blob: b94fe1b3da435f34f567e8a23358ec18ccd452ad [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090016#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070022#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/file.h>
24#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28#include <linux/mm_inline.h>
29#include <linux/pagevec.h>
30#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
35#include <linux/notifier.h>
36#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080037#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070038#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080039#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080040#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070041#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070042#include <linux/sysctl.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043
44#include <asm/tlbflush.h>
45#include <asm/div64.h>
46
47#include <linux/swapops.h>
48
Nick Piggin0f8053a2006-03-22 00:08:33 -080049#include "internal.h"
50
Linus Torvalds1da177e2005-04-16 15:20:36 -070051struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070052 /* Incremented by the number of inactive pages that were scanned */
53 unsigned long nr_scanned;
54
Rik van Riela79311c2009-01-06 14:40:01 -080055 /* Number of pages freed so far during a call to shrink_zones() */
56 unsigned long nr_reclaimed;
57
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080058 /* How many pages shrink_list() should reclaim */
59 unsigned long nr_to_reclaim;
60
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080061 unsigned long hibernation_mode;
62
Linus Torvalds1da177e2005-04-16 15:20:36 -070063 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040064 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070065
66 int may_writepage;
67
Johannes Weinera6dc60f2009-03-31 15:19:30 -070068 /* Can mapped pages be reclaimed? */
69 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080070
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070071 /* Can pages be swapped as part of reclaim? */
72 int may_swap;
73
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070074 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070075
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070076 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080077
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070078 /*
79 * Intend to reclaim enough contenious memory rather than to reclaim
80 * enough amount memory. I.e, it's the mode for high order allocation.
81 */
82 bool lumpy_reclaim_mode;
83
Balbir Singh66e17072008-02-07 00:13:56 -080084 /* Which cgroup do we reclaim from */
85 struct mem_cgroup *mem_cgroup;
86
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -070087 /*
88 * Nodemask of nodes allowed by the caller. If NULL, all nodes
89 * are scanned.
90 */
91 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070092};
93
Linus Torvalds1da177e2005-04-16 15:20:36 -070094#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
95
96#ifdef ARCH_HAS_PREFETCH
97#define prefetch_prev_lru_page(_page, _base, _field) \
98 do { \
99 if ((_page)->lru.prev != _base) { \
100 struct page *prev; \
101 \
102 prev = lru_to_page(&(_page->lru)); \
103 prefetch(&prev->_field); \
104 } \
105 } while (0)
106#else
107#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
108#endif
109
110#ifdef ARCH_HAS_PREFETCHW
111#define prefetchw_prev_lru_page(_page, _base, _field) \
112 do { \
113 if ((_page)->lru.prev != _base) { \
114 struct page *prev; \
115 \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetchw(&prev->_field); \
118 } \
119 } while (0)
120#else
121#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
122#endif
123
124/*
125 * From 0 .. 100. Higher means more swappy.
126 */
127int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700128long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129
130static LIST_HEAD(shrinker_list);
131static DECLARE_RWSEM(shrinker_rwsem);
132
Balbir Singh00f0b822008-03-04 14:28:39 -0800133#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800134#define scanning_global_lru(sc) (!(sc)->mem_cgroup)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800135#else
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800136#define scanning_global_lru(sc) (1)
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800137#endif
138
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800139static struct zone_reclaim_stat *get_reclaim_stat(struct zone *zone,
140 struct scan_control *sc)
141{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800142 if (!scanning_global_lru(sc))
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -0800143 return mem_cgroup_get_reclaim_stat(sc->mem_cgroup, zone);
144
KOSAKI Motohiro6e901572009-01-07 18:08:15 -0800145 return &zone->reclaim_stat;
146}
147
Vincent Li0b217672009-09-21 17:03:09 -0700148static unsigned long zone_nr_lru_pages(struct zone *zone,
149 struct scan_control *sc, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800150{
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -0800151 if (!scanning_global_lru(sc))
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800152 return mem_cgroup_zone_nr_pages(sc->mem_cgroup, zone, lru);
153
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800154 return zone_page_state(zone, NR_LRU_BASE + lru);
155}
156
157
Linus Torvalds1da177e2005-04-16 15:20:36 -0700158/*
159 * Add a shrinker callback to be called from the vm
160 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700161void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700163 shrinker->nr = 0;
164 down_write(&shrinker_rwsem);
165 list_add_tail(&shrinker->list, &shrinker_list);
166 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700167}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700168EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700169
170/*
171 * Remove one
172 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700173void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174{
175 down_write(&shrinker_rwsem);
176 list_del(&shrinker->list);
177 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700178}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700179EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700180
181#define SHRINK_BATCH 128
182/*
183 * Call the shrink functions to age shrinkable caches
184 *
185 * Here we assume it costs one seek to replace a lru page and that it also
186 * takes a seek to recreate a cache object. With this in mind we age equal
187 * percentages of the lru and ageable caches. This should balance the seeks
188 * generated by these structures.
189 *
Simon Arlott183ff222007-10-20 01:27:18 +0200190 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 * slab to avoid swapping.
192 *
193 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
194 *
195 * `lru_pages' represents the number of on-LRU pages in all the zones which
196 * are eligible for the caller's allocation attempt. It is used for balancing
197 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700198 *
199 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700200 */
Andrew Morton69e05942006-03-22 00:08:19 -0800201unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
202 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203{
204 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800205 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206
207 if (scanned == 0)
208 scanned = SWAP_CLUSTER_MAX;
209
210 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700211 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700212
213 list_for_each_entry(shrinker, &shrinker_list, list) {
214 unsigned long long delta;
215 unsigned long total_scan;
Dave Chinner7f8275d2010-07-19 14:56:17 +1000216 unsigned long max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217
Dave Chinner7f8275d2010-07-19 14:56:17 +1000218 max_pass = (*shrinker->shrink)(shrinker, 0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700219 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800220 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221 do_div(delta, lru_pages + 1);
222 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800223 if (shrinker->nr < 0) {
David Rientjes88c3bd72009-03-31 15:23:29 -0700224 printk(KERN_ERR "shrink_slab: %pF negative objects to "
225 "delete nr=%ld\n",
226 shrinker->shrink, shrinker->nr);
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800227 shrinker->nr = max_pass;
228 }
229
230 /*
231 * Avoid risking looping forever due to too large nr value:
232 * never try to free more than twice the estimate number of
233 * freeable entries.
234 */
235 if (shrinker->nr > max_pass * 2)
236 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700237
238 total_scan = shrinker->nr;
239 shrinker->nr = 0;
240
241 while (total_scan >= SHRINK_BATCH) {
242 long this_scan = SHRINK_BATCH;
243 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700244 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700245
Dave Chinner7f8275d2010-07-19 14:56:17 +1000246 nr_before = (*shrinker->shrink)(shrinker, 0, gfp_mask);
247 shrink_ret = (*shrinker->shrink)(shrinker, this_scan,
248 gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700249 if (shrink_ret == -1)
250 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700251 if (shrink_ret < nr_before)
252 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700253 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700254 total_scan -= this_scan;
255
256 cond_resched();
257 }
258
259 shrinker->nr += total_scan;
260 }
261 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700262 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700263}
264
Linus Torvalds1da177e2005-04-16 15:20:36 -0700265static inline int is_page_cache_freeable(struct page *page)
266{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700267 /*
268 * A freeable page cache page is referenced only by the caller
269 * that isolated the page, the page cache radix tree and
270 * optional buffer heads at page->private.
271 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700272 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700273}
274
275static int may_write_to_queue(struct backing_dev_info *bdi)
276{
Christoph Lameter930d9152006-01-08 01:00:47 -0800277 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700278 return 1;
279 if (!bdi_write_congested(bdi))
280 return 1;
281 if (bdi == current->backing_dev_info)
282 return 1;
283 return 0;
284}
285
286/*
287 * We detected a synchronous write error writing a page out. Probably
288 * -ENOSPC. We need to propagate that into the address_space for a subsequent
289 * fsync(), msync() or close().
290 *
291 * The tricky part is that after writepage we cannot touch the mapping: nothing
292 * prevents it from being freed up. But we have a ref on the page and once
293 * that page is locked, the mapping is pinned.
294 *
295 * We're allowed to run sleeping lock_page() here because we know the caller has
296 * __GFP_FS.
297 */
298static void handle_write_error(struct address_space *mapping,
299 struct page *page, int error)
300{
Nick Piggina6aa62a2010-07-20 13:24:25 -0700301 lock_page_nosync(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700302 if (page_mapping(page) == mapping)
303 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700304 unlock_page(page);
305}
306
Andy Whitcroftc661b072007-08-22 14:01:26 -0700307/* Request for sync pageout. */
308enum pageout_io {
309 PAGEOUT_IO_ASYNC,
310 PAGEOUT_IO_SYNC,
311};
312
Christoph Lameter04e62a22006-06-23 02:03:38 -0700313/* possible outcome of pageout() */
314typedef enum {
315 /* failed to write page out, page is locked */
316 PAGE_KEEP,
317 /* move page to the active list, page is locked */
318 PAGE_ACTIVATE,
319 /* page has been sent to the disk successfully, page is unlocked */
320 PAGE_SUCCESS,
321 /* page is clean and locked */
322 PAGE_CLEAN,
323} pageout_t;
324
Linus Torvalds1da177e2005-04-16 15:20:36 -0700325/*
Andrew Morton1742f192006-03-22 00:08:21 -0800326 * pageout is called by shrink_page_list() for each dirty page.
327 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700328 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700329static pageout_t pageout(struct page *page, struct address_space *mapping,
330 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331{
332 /*
333 * If the page is dirty, only perform writeback if that write
334 * will be non-blocking. To prevent this allocation from being
335 * stalled by pagecache activity. But note that there may be
336 * stalls if we need to run get_block(). We could test
337 * PagePrivate for that.
338 *
Vincent Li6aceb532009-12-14 17:58:49 -0800339 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 * this page's queue, we can perform writeback even if that
341 * will block.
342 *
343 * If the page is swapcache, write it back even if that would
344 * block, for some throttling. This happens by accident, because
345 * swap_backing_dev_info is bust: it doesn't reflect the
346 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700347 */
348 if (!is_page_cache_freeable(page))
349 return PAGE_KEEP;
350 if (!mapping) {
351 /*
352 * Some data journaling orphaned pages can have
353 * page->mapping == NULL while being dirty with clean buffers.
354 */
David Howells266cf652009-04-03 16:42:36 +0100355 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356 if (try_to_free_buffers(page)) {
357 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700358 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359 return PAGE_CLEAN;
360 }
361 }
362 return PAGE_KEEP;
363 }
364 if (mapping->a_ops->writepage == NULL)
365 return PAGE_ACTIVATE;
366 if (!may_write_to_queue(mapping->backing_dev_info))
367 return PAGE_KEEP;
368
369 if (clear_page_dirty_for_io(page)) {
370 int res;
371 struct writeback_control wbc = {
372 .sync_mode = WB_SYNC_NONE,
373 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700374 .range_start = 0,
375 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700376 .nonblocking = 1,
377 .for_reclaim = 1,
378 };
379
380 SetPageReclaim(page);
381 res = mapping->a_ops->writepage(page, &wbc);
382 if (res < 0)
383 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800384 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 ClearPageReclaim(page);
386 return PAGE_ACTIVATE;
387 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700388
389 /*
390 * Wait on writeback if requested to. This happens when
391 * direct reclaiming a large contiguous area and the
392 * first attempt to free a range of pages fails.
393 */
394 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
395 wait_on_page_writeback(page);
396
Linus Torvalds1da177e2005-04-16 15:20:36 -0700397 if (!PageWriteback(page)) {
398 /* synchronous write or broken a_ops? */
399 ClearPageReclaim(page);
400 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700401 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700402 return PAGE_SUCCESS;
403 }
404
405 return PAGE_CLEAN;
406}
407
Andrew Mortona649fd92006-10-17 00:09:36 -0700408/*
Nick Piggine2867812008-07-25 19:45:30 -0700409 * Same as remove_mapping, but if the page is removed from the mapping, it
410 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700411 */
Nick Piggine2867812008-07-25 19:45:30 -0700412static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800413{
Nick Piggin28e4d962006-09-25 23:31:23 -0700414 BUG_ON(!PageLocked(page));
415 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800416
Nick Piggin19fd6232008-07-25 19:45:32 -0700417 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800418 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700419 * The non racy check for a busy page.
420 *
421 * Must be careful with the order of the tests. When someone has
422 * a ref to the page, it may be possible that they dirty it then
423 * drop the reference. So if PageDirty is tested before page_count
424 * here, then the following race may occur:
425 *
426 * get_user_pages(&page);
427 * [user mapping goes away]
428 * write_to(page);
429 * !PageDirty(page) [good]
430 * SetPageDirty(page);
431 * put_page(page);
432 * !page_count(page) [good, discard it]
433 *
434 * [oops, our write_to data is lost]
435 *
436 * Reversing the order of the tests ensures such a situation cannot
437 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
438 * load is not satisfied before that of page->_count.
439 *
440 * Note that if SetPageDirty is always performed via set_page_dirty,
441 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800442 */
Nick Piggine2867812008-07-25 19:45:30 -0700443 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800444 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700445 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
446 if (unlikely(PageDirty(page))) {
447 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800448 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700449 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800450
451 if (PageSwapCache(page)) {
452 swp_entry_t swap = { .val = page_private(page) };
453 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700454 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700455 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700456 } else {
457 __remove_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700458 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700459 mem_cgroup_uncharge_cache_page(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800460 }
461
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800462 return 1;
463
464cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700465 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800466 return 0;
467}
468
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469/*
Nick Piggine2867812008-07-25 19:45:30 -0700470 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
471 * someone else has a ref on the page, abort and return 0. If it was
472 * successfully detached, return 1. Assumes the caller has a single ref on
473 * this page.
474 */
475int remove_mapping(struct address_space *mapping, struct page *page)
476{
477 if (__remove_mapping(mapping, page)) {
478 /*
479 * Unfreezing the refcount with 1 rather than 2 effectively
480 * drops the pagecache ref for us without requiring another
481 * atomic operation.
482 */
483 page_unfreeze_refs(page, 1);
484 return 1;
485 }
486 return 0;
487}
488
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700489/**
490 * putback_lru_page - put previously isolated page onto appropriate LRU list
491 * @page: page to be put back to appropriate lru list
492 *
493 * Add previously isolated @page to appropriate LRU list.
494 * Page may still be unevictable for other reasons.
495 *
496 * lru_lock must not be held, interrupts must be enabled.
497 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700498void putback_lru_page(struct page *page)
499{
500 int lru;
501 int active = !!TestClearPageActive(page);
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700502 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700503
504 VM_BUG_ON(PageLRU(page));
505
506redo:
507 ClearPageUnevictable(page);
508
509 if (page_evictable(page, NULL)) {
510 /*
511 * For evictable pages, we can use the cache.
512 * In event of a race, worst case is we end up with an
513 * unevictable page on [in]active list.
514 * We know how to handle that.
515 */
Johannes Weiner401a8e12009-09-21 17:02:58 -0700516 lru = active + page_lru_base_type(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700517 lru_cache_add_lru(page, lru);
518 } else {
519 /*
520 * Put unevictable pages directly on zone's unevictable
521 * list.
522 */
523 lru = LRU_UNEVICTABLE;
524 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700525 /*
526 * When racing with an mlock clearing (page is
527 * unlocked), make sure that if the other thread does
528 * not observe our setting of PG_lru and fails
529 * isolation, we see PG_mlocked cleared below and move
530 * the page back to the evictable list.
531 *
532 * The other side is TestClearPageMlocked().
533 */
534 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700535 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700536
537 /*
538 * page's status can change while we move it among lru. If an evictable
539 * page is on unevictable list, it never be freed. To avoid that,
540 * check after we added it to the list, again.
541 */
542 if (lru == LRU_UNEVICTABLE && page_evictable(page, NULL)) {
543 if (!isolate_lru_page(page)) {
544 put_page(page);
545 goto redo;
546 }
547 /* This means someone else dropped this page from LRU
548 * So, it will be freed or putback to LRU again. There is
549 * nothing to do here.
550 */
551 }
552
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700553 if (was_unevictable && lru != LRU_UNEVICTABLE)
554 count_vm_event(UNEVICTABLE_PGRESCUED);
555 else if (!was_unevictable && lru == LRU_UNEVICTABLE)
556 count_vm_event(UNEVICTABLE_PGCULLED);
557
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700558 put_page(page); /* drop ref from isolate */
559}
560
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800561enum page_references {
562 PAGEREF_RECLAIM,
563 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner64574742010-03-05 13:42:22 -0800564 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800565 PAGEREF_ACTIVATE,
566};
567
568static enum page_references page_check_references(struct page *page,
569 struct scan_control *sc)
570{
Johannes Weiner64574742010-03-05 13:42:22 -0800571 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800572 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800573
Johannes Weiner64574742010-03-05 13:42:22 -0800574 referenced_ptes = page_referenced(page, 1, sc->mem_cgroup, &vm_flags);
575 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800576
577 /* Lumpy reclaim - ignore references */
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -0700578 if (sc->lumpy_reclaim_mode)
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800579 return PAGEREF_RECLAIM;
580
581 /*
582 * Mlock lost the isolation race with us. Let try_to_unmap()
583 * move the page to the unevictable list.
584 */
585 if (vm_flags & VM_LOCKED)
586 return PAGEREF_RECLAIM;
587
Johannes Weiner64574742010-03-05 13:42:22 -0800588 if (referenced_ptes) {
589 if (PageAnon(page))
590 return PAGEREF_ACTIVATE;
591 /*
592 * All mapped pages start out with page table
593 * references from the instantiating fault, so we need
594 * to look twice if a mapped file page is used more
595 * than once.
596 *
597 * Mark it and spare it for another trip around the
598 * inactive list. Another page table reference will
599 * lead to its activation.
600 *
601 * Note: the mark is set for activated pages as well
602 * so that recently deactivated but used pages are
603 * quickly recovered.
604 */
605 SetPageReferenced(page);
606
607 if (referenced_page)
608 return PAGEREF_ACTIVATE;
609
610 return PAGEREF_KEEP;
611 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800612
613 /* Reclaim if clean, defer dirty pages to writeback */
Johannes Weiner64574742010-03-05 13:42:22 -0800614 if (referenced_page)
615 return PAGEREF_RECLAIM_CLEAN;
616
617 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800618}
619
Nick Piggine2867812008-07-25 19:45:30 -0700620/*
Andrew Morton1742f192006-03-22 00:08:21 -0800621 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 */
Andrew Morton1742f192006-03-22 00:08:21 -0800623static unsigned long shrink_page_list(struct list_head *page_list,
Andy Whitcroftc661b072007-08-22 14:01:26 -0700624 struct scan_control *sc,
625 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700626{
627 LIST_HEAD(ret_pages);
628 struct pagevec freed_pvec;
629 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800630 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700631
632 cond_resched();
633
634 pagevec_init(&freed_pvec, 1);
635 while (!list_empty(page_list)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800636 enum page_references references;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700637 struct address_space *mapping;
638 struct page *page;
639 int may_enter_fs;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640
641 cond_resched();
642
643 page = lru_to_page(page_list);
644 list_del(&page->lru);
645
Nick Piggin529ae9a2008-08-02 12:01:03 +0200646 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 goto keep;
648
Nick Piggin725d7042006-09-25 23:30:55 -0700649 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650
651 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800652
Nick Pigginb291f002008-10-18 20:26:44 -0700653 if (unlikely(!page_evictable(page, NULL)))
654 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700655
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700656 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800657 goto keep_locked;
658
Linus Torvalds1da177e2005-04-16 15:20:36 -0700659 /* Double the slab pressure for mapped and swapcache pages */
660 if (page_mapped(page) || PageSwapCache(page))
661 sc->nr_scanned++;
662
Andy Whitcroftc661b072007-08-22 14:01:26 -0700663 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
664 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
665
666 if (PageWriteback(page)) {
667 /*
668 * Synchronous reclaim is performed in two passes,
669 * first an asynchronous pass over the list to
670 * start parallel writeback, and a second synchronous
671 * pass to wait for the IO to complete. Wait here
672 * for any page for which writeback has already
673 * started.
674 */
675 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
676 wait_on_page_writeback(page);
Andrew Morton4dd4b922008-03-24 12:29:52 -0700677 else
Andy Whitcroftc661b072007-08-22 14:01:26 -0700678 goto keep_locked;
679 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700680
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800681 references = page_check_references(page, sc);
682 switch (references) {
683 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700684 goto activate_locked;
Johannes Weiner64574742010-03-05 13:42:22 -0800685 case PAGEREF_KEEP:
686 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800687 case PAGEREF_RECLAIM:
688 case PAGEREF_RECLAIM_CLEAN:
689 ; /* try to reclaim the page below */
690 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691
Linus Torvalds1da177e2005-04-16 15:20:36 -0700692 /*
693 * Anonymous process memory has backing store?
694 * Try to allocate it some swap space here.
695 */
Nick Pigginb291f002008-10-18 20:26:44 -0700696 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800697 if (!(sc->gfp_mask & __GFP_IO))
698 goto keep_locked;
Hugh Dickinsac47b002009-01-06 14:39:39 -0800699 if (!add_to_swap(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800701 may_enter_fs = 1;
Nick Pigginb291f002008-10-18 20:26:44 -0700702 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700703
704 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700705
706 /*
707 * The page is mapped into the page tables of one or more
708 * processes. Try to unmap it here.
709 */
710 if (page_mapped(page) && mapping) {
Andi Kleen14fa31b2009-09-16 11:50:10 +0200711 switch (try_to_unmap(page, TTU_UNMAP)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700712 case SWAP_FAIL:
713 goto activate_locked;
714 case SWAP_AGAIN:
715 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700716 case SWAP_MLOCK:
717 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700718 case SWAP_SUCCESS:
719 ; /* try to free the page below */
720 }
721 }
722
723 if (PageDirty(page)) {
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800724 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700725 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700726 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700727 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800728 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700729 goto keep_locked;
730
731 /* Page is dirty, try to write it out here */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700732 switch (pageout(page, mapping, sync_writeback)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700733 case PAGE_KEEP:
734 goto keep_locked;
735 case PAGE_ACTIVATE:
736 goto activate_locked;
737 case PAGE_SUCCESS:
Andrew Morton4dd4b922008-03-24 12:29:52 -0700738 if (PageWriteback(page) || PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700739 goto keep;
740 /*
741 * A synchronous write - probably a ramdisk. Go
742 * ahead and try to reclaim the page.
743 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200744 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700745 goto keep;
746 if (PageDirty(page) || PageWriteback(page))
747 goto keep_locked;
748 mapping = page_mapping(page);
749 case PAGE_CLEAN:
750 ; /* try to free the page below */
751 }
752 }
753
754 /*
755 * If the page has buffers, try to free the buffer mappings
756 * associated with this page. If we succeed we try to free
757 * the page as well.
758 *
759 * We do this even if the page is PageDirty().
760 * try_to_release_page() does not perform I/O, but it is
761 * possible for a page to have PageDirty set, but it is actually
762 * clean (all its buffers are clean). This happens if the
763 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700764 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700765 * try_to_release_page() will discover that cleanness and will
766 * drop the buffers and mark the page clean - it can be freed.
767 *
768 * Rarely, pages can have buffers and no ->mapping. These are
769 * the pages which were not successfully invalidated in
770 * truncate_complete_page(). We try to drop those buffers here
771 * and if that worked, and the page is no longer mapped into
772 * process address space (page_count == 1) it can be freed.
773 * Otherwise, leave the page on the LRU so it is swappable.
774 */
David Howells266cf652009-04-03 16:42:36 +0100775 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700776 if (!try_to_release_page(page, sc->gfp_mask))
777 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700778 if (!mapping && page_count(page) == 1) {
779 unlock_page(page);
780 if (put_page_testzero(page))
781 goto free_it;
782 else {
783 /*
784 * rare race with speculative reference.
785 * the speculative reference will free
786 * this page shortly, so we may
787 * increment nr_reclaimed here (and
788 * leave it off the LRU).
789 */
790 nr_reclaimed++;
791 continue;
792 }
793 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794 }
795
Nick Piggine2867812008-07-25 19:45:30 -0700796 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800797 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700798
Nick Piggina978d6f2008-10-18 20:26:58 -0700799 /*
800 * At this point, we have no other references and there is
801 * no way to pick any more up (removed from LRU, removed
802 * from pagecache). Can use non-atomic bitops now (and
803 * we obviously don't have to worry about waking up a process
804 * waiting on the page lock, because there are no references.
805 */
806 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -0700807free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800808 nr_reclaimed++;
Nick Piggine2867812008-07-25 19:45:30 -0700809 if (!pagevec_add(&freed_pvec, page)) {
810 __pagevec_free(&freed_pvec);
811 pagevec_reinit(&freed_pvec);
812 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 continue;
814
Nick Pigginb291f002008-10-18 20:26:44 -0700815cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -0800816 if (PageSwapCache(page))
817 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -0700818 unlock_page(page);
819 putback_lru_page(page);
820 continue;
821
Linus Torvalds1da177e2005-04-16 15:20:36 -0700822activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -0700823 /* Not a candidate for swapping, so reclaim swap space. */
824 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -0800825 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700826 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700827 SetPageActive(page);
828 pgactivate++;
829keep_locked:
830 unlock_page(page);
831keep:
832 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -0700833 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834 }
835 list_splice(&ret_pages, page_list);
836 if (pagevec_count(&freed_pvec))
Nick Piggine2867812008-07-25 19:45:30 -0700837 __pagevec_free(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700838 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800839 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700840}
841
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700842/*
843 * Attempt to remove the specified page from its LRU. Only take this page
844 * if it is of the appropriate PageActive status. Pages which are being
845 * freed elsewhere are also ignored.
846 *
847 * page: page to consider
848 * mode: one of the LRU isolation modes defined above
849 *
850 * returns 0 on success, -ve errno on failure.
851 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700852int __isolate_lru_page(struct page *page, int mode, int file)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700853{
854 int ret = -EINVAL;
855
856 /* Only take pages on the LRU. */
857 if (!PageLRU(page))
858 return ret;
859
860 /*
861 * When checking the active state, we need to be sure we are
862 * dealing with comparible boolean values. Take the logical not
863 * of each.
864 */
865 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
866 return ret;
867
Johannes Weiner6c0b1352009-09-21 17:02:59 -0700868 if (mode != ISOLATE_BOTH && page_is_file_cache(page) != file)
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700869 return ret;
870
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700871 /*
872 * When this function is being called for lumpy reclaim, we
873 * initially look into all LRU pages, active, inactive and
874 * unevictable; only give shrink_page_list evictable pages.
875 */
876 if (PageUnevictable(page))
877 return ret;
878
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700879 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -0800880
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700881 if (likely(get_page_unless_zero(page))) {
882 /*
883 * Be careful not to clear PageLRU until after we're
884 * sure the page is not being freed elsewhere -- the
885 * page release code relies on it.
886 */
887 ClearPageLRU(page);
888 ret = 0;
889 }
890
891 return ret;
892}
893
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800894/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 * zone->lru_lock is heavily contended. Some of the functions that
896 * shrink the lists perform better by taking out a batch of pages
897 * and working on them outside the LRU lock.
898 *
899 * For pagecache intensive workloads, this function is the hottest
900 * spot in the kernel (apart from copy_*_user functions).
901 *
902 * Appropriate locks must be held before calling this function.
903 *
904 * @nr_to_scan: The number of pages to look through on the list.
905 * @src: The LRU list to pull pages off.
906 * @dst: The temp list to put pages on to.
907 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700908 * @order: The caller's attempted allocation order
909 * @mode: One of the LRU isolation modes
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700910 * @file: True [1] if isolating file [!anon] pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700911 *
912 * returns how many pages were moved onto *@dst.
913 */
Andrew Morton69e05942006-03-22 00:08:19 -0800914static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
915 struct list_head *src, struct list_head *dst,
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700916 unsigned long *scanned, int order, int mode, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700917{
Andrew Morton69e05942006-03-22 00:08:19 -0800918 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800919 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700920
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800921 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700922 struct page *page;
923 unsigned long pfn;
924 unsigned long end_pfn;
925 unsigned long page_pfn;
926 int zone_id;
927
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 page = lru_to_page(src);
929 prefetchw_prev_lru_page(page, src, flags);
930
Nick Piggin725d7042006-09-25 23:30:55 -0700931 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800932
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700933 switch (__isolate_lru_page(page, mode, file)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700934 case 0:
935 list_move(&page->lru, dst);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700936 mem_cgroup_del_lru(page);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800937 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700938 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800939
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700940 case -EBUSY:
941 /* else it is being freed elsewhere */
942 list_move(&page->lru, src);
KAMEZAWA Hiroyuki2ffebca2009-06-17 16:27:21 -0700943 mem_cgroup_rotate_lru_list(page, page_lru(page));
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700944 continue;
945
946 default:
947 BUG();
948 }
949
950 if (!order)
951 continue;
952
953 /*
954 * Attempt to take all pages in the order aligned region
955 * surrounding the tag page. Only take those pages of
956 * the same active state as that tag page. We may safely
957 * round the target page pfn down to the requested order
958 * as the mem_map is guarenteed valid out to MAX_ORDER,
959 * where that page is in a different zone we will detect
960 * it from its zone id and abort this block scan.
961 */
962 zone_id = page_zone_id(page);
963 page_pfn = page_to_pfn(page);
964 pfn = page_pfn & ~((1 << order) - 1);
965 end_pfn = pfn + (1 << order);
966 for (; pfn < end_pfn; pfn++) {
967 struct page *cursor_page;
968
969 /* The target page is in the block, ignore it. */
970 if (unlikely(pfn == page_pfn))
971 continue;
972
973 /* Avoid holes within the zone. */
974 if (unlikely(!pfn_valid_within(pfn)))
975 break;
976
977 cursor_page = pfn_to_page(pfn);
Rik van Riel4f98a2f2008-10-18 20:26:32 -0700978
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700979 /* Check that we have not crossed a zone boundary. */
980 if (unlikely(page_zone_id(cursor_page) != zone_id))
981 continue;
Minchan Kimde2e7562009-09-21 17:01:43 -0700982
983 /*
984 * If we don't have enough swap space, reclaiming of
985 * anon page which don't already have a swap slot is
986 * pointless.
987 */
988 if (nr_swap_pages <= 0 && PageAnon(cursor_page) &&
989 !PageSwapCache(cursor_page))
990 continue;
991
KAMEZAWA Hiroyukiee993b12009-06-16 15:33:24 -0700992 if (__isolate_lru_page(cursor_page, mode, file) == 0) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700993 list_move(&cursor_page->lru, dst);
KAMEZAWA Hiroyukicb4cbcf2009-06-23 08:57:55 +0900994 mem_cgroup_del_lru(cursor_page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700995 nr_taken++;
996 scan++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700997 }
998 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999 }
1000
1001 *scanned = scan;
1002 return nr_taken;
1003}
1004
Balbir Singh66e17072008-02-07 00:13:56 -08001005static unsigned long isolate_pages_global(unsigned long nr,
1006 struct list_head *dst,
1007 unsigned long *scanned, int order,
1008 int mode, struct zone *z,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001009 int active, int file)
Balbir Singh66e17072008-02-07 00:13:56 -08001010{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001011 int lru = LRU_BASE;
Balbir Singh66e17072008-02-07 00:13:56 -08001012 if (active)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001013 lru += LRU_ACTIVE;
1014 if (file)
1015 lru += LRU_FILE;
1016 return isolate_lru_pages(nr, &z->lru[lru].list, dst, scanned, order,
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001017 mode, file);
Balbir Singh66e17072008-02-07 00:13:56 -08001018}
1019
Linus Torvalds1da177e2005-04-16 15:20:36 -07001020/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001021 * clear_active_flags() is a helper for shrink_active_list(), clearing
1022 * any active bits from the pages in the list.
1023 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001024static unsigned long clear_active_flags(struct list_head *page_list,
1025 unsigned int *count)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001026{
1027 int nr_active = 0;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001028 int lru;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001029 struct page *page;
1030
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001031 list_for_each_entry(page, page_list, lru) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07001032 lru = page_lru_base_type(page);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001033 if (PageActive(page)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001034 lru += LRU_ACTIVE;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001035 ClearPageActive(page);
1036 nr_active++;
1037 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001038 count[lru]++;
1039 }
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001040
1041 return nr_active;
1042}
1043
Nick Piggin62695a82008-10-18 20:26:09 -07001044/**
1045 * isolate_lru_page - tries to isolate a page from its LRU list
1046 * @page: page to isolate from its LRU list
1047 *
1048 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1049 * vmstat statistic corresponding to whatever LRU list the page was on.
1050 *
1051 * Returns 0 if the page was removed from an LRU list.
1052 * Returns -EBUSY if the page was not on an LRU list.
1053 *
1054 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001055 * the active list, it will have PageActive set. If it was found on
1056 * the unevictable list, it will have the PageUnevictable bit set. That flag
1057 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001058 *
1059 * The vmstat statistic corresponding to the list on which the page was
1060 * found will be decremented.
1061 *
1062 * Restrictions:
1063 * (1) Must be called with an elevated refcount on the page. This is a
1064 * fundamentnal difference from isolate_lru_pages (which is called
1065 * without a stable reference).
1066 * (2) the lru_lock must not be held.
1067 * (3) interrupts must be enabled.
1068 */
1069int isolate_lru_page(struct page *page)
1070{
1071 int ret = -EBUSY;
1072
1073 if (PageLRU(page)) {
1074 struct zone *zone = page_zone(page);
1075
1076 spin_lock_irq(&zone->lru_lock);
1077 if (PageLRU(page) && get_page_unless_zero(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001078 int lru = page_lru(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001079 ret = 0;
1080 ClearPageLRU(page);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001081
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001082 del_page_from_lru_list(zone, page, lru);
Nick Piggin62695a82008-10-18 20:26:09 -07001083 }
1084 spin_unlock_irq(&zone->lru_lock);
1085 }
1086 return ret;
1087}
1088
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001089/*
Rik van Riel35cd7812009-09-21 17:01:38 -07001090 * Are there way too many processes in the direct reclaim path already?
1091 */
1092static int too_many_isolated(struct zone *zone, int file,
1093 struct scan_control *sc)
1094{
1095 unsigned long inactive, isolated;
1096
1097 if (current_is_kswapd())
1098 return 0;
1099
1100 if (!scanning_global_lru(sc))
1101 return 0;
1102
1103 if (file) {
1104 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1105 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1106 } else {
1107 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1108 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1109 }
1110
1111 return isolated > inactive;
1112}
1113
1114/*
Andrew Morton1742f192006-03-22 00:08:21 -08001115 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1116 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 */
Andrew Morton1742f192006-03-22 00:08:21 -08001118static unsigned long shrink_inactive_list(unsigned long max_scan,
Rik van Riel33c120e2008-10-18 20:26:36 -07001119 struct zone *zone, struct scan_control *sc,
1120 int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121{
1122 LIST_HEAD(page_list);
1123 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -08001124 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001125 unsigned long nr_reclaimed = 0;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001126 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001127
Rik van Riel35cd7812009-09-21 17:01:38 -07001128 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001129 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001130
1131 /* We are about to die and free our memory. Return now. */
1132 if (fatal_signal_pending(current))
1133 return SWAP_CLUSTER_MAX;
1134 }
1135
Linus Torvalds1da177e2005-04-16 15:20:36 -07001136
1137 pagevec_init(&pvec, 1);
1138
1139 lru_add_drain();
1140 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -08001141 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001142 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -08001143 unsigned long nr_taken;
1144 unsigned long nr_scan;
1145 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001146 unsigned long nr_active;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001147 unsigned int count[NR_LRU_LISTS] = { 0, };
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -07001148 int mode = sc->lumpy_reclaim_mode ? ISOLATE_BOTH : ISOLATE_INACTIVE;
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001149 unsigned long nr_anon;
1150 unsigned long nr_file;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001151
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001152 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001153 nr_taken = isolate_pages_global(SWAP_CLUSTER_MAX,
1154 &page_list, &nr_scan,
1155 sc->order, mode,
1156 zone, 0, file);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001157 zone->pages_scanned += nr_scan;
1158 if (current_is_kswapd())
1159 __count_zone_vm_events(PGSCAN_KSWAPD, zone,
1160 nr_scan);
1161 else
1162 __count_zone_vm_events(PGSCAN_DIRECT, zone,
1163 nr_scan);
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001164 } else {
1165 nr_taken = mem_cgroup_isolate_pages(SWAP_CLUSTER_MAX,
1166 &page_list, &nr_scan,
1167 sc->order, mode,
1168 zone, sc->mem_cgroup,
1169 0, file);
1170 /*
1171 * mem_cgroup_isolate_pages() keeps track of
1172 * scanned pages on its own.
1173 */
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001174 }
1175
1176 if (nr_taken == 0)
1177 goto done;
1178
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001179 nr_active = clear_active_flags(&page_list, count);
Andy Whitcrofte9187bd2007-08-22 14:01:25 -07001180 __count_vm_events(PGDEACTIVATE, nr_active);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001181
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001182 __mod_zone_page_state(zone, NR_ACTIVE_FILE,
1183 -count[LRU_ACTIVE_FILE]);
1184 __mod_zone_page_state(zone, NR_INACTIVE_FILE,
1185 -count[LRU_INACTIVE_FILE]);
1186 __mod_zone_page_state(zone, NR_ACTIVE_ANON,
1187 -count[LRU_ACTIVE_ANON]);
1188 __mod_zone_page_state(zone, NR_INACTIVE_ANON,
1189 -count[LRU_INACTIVE_ANON]);
1190
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001191 nr_anon = count[LRU_ACTIVE_ANON] + count[LRU_INACTIVE_ANON];
1192 nr_file = count[LRU_ACTIVE_FILE] + count[LRU_INACTIVE_FILE];
1193 __mod_zone_page_state(zone, NR_ISOLATED_ANON, nr_anon);
1194 __mod_zone_page_state(zone, NR_ISOLATED_FILE, nr_file);
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001195
Huang Shijie62c0c2f2009-12-14 17:59:48 -08001196 reclaim_stat->recent_scanned[0] += nr_anon;
1197 reclaim_stat->recent_scanned[1] += nr_file;
KOSAKI Motohiro3e2f41f2009-01-07 18:08:20 -08001198
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 spin_unlock_irq(&zone->lru_lock);
1200
Andrew Morton69e05942006-03-22 00:08:19 -08001201 nr_scanned += nr_scan;
Andy Whitcroftc661b072007-08-22 14:01:26 -07001202 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
1203
1204 /*
1205 * If we are direct reclaiming for contiguous pages and we do
1206 * not reclaim everything in the list, try again and wait
1207 * for IO to complete. This will stall high-order allocations
1208 * but that should be acceptable to the caller
1209 */
1210 if (nr_freed < nr_taken && !current_is_kswapd() &&
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -07001211 sc->lumpy_reclaim_mode) {
Jens Axboe8aa7e842009-07-09 14:52:32 +02001212 congestion_wait(BLK_RW_ASYNC, HZ/10);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001213
1214 /*
1215 * The attempt at page out may have made some
1216 * of the pages active, mark them inactive again.
1217 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001218 nr_active = clear_active_flags(&page_list, count);
Andy Whitcroftc661b072007-08-22 14:01:26 -07001219 count_vm_events(PGDEACTIVATE, nr_active);
1220
1221 nr_freed += shrink_page_list(&page_list, sc,
1222 PAGEOUT_IO_SYNC);
1223 }
1224
Andrew Morton05ff5132006-03-22 00:08:20 -08001225 nr_reclaimed += nr_freed;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001226
Nick Piggina74609f2006-01-06 00:11:20 -08001227 local_irq_disable();
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001228 if (current_is_kswapd())
Christoph Lameterf8891e52006-06-30 01:55:45 -07001229 __count_vm_events(KSWAPD_STEAL, nr_freed);
Shantanu Goel918d3f92006-12-29 16:48:59 -08001230 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -08001231
1232 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001233 /*
1234 * Put back any unfreeable pages.
1235 */
1236 while (!list_empty(&page_list)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001237 int lru;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001238 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -07001239 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 list_del(&page->lru);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001241 if (unlikely(!page_evictable(page, NULL))) {
1242 spin_unlock_irq(&zone->lru_lock);
1243 putback_lru_page(page);
1244 spin_lock_irq(&zone->lru_lock);
1245 continue;
1246 }
1247 SetPageLRU(page);
1248 lru = page_lru(page);
1249 add_page_to_lru_list(zone, page, lru);
KOSAKI Motohiro74a1c482009-09-21 17:01:45 -07001250 if (is_active_lru(lru)) {
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001251 int file = is_file_lru(lru);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001252 reclaim_stat->recent_rotated[file]++;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001253 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254 if (!pagevec_add(&pvec, page)) {
1255 spin_unlock_irq(&zone->lru_lock);
1256 __pagevec_release(&pvec);
1257 spin_lock_irq(&zone->lru_lock);
1258 }
1259 }
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001260 __mod_zone_page_state(zone, NR_ISOLATED_ANON, -nr_anon);
1261 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -nr_file);
1262
Andrew Morton69e05942006-03-22 00:08:19 -08001263 } while (nr_scanned < max_scan);
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001264
Linus Torvalds1da177e2005-04-16 15:20:36 -07001265done:
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001266 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001267 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -08001268 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001269}
1270
Martin Bligh3bb1a852006-10-28 10:38:24 -07001271/*
1272 * We are about to scan this zone at a certain priority level. If that priority
1273 * level is smaller (ie: more urgent) than the previous priority, then note
1274 * that priority level within the zone. This is done so that when the next
1275 * process comes in to scan this zone, it will immediately start out at this
1276 * priority level rather than having to build up its own scanning priority.
1277 * Here, this priority affects only the reclaim-mapped threshold.
1278 */
1279static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1280{
1281 if (priority < zone->prev_priority)
1282 zone->prev_priority = priority;
1283}
1284
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001285/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 * This moves pages from the active list to the inactive list.
1287 *
1288 * We move them the other way if the page is referenced by one or more
1289 * processes, from rmap.
1290 *
1291 * If the pages are mostly unmapped, the processing is fast and it is
1292 * appropriate to hold zone->lru_lock across the whole operation. But if
1293 * the pages are mapped, the processing is slow (page_referenced()) so we
1294 * should drop zone->lru_lock around each page. It's impossible to balance
1295 * this, so instead we remove the pages from the LRU while processing them.
1296 * It is safe to rely on PG_active against the non-LRU pages in here because
1297 * nobody will play with that bit on a non-LRU page.
1298 *
1299 * The downside is that we have to touch page->_count against each page.
1300 * But we had to alter page->flags anyway.
1301 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001302
Wu Fengguang3eb41402009-06-16 15:33:13 -07001303static void move_active_pages_to_lru(struct zone *zone,
1304 struct list_head *list,
1305 enum lru_list lru)
1306{
1307 unsigned long pgmoved = 0;
1308 struct pagevec pvec;
1309 struct page *page;
1310
1311 pagevec_init(&pvec, 1);
1312
1313 while (!list_empty(list)) {
1314 page = lru_to_page(list);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001315
1316 VM_BUG_ON(PageLRU(page));
1317 SetPageLRU(page);
1318
Wu Fengguang3eb41402009-06-16 15:33:13 -07001319 list_move(&page->lru, &zone->lru[lru].list);
1320 mem_cgroup_add_lru_list(page, lru);
1321 pgmoved++;
1322
1323 if (!pagevec_add(&pvec, page) || list_empty(list)) {
1324 spin_unlock_irq(&zone->lru_lock);
1325 if (buffer_heads_over_limit)
1326 pagevec_strip(&pvec);
1327 __pagevec_release(&pvec);
1328 spin_lock_irq(&zone->lru_lock);
1329 }
1330 }
1331 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1332 if (!is_active_lru(lru))
1333 __count_vm_events(PGDEACTIVATE, pgmoved);
1334}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001335
Andrew Morton1742f192006-03-22 00:08:21 -08001336static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001337 struct scan_control *sc, int priority, int file)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001338{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001339 unsigned long nr_taken;
Andrew Morton69e05942006-03-22 00:08:19 -08001340 unsigned long pgscanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001341 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001342 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001343 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001344 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345 struct page *page;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001346 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001347 unsigned long nr_rotated = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001348
1349 lru_add_drain();
1350 spin_lock_irq(&zone->lru_lock);
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001351 if (scanning_global_lru(sc)) {
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001352 nr_taken = isolate_pages_global(nr_pages, &l_hold,
1353 &pgscanned, sc->order,
1354 ISOLATE_ACTIVE, zone,
1355 1, file);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001356 zone->pages_scanned += pgscanned;
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001357 } else {
1358 nr_taken = mem_cgroup_isolate_pages(nr_pages, &l_hold,
1359 &pgscanned, sc->order,
1360 ISOLATE_ACTIVE, zone,
1361 sc->mem_cgroup, 1, file);
1362 /*
1363 * mem_cgroup_isolate_pages() keeps track of
1364 * scanned pages on its own.
1365 */
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001366 }
Johannes Weiner8b25c6d2010-05-24 14:32:40 -07001367
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001368 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001369
Wu Fengguang3eb41402009-06-16 15:33:13 -07001370 __count_zone_vm_events(PGREFILL, zone, pgscanned);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001371 if (file)
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001372 __mod_zone_page_state(zone, NR_ACTIVE_FILE, -nr_taken);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001373 else
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001374 __mod_zone_page_state(zone, NR_ACTIVE_ANON, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001375 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001376 spin_unlock_irq(&zone->lru_lock);
1377
Linus Torvalds1da177e2005-04-16 15:20:36 -07001378 while (!list_empty(&l_hold)) {
1379 cond_resched();
1380 page = lru_to_page(&l_hold);
1381 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001382
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001383 if (unlikely(!page_evictable(page, NULL))) {
1384 putback_lru_page(page);
1385 continue;
1386 }
1387
Johannes Weiner64574742010-03-05 13:42:22 -08001388 if (page_referenced(page, 0, sc->mem_cgroup, &vm_flags)) {
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001389 nr_rotated++;
Wu Fengguang8cab4752009-06-16 15:33:12 -07001390 /*
1391 * Identify referenced, file-backed active pages and
1392 * give them one more trip around the active list. So
1393 * that executable code get better chances to stay in
1394 * memory under moderate memory pressure. Anon pages
1395 * are not likely to be evicted by use-once streaming
1396 * IO, plus JVM can create lots of anon VM_EXEC pages,
1397 * so we ignore them here.
1398 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001399 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001400 list_add(&page->lru, &l_active);
1401 continue;
1402 }
1403 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001404
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001405 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001406 list_add(&page->lru, &l_inactive);
1407 }
1408
Andrew Mortonb5557492009-01-06 14:40:13 -08001409 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001410 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001411 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001412 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001413 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001414 * Count referenced pages from currently used mappings as rotated,
1415 * even though only some of them are actually re-activated. This
1416 * helps balance scan pressure between file and anonymous pages in
1417 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001418 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001419 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001420
Wu Fengguang3eb41402009-06-16 15:33:13 -07001421 move_active_pages_to_lru(zone, &l_active,
1422 LRU_ACTIVE + file * LRU_FILE);
1423 move_active_pages_to_lru(zone, &l_inactive,
1424 LRU_BASE + file * LRU_FILE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001425 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001426 spin_unlock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001427}
1428
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001429static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001430{
1431 unsigned long active, inactive;
1432
1433 active = zone_page_state(zone, NR_ACTIVE_ANON);
1434 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1435
1436 if (inactive * zone->inactive_ratio < active)
1437 return 1;
1438
1439 return 0;
1440}
1441
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001442/**
1443 * inactive_anon_is_low - check if anonymous pages need to be deactivated
1444 * @zone: zone to check
1445 * @sc: scan control of this context
1446 *
1447 * Returns true if the zone does not have enough inactive anon pages,
1448 * meaning some active anon pages need to be deactivated.
1449 */
1450static int inactive_anon_is_low(struct zone *zone, struct scan_control *sc)
1451{
1452 int low;
1453
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001454 if (scanning_global_lru(sc))
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001455 low = inactive_anon_is_low_global(zone);
1456 else
KOSAKI Motohiroc772be92009-01-07 18:08:25 -08001457 low = mem_cgroup_inactive_anon_is_low(sc->mem_cgroup);
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001458 return low;
1459}
1460
Rik van Riel56e49d22009-06-16 15:32:28 -07001461static int inactive_file_is_low_global(struct zone *zone)
1462{
1463 unsigned long active, inactive;
1464
1465 active = zone_page_state(zone, NR_ACTIVE_FILE);
1466 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1467
1468 return (active > inactive);
1469}
1470
1471/**
1472 * inactive_file_is_low - check if file pages need to be deactivated
1473 * @zone: zone to check
1474 * @sc: scan control of this context
1475 *
1476 * When the system is doing streaming IO, memory pressure here
1477 * ensures that active file pages get deactivated, until more
1478 * than half of the file pages are on the inactive list.
1479 *
1480 * Once we get to that situation, protect the system's working
1481 * set from being evicted by disabling active file page aging.
1482 *
1483 * This uses a different ratio than the anonymous pages, because
1484 * the page cache uses a use-once replacement algorithm.
1485 */
1486static int inactive_file_is_low(struct zone *zone, struct scan_control *sc)
1487{
1488 int low;
1489
1490 if (scanning_global_lru(sc))
1491 low = inactive_file_is_low_global(zone);
1492 else
1493 low = mem_cgroup_inactive_file_is_low(sc->mem_cgroup);
1494 return low;
1495}
1496
Rik van Rielb39415b2009-12-14 17:59:48 -08001497static int inactive_list_is_low(struct zone *zone, struct scan_control *sc,
1498 int file)
1499{
1500 if (file)
1501 return inactive_file_is_low(zone, sc);
1502 else
1503 return inactive_anon_is_low(zone, sc);
1504}
1505
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001506static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Christoph Lameterb69408e2008-10-18 20:26:14 -07001507 struct zone *zone, struct scan_control *sc, int priority)
1508{
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001509 int file = is_file_lru(lru);
1510
Rik van Rielb39415b2009-12-14 17:59:48 -08001511 if (is_active_lru(lru)) {
1512 if (inactive_list_is_low(zone, sc, file))
1513 shrink_active_list(nr_to_scan, zone, sc, priority, file);
Rik van Riel556adec2008-10-18 20:26:34 -07001514 return 0;
1515 }
1516
Rik van Riel33c120e2008-10-18 20:26:36 -07001517 return shrink_inactive_list(nr_to_scan, zone, sc, priority, file);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001518}
1519
Linus Torvalds1da177e2005-04-16 15:20:36 -07001520/*
Shaohua Li76a33fc2010-05-24 14:32:36 -07001521 * Smallish @nr_to_scan's are deposited in @nr_saved_scan,
1522 * until we collected @swap_cluster_max pages to scan.
1523 */
1524static unsigned long nr_scan_try_batch(unsigned long nr_to_scan,
1525 unsigned long *nr_saved_scan)
1526{
1527 unsigned long nr;
1528
1529 *nr_saved_scan += nr_to_scan;
1530 nr = *nr_saved_scan;
1531
1532 if (nr >= SWAP_CLUSTER_MAX)
1533 *nr_saved_scan = 0;
1534 else
1535 nr = 0;
1536
1537 return nr;
1538}
1539
1540/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001541 * Determine how aggressively the anon and file LRU lists should be
1542 * scanned. The relative value of each set of LRU lists is determined
1543 * by looking at the fraction of the pages scanned we did rotate back
1544 * onto the active list instead of evict.
1545 *
Shaohua Li76a33fc2010-05-24 14:32:36 -07001546 * nr[0] = anon pages to scan; nr[1] = file pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001547 */
Shaohua Li76a33fc2010-05-24 14:32:36 -07001548static void get_scan_count(struct zone *zone, struct scan_control *sc,
1549 unsigned long *nr, int priority)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001550{
1551 unsigned long anon, file, free;
1552 unsigned long anon_prio, file_prio;
1553 unsigned long ap, fp;
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001554 struct zone_reclaim_stat *reclaim_stat = get_reclaim_stat(zone, sc);
Shaohua Li76a33fc2010-05-24 14:32:36 -07001555 u64 fraction[2], denominator;
1556 enum lru_list l;
1557 int noswap = 0;
1558
1559 /* If we have no swap space, do not bother scanning anon pages. */
1560 if (!sc->may_swap || (nr_swap_pages <= 0)) {
1561 noswap = 1;
1562 fraction[0] = 0;
1563 fraction[1] = 1;
1564 denominator = 1;
1565 goto out;
1566 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001567
Vincent Li0b217672009-09-21 17:03:09 -07001568 anon = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_ANON) +
1569 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_ANON);
1570 file = zone_nr_lru_pages(zone, sc, LRU_ACTIVE_FILE) +
1571 zone_nr_lru_pages(zone, sc, LRU_INACTIVE_FILE);
Hugh Dickinsb9627162009-01-06 14:39:41 -08001572
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001573 if (scanning_global_lru(sc)) {
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001574 free = zone_page_state(zone, NR_FREE_PAGES);
1575 /* If we have very few page cache pages,
1576 force-scan anon pages. */
Mel Gorman41858962009-06-16 15:32:12 -07001577 if (unlikely(file + free <= high_wmark_pages(zone))) {
Shaohua Li76a33fc2010-05-24 14:32:36 -07001578 fraction[0] = 1;
1579 fraction[1] = 0;
1580 denominator = 1;
1581 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001582 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001583 }
1584
1585 /*
1586 * OK, so we have swap space and a fair amount of page cache
1587 * pages. We use the recently rotated / recently scanned
1588 * ratios to determine how valuable each cache is.
1589 *
1590 * Because workloads change over time (and to avoid overflow)
1591 * we keep these statistics as a floating average, which ends
1592 * up weighing recent references more than old ones.
1593 *
1594 * anon in [0], file in [1]
1595 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001596 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001597 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001598 reclaim_stat->recent_scanned[0] /= 2;
1599 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001600 spin_unlock_irq(&zone->lru_lock);
1601 }
1602
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001603 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001604 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001605 reclaim_stat->recent_scanned[1] /= 2;
1606 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001607 spin_unlock_irq(&zone->lru_lock);
1608 }
1609
1610 /*
1611 * With swappiness at 100, anonymous and file have the same priority.
1612 * This scanning priority is essentially the inverse of IO cost.
1613 */
1614 anon_prio = sc->swappiness;
1615 file_prio = 200 - sc->swappiness;
1616
1617 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001618 * The amount of pressure on anon vs file pages is inversely
1619 * proportional to the fraction of recently scanned pages on
1620 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001621 */
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001622 ap = (anon_prio + 1) * (reclaim_stat->recent_scanned[0] + 1);
1623 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001624
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001625 fp = (file_prio + 1) * (reclaim_stat->recent_scanned[1] + 1);
1626 fp /= reclaim_stat->recent_rotated[1] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001627
Shaohua Li76a33fc2010-05-24 14:32:36 -07001628 fraction[0] = ap;
1629 fraction[1] = fp;
1630 denominator = ap + fp + 1;
1631out:
1632 for_each_evictable_lru(l) {
1633 int file = is_file_lru(l);
1634 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001635
Shaohua Li76a33fc2010-05-24 14:32:36 -07001636 scan = zone_nr_lru_pages(zone, sc, l);
1637 if (priority || noswap) {
1638 scan >>= priority;
1639 scan = div64_u64(scan * fraction[file], denominator);
1640 }
1641 nr[l] = nr_scan_try_batch(scan,
1642 &reclaim_stat->nr_saved_scan[l]);
1643 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001644}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001645
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -07001646static void set_lumpy_reclaim_mode(int priority, struct scan_control *sc)
1647{
1648 /*
1649 * If we need a large contiguous chunk of memory, or have
1650 * trouble getting a small set of contiguous pages, we
1651 * will reclaim both active and inactive pages.
1652 */
1653 if (sc->order > PAGE_ALLOC_COSTLY_ORDER)
1654 sc->lumpy_reclaim_mode = 1;
1655 else if (sc->order && priority < DEF_PRIORITY - 2)
1656 sc->lumpy_reclaim_mode = 1;
1657 else
1658 sc->lumpy_reclaim_mode = 0;
1659}
1660
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001661/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001662 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1663 */
Rik van Riela79311c2009-01-06 14:40:01 -08001664static void shrink_zone(int priority, struct zone *zone,
Andrew Morton05ff5132006-03-22 00:08:20 -08001665 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001666{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001667 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001668 unsigned long nr_to_scan;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001669 enum lru_list l;
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001670 unsigned long nr_reclaimed = sc->nr_reclaimed;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001671 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672
Shaohua Li76a33fc2010-05-24 14:32:36 -07001673 get_scan_count(zone, sc, nr, priority);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001674
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -07001675 set_lumpy_reclaim_mode(priority, sc);
1676
Rik van Riel556adec2008-10-18 20:26:34 -07001677 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
1678 nr[LRU_INACTIVE_FILE]) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001679 for_each_evictable_lru(l) {
Christoph Lameterb69408e2008-10-18 20:26:14 -07001680 if (nr[l]) {
KOSAKI Motohiroece74b22009-12-14 17:59:14 -08001681 nr_to_scan = min_t(unsigned long,
1682 nr[l], SWAP_CLUSTER_MAX);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001683 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001684
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001685 nr_reclaimed += shrink_list(l, nr_to_scan,
1686 zone, sc, priority);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001687 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001688 }
Rik van Riela79311c2009-01-06 14:40:01 -08001689 /*
1690 * On large memory systems, scan >> priority can become
1691 * really large. This is fine for the starting priority;
1692 * we want to put equal scanning pressure on each zone.
1693 * However, if the VM has a harder time of freeing pages,
1694 * with multiple processes reclaiming pages, the total
1695 * freeing target can get unreasonably large.
1696 */
KOSAKI Motohiro338fde92009-12-14 17:59:15 -08001697 if (nr_reclaimed >= nr_to_reclaim && priority < DEF_PRIORITY)
Rik van Riela79311c2009-01-06 14:40:01 -08001698 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 }
1700
KOSAKI Motohiro01dbe5c2009-01-06 14:40:02 -08001701 sc->nr_reclaimed = nr_reclaimed;
1702
Rik van Riel556adec2008-10-18 20:26:34 -07001703 /*
1704 * Even if we did not try to evict anon pages at all, we want to
1705 * rebalance the anon lru active/inactive ratio.
1706 */
MinChan Kim69c85482009-06-16 15:32:44 -07001707 if (inactive_anon_is_low(zone, sc) && nr_swap_pages > 0)
Rik van Riel556adec2008-10-18 20:26:34 -07001708 shrink_active_list(SWAP_CLUSTER_MAX, zone, sc, priority, 0);
1709
Andrew Morton232ea4d2007-02-28 20:13:21 -08001710 throttle_vm_writeout(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711}
1712
1713/*
1714 * This is the direct reclaim path, for page-allocating processes. We only
1715 * try to reclaim pages from zones which will satisfy the caller's allocation
1716 * request.
1717 *
Mel Gorman41858962009-06-16 15:32:12 -07001718 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
1719 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001720 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1721 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07001722 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
1723 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
1724 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001725 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001726 * If a zone is deemed to be full of pinned pages then just give it a light
1727 * scan then give up on it.
1728 */
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001729static bool shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001730 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001731{
Mel Gorman54a6eb52008-04-28 02:12:16 -07001732 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Mel Gormandd1a2392008-04-28 02:12:17 -07001733 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001734 struct zone *zone;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001735 bool all_unreclaimable = true;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001736
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001737 for_each_zone_zonelist_nodemask(zone, z, zonelist, high_zoneidx,
1738 sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001739 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001741 /*
1742 * Take care memory controller reclaiming has small influence
1743 * to global LRU.
1744 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001745 if (scanning_global_lru(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001746 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1747 continue;
1748 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001749
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08001750 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001751 continue; /* Let kswapd poll it */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001752 } else {
1753 /*
1754 * Ignore cpuset limitation here. We just want to reduce
1755 * # of used pages by us regardless of memory shortage.
1756 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001757 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1758 priority);
1759 }
Nick Piggin408d8542006-09-25 23:31:27 -07001760
Rik van Riela79311c2009-01-06 14:40:01 -08001761 shrink_zone(priority, zone, sc);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001762 all_unreclaimable = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001763 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001764 return all_unreclaimable;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001765}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001766
Linus Torvalds1da177e2005-04-16 15:20:36 -07001767/*
1768 * This is the main entry point to direct page reclaim.
1769 *
1770 * If a full scan of the inactive list fails to free enough memory then we
1771 * are "out of memory" and something needs to be killed.
1772 *
1773 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1774 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02001775 * caller can't do much about. We kick the writeback threads and take explicit
1776 * naps in the hope that some of these pages can be written. But if the
1777 * allocating task holds filesystem locks which prevent writeout this might not
1778 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001779 *
1780 * returns: 0, if no pages reclaimed
1781 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001782 */
Mel Gormandac1d272008-04-28 02:12:12 -07001783static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001784 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001785{
1786 int priority;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001787 bool all_unreclaimable;
Andrew Morton69e05942006-03-22 00:08:19 -08001788 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001790 unsigned long lru_pages = 0;
Mel Gormandd1a2392008-04-28 02:12:17 -07001791 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001792 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -07001793 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001794 unsigned long writeback_threshold;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795
Miao Xiec0ff7452010-05-24 14:32:08 -07001796 get_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07001797 delayacct_freepages_start();
1798
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001799 if (scanning_global_lru(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001800 count_vm_event(ALLOCSTALL);
1801 /*
1802 * mem_cgroup will not do shrink_slab.
1803 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001804 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001805 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001806
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001807 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1808 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001809
Wu Fengguangadea02a2009-09-21 17:01:42 -07001810 lru_pages += zone_reclaimable_pages(zone);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001811 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001812 }
1813
1814 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001815 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001816 if (!priority)
1817 disable_swap_token();
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001818 all_unreclaimable = shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001819 /*
1820 * Don't shrink slabs when reclaiming memory from
1821 * over limit cgroups
1822 */
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001823 if (scanning_global_lru(sc)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07001824 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001825 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08001826 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001827 reclaim_state->reclaimed_slab = 0;
1828 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001829 }
Balbir Singh66e17072008-02-07 00:13:56 -08001830 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001831 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001832 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833
1834 /*
1835 * Try to write back as many pages as we just scanned. This
1836 * tends to cause slow streaming writers to write data to the
1837 * disk smoothly, at the dirtying rate, which is nice. But
1838 * that's undesirable in laptop mode, where we *want* lumpy
1839 * writeout. So in laptop mode, write out the whole world.
1840 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001841 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
1842 if (total_scanned > writeback_threshold) {
Jens Axboe03ba3782009-09-09 09:08:54 +02001843 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001844 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001845 }
1846
1847 /* Take a nap, wait for some writeback to complete */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08001848 if (!sc->hibernation_mode && sc->nr_scanned &&
1849 priority < DEF_PRIORITY - 2)
Jens Axboe8aa7e842009-07-09 14:52:32 +02001850 congestion_wait(BLK_RW_ASYNC, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001851 }
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001852
Linus Torvalds1da177e2005-04-16 15:20:36 -07001853out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07001854 /*
1855 * Now that we've scanned all the zones at this priority level, note
1856 * that level within the zone so that the next thread which performs
1857 * scanning of this zone will immediately start out at this priority
1858 * level. This affects only the decision whether or not to bring
1859 * mapped pages onto the inactive list.
1860 */
1861 if (priority < 0)
1862 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001863
KAMEZAWA Hiroyukie72e2bd2009-01-07 18:08:23 -08001864 if (scanning_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001865 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001867 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1868 continue;
1869
1870 zone->prev_priority = priority;
1871 }
1872 } else
1873 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1874
Keika Kobayashi873b4772008-07-25 01:48:52 -07001875 delayacct_freepages_end();
Miao Xiec0ff7452010-05-24 14:32:08 -07001876 put_mems_allowed();
Keika Kobayashi873b4772008-07-25 01:48:52 -07001877
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07001878 if (sc->nr_reclaimed)
1879 return sc->nr_reclaimed;
1880
1881 /* top priority shrink_zones still had more to do? don't OOM, then */
1882 if (scanning_global_lru(sc) && !all_unreclaimable)
1883 return 1;
1884
1885 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886}
1887
Mel Gormandac1d272008-04-28 02:12:12 -07001888unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001889 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08001890{
1891 struct scan_control sc = {
1892 .gfp_mask = gfp_mask,
1893 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001894 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001895 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001896 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08001897 .swappiness = vm_swappiness,
1898 .order = order,
1899 .mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001900 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08001901 };
1902
Mel Gormandd1a2392008-04-28 02:12:17 -07001903 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001904}
1905
Balbir Singh00f0b822008-03-04 14:28:39 -08001906#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08001907
Balbir Singh4e416952009-09-23 15:56:39 -07001908unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *mem,
1909 gfp_t gfp_mask, bool noswap,
1910 unsigned int swappiness,
1911 struct zone *zone, int nid)
1912{
1913 struct scan_control sc = {
1914 .may_writepage = !laptop_mode,
1915 .may_unmap = 1,
1916 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07001917 .swappiness = swappiness,
1918 .order = 0,
1919 .mem_cgroup = mem,
Balbir Singh4e416952009-09-23 15:56:39 -07001920 };
1921 nodemask_t nm = nodemask_of_node(nid);
1922
1923 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1924 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1925 sc.nodemask = &nm;
1926 sc.nr_reclaimed = 0;
1927 sc.nr_scanned = 0;
1928 /*
1929 * NOTE: Although we can get the priority field, using it
1930 * here is not a good idea, since it limits the pages we can scan.
1931 * if we don't reclaim here, the shrink_zone from balance_pgdat
1932 * will pick up pages from other mem cgroup's as well. We hack
1933 * the priority and make it zero.
1934 */
1935 shrink_zone(0, zone, &sc);
1936 return sc.nr_reclaimed;
1937}
1938
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001939unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001940 gfp_t gfp_mask,
1941 bool noswap,
1942 unsigned int swappiness)
Balbir Singh66e17072008-02-07 00:13:56 -08001943{
Balbir Singh4e416952009-09-23 15:56:39 -07001944 struct zonelist *zonelist;
Balbir Singh66e17072008-02-07 00:13:56 -08001945 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08001946 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07001947 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07001948 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08001949 .nr_to_reclaim = SWAP_CLUSTER_MAX,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08001950 .swappiness = swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08001951 .order = 0,
1952 .mem_cgroup = mem_cont,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07001953 .nodemask = NULL, /* we don't care the placement */
Balbir Singh66e17072008-02-07 00:13:56 -08001954 };
Balbir Singh66e17072008-02-07 00:13:56 -08001955
Mel Gormandd1a2392008-04-28 02:12:17 -07001956 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1957 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1958 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1959 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001960}
1961#endif
1962
Mel Gormanf50de2d2009-12-14 17:58:53 -08001963/* is kswapd sleeping prematurely? */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001964static int sleeping_prematurely(pg_data_t *pgdat, int order, long remaining)
Mel Gormanf50de2d2009-12-14 17:58:53 -08001965{
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001966 int i;
Mel Gormanf50de2d2009-12-14 17:58:53 -08001967
1968 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
1969 if (remaining)
1970 return 1;
1971
1972 /* If after HZ/10, a zone is below the high mark, it's premature */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001973 for (i = 0; i < pgdat->nr_zones; i++) {
1974 struct zone *zone = pgdat->node_zones + i;
1975
1976 if (!populated_zone(zone))
1977 continue;
1978
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08001979 if (zone->all_unreclaimable)
KOSAKI Motohirode3fab32010-01-15 17:01:25 -08001980 continue;
1981
Mel Gormanf50de2d2009-12-14 17:58:53 -08001982 if (!zone_watermark_ok(zone, order, high_wmark_pages(zone),
1983 0, 0))
1984 return 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08001985 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08001986
1987 return 0;
1988}
1989
Linus Torvalds1da177e2005-04-16 15:20:36 -07001990/*
1991 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07001992 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07001993 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 * Returns the number of pages which were actually freed.
1995 *
1996 * There is special handling here for zones which are full of pinned pages.
1997 * This can happen if the pages are all mlocked, or if they are all used by
1998 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1999 * What we do is to detect the case where all pages in the zone have been
2000 * scanned twice and there has been zero successful reclaim. Mark the zone as
2001 * dead and from now on, only perform a short scan. Basically we're polling
2002 * the zone for when the problem goes away.
2003 *
2004 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002005 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2006 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2007 * lower zones regardless of the number of free pages in the lower zones. This
2008 * interoperates with the page allocator fallback scheme to ensure that aging
2009 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002010 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002011static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002012{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013 int all_zones_ok;
2014 int priority;
2015 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08002016 unsigned long total_scanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002017 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08002018 struct scan_control sc = {
2019 .gfp_mask = GFP_KERNEL,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002020 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002021 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002022 /*
2023 * kswapd doesn't want to be bailed out while reclaim. because
2024 * we want to put equal scanning pressure on each zone.
2025 */
2026 .nr_to_reclaim = ULONG_MAX,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002027 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002028 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08002029 .mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002030 };
Martin Bligh3bb1a852006-10-28 10:38:24 -07002031 /*
2032 * temp_priority is used to remember the scanning priority at which
Mel Gorman41858962009-06-16 15:32:12 -07002033 * this zone was successfully refilled to
2034 * free_pages == high_wmark_pages(zone).
Martin Bligh3bb1a852006-10-28 10:38:24 -07002035 */
2036 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002037
2038loop_again:
2039 total_scanned = 0;
Rik van Riela79311c2009-01-06 14:40:01 -08002040 sc.nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07002041 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07002042 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002043
Martin Bligh3bb1a852006-10-28 10:38:24 -07002044 for (i = 0; i < pgdat->nr_zones; i++)
2045 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002046
2047 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
2048 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
2049 unsigned long lru_pages = 0;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002050 int has_under_min_watermark_zone = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002051
Rik van Rielf7b7fd82005-11-28 13:44:07 -08002052 /* The swap token gets in the way of swapout... */
2053 if (!priority)
2054 disable_swap_token();
2055
Linus Torvalds1da177e2005-04-16 15:20:36 -07002056 all_zones_ok = 1;
2057
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002058 /*
2059 * Scan in the highmem->dma direction for the highest
2060 * zone which needs scanning
2061 */
2062 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2063 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002064
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002065 if (!populated_zone(zone))
2066 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002067
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002068 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002069 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002070
Rik van Riel556adec2008-10-18 20:26:34 -07002071 /*
2072 * Do some background aging of the anon list, to give
2073 * pages a chance to be referenced before reclaiming.
2074 */
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08002075 if (inactive_anon_is_low(zone, &sc))
Rik van Riel556adec2008-10-18 20:26:34 -07002076 shrink_active_list(SWAP_CLUSTER_MAX, zone,
2077 &sc, priority, 0);
2078
Mel Gorman41858962009-06-16 15:32:12 -07002079 if (!zone_watermark_ok(zone, order,
2080 high_wmark_pages(zone), 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002081 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002082 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002083 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002084 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002085 if (i < 0)
2086 goto out;
2087
Linus Torvalds1da177e2005-04-16 15:20:36 -07002088 for (i = 0; i <= end_zone; i++) {
2089 struct zone *zone = pgdat->node_zones + i;
2090
Wu Fengguangadea02a2009-09-21 17:01:42 -07002091 lru_pages += zone_reclaimable_pages(zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002092 }
2093
2094 /*
2095 * Now scan the zone in the dma->highmem direction, stopping
2096 * at the last zone which needs scanning.
2097 *
2098 * We do this because the page allocator works in the opposite
2099 * direction. This prevents the page allocator from allocating
2100 * pages behind kswapd's direction of progress, which would
2101 * cause too much scanning of the lower zones.
2102 */
2103 for (i = 0; i <= end_zone; i++) {
2104 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002105 int nr_slab;
Balbir Singh4e416952009-09-23 15:56:39 -07002106 int nid, zid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002107
Con Kolivasf3fe6512006-01-06 00:11:15 -08002108 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002109 continue;
2110
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002111 if (zone->all_unreclaimable && priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112 continue;
2113
Martin Bligh3bb1a852006-10-28 10:38:24 -07002114 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002115 sc.nr_scanned = 0;
Martin Bligh3bb1a852006-10-28 10:38:24 -07002116 note_zone_scanning_priority(zone, priority);
Balbir Singh4e416952009-09-23 15:56:39 -07002117
2118 nid = pgdat->node_id;
2119 zid = zone_idx(zone);
2120 /*
2121 * Call soft limit reclaim before calling shrink_zone.
2122 * For now we ignore the return value
2123 */
2124 mem_cgroup_soft_limit_reclaim(zone, order, sc.gfp_mask,
2125 nid, zid);
Rik van Riel32a43302007-10-16 01:24:50 -07002126 /*
2127 * We put equal pressure on every zone, unless one
2128 * zone has way too many pages free already.
2129 */
Mel Gorman41858962009-06-16 15:32:12 -07002130 if (!zone_watermark_ok(zone, order,
2131 8*high_wmark_pages(zone), end_zone, 0))
Rik van Riela79311c2009-01-06 14:40:01 -08002132 shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002133 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07002134 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
2135 lru_pages);
Rik van Riela79311c2009-01-06 14:40:01 -08002136 sc.nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002137 total_scanned += sc.nr_scanned;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002138 if (zone->all_unreclaimable)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002139 continue;
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002140 if (nr_slab == 0 &&
2141 zone->pages_scanned >= (zone_reclaimable_pages(zone) * 6))
2142 zone->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002143 /*
2144 * If we've done a decent amount of scanning and
2145 * the reclaim ratio is low, start doing writepage
2146 * even in laptop mode
2147 */
2148 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Rik van Riela79311c2009-01-06 14:40:01 -08002149 total_scanned > sc.nr_reclaimed + sc.nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002150 sc.may_writepage = 1;
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002151
Minchan Kim45973d72010-03-05 13:41:45 -08002152 if (!zone_watermark_ok(zone, order,
2153 high_wmark_pages(zone), end_zone, 0)) {
2154 all_zones_ok = 0;
2155 /*
2156 * We are still under min water mark. This
2157 * means that we have a GFP_ATOMIC allocation
2158 * failure risk. Hurry up!
2159 */
2160 if (!zone_watermark_ok(zone, order,
2161 min_wmark_pages(zone), end_zone, 0))
2162 has_under_min_watermark_zone = 1;
2163 }
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002164
Linus Torvalds1da177e2005-04-16 15:20:36 -07002165 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002166 if (all_zones_ok)
2167 break; /* kswapd: all done */
2168 /*
2169 * OK, kswapd is getting into trouble. Take a nap, then take
2170 * another pass across the zones.
2171 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002172 if (total_scanned && (priority < DEF_PRIORITY - 2)) {
2173 if (has_under_min_watermark_zone)
2174 count_vm_event(KSWAPD_SKIP_CONGESTION_WAIT);
2175 else
2176 congestion_wait(BLK_RW_ASYNC, HZ/10);
2177 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002178
2179 /*
2180 * We do this so kswapd doesn't build up large priorities for
2181 * example when it is freeing in parallel with allocators. It
2182 * matches the direct reclaim path behaviour in terms of impact
2183 * on zone->*_priority.
2184 */
Rik van Riela79311c2009-01-06 14:40:01 -08002185 if (sc.nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002186 break;
2187 }
2188out:
Martin Bligh3bb1a852006-10-28 10:38:24 -07002189 /*
2190 * Note within each zone the priority level at which this zone was
2191 * brought into a happy state. So that the next thread which scans this
2192 * zone will start out at that priority level.
2193 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002194 for (i = 0; i < pgdat->nr_zones; i++) {
2195 struct zone *zone = pgdat->node_zones + i;
2196
Martin Bligh3bb1a852006-10-28 10:38:24 -07002197 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07002198 }
2199 if (!all_zones_ok) {
2200 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08002201
2202 try_to_freeze();
2203
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08002204 /*
2205 * Fragmentation may mean that the system cannot be
2206 * rebalanced for high-order allocations in all zones.
2207 * At this point, if nr_reclaimed < SWAP_CLUSTER_MAX,
2208 * it means the zones have been fully scanned and are still
2209 * not balanced. For high-order allocations, there is
2210 * little point trying all over again as kswapd may
2211 * infinite loop.
2212 *
2213 * Instead, recheck all watermarks at order-0 as they
2214 * are the most important. If watermarks are ok, kswapd will go
2215 * back to sleep. High-order users can still perform direct
2216 * reclaim if they wish.
2217 */
2218 if (sc.nr_reclaimed < SWAP_CLUSTER_MAX)
2219 order = sc.order = 0;
2220
Linus Torvalds1da177e2005-04-16 15:20:36 -07002221 goto loop_again;
2222 }
2223
Rik van Riela79311c2009-01-06 14:40:01 -08002224 return sc.nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002225}
2226
2227/*
2228 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002229 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002230 *
2231 * This basically trickles out pages so that we have _some_
2232 * free memory available even if there is no other activity
2233 * that frees anything up. This is needed for things like routing
2234 * etc, where we otherwise might have all activity going on in
2235 * asynchronous contexts that cannot page things out.
2236 *
2237 * If there are applications that are active memory-allocators
2238 * (most normal use), this basically shouldn't matter.
2239 */
2240static int kswapd(void *p)
2241{
2242 unsigned long order;
2243 pg_data_t *pgdat = (pg_data_t*)p;
2244 struct task_struct *tsk = current;
2245 DEFINE_WAIT(wait);
2246 struct reclaim_state reclaim_state = {
2247 .reclaimed_slab = 0,
2248 };
Rusty Russella70f7302009-03-13 14:49:46 +10302249 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002250
Nick Piggincf40bd12009-01-21 08:12:39 +01002251 lockdep_set_current_reclaim_state(GFP_KERNEL);
2252
Rusty Russell174596a2009-01-01 10:12:29 +10302253 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07002254 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255 current->reclaim_state = &reclaim_state;
2256
2257 /*
2258 * Tell the memory management that we're a "memory allocator",
2259 * and that if we need more memory we should get access to it
2260 * regardless (see "__alloc_pages()"). "kswapd" should
2261 * never get caught in the normal page freeing logic.
2262 *
2263 * (Kswapd normally doesn't need memory anyway, but sometimes
2264 * you need a small amount of memory in order to be able to
2265 * page out something else, and this flag essentially protects
2266 * us from recursively trying to free more memory as we're
2267 * trying to free the first piece of memory in the first place).
2268 */
Christoph Lameter930d9152006-01-08 01:00:47 -08002269 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07002270 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002271
2272 order = 0;
2273 for ( ; ; ) {
2274 unsigned long new_order;
David Rientjes8fe23e02009-12-14 17:58:33 -08002275 int ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07002276
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2278 new_order = pgdat->kswapd_max_order;
2279 pgdat->kswapd_max_order = 0;
2280 if (order < new_order) {
2281 /*
2282 * Don't sleep if someone wants a larger 'order'
2283 * allocation
2284 */
2285 order = new_order;
2286 } else {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002287 if (!freezing(current) && !kthread_should_stop()) {
2288 long remaining = 0;
2289
2290 /* Try to sleep for a short interval */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002291 if (!sleeping_prematurely(pgdat, order, remaining)) {
Mel Gormanf50de2d2009-12-14 17:58:53 -08002292 remaining = schedule_timeout(HZ/10);
2293 finish_wait(&pgdat->kswapd_wait, &wait);
2294 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
2295 }
2296
2297 /*
2298 * After a short sleep, check if it was a
2299 * premature sleep. If not, then go fully
2300 * to sleep until explicitly woken up
2301 */
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002302 if (!sleeping_prematurely(pgdat, order, remaining))
Mel Gormanf50de2d2009-12-14 17:58:53 -08002303 schedule();
2304 else {
2305 if (remaining)
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002306 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002307 else
KOSAKI Motohirobb3ab592009-12-14 17:58:55 -08002308 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002309 }
2310 }
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002311
Linus Torvalds1da177e2005-04-16 15:20:36 -07002312 order = pgdat->kswapd_max_order;
2313 }
2314 finish_wait(&pgdat->kswapd_wait, &wait);
2315
David Rientjes8fe23e02009-12-14 17:58:33 -08002316 ret = try_to_freeze();
2317 if (kthread_should_stop())
2318 break;
2319
2320 /*
2321 * We can speed up thawing tasks if we don't call balance_pgdat
2322 * after returning from the refrigerator
2323 */
2324 if (!ret)
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07002325 balance_pgdat(pgdat, order);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326 }
2327 return 0;
2328}
2329
2330/*
2331 * A zone is low on free memory, so wake its kswapd task to service it.
2332 */
2333void wakeup_kswapd(struct zone *zone, int order)
2334{
2335 pg_data_t *pgdat;
2336
Con Kolivasf3fe6512006-01-06 00:11:15 -08002337 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 return;
2339
2340 pgdat = zone->zone_pgdat;
Mel Gorman41858962009-06-16 15:32:12 -07002341 if (zone_watermark_ok(zone, order, low_wmark_pages(zone), 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342 return;
2343 if (pgdat->kswapd_max_order < order)
2344 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08002345 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002347 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002348 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07002349 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002350}
2351
Wu Fengguangadea02a2009-09-21 17:01:42 -07002352/*
2353 * The reclaimable count would be mostly accurate.
2354 * The less reclaimable pages may be
2355 * - mlocked pages, which will be moved to unevictable list when encountered
2356 * - mapped pages, which may require several travels to be reclaimed
2357 * - dirty pages, which is not "instantly" reclaimable
2358 */
2359unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002360{
Wu Fengguangadea02a2009-09-21 17:01:42 -07002361 int nr;
2362
2363 nr = global_page_state(NR_ACTIVE_FILE) +
2364 global_page_state(NR_INACTIVE_FILE);
2365
2366 if (nr_swap_pages > 0)
2367 nr += global_page_state(NR_ACTIVE_ANON) +
2368 global_page_state(NR_INACTIVE_ANON);
2369
2370 return nr;
2371}
2372
2373unsigned long zone_reclaimable_pages(struct zone *zone)
2374{
2375 int nr;
2376
2377 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
2378 zone_page_state(zone, NR_INACTIVE_FILE);
2379
2380 if (nr_swap_pages > 0)
2381 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
2382 zone_page_state(zone, NR_INACTIVE_ANON);
2383
2384 return nr;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002385}
2386
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002387#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002389 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002390 * freed pages.
2391 *
2392 * Rather than trying to age LRUs the aim is to preserve the overall
2393 * LRU order by reclaiming preferentially
2394 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002396unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002397{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002398 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002399 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002400 .gfp_mask = GFP_HIGHUSER_MOVABLE,
2401 .may_swap = 1,
2402 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002403 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002404 .nr_to_reclaim = nr_to_reclaim,
2405 .hibernation_mode = 1,
2406 .swappiness = vm_swappiness,
2407 .order = 0,
Linus Torvalds1da177e2005-04-16 15:20:36 -07002408 };
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002409 struct zonelist * zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
2410 struct task_struct *p = current;
2411 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002412
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002413 p->flags |= PF_MEMALLOC;
2414 lockdep_set_current_reclaim_state(sc.gfp_mask);
2415 reclaim_state.reclaimed_slab = 0;
2416 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08002417
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002418 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002419
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002420 p->reclaim_state = NULL;
2421 lockdep_clear_current_reclaim_state();
2422 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002423
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08002424 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02002426#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002427
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428/* It's optimal to keep kswapds on the same CPUs as their memory, but
2429 not required for correctness. So if the last cpu in a node goes
2430 away, we get changed to run anywhere: as the first one comes back,
2431 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07002432static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08002433 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002434{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002435 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002436
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07002437 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07002438 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07002439 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10302440 const struct cpumask *mask;
2441
2442 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07002443
Rusty Russell3e597942009-01-01 10:12:24 +10302444 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002445 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07002446 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447 }
2448 }
2449 return NOTIFY_OK;
2450}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002451
Yasunori Goto3218ae12006-06-27 02:53:33 -07002452/*
2453 * This kswapd start function will be called by init and node-hot-add.
2454 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
2455 */
2456int kswapd_run(int nid)
2457{
2458 pg_data_t *pgdat = NODE_DATA(nid);
2459 int ret = 0;
2460
2461 if (pgdat->kswapd)
2462 return 0;
2463
2464 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
2465 if (IS_ERR(pgdat->kswapd)) {
2466 /* failure at boot is fatal */
2467 BUG_ON(system_state == SYSTEM_BOOTING);
2468 printk("Failed to start kswapd on node %d\n",nid);
2469 ret = -1;
2470 }
2471 return ret;
2472}
2473
David Rientjes8fe23e02009-12-14 17:58:33 -08002474/*
2475 * Called by memory hotplug when all memory in a node is offlined.
2476 */
2477void kswapd_stop(int nid)
2478{
2479 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
2480
2481 if (kswapd)
2482 kthread_stop(kswapd);
2483}
2484
Linus Torvalds1da177e2005-04-16 15:20:36 -07002485static int __init kswapd_init(void)
2486{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002487 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002488
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002490 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002491 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002492 hotcpu_notifier(cpu_callback, 0);
2493 return 0;
2494}
2495
2496module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002497
2498#ifdef CONFIG_NUMA
2499/*
2500 * Zone reclaim mode
2501 *
2502 * If non-zero call zone_reclaim when the number of free pages falls below
2503 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002504 */
2505int zone_reclaim_mode __read_mostly;
2506
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002507#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002508#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002509#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2510#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2511
Christoph Lameter9eeff232006-01-18 17:42:31 -08002512/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002513 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2514 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2515 * a zone.
2516 */
2517#define ZONE_RECLAIM_PRIORITY 4
2518
Christoph Lameter9eeff232006-01-18 17:42:31 -08002519/*
Christoph Lameter96146342006-07-03 00:24:13 -07002520 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2521 * occur.
2522 */
2523int sysctl_min_unmapped_ratio = 1;
2524
2525/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002526 * If the number of slab pages in a zone grows beyond this percentage then
2527 * slab reclaim needs to occur.
2528 */
2529int sysctl_min_slab_ratio = 5;
2530
Mel Gorman90afa5d2009-06-16 15:33:20 -07002531static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
2532{
2533 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
2534 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
2535 zone_page_state(zone, NR_ACTIVE_FILE);
2536
2537 /*
2538 * It's possible for there to be more file mapped pages than
2539 * accounted for by the pages on the file LRU lists because
2540 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
2541 */
2542 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
2543}
2544
2545/* Work out how many page cache pages we can reclaim in this reclaim_mode */
2546static long zone_pagecache_reclaimable(struct zone *zone)
2547{
2548 long nr_pagecache_reclaimable;
2549 long delta = 0;
2550
2551 /*
2552 * If RECLAIM_SWAP is set, then all file pages are considered
2553 * potentially reclaimable. Otherwise, we have to worry about
2554 * pages like swapcache and zone_unmapped_file_pages() provides
2555 * a better estimate
2556 */
2557 if (zone_reclaim_mode & RECLAIM_SWAP)
2558 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
2559 else
2560 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
2561
2562 /* If we can't clean pages, remove dirty pages from consideration */
2563 if (!(zone_reclaim_mode & RECLAIM_WRITE))
2564 delta += zone_page_state(zone, NR_FILE_DIRTY);
2565
2566 /* Watch for any possible underflows due to delta */
2567 if (unlikely(delta > nr_pagecache_reclaimable))
2568 delta = nr_pagecache_reclaimable;
2569
2570 return nr_pagecache_reclaimable - delta;
2571}
2572
Christoph Lameter0ff38492006-09-25 23:31:52 -07002573/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002574 * Try to free up some pages from this zone through reclaim.
2575 */
Andrew Morton179e9632006-03-22 00:08:18 -08002576static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002577{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002578 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002579 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002580 struct task_struct *p = current;
2581 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002582 int priority;
Andrew Morton179e9632006-03-22 00:08:18 -08002583 struct scan_control sc = {
2584 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002585 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002586 .may_swap = 1,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002587 .nr_to_reclaim = max_t(unsigned long, nr_pages,
2588 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002589 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002590 .swappiness = vm_swappiness,
Johannes Weinerbd2f6192009-03-31 15:19:38 -07002591 .order = order,
Andrew Morton179e9632006-03-22 00:08:18 -08002592 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07002593 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002594
2595 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08002596 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002597 /*
2598 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2599 * and we also need to be able to write out pages for RECLAIM_WRITE
2600 * and RECLAIM_SWAP.
2601 */
2602 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002603 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08002604 reclaim_state.reclaimed_slab = 0;
2605 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002606
Mel Gorman90afa5d2009-06-16 15:33:20 -07002607 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07002608 /*
2609 * Free memory by calling shrink zone with increasing
2610 * priorities until we have enough memory freed.
2611 */
2612 priority = ZONE_RECLAIM_PRIORITY;
2613 do {
Martin Bligh3bb1a852006-10-28 10:38:24 -07002614 note_zone_scanning_priority(zone, priority);
Rik van Riela79311c2009-01-06 14:40:01 -08002615 shrink_zone(priority, zone, &sc);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002616 priority--;
Rik van Riela79311c2009-01-06 14:40:01 -08002617 } while (priority >= 0 && sc.nr_reclaimed < nr_pages);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002618 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002619
Christoph Lameter83e33a42006-09-25 23:31:53 -07002620 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2621 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002622 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002623 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002624 * many pages were freed in this zone. So we take the current
2625 * number of slab pages and shake the slab until it is reduced
2626 * by the same nr_pages that we used for reclaiming unmapped
2627 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002628 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002629 * Note that shrink_slab will free memory on all zones and may
2630 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002631 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07002632 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07002633 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2634 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07002635 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07002636
2637 /*
2638 * Update nr_reclaimed by the number of slab pages we
2639 * reclaimed from this zone.
2640 */
Rik van Riela79311c2009-01-06 14:40:01 -08002641 sc.nr_reclaimed += slab_reclaimable -
Christoph Lameter83e33a42006-09-25 23:31:53 -07002642 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002643 }
2644
Christoph Lameter9eeff232006-01-18 17:42:31 -08002645 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002646 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08002647 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08002648 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002649}
Andrew Morton179e9632006-03-22 00:08:18 -08002650
2651int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2652{
Andrew Morton179e9632006-03-22 00:08:18 -08002653 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002654 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002655
2656 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002657 * Zone reclaim reclaims unmapped file backed pages and
2658 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002659 *
Christoph Lameter96146342006-07-03 00:24:13 -07002660 * A small portion of unmapped file backed pages is needed for
2661 * file I/O otherwise pages read by file I/O will be immediately
2662 * thrown out if the zone is overallocated. So we do not reclaim
2663 * if less than a specified percentage of the zone is used by
2664 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002665 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07002666 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
2667 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002668 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08002669
KOSAKI Motohiro93e4a892010-03-05 13:41:55 -08002670 if (zone->all_unreclaimable)
Mel Gormanfa5e0842009-06-16 15:33:22 -07002671 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07002672
Andrew Morton179e9632006-03-22 00:08:18 -08002673 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002674 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002675 */
David Rientjesd773ed62007-10-16 23:26:01 -07002676 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002677 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08002678
2679 /*
2680 * Only run zone reclaim on the local zone or on zones that do not
2681 * have associated processors. This will favor the local processor
2682 * over remote processors and spread off node memory allocations
2683 * as wide as possible.
2684 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002685 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002686 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07002687 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07002688
2689 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07002690 return ZONE_RECLAIM_NOSCAN;
2691
David Rientjesd773ed62007-10-16 23:26:01 -07002692 ret = __zone_reclaim(zone, gfp_mask, order);
2693 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2694
Mel Gorman24cf725182009-06-16 15:33:23 -07002695 if (!ret)
2696 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
2697
David Rientjesd773ed62007-10-16 23:26:01 -07002698 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002699}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002700#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002701
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002702/*
2703 * page_evictable - test whether a page is evictable
2704 * @page: the page to test
2705 * @vma: the VMA in which the page is or will be mapped, may be NULL
2706 *
2707 * Test whether page is evictable--i.e., should be placed on active/inactive
Nick Pigginb291f002008-10-18 20:26:44 -07002708 * lists vs unevictable list. The vma argument is !NULL when called from the
2709 * fault path to determine how to instantate a new page.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002710 *
2711 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002712 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07002713 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002714 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002715 */
2716int page_evictable(struct page *page, struct vm_area_struct *vma)
2717{
2718
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07002719 if (mapping_unevictable(page_mapping(page)))
2720 return 0;
2721
Nick Pigginb291f002008-10-18 20:26:44 -07002722 if (PageMlocked(page) || (vma && is_mlocked_vma(vma, page)))
2723 return 0;
Lee Schermerhorn894bc312008-10-18 20:26:39 -07002724
2725 return 1;
2726}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002727
2728/**
2729 * check_move_unevictable_page - check page for evictability and move to appropriate zone lru list
2730 * @page: page to check evictability and move to appropriate lru list
2731 * @zone: zone page is in
2732 *
2733 * Checks a page for evictability and moves the page to the appropriate
2734 * zone lru list.
2735 *
2736 * Restrictions: zone->lru_lock must be held, page must be on LRU and must
2737 * have PageUnevictable set.
2738 */
2739static void check_move_unevictable_page(struct page *page, struct zone *zone)
2740{
2741 VM_BUG_ON(PageActive(page));
2742
2743retry:
2744 ClearPageUnevictable(page);
2745 if (page_evictable(page, NULL)) {
Johannes Weiner401a8e12009-09-21 17:02:58 -07002746 enum lru_list l = page_lru_base_type(page);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002747
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002748 __dec_zone_state(zone, NR_UNEVICTABLE);
2749 list_move(&page->lru, &zone->lru[l].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002750 mem_cgroup_move_lists(page, LRU_UNEVICTABLE, l);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002751 __inc_zone_state(zone, NR_INACTIVE_ANON + l);
2752 __count_vm_event(UNEVICTABLE_PGRESCUED);
2753 } else {
2754 /*
2755 * rotate unevictable list
2756 */
2757 SetPageUnevictable(page);
2758 list_move(&page->lru, &zone->lru[LRU_UNEVICTABLE].list);
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08002759 mem_cgroup_rotate_lru_list(page, LRU_UNEVICTABLE);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07002760 if (page_evictable(page, NULL))
2761 goto retry;
2762 }
2763}
2764
2765/**
2766 * scan_mapping_unevictable_pages - scan an address space for evictable pages
2767 * @mapping: struct address_space to scan for evictable pages
2768 *
2769 * Scan all pages in mapping. Check unevictable pages for
2770 * evictability and move them to the appropriate zone lru list.
2771 */
2772void scan_mapping_unevictable_pages(struct address_space *mapping)
2773{
2774 pgoff_t next = 0;
2775 pgoff_t end = (i_size_read(mapping->host) + PAGE_CACHE_SIZE - 1) >>
2776 PAGE_CACHE_SHIFT;
2777 struct zone *zone;
2778 struct pagevec pvec;
2779
2780 if (mapping->nrpages == 0)
2781 return;
2782
2783 pagevec_init(&pvec, 0);
2784 while (next < end &&
2785 pagevec_lookup(&pvec, mapping, next, PAGEVEC_SIZE)) {
2786 int i;
2787 int pg_scanned = 0;
2788
2789 zone = NULL;
2790
2791 for (i = 0; i < pagevec_count(&pvec); i++) {
2792 struct page *page = pvec.pages[i];
2793 pgoff_t page_index = page->index;
2794 struct zone *pagezone = page_zone(page);
2795
2796 pg_scanned++;
2797 if (page_index > next)
2798 next = page_index;
2799 next++;
2800
2801 if (pagezone != zone) {
2802 if (zone)
2803 spin_unlock_irq(&zone->lru_lock);
2804 zone = pagezone;
2805 spin_lock_irq(&zone->lru_lock);
2806 }
2807
2808 if (PageLRU(page) && PageUnevictable(page))
2809 check_move_unevictable_page(page, zone);
2810 }
2811 if (zone)
2812 spin_unlock_irq(&zone->lru_lock);
2813 pagevec_release(&pvec);
2814
2815 count_vm_events(UNEVICTABLE_PGSCANNED, pg_scanned);
2816 }
2817
2818}
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002819
2820/**
2821 * scan_zone_unevictable_pages - check unevictable list for evictable pages
2822 * @zone - zone of which to scan the unevictable list
2823 *
2824 * Scan @zone's unevictable LRU lists to check for pages that have become
2825 * evictable. Move those that have to @zone's inactive list where they
2826 * become candidates for reclaim, unless shrink_inactive_zone() decides
2827 * to reactivate them. Pages that are still unevictable are rotated
2828 * back onto @zone's unevictable list.
2829 */
2830#define SCAN_UNEVICTABLE_BATCH_SIZE 16UL /* arbitrary lock hold batch size */
KOSAKI Motohiro14b90b22009-01-06 14:39:45 -08002831static void scan_zone_unevictable_pages(struct zone *zone)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002832{
2833 struct list_head *l_unevictable = &zone->lru[LRU_UNEVICTABLE].list;
2834 unsigned long scan;
2835 unsigned long nr_to_scan = zone_page_state(zone, NR_UNEVICTABLE);
2836
2837 while (nr_to_scan > 0) {
2838 unsigned long batch_size = min(nr_to_scan,
2839 SCAN_UNEVICTABLE_BATCH_SIZE);
2840
2841 spin_lock_irq(&zone->lru_lock);
2842 for (scan = 0; scan < batch_size; scan++) {
2843 struct page *page = lru_to_page(l_unevictable);
2844
2845 if (!trylock_page(page))
2846 continue;
2847
2848 prefetchw_prev_lru_page(page, l_unevictable, flags);
2849
2850 if (likely(PageLRU(page) && PageUnevictable(page)))
2851 check_move_unevictable_page(page, zone);
2852
2853 unlock_page(page);
2854 }
2855 spin_unlock_irq(&zone->lru_lock);
2856
2857 nr_to_scan -= batch_size;
2858 }
2859}
2860
2861
2862/**
2863 * scan_all_zones_unevictable_pages - scan all unevictable lists for evictable pages
2864 *
2865 * A really big hammer: scan all zones' unevictable LRU lists to check for
2866 * pages that have become evictable. Move those back to the zones'
2867 * inactive list where they become candidates for reclaim.
2868 * This occurs when, e.g., we have unswappable pages on the unevictable lists,
2869 * and we add swap to the system. As such, it runs in the context of a task
2870 * that has possibly/probably made some previously unevictable pages
2871 * evictable.
2872 */
KOSAKI Motohiroff301532009-01-06 14:39:44 -08002873static void scan_all_zones_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002874{
2875 struct zone *zone;
2876
2877 for_each_zone(zone) {
2878 scan_zone_unevictable_pages(zone);
2879 }
2880}
2881
2882/*
2883 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
2884 * all nodes' unevictable lists for evictable pages
2885 */
2886unsigned long scan_unevictable_pages;
2887
2888int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002889 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002890 size_t *length, loff_t *ppos)
2891{
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07002892 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07002893
2894 if (write && *(unsigned long *)table->data)
2895 scan_all_zones_unevictable_pages();
2896
2897 scan_unevictable_pages = 0;
2898 return 0;
2899}
2900
2901/*
2902 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
2903 * a specified node's per zone unevictable lists for evictable pages.
2904 */
2905
2906static ssize_t read_scan_unevictable_node(struct sys_device *dev,
2907 struct sysdev_attribute *attr,
2908 char *buf)
2909{
2910 return sprintf(buf, "0\n"); /* always zero; should fit... */
2911}
2912
2913static ssize_t write_scan_unevictable_node(struct sys_device *dev,
2914 struct sysdev_attribute *attr,
2915 const char *buf, size_t count)
2916{
2917 struct zone *node_zones = NODE_DATA(dev->id)->node_zones;
2918 struct zone *zone;
2919 unsigned long res;
2920 unsigned long req = strict_strtoul(buf, 10, &res);
2921
2922 if (!req)
2923 return 1; /* zero is no-op */
2924
2925 for (zone = node_zones; zone - node_zones < MAX_NR_ZONES; ++zone) {
2926 if (!populated_zone(zone))
2927 continue;
2928 scan_zone_unevictable_pages(zone);
2929 }
2930 return 1;
2931}
2932
2933
2934static SYSDEV_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
2935 read_scan_unevictable_node,
2936 write_scan_unevictable_node);
2937
2938int scan_unevictable_register_node(struct node *node)
2939{
2940 return sysdev_create_file(&node->sysdev, &attr_scan_unevictable_pages);
2941}
2942
2943void scan_unevictable_unregister_node(struct node *node)
2944{
2945 sysdev_remove_file(&node->sysdev, &attr_scan_unevictable_pages);
2946}
2947