blob: 4046434046e68b0a45864c019a095fab088420ee [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
16#include <linux/slab.h>
17#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070022#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/file.h>
24#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28#include <linux/mm_inline.h>
29#include <linux/pagevec.h>
30#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
35#include <linux/notifier.h>
36#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080037#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070038#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080039#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080040#include <linux/memcontrol.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070041
42#include <asm/tlbflush.h>
43#include <asm/div64.h>
44
45#include <linux/swapops.h>
46
Nick Piggin0f8053a2006-03-22 00:08:33 -080047#include "internal.h"
48
Linus Torvalds1da177e2005-04-16 15:20:36 -070049struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070050 /* Incremented by the number of inactive pages that were scanned */
51 unsigned long nr_scanned;
52
Linus Torvalds1da177e2005-04-16 15:20:36 -070053 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040054 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070055
56 int may_writepage;
57
Christoph Lameterf1fd1062006-01-18 17:42:30 -080058 /* Can pages be swapped as part of reclaim? */
59 int may_swap;
60
Linus Torvalds1da177e2005-04-16 15:20:36 -070061 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
62 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
63 * In this context, it doesn't matter that we scan the
64 * whole list at once. */
65 int swap_cluster_max;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070066
67 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070068
69 int all_unreclaimable;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070070
71 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080072
73 /* Which cgroup do we reclaim from */
74 struct mem_cgroup *mem_cgroup;
75
76 /* Pluggable isolate pages callback */
77 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
78 unsigned long *scanned, int order, int mode,
79 struct zone *z, struct mem_cgroup *mem_cont,
80 int active);
Linus Torvalds1da177e2005-04-16 15:20:36 -070081};
82
Linus Torvalds1da177e2005-04-16 15:20:36 -070083#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
84
85#ifdef ARCH_HAS_PREFETCH
86#define prefetch_prev_lru_page(_page, _base, _field) \
87 do { \
88 if ((_page)->lru.prev != _base) { \
89 struct page *prev; \
90 \
91 prev = lru_to_page(&(_page->lru)); \
92 prefetch(&prev->_field); \
93 } \
94 } while (0)
95#else
96#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
97#endif
98
99#ifdef ARCH_HAS_PREFETCHW
100#define prefetchw_prev_lru_page(_page, _base, _field) \
101 do { \
102 if ((_page)->lru.prev != _base) { \
103 struct page *prev; \
104 \
105 prev = lru_to_page(&(_page->lru)); \
106 prefetchw(&prev->_field); \
107 } \
108 } while (0)
109#else
110#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
111#endif
112
113/*
114 * From 0 .. 100. Higher means more swappy.
115 */
116int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700117long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700118
119static LIST_HEAD(shrinker_list);
120static DECLARE_RWSEM(shrinker_rwsem);
121
Balbir Singh00f0b822008-03-04 14:28:39 -0800122#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800123#define scan_global_lru(sc) (!(sc)->mem_cgroup)
124#else
125#define scan_global_lru(sc) (1)
126#endif
127
Linus Torvalds1da177e2005-04-16 15:20:36 -0700128/*
129 * Add a shrinker callback to be called from the vm
130 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700131void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700132{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700133 shrinker->nr = 0;
134 down_write(&shrinker_rwsem);
135 list_add_tail(&shrinker->list, &shrinker_list);
136 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700137}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700138EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700139
140/*
141 * Remove one
142 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700143void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700144{
145 down_write(&shrinker_rwsem);
146 list_del(&shrinker->list);
147 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700148}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700149EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700150
151#define SHRINK_BATCH 128
152/*
153 * Call the shrink functions to age shrinkable caches
154 *
155 * Here we assume it costs one seek to replace a lru page and that it also
156 * takes a seek to recreate a cache object. With this in mind we age equal
157 * percentages of the lru and ageable caches. This should balance the seeks
158 * generated by these structures.
159 *
Simon Arlott183ff222007-10-20 01:27:18 +0200160 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161 * slab to avoid swapping.
162 *
163 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
164 *
165 * `lru_pages' represents the number of on-LRU pages in all the zones which
166 * are eligible for the caller's allocation attempt. It is used for balancing
167 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700168 *
169 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700170 */
Andrew Morton69e05942006-03-22 00:08:19 -0800171unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
172 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700173{
174 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800175 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700176
177 if (scanned == 0)
178 scanned = SWAP_CLUSTER_MAX;
179
180 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700181 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700182
183 list_for_each_entry(shrinker, &shrinker_list, list) {
184 unsigned long long delta;
185 unsigned long total_scan;
Rusty Russell8e1f9362007-07-17 04:03:17 -0700186 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700187
188 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800189 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700190 do_div(delta, lru_pages + 1);
191 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800192 if (shrinker->nr < 0) {
193 printk(KERN_ERR "%s: nr=%ld\n",
194 __FUNCTION__, shrinker->nr);
195 shrinker->nr = max_pass;
196 }
197
198 /*
199 * Avoid risking looping forever due to too large nr value:
200 * never try to free more than twice the estimate number of
201 * freeable entries.
202 */
203 if (shrinker->nr > max_pass * 2)
204 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205
206 total_scan = shrinker->nr;
207 shrinker->nr = 0;
208
209 while (total_scan >= SHRINK_BATCH) {
210 long this_scan = SHRINK_BATCH;
211 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700212 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700213
Rusty Russell8e1f9362007-07-17 04:03:17 -0700214 nr_before = (*shrinker->shrink)(0, gfp_mask);
215 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700216 if (shrink_ret == -1)
217 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700218 if (shrink_ret < nr_before)
219 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700220 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700221 total_scan -= this_scan;
222
223 cond_resched();
224 }
225
226 shrinker->nr += total_scan;
227 }
228 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700229 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700230}
231
232/* Called without lock on whether page is mapped, so answer is unstable */
233static inline int page_mapping_inuse(struct page *page)
234{
235 struct address_space *mapping;
236
237 /* Page is in somebody's page tables. */
238 if (page_mapped(page))
239 return 1;
240
241 /* Be more reluctant to reclaim swapcache than pagecache */
242 if (PageSwapCache(page))
243 return 1;
244
245 mapping = page_mapping(page);
246 if (!mapping)
247 return 0;
248
249 /* File is mmap'd by somebody? */
250 return mapping_mapped(mapping);
251}
252
253static inline int is_page_cache_freeable(struct page *page)
254{
255 return page_count(page) - !!PagePrivate(page) == 2;
256}
257
258static int may_write_to_queue(struct backing_dev_info *bdi)
259{
Christoph Lameter930d9152006-01-08 01:00:47 -0800260 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700261 return 1;
262 if (!bdi_write_congested(bdi))
263 return 1;
264 if (bdi == current->backing_dev_info)
265 return 1;
266 return 0;
267}
268
269/*
270 * We detected a synchronous write error writing a page out. Probably
271 * -ENOSPC. We need to propagate that into the address_space for a subsequent
272 * fsync(), msync() or close().
273 *
274 * The tricky part is that after writepage we cannot touch the mapping: nothing
275 * prevents it from being freed up. But we have a ref on the page and once
276 * that page is locked, the mapping is pinned.
277 *
278 * We're allowed to run sleeping lock_page() here because we know the caller has
279 * __GFP_FS.
280 */
281static void handle_write_error(struct address_space *mapping,
282 struct page *page, int error)
283{
284 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700285 if (page_mapping(page) == mapping)
286 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700287 unlock_page(page);
288}
289
Andy Whitcroftc661b072007-08-22 14:01:26 -0700290/* Request for sync pageout. */
291enum pageout_io {
292 PAGEOUT_IO_ASYNC,
293 PAGEOUT_IO_SYNC,
294};
295
Christoph Lameter04e62a22006-06-23 02:03:38 -0700296/* possible outcome of pageout() */
297typedef enum {
298 /* failed to write page out, page is locked */
299 PAGE_KEEP,
300 /* move page to the active list, page is locked */
301 PAGE_ACTIVATE,
302 /* page has been sent to the disk successfully, page is unlocked */
303 PAGE_SUCCESS,
304 /* page is clean and locked */
305 PAGE_CLEAN,
306} pageout_t;
307
Linus Torvalds1da177e2005-04-16 15:20:36 -0700308/*
Andrew Morton1742f192006-03-22 00:08:21 -0800309 * pageout is called by shrink_page_list() for each dirty page.
310 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700311 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700312static pageout_t pageout(struct page *page, struct address_space *mapping,
313 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314{
315 /*
316 * If the page is dirty, only perform writeback if that write
317 * will be non-blocking. To prevent this allocation from being
318 * stalled by pagecache activity. But note that there may be
319 * stalls if we need to run get_block(). We could test
320 * PagePrivate for that.
321 *
322 * If this process is currently in generic_file_write() against
323 * this page's queue, we can perform writeback even if that
324 * will block.
325 *
326 * If the page is swapcache, write it back even if that would
327 * block, for some throttling. This happens by accident, because
328 * swap_backing_dev_info is bust: it doesn't reflect the
329 * congestion state of the swapdevs. Easy to fix, if needed.
330 * See swapfile.c:page_queue_congested().
331 */
332 if (!is_page_cache_freeable(page))
333 return PAGE_KEEP;
334 if (!mapping) {
335 /*
336 * Some data journaling orphaned pages can have
337 * page->mapping == NULL while being dirty with clean buffers.
338 */
akpm@osdl.org323aca62005-04-16 15:24:06 -0700339 if (PagePrivate(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 if (try_to_free_buffers(page)) {
341 ClearPageDirty(page);
342 printk("%s: orphaned page\n", __FUNCTION__);
343 return PAGE_CLEAN;
344 }
345 }
346 return PAGE_KEEP;
347 }
348 if (mapping->a_ops->writepage == NULL)
349 return PAGE_ACTIVATE;
350 if (!may_write_to_queue(mapping->backing_dev_info))
351 return PAGE_KEEP;
352
353 if (clear_page_dirty_for_io(page)) {
354 int res;
355 struct writeback_control wbc = {
356 .sync_mode = WB_SYNC_NONE,
357 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700358 .range_start = 0,
359 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700360 .nonblocking = 1,
361 .for_reclaim = 1,
362 };
363
364 SetPageReclaim(page);
365 res = mapping->a_ops->writepage(page, &wbc);
366 if (res < 0)
367 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800368 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700369 ClearPageReclaim(page);
370 return PAGE_ACTIVATE;
371 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700372
373 /*
374 * Wait on writeback if requested to. This happens when
375 * direct reclaiming a large contiguous area and the
376 * first attempt to free a range of pages fails.
377 */
378 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
379 wait_on_page_writeback(page);
380
Linus Torvalds1da177e2005-04-16 15:20:36 -0700381 if (!PageWriteback(page)) {
382 /* synchronous write or broken a_ops? */
383 ClearPageReclaim(page);
384 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700385 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386 return PAGE_SUCCESS;
387 }
388
389 return PAGE_CLEAN;
390}
391
Andrew Mortona649fd92006-10-17 00:09:36 -0700392/*
393 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
394 * someone else has a ref on the page, abort and return 0. If it was
395 * successfully detached, return 1. Assumes the caller has a single ref on
396 * this page.
397 */
Christoph Lameterb20a3502006-03-22 00:09:12 -0800398int remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800399{
Nick Piggin28e4d962006-09-25 23:31:23 -0700400 BUG_ON(!PageLocked(page));
401 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800402
403 write_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800404 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700405 * The non racy check for a busy page.
406 *
407 * Must be careful with the order of the tests. When someone has
408 * a ref to the page, it may be possible that they dirty it then
409 * drop the reference. So if PageDirty is tested before page_count
410 * here, then the following race may occur:
411 *
412 * get_user_pages(&page);
413 * [user mapping goes away]
414 * write_to(page);
415 * !PageDirty(page) [good]
416 * SetPageDirty(page);
417 * put_page(page);
418 * !page_count(page) [good, discard it]
419 *
420 * [oops, our write_to data is lost]
421 *
422 * Reversing the order of the tests ensures such a situation cannot
423 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
424 * load is not satisfied before that of page->_count.
425 *
426 * Note that if SetPageDirty is always performed via set_page_dirty,
427 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800428 */
429 if (unlikely(page_count(page) != 2))
430 goto cannot_free;
431 smp_rmb();
432 if (unlikely(PageDirty(page)))
433 goto cannot_free;
434
435 if (PageSwapCache(page)) {
436 swp_entry_t swap = { .val = page_private(page) };
437 __delete_from_swap_cache(page);
438 write_unlock_irq(&mapping->tree_lock);
439 swap_free(swap);
440 __put_page(page); /* The pagecache ref */
441 return 1;
442 }
443
444 __remove_from_page_cache(page);
445 write_unlock_irq(&mapping->tree_lock);
446 __put_page(page);
447 return 1;
448
449cannot_free:
450 write_unlock_irq(&mapping->tree_lock);
451 return 0;
452}
453
Linus Torvalds1da177e2005-04-16 15:20:36 -0700454/*
Andrew Morton1742f192006-03-22 00:08:21 -0800455 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700456 */
Andrew Morton1742f192006-03-22 00:08:21 -0800457static unsigned long shrink_page_list(struct list_head *page_list,
Andy Whitcroftc661b072007-08-22 14:01:26 -0700458 struct scan_control *sc,
459 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460{
461 LIST_HEAD(ret_pages);
462 struct pagevec freed_pvec;
463 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800464 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700465
466 cond_resched();
467
468 pagevec_init(&freed_pvec, 1);
469 while (!list_empty(page_list)) {
470 struct address_space *mapping;
471 struct page *page;
472 int may_enter_fs;
473 int referenced;
474
475 cond_resched();
476
477 page = lru_to_page(page_list);
478 list_del(&page->lru);
479
480 if (TestSetPageLocked(page))
481 goto keep;
482
Nick Piggin725d7042006-09-25 23:30:55 -0700483 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484
485 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800486
487 if (!sc->may_swap && page_mapped(page))
488 goto keep_locked;
489
Linus Torvalds1da177e2005-04-16 15:20:36 -0700490 /* Double the slab pressure for mapped and swapcache pages */
491 if (page_mapped(page) || PageSwapCache(page))
492 sc->nr_scanned++;
493
Andy Whitcroftc661b072007-08-22 14:01:26 -0700494 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
495 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
496
497 if (PageWriteback(page)) {
498 /*
499 * Synchronous reclaim is performed in two passes,
500 * first an asynchronous pass over the list to
501 * start parallel writeback, and a second synchronous
502 * pass to wait for the IO to complete. Wait here
503 * for any page for which writeback has already
504 * started.
505 */
506 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
507 wait_on_page_writeback(page);
Andrew Morton4dd4b922008-03-24 12:29:52 -0700508 else
Andy Whitcroftc661b072007-08-22 14:01:26 -0700509 goto keep_locked;
510 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700511
Balbir Singhbed71612008-02-07 00:14:01 -0800512 referenced = page_referenced(page, 1, sc->mem_cgroup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700513 /* In active use or really unfreeable? Activate it. */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700514 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
515 referenced && page_mapping_inuse(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 goto activate_locked;
517
518#ifdef CONFIG_SWAP
519 /*
520 * Anonymous process memory has backing store?
521 * Try to allocate it some swap space here.
522 */
Christoph Lameter6e5ef1a2006-03-22 00:08:45 -0800523 if (PageAnon(page) && !PageSwapCache(page))
Christoph Lameter1480a542006-01-08 01:00:53 -0800524 if (!add_to_swap(page, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700525 goto activate_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700526#endif /* CONFIG_SWAP */
527
528 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529
530 /*
531 * The page is mapped into the page tables of one or more
532 * processes. Try to unmap it here.
533 */
534 if (page_mapped(page) && mapping) {
Christoph Lametera48d07a2006-02-01 03:05:38 -0800535 switch (try_to_unmap(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700536 case SWAP_FAIL:
537 goto activate_locked;
538 case SWAP_AGAIN:
539 goto keep_locked;
540 case SWAP_SUCCESS:
541 ; /* try to free the page below */
542 }
543 }
544
545 if (PageDirty(page)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700546 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700547 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700548 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700549 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800550 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700551 goto keep_locked;
552
553 /* Page is dirty, try to write it out here */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700554 switch (pageout(page, mapping, sync_writeback)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555 case PAGE_KEEP:
556 goto keep_locked;
557 case PAGE_ACTIVATE:
558 goto activate_locked;
559 case PAGE_SUCCESS:
Andrew Morton4dd4b922008-03-24 12:29:52 -0700560 if (PageWriteback(page) || PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700561 goto keep;
562 /*
563 * A synchronous write - probably a ramdisk. Go
564 * ahead and try to reclaim the page.
565 */
566 if (TestSetPageLocked(page))
567 goto keep;
568 if (PageDirty(page) || PageWriteback(page))
569 goto keep_locked;
570 mapping = page_mapping(page);
571 case PAGE_CLEAN:
572 ; /* try to free the page below */
573 }
574 }
575
576 /*
577 * If the page has buffers, try to free the buffer mappings
578 * associated with this page. If we succeed we try to free
579 * the page as well.
580 *
581 * We do this even if the page is PageDirty().
582 * try_to_release_page() does not perform I/O, but it is
583 * possible for a page to have PageDirty set, but it is actually
584 * clean (all its buffers are clean). This happens if the
585 * buffers were written out directly, with submit_bh(). ext3
586 * will do this, as well as the blockdev mapping.
587 * try_to_release_page() will discover that cleanness and will
588 * drop the buffers and mark the page clean - it can be freed.
589 *
590 * Rarely, pages can have buffers and no ->mapping. These are
591 * the pages which were not successfully invalidated in
592 * truncate_complete_page(). We try to drop those buffers here
593 * and if that worked, and the page is no longer mapped into
594 * process address space (page_count == 1) it can be freed.
595 * Otherwise, leave the page on the LRU so it is swappable.
596 */
597 if (PagePrivate(page)) {
598 if (!try_to_release_page(page, sc->gfp_mask))
599 goto activate_locked;
600 if (!mapping && page_count(page) == 1)
601 goto free_it;
602 }
603
Nick Piggin28e4d962006-09-25 23:31:23 -0700604 if (!mapping || !remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800605 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606
607free_it:
608 unlock_page(page);
Andrew Morton05ff5132006-03-22 00:08:20 -0800609 nr_reclaimed++;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 if (!pagevec_add(&freed_pvec, page))
611 __pagevec_release_nonlru(&freed_pvec);
612 continue;
613
614activate_locked:
615 SetPageActive(page);
616 pgactivate++;
617keep_locked:
618 unlock_page(page);
619keep:
620 list_add(&page->lru, &ret_pages);
Nick Piggin725d7042006-09-25 23:30:55 -0700621 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700622 }
623 list_splice(&ret_pages, page_list);
624 if (pagevec_count(&freed_pvec))
625 __pagevec_release_nonlru(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700626 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800627 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628}
629
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700630/* LRU Isolation modes. */
631#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
632#define ISOLATE_ACTIVE 1 /* Isolate active pages. */
633#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
634
635/*
636 * Attempt to remove the specified page from its LRU. Only take this page
637 * if it is of the appropriate PageActive status. Pages which are being
638 * freed elsewhere are also ignored.
639 *
640 * page: page to consider
641 * mode: one of the LRU isolation modes defined above
642 *
643 * returns 0 on success, -ve errno on failure.
644 */
Balbir Singh66e17072008-02-07 00:13:56 -0800645int __isolate_lru_page(struct page *page, int mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700646{
647 int ret = -EINVAL;
648
649 /* Only take pages on the LRU. */
650 if (!PageLRU(page))
651 return ret;
652
653 /*
654 * When checking the active state, we need to be sure we are
655 * dealing with comparible boolean values. Take the logical not
656 * of each.
657 */
658 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
659 return ret;
660
661 ret = -EBUSY;
662 if (likely(get_page_unless_zero(page))) {
663 /*
664 * Be careful not to clear PageLRU until after we're
665 * sure the page is not being freed elsewhere -- the
666 * page release code relies on it.
667 */
668 ClearPageLRU(page);
669 ret = 0;
670 }
671
672 return ret;
673}
674
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800675/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700676 * zone->lru_lock is heavily contended. Some of the functions that
677 * shrink the lists perform better by taking out a batch of pages
678 * and working on them outside the LRU lock.
679 *
680 * For pagecache intensive workloads, this function is the hottest
681 * spot in the kernel (apart from copy_*_user functions).
682 *
683 * Appropriate locks must be held before calling this function.
684 *
685 * @nr_to_scan: The number of pages to look through on the list.
686 * @src: The LRU list to pull pages off.
687 * @dst: The temp list to put pages on to.
688 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700689 * @order: The caller's attempted allocation order
690 * @mode: One of the LRU isolation modes
Linus Torvalds1da177e2005-04-16 15:20:36 -0700691 *
692 * returns how many pages were moved onto *@dst.
693 */
Andrew Morton69e05942006-03-22 00:08:19 -0800694static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
695 struct list_head *src, struct list_head *dst,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700696 unsigned long *scanned, int order, int mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700697{
Andrew Morton69e05942006-03-22 00:08:19 -0800698 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800699 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700700
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800701 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700702 struct page *page;
703 unsigned long pfn;
704 unsigned long end_pfn;
705 unsigned long page_pfn;
706 int zone_id;
707
Linus Torvalds1da177e2005-04-16 15:20:36 -0700708 page = lru_to_page(src);
709 prefetchw_prev_lru_page(page, src, flags);
710
Nick Piggin725d7042006-09-25 23:30:55 -0700711 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800712
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700713 switch (__isolate_lru_page(page, mode)) {
714 case 0:
715 list_move(&page->lru, dst);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800716 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700717 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800718
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700719 case -EBUSY:
720 /* else it is being freed elsewhere */
721 list_move(&page->lru, src);
722 continue;
723
724 default:
725 BUG();
726 }
727
728 if (!order)
729 continue;
730
731 /*
732 * Attempt to take all pages in the order aligned region
733 * surrounding the tag page. Only take those pages of
734 * the same active state as that tag page. We may safely
735 * round the target page pfn down to the requested order
736 * as the mem_map is guarenteed valid out to MAX_ORDER,
737 * where that page is in a different zone we will detect
738 * it from its zone id and abort this block scan.
739 */
740 zone_id = page_zone_id(page);
741 page_pfn = page_to_pfn(page);
742 pfn = page_pfn & ~((1 << order) - 1);
743 end_pfn = pfn + (1 << order);
744 for (; pfn < end_pfn; pfn++) {
745 struct page *cursor_page;
746
747 /* The target page is in the block, ignore it. */
748 if (unlikely(pfn == page_pfn))
749 continue;
750
751 /* Avoid holes within the zone. */
752 if (unlikely(!pfn_valid_within(pfn)))
753 break;
754
755 cursor_page = pfn_to_page(pfn);
756 /* Check that we have not crossed a zone boundary. */
757 if (unlikely(page_zone_id(cursor_page) != zone_id))
758 continue;
759 switch (__isolate_lru_page(cursor_page, mode)) {
760 case 0:
761 list_move(&cursor_page->lru, dst);
762 nr_taken++;
763 scan++;
764 break;
765
766 case -EBUSY:
767 /* else it is being freed elsewhere */
768 list_move(&cursor_page->lru, src);
769 default:
770 break;
771 }
772 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700773 }
774
775 *scanned = scan;
776 return nr_taken;
777}
778
Balbir Singh66e17072008-02-07 00:13:56 -0800779static unsigned long isolate_pages_global(unsigned long nr,
780 struct list_head *dst,
781 unsigned long *scanned, int order,
782 int mode, struct zone *z,
783 struct mem_cgroup *mem_cont,
784 int active)
785{
786 if (active)
787 return isolate_lru_pages(nr, &z->active_list, dst,
788 scanned, order, mode);
789 else
790 return isolate_lru_pages(nr, &z->inactive_list, dst,
791 scanned, order, mode);
792}
793
Linus Torvalds1da177e2005-04-16 15:20:36 -0700794/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700795 * clear_active_flags() is a helper for shrink_active_list(), clearing
796 * any active bits from the pages in the list.
797 */
798static unsigned long clear_active_flags(struct list_head *page_list)
799{
800 int nr_active = 0;
801 struct page *page;
802
803 list_for_each_entry(page, page_list, lru)
804 if (PageActive(page)) {
805 ClearPageActive(page);
806 nr_active++;
807 }
808
809 return nr_active;
810}
811
812/*
Andrew Morton1742f192006-03-22 00:08:21 -0800813 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
814 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700815 */
Andrew Morton1742f192006-03-22 00:08:21 -0800816static unsigned long shrink_inactive_list(unsigned long max_scan,
817 struct zone *zone, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700818{
819 LIST_HEAD(page_list);
820 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -0800821 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800822 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700823
824 pagevec_init(&pvec, 1);
825
826 lru_add_drain();
827 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -0800828 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -0800830 unsigned long nr_taken;
831 unsigned long nr_scan;
832 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700833 unsigned long nr_active;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700834
Balbir Singh66e17072008-02-07 00:13:56 -0800835 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700836 &page_list, &nr_scan, sc->order,
837 (sc->order > PAGE_ALLOC_COSTLY_ORDER)?
Balbir Singh66e17072008-02-07 00:13:56 -0800838 ISOLATE_BOTH : ISOLATE_INACTIVE,
839 zone, sc->mem_cgroup, 0);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700840 nr_active = clear_active_flags(&page_list);
Andy Whitcrofte9187bd2007-08-22 14:01:25 -0700841 __count_vm_events(PGDEACTIVATE, nr_active);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700842
843 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
844 __mod_zone_page_state(zone, NR_INACTIVE,
845 -(nr_taken - nr_active));
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -0800846 if (scan_global_lru(sc))
847 zone->pages_scanned += nr_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700848 spin_unlock_irq(&zone->lru_lock);
849
Andrew Morton69e05942006-03-22 00:08:19 -0800850 nr_scanned += nr_scan;
Andy Whitcroftc661b072007-08-22 14:01:26 -0700851 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
852
853 /*
854 * If we are direct reclaiming for contiguous pages and we do
855 * not reclaim everything in the list, try again and wait
856 * for IO to complete. This will stall high-order allocations
857 * but that should be acceptable to the caller
858 */
859 if (nr_freed < nr_taken && !current_is_kswapd() &&
860 sc->order > PAGE_ALLOC_COSTLY_ORDER) {
861 congestion_wait(WRITE, HZ/10);
862
863 /*
864 * The attempt at page out may have made some
865 * of the pages active, mark them inactive again.
866 */
867 nr_active = clear_active_flags(&page_list);
868 count_vm_events(PGDEACTIVATE, nr_active);
869
870 nr_freed += shrink_page_list(&page_list, sc,
871 PAGEOUT_IO_SYNC);
872 }
873
Andrew Morton05ff5132006-03-22 00:08:20 -0800874 nr_reclaimed += nr_freed;
Nick Piggina74609f2006-01-06 00:11:20 -0800875 local_irq_disable();
876 if (current_is_kswapd()) {
Christoph Lameterf8891e52006-06-30 01:55:45 -0700877 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
878 __count_vm_events(KSWAPD_STEAL, nr_freed);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -0800879 } else if (scan_global_lru(sc))
Christoph Lameterf8891e52006-06-30 01:55:45 -0700880 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -0800881
Shantanu Goel918d3f92006-12-29 16:48:59 -0800882 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -0800883
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800884 if (nr_taken == 0)
885 goto done;
886
Nick Piggina74609f2006-01-06 00:11:20 -0800887 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700888 /*
889 * Put back any unfreeable pages.
890 */
891 while (!list_empty(&page_list)) {
892 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -0700893 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800894 SetPageLRU(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 list_del(&page->lru);
896 if (PageActive(page))
897 add_page_to_active_list(zone, page);
898 else
899 add_page_to_inactive_list(zone, page);
900 if (!pagevec_add(&pvec, page)) {
901 spin_unlock_irq(&zone->lru_lock);
902 __pagevec_release(&pvec);
903 spin_lock_irq(&zone->lru_lock);
904 }
905 }
Andrew Morton69e05942006-03-22 00:08:19 -0800906 } while (nr_scanned < max_scan);
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800907 spin_unlock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908done:
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800909 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700910 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -0800911 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700912}
913
Martin Bligh3bb1a8522006-10-28 10:38:24 -0700914/*
915 * We are about to scan this zone at a certain priority level. If that priority
916 * level is smaller (ie: more urgent) than the previous priority, then note
917 * that priority level within the zone. This is done so that when the next
918 * process comes in to scan this zone, it will immediately start out at this
919 * priority level rather than having to build up its own scanning priority.
920 * Here, this priority affects only the reclaim-mapped threshold.
921 */
922static inline void note_zone_scanning_priority(struct zone *zone, int priority)
923{
924 if (priority < zone->prev_priority)
925 zone->prev_priority = priority;
926}
927
Nick Piggin4ff1ffb2006-09-25 23:31:28 -0700928static inline int zone_is_near_oom(struct zone *zone)
929{
Christoph Lameterc8785382007-02-10 01:43:01 -0800930 return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
931 + zone_page_state(zone, NR_INACTIVE))*3;
Nick Piggin4ff1ffb2006-09-25 23:31:28 -0700932}
933
Linus Torvalds1da177e2005-04-16 15:20:36 -0700934/*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -0800935 * Determine we should try to reclaim mapped pages.
936 * This is called only when sc->mem_cgroup is NULL.
937 */
938static int calc_reclaim_mapped(struct scan_control *sc, struct zone *zone,
939 int priority)
940{
941 long mapped_ratio;
942 long distress;
943 long swap_tendency;
944 long imbalance;
945 int reclaim_mapped = 0;
946 int prev_priority;
947
948 if (scan_global_lru(sc) && zone_is_near_oom(zone))
949 return 1;
950 /*
951 * `distress' is a measure of how much trouble we're having
952 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
953 */
954 if (scan_global_lru(sc))
955 prev_priority = zone->prev_priority;
956 else
957 prev_priority = mem_cgroup_get_reclaim_priority(sc->mem_cgroup);
958
959 distress = 100 >> min(prev_priority, priority);
960
961 /*
962 * The point of this algorithm is to decide when to start
963 * reclaiming mapped memory instead of just pagecache. Work out
964 * how much memory
965 * is mapped.
966 */
967 if (scan_global_lru(sc))
968 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
969 global_page_state(NR_ANON_PAGES)) * 100) /
970 vm_total_pages;
971 else
972 mapped_ratio = mem_cgroup_calc_mapped_ratio(sc->mem_cgroup);
973
974 /*
975 * Now decide how much we really want to unmap some pages. The
976 * mapped ratio is downgraded - just because there's a lot of
977 * mapped memory doesn't necessarily mean that page reclaim
978 * isn't succeeding.
979 *
980 * The distress ratio is important - we don't want to start
981 * going oom.
982 *
983 * A 100% value of vm_swappiness overrides this algorithm
984 * altogether.
985 */
986 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
987
988 /*
989 * If there's huge imbalance between active and inactive
990 * (think active 100 times larger than inactive) we should
991 * become more permissive, or the system will take too much
992 * cpu before it start swapping during memory pressure.
993 * Distress is about avoiding early-oom, this is about
994 * making swappiness graceful despite setting it to low
995 * values.
996 *
997 * Avoid div by zero with nr_inactive+1, and max resulting
998 * value is vm_total_pages.
999 */
1000 if (scan_global_lru(sc)) {
1001 imbalance = zone_page_state(zone, NR_ACTIVE);
1002 imbalance /= zone_page_state(zone, NR_INACTIVE) + 1;
1003 } else
1004 imbalance = mem_cgroup_reclaim_imbalance(sc->mem_cgroup);
1005
1006 /*
1007 * Reduce the effect of imbalance if swappiness is low,
1008 * this means for a swappiness very low, the imbalance
1009 * must be much higher than 100 for this logic to make
1010 * the difference.
1011 *
1012 * Max temporary value is vm_total_pages*100.
1013 */
1014 imbalance *= (vm_swappiness + 1);
1015 imbalance /= 100;
1016
1017 /*
1018 * If not much of the ram is mapped, makes the imbalance
1019 * less relevant, it's high priority we refill the inactive
1020 * list with mapped pages only in presence of high ratio of
1021 * mapped pages.
1022 *
1023 * Max temporary value is vm_total_pages*100.
1024 */
1025 imbalance *= mapped_ratio;
1026 imbalance /= 100;
1027
1028 /* apply imbalance feedback to swap_tendency */
1029 swap_tendency += imbalance;
1030
1031 /*
1032 * Now use this metric to decide whether to start moving mapped
1033 * memory onto the inactive list.
1034 */
1035 if (swap_tendency >= 100)
1036 reclaim_mapped = 1;
1037
1038 return reclaim_mapped;
1039}
1040
1041/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001042 * This moves pages from the active list to the inactive list.
1043 *
1044 * We move them the other way if the page is referenced by one or more
1045 * processes, from rmap.
1046 *
1047 * If the pages are mostly unmapped, the processing is fast and it is
1048 * appropriate to hold zone->lru_lock across the whole operation. But if
1049 * the pages are mapped, the processing is slow (page_referenced()) so we
1050 * should drop zone->lru_lock around each page. It's impossible to balance
1051 * this, so instead we remove the pages from the LRU while processing them.
1052 * It is safe to rely on PG_active against the non-LRU pages in here because
1053 * nobody will play with that bit on a non-LRU page.
1054 *
1055 * The downside is that we have to touch page->_count against each page.
1056 * But we had to alter page->flags anyway.
1057 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001058
1059
Andrew Morton1742f192006-03-22 00:08:21 -08001060static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Martin Blighbbdb3962006-10-28 10:38:25 -07001061 struct scan_control *sc, int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001062{
Andrew Morton69e05942006-03-22 00:08:19 -08001063 unsigned long pgmoved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001064 int pgdeactivate = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001065 unsigned long pgscanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066 LIST_HEAD(l_hold); /* The pages which were snipped off */
1067 LIST_HEAD(l_inactive); /* Pages to go onto the inactive_list */
1068 LIST_HEAD(l_active); /* Pages to go onto the active_list */
1069 struct page *page;
1070 struct pagevec pvec;
1071 int reclaim_mapped = 0;
Christoph Lameter2903fb12006-02-11 17:55:55 -08001072
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001073 if (sc->may_swap)
1074 reclaim_mapped = calc_reclaim_mapped(sc, zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001075
1076 lru_add_drain();
1077 spin_lock_irq(&zone->lru_lock);
Balbir Singh66e17072008-02-07 00:13:56 -08001078 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1079 ISOLATE_ACTIVE, zone,
1080 sc->mem_cgroup, 1);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001081 /*
1082 * zone->pages_scanned is used for detect zone's oom
1083 * mem_cgroup remembers nr_scan by itself.
1084 */
1085 if (scan_global_lru(sc))
1086 zone->pages_scanned += pgscanned;
1087
Christoph Lameterc8785382007-02-10 01:43:01 -08001088 __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001089 spin_unlock_irq(&zone->lru_lock);
1090
Linus Torvalds1da177e2005-04-16 15:20:36 -07001091 while (!list_empty(&l_hold)) {
1092 cond_resched();
1093 page = lru_to_page(&l_hold);
1094 list_del(&page->lru);
1095 if (page_mapped(page)) {
1096 if (!reclaim_mapped ||
1097 (total_swap_pages == 0 && PageAnon(page)) ||
Balbir Singhbed71612008-02-07 00:14:01 -08001098 page_referenced(page, 0, sc->mem_cgroup)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001099 list_add(&page->lru, &l_active);
1100 continue;
1101 }
1102 }
1103 list_add(&page->lru, &l_inactive);
1104 }
1105
1106 pagevec_init(&pvec, 1);
1107 pgmoved = 0;
1108 spin_lock_irq(&zone->lru_lock);
1109 while (!list_empty(&l_inactive)) {
1110 page = lru_to_page(&l_inactive);
1111 prefetchw_prev_lru_page(page, &l_inactive, flags);
Nick Piggin725d7042006-09-25 23:30:55 -07001112 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001113 SetPageLRU(page);
Nick Piggin725d7042006-09-25 23:30:55 -07001114 VM_BUG_ON(!PageActive(page));
Nick Piggin4c84cac2006-03-22 00:08:00 -08001115 ClearPageActive(page);
1116
Linus Torvalds1da177e2005-04-16 15:20:36 -07001117 list_move(&page->lru, &zone->inactive_list);
Hugh Dickins427d5412008-03-04 14:29:03 -08001118 mem_cgroup_move_lists(page, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 pgmoved++;
1120 if (!pagevec_add(&pvec, page)) {
Christoph Lameterc8785382007-02-10 01:43:01 -08001121 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001122 spin_unlock_irq(&zone->lru_lock);
1123 pgdeactivate += pgmoved;
1124 pgmoved = 0;
1125 if (buffer_heads_over_limit)
1126 pagevec_strip(&pvec);
1127 __pagevec_release(&pvec);
1128 spin_lock_irq(&zone->lru_lock);
1129 }
1130 }
Christoph Lameterc8785382007-02-10 01:43:01 -08001131 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001132 pgdeactivate += pgmoved;
1133 if (buffer_heads_over_limit) {
1134 spin_unlock_irq(&zone->lru_lock);
1135 pagevec_strip(&pvec);
1136 spin_lock_irq(&zone->lru_lock);
1137 }
1138
1139 pgmoved = 0;
1140 while (!list_empty(&l_active)) {
1141 page = lru_to_page(&l_active);
1142 prefetchw_prev_lru_page(page, &l_active, flags);
Nick Piggin725d7042006-09-25 23:30:55 -07001143 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001144 SetPageLRU(page);
Nick Piggin725d7042006-09-25 23:30:55 -07001145 VM_BUG_ON(!PageActive(page));
Hugh Dickins427d5412008-03-04 14:29:03 -08001146
Linus Torvalds1da177e2005-04-16 15:20:36 -07001147 list_move(&page->lru, &zone->active_list);
Hugh Dickins427d5412008-03-04 14:29:03 -08001148 mem_cgroup_move_lists(page, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001149 pgmoved++;
1150 if (!pagevec_add(&pvec, page)) {
Christoph Lameterc8785382007-02-10 01:43:01 -08001151 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152 pgmoved = 0;
1153 spin_unlock_irq(&zone->lru_lock);
1154 __pagevec_release(&pvec);
1155 spin_lock_irq(&zone->lru_lock);
1156 }
1157 }
Christoph Lameterc8785382007-02-10 01:43:01 -08001158 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001159
Christoph Lameterf8891e52006-06-30 01:55:45 -07001160 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1161 __count_vm_events(PGDEACTIVATE, pgdeactivate);
1162 spin_unlock_irq(&zone->lru_lock);
Nick Piggina74609f2006-01-06 00:11:20 -08001163
1164 pagevec_release(&pvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001165}
1166
1167/*
1168 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1169 */
Andrew Morton05ff5132006-03-22 00:08:20 -08001170static unsigned long shrink_zone(int priority, struct zone *zone,
1171 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001172{
1173 unsigned long nr_active;
1174 unsigned long nr_inactive;
Christoph Lameter86959492006-03-22 00:08:18 -08001175 unsigned long nr_to_scan;
Andrew Morton05ff5132006-03-22 00:08:20 -08001176 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001177
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001178 if (scan_global_lru(sc)) {
1179 /*
1180 * Add one to nr_to_scan just to make sure that the kernel
1181 * will slowly sift through the active list.
1182 */
1183 zone->nr_scan_active +=
1184 (zone_page_state(zone, NR_ACTIVE) >> priority) + 1;
1185 nr_active = zone->nr_scan_active;
1186 zone->nr_scan_inactive +=
1187 (zone_page_state(zone, NR_INACTIVE) >> priority) + 1;
1188 nr_inactive = zone->nr_scan_inactive;
1189 if (nr_inactive >= sc->swap_cluster_max)
1190 zone->nr_scan_inactive = 0;
1191 else
1192 nr_inactive = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001193
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001194 if (nr_active >= sc->swap_cluster_max)
1195 zone->nr_scan_active = 0;
1196 else
1197 nr_active = 0;
1198 } else {
1199 /*
1200 * This reclaim occurs not because zone memory shortage but
1201 * because memory controller hits its limit.
1202 * Then, don't modify zone reclaim related data.
1203 */
1204 nr_active = mem_cgroup_calc_reclaim_active(sc->mem_cgroup,
1205 zone, priority);
1206
1207 nr_inactive = mem_cgroup_calc_reclaim_inactive(sc->mem_cgroup,
1208 zone, priority);
1209 }
1210
Linus Torvalds1da177e2005-04-16 15:20:36 -07001211
Linus Torvalds1da177e2005-04-16 15:20:36 -07001212 while (nr_active || nr_inactive) {
1213 if (nr_active) {
Christoph Lameter86959492006-03-22 00:08:18 -08001214 nr_to_scan = min(nr_active,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001215 (unsigned long)sc->swap_cluster_max);
Christoph Lameter86959492006-03-22 00:08:18 -08001216 nr_active -= nr_to_scan;
Martin Blighbbdb3962006-10-28 10:38:25 -07001217 shrink_active_list(nr_to_scan, zone, sc, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001218 }
1219
1220 if (nr_inactive) {
Christoph Lameter86959492006-03-22 00:08:18 -08001221 nr_to_scan = min(nr_inactive,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001222 (unsigned long)sc->swap_cluster_max);
Christoph Lameter86959492006-03-22 00:08:18 -08001223 nr_inactive -= nr_to_scan;
Andrew Morton1742f192006-03-22 00:08:21 -08001224 nr_reclaimed += shrink_inactive_list(nr_to_scan, zone,
1225 sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 }
1227 }
1228
Andrew Morton232ea4d2007-02-28 20:13:21 -08001229 throttle_vm_writeout(sc->gfp_mask);
Andrew Morton05ff5132006-03-22 00:08:20 -08001230 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231}
1232
1233/*
1234 * This is the direct reclaim path, for page-allocating processes. We only
1235 * try to reclaim pages from zones which will satisfy the caller's allocation
1236 * request.
1237 *
1238 * We reclaim from a zone even if that zone is over pages_high. Because:
1239 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1240 * allocation or
1241 * b) The zones may be over pages_high but they must go *over* pages_high to
1242 * satisfy the `incremental min' zone defense algorithm.
1243 *
1244 * Returns the number of reclaimed pages.
1245 *
1246 * If a zone is deemed to be full of pinned pages then just give it a light
1247 * scan then give up on it.
1248 */
Andrew Morton1742f192006-03-22 00:08:21 -08001249static unsigned long shrink_zones(int priority, struct zone **zones,
Andrew Morton05ff5132006-03-22 00:08:20 -08001250 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001251{
Andrew Morton05ff5132006-03-22 00:08:20 -08001252 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253 int i;
1254
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001255
Nick Piggin408d8542006-09-25 23:31:27 -07001256 sc->all_unreclaimable = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001257 for (i = 0; zones[i] != NULL; i++) {
1258 struct zone *zone = zones[i];
1259
Con Kolivasf3fe6512006-01-06 00:11:15 -08001260 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001262 /*
1263 * Take care memory controller reclaiming has small influence
1264 * to global LRU.
1265 */
1266 if (scan_global_lru(sc)) {
1267 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1268 continue;
1269 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001270
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001271 if (zone_is_all_unreclaimable(zone) &&
1272 priority != DEF_PRIORITY)
1273 continue; /* Let kswapd poll it */
1274 sc->all_unreclaimable = 0;
1275 } else {
1276 /*
1277 * Ignore cpuset limitation here. We just want to reduce
1278 * # of used pages by us regardless of memory shortage.
1279 */
1280 sc->all_unreclaimable = 0;
1281 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1282 priority);
1283 }
Nick Piggin408d8542006-09-25 23:31:27 -07001284
Andrew Morton05ff5132006-03-22 00:08:20 -08001285 nr_reclaimed += shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001286 }
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001287
Andrew Morton05ff5132006-03-22 00:08:20 -08001288 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001289}
1290
1291/*
1292 * This is the main entry point to direct page reclaim.
1293 *
1294 * If a full scan of the inactive list fails to free enough memory then we
1295 * are "out of memory" and something needs to be killed.
1296 *
1297 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1298 * high - the zone may be full of dirty or under-writeback pages, which this
1299 * caller can't do much about. We kick pdflush and take explicit naps in the
1300 * hope that some of these pages can be written. But if the allocating task
1301 * holds filesystem locks which prevent writeout this might not work, and the
1302 * allocation attempt will fail.
1303 */
Balbir Singh66e17072008-02-07 00:13:56 -08001304static unsigned long do_try_to_free_pages(struct zone **zones, gfp_t gfp_mask,
1305 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001306{
1307 int priority;
1308 int ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001309 unsigned long total_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001310 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001311 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001312 unsigned long lru_pages = 0;
1313 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001314
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001315 if (scan_global_lru(sc))
1316 count_vm_event(ALLOCSTALL);
1317 /*
1318 * mem_cgroup will not do shrink_slab.
1319 */
1320 if (scan_global_lru(sc)) {
1321 for (i = 0; zones[i] != NULL; i++) {
1322 struct zone *zone = zones[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001323
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001324 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1325 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001326
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001327 lru_pages += zone_page_state(zone, NR_ACTIVE)
1328 + zone_page_state(zone, NR_INACTIVE);
1329 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001330 }
1331
1332 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001333 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001334 if (!priority)
1335 disable_swap_token();
Balbir Singh66e17072008-02-07 00:13:56 -08001336 nr_reclaimed += shrink_zones(priority, zones, sc);
1337 /*
1338 * Don't shrink slabs when reclaiming memory from
1339 * over limit cgroups
1340 */
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001341 if (scan_global_lru(sc)) {
Balbir Singh66e17072008-02-07 00:13:56 -08001342 shrink_slab(sc->nr_scanned, gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001343 if (reclaim_state) {
1344 nr_reclaimed += reclaim_state->reclaimed_slab;
1345 reclaim_state->reclaimed_slab = 0;
1346 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001347 }
Balbir Singh66e17072008-02-07 00:13:56 -08001348 total_scanned += sc->nr_scanned;
1349 if (nr_reclaimed >= sc->swap_cluster_max) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001350 ret = 1;
1351 goto out;
1352 }
1353
1354 /*
1355 * Try to write back as many pages as we just scanned. This
1356 * tends to cause slow streaming writers to write data to the
1357 * disk smoothly, at the dirtying rate, which is nice. But
1358 * that's undesirable in laptop mode, where we *want* lumpy
1359 * writeout. So in laptop mode, write out the whole world.
1360 */
Balbir Singh66e17072008-02-07 00:13:56 -08001361 if (total_scanned > sc->swap_cluster_max +
1362 sc->swap_cluster_max / 2) {
Pekka J Enberg687a21c2005-06-28 20:44:55 -07001363 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001364 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001365 }
1366
1367 /* Take a nap, wait for some writeback to complete */
Andrew Morton4dd4b922008-03-24 12:29:52 -07001368 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001369 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001370 }
Nick Piggin408d8542006-09-25 23:31:27 -07001371 /* top priority shrink_caches still had more to do? don't OOM, then */
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001372 if (!sc->all_unreclaimable && scan_global_lru(sc))
Nick Piggin408d8542006-09-25 23:31:27 -07001373 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001374out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001375 /*
1376 * Now that we've scanned all the zones at this priority level, note
1377 * that level within the zone so that the next thread which performs
1378 * scanning of this zone will immediately start out at this priority
1379 * level. This affects only the decision whether or not to bring
1380 * mapped pages onto the inactive list.
1381 */
1382 if (priority < 0)
1383 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001385 if (scan_global_lru(sc)) {
1386 for (i = 0; zones[i] != NULL; i++) {
1387 struct zone *zone = zones[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001388
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001389 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1390 continue;
1391
1392 zone->prev_priority = priority;
1393 }
1394 } else
1395 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1396
Linus Torvalds1da177e2005-04-16 15:20:36 -07001397 return ret;
1398}
1399
Balbir Singh66e17072008-02-07 00:13:56 -08001400unsigned long try_to_free_pages(struct zone **zones, int order, gfp_t gfp_mask)
1401{
1402 struct scan_control sc = {
1403 .gfp_mask = gfp_mask,
1404 .may_writepage = !laptop_mode,
1405 .swap_cluster_max = SWAP_CLUSTER_MAX,
1406 .may_swap = 1,
1407 .swappiness = vm_swappiness,
1408 .order = order,
1409 .mem_cgroup = NULL,
1410 .isolate_pages = isolate_pages_global,
1411 };
1412
1413 return do_try_to_free_pages(zones, gfp_mask, &sc);
1414}
1415
Balbir Singh00f0b822008-03-04 14:28:39 -08001416#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08001417
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001418unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1419 gfp_t gfp_mask)
Balbir Singh66e17072008-02-07 00:13:56 -08001420{
1421 struct scan_control sc = {
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001422 .gfp_mask = gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08001423 .may_writepage = !laptop_mode,
1424 .may_swap = 1,
1425 .swap_cluster_max = SWAP_CLUSTER_MAX,
1426 .swappiness = vm_swappiness,
1427 .order = 0,
1428 .mem_cgroup = mem_cont,
1429 .isolate_pages = mem_cgroup_isolate_pages,
1430 };
Balbir Singh66e17072008-02-07 00:13:56 -08001431 struct zone **zones;
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001432 int target_zone = gfp_zone(GFP_HIGHUSER_MOVABLE);
Balbir Singh66e17072008-02-07 00:13:56 -08001433
KAMEZAWA Hiroyuki417eead2008-02-07 00:14:14 -08001434 zones = NODE_DATA(numa_node_id())->node_zonelists[target_zone].zones;
1435 if (do_try_to_free_pages(zones, sc.gfp_mask, &sc))
1436 return 1;
Balbir Singh66e17072008-02-07 00:13:56 -08001437 return 0;
1438}
1439#endif
1440
Linus Torvalds1da177e2005-04-16 15:20:36 -07001441/*
1442 * For kswapd, balance_pgdat() will work across all this node's zones until
1443 * they are all at pages_high.
1444 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001445 * Returns the number of pages which were actually freed.
1446 *
1447 * There is special handling here for zones which are full of pinned pages.
1448 * This can happen if the pages are all mlocked, or if they are all used by
1449 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1450 * What we do is to detect the case where all pages in the zone have been
1451 * scanned twice and there has been zero successful reclaim. Mark the zone as
1452 * dead and from now on, only perform a short scan. Basically we're polling
1453 * the zone for when the problem goes away.
1454 *
1455 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1456 * zones which have free_pages > pages_high, but once a zone is found to have
1457 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1458 * of the number of free pages in the lower zones. This interoperates with
1459 * the page allocator fallback scheme to ensure that aging of pages is balanced
1460 * across the zones.
1461 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001462static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001463{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 int all_zones_ok;
1465 int priority;
1466 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08001467 unsigned long total_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001468 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001469 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08001470 struct scan_control sc = {
1471 .gfp_mask = GFP_KERNEL,
1472 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001473 .swap_cluster_max = SWAP_CLUSTER_MAX,
1474 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001475 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08001476 .mem_cgroup = NULL,
1477 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08001478 };
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001479 /*
1480 * temp_priority is used to remember the scanning priority at which
1481 * this zone was successfully refilled to free_pages == pages_high.
1482 */
1483 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001484
1485loop_again:
1486 total_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001487 nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07001488 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001489 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001490
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001491 for (i = 0; i < pgdat->nr_zones; i++)
1492 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001493
1494 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1495 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1496 unsigned long lru_pages = 0;
1497
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001498 /* The swap token gets in the way of swapout... */
1499 if (!priority)
1500 disable_swap_token();
1501
Linus Torvalds1da177e2005-04-16 15:20:36 -07001502 all_zones_ok = 1;
1503
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001504 /*
1505 * Scan in the highmem->dma direction for the highest
1506 * zone which needs scanning
1507 */
1508 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1509 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001510
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001511 if (!populated_zone(zone))
1512 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001513
David Rientjese815af92007-10-16 23:25:54 -07001514 if (zone_is_all_unreclaimable(zone) &&
1515 priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001516 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001517
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001518 if (!zone_watermark_ok(zone, order, zone->pages_high,
1519 0, 0)) {
1520 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001521 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001523 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001524 if (i < 0)
1525 goto out;
1526
Linus Torvalds1da177e2005-04-16 15:20:36 -07001527 for (i = 0; i <= end_zone; i++) {
1528 struct zone *zone = pgdat->node_zones + i;
1529
Christoph Lameterc8785382007-02-10 01:43:01 -08001530 lru_pages += zone_page_state(zone, NR_ACTIVE)
1531 + zone_page_state(zone, NR_INACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001532 }
1533
1534 /*
1535 * Now scan the zone in the dma->highmem direction, stopping
1536 * at the last zone which needs scanning.
1537 *
1538 * We do this because the page allocator works in the opposite
1539 * direction. This prevents the page allocator from allocating
1540 * pages behind kswapd's direction of progress, which would
1541 * cause too much scanning of the lower zones.
1542 */
1543 for (i = 0; i <= end_zone; i++) {
1544 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001545 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001546
Con Kolivasf3fe6512006-01-06 00:11:15 -08001547 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001548 continue;
1549
David Rientjese815af92007-10-16 23:25:54 -07001550 if (zone_is_all_unreclaimable(zone) &&
1551 priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001552 continue;
1553
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001554 if (!zone_watermark_ok(zone, order, zone->pages_high,
1555 end_zone, 0))
1556 all_zones_ok = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001557 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558 sc.nr_scanned = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001559 note_zone_scanning_priority(zone, priority);
Rik van Riel32a43302007-10-16 01:24:50 -07001560 /*
1561 * We put equal pressure on every zone, unless one
1562 * zone has way too many pages free already.
1563 */
1564 if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1565 end_zone, 0))
1566 nr_reclaimed += shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001567 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001568 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1569 lru_pages);
Andrew Morton05ff5132006-03-22 00:08:20 -08001570 nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571 total_scanned += sc.nr_scanned;
David Rientjese815af92007-10-16 23:25:54 -07001572 if (zone_is_all_unreclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001574 if (nr_slab == 0 && zone->pages_scanned >=
Christoph Lameterc8785382007-02-10 01:43:01 -08001575 (zone_page_state(zone, NR_ACTIVE)
1576 + zone_page_state(zone, NR_INACTIVE)) * 6)
David Rientjese815af92007-10-16 23:25:54 -07001577 zone_set_flag(zone,
1578 ZONE_ALL_UNRECLAIMABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001579 /*
1580 * If we've done a decent amount of scanning and
1581 * the reclaim ratio is low, start doing writepage
1582 * even in laptop mode
1583 */
1584 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Andrew Morton05ff5132006-03-22 00:08:20 -08001585 total_scanned > nr_reclaimed + nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001586 sc.may_writepage = 1;
1587 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001588 if (all_zones_ok)
1589 break; /* kswapd: all done */
1590 /*
1591 * OK, kswapd is getting into trouble. Take a nap, then take
1592 * another pass across the zones.
1593 */
Andrew Morton4dd4b922008-03-24 12:29:52 -07001594 if (total_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001595 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001596
1597 /*
1598 * We do this so kswapd doesn't build up large priorities for
1599 * example when it is freeing in parallel with allocators. It
1600 * matches the direct reclaim path behaviour in terms of impact
1601 * on zone->*_priority.
1602 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001603 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 break;
1605 }
1606out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001607 /*
1608 * Note within each zone the priority level at which this zone was
1609 * brought into a happy state. So that the next thread which scans this
1610 * zone will start out at that priority level.
1611 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001612 for (i = 0; i < pgdat->nr_zones; i++) {
1613 struct zone *zone = pgdat->node_zones + i;
1614
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001615 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001616 }
1617 if (!all_zones_ok) {
1618 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08001619
1620 try_to_freeze();
1621
Linus Torvalds1da177e2005-04-16 15:20:36 -07001622 goto loop_again;
1623 }
1624
Andrew Morton05ff5132006-03-22 00:08:20 -08001625 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626}
1627
1628/*
1629 * The background pageout daemon, started as a kernel thread
1630 * from the init process.
1631 *
1632 * This basically trickles out pages so that we have _some_
1633 * free memory available even if there is no other activity
1634 * that frees anything up. This is needed for things like routing
1635 * etc, where we otherwise might have all activity going on in
1636 * asynchronous contexts that cannot page things out.
1637 *
1638 * If there are applications that are active memory-allocators
1639 * (most normal use), this basically shouldn't matter.
1640 */
1641static int kswapd(void *p)
1642{
1643 unsigned long order;
1644 pg_data_t *pgdat = (pg_data_t*)p;
1645 struct task_struct *tsk = current;
1646 DEFINE_WAIT(wait);
1647 struct reclaim_state reclaim_state = {
1648 .reclaimed_slab = 0,
1649 };
1650 cpumask_t cpumask;
1651
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652 cpumask = node_to_cpumask(pgdat->node_id);
1653 if (!cpus_empty(cpumask))
1654 set_cpus_allowed(tsk, cpumask);
1655 current->reclaim_state = &reclaim_state;
1656
1657 /*
1658 * Tell the memory management that we're a "memory allocator",
1659 * and that if we need more memory we should get access to it
1660 * regardless (see "__alloc_pages()"). "kswapd" should
1661 * never get caught in the normal page freeing logic.
1662 *
1663 * (Kswapd normally doesn't need memory anyway, but sometimes
1664 * you need a small amount of memory in order to be able to
1665 * page out something else, and this flag essentially protects
1666 * us from recursively trying to free more memory as we're
1667 * trying to free the first piece of memory in the first place).
1668 */
Christoph Lameter930d9152006-01-08 01:00:47 -08001669 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07001670 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001671
1672 order = 0;
1673 for ( ; ; ) {
1674 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07001675
Linus Torvalds1da177e2005-04-16 15:20:36 -07001676 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1677 new_order = pgdat->kswapd_max_order;
1678 pgdat->kswapd_max_order = 0;
1679 if (order < new_order) {
1680 /*
1681 * Don't sleep if someone wants a larger 'order'
1682 * allocation
1683 */
1684 order = new_order;
1685 } else {
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07001686 if (!freezing(current))
1687 schedule();
1688
Linus Torvalds1da177e2005-04-16 15:20:36 -07001689 order = pgdat->kswapd_max_order;
1690 }
1691 finish_wait(&pgdat->kswapd_wait, &wait);
1692
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07001693 if (!try_to_freeze()) {
1694 /* We can speed up thawing tasks if we don't call
1695 * balance_pgdat after returning from the refrigerator
1696 */
1697 balance_pgdat(pgdat, order);
1698 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001699 }
1700 return 0;
1701}
1702
1703/*
1704 * A zone is low on free memory, so wake its kswapd task to service it.
1705 */
1706void wakeup_kswapd(struct zone *zone, int order)
1707{
1708 pg_data_t *pgdat;
1709
Con Kolivasf3fe6512006-01-06 00:11:15 -08001710 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001711 return;
1712
1713 pgdat = zone->zone_pgdat;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001714 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001715 return;
1716 if (pgdat->kswapd_max_order < order)
1717 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08001718 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001719 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001720 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001721 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001722 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001723}
1724
1725#ifdef CONFIG_PM
1726/*
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001727 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
1728 * from LRU lists system-wide, for given pass and priority, and returns the
1729 * number of reclaimed pages
1730 *
1731 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1732 */
Nigel Cunninghame07aa052006-12-22 01:07:21 -08001733static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1734 int pass, struct scan_control *sc)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001735{
1736 struct zone *zone;
1737 unsigned long nr_to_scan, ret = 0;
1738
1739 for_each_zone(zone) {
1740
1741 if (!populated_zone(zone))
1742 continue;
1743
David Rientjese815af92007-10-16 23:25:54 -07001744 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001745 continue;
1746
1747 /* For pass = 0 we don't shrink the active list */
1748 if (pass > 0) {
Christoph Lameterc8785382007-02-10 01:43:01 -08001749 zone->nr_scan_active +=
1750 (zone_page_state(zone, NR_ACTIVE) >> prio) + 1;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001751 if (zone->nr_scan_active >= nr_pages || pass > 3) {
1752 zone->nr_scan_active = 0;
Christoph Lameterc8785382007-02-10 01:43:01 -08001753 nr_to_scan = min(nr_pages,
1754 zone_page_state(zone, NR_ACTIVE));
Martin Blighbbdb3962006-10-28 10:38:25 -07001755 shrink_active_list(nr_to_scan, zone, sc, prio);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001756 }
1757 }
1758
Christoph Lameterc8785382007-02-10 01:43:01 -08001759 zone->nr_scan_inactive +=
1760 (zone_page_state(zone, NR_INACTIVE) >> prio) + 1;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001761 if (zone->nr_scan_inactive >= nr_pages || pass > 3) {
1762 zone->nr_scan_inactive = 0;
Christoph Lameterc8785382007-02-10 01:43:01 -08001763 nr_to_scan = min(nr_pages,
1764 zone_page_state(zone, NR_INACTIVE));
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001765 ret += shrink_inactive_list(nr_to_scan, zone, sc);
1766 if (ret >= nr_pages)
1767 return ret;
1768 }
1769 }
1770
1771 return ret;
1772}
1773
Andrew Morton76395d32007-01-05 16:37:05 -08001774static unsigned long count_lru_pages(void)
1775{
Christoph Lameterc8785382007-02-10 01:43:01 -08001776 return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
Andrew Morton76395d32007-01-05 16:37:05 -08001777}
1778
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001779/*
1780 * Try to free `nr_pages' of memory, system-wide, and return the number of
1781 * freed pages.
1782 *
1783 * Rather than trying to age LRUs the aim is to preserve the overall
1784 * LRU order by reclaiming preferentially
1785 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07001786 */
Andrew Morton69e05942006-03-22 00:08:19 -08001787unsigned long shrink_all_memory(unsigned long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001788{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001789 unsigned long lru_pages, nr_slab;
Andrew Morton69e05942006-03-22 00:08:19 -08001790 unsigned long ret = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001791 int pass;
1792 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001793 struct scan_control sc = {
1794 .gfp_mask = GFP_KERNEL,
1795 .may_swap = 0,
1796 .swap_cluster_max = nr_pages,
1797 .may_writepage = 1,
1798 .swappiness = vm_swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08001799 .isolate_pages = isolate_pages_global,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001800 };
1801
1802 current->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08001803
Andrew Morton76395d32007-01-05 16:37:05 -08001804 lru_pages = count_lru_pages();
Christoph Lameter972d1a72006-09-25 23:31:51 -07001805 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001806 /* If slab caches are huge, it's better to hit them first */
1807 while (nr_slab >= lru_pages) {
1808 reclaim_state.reclaimed_slab = 0;
1809 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1810 if (!reclaim_state.reclaimed_slab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001811 break;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001812
1813 ret += reclaim_state.reclaimed_slab;
1814 if (ret >= nr_pages)
1815 goto out;
1816
1817 nr_slab -= reclaim_state.reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001818 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001819
1820 /*
1821 * We try to shrink LRUs in 5 passes:
1822 * 0 = Reclaim from inactive_list only
1823 * 1 = Reclaim from active list but don't reclaim mapped
1824 * 2 = 2nd pass of type 1
1825 * 3 = Reclaim mapped (normal reclaim)
1826 * 4 = 2nd pass of type 3
1827 */
1828 for (pass = 0; pass < 5; pass++) {
1829 int prio;
1830
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001831 /* Force reclaiming mapped pages in the passes #3 and #4 */
1832 if (pass > 2) {
1833 sc.may_swap = 1;
1834 sc.swappiness = 100;
1835 }
1836
1837 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1838 unsigned long nr_to_scan = nr_pages - ret;
1839
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001840 sc.nr_scanned = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001841 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1842 if (ret >= nr_pages)
1843 goto out;
1844
1845 reclaim_state.reclaimed_slab = 0;
Andrew Morton76395d32007-01-05 16:37:05 -08001846 shrink_slab(sc.nr_scanned, sc.gfp_mask,
1847 count_lru_pages());
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001848 ret += reclaim_state.reclaimed_slab;
1849 if (ret >= nr_pages)
1850 goto out;
1851
1852 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001853 congestion_wait(WRITE, HZ / 10);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001854 }
Rafael J. Wysocki248a0302006-03-22 00:09:04 -08001855 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001856
1857 /*
1858 * If ret = 0, we could not shrink LRUs, but there may be something
1859 * in slab caches
1860 */
Andrew Morton76395d32007-01-05 16:37:05 -08001861 if (!ret) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001862 do {
1863 reclaim_state.reclaimed_slab = 0;
Andrew Morton76395d32007-01-05 16:37:05 -08001864 shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages());
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001865 ret += reclaim_state.reclaimed_slab;
1866 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
Andrew Morton76395d32007-01-05 16:37:05 -08001867 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001868
1869out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870 current->reclaim_state = NULL;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001871
Linus Torvalds1da177e2005-04-16 15:20:36 -07001872 return ret;
1873}
1874#endif
1875
Linus Torvalds1da177e2005-04-16 15:20:36 -07001876/* It's optimal to keep kswapds on the same CPUs as their memory, but
1877 not required for correctness. So if the last cpu in a node goes
1878 away, we get changed to run anywhere: as the first one comes back,
1879 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07001880static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08001881 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882{
1883 pg_data_t *pgdat;
1884 cpumask_t mask;
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07001885 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001887 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07001888 for_each_node_state(nid, N_HIGH_MEMORY) {
1889 pgdat = NODE_DATA(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001890 mask = node_to_cpumask(pgdat->node_id);
1891 if (any_online_cpu(mask) != NR_CPUS)
1892 /* One of our CPUs online: restore mask */
1893 set_cpus_allowed(pgdat->kswapd, mask);
1894 }
1895 }
1896 return NOTIFY_OK;
1897}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001898
Yasunori Goto3218ae12006-06-27 02:53:33 -07001899/*
1900 * This kswapd start function will be called by init and node-hot-add.
1901 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1902 */
1903int kswapd_run(int nid)
1904{
1905 pg_data_t *pgdat = NODE_DATA(nid);
1906 int ret = 0;
1907
1908 if (pgdat->kswapd)
1909 return 0;
1910
1911 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1912 if (IS_ERR(pgdat->kswapd)) {
1913 /* failure at boot is fatal */
1914 BUG_ON(system_state == SYSTEM_BOOTING);
1915 printk("Failed to start kswapd on node %d\n",nid);
1916 ret = -1;
1917 }
1918 return ret;
1919}
1920
Linus Torvalds1da177e2005-04-16 15:20:36 -07001921static int __init kswapd_init(void)
1922{
Yasunori Goto3218ae12006-06-27 02:53:33 -07001923 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08001924
Linus Torvalds1da177e2005-04-16 15:20:36 -07001925 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07001926 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07001927 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001928 hotcpu_notifier(cpu_callback, 0);
1929 return 0;
1930}
1931
1932module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08001933
1934#ifdef CONFIG_NUMA
1935/*
1936 * Zone reclaim mode
1937 *
1938 * If non-zero call zone_reclaim when the number of free pages falls below
1939 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08001940 */
1941int zone_reclaim_mode __read_mostly;
1942
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08001943#define RECLAIM_OFF 0
1944#define RECLAIM_ZONE (1<<0) /* Run shrink_cache on the zone */
1945#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
1946#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
1947
Christoph Lameter9eeff232006-01-18 17:42:31 -08001948/*
Christoph Lametera92f7122006-02-01 03:05:32 -08001949 * Priority for ZONE_RECLAIM. This determines the fraction of pages
1950 * of a node considered for each zone_reclaim. 4 scans 1/16th of
1951 * a zone.
1952 */
1953#define ZONE_RECLAIM_PRIORITY 4
1954
Christoph Lameter9eeff232006-01-18 17:42:31 -08001955/*
Christoph Lameter96146342006-07-03 00:24:13 -07001956 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
1957 * occur.
1958 */
1959int sysctl_min_unmapped_ratio = 1;
1960
1961/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07001962 * If the number of slab pages in a zone grows beyond this percentage then
1963 * slab reclaim needs to occur.
1964 */
1965int sysctl_min_slab_ratio = 5;
1966
1967/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08001968 * Try to free up some pages from this zone through reclaim.
1969 */
Andrew Morton179e9632006-03-22 00:08:18 -08001970static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08001971{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08001972 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08001973 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001974 struct task_struct *p = current;
1975 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08001976 int priority;
Andrew Morton05ff5132006-03-22 00:08:20 -08001977 unsigned long nr_reclaimed = 0;
Andrew Morton179e9632006-03-22 00:08:18 -08001978 struct scan_control sc = {
1979 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
1980 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
Andrew Morton69e05942006-03-22 00:08:19 -08001981 .swap_cluster_max = max_t(unsigned long, nr_pages,
1982 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08001983 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001984 .swappiness = vm_swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08001985 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08001986 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07001987 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001988
1989 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08001990 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08001991 /*
1992 * We need to be able to allocate from the reserves for RECLAIM_SWAP
1993 * and we also need to be able to write out pages for RECLAIM_WRITE
1994 * and RECLAIM_SWAP.
1995 */
1996 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08001997 reclaim_state.reclaimed_slab = 0;
1998 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08001999
Christoph Lameter0ff38492006-09-25 23:31:52 -07002000 if (zone_page_state(zone, NR_FILE_PAGES) -
2001 zone_page_state(zone, NR_FILE_MAPPED) >
2002 zone->min_unmapped_pages) {
2003 /*
2004 * Free memory by calling shrink zone with increasing
2005 * priorities until we have enough memory freed.
2006 */
2007 priority = ZONE_RECLAIM_PRIORITY;
2008 do {
Martin Bligh3bb1a8522006-10-28 10:38:24 -07002009 note_zone_scanning_priority(zone, priority);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002010 nr_reclaimed += shrink_zone(priority, zone, &sc);
2011 priority--;
2012 } while (priority >= 0 && nr_reclaimed < nr_pages);
2013 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002014
Christoph Lameter83e33a42006-09-25 23:31:53 -07002015 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2016 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002017 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002018 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002019 * many pages were freed in this zone. So we take the current
2020 * number of slab pages and shake the slab until it is reduced
2021 * by the same nr_pages that we used for reclaiming unmapped
2022 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002023 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002024 * Note that shrink_slab will free memory on all zones and may
2025 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002026 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07002027 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07002028 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2029 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07002030 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07002031
2032 /*
2033 * Update nr_reclaimed by the number of slab pages we
2034 * reclaimed from this zone.
2035 */
2036 nr_reclaimed += slab_reclaimable -
2037 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002038 }
2039
Christoph Lameter9eeff232006-01-18 17:42:31 -08002040 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002041 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
Andrew Morton05ff5132006-03-22 00:08:20 -08002042 return nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002043}
Andrew Morton179e9632006-03-22 00:08:18 -08002044
2045int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2046{
Andrew Morton179e9632006-03-22 00:08:18 -08002047 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002048 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002049
2050 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002051 * Zone reclaim reclaims unmapped file backed pages and
2052 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002053 *
Christoph Lameter96146342006-07-03 00:24:13 -07002054 * A small portion of unmapped file backed pages is needed for
2055 * file I/O otherwise pages read by file I/O will be immediately
2056 * thrown out if the zone is overallocated. So we do not reclaim
2057 * if less than a specified percentage of the zone is used by
2058 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002059 */
Christoph Lameter34aa1332006-06-30 01:55:37 -07002060 if (zone_page_state(zone, NR_FILE_PAGES) -
Christoph Lameter0ff38492006-09-25 23:31:52 -07002061 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2062 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2063 <= zone->min_slab_pages)
Christoph Lameter96146342006-07-03 00:24:13 -07002064 return 0;
Andrew Morton179e9632006-03-22 00:08:18 -08002065
David Rientjesd773ed62007-10-16 23:26:01 -07002066 if (zone_is_all_unreclaimable(zone))
2067 return 0;
2068
Andrew Morton179e9632006-03-22 00:08:18 -08002069 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002070 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002071 */
David Rientjesd773ed62007-10-16 23:26:01 -07002072 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Andrew Morton179e9632006-03-22 00:08:18 -08002073 return 0;
2074
2075 /*
2076 * Only run zone reclaim on the local zone or on zones that do not
2077 * have associated processors. This will favor the local processor
2078 * over remote processors and spread off node memory allocations
2079 * as wide as possible.
2080 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002081 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002082 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Andrew Morton179e9632006-03-22 00:08:18 -08002083 return 0;
David Rientjesd773ed62007-10-16 23:26:01 -07002084
2085 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2086 return 0;
2087 ret = __zone_reclaim(zone, gfp_mask, order);
2088 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2089
2090 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002091}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002092#endif