blob: 06e72693b4587a6181b002dec4f14ac5a7d4ba06 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
16#include <linux/slab.h>
17#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070022#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/file.h>
24#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28#include <linux/mm_inline.h>
29#include <linux/pagevec.h>
30#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
35#include <linux/notifier.h>
36#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080037#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070038#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080039#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080040#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070041#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070042#include <linux/sysctl.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44#include <asm/tlbflush.h>
45#include <asm/div64.h>
46
47#include <linux/swapops.h>
48
Nick Piggin0f8053a2006-03-22 00:08:33 -080049#include "internal.h"
50
Linus Torvalds1da177e2005-04-16 15:20:36 -070051struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 /* Incremented by the number of inactive pages that were scanned */
53 unsigned long nr_scanned;
54
Rik van Riela79311c2009-01-06 14:40:01 -080055 /* Number of pages freed so far during a call to shrink_zones() */
56 unsigned long nr_reclaimed;
57
Linus Torvalds1da177e2005-04-16 15:20:36 -070058 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040059 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070060
61 int may_writepage;
62
Johannes Weinera6dc60f2009-03-31 15:19:30 -070063 /* Can mapped pages be reclaimed? */
64 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080065
Linus Torvalds1da177e2005-04-16 15:20:36 -070066 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
67 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
68 * In this context, it doesn't matter that we scan the
69 * whole list at once. */
70 int swap_cluster_max;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070071
72 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070073
74 int all_unreclaimable;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070075
76 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080077
78 /* Which cgroup do we reclaim from */
79 struct mem_cgroup *mem_cgroup;
80
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -070081 /*
82 * Nodemask of nodes allowed by the caller. If NULL, all nodes
83 * are scanned.
84 */
85 nodemask_t *nodemask;
86
Balbir Singh66e17072008-02-07 00:13:56 -080087 /* Pluggable isolate pages callback */
88 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
89 unsigned long *scanned, int order, int mode,
90 struct zone *z, struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -070091 int active, int file);
Linus Torvalds1da177e2005-04-16 15:20:36 -070092};
93
Linus Torvalds1da177e2005-04-16 15:20:36 -070094#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
95
96#ifdef ARCH_HAS_PREFETCH
97#define prefetch_prev_lru_page(_page, _base, _field) \
98 do { \
99 if ((_page)->lru.prev != _base) { \
100 struct page *prev; \
101 \
102 prev = lru_to_page(&(_page->lru)); \
103 prefetch(&prev->_field); \
104 } \
105 } while (0)
106#else
107#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
108#endif
109
110#ifdef ARCH_HAS_PREFETCHW
111#define prefetchw_prev_lru_page(_page, _base, _field) \
112 do { \
113 if ((_page)->lru.prev != _base) { \
114 struct page *prev; \
115 \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetchw(&prev->_field); \
118 } \
119 } while (0)
120#else
121#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
122#endif
123
124/*
125 * From 0 .. 100. Higher means more swappy.
126 */
127int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700128long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129
130static LIST_HEAD(shrinker_list);
131static DECLARE_RWSEM(shrinker_rwsem);
132
Balbir Singh00f0b822008-03-04 14:28:39 -0800133#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800134#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800135#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800136#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800137#endif
138
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800139static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
140 struct scan_control *sc)
141{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800142 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800143 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
144
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800145 return &zone->reclaim_stat;
146}
147
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800148static unsigned long zone_nr_pages(struct zone *zone, struct scan_control *sc,
149 enum lru_list lru)
150{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800151 if (!scanning_global_lru(sc))
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800152 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
153
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800154 return zone_page_state(zone, NR_LRU_BASE + lru);
155}
156
157
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158/*
159 * Add a shrinker callback to be called from the vm
160 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700161void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700163 shrinker->nr = 0;
164 down_write(&shrinker_rwsem);
165 list_add_tail(&shrinker->list, &shrinker_list);
166 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700168EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169
170/*
171 * Remove one
172 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700173void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174{
175 down_write(&shrinker_rwsem);
176 list_del(&shrinker->list);
177 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700179EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180
181#define SHRINK_BATCH 128
182/*
183 * Call the shrink functions to age shrinkable caches
184 *
185 * Here we assume it costs one seek to replace a lru page and that it also
186 * takes a seek to recreate a cache object. With this in mind we age equal
187 * percentages of the lru and ageable caches. This should balance the seeks
188 * generated by these structures.
189 *
Simon Arlott183ff222007-10-20 01:27:18 +0200190 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 * slab to avoid swapping.
192 *
193 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
194 *
195 * `lru_pages' represents the number of on-LRU pages in all the zones which
196 * are eligible for the caller's allocation attempt. It is used for balancing
197 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700198 *
199 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 */
Andrew Morton69e05942006-03-22 00:08:19 -0800201unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
202 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203{
204 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800205 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206
207 if (scanned == 0)
208 scanned = SWAP_CLUSTER_MAX;
209
210 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700211 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212
213 list_for_each_entry(shrinker, &shrinker_list, list) {
214 unsigned long long delta;
215 unsigned long total_scan;
Rusty Russell8e1f9362007-07-17 04:03:17 -0700216 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217
218 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800219 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700220 do_div(delta, lru_pages + 1);
221 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800222 if (shrinker->nr < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700223 printk(KERN_ERR "shrink_slab: %pF negative objects to "
224 "delete nr=%ld\n",
225 shrinker->shrink, shrinker->nr);
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800226 shrinker->nr = max_pass;
227 }
228
229 /*
230 * Avoid risking looping forever due to too large nr value:
231 * never try to free more than twice the estimate number of
232 * freeable entries.
233 */
234 if (shrinker->nr > max_pass * 2)
235 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700236
237 total_scan = shrinker->nr;
238 shrinker->nr = 0;
239
240 while (total_scan >= SHRINK_BATCH) {
241 long this_scan = SHRINK_BATCH;
242 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700243 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700244
Rusty Russell8e1f9362007-07-17 04:03:17 -0700245 nr_before = (*shrinker->shrink)(0, gfp_mask);
246 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700247 if (shrink_ret == -1)
248 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700249 if (shrink_ret < nr_before)
250 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700251 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700252 total_scan -= this_scan;
253
254 cond_resched();
255 }
256
257 shrinker->nr += total_scan;
258 }
259 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700260 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261}
262
263/* Called without lock on whether page is mapped, so answer is unstable */
264static inline int page_mapping_inuse(struct page *page)
265{
266 struct address_space *mapping;
267
268 /* Page is in somebody's page tables. */
269 if (page_mapped(page))
270 return 1;
271
272 /* Be more reluctant to reclaim swapcache than pagecache */
273 if (PageSwapCache(page))
274 return 1;
275
276 mapping = page_mapping(page);
277 if (!mapping)
278 return 0;
279
280 /* File is mmap'd by somebody? */
281 return mapping_mapped(mapping);
282}
283
284static inline int is_page_cache_freeable(struct page *page)
285{
286 return page_count(page) - !!PagePrivate(page) == 2;
287}
288
289static int may_write_to_queue(struct backing_dev_info *bdi)
290{
Christoph Lameter930d9152006-01-08 01:00:47 -0800291 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700292 return 1;
293 if (!bdi_write_congested(bdi))
294 return 1;
295 if (bdi == current->backing_dev_info)
296 return 1;
297 return 0;
298}
299
300/*
301 * We detected a synchronous write error writing a page out. Probably
302 * -ENOSPC. We need to propagate that into the address_space for a subsequent
303 * fsync(), msync() or close().
304 *
305 * The tricky part is that after writepage we cannot touch the mapping: nothing
306 * prevents it from being freed up. But we have a ref on the page and once
307 * that page is locked, the mapping is pinned.
308 *
309 * We're allowed to run sleeping lock_page() here because we know the caller has
310 * __GFP_FS.
311 */
312static void handle_write_error(struct address_space *mapping,
313 struct page *page, int error)
314{
315 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700316 if (page_mapping(page) == mapping)
317 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700318 unlock_page(page);
319}
320
Andy Whitcroftc661b072007-08-22 14:01:26 -0700321/* Request for sync pageout. */
322enum pageout_io {
323 PAGEOUT_IO_ASYNC,
324 PAGEOUT_IO_SYNC,
325};
326
Christoph Lameter04e62a22006-06-23 02:03:38 -0700327/* possible outcome of pageout() */
328typedef enum {
329 /* failed to write page out, page is locked */
330 PAGE_KEEP,
331 /* move page to the active list, page is locked */
332 PAGE_ACTIVATE,
333 /* page has been sent to the disk successfully, page is unlocked */
334 PAGE_SUCCESS,
335 /* page is clean and locked */
336 PAGE_CLEAN,
337} pageout_t;
338
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339/*
Andrew Morton1742f192006-03-22 00:08:21 -0800340 * pageout is called by shrink_page_list() for each dirty page.
341 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700343static pageout_t pageout(struct page *page, struct address_space *mapping,
344 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700345{
346 /*
347 * If the page is dirty, only perform writeback if that write
348 * will be non-blocking. To prevent this allocation from being
349 * stalled by pagecache activity. But note that there may be
350 * stalls if we need to run get_block(). We could test
351 * PagePrivate for that.
352 *
353 * If this process is currently in generic_file_write() against
354 * this page's queue, we can perform writeback even if that
355 * will block.
356 *
357 * If the page is swapcache, write it back even if that would
358 * block, for some throttling. This happens by accident, because
359 * swap_backing_dev_info is bust: it doesn't reflect the
360 * congestion state of the swapdevs. Easy to fix, if needed.
361 * See swapfile.c:page_queue_congested().
362 */
363 if (!is_page_cache_freeable(page))
364 return PAGE_KEEP;
365 if (!mapping) {
366 /*
367 * Some data journaling orphaned pages can have
368 * page->mapping == NULL while being dirty with clean buffers.
369 */
akpm@osdl.org323aca62005-04-16 15:24:06 -0700370 if (PagePrivate(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700371 if (try_to_free_buffers(page)) {
372 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700373 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374 return PAGE_CLEAN;
375 }
376 }
377 return PAGE_KEEP;
378 }
379 if (mapping->a_ops->writepage == NULL)
380 return PAGE_ACTIVATE;
381 if (!may_write_to_queue(mapping->backing_dev_info))
382 return PAGE_KEEP;
383
384 if (clear_page_dirty_for_io(page)) {
385 int res;
386 struct writeback_control wbc = {
387 .sync_mode = WB_SYNC_NONE,
388 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700389 .range_start = 0,
390 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700391 .nonblocking = 1,
392 .for_reclaim = 1,
393 };
394
395 SetPageReclaim(page);
396 res = mapping->a_ops->writepage(page, &wbc);
397 if (res < 0)
398 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800399 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400 ClearPageReclaim(page);
401 return PAGE_ACTIVATE;
402 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700403
404 /*
405 * Wait on writeback if requested to. This happens when
406 * direct reclaiming a large contiguous area and the
407 * first attempt to free a range of pages fails.
408 */
409 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
410 wait_on_page_writeback(page);
411
Linus Torvalds1da177e2005-04-16 15:20:36 -0700412 if (!PageWriteback(page)) {
413 /* synchronous write or broken a_ops? */
414 ClearPageReclaim(page);
415 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700416 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700417 return PAGE_SUCCESS;
418 }
419
420 return PAGE_CLEAN;
421}
422
Andrew Mortona649fd92006-10-17 00:09:36 -0700423/*
Nick Piggine2867812008-07-25 19:45:30 -0700424 * Same as remove_mapping, but if the page is removed from the mapping, it
425 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700426 */
Nick Piggine2867812008-07-25 19:45:30 -0700427static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800428{
Nick Piggin28e4d962006-09-25 23:31:23 -0700429 BUG_ON(!PageLocked(page));
430 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800431
Nick Piggin19fd6232008-07-25 19:45:32 -0700432 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800433 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700434 * The non racy check for a busy page.
435 *
436 * Must be careful with the order of the tests. When someone has
437 * a ref to the page, it may be possible that they dirty it then
438 * drop the reference. So if PageDirty is tested before page_count
439 * here, then the following race may occur:
440 *
441 * get_user_pages(&page);
442 * [user mapping goes away]
443 * write_to(page);
444 * !PageDirty(page) [good]
445 * SetPageDirty(page);
446 * put_page(page);
447 * !page_count(page) [good, discard it]
448 *
449 * [oops, our write_to data is lost]
450 *
451 * Reversing the order of the tests ensures such a situation cannot
452 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
453 * load is not satisfied before that of page->_count.
454 *
455 * Note that if SetPageDirty is always performed via set_page_dirty,
456 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800457 */
Nick Piggine2867812008-07-25 19:45:30 -0700458 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800459 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700460 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
461 if (unlikely(PageDirty(page))) {
462 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800463 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700464 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800465
466 if (PageSwapCache(page)) {
467 swp_entry_t swap = { .val = page_private(page) };
468 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700469 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800470 swap_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700471 } else {
472 __remove_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700473 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800474 }
475
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800476 return 1;
477
478cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700479 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800480 return 0;
481}
482
Linus Torvalds1da177e2005-04-16 15:20:36 -0700483/*
Nick Piggine2867812008-07-25 19:45:30 -0700484 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
485 * someone else has a ref on the page, abort and return 0. If it was
486 * successfully detached, return 1. Assumes the caller has a single ref on
487 * this page.
488 */
489int remove_mapping(struct address_space *mapping, struct page *page)
490{
491 if (__remove_mapping(mapping, page)) {
492 /*
493 * Unfreezing the refcount with 1 rather than 2 effectively
494 * drops the pagecache ref for us without requiring another
495 * atomic operation.
496 */
497 page_unfreeze_refs(page, 1);
498 return 1;
499 }
500 return 0;
501}
502
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700503/**
504 * putback_lru_page - put previously isolated page onto appropriate LRU list
505 * @page: page to be put back to appropriate lru list
506 *
507 * Add previously isolated @page to appropriate LRU list.
508 * Page may still be unevictable for other reasons.
509 *
510 * lru_lock must not be held, interrupts must be enabled.
511 */
512#ifdef CONFIG_UNEVICTABLE_LRU
513void putback_lru_page(struct page *page)
514{
515 int lru;
516 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700517 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700518
519 VM_BUG_ON(PageLRU(page));
520
521redo:
522 ClearPageUnevictable(page);
523
524 if (page_evictable(page, NULL)) {
525 /*
526 * For evictable pages, we can use the cache.
527 * In event of a race, worst case is we end up with an
528 * unevictable page on [in]active list.
529 * We know how to handle that.
530 */
531 lru = active + page_is_file_cache(page);
532 lru_cache_add_lru(page, lru);
533 } else {
534 /*
535 * Put unevictable pages directly on zone's unevictable
536 * list.
537 */
538 lru = LRU_UNEVICTABLE;
539 add_page_to_unevictable_list(page);
540 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700541
542 /*
543 * page's status can change while we move it among lru. If an evictable
544 * page is on unevictable list, it never be freed. To avoid that,
545 * check after we added it to the list, again.
546 */
547 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
548 if (!isolate_lru_page(page)) {
549 put_page(page);
550 goto redo;
551 }
552 /* This means someone else dropped this page from LRU
553 * So, it will be freed or putback to LRU again. There is
554 * nothing to do here.
555 */
556 }
557
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700558 if (was_unevictable && lru != LRU_UNEVICTABLE)
559 count_vm_event(UNEVICTABLE_PGRESCUED);
560 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
561 count_vm_event(UNEVICTABLE_PGCULLED);
562
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700563 put_page(page); /* drop ref from isolate */
564}
565
566#else /* CONFIG_UNEVICTABLE_LRU */
567
568void putback_lru_page(struct page *page)
569{
570 int lru;
571 VM_BUG_ON(PageLRU(page));
572
573 lru = !!TestClearPageActive(page) + page_is_file_cache(page);
574 lru_cache_add_lru(page, lru);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700575 put_page(page);
576}
577#endif /* CONFIG_UNEVICTABLE_LRU */
578
579
Nick Piggine2867812008-07-25 19:45:30 -0700580/*
Andrew Morton1742f192006-03-22 00:08:21 -0800581 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700582 */
Andrew Morton1742f192006-03-22 00:08:21 -0800583static unsigned long shrink_page_list(struct list_head *page_list,
Andy Whitcroftc661b072007-08-22 14:01:26 -0700584 struct scan_control *sc,
585 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586{
587 LIST_HEAD(ret_pages);
588 struct pagevec freed_pvec;
589 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800590 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700591
592 cond_resched();
593
594 pagevec_init(&freed_pvec, 1);
595 while (!list_empty(page_list)) {
596 struct address_space *mapping;
597 struct page *page;
598 int may_enter_fs;
599 int referenced;
600
601 cond_resched();
602
603 page = lru_to_page(page_list);
604 list_del(&page->lru);
605
Nick Piggin529ae9a2008-08-02 12:01:03 +0200606 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607 goto keep;
608
Nick Piggin725d7042006-09-25 23:30:55 -0700609 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610
611 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800612
Nick Pigginb291f002008-10-18 20:26:44 -0700613 if (unlikely(!page_evictable(page, NULL)))
614 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700615
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700616 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800617 goto keep_locked;
618
Linus Torvalds1da177e2005-04-16 15:20:36 -0700619 /* Double the slab pressure for mapped and swapcache pages */
620 if (page_mapped(page) || PageSwapCache(page))
621 sc->nr_scanned++;
622
Andy Whitcroftc661b072007-08-22 14:01:26 -0700623 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
624 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
625
626 if (PageWriteback(page)) {
627 /*
628 * Synchronous reclaim is performed in two passes,
629 * first an asynchronous pass over the list to
630 * start parallel writeback, and a second synchronous
631 * pass to wait for the IO to complete. Wait here
632 * for any page for which writeback has already
633 * started.
634 */
635 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
636 wait_on_page_writeback(page);
Andrew Morton4dd4b922008-03-24 12:29:52 -0700637 else
Andy Whitcroftc661b072007-08-22 14:01:26 -0700638 goto keep_locked;
639 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640
Balbir Singhbed71612008-02-07 00:14:01 -0800641 referenced = page_referenced(page, 1, sc->mem_cgroup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642 /* In active use or really unfreeable? Activate it. */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700643 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
644 referenced && page_mapping_inuse(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700645 goto activate_locked;
646
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 /*
648 * Anonymous process memory has backing store?
649 * Try to allocate it some swap space here.
650 */
Nick Pigginb291f002008-10-18 20:26:44 -0700651 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800652 if (!(sc->gfp_mask & __GFP_IO))
653 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800654 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800656 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700657 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700658
659 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700660
661 /*
662 * The page is mapped into the page tables of one or more
663 * processes. Try to unmap it here.
664 */
665 if (page_mapped(page) && mapping) {
Christoph Lametera48d07a2006-02-01 03:05:38 -0800666 switch (try_to_unmap(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700667 case SWAP_FAIL:
668 goto activate_locked;
669 case SWAP_AGAIN:
670 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700671 case SWAP_MLOCK:
672 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700673 case SWAP_SUCCESS:
674 ; /* try to free the page below */
675 }
676 }
677
678 if (PageDirty(page)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700679 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700681 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700682 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800683 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 goto keep_locked;
685
686 /* Page is dirty, try to write it out here */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700687 switch (pageout(page, mapping, sync_writeback)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700688 case PAGE_KEEP:
689 goto keep_locked;
690 case PAGE_ACTIVATE:
691 goto activate_locked;
692 case PAGE_SUCCESS:
Andrew Morton4dd4b922008-03-24 12:29:52 -0700693 if (PageWriteback(page) || PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700694 goto keep;
695 /*
696 * A synchronous write - probably a ramdisk. Go
697 * ahead and try to reclaim the page.
698 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200699 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 goto keep;
701 if (PageDirty(page) || PageWriteback(page))
702 goto keep_locked;
703 mapping = page_mapping(page);
704 case PAGE_CLEAN:
705 ; /* try to free the page below */
706 }
707 }
708
709 /*
710 * If the page has buffers, try to free the buffer mappings
711 * associated with this page. If we succeed we try to free
712 * the page as well.
713 *
714 * We do this even if the page is PageDirty().
715 * try_to_release_page() does not perform I/O, but it is
716 * possible for a page to have PageDirty set, but it is actually
717 * clean (all its buffers are clean). This happens if the
718 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700719 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700720 * try_to_release_page() will discover that cleanness and will
721 * drop the buffers and mark the page clean - it can be freed.
722 *
723 * Rarely, pages can have buffers and no ->mapping. These are
724 * the pages which were not successfully invalidated in
725 * truncate_complete_page(). We try to drop those buffers here
726 * and if that worked, and the page is no longer mapped into
727 * process address space (page_count == 1) it can be freed.
728 * Otherwise, leave the page on the LRU so it is swappable.
729 */
730 if (PagePrivate(page)) {
731 if (!try_to_release_page(page, sc->gfp_mask))
732 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700733 if (!mapping && page_count(page) == 1) {
734 unlock_page(page);
735 if (put_page_testzero(page))
736 goto free_it;
737 else {
738 /*
739 * rare race with speculative reference.
740 * the speculative reference will free
741 * this page shortly, so we may
742 * increment nr_reclaimed here (and
743 * leave it off the LRU).
744 */
745 nr_reclaimed++;
746 continue;
747 }
748 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700749 }
750
Nick Piggine2867812008-07-25 19:45:30 -0700751 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800752 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700753
Nick Piggina978d6f2008-10-18 20:26:58 -0700754 /*
755 * At this point, we have no other references and there is
756 * no way to pick any more up (removed from LRU, removed
757 * from pagecache). Can use non-atomic bitops now (and
758 * we obviously don't have to worry about waking up a process
759 * waiting on the page lock, because there are no references.
760 */
761 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700762free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800763 nr_reclaimed++;
Nick Piggine2867812008-07-25 19:45:30 -0700764 if (!pagevec_add(&freed_pvec, page)) {
765 __pagevec_free(&freed_pvec);
766 pagevec_reinit(&freed_pvec);
767 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768 continue;
769
Nick Pigginb291f002008-10-18 20:26:44 -0700770cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800771 if (PageSwapCache(page))
772 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700773 unlock_page(page);
774 putback_lru_page(page);
775 continue;
776
Linus Torvalds1da177e2005-04-16 15:20:36 -0700777activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700778 /* Not a candidate for swapping, so reclaim swap space. */
779 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800780 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700781 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 SetPageActive(page);
783 pgactivate++;
784keep_locked:
785 unlock_page(page);
786keep:
787 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700788 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 }
790 list_splice(&ret_pages, page_list);
791 if (pagevec_count(&freed_pvec))
Nick Piggine2867812008-07-25 19:45:30 -0700792 __pagevec_free(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700793 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800794 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795}
796
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700797/* LRU Isolation modes. */
798#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
799#define ISOLATE_ACTIVE 1 /* Isolate active pages. */
800#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
801
802/*
803 * Attempt to remove the specified page from its LRU. Only take this page
804 * if it is of the appropriate PageActive status. Pages which are being
805 * freed elsewhere are also ignored.
806 *
807 * page: page to consider
808 * mode: one of the LRU isolation modes defined above
809 *
810 * returns 0 on success, -ve errno on failure.
811 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700812int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700813{
814 int ret = -EINVAL;
815
816 /* Only take pages on the LRU. */
817 if (!PageLRU(page))
818 return ret;
819
820 /*
821 * When checking the active state, we need to be sure we are
822 * dealing with comparible boolean values. Take the logical not
823 * of each.
824 */
825 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
826 return ret;
827
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700828 if (mode != ISOLATE_BOTH && (!page_is_file_cache(page) != !file))
829 return ret;
830
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700831 /*
832 * When this function is being called for lumpy reclaim, we
833 * initially look into all LRU pages, active, inactive and
834 * unevictable; only give shrink_page_list evictable pages.
835 */
836 if (PageUnevictable(page))
837 return ret;
838
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700839 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800840
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700841 if (likely(get_page_unless_zero(page))) {
842 /*
843 * Be careful not to clear PageLRU until after we're
844 * sure the page is not being freed elsewhere -- the
845 * page release code relies on it.
846 */
847 ClearPageLRU(page);
848 ret = 0;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800849 mem_cgroup_del_lru(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700850 }
851
852 return ret;
853}
854
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800855/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856 * zone->lru_lock is heavily contended. Some of the functions that
857 * shrink the lists perform better by taking out a batch of pages
858 * and working on them outside the LRU lock.
859 *
860 * For pagecache intensive workloads, this function is the hottest
861 * spot in the kernel (apart from copy_*_user functions).
862 *
863 * Appropriate locks must be held before calling this function.
864 *
865 * @nr_to_scan: The number of pages to look through on the list.
866 * @src: The LRU list to pull pages off.
867 * @dst: The temp list to put pages on to.
868 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700869 * @order: The caller's attempted allocation order
870 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700871 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700872 *
873 * returns how many pages were moved onto *@dst.
874 */
Andrew Morton69e05942006-03-22 00:08:19 -0800875static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
876 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700877 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700878{
Andrew Morton69e05942006-03-22 00:08:19 -0800879 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800880 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700881
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800882 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700883 struct page *page;
884 unsigned long pfn;
885 unsigned long end_pfn;
886 unsigned long page_pfn;
887 int zone_id;
888
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889 page = lru_to_page(src);
890 prefetchw_prev_lru_page(page, src, flags);
891
Nick Piggin725d7042006-09-25 23:30:55 -0700892 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800893
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700894 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700895 case 0:
896 list_move(&page->lru, dst);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800897 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700898 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800899
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700900 case -EBUSY:
901 /* else it is being freed elsewhere */
902 list_move(&page->lru, src);
903 continue;
904
905 default:
906 BUG();
907 }
908
909 if (!order)
910 continue;
911
912 /*
913 * Attempt to take all pages in the order aligned region
914 * surrounding the tag page. Only take those pages of
915 * the same active state as that tag page. We may safely
916 * round the target page pfn down to the requested order
917 * as the mem_map is guarenteed valid out to MAX_ORDER,
918 * where that page is in a different zone we will detect
919 * it from its zone id and abort this block scan.
920 */
921 zone_id = page_zone_id(page);
922 page_pfn = page_to_pfn(page);
923 pfn = page_pfn & ~((1 << order) - 1);
924 end_pfn = pfn + (1 << order);
925 for (; pfn < end_pfn; pfn++) {
926 struct page *cursor_page;
927
928 /* The target page is in the block, ignore it. */
929 if (unlikely(pfn == page_pfn))
930 continue;
931
932 /* Avoid holes within the zone. */
933 if (unlikely(!pfn_valid_within(pfn)))
934 break;
935
936 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700937
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700938 /* Check that we have not crossed a zone boundary. */
939 if (unlikely(page_zone_id(cursor_page) != zone_id))
940 continue;
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700941 switch (__isolate_lru_page(cursor_page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700942 case 0:
943 list_move(&cursor_page->lru, dst);
944 nr_taken++;
945 scan++;
946 break;
947
948 case -EBUSY:
949 /* else it is being freed elsewhere */
950 list_move(&cursor_page->lru, src);
951 default:
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700952 break; /* ! on LRU or wrong list */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700953 }
954 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 }
956
957 *scanned = scan;
958 return nr_taken;
959}
960
Balbir Singh66e17072008-02-07 00:13:56 -0800961static unsigned long isolate_pages_global(unsigned long nr,
962 struct list_head *dst,
963 unsigned long *scanned, int order,
964 int mode, struct zone *z,
965 struct mem_cgroup *mem_cont,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700966 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -0800967{
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700968 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -0800969 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700970 lru += LRU_ACTIVE;
971 if (file)
972 lru += LRU_FILE;
973 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
974 mode, !!file);
Balbir Singh66e17072008-02-07 00:13:56 -0800975}
976
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700978 * clear_active_flags() is a helper for shrink_active_list(), clearing
979 * any active bits from the pages in the list.
980 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700981static unsigned long clear_active_flags(struct list_head *page_list,
982 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700983{
984 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700985 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700986 struct page *page;
987
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700988 list_for_each_entry(page, page_list, lru) {
989 lru = page_is_file_cache(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700990 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700991 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700992 ClearPageActive(page);
993 nr_active++;
994 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700995 count[lru]++;
996 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700997
998 return nr_active;
999}
1000
Nick Piggin62695a82008-10-18 20:26:09 -07001001/**
1002 * isolate_lru_page - tries to isolate a page from its LRU list
1003 * @page: page to isolate from its LRU list
1004 *
1005 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1006 * vmstat statistic corresponding to whatever LRU list the page was on.
1007 *
1008 * Returns 0 if the page was removed from an LRU list.
1009 * Returns -EBUSY if the page was not on an LRU list.
1010 *
1011 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001012 * the active list, it will have PageActive set. If it was found on
1013 * the unevictable list, it will have the PageUnevictable bit set. That flag
1014 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001015 *
1016 * The vmstat statistic corresponding to the list on which the page was
1017 * found will be decremented.
1018 *
1019 * Restrictions:
1020 * (1) Must be called with an elevated refcount on the page. This is a
1021 * fundamentnal difference from isolate_lru_pages (which is called
1022 * without a stable reference).
1023 * (2) the lru_lock must not be held.
1024 * (3) interrupts must be enabled.
1025 */
1026int isolate_lru_page(struct page *page)
1027{
1028 int ret = -EBUSY;
1029
1030 if (PageLRU(page)) {
1031 struct zone *zone = page_zone(page);
1032
1033 spin_lock_irq(&zone->lru_lock);
1034 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001035 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001036 ret = 0;
1037 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001038
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001039 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001040 }
1041 spin_unlock_irq(&zone->lru_lock);
1042 }
1043 return ret;
1044}
1045
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001046/*
Andrew Morton1742f192006-03-22 00:08:21 -08001047 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1048 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001049 */
Andrew Morton1742f192006-03-22 00:08:21 -08001050static unsigned long shrink_inactive_list(unsigned long max_scan,
Rik van Riel33c120e2008-10-18 20:26:36 -07001051 struct zone *zone, struct scan_control *sc,
1052 int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001053{
1054 LIST_HEAD(page_list);
1055 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -08001056 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001057 unsigned long nr_reclaimed = 0;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001058 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001059
1060 pagevec_init(&pvec, 1);
1061
1062 lru_add_drain();
1063 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -08001064 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001065 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -08001066 unsigned long nr_taken;
1067 unsigned long nr_scan;
1068 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001069 unsigned long nr_active;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001070 unsigned int count[NR_LRU_LISTS] = { 0, };
Rik van Riel33c120e2008-10-18 20:26:36 -07001071 int mode = ISOLATE_INACTIVE;
1072
1073 /*
1074 * If we need a large contiguous chunk of memory, or have
1075 * trouble getting a small set of contiguous pages, we
1076 * will reclaim both active and inactive pages.
1077 *
1078 * We use the same threshold as pageout congestion_wait below.
1079 */
1080 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1081 mode = ISOLATE_BOTH;
1082 else if (sc->order && priority < DEF_PRIORITY - 2)
1083 mode = ISOLATE_BOTH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001084
Balbir Singh66e17072008-02-07 00:13:56 -08001085 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001086 &page_list, &nr_scan, sc->order, mode,
1087 zone, sc->mem_cgroup, 0, file);
1088 nr_active = clear_active_flags(&page_list, count);
Andy Whitcrofte9187bd2007-08-22 14:01:25 -07001089 __count_vm_events(PGDEACTIVATE, nr_active);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001090
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001091 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1092 -count[LRU_ACTIVE_FILE]);
1093 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1094 -count[LRU_INACTIVE_FILE]);
1095 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1096 -count[LRU_ACTIVE_ANON]);
1097 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1098 -count[LRU_INACTIVE_ANON]);
1099
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001100 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001101 zone->pages_scanned += nr_scan;
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001102
1103 reclaim_stat->recent_scanned[0] += count[LRU_INACTIVE_ANON];
1104 reclaim_stat->recent_scanned[0] += count[LRU_ACTIVE_ANON];
1105 reclaim_stat->recent_scanned[1] += count[LRU_INACTIVE_FILE];
1106 reclaim_stat->recent_scanned[1] += count[LRU_ACTIVE_FILE];
1107
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 spin_unlock_irq(&zone->lru_lock);
1109
Andrew Morton69e05942006-03-22 00:08:19 -08001110 nr_scanned += nr_scan;
Andy Whitcroftc661b072007-08-22 14:01:26 -07001111 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1112
1113 /*
1114 * If we are direct reclaiming for contiguous pages and we do
1115 * not reclaim everything in the list, try again and wait
1116 * for IO to complete. This will stall high-order allocations
1117 * but that should be acceptable to the caller
1118 */
1119 if (nr_freed < nr_taken && !current_is_kswapd() &&
1120 sc->order > PAGE_ALLOC_COSTLY_ORDER) {
1121 congestion_wait(WRITE, HZ/10);
1122
1123 /*
1124 * The attempt at page out may have made some
1125 * of the pages active, mark them inactive again.
1126 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001127 nr_active = clear_active_flags(&page_list, count);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001128 count_vm_events(PGDEACTIVATE, nr_active);
1129
1130 nr_freed += shrink_page_list(&page_list, sc,
1131 PAGEOUT_IO_SYNC);
1132 }
1133
Andrew Morton05ff5132006-03-22 00:08:20 -08001134 nr_reclaimed += nr_freed;
Nick Piggina74609f2006-01-06 00:11:20 -08001135 local_irq_disable();
1136 if (current_is_kswapd()) {
Christoph Lameterf8891e52006-06-30 01:55:45 -07001137 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
1138 __count_vm_events(KSWAPD_STEAL, nr_freed);
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001139 } else if (scanning_global_lru(sc))
Christoph Lameterf8891e52006-06-30 01:55:45 -07001140 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001141
Shantanu Goel918d3f92006-12-29 16:48:59 -08001142 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -08001143
Wu Fengguangfb8d14e2006-03-22 00:08:28 -08001144 if (nr_taken == 0)
1145 goto done;
1146
Nick Piggina74609f2006-01-06 00:11:20 -08001147 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001148 /*
1149 * Put back any unfreeable pages.
1150 */
1151 while (!list_empty(&page_list)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001152 int lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001153 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -07001154 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001155 list_del(&page->lru);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001156 if (unlikely(!page_evictable(page, NULL))) {
1157 spin_unlock_irq(&zone->lru_lock);
1158 putback_lru_page(page);
1159 spin_lock_irq(&zone->lru_lock);
1160 continue;
1161 }
1162 SetPageLRU(page);
1163 lru = page_lru(page);
1164 add_page_to_lru_list(zone, page, lru);
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001165 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001166 int file = !!page_is_file_cache(page);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001167 reclaim_stat->recent_rotated[file]++;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001168 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001169 if (!pagevec_add(&pvec, page)) {
1170 spin_unlock_irq(&zone->lru_lock);
1171 __pagevec_release(&pvec);
1172 spin_lock_irq(&zone->lru_lock);
1173 }
1174 }
Andrew Morton69e05942006-03-22 00:08:19 -08001175 } while (nr_scanned < max_scan);
Wu Fengguangfb8d14e2006-03-22 00:08:28 -08001176 spin_unlock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177done:
Wu Fengguangfb8d14e2006-03-22 00:08:28 -08001178 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001179 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -08001180 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001181}
1182
Martin Bligh3bb1a852006-10-28 10:38:24 -07001183/*
1184 * We are about to scan this zone at a certain priority level. If that priority
1185 * level is smaller (ie: more urgent) than the previous priority, then note
1186 * that priority level within the zone. This is done so that when the next
1187 * process comes in to scan this zone, it will immediately start out at this
1188 * priority level rather than having to build up its own scanning priority.
1189 * Here, this priority affects only the reclaim-mapped threshold.
1190 */
1191static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1192{
1193 if (priority < zone->prev_priority)
1194 zone->prev_priority = priority;
1195}
1196
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001197/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001198 * This moves pages from the active list to the inactive list.
1199 *
1200 * We move them the other way if the page is referenced by one or more
1201 * processes, from rmap.
1202 *
1203 * If the pages are mostly unmapped, the processing is fast and it is
1204 * appropriate to hold zone->lru_lock across the whole operation. But if
1205 * the pages are mapped, the processing is slow (page_referenced()) so we
1206 * should drop zone->lru_lock around each page. It's impossible to balance
1207 * this, so instead we remove the pages from the LRU while processing them.
1208 * It is safe to rely on PG_active against the non-LRU pages in here because
1209 * nobody will play with that bit on a non-LRU page.
1210 *
1211 * The downside is that we have to touch page->_count against each page.
1212 * But we had to alter page->flags anyway.
1213 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001214
1215
Andrew Morton1742f192006-03-22 00:08:21 -08001216static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001217 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218{
Andrew Morton69e05942006-03-22 00:08:19 -08001219 unsigned long pgmoved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001220 int pgdeactivate = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001221 unsigned long pgscanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222 LIST_HEAD(l_hold); /* The pages which were snipped off */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001223 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 struct page *page;
1225 struct pagevec pvec;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001226 enum lru_list lru;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001227 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001228
1229 lru_add_drain();
1230 spin_lock_irq(&zone->lru_lock);
Balbir Singh66e17072008-02-07 00:13:56 -08001231 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1232 ISOLATE_ACTIVE, zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001233 sc->mem_cgroup, 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001234 /*
1235 * zone->pages_scanned is used for detect zone's oom
1236 * mem_cgroup remembers nr_scan by itself.
1237 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001238 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001239 zone->pages_scanned += pgscanned;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001240 }
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001241 reclaim_stat->recent_scanned[!!file] += pgmoved;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001242
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001243 if (file)
1244 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -pgmoved);
1245 else
1246 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001247 spin_unlock_irq(&zone->lru_lock);
1248
Rik van Riel556adec2008-10-18 20:26:34 -07001249 pgmoved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001250 while (!list_empty(&l_hold)) {
1251 cond_resched();
1252 page = lru_to_page(&l_hold);
1253 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001254
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001255 if (unlikely(!page_evictable(page, NULL))) {
1256 putback_lru_page(page);
1257 continue;
1258 }
1259
Rik van Riel7e9cd482008-10-18 20:26:35 -07001260 /* page_referenced clears PageReferenced */
1261 if (page_mapping_inuse(page) &&
1262 page_referenced(page, 0, sc->mem_cgroup))
1263 pgmoved++;
1264
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265 list_add(&page->lru, &l_inactive);
1266 }
1267
Andrew Mortonb5557492009-01-06 14:40:13 -08001268 /*
1269 * Move the pages to the [file or anon] inactive list.
1270 */
1271 pagevec_init(&pvec, 1);
Andrew Mortonb5557492009-01-06 14:40:13 -08001272 lru = LRU_BASE + file * LRU_FILE;
1273
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001274 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001275 /*
Rik van Riel7e9cd482008-10-18 20:26:35 -07001276 * Count referenced pages from currently used mappings as
1277 * rotated, even though they are moved to the inactive list.
1278 * This helps balance scan pressure between file and anonymous
1279 * pages in get_scan_ratio.
1280 */
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001281 reclaim_stat->recent_rotated[!!file] += pgmoved;
Rik van Riel556adec2008-10-18 20:26:34 -07001282
Daisuke Nishimura1d885522009-03-13 13:52:00 -07001283 pgmoved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 while (!list_empty(&l_inactive)) {
1285 page = lru_to_page(&l_inactive);
1286 prefetchw_prev_lru_page(page, &l_inactive, flags);
Nick Piggin725d7042006-09-25 23:30:55 -07001287 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001288 SetPageLRU(page);
Nick Piggin725d7042006-09-25 23:30:55 -07001289 VM_BUG_ON(!PageActive(page));
Nick Piggin4c84cac2006-03-22 00:08:00 -08001290 ClearPageActive(page);
1291
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001292 list_move(&page->lru, &zone->lru[lru].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001293 mem_cgroup_add_lru_list(page, lru);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001294 pgmoved++;
1295 if (!pagevec_add(&pvec, page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001296 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001297 spin_unlock_irq(&zone->lru_lock);
1298 pgdeactivate += pgmoved;
1299 pgmoved = 0;
1300 if (buffer_heads_over_limit)
1301 pagevec_strip(&pvec);
1302 __pagevec_release(&pvec);
1303 spin_lock_irq(&zone->lru_lock);
1304 }
1305 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001306 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307 pgdeactivate += pgmoved;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001308 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1309 __count_vm_events(PGDEACTIVATE, pgdeactivate);
1310 spin_unlock_irq(&zone->lru_lock);
Johannes Weiner24434622009-03-31 15:23:12 -07001311 if (buffer_heads_over_limit)
1312 pagevec_strip(&pvec);
Nick Piggina74609f2006-01-06 00:11:20 -08001313 pagevec_release(&pvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314}
1315
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001316static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001317{
1318 unsigned long active, inactive;
1319
1320 active = zone_page_state(zone, NR_ACTIVE_ANON);
1321 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1322
1323 if (inactive * zone->inactive_ratio < active)
1324 return 1;
1325
1326 return 0;
1327}
1328
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001329/**
1330 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1331 * @zone: zone to check
1332 * @sc: scan control of this context
1333 *
1334 * Returns true if the zone does not have enough inactive anon pages,
1335 * meaning some active anon pages need to be deactivated.
1336 */
1337static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1338{
1339 int low;
1340
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001341 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001342 low = inactive_anon_is_low_global(zone);
1343 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001344 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001345 return low;
1346}
1347
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001348static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001349 struct zone *zone, struct scan_control *sc, int priority)
1350{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001351 int file = is_file_lru(lru);
1352
Rik van Riel556adec2008-10-18 20:26:34 -07001353 if (lru == LRU_ACTIVE_FILE) {
1354 shrink_active_list(nr_to_scan, zone, sc, priority, file);
1355 return 0;
1356 }
1357
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001358 if (lru == LRU_ACTIVE_ANON && inactive_anon_is_low(zone, sc)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001359 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001360 return 0;
1361 }
Rik van Riel33c120e2008-10-18 20:26:36 -07001362 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001363}
1364
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001366 * Determine how aggressively the anon and file LRU lists should be
1367 * scanned. The relative value of each set of LRU lists is determined
1368 * by looking at the fraction of the pages scanned we did rotate back
1369 * onto the active list instead of evict.
1370 *
1371 * percent[0] specifies how much pressure to put on ram/swap backed
1372 * memory, while percent[1] determines pressure on the file LRUs.
1373 */
1374static void get_scan_ratio(struct zone *zone, struct scan_control *sc,
1375 unsigned long *percent)
1376{
1377 unsigned long anon, file, free;
1378 unsigned long anon_prio, file_prio;
1379 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001380 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001381
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001382 /* If we have no swap space, do not bother scanning anon pages. */
1383 if (nr_swap_pages <= 0) {
1384 percent[0] = 0;
1385 percent[1] = 100;
1386 return;
1387 }
1388
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -08001389 anon = zone_nr_pages(zone, sc, LRU_ACTIVE_ANON) +
1390 zone_nr_pages(zone, sc, LRU_INACTIVE_ANON);
1391 file = zone_nr_pages(zone, sc, LRU_ACTIVE_FILE) +
1392 zone_nr_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001393
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001394 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001395 free = zone_page_state(zone, NR_FREE_PAGES);
1396 /* If we have very few page cache pages,
1397 force-scan anon pages. */
1398 if (unlikely(file + free <= zone->pages_high)) {
1399 percent[0] = 100;
1400 percent[1] = 0;
1401 return;
1402 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001403 }
1404
1405 /*
1406 * OK, so we have swap space and a fair amount of page cache
1407 * pages. We use the recently rotated / recently scanned
1408 * ratios to determine how valuable each cache is.
1409 *
1410 * Because workloads change over time (and to avoid overflow)
1411 * we keep these statistics as a floating average, which ends
1412 * up weighing recent references more than old ones.
1413 *
1414 * anon in [0], file in [1]
1415 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001416 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001417 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001418 reclaim_stat->recent_scanned[0] /= 2;
1419 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001420 spin_unlock_irq(&zone->lru_lock);
1421 }
1422
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001423 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001424 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001425 reclaim_stat->recent_scanned[1] /= 2;
1426 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001427 spin_unlock_irq(&zone->lru_lock);
1428 }
1429
1430 /*
1431 * With swappiness at 100, anonymous and file have the same priority.
1432 * This scanning priority is essentially the inverse of IO cost.
1433 */
1434 anon_prio = sc->swappiness;
1435 file_prio = 200 - sc->swappiness;
1436
1437 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001438 * The amount of pressure on anon vs file pages is inversely
1439 * proportional to the fraction of recently scanned pages on
1440 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001441 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001442 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1443 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001444
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001445 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1446 fp /= reclaim_stat->recent_rotated[1] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001447
1448 /* Normalize to percentages */
1449 percent[0] = 100 * ap / (ap + fp + 1);
1450 percent[1] = 100 - percent[0];
1451}
1452
1453
1454/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1456 */
Rik van Riela79311c2009-01-06 14:40:01 -08001457static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001458 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001459{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001460 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001461 unsigned long nr_to_scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001462 unsigned long percent[2]; /* anon @ 0; file @ 1 */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001463 enum lru_list l;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001464 unsigned long nr_reclaimed = sc->nr_reclaimed;
1465 unsigned long swap_cluster_max = sc->swap_cluster_max;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001466
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001467 get_scan_ratio(zone, sc, percent);
1468
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001469 for_each_evictable_lru(l) {
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001470 int file = is_file_lru(l);
1471 int scan;
Johannes Weinere0f79b82008-10-18 20:26:55 -07001472
KOSAKI Motohirof272b7b2009-03-12 14:31:36 -07001473 scan = zone_nr_pages(zone, sc, l);
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001474 if (priority) {
1475 scan >>= priority;
1476 scan = (scan * percent[file]) / 100;
1477 }
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001478 if (scanning_global_lru(sc)) {
Johannes Weinere0f79b82008-10-18 20:26:55 -07001479 zone->lru[l].nr_scan += scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001480 nr[l] = zone->lru[l].nr_scan;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001481 if (nr[l] >= swap_cluster_max)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001482 zone->lru[l].nr_scan = 0;
1483 else
1484 nr[l] = 0;
KOSAKI Motohiro9439c1c2009-01-07 18:08:21 -08001485 } else
1486 nr[l] = scan;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001487 }
1488
Rik van Riel556adec2008-10-18 20:26:34 -07001489 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1490 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001491 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001492 if (nr[l]) {
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001493 nr_to_scan = min(nr[l], swap_cluster_max);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001494 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001495
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001496 nr_reclaimed += shrink_list(l, nr_to_scan,
1497 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001498 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001499 }
Rik van Riela79311c2009-01-06 14:40:01 -08001500 /*
1501 * On large memory systems, scan >> priority can become
1502 * really large. This is fine for the starting priority;
1503 * we want to put equal scanning pressure on each zone.
1504 * However, if the VM has a harder time of freeing pages,
1505 * with multiple processes reclaiming pages, the total
1506 * freeing target can get unreasonably large.
1507 */
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001508 if (nr_reclaimed > swap_cluster_max &&
Rik van Riela79311c2009-01-06 14:40:01 -08001509 priority < DEF_PRIORITY && !current_is_kswapd())
1510 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001511 }
1512
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001513 sc->nr_reclaimed = nr_reclaimed;
1514
Rik van Riel556adec2008-10-18 20:26:34 -07001515 /*
1516 * Even if we did not try to evict anon pages at all, we want to
1517 * rebalance the anon lru active/inactive ratio.
1518 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001519 if (inactive_anon_is_low(zone, sc))
Rik van Riel556adec2008-10-18 20:26:34 -07001520 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1521
Andrew Morton232ea4d2007-02-28 20:13:21 -08001522 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523}
1524
1525/*
1526 * This is the direct reclaim path, for page-allocating processes. We only
1527 * try to reclaim pages from zones which will satisfy the caller's allocation
1528 * request.
1529 *
1530 * We reclaim from a zone even if that zone is over pages_high. Because:
1531 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1532 * allocation or
1533 * b) The zones may be over pages_high but they must go *over* pages_high to
1534 * satisfy the `incremental min' zone defense algorithm.
1535 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001536 * If a zone is deemed to be full of pinned pages then just give it a light
1537 * scan then give up on it.
1538 */
Rik van Riela79311c2009-01-06 14:40:01 -08001539static void shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001540 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001541{
Mel Gorman54a6eb52008-04-28 02:12:16 -07001542 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Mel Gormandd1a2392008-04-28 02:12:17 -07001543 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001544 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001545
Nick Piggin408d8542006-09-25 23:31:27 -07001546 sc->all_unreclaimable = 1;
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001547 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1548 sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001549 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001551 /*
1552 * Take care memory controller reclaiming has small influence
1553 * to global LRU.
1554 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001555 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001556 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1557 continue;
1558 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001559
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001560 if (zone_is_all_unreclaimable(zone) &&
1561 priority != DEF_PRIORITY)
1562 continue; /* Let kswapd poll it */
1563 sc->all_unreclaimable = 0;
1564 } else {
1565 /*
1566 * Ignore cpuset limitation here. We just want to reduce
1567 * # of used pages by us regardless of memory shortage.
1568 */
1569 sc->all_unreclaimable = 0;
1570 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1571 priority);
1572 }
Nick Piggin408d8542006-09-25 23:31:27 -07001573
Rik van Riela79311c2009-01-06 14:40:01 -08001574 shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001575 }
1576}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001577
Linus Torvalds1da177e2005-04-16 15:20:36 -07001578/*
1579 * This is the main entry point to direct page reclaim.
1580 *
1581 * If a full scan of the inactive list fails to free enough memory then we
1582 * are "out of memory" and something needs to be killed.
1583 *
1584 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1585 * high - the zone may be full of dirty or under-writeback pages, which this
1586 * caller can't do much about. We kick pdflush and take explicit naps in the
1587 * hope that some of these pages can be written. But if the allocating task
1588 * holds filesystem locks which prevent writeout this might not work, and the
1589 * allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001590 *
1591 * returns: 0, if no pages reclaimed
1592 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001593 */
Mel Gormandac1d272008-04-28 02:12:12 -07001594static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001595 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596{
1597 int priority;
kosaki.motohiro@jp.fujitsu.comc700be32008-06-12 15:21:27 -07001598 unsigned long ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001599 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001600 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001601 unsigned long lru_pages = 0;
Mel Gormandd1a2392008-04-28 02:12:17 -07001602 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001603 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -07001604 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001605
Keika Kobayashi873b4772008-07-25 01:48:52 -07001606 delayacct_freepages_start();
1607
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001608 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001609 count_vm_event(ALLOCSTALL);
1610 /*
1611 * mem_cgroup will not do shrink_slab.
1612 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001613 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001614 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001615
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001616 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1617 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001618
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001619 lru_pages += zone_lru_pages(zone);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001620 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001621 }
1622
1623 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001624 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001625 if (!priority)
1626 disable_swap_token();
Rik van Riela79311c2009-01-06 14:40:01 -08001627 shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001628 /*
1629 * Don't shrink slabs when reclaiming memory from
1630 * over limit cgroups
1631 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001632 if (scanning_global_lru(sc)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07001633 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001634 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08001635 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001636 reclaim_state->reclaimed_slab = 0;
1637 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001638 }
Balbir Singh66e17072008-02-07 00:13:56 -08001639 total_scanned += sc->nr_scanned;
Rik van Riela79311c2009-01-06 14:40:01 -08001640 if (sc->nr_reclaimed >= sc->swap_cluster_max) {
1641 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001642 goto out;
1643 }
1644
1645 /*
1646 * Try to write back as many pages as we just scanned. This
1647 * tends to cause slow streaming writers to write data to the
1648 * disk smoothly, at the dirtying rate, which is nice. But
1649 * that's undesirable in laptop mode, where we *want* lumpy
1650 * writeout. So in laptop mode, write out the whole world.
1651 */
Balbir Singh66e17072008-02-07 00:13:56 -08001652 if (total_scanned > sc->swap_cluster_max +
1653 sc->swap_cluster_max / 2) {
Pekka J Enberg687a21c2005-06-28 20:44:55 -07001654 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001655 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001656 }
1657
1658 /* Take a nap, wait for some writeback to complete */
Andrew Morton4dd4b922008-03-24 12:29:52 -07001659 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001660 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001661 }
Fernando Luis Vazquez Cao87547ee2008-07-29 22:33:42 -07001662 /* top priority shrink_zones still had more to do? don't OOM, then */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001663 if (!sc->all_unreclaimable && scanning_global_lru(sc))
Rik van Riela79311c2009-01-06 14:40:01 -08001664 ret = sc->nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001665out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07001666 /*
1667 * Now that we've scanned all the zones at this priority level, note
1668 * that level within the zone so that the next thread which performs
1669 * scanning of this zone will immediately start out at this priority
1670 * level. This affects only the decision whether or not to bring
1671 * mapped pages onto the inactive list.
1672 */
1673 if (priority < 0)
1674 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001675
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001676 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001677 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001678
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001679 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1680 continue;
1681
1682 zone->prev_priority = priority;
1683 }
1684 } else
1685 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1686
Keika Kobayashi873b4772008-07-25 01:48:52 -07001687 delayacct_freepages_end();
1688
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 return ret;
1690}
1691
Mel Gormandac1d272008-04-28 02:12:12 -07001692unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001693 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08001694{
1695 struct scan_control sc = {
1696 .gfp_mask = gfp_mask,
1697 .may_writepage = !laptop_mode,
1698 .swap_cluster_max = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001699 .may_unmap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08001700 .swappiness = vm_swappiness,
1701 .order = order,
1702 .mem_cgroup = NULL,
1703 .isolate_pages = isolate_pages_global,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001704 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08001705 };
1706
Mel Gormandd1a2392008-04-28 02:12:17 -07001707 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001708}
1709
Balbir Singh00f0b822008-03-04 14:28:39 -08001710#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08001711
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001712unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001713 gfp_t gfp_mask,
1714 bool noswap,
1715 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08001716{
1717 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08001718 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001719 .may_unmap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08001720 .swap_cluster_max = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001721 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08001722 .order = 0,
1723 .mem_cgroup = mem_cont,
1724 .isolate_pages = mem_cgroup_isolate_pages,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001725 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08001726 };
Mel Gormandac1d272008-04-28 02:12:12 -07001727 struct zonelist *zonelist;
Balbir Singh66e17072008-02-07 00:13:56 -08001728
KAMEZAWA Hiroyuki8c7c6e32009-01-07 18:08:00 -08001729 if (noswap)
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001730 sc.may_unmap = 0;
KAMEZAWA Hiroyuki8c7c6e32009-01-07 18:08:00 -08001731
Mel Gormandd1a2392008-04-28 02:12:17 -07001732 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1733 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1734 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1735 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001736}
1737#endif
1738
Linus Torvalds1da177e2005-04-16 15:20:36 -07001739/*
1740 * For kswapd, balance_pgdat() will work across all this node's zones until
1741 * they are all at pages_high.
1742 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 * Returns the number of pages which were actually freed.
1744 *
1745 * There is special handling here for zones which are full of pinned pages.
1746 * This can happen if the pages are all mlocked, or if they are all used by
1747 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1748 * What we do is to detect the case where all pages in the zone have been
1749 * scanned twice and there has been zero successful reclaim. Mark the zone as
1750 * dead and from now on, only perform a short scan. Basically we're polling
1751 * the zone for when the problem goes away.
1752 *
1753 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1754 * zones which have free_pages > pages_high, but once a zone is found to have
1755 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1756 * of the number of free pages in the lower zones. This interoperates with
1757 * the page allocator fallback scheme to ensure that aging of pages is balanced
1758 * across the zones.
1759 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001760static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001761{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001762 int all_zones_ok;
1763 int priority;
1764 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08001765 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001766 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08001767 struct scan_control sc = {
1768 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001769 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001770 .swap_cluster_max = SWAP_CLUSTER_MAX,
1771 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001772 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08001773 .mem_cgroup = NULL,
1774 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08001775 };
Martin Bligh3bb1a852006-10-28 10:38:24 -07001776 /*
1777 * temp_priority is used to remember the scanning priority at which
1778 * this zone was successfully refilled to free_pages == pages_high.
1779 */
1780 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001781
1782loop_again:
1783 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08001784 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07001785 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001786 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001787
Martin Bligh3bb1a852006-10-28 10:38:24 -07001788 for (i = 0; i < pgdat->nr_zones; i++)
1789 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790
1791 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1792 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1793 unsigned long lru_pages = 0;
1794
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001795 /* The swap token gets in the way of swapout... */
1796 if (!priority)
1797 disable_swap_token();
1798
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 all_zones_ok = 1;
1800
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001801 /*
1802 * Scan in the highmem->dma direction for the highest
1803 * zone which needs scanning
1804 */
1805 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1806 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001807
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001808 if (!populated_zone(zone))
1809 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001810
David Rientjese815af92007-10-16 23:25:54 -07001811 if (zone_is_all_unreclaimable(zone) &&
1812 priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001813 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001814
Rik van Riel556adec2008-10-18 20:26:34 -07001815 /*
1816 * Do some background aging of the anon list, to give
1817 * pages a chance to be referenced before reclaiming.
1818 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001819 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07001820 shrink_active_list(SWAP_CLUSTER_MAX, zone,
1821 &sc, priority, 0);
1822
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001823 if (!zone_watermark_ok(zone, order, zone->pages_high,
1824 0, 0)) {
1825 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001826 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001827 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001828 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001829 if (i < 0)
1830 goto out;
1831
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 for (i = 0; i <= end_zone; i++) {
1833 struct zone *zone = pgdat->node_zones + i;
1834
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001835 lru_pages += zone_lru_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001836 }
1837
1838 /*
1839 * Now scan the zone in the dma->highmem direction, stopping
1840 * at the last zone which needs scanning.
1841 *
1842 * We do this because the page allocator works in the opposite
1843 * direction. This prevents the page allocator from allocating
1844 * pages behind kswapd's direction of progress, which would
1845 * cause too much scanning of the lower zones.
1846 */
1847 for (i = 0; i <= end_zone; i++) {
1848 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001849 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001850
Con Kolivasf3fe6512006-01-06 00:11:15 -08001851 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001852 continue;
1853
David Rientjese815af92007-10-16 23:25:54 -07001854 if (zone_is_all_unreclaimable(zone) &&
1855 priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856 continue;
1857
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001858 if (!zone_watermark_ok(zone, order, zone->pages_high,
1859 end_zone, 0))
1860 all_zones_ok = 0;
Martin Bligh3bb1a852006-10-28 10:38:24 -07001861 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862 sc.nr_scanned = 0;
Martin Bligh3bb1a852006-10-28 10:38:24 -07001863 note_zone_scanning_priority(zone, priority);
Rik van Riel32a43302007-10-16 01:24:50 -07001864 /*
1865 * We put equal pressure on every zone, unless one
1866 * zone has way too many pages free already.
1867 */
1868 if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1869 end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08001870 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001872 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1873 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08001874 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001875 total_scanned += sc.nr_scanned;
David Rientjese815af92007-10-16 23:25:54 -07001876 if (zone_is_all_unreclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001878 if (nr_slab == 0 && zone->pages_scanned >=
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001879 (zone_lru_pages(zone) * 6))
David Rientjese815af92007-10-16 23:25:54 -07001880 zone_set_flag(zone,
1881 ZONE_ALL_UNRECLAIMABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 /*
1883 * If we've done a decent amount of scanning and
1884 * the reclaim ratio is low, start doing writepage
1885 * even in laptop mode
1886 */
1887 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08001888 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 sc.may_writepage = 1;
1890 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001891 if (all_zones_ok)
1892 break; /* kswapd: all done */
1893 /*
1894 * OK, kswapd is getting into trouble. Take a nap, then take
1895 * another pass across the zones.
1896 */
Andrew Morton4dd4b922008-03-24 12:29:52 -07001897 if (total_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001898 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899
1900 /*
1901 * We do this so kswapd doesn't build up large priorities for
1902 * example when it is freeing in parallel with allocators. It
1903 * matches the direct reclaim path behaviour in terms of impact
1904 * on zone->*_priority.
1905 */
Rik van Riela79311c2009-01-06 14:40:01 -08001906 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001907 break;
1908 }
1909out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07001910 /*
1911 * Note within each zone the priority level at which this zone was
1912 * brought into a happy state. So that the next thread which scans this
1913 * zone will start out at that priority level.
1914 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001915 for (i = 0; i < pgdat->nr_zones; i++) {
1916 struct zone *zone = pgdat->node_zones + i;
1917
Martin Bligh3bb1a852006-10-28 10:38:24 -07001918 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001919 }
1920 if (!all_zones_ok) {
1921 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08001922
1923 try_to_freeze();
1924
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08001925 /*
1926 * Fragmentation may mean that the system cannot be
1927 * rebalanced for high-order allocations in all zones.
1928 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
1929 * it means the zones have been fully scanned and are still
1930 * not balanced. For high-order allocations, there is
1931 * little point trying all over again as kswapd may
1932 * infinite loop.
1933 *
1934 * Instead, recheck all watermarks at order-0 as they
1935 * are the most important. If watermarks are ok, kswapd will go
1936 * back to sleep. High-order users can still perform direct
1937 * reclaim if they wish.
1938 */
1939 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
1940 order = sc.order = 0;
1941
Linus Torvalds1da177e2005-04-16 15:20:36 -07001942 goto loop_again;
1943 }
1944
Rik van Riela79311c2009-01-06 14:40:01 -08001945 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001946}
1947
1948/*
1949 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001950 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001951 *
1952 * This basically trickles out pages so that we have _some_
1953 * free memory available even if there is no other activity
1954 * that frees anything up. This is needed for things like routing
1955 * etc, where we otherwise might have all activity going on in
1956 * asynchronous contexts that cannot page things out.
1957 *
1958 * If there are applications that are active memory-allocators
1959 * (most normal use), this basically shouldn't matter.
1960 */
1961static int kswapd(void *p)
1962{
1963 unsigned long order;
1964 pg_data_t *pgdat = (pg_data_t*)p;
1965 struct task_struct *tsk = current;
1966 DEFINE_WAIT(wait);
1967 struct reclaim_state reclaim_state = {
1968 .reclaimed_slab = 0,
1969 };
Mike Travisc5f59f02008-04-04 18:11:10 -07001970 node_to_cpumask_ptr(cpumask, pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971
Nick Piggincf40bd12009-01-21 08:12:39 +01001972 lockdep_set_current_reclaim_state(GFP_KERNEL);
1973
Rusty Russell174596a2009-01-01 10:12:29 +10301974 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07001975 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001976 current->reclaim_state = &reclaim_state;
1977
1978 /*
1979 * Tell the memory management that we're a "memory allocator",
1980 * and that if we need more memory we should get access to it
1981 * regardless (see "__alloc_pages()"). "kswapd" should
1982 * never get caught in the normal page freeing logic.
1983 *
1984 * (Kswapd normally doesn't need memory anyway, but sometimes
1985 * you need a small amount of memory in order to be able to
1986 * page out something else, and this flag essentially protects
1987 * us from recursively trying to free more memory as we're
1988 * trying to free the first piece of memory in the first place).
1989 */
Christoph Lameter930d9152006-01-08 01:00:47 -08001990 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07001991 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992
1993 order = 0;
1994 for ( ; ; ) {
1995 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07001996
Linus Torvalds1da177e2005-04-16 15:20:36 -07001997 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1998 new_order = pgdat->kswapd_max_order;
1999 pgdat->kswapd_max_order = 0;
2000 if (order < new_order) {
2001 /*
2002 * Don't sleep if someone wants a larger 'order'
2003 * allocation
2004 */
2005 order = new_order;
2006 } else {
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002007 if (!freezing(current))
2008 schedule();
2009
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 order = pgdat->kswapd_max_order;
2011 }
2012 finish_wait(&pgdat->kswapd_wait, &wait);
2013
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002014 if (!try_to_freeze()) {
2015 /* We can speed up thawing tasks if we don't call
2016 * balance_pgdat after returning from the refrigerator
2017 */
2018 balance_pgdat(pgdat, order);
2019 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002020 }
2021 return 0;
2022}
2023
2024/*
2025 * A zone is low on free memory, so wake its kswapd task to service it.
2026 */
2027void wakeup_kswapd(struct zone *zone, int order)
2028{
2029 pg_data_t *pgdat;
2030
Con Kolivasf3fe6512006-01-06 00:11:15 -08002031 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002032 return;
2033
2034 pgdat = zone->zone_pgdat;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08002035 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002036 return;
2037 if (pgdat->kswapd_max_order < order)
2038 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08002039 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002040 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002041 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002042 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002043 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002044}
2045
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002046unsigned long global_lru_pages(void)
2047{
2048 return global_page_state(NR_ACTIVE_ANON)
2049 + global_page_state(NR_ACTIVE_FILE)
2050 + global_page_state(NR_INACTIVE_ANON)
2051 + global_page_state(NR_INACTIVE_FILE);
2052}
2053
Linus Torvalds1da177e2005-04-16 15:20:36 -07002054#ifdef CONFIG_PM
2055/*
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002056 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
MinChan Kimd9796772009-03-31 15:19:34 -07002057 * from LRU lists system-wide, for given pass and priority.
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002058 *
2059 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
2060 */
MinChan Kimd9796772009-03-31 15:19:34 -07002061static void shrink_all_zones(unsigned long nr_pages, int prio,
Nigel Cunninghame07aa052006-12-22 01:07:21 -08002062 int pass, struct scan_control *sc)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002063{
2064 struct zone *zone;
MinChan Kimd9796772009-03-31 15:19:34 -07002065 unsigned long nr_reclaimed = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002066
KOSAKI Motohiroee99c712009-03-31 15:19:31 -07002067 for_each_populated_zone(zone) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002068 enum lru_list l;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002069
David Rientjese815af92007-10-16 23:25:54 -07002070 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002071 continue;
2072
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002073 for_each_evictable_lru(l) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002074 enum zone_stat_item ls = NR_LRU_BASE + l;
2075 unsigned long lru_pages = zone_page_state(zone, ls);
2076
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002077 /* For pass = 0, we don't shrink the active list */
Johannes Weiner0cb57252009-02-14 02:04:10 +01002078 if (pass == 0 && (l == LRU_ACTIVE_ANON ||
2079 l == LRU_ACTIVE_FILE))
Christoph Lameterb69408e2008-10-18 20:26:14 -07002080 continue;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002081
Johannes Weiner0cb57252009-02-14 02:04:10 +01002082 zone->lru[l].nr_scan += (lru_pages >> prio) + 1;
Christoph Lameterb69408e2008-10-18 20:26:14 -07002083 if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
Johannes Weiner0cb57252009-02-14 02:04:10 +01002084 unsigned long nr_to_scan;
2085
Christoph Lameterb69408e2008-10-18 20:26:14 -07002086 zone->lru[l].nr_scan = 0;
Johannes Weiner0cb57252009-02-14 02:04:10 +01002087 nr_to_scan = min(nr_pages, lru_pages);
MinChan Kimd9796772009-03-31 15:19:34 -07002088 nr_reclaimed += shrink_list(l, nr_to_scan, zone,
Christoph Lameterb69408e2008-10-18 20:26:14 -07002089 sc, prio);
MinChan Kimd9796772009-03-31 15:19:34 -07002090 if (nr_reclaimed >= nr_pages) {
2091 sc->nr_reclaimed = nr_reclaimed;
2092 return;
2093 }
Christoph Lameterb69408e2008-10-18 20:26:14 -07002094 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002095 }
2096 }
MinChan Kimd9796772009-03-31 15:19:34 -07002097 sc->nr_reclaimed = nr_reclaimed;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002098}
2099
2100/*
2101 * Try to free `nr_pages' of memory, system-wide, and return the number of
2102 * freed pages.
2103 *
2104 * Rather than trying to age LRUs the aim is to preserve the overall
2105 * LRU order by reclaiming preferentially
2106 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107 */
Andrew Morton69e05942006-03-22 00:08:19 -08002108unsigned long shrink_all_memory(unsigned long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002110 unsigned long lru_pages, nr_slab;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002111 int pass;
2112 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002113 struct scan_control sc = {
2114 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002115 .may_unmap = 0,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002116 .may_writepage = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002117 .isolate_pages = isolate_pages_global,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002118 };
2119
2120 current->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002121
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002122 lru_pages = global_lru_pages();
Christoph Lameter972d1a72006-09-25 23:31:51 -07002123 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002124 /* If slab caches are huge, it's better to hit them first */
2125 while (nr_slab >= lru_pages) {
2126 reclaim_state.reclaimed_slab = 0;
2127 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
2128 if (!reclaim_state.reclaimed_slab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002129 break;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002130
MinChan Kimd9796772009-03-31 15:19:34 -07002131 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2132 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002133 goto out;
2134
2135 nr_slab -= reclaim_state.reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002136 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002137
2138 /*
2139 * We try to shrink LRUs in 5 passes:
2140 * 0 = Reclaim from inactive_list only
2141 * 1 = Reclaim from active list but don't reclaim mapped
2142 * 2 = 2nd pass of type 1
2143 * 3 = Reclaim mapped (normal reclaim)
2144 * 4 = 2nd pass of type 3
2145 */
2146 for (pass = 0; pass < 5; pass++) {
2147 int prio;
2148
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002149 /* Force reclaiming mapped pages in the passes #3 and #4 */
Johannes Weiner30491032009-02-14 02:03:08 +01002150 if (pass > 2)
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002151 sc.may_unmap = 1;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002152
2153 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
MinChan Kimd9796772009-03-31 15:19:34 -07002154 unsigned long nr_to_scan = nr_pages - sc.nr_reclaimed;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002155
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002156 sc.nr_scanned = 0;
Johannes Weiner9786bf82009-03-31 15:19:35 -07002157 sc.swap_cluster_max = nr_to_scan;
MinChan Kimd9796772009-03-31 15:19:34 -07002158 shrink_all_zones(nr_to_scan, prio, pass, &sc);
2159 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002160 goto out;
2161
2162 reclaim_state.reclaimed_slab = 0;
Andrew Morton76395d32007-01-05 16:37:05 -08002163 shrink_slab(sc.nr_scanned, sc.gfp_mask,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002164 global_lru_pages());
MinChan Kimd9796772009-03-31 15:19:34 -07002165 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2166 if (sc.nr_reclaimed >= nr_pages)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002167 goto out;
2168
2169 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07002170 congestion_wait(WRITE, HZ / 10);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002171 }
Rafael J. Wysocki248a0302006-03-22 00:09:04 -08002172 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002173
2174 /*
MinChan Kimd9796772009-03-31 15:19:34 -07002175 * If sc.nr_reclaimed = 0, we could not shrink LRUs, but there may be
2176 * something in slab caches
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002177 */
MinChan Kimd9796772009-03-31 15:19:34 -07002178 if (!sc.nr_reclaimed) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002179 do {
2180 reclaim_state.reclaimed_slab = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002181 shrink_slab(nr_pages, sc.gfp_mask, global_lru_pages());
MinChan Kimd9796772009-03-31 15:19:34 -07002182 sc.nr_reclaimed += reclaim_state.reclaimed_slab;
2183 } while (sc.nr_reclaimed < nr_pages &&
2184 reclaim_state.reclaimed_slab > 0);
Andrew Morton76395d32007-01-05 16:37:05 -08002185 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002186
MinChan Kimd9796772009-03-31 15:19:34 -07002187
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002188out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002189 current->reclaim_state = NULL;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002190
MinChan Kimd9796772009-03-31 15:19:34 -07002191 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002192}
2193#endif
2194
Linus Torvalds1da177e2005-04-16 15:20:36 -07002195/* It's optimal to keep kswapds on the same CPUs as their memory, but
2196 not required for correctness. So if the last cpu in a node goes
2197 away, we get changed to run anywhere: as the first one comes back,
2198 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002199static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002200 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002202 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002203
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002204 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002205 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002206 pg_data_t *pgdat = NODE_DATA(nid);
2207 node_to_cpumask_ptr(mask, pgdat->node_id);
2208
Rusty Russell3e597942009-01-01 10:12:24 +10302209 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002210 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002211 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002212 }
2213 }
2214 return NOTIFY_OK;
2215}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002216
Yasunori Goto3218ae12006-06-27 02:53:33 -07002217/*
2218 * This kswapd start function will be called by init and node-hot-add.
2219 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2220 */
2221int kswapd_run(int nid)
2222{
2223 pg_data_t *pgdat = NODE_DATA(nid);
2224 int ret = 0;
2225
2226 if (pgdat->kswapd)
2227 return 0;
2228
2229 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2230 if (IS_ERR(pgdat->kswapd)) {
2231 /* failure at boot is fatal */
2232 BUG_ON(system_state == SYSTEM_BOOTING);
2233 printk("Failed to start kswapd on node %d\n",nid);
2234 ret = -1;
2235 }
2236 return ret;
2237}
2238
Linus Torvalds1da177e2005-04-16 15:20:36 -07002239static int __init kswapd_init(void)
2240{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002241 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002242
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002244 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002245 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002246 hotcpu_notifier(cpu_callback, 0);
2247 return 0;
2248}
2249
2250module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002251
2252#ifdef CONFIG_NUMA
2253/*
2254 * Zone reclaim mode
2255 *
2256 * If non-zero call zone_reclaim when the number of free pages falls below
2257 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002258 */
2259int zone_reclaim_mode __read_mostly;
2260
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002261#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002262#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002263#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2264#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2265
Christoph Lameter9eeff232006-01-18 17:42:31 -08002266/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002267 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2268 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2269 * a zone.
2270 */
2271#define ZONE_RECLAIM_PRIORITY 4
2272
Christoph Lameter9eeff232006-01-18 17:42:31 -08002273/*
Christoph Lameter96146342006-07-03 00:24:13 -07002274 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2275 * occur.
2276 */
2277int sysctl_min_unmapped_ratio = 1;
2278
2279/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002280 * If the number of slab pages in a zone grows beyond this percentage then
2281 * slab reclaim needs to occur.
2282 */
2283int sysctl_min_slab_ratio = 5;
2284
2285/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002286 * Try to free up some pages from this zone through reclaim.
2287 */
Andrew Morton179e9632006-03-22 00:08:18 -08002288static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002289{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002290 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002291 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002292 struct task_struct *p = current;
2293 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002294 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002295 struct scan_control sc = {
2296 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002297 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
Andrew Morton69e05942006-03-22 00:08:19 -08002298 .swap_cluster_max = max_t(unsigned long, nr_pages,
2299 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002300 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002301 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002302 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002303 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08002304 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07002305 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002306
2307 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08002308 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002309 /*
2310 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2311 * and we also need to be able to write out pages for RECLAIM_WRITE
2312 * and RECLAIM_SWAP.
2313 */
2314 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002315 reclaim_state.reclaimed_slab = 0;
2316 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002317
Christoph Lameter0ff38492006-09-25 23:31:52 -07002318 if (zone_page_state(zone, NR_FILE_PAGES) -
2319 zone_page_state(zone, NR_FILE_MAPPED) >
2320 zone->min_unmapped_pages) {
2321 /*
2322 * Free memory by calling shrink zone with increasing
2323 * priorities until we have enough memory freed.
2324 */
2325 priority = ZONE_RECLAIM_PRIORITY;
2326 do {
Martin Bligh3bb1a852006-10-28 10:38:24 -07002327 note_zone_scanning_priority(zone, priority);
Rik van Riela79311c2009-01-06 14:40:01 -08002328 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002329 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002330 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002331 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002332
Christoph Lameter83e33a42006-09-25 23:31:53 -07002333 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2334 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002335 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002336 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002337 * many pages were freed in this zone. So we take the current
2338 * number of slab pages and shake the slab until it is reduced
2339 * by the same nr_pages that we used for reclaiming unmapped
2340 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002341 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002342 * Note that shrink_slab will free memory on all zones and may
2343 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002344 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07002345 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07002346 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2347 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07002348 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07002349
2350 /*
2351 * Update nr_reclaimed by the number of slab pages we
2352 * reclaimed from this zone.
2353 */
Rik van Riela79311c2009-01-06 14:40:01 -08002354 sc.nr_reclaimed += slab_reclaimable -
Christoph Lameter83e33a42006-09-25 23:31:53 -07002355 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002356 }
2357
Christoph Lameter9eeff232006-01-18 17:42:31 -08002358 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002359 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
Rik van Riela79311c2009-01-06 14:40:01 -08002360 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002361}
Andrew Morton179e9632006-03-22 00:08:18 -08002362
2363int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2364{
Andrew Morton179e9632006-03-22 00:08:18 -08002365 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002366 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002367
2368 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002369 * Zone reclaim reclaims unmapped file backed pages and
2370 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002371 *
Christoph Lameter96146342006-07-03 00:24:13 -07002372 * A small portion of unmapped file backed pages is needed for
2373 * file I/O otherwise pages read by file I/O will be immediately
2374 * thrown out if the zone is overallocated. So we do not reclaim
2375 * if less than a specified percentage of the zone is used by
2376 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002377 */
Christoph Lameter34aa1332006-06-30 01:55:37 -07002378 if (zone_page_state(zone, NR_FILE_PAGES) -
Christoph Lameter0ff38492006-09-25 23:31:52 -07002379 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2380 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2381 <= zone->min_slab_pages)
Christoph Lameter96146342006-07-03 00:24:13 -07002382 return 0;
Andrew Morton179e9632006-03-22 00:08:18 -08002383
David Rientjesd773ed62007-10-16 23:26:01 -07002384 if (zone_is_all_unreclaimable(zone))
2385 return 0;
2386
Andrew Morton179e9632006-03-22 00:08:18 -08002387 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002388 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002389 */
David Rientjesd773ed62007-10-16 23:26:01 -07002390 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Andrew Morton179e9632006-03-22 00:08:18 -08002391 return 0;
2392
2393 /*
2394 * Only run zone reclaim on the local zone or on zones that do not
2395 * have associated processors. This will favor the local processor
2396 * over remote processors and spread off node memory allocations
2397 * as wide as possible.
2398 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002399 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002400 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Andrew Morton179e9632006-03-22 00:08:18 -08002401 return 0;
David Rientjesd773ed62007-10-16 23:26:01 -07002402
2403 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2404 return 0;
2405 ret = __zone_reclaim(zone, gfp_mask, order);
2406 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2407
2408 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002409}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002410#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002411
2412#ifdef CONFIG_UNEVICTABLE_LRU
2413/*
2414 * page_evictable - test whether a page is evictable
2415 * @page: the page to test
2416 * @vma: the VMA in which the page is or will be mapped, may be NULL
2417 *
2418 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07002419 * lists vs unevictable list. The vma argument is !NULL when called from the
2420 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002421 *
2422 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002423 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07002424 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002425 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002426 */
2427int page_evictable(struct page *page, struct vm_area_struct *vma)
2428{
2429
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002430 if (mapping_unevictable(page_mapping(page)))
2431 return 0;
2432
Nick Pigginb291f002008-10-18 20:26:44 -07002433 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2434 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002435
2436 return 1;
2437}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002438
2439/**
2440 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2441 * @page: page to check evictability and move to appropriate lru list
2442 * @zone: zone page is in
2443 *
2444 * Checks a page for evictability and moves the page to the appropriate
2445 * zone lru list.
2446 *
2447 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2448 * have PageUnevictable set.
2449 */
2450static void check_move_unevictable_page(struct page *page, struct zone *zone)
2451{
2452 VM_BUG_ON(PageActive(page));
2453
2454retry:
2455 ClearPageUnevictable(page);
2456 if (page_evictable(page, NULL)) {
2457 enum lru_list l = LRU_INACTIVE_ANON + page_is_file_cache(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002458
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002459 __dec_zone_state(zone, NR_UNEVICTABLE);
2460 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002461 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002462 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2463 __count_vm_event(UNEVICTABLE_PGRESCUED);
2464 } else {
2465 /*
2466 * rotate unevictable list
2467 */
2468 SetPageUnevictable(page);
2469 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002470 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002471 if (page_evictable(page, NULL))
2472 goto retry;
2473 }
2474}
2475
2476/**
2477 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2478 * @mapping: struct address_space to scan for evictable pages
2479 *
2480 * Scan all pages in mapping. Check unevictable pages for
2481 * evictability and move them to the appropriate zone lru list.
2482 */
2483void scan_mapping_unevictable_pages(struct address_space *mapping)
2484{
2485 pgoff_t next = 0;
2486 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2487 PAGE_CACHE_SHIFT;
2488 struct zone *zone;
2489 struct pagevec pvec;
2490
2491 if (mapping->nrpages == 0)
2492 return;
2493
2494 pagevec_init(&pvec, 0);
2495 while (next < end &&
2496 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2497 int i;
2498 int pg_scanned = 0;
2499
2500 zone = NULL;
2501
2502 for (i = 0; i < pagevec_count(&pvec); i++) {
2503 struct page *page = pvec.pages[i];
2504 pgoff_t page_index = page->index;
2505 struct zone *pagezone = page_zone(page);
2506
2507 pg_scanned++;
2508 if (page_index > next)
2509 next = page_index;
2510 next++;
2511
2512 if (pagezone != zone) {
2513 if (zone)
2514 spin_unlock_irq(&zone->lru_lock);
2515 zone = pagezone;
2516 spin_lock_irq(&zone->lru_lock);
2517 }
2518
2519 if (PageLRU(page) && PageUnevictable(page))
2520 check_move_unevictable_page(page, zone);
2521 }
2522 if (zone)
2523 spin_unlock_irq(&zone->lru_lock);
2524 pagevec_release(&pvec);
2525
2526 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2527 }
2528
2529}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002530
2531/**
2532 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2533 * @zone - zone of which to scan the unevictable list
2534 *
2535 * Scan @zone's unevictable LRU lists to check for pages that have become
2536 * evictable. Move those that have to @zone's inactive list where they
2537 * become candidates for reclaim, unless shrink_inactive_zone() decides
2538 * to reactivate them. Pages that are still unevictable are rotated
2539 * back onto @zone's unevictable list.
2540 */
2541#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08002542static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002543{
2544 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2545 unsigned long scan;
2546 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2547
2548 while (nr_to_scan > 0) {
2549 unsigned long batch_size = min(nr_to_scan,
2550 SCAN_UNEVICTABLE_BATCH_SIZE);
2551
2552 spin_lock_irq(&zone->lru_lock);
2553 for (scan = 0; scan < batch_size; scan++) {
2554 struct page *page = lru_to_page(l_unevictable);
2555
2556 if (!trylock_page(page))
2557 continue;
2558
2559 prefetchw_prev_lru_page(page, l_unevictable, flags);
2560
2561 if (likely(PageLRU(page) && PageUnevictable(page)))
2562 check_move_unevictable_page(page, zone);
2563
2564 unlock_page(page);
2565 }
2566 spin_unlock_irq(&zone->lru_lock);
2567
2568 nr_to_scan -= batch_size;
2569 }
2570}
2571
2572
2573/**
2574 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2575 *
2576 * A really big hammer: scan all zones' unevictable LRU lists to check for
2577 * pages that have become evictable. Move those back to the zones'
2578 * inactive list where they become candidates for reclaim.
2579 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2580 * and we add swap to the system. As such, it runs in the context of a task
2581 * that has possibly/probably made some previously unevictable pages
2582 * evictable.
2583 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08002584static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002585{
2586 struct zone *zone;
2587
2588 for_each_zone(zone) {
2589 scan_zone_unevictable_pages(zone);
2590 }
2591}
2592
2593/*
2594 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2595 * all nodes' unevictable lists for evictable pages
2596 */
2597unsigned long scan_unevictable_pages;
2598
2599int scan_unevictable_handler(struct ctl_table *table, int write,
2600 struct file *file, void __user *buffer,
2601 size_t *length, loff_t *ppos)
2602{
2603 proc_doulongvec_minmax(table, write, file, buffer, length, ppos);
2604
2605 if (write && *(unsigned long *)table->data)
2606 scan_all_zones_unevictable_pages();
2607
2608 scan_unevictable_pages = 0;
2609 return 0;
2610}
2611
2612/*
2613 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2614 * a specified node's per zone unevictable lists for evictable pages.
2615 */
2616
2617static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2618 struct sysdev_attribute *attr,
2619 char *buf)
2620{
2621 return sprintf(buf, "0\n"); /* always zero; should fit... */
2622}
2623
2624static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2625 struct sysdev_attribute *attr,
2626 const char *buf, size_t count)
2627{
2628 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2629 struct zone *zone;
2630 unsigned long res;
2631 unsigned long req = strict_strtoul(buf, 10, &res);
2632
2633 if (!req)
2634 return 1; /* zero is no-op */
2635
2636 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2637 if (!populated_zone(zone))
2638 continue;
2639 scan_zone_unevictable_pages(zone);
2640 }
2641 return 1;
2642}
2643
2644
2645static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2646 read_scan_unevictable_node,
2647 write_scan_unevictable_node);
2648
2649int scan_unevictable_register_node(struct node *node)
2650{
2651 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2652}
2653
2654void scan_unevictable_unregister_node(struct node *node)
2655{
2656 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2657}
2658
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002659#endif