_global_lru(sc)) {
1369 nr_taken = isolate_pages_global(nr_to_scan,
1370 &page_list, &nr_scanned, sc->order,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001371 sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1372 ISOLATE_INACTIVE : ISOLATE_BOTH,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001373 zone, 0, file);
1374 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001375 if (current_is_kswapd())
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001376 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1377 nr_scanned);
1378 else
1379 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1380 nr_scanned);
1381 } else {
1382 nr_taken = mem_cgroup_isolate_pages(nr_to_scan,
1383 &page_list, &nr_scanned, sc->order,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001384 sc->lumpy_reclaim_mode == LUMPY_MODE_NONE ?
1385 ISOLATE_INACTIVE : ISOLATE_BOTH,
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001386 zone, sc->mem_cgroup,
1387 0, file);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388 /*
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001389 * mem_cgroup_isolate_pages() keeps track of
1390 * scanned pages on its own.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001391 */
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001392 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001393
Mel Gorman66635622010-08-09 17:19:30 -07001394 if (nr_taken == 0) {
1395 spin_unlock_irq(&zone->lru_lock);
1396 return 0;
1397 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001398
Mel Gorman1489fa12010-08-09 17:19:33 -07001399 update_isolated_counts(zone, sc, &nr_anon, &nr_file, &page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001400
1401 spin_unlock_irq(&zone->lru_lock);
1402
Mel Gorman0e093d992010-10-26 14:21:45 -07001403 nr_reclaimed = shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001404
Wu Fengguange31f36982010-08-09 17:20:01 -07001405 /* Check if we should syncronously wait for writeback */
1406 if (should_reclaim_stall(nr_taken, nr_reclaimed, priority, sc)) {
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001407 set_lumpy_reclaim_mode(priority, sc, true);
Mel Gorman0e093d992010-10-26 14:21:45 -07001408 nr_reclaimed += shrink_page_list(&page_list, zone, sc);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001409 }
1410
1411 local_irq_disable();
1412 if (current_is_kswapd())
1413 __count_vm_events(KSWAPD_STEAL, nr_reclaimed);
1414 __count_zone_vm_events(PGSTEAL, zone, nr_reclaimed);
1415
Mel Gorman1489fa12010-08-09 17:19:33 -07001416 putback_lru_pages(zone, sc, nr_anon, nr_file, &page_list);
Mel Gormane11da5b2010-10-26 14:21:40 -07001417
1418 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1419 zone_idx(zone),
1420 nr_scanned, nr_reclaimed,
1421 priority,
1422 trace_shrink_flags(file, sc->lumpy_reclaim_mode));
Andrew Morton05ff5132006-03-22 00:08:20 -08001423 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001424}
1425
Martin Bligh3bb1a852006-10-28 10:38:24 -07001426/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427 * This moves pages from the active list to the inactive list.
1428 *
1429 * We move them the other way if the page is referenced by one or more
1430 * processes, from rmap.
1431 *
1432 * If the pages are mostly unmapped, the processing is fast and it is
1433 * appropriate to hold zone->lru_lock across the whole operation. But if
1434 * the pages are mapped, the processing is slow (page_referenced()) so we
1435 * should drop zone->lru_lock around each page. It's impossible to balance
1436 * this, so instead we remove the pages from the LRU while processing them.
1437 * It is safe to rely on PG_active against the non-LRU pages in here because
1438 * nobody will play with that bit on a non-LRU page.
1439 *
1440 * The downside is that we have to touch page->_count against each page.
1441 * But we had to alter page->flags anyway.
1442 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001443
Wu Fengguang3eb41402009-06-16 15:33:13 -07001444static void move_active_pages_to_lru(struct zone *zone,
1445 struct list_head *list,
1446 enum lru_list lru)
1447{
1448 unsigned long pgmoved = 0;
1449 struct pagevec pvec;
1450 struct page *page;
1451
1452 pagevec_init(&pvec, 1);
1453
1454 while (!list_empty(list)) {
1455 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001456
1457 VM_BUG_ON(PageLRU(page));
1458 SetPageLRU(page);
1459
Wu Fengguang3eb41402009-06-16 15:33:13 -07001460 list_move(&page->lru, &zone->lru[lru].list);
1461 mem_cgroup_add_lru_list(page, lru);
1462 pgmoved++;
1463
1464 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1465 spin_unlock_irq(&zone->lru_lock);
1466 if (buffer_heads_over_limit)
1467 pagevec_strip(&pvec);
1468 __pagevec_release(&pvec);
1469 spin_lock_irq(&zone->lru_lock);
1470 }
1471 }
1472 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1473 if (!is_active_lru(lru))
1474 __count_vm_events(PGDEACTIVATE, pgmoved);
1475}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001476
Andrew Morton1742f192006-03-22 00:08:21 -08001477static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001478 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001480 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001481 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001482 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001483 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001484 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001485 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001486 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001487 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001488 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001489
1490 lru_add_drain();
1491 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001492 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001493 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1494 &pgscanned, sc->order,
1495 ISOLATE_ACTIVE, zone,
1496 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001497 zone->pages_scanned += pgscanned;
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001498 } else {
1499 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1500 &pgscanned, sc->order,
1501 ISOLATE_ACTIVE, zone,
1502 sc->mem_cgroup, 1, file);
1503 /*
1504 * mem_cgroup_isolate_pages() keeps track of
1505 * scanned pages on its own.
1506 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001507 }
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001508
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001509 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001510
Wu Fengguang3eb41402009-06-16 15:33:13 -07001511 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001512 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001513 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001514 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001515 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001516 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517 spin_unlock_irq(&zone->lru_lock);
1518
Linus Torvalds1da177e2005-04-16 15:20:36 -07001519 while (!list_empty(&l_hold)) {
1520 cond_resched();
1521 page = lru_to_page(&l_hold);
1522 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001523
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001524 if (unlikely(!page_evictable(page, NULL))) {
1525 putback_lru_page(page);
1526 continue;
1527 }
1528
Johannes Weiner64574742010-03-05 13:42:22 -08001529 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001530 nr_rotated++;
Wu Fengguang8cab4752009-06-16 15:33:12 -07001531 /*
1532 * Identify referenced, file-backed active pages and
1533 * give them one more trip around the active list. So
1534 * that executable code get better chances to stay in
1535 * memory under moderate memory pressure. Anon pages
1536 * are not likely to be evicted by use-once streaming
1537 * IO, plus JVM can create lots of anon VM_EXEC pages,
1538 * so we ignore them here.
1539 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001540 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001541 list_add(&page->lru, &l_active);
1542 continue;
1543 }
1544 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001545
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001546 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001547 list_add(&page->lru, &l_inactive);
1548 }
1549
Andrew Mortonb5557492009-01-06 14:40:13 -08001550 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001551 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001552 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001553 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001554 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001555 * Count referenced pages from currently used mappings as rotated,
1556 * even though only some of them are actually re-activated. This
1557 * helps balance scan pressure between file and anonymous pages in
1558 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001559 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001560 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001561
Wu Fengguang3eb41402009-06-16 15:33:13 -07001562 move_active_pages_to_lru(zone, &l_active,
1563 LRU_ACTIVE + file * LRU_FILE);
1564 move_active_pages_to_lru(zone, &l_inactive,
1565 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001566 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001567 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001568}
1569
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001570#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001571static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001572{
1573 unsigned long active, inactive;
1574
1575 active = zone_page_state(zone, NR_ACTIVE_ANON);
1576 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1577
1578 if (inactive * zone->inactive_ratio < active)
1579 return 1;
1580
1581 return 0;
1582}
1583
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001584/**
1585 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1586 * @zone: zone to check
1587 * @sc: scan control of this context
1588 *
1589 * Returns true if the zone does not have enough inactive anon pages,
1590 * meaning some active anon pages need to be deactivated.
1591 */
1592static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1593{
1594 int low;
1595
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001596 /*
1597 * If we don't have swap space, anonymous page deactivation
1598 * is pointless.
1599 */
1600 if (!total_swap_pages)
1601 return 0;
1602
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001603 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001604 low = inactive_anon_is_low_global(zone);
1605 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001606 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001607 return low;
1608}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001609#else
1610static inline int inactive_anon_is_low(struct zone *zone,
1611 struct scan_control *sc)
1612{
1613 return 0;
1614}
1615#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001616
Rik van Riel56e49d22009-06-16 15:32:28 -07001617static int inactive_file_is_low_global(struct zone *zone)
1618{
1619 unsigned long active, inactive;
1620
1621 active = zone_page_state(zone, NR_ACTIVE_FILE);
1622 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1623
1624 return (active > inactive);
1625}
1626
1627/**
1628 * inactive_file_is_low - check if file pages need to be deactivated
1629 * @zone: zone to check
1630 * @sc: scan control of this context
1631 *
1632 * When the system is doing streaming IO, memory pressure here
1633 * ensures that active file pages get deactivated, until more
1634 * than half of the file pages are on the inactive list.
1635 *
1636 * Once we get to that situation, protect the system's working
1637 * set from being evicted by disabling active file page aging.
1638 *
1639 * This uses a different ratio than the anonymous pages, because
1640 * the page cache uses a use-once replacement algorithm.
1641 */
1642static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1643{
1644 int low;
1645
1646 if (scanning_global_lru(sc))
1647 low = inactive_file_is_low_global(zone);
1648 else
1649 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1650 return low;
1651}
1652
Rik van Rielb39415b2009-12-14 17:59:48 -08001653static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1654 int file)
1655{
1656 if (file)
1657 return inactive_file_is_low(zone, sc);
1658 else
1659 return inactive_anon_is_low(zone, sc);
1660}
1661
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001662static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001663 struct zone *zone, struct scan_control *sc, int priority)
1664{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001665 int file = is_file_lru(lru);
1666
Rik van Rielb39415b2009-12-14 17:59:48 -08001667 if (is_active_lru(lru)) {
1668 if (inactive_list_is_low(zone, sc, file))
1669 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001670 return 0;
1671 }
1672
Rik van Riel33c120e2008-10-18 20:26:36 -07001673 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001674}
1675
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676/*
Shaohua Li76a33fc2010-05-24 14:32:36 -07001677 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1678 * until we collected @swap_cluster_max pages to scan.
1679 */
1680static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1681 unsigned long *nr_saved_scan)
1682{
1683 unsigned long nr;
1684
1685 *nr_saved_scan += nr_to_scan;
1686 nr = *nr_saved_scan;
1687
1688 if (nr >= SWAP_CLUSTER_MAX)
1689 *nr_saved_scan = 0;
1690 else
1691 nr = 0;
1692
1693 return nr;
1694}
1695
1696/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001697 * Determine how aggressively the anon and file LRU lists should be
1698 * scanned. The relative value of each set of LRU lists is determined
1699 * by looking at the fraction of the pages scanned we did rotate back
1700 * onto the active list instead of evict.
1701 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001702 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001703 */
Shaohua Li76a33fc2010-05-24 14:32:36 -07001704static void get_scan_count(struct zone *zone, struct scan_control *sc,
1705 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001706{
1707 unsigned long anon, file, free;
1708 unsigned long anon_prio, file_prio;
1709 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001710 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001711 u64 fraction[2], denominator;
1712 enum lru_list l;
1713 int noswap = 0;
1714
1715 /* If we have no swap space, do not bother scanning anon pages. */
1716 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1717 noswap = 1;
1718 fraction[0] = 0;
1719 fraction[1] = 1;
1720 denominator = 1;
1721 goto out;
1722 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001723
Vincent Li0b217672009-09-21 17:03:09 -07001724 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1725 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1726 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1727 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001728
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001729 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001730 free = zone_page_state(zone, NR_FREE_PAGES);
1731 /* If we have very few page cache pages,
1732 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001733 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001734 fraction[0] = 1;
1735 fraction[1] = 0;
1736 denominator = 1;
1737 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001738 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001739 }
1740
1741 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001742 * With swappiness at 100, anonymous and file have the same priority.
1743 * This scanning priority is essentially the inverse of IO cost.
1744 */
1745 anon_prio = sc->swappiness;
1746 file_prio = 200 - sc->swappiness;
1747
1748 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001749 * OK, so we have swap space and a fair amount of page cache
1750 * pages. We use the recently rotated / recently scanned
1751 * ratios to determine how valuable each cache is.
1752 *
1753 * Because workloads change over time (and to avoid overflow)
1754 * we keep these statistics as a floating average, which ends
1755 * up weighing recent references more than old ones.
1756 *
1757 * anon in [0], file in [1]
1758 */
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001759 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001760 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001761 reclaim_stat->recent_scanned[0] /= 2;
1762 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001763 }
1764
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001765 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001766 reclaim_stat->recent_scanned[1] /= 2;
1767 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001768 }
1769
1770 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001771 * The amount of pressure on anon vs file pages is inversely
1772 * proportional to the fraction of recently scanned pages on
1773 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001774 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001775 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1776 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001777
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001778 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1779 fp /= reclaim_stat->recent_rotated[1] + 1;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001780 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001781
Shaohua Li76a33fc2010-05-24 14:32:36 -07001782 fraction[0] = ap;
1783 fraction[1] = fp;
1784 denominator = ap + fp + 1;
1785out:
1786 for_each_evictable_lru(l) {
1787 int file = is_file_lru(l);
1788 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001789
Shaohua Li76a33fc2010-05-24 14:32:36 -07001790 scan = zone_nr_lru_pages(zone, sc, l);
1791 if (priority || noswap) {
1792 scan >>= priority;
1793 scan = div64_u64(scan * fraction[file], denominator);
1794 }
1795 nr[l] = nr_scan_try_batch(scan,
1796 &reclaim_stat->nr_saved_scan[l]);
1797 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001798}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001799
1800/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1802 */
Rik van Riela79311c2009-01-06 14:40:01 -08001803static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001804 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001805{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001806 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001807 unsigned long nr_to_scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001808 enum lru_list l;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001809 unsigned long nr_reclaimed = sc->nr_reclaimed;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001810 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811
Shaohua Li76a33fc2010-05-24 14:32:36 -07001812 get_scan_count(zone, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001813
Rik van Riel556adec2008-10-18 20:26:34 -07001814 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1815 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001816 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001817 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001818 nr_to_scan = min_t(unsigned long,
1819 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001820 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001821
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001822 nr_reclaimed += shrink_list(l, nr_to_scan,
1823 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001824 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001825 }
Rik van Riela79311c2009-01-06 14:40:01 -08001826 /*
1827 * On large memory systems, scan >> priority can become
1828 * really large. This is fine for the starting priority;
1829 * we want to put equal scanning pressure on each zone.
1830 * However, if the VM has a harder time of freeing pages,
1831 * with multiple processes reclaiming pages, the total
1832 * freeing target can get unreasonably large.
1833 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08001834 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001835 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 }
1837
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001838 sc->nr_reclaimed = nr_reclaimed;
1839
Rik van Riel556adec2008-10-18 20:26:34 -07001840 /*
1841 * Even if we did not try to evict anon pages at all, we want to
1842 * rebalance the anon lru active/inactive ratio.
1843 */
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001844 if (inactive_anon_is_low(zone, sc))
Rik van Riel556adec2008-10-18 20:26:34 -07001845 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1846
Andrew Morton232ea4d2007-02-28 20:13:21 -08001847 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001848}
1849
1850/*
1851 * This is the direct reclaim path, for page-allocating processes. We only
1852 * try to reclaim pages from zones which will satisfy the caller's allocation
1853 * request.
1854 *
Mel Gorman41858962009-06-16 15:32:12 -07001855 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1856 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001857 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1858 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001859 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1860 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1861 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863 * If a zone is deemed to be full of pinned pages then just give it a light
1864 * scan then give up on it.
1865 */
Minchan Kimd1908362010-09-22 13:05:01 -07001866static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001867 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001868{
Mel Gormandd1a2392008-04-28 02:12:17 -07001869 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001870 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001871
Mel Gormand4debc62010-08-09 17:19:29 -07001872 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1873 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001874 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001876 /*
1877 * Take care memory controller reclaiming has small influence
1878 * to global LRU.
1879 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001880 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001881 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1882 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08001883 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001884 continue; /* Let kswapd poll it */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001885 }
Nick Piggin408d8542006-09-25 23:31:27 -07001886
Rik van Riela79311c2009-01-06 14:40:01 -08001887 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001888 }
Minchan Kimd1908362010-09-22 13:05:01 -07001889}
1890
1891static bool zone_reclaimable(struct zone *zone)
1892{
1893 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
1894}
1895
1896/*
1897 * As hibernation is going on, kswapd is freezed so that it can't mark
1898 * the zone into all_unreclaimable. It can't handle OOM during hibernation.
1899 * So let's check zone's unreclaimable in direct reclaim as well as kswapd.
1900 */
1901static bool all_unreclaimable(struct zonelist *zonelist,
1902 struct scan_control *sc)
1903{
1904 struct zoneref *z;
1905 struct zone *zone;
1906 bool all_unreclaimable = true;
1907
1908 for_each_zone_zonelist_nodemask(zone, z, zonelist,
1909 gfp_zone(sc->gfp_mask), sc->nodemask) {
1910 if (!populated_zone(zone))
1911 continue;
1912 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1913 continue;
1914 if (zone_reclaimable(zone)) {
1915 all_unreclaimable = false;
1916 break;
1917 }
1918 }
1919
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001920 return all_unreclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001922
Linus Torvalds1da177e2005-04-16 15:20:36 -07001923/*
1924 * This is the main entry point to direct page reclaim.
1925 *
1926 * If a full scan of the inactive list fails to free enough memory then we
1927 * are "out of memory" and something needs to be killed.
1928 *
1929 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1930 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02001931 * caller can't do much about. We kick the writeback threads and take explicit
1932 * naps in the hope that some of these pages can be written. But if the
1933 * allocating task holds filesystem locks which prevent writeout this might not
1934 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001935 *
1936 * returns: 0, if no pages reclaimed
1937 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001938 */
Mel Gormandac1d272008-04-28 02:12:12 -07001939static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001940 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941{
1942 int priority;
Andrew Morton69e05942006-03-22 00:08:19 -08001943 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001944 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07001945 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001946 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001947 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948
Miao Xiec0ff7452010-05-24 14:32:08 -07001949 get_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07001950 delayacct_freepages_start();
1951
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001952 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001953 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954
1955 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001956 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001957 if (!priority)
1958 disable_swap_token();
Minchan Kimd1908362010-09-22 13:05:01 -07001959 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001960 /*
1961 * Don't shrink slabs when reclaiming memory from
1962 * over limit cgroups
1963 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001964 if (scanning_global_lru(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07001965 unsigned long lru_pages = 0;
Mel Gormand4debc62010-08-09 17:19:29 -07001966 for_each_zone_zonelist(zone, z, zonelist,
1967 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07001968 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1969 continue;
1970
1971 lru_pages += zone_reclaimable_pages(zone);
1972 }
1973
Mel Gormandd1a2392008-04-28 02:12:17 -07001974 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001975 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08001976 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001977 reclaim_state->reclaimed_slab = 0;
1978 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001979 }
Balbir Singh66e17072008-02-07 00:13:56 -08001980 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001981 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001982 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001983
1984 /*
1985 * Try to write back as many pages as we just scanned. This
1986 * tends to cause slow streaming writers to write data to the
1987 * disk smoothly, at the dirtying rate, which is nice. But
1988 * that's undesirable in laptop mode, where we *want* lumpy
1989 * writeout. So in laptop mode, write out the whole world.
1990 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001991 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1992 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001993 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001994 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001995 }
1996
1997 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08001998 if (!sc->hibernation_mode && sc->nr_scanned &&
Mel Gorman0e093d992010-10-26 14:21:45 -07001999 priority < DEF_PRIORITY - 2) {
2000 struct zone *preferred_zone;
2001
2002 first_zones_zonelist(zonelist, gfp_zone(sc->gfp_mask),
2003 NULL, &preferred_zone);
2004 wait_iff_congested(preferred_zone, BLK_RW_ASYNC, HZ/10);
2005 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002006 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002007
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002009 delayacct_freepages_end();
Miao Xiec0ff7452010-05-24 14:32:08 -07002010 put_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07002011
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002012 if (sc->nr_reclaimed)
2013 return sc->nr_reclaimed;
2014
2015 /* top priority shrink_zones still had more to do? don't OOM, then */
Minchan Kimd1908362010-09-22 13:05:01 -07002016 if (scanning_global_lru(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002017 return 1;
2018
2019 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020}
2021
Mel Gormandac1d272008-04-28 02:12:12 -07002022unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002023 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002024{
Mel Gorman33906bc2010-08-09 17:19:16 -07002025 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002026 struct scan_control sc = {
2027 .gfp_mask = gfp_mask,
2028 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002029 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002030 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002031 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002032 .swappiness = vm_swappiness,
2033 .order = order,
2034 .mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002035 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002036 };
2037
Mel Gorman33906bc2010-08-09 17:19:16 -07002038 trace_mm_vmscan_direct_reclaim_begin(order,
2039 sc.may_writepage,
2040 gfp_mask);
2041
2042 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2043
2044 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2045
2046 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002047}
2048
Balbir Singh00f0b822008-03-04 14:28:39 -08002049#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08002050
Balbir Singh4e416952009-09-23 15:56:39 -07002051unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
2052 gfp_t gfp_mask, bool noswap,
2053 unsigned int swappiness,
KOSAKI Motohiro14fec792010-08-10 18:03:05 -07002054 struct zone *zone)
Balbir Singh4e416952009-09-23 15:56:39 -07002055{
2056 struct scan_control sc = {
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002057 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002058 .may_writepage = !laptop_mode,
2059 .may_unmap = 1,
2060 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002061 .swappiness = swappiness,
2062 .order = 0,
2063 .mem_cgroup = mem,
Balbir Singh4e416952009-09-23 15:56:39 -07002064 };
Balbir Singh4e416952009-09-23 15:56:39 -07002065 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2066 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002067
2068 trace_mm_vmscan_memcg_softlimit_reclaim_begin(0,
2069 sc.may_writepage,
2070 sc.gfp_mask);
2071
Balbir Singh4e416952009-09-23 15:56:39 -07002072 /*
2073 * NOTE: Although we can get the priority field, using it
2074 * here is not a good idea, since it limits the pages we can scan.
2075 * if we don't reclaim here, the shrink_zone from balance_pgdat
2076 * will pick up pages from other mem cgroup's as well. We hack
2077 * the priority and make it zero.
2078 */
2079 shrink_zone(0, zone, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002080
2081 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2082
Balbir Singh4e416952009-09-23 15:56:39 -07002083 return sc.nr_reclaimed;
2084}
2085
Balbir Singhe1a1cd52008-02-07 00:14:02 -08002086unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002087 gfp_t gfp_mask,
2088 bool noswap,
2089 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08002090{
Balbir Singh4e416952009-09-23 15:56:39 -07002091 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002092 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002093 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002094 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002095 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002096 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002097 .nr_to_reclaim = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002098 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08002099 .order = 0,
2100 .mem_cgroup = mem_cont,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002101 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08002102 };
Balbir Singh66e17072008-02-07 00:13:56 -08002103
Mel Gormandd1a2392008-04-28 02:12:17 -07002104 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2105 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
2106 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002107
2108 trace_mm_vmscan_memcg_reclaim_begin(0,
2109 sc.may_writepage,
2110 sc.gfp_mask);
2111
2112 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
2113
2114 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2115
2116 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002117}
2118#endif
2119
Mel Gormanf50de2d2009-12-14 17:58:53 -08002120/* is kswapd sleeping prematurely? */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002121static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002122{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002123 int i;
Mel Gormanf50de2d2009-12-14 17:58:53 -08002124
2125 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2126 if (remaining)
2127 return 1;
2128
2129 /* If after HZ/10, a zone is below the high mark, it's premature */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002130 for (i = 0; i < pgdat->nr_zones; i++) {
2131 struct zone *zone = pgdat->node_zones + i;
2132
2133 if (!populated_zone(zone))
2134 continue;
2135
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002136 if (zone->all_unreclaimable)
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08002137 continue;
2138
Mel Gormanf50de2d2009-12-14 17:58:53 -08002139 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
2140 0, 0))
2141 return 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002142 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002143
2144 return 0;
2145}
2146
Linus Torvalds1da177e2005-04-16 15:20:36 -07002147/*
2148 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002149 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002151 * Returns the number of pages which were actually freed.
2152 *
2153 * There is special handling here for zones which are full of pinned pages.
2154 * This can happen if the pages are all mlocked, or if they are all used by
2155 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2156 * What we do is to detect the case where all pages in the zone have been
2157 * scanned twice and there has been zero successful reclaim. Mark the zone as
2158 * dead and from now on, only perform a short scan. Basically we're polling
2159 * the zone for when the problem goes away.
2160 *
2161 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002162 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2163 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2164 * lower zones regardless of the number of free pages in the lower zones. This
2165 * interoperates with the page allocator fallback scheme to ensure that aging
2166 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002167 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002168static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002169{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002170 int all_zones_ok;
2171 int priority;
2172 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08002173 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002174 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08002175 struct scan_control sc = {
2176 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002177 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002178 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002179 /*
2180 * kswapd doesn't want to be bailed out while reclaim. because
2181 * we want to put equal scanning pressure on each zone.
2182 */
2183 .nr_to_reclaim = ULONG_MAX,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002184 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002185 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002186 .mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002187 };
Linus Torvalds1da177e2005-04-16 15:20:36 -07002188loop_again:
2189 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002190 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002191 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002192 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002193
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2195 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2196 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002197 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002199 /* The swap token gets in the way of swapout... */
2200 if (!priority)
2201 disable_swap_token();
2202
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203 all_zones_ok = 1;
2204
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002205 /*
2206 * Scan in the highmem->dma direction for the highest
2207 * zone which needs scanning
2208 */
2209 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2210 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002211
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002212 if (!populated_zone(zone))
2213 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002215 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002216 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217
Rik van Riel556adec2008-10-18 20:26:34 -07002218 /*
2219 * Do some background aging of the anon list, to give
2220 * pages a chance to be referenced before reclaiming.
2221 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002222 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002223 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2224 &sc, priority, 0);
2225
Mel Gorman41858962009-06-16 15:32:12 -07002226 if (!zone_watermark_ok(zone, order,
2227 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002228 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002229 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002231 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002232 if (i < 0)
2233 goto out;
2234
Linus Torvalds1da177e2005-04-16 15:20:36 -07002235 for (i = 0; i <= end_zone; i++) {
2236 struct zone *zone = pgdat->node_zones + i;
2237
Wu Fengguangadea02a2009-09-21 17:01:42 -07002238 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239 }
2240
2241 /*
2242 * Now scan the zone in the dma->highmem direction, stopping
2243 * at the last zone which needs scanning.
2244 *
2245 * We do this because the page allocator works in the opposite
2246 * direction. This prevents the page allocator from allocating
2247 * pages behind kswapd's direction of progress, which would
2248 * cause too much scanning of the lower zones.
2249 */
2250 for (i = 0; i <= end_zone; i++) {
2251 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002252 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253
Con Kolivasf3fe6512006-01-06 00:11:15 -08002254 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 continue;
2256
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002257 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002258 continue;
2259
Linus Torvalds1da177e2005-04-16 15:20:36 -07002260 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07002261
Balbir Singh4e416952009-09-23 15:56:39 -07002262 /*
2263 * Call soft limit reclaim before calling shrink_zone.
2264 * For now we ignore the return value
2265 */
KOSAKI Motohiro00918b62010-08-10 18:03:05 -07002266 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask);
2267
Rik van Riel32a43302007-10-16 01:24:50 -07002268 /*
2269 * We put equal pressure on every zone, unless one
2270 * zone has way too many pages free already.
2271 */
Mel Gorman41858962009-06-16 15:32:12 -07002272 if (!zone_watermark_ok(zone, order,
2273 8*high_wmark_pages(zone), end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002274 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002276 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2277 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002278 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002279 total_scanned += sc.nr_scanned;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002280 if (zone->all_unreclaimable)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002281 continue;
Minchan Kimd1908362010-09-22 13:05:01 -07002282 if (nr_slab == 0 && !zone_reclaimable(zone))
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002283 zone->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002284 /*
2285 * If we've done a decent amount of scanning and
2286 * the reclaim ratio is low, start doing writepage
2287 * even in laptop mode
2288 */
2289 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002290 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002291 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002292
Minchan Kim45973d72010-03-05 13:41:45 -08002293 if (!zone_watermark_ok(zone, order,
2294 high_wmark_pages(zone), end_zone, 0)) {
2295 all_zones_ok = 0;
2296 /*
2297 * We are still under min water mark. This
2298 * means that we have a GFP_ATOMIC allocation
2299 * failure risk. Hurry up!
2300 */
2301 if (!zone_watermark_ok(zone, order,
2302 min_wmark_pages(zone), end_zone, 0))
2303 has_under_min_watermark_zone = 1;
Mel Gorman0e093d992010-10-26 14:21:45 -07002304 } else {
2305 /*
2306 * If a zone reaches its high watermark,
2307 * consider it to be no longer congested. It's
2308 * possible there are dirty pages backed by
2309 * congested BDIs but as pressure is relieved,
2310 * spectulatively avoid congestion waits
2311 */
2312 zone_clear_flag(zone, ZONE_CONGESTED);
Minchan Kim45973d72010-03-05 13:41:45 -08002313 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002314
Linus Torvalds1da177e2005-04-16 15:20:36 -07002315 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002316 if (all_zones_ok)
2317 break; /* kswapd: all done */
2318 /*
2319 * OK, kswapd is getting into trouble. Take a nap, then take
2320 * another pass across the zones.
2321 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002322 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2323 if (has_under_min_watermark_zone)
2324 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2325 else
2326 congestion_wait(BLK_RW_ASYNC, HZ/10);
2327 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002328
2329 /*
2330 * We do this so kswapd doesn't build up large priorities for
2331 * example when it is freeing in parallel with allocators. It
2332 * matches the direct reclaim path behaviour in terms of impact
2333 * on zone->*_priority.
2334 */
Rik van Riela79311c2009-01-06 14:40:01 -08002335 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002336 break;
2337 }
2338out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002339 if (!all_zones_ok) {
2340 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002341
2342 try_to_freeze();
2343
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002344 /*
2345 * Fragmentation may mean that the system cannot be
2346 * rebalanced for high-order allocations in all zones.
2347 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2348 * it means the zones have been fully scanned and are still
2349 * not balanced. For high-order allocations, there is
2350 * little point trying all over again as kswapd may
2351 * infinite loop.
2352 *
2353 * Instead, recheck all watermarks at order-0 as they
2354 * are the most important. If watermarks are ok, kswapd will go
2355 * back to sleep. High-order users can still perform direct
2356 * reclaim if they wish.
2357 */
2358 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2359 order = sc.order = 0;
2360
Linus Torvalds1da177e2005-04-16 15:20:36 -07002361 goto loop_again;
2362 }
2363
Rik van Riela79311c2009-01-06 14:40:01 -08002364 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002365}
2366
2367/*
2368 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002369 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002370 *
2371 * This basically trickles out pages so that we have _some_
2372 * free memory available even if there is no other activity
2373 * that frees anything up. This is needed for things like routing
2374 * etc, where we otherwise might have all activity going on in
2375 * asynchronous contexts that cannot page things out.
2376 *
2377 * If there are applications that are active memory-allocators
2378 * (most normal use), this basically shouldn't matter.
2379 */
2380static int kswapd(void *p)
2381{
2382 unsigned long order;
2383 pg_data_t *pgdat = (pg_data_t*)p;
2384 struct task_struct *tsk = current;
2385 DEFINE_WAIT(wait);
2386 struct reclaim_state reclaim_state = {
2387 .reclaimed_slab = 0,
2388 };
Rusty Russella70f7302009-03-13 14:49:46 +10302389 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002390
Nick Piggincf40bd12009-01-21 08:12:39 +01002391 lockdep_set_current_reclaim_state(GFP_KERNEL);
2392
Rusty Russell174596a2009-01-01 10:12:29 +10302393 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002394 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395 current->reclaim_state = &reclaim_state;
2396
2397 /*
2398 * Tell the memory management that we're a "memory allocator",
2399 * and that if we need more memory we should get access to it
2400 * regardless (see "__alloc_pages()"). "kswapd" should
2401 * never get caught in the normal page freeing logic.
2402 *
2403 * (Kswapd normally doesn't need memory anyway, but sometimes
2404 * you need a small amount of memory in order to be able to
2405 * page out something else, and this flag essentially protects
2406 * us from recursively trying to free more memory as we're
2407 * trying to free the first piece of memory in the first place).
2408 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002409 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002410 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411
2412 order = 0;
2413 for ( ; ; ) {
2414 unsigned long new_order;
David Rientjes8fe23e02009-12-14 17:58:33 -08002415 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002416
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2418 new_order = pgdat->kswapd_max_order;
2419 pgdat->kswapd_max_order = 0;
2420 if (order < new_order) {
2421 /*
2422 * Don't sleep if someone wants a larger 'order'
2423 * allocation
2424 */
2425 order = new_order;
2426 } else {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002427 if (!freezing(current) && !kthread_should_stop()) {
2428 long remaining = 0;
2429
2430 /* Try to sleep for a short interval */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002431 if (!sleeping_prematurely(pgdat, order, remaining)) {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002432 remaining = schedule_timeout(HZ/10);
2433 finish_wait(&pgdat->kswapd_wait, &wait);
2434 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2435 }
2436
2437 /*
2438 * After a short sleep, check if it was a
2439 * premature sleep. If not, then go fully
2440 * to sleep until explicitly woken up
2441 */
Mel Gorman33906bc2010-08-09 17:19:16 -07002442 if (!sleeping_prematurely(pgdat, order, remaining)) {
2443 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002444 schedule();
Mel Gorman33906bc2010-08-09 17:19:16 -07002445 } else {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002446 if (remaining)
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002447 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002448 else
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002449 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002450 }
2451 }
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002452
Linus Torvalds1da177e2005-04-16 15:20:36 -07002453 order = pgdat->kswapd_max_order;
2454 }
2455 finish_wait(&pgdat->kswapd_wait, &wait);
2456
David Rientjes8fe23e02009-12-14 17:58:33 -08002457 ret = try_to_freeze();
2458 if (kthread_should_stop())
2459 break;
2460
2461 /*
2462 * We can speed up thawing tasks if we don't call balance_pgdat
2463 * after returning from the refrigerator
2464 */
Mel Gorman33906bc2010-08-09 17:19:16 -07002465 if (!ret) {
2466 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002467 balance_pgdat(pgdat, order);
Mel Gorman33906bc2010-08-09 17:19:16 -07002468 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002469 }
2470 return 0;
2471}
2472
2473/*
2474 * A zone is low on free memory, so wake its kswapd task to service it.
2475 */
2476void wakeup_kswapd(struct zone *zone, int order)
2477{
2478 pg_data_t *pgdat;
2479
Con Kolivasf3fe6512006-01-06 00:11:15 -08002480 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002481 return;
2482
2483 pgdat = zone->zone_pgdat;
Mel Gorman41858962009-06-16 15:32:12 -07002484 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485 return;
2486 if (pgdat->kswapd_max_order < order)
2487 pgdat->kswapd_max_order = order;
Mel Gorman33906bc2010-08-09 17:19:16 -07002488 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Paul Jackson02a0e532006-12-13 00:34:25 -08002489 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002490 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002491 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002493 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002494}
2495
Wu Fengguangadea02a2009-09-21 17:01:42 -07002496/*
2497 * The reclaimable count would be mostly accurate.
2498 * The less reclaimable pages may be
2499 * - mlocked pages, which will be moved to unevictable list when encountered
2500 * - mapped pages, which may require several travels to be reclaimed
2501 * - dirty pages, which is not "instantly" reclaimable
2502 */
2503unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002504{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002505 int nr;
2506
2507 nr = global_page_state(NR_ACTIVE_FILE) +
2508 global_page_state(NR_INACTIVE_FILE);
2509
2510 if (nr_swap_pages > 0)
2511 nr += global_page_state(NR_ACTIVE_ANON) +
2512 global_page_state(NR_INACTIVE_ANON);
2513
2514 return nr;
2515}
2516
2517unsigned long zone_reclaimable_pages(struct zone *zone)
2518{
2519 int nr;
2520
2521 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2522 zone_page_state(zone, NR_INACTIVE_FILE);
2523
2524 if (nr_swap_pages > 0)
2525 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2526 zone_page_state(zone, NR_INACTIVE_ANON);
2527
2528 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002529}
2530
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002531#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002532/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002533 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002534 * freed pages.
2535 *
2536 * Rather than trying to age LRUs the aim is to preserve the overall
2537 * LRU order by reclaiming preferentially
2538 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002539 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002540unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002541{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002542 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002543 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002544 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2545 .may_swap = 1,
2546 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002547 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002548 .nr_to_reclaim = nr_to_reclaim,
2549 .hibernation_mode = 1,
2550 .swappiness = vm_swappiness,
2551 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552 };
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002553 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2554 struct task_struct *p = current;
2555 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002557 p->flags |= PF_MEMALLOC;
2558 lockdep_set_current_reclaim_state(sc.gfp_mask);
2559 reclaim_state.reclaimed_slab = 0;
2560 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002561
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002562 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002563
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002564 p->reclaim_state = NULL;
2565 lockdep_clear_current_reclaim_state();
2566 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002567
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002568 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002569}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002570#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571
Linus Torvalds1da177e2005-04-16 15:20:36 -07002572/* It's optimal to keep kswapds on the same CPUs as their memory, but
2573 not required for correctness. So if the last cpu in a node goes
2574 away, we get changed to run anywhere: as the first one comes back,
2575 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002576static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002577 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002578{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002579 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002580
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002581 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002582 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002583 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302584 const struct cpumask *mask;
2585
2586 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002587
Rusty Russell3e597942009-01-01 10:12:24 +10302588 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002589 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002590 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002591 }
2592 }
2593 return NOTIFY_OK;
2594}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002595
Yasunori Goto3218ae12006-06-27 02:53:33 -07002596/*
2597 * This kswapd start function will be called by init and node-hot-add.
2598 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2599 */
2600int kswapd_run(int nid)
2601{
2602 pg_data_t *pgdat = NODE_DATA(nid);
2603 int ret = 0;
2604
2605 if (pgdat->kswapd)
2606 return 0;
2607
2608 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2609 if (IS_ERR(pgdat->kswapd)) {
2610 /* failure at boot is fatal */
2611 BUG_ON(system_state == SYSTEM_BOOTING);
2612 printk("Failed to start kswapd on node %d\n",nid);
2613 ret = -1;
2614 }
2615 return ret;
2616}
2617
David Rientjes8fe23e02009-12-14 17:58:33 -08002618/*
2619 * Called by memory hotplug when all memory in a node is offlined.
2620 */
2621void kswapd_stop(int nid)
2622{
2623 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2624
2625 if (kswapd)
2626 kthread_stop(kswapd);
2627}
2628
Linus Torvalds1da177e2005-04-16 15:20:36 -07002629static int __init kswapd_init(void)
2630{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002631 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002632
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002634 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002635 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002636 hotcpu_notifier(cpu_callback, 0);
2637 return 0;
2638}
2639
2640module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002641
2642#ifdef CONFIG_NUMA
2643/*
2644 * Zone reclaim mode
2645 *
2646 * If non-zero call zone_reclaim when the number of free pages falls below
2647 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002648 */
2649int zone_reclaim_mode __read_mostly;
2650
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002651#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002652#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002653#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2654#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2655
Christoph Lameter9eeff232006-01-18 17:42:31 -08002656/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002657 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2658 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2659 * a zone.
2660 */
2661#define ZONE_RECLAIM_PRIORITY 4
2662
Christoph Lameter9eeff232006-01-18 17:42:31 -08002663/*
Christoph Lameter96146342006-07-03 00:24:13 -07002664 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2665 * occur.
2666 */
2667int sysctl_min_unmapped_ratio = 1;
2668
2669/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002670 * If the number of slab pages in a zone grows beyond this percentage then
2671 * slab reclaim needs to occur.
2672 */
2673int sysctl_min_slab_ratio = 5;
2674
Mel Gorman90afa5d2009-06-16 15:33:20 -07002675static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2676{
2677 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2678 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2679 zone_page_state(zone, NR_ACTIVE_FILE);
2680
2681 /*
2682 * It's possible for there to be more file mapped pages than
2683 * accounted for by the pages on the file LRU lists because
2684 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2685 */
2686 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2687}
2688
2689/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2690static long zone_pagecache_reclaimable(struct zone *zone)
2691{
2692 long nr_pagecache_reclaimable;
2693 long delta = 0;
2694
2695 /*
2696 * If RECLAIM_SWAP is set, then all file pages are considered
2697 * potentially reclaimable. Otherwise, we have to worry about
2698 * pages like swapcache and zone_unmapped_file_pages() provides
2699 * a better estimate
2700 */
2701 if (zone_reclaim_mode & RECLAIM_SWAP)
2702 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2703 else
2704 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2705
2706 /* If we can't clean pages, remove dirty pages from consideration */
2707 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2708 delta += zone_page_state(zone, NR_FILE_DIRTY);
2709
2710 /* Watch for any possible underflows due to delta */
2711 if (unlikely(delta > nr_pagecache_reclaimable))
2712 delta = nr_pagecache_reclaimable;
2713
2714 return nr_pagecache_reclaimable - delta;
2715}
2716
Christoph Lameter0ff38492006-09-25 23:31:52 -07002717/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002718 * Try to free up some pages from this zone through reclaim.
2719 */
Andrew Morton179e9632006-03-22 00:08:18 -08002720static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002721{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002722 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002723 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002724 struct task_struct *p = current;
2725 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002726 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002727 struct scan_control sc = {
2728 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002729 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002730 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002731 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2732 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002733 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002734 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002735 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08002736 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07002737 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002738
Christoph Lameter9eeff232006-01-18 17:42:31 -08002739 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002740 /*
2741 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2742 * and we also need to be able to write out pages for RECLAIM_WRITE
2743 * and RECLAIM_SWAP.
2744 */
2745 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002746 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08002747 reclaim_state.reclaimed_slab = 0;
2748 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002749
Mel Gorman90afa5d2009-06-16 15:33:20 -07002750 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07002751 /*
2752 * Free memory by calling shrink zone with increasing
2753 * priorities until we have enough memory freed.
2754 */
2755 priority = ZONE_RECLAIM_PRIORITY;
2756 do {
Rik van Riela79311c2009-01-06 14:40:01 -08002757 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002758 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002759 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002760 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002761
KOSAKI Motohiro15748042010-08-09 17:19:50 -07002762 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2763 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002764 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002765 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002766 * many pages were freed in this zone. So we take the current
2767 * number of slab pages and shake the slab until it is reduced
2768 * by the same nr_pages that we used for reclaiming unmapped
2769 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002770 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002771 * Note that shrink_slab will free memory on all zones and may
2772 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002773 */
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07002774 for (;;) {
2775 unsigned long lru_pages = zone_reclaimable_pages(zone);
2776
2777 /* No reclaimable slab or very low memory pressure */
2778 if (!shrink_slab(sc.nr_scanned, gfp_mask, lru_pages))
2779 break;
2780
2781 /* Freed enough memory */
2782 nr_slab_pages1 = zone_page_state(zone,
2783 NR_SLAB_RECLAIMABLE);
2784 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
2785 break;
2786 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07002787
2788 /*
2789 * Update nr_reclaimed by the number of slab pages we
2790 * reclaimed from this zone.
2791 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07002792 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2793 if (nr_slab_pages1 < nr_slab_pages0)
2794 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002795 }
2796
Christoph Lameter9eeff232006-01-18 17:42:31 -08002797 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002798 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002799 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08002800 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002801}
Andrew Morton179e9632006-03-22 00:08:18 -08002802
2803int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2804{
Andrew Morton179e9632006-03-22 00:08:18 -08002805 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002806 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002807
2808 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002809 * Zone reclaim reclaims unmapped file backed pages and
2810 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002811 *
Christoph Lameter96146342006-07-03 00:24:13 -07002812 * A small portion of unmapped file backed pages is needed for
2813 * file I/O otherwise pages read by file I/O will be immediately
2814 * thrown out if the zone is overallocated. So we do not reclaim
2815 * if less than a specified percentage of the zone is used by
2816 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002817 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07002818 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2819 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002820 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08002821
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002822 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002823 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07002824
Andrew Morton179e9632006-03-22 00:08:18 -08002825 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002826 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002827 */
David Rientjesd773ed62007-10-16 23:26:01 -07002828 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002829 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08002830
2831 /*
2832 * Only run zone reclaim on the local zone or on zones that do not
2833 * have associated processors. This will favor the local processor
2834 * over remote processors and spread off node memory allocations
2835 * as wide as possible.
2836 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002837 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002838 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07002839 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07002840
2841 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002842 return ZONE_RECLAIM_NOSCAN;
2843
David Rientjesd773ed62007-10-16 23:26:01 -07002844 ret = __zone_reclaim(zone, gfp_mask, order);
2845 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2846
Mel Gorman24cf725182009-06-16 15:33:23 -07002847 if (!ret)
2848 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2849
David Rientjesd773ed62007-10-16 23:26:01 -07002850 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002851}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002852#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002853
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002854/*
2855 * page_evictable - test whether a page is evictable
2856 * @page: the page to test
2857 * @vma: the VMA in which the page is or will be mapped, may be NULL
2858 *
2859 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07002860 * lists vs unevictable list. The vma argument is !NULL when called from the
2861 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002862 *
2863 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002864 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07002865 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002866 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002867 */
2868int page_evictable(struct page *page, struct vm_area_struct *vma)
2869{
2870
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002871 if (mapping_unevictable(page_mapping(page)))
2872 return 0;
2873
Nick Pigginb291f002008-10-18 20:26:44 -07002874 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2875 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002876
2877 return 1;
2878}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002879
2880/**
2881 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2882 * @page: page to check evictability and move to appropriate lru list
2883 * @zone: zone page is in
2884 *
2885 * Checks a page for evictability and moves the page to the appropriate
2886 * zone lru list.
2887 *
2888 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2889 * have PageUnevictable set.
2890 */
2891static void check_move_unevictable_page(struct page *page, struct zone *zone)
2892{
2893 VM_BUG_ON(PageActive(page));
2894
2895retry:
2896 ClearPageUnevictable(page);
2897 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07002898 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002899
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002900 __dec_zone_state(zone, NR_UNEVICTABLE);
2901 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002902 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002903 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2904 __count_vm_event(UNEVICTABLE_PGRESCUED);
2905 } else {
2906 /*
2907 * rotate unevictable list
2908 */
2909 SetPageUnevictable(page);
2910 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002911 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002912 if (page_evictable(page, NULL))
2913 goto retry;
2914 }
2915}
2916
2917/**
2918 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2919 * @mapping: struct address_space to scan for evictable pages
2920 *
2921 * Scan all pages in mapping. Check unevictable pages for
2922 * evictability and move them to the appropriate zone lru list.
2923 */
2924void scan_mapping_unevictable_pages(struct address_space *mapping)
2925{
2926 pgoff_t next = 0;
2927 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2928 PAGE_CACHE_SHIFT;
2929 struct zone *zone;
2930 struct pagevec pvec;
2931
2932 if (mapping->nrpages == 0)
2933 return;
2934
2935 pagevec_init(&pvec, 0);
2936 while (next < end &&
2937 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2938 int i;
2939 int pg_scanned = 0;
2940
2941 zone = NULL;
2942
2943 for (i = 0; i < pagevec_count(&pvec); i++) {
2944 struct page *page = pvec.pages[i];
2945 pgoff_t page_index = page->index;
2946 struct zone *pagezone = page_zone(page);
2947
2948 pg_scanned++;
2949 if (page_index > next)
2950 next = page_index;
2951 next++;
2952
2953 if (pagezone != zone) {
2954 if (zone)
2955 spin_unlock_irq(&zone->lru_lock);
2956 zone = pagezone;
2957 spin_lock_irq(&zone->lru_lock);
2958 }
2959
2960 if (PageLRU(page) && PageUnevictable(page))
2961 check_move_unevictable_page(page, zone);
2962 }
2963 if (zone)
2964 spin_unlock_irq(&zone->lru_lock);
2965 pagevec_release(&pvec);
2966
2967 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2968 }
2969
2970}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002971
2972/**
2973 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2974 * @zone - zone of which to scan the unevictable list
2975 *
2976 * Scan @zone's unevictable LRU lists to check for pages that have become
2977 * evictable. Move those that have to @zone's inactive list where they
2978 * become candidates for reclaim, unless shrink_inactive_zone() decides
2979 * to reactivate them. Pages that are still unevictable are rotated
2980 * back onto @zone's unevictable list.
2981 */
2982#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08002983static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002984{
2985 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2986 unsigned long scan;
2987 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2988
2989 while (nr_to_scan > 0) {
2990 unsigned long batch_size = min(nr_to_scan,
2991 SCAN_UNEVICTABLE_BATCH_SIZE);
2992
2993 spin_lock_irq(&zone->lru_lock);
2994 for (scan = 0; scan < batch_size; scan++) {
2995 struct page *page = lru_to_page(l_unevictable);
2996
2997 if (!trylock_page(page))
2998 continue;
2999
3000 prefetchw_prev_lru_page(page, l_unevictable, flags);
3001
3002 if (likely(PageLRU(page) && PageUnevictable(page)))
3003 check_move_unevictable_page(page, zone);
3004
3005 unlock_page(page);
3006 }
3007 spin_unlock_irq(&zone->lru_lock);
3008
3009 nr_to_scan -= batch_size;
3010 }
3011}
3012
3013
3014/**
3015 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
3016 *
3017 * A really big hammer: scan all zones' unevictable LRU lists to check for
3018 * pages that have become evictable. Move those back to the zones'
3019 * inactive list where they become candidates for reclaim.
3020 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
3021 * and we add swap to the system. As such, it runs in the context of a task
3022 * that has possibly/probably made some previously unevictable pages
3023 * evictable.
3024 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08003025static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003026{
3027 struct zone *zone;
3028
3029 for_each_zone(zone) {
3030 scan_zone_unevictable_pages(zone);
3031 }
3032}
3033
3034/*
3035 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3036 * all nodes' unevictable lists for evictable pages
3037 */
3038unsigned long scan_unevictable_pages;
3039
3040int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003041 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003042 size_t *length, loff_t *ppos)
3043{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003044 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003045
3046 if (write && *(unsigned long *)table->data)
3047 scan_all_zones_unevictable_pages();
3048
3049 scan_unevictable_pages = 0;
3050 return 0;
3051}
3052
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003053#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003054/*
3055 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3056 * a specified node's per zone unevictable lists for evictable pages.
3057 */
3058
3059static ssize_t read_scan_unevictable_node(struct sys_device *dev,
3060 struct sysdev_attribute *attr,
3061 char *buf)
3062{
3063 return sprintf(buf, "0\n"); /* always zero; should fit... */
3064}
3065
3066static ssize_t write_scan_unevictable_node(struct sys_device *dev,
3067 struct sysdev_attribute *attr,
3068 const char *buf, size_t count)
3069{
3070 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
3071 struct zone *zone;
3072 unsigned long res;
3073 unsigned long req = strict_strtoul(buf, 10, &res);
3074
3075 if (!req)
3076 return 1; /* zero is no-op */
3077
3078 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
3079 if (!populated_zone(zone))
3080 continue;
3081 scan_zone_unevictable_pages(zone);
3082 }
3083 return 1;
3084}
3085
3086
3087static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
3088 read_scan_unevictable_node,
3089 write_scan_unevictable_node);
3090
3091int scan_unevictable_register_node(struct node *node)
3092{
3093 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
3094}
3095
3096void scan_unevictable_unregister_node(struct node *node)
3097{
3098 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
3099}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003100#endif