blob: 46fdaa546b8d0493acb74b2b428693d46e0b14b3 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
16#include <linux/slab.h>
17#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070022#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070023#include <linux/file.h>
24#include <linux/writeback.h>
25#include <linux/blkdev.h>
26#include <linux/buffer_head.h> /* for try_to_release_page(),
27 buffer_heads_over_limit */
28#include <linux/mm_inline.h>
29#include <linux/pagevec.h>
30#include <linux/backing-dev.h>
31#include <linux/rmap.h>
32#include <linux/topology.h>
33#include <linux/cpu.h>
34#include <linux/cpuset.h>
35#include <linux/notifier.h>
36#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080037#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070038#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080039#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080040#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070041#include <linux/delayacct.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070042
43#include <asm/tlbflush.h>
44#include <asm/div64.h>
45
46#include <linux/swapops.h>
47
Nick Piggin0f8053a2006-03-22 00:08:33 -080048#include "internal.h"
49
Linus Torvalds1da177e2005-04-16 15:20:36 -070050struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070051 /* Incremented by the number of inactive pages that were scanned */
52 unsigned long nr_scanned;
53
Linus Torvalds1da177e2005-04-16 15:20:36 -070054 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040055 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070056
57 int may_writepage;
58
Christoph Lameterf1fd1062006-01-18 17:42:30 -080059 /* Can pages be swapped as part of reclaim? */
60 int may_swap;
61
Linus Torvalds1da177e2005-04-16 15:20:36 -070062 /* This context's SWAP_CLUSTER_MAX. If freeing memory for
63 * suspend, we effectively ignore SWAP_CLUSTER_MAX.
64 * In this context, it doesn't matter that we scan the
65 * whole list at once. */
66 int swap_cluster_max;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -070067
68 int swappiness;
Nick Piggin408d8542006-09-25 23:31:27 -070069
70 int all_unreclaimable;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070071
72 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080073
74 /* Which cgroup do we reclaim from */
75 struct mem_cgroup *mem_cgroup;
76
77 /* Pluggable isolate pages callback */
78 unsigned long (*isolate_pages)(unsigned long nr, struct list_head *dst,
79 unsigned long *scanned, int order, int mode,
80 struct zone *z, struct mem_cgroup *mem_cont,
81 int active);
Linus Torvalds1da177e2005-04-16 15:20:36 -070082};
83
Linus Torvalds1da177e2005-04-16 15:20:36 -070084#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
85
86#ifdef ARCH_HAS_PREFETCH
87#define prefetch_prev_lru_page(_page, _base, _field) \
88 do { \
89 if ((_page)->lru.prev != _base) { \
90 struct page *prev; \
91 \
92 prev = lru_to_page(&(_page->lru)); \
93 prefetch(&prev->_field); \
94 } \
95 } while (0)
96#else
97#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
98#endif
99
100#ifdef ARCH_HAS_PREFETCHW
101#define prefetchw_prev_lru_page(_page, _base, _field) \
102 do { \
103 if ((_page)->lru.prev != _base) { \
104 struct page *prev; \
105 \
106 prev = lru_to_page(&(_page->lru)); \
107 prefetchw(&prev->_field); \
108 } \
109 } while (0)
110#else
111#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
112#endif
113
114/*
115 * From 0 .. 100. Higher means more swappy.
116 */
117int vm_swappiness = 60;
Andrew Mortonbd1e22b2006-06-23 02:03:47 -0700118long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700119
120static LIST_HEAD(shrinker_list);
121static DECLARE_RWSEM(shrinker_rwsem);
122
Balbir Singh00f0b822008-03-04 14:28:39 -0800123#ifdef CONFIG_CGROUP_MEM_RES_CTLR
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800124#define scan_global_lru(sc) (!(sc)->mem_cgroup)
125#else
126#define scan_global_lru(sc) (1)
127#endif
128
Linus Torvalds1da177e2005-04-16 15:20:36 -0700129/*
130 * Add a shrinker callback to be called from the vm
131 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700132void register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700133{
Rusty Russell8e1f9362007-07-17 04:03:17 -0700134 shrinker->nr = 0;
135 down_write(&shrinker_rwsem);
136 list_add_tail(&shrinker->list, &shrinker_list);
137 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700138}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700139EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700140
141/*
142 * Remove one
143 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700144void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700145{
146 down_write(&shrinker_rwsem);
147 list_del(&shrinker->list);
148 up_write(&shrinker_rwsem);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700149}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700150EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700151
152#define SHRINK_BATCH 128
153/*
154 * Call the shrink functions to age shrinkable caches
155 *
156 * Here we assume it costs one seek to replace a lru page and that it also
157 * takes a seek to recreate a cache object. With this in mind we age equal
158 * percentages of the lru and ageable caches. This should balance the seeks
159 * generated by these structures.
160 *
Simon Arlott183ff222007-10-20 01:27:18 +0200161 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700162 * slab to avoid swapping.
163 *
164 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
165 *
166 * `lru_pages' represents the number of on-LRU pages in all the zones which
167 * are eligible for the caller's allocation attempt. It is used for balancing
168 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700169 *
170 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700171 */
Andrew Morton69e05942006-03-22 00:08:19 -0800172unsigned long shrink_slab(unsigned long scanned, gfp_t gfp_mask,
173 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700174{
175 struct shrinker *shrinker;
Andrew Morton69e05942006-03-22 00:08:19 -0800176 unsigned long ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177
178 if (scanned == 0)
179 scanned = SWAP_CLUSTER_MAX;
180
181 if (!down_read_trylock(&shrinker_rwsem))
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700182 return 1; /* Assume we'll be able to shrink next time */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700183
184 list_for_each_entry(shrinker, &shrinker_list, list) {
185 unsigned long long delta;
186 unsigned long total_scan;
Rusty Russell8e1f9362007-07-17 04:03:17 -0700187 unsigned long max_pass = (*shrinker->shrink)(0, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700188
189 delta = (4 * scanned) / shrinker->seeks;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800190 delta *= max_pass;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700191 do_div(delta, lru_pages + 1);
192 shrinker->nr += delta;
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800193 if (shrinker->nr < 0) {
194 printk(KERN_ERR "%s: nr=%ld\n",
Harvey Harrisond40cee22008-04-30 00:55:07 -0700195 __func__, shrinker->nr);
Andrea Arcangeliea164d72005-11-28 13:44:15 -0800196 shrinker->nr = max_pass;
197 }
198
199 /*
200 * Avoid risking looping forever due to too large nr value:
201 * never try to free more than twice the estimate number of
202 * freeable entries.
203 */
204 if (shrinker->nr > max_pass * 2)
205 shrinker->nr = max_pass * 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700206
207 total_scan = shrinker->nr;
208 shrinker->nr = 0;
209
210 while (total_scan >= SHRINK_BATCH) {
211 long this_scan = SHRINK_BATCH;
212 int shrink_ret;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700213 int nr_before;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700214
Rusty Russell8e1f9362007-07-17 04:03:17 -0700215 nr_before = (*shrinker->shrink)(0, gfp_mask);
216 shrink_ret = (*shrinker->shrink)(this_scan, gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217 if (shrink_ret == -1)
218 break;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700219 if (shrink_ret < nr_before)
220 ret += nr_before - shrink_ret;
Christoph Lameterf8891e52006-06-30 01:55:45 -0700221 count_vm_events(SLABS_SCANNED, this_scan);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700222 total_scan -= this_scan;
223
224 cond_resched();
225 }
226
227 shrinker->nr += total_scan;
228 }
229 up_read(&shrinker_rwsem);
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700230 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700231}
232
233/* Called without lock on whether page is mapped, so answer is unstable */
234static inline int page_mapping_inuse(struct page *page)
235{
236 struct address_space *mapping;
237
238 /* Page is in somebody's page tables. */
239 if (page_mapped(page))
240 return 1;
241
242 /* Be more reluctant to reclaim swapcache than pagecache */
243 if (PageSwapCache(page))
244 return 1;
245
246 mapping = page_mapping(page);
247 if (!mapping)
248 return 0;
249
250 /* File is mmap'd by somebody? */
251 return mapping_mapped(mapping);
252}
253
254static inline int is_page_cache_freeable(struct page *page)
255{
256 return page_count(page) - !!PagePrivate(page) == 2;
257}
258
259static int may_write_to_queue(struct backing_dev_info *bdi)
260{
Christoph Lameter930d9152006-01-08 01:00:47 -0800261 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700262 return 1;
263 if (!bdi_write_congested(bdi))
264 return 1;
265 if (bdi == current->backing_dev_info)
266 return 1;
267 return 0;
268}
269
270/*
271 * We detected a synchronous write error writing a page out. Probably
272 * -ENOSPC. We need to propagate that into the address_space for a subsequent
273 * fsync(), msync() or close().
274 *
275 * The tricky part is that after writepage we cannot touch the mapping: nothing
276 * prevents it from being freed up. But we have a ref on the page and once
277 * that page is locked, the mapping is pinned.
278 *
279 * We're allowed to run sleeping lock_page() here because we know the caller has
280 * __GFP_FS.
281 */
282static void handle_write_error(struct address_space *mapping,
283 struct page *page, int error)
284{
285 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700286 if (page_mapping(page) == mapping)
287 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700288 unlock_page(page);
289}
290
Andy Whitcroftc661b072007-08-22 14:01:26 -0700291/* Request for sync pageout. */
292enum pageout_io {
293 PAGEOUT_IO_ASYNC,
294 PAGEOUT_IO_SYNC,
295};
296
Christoph Lameter04e62a22006-06-23 02:03:38 -0700297/* possible outcome of pageout() */
298typedef enum {
299 /* failed to write page out, page is locked */
300 PAGE_KEEP,
301 /* move page to the active list, page is locked */
302 PAGE_ACTIVATE,
303 /* page has been sent to the disk successfully, page is unlocked */
304 PAGE_SUCCESS,
305 /* page is clean and locked */
306 PAGE_CLEAN,
307} pageout_t;
308
Linus Torvalds1da177e2005-04-16 15:20:36 -0700309/*
Andrew Morton1742f192006-03-22 00:08:21 -0800310 * pageout is called by shrink_page_list() for each dirty page.
311 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700313static pageout_t pageout(struct page *page, struct address_space *mapping,
314 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700315{
316 /*
317 * If the page is dirty, only perform writeback if that write
318 * will be non-blocking. To prevent this allocation from being
319 * stalled by pagecache activity. But note that there may be
320 * stalls if we need to run get_block(). We could test
321 * PagePrivate for that.
322 *
323 * If this process is currently in generic_file_write() against
324 * this page's queue, we can perform writeback even if that
325 * will block.
326 *
327 * If the page is swapcache, write it back even if that would
328 * block, for some throttling. This happens by accident, because
329 * swap_backing_dev_info is bust: it doesn't reflect the
330 * congestion state of the swapdevs. Easy to fix, if needed.
331 * See swapfile.c:page_queue_congested().
332 */
333 if (!is_page_cache_freeable(page))
334 return PAGE_KEEP;
335 if (!mapping) {
336 /*
337 * Some data journaling orphaned pages can have
338 * page->mapping == NULL while being dirty with clean buffers.
339 */
akpm@osdl.org323aca62005-04-16 15:24:06 -0700340 if (PagePrivate(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700341 if (try_to_free_buffers(page)) {
342 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700343 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700344 return PAGE_CLEAN;
345 }
346 }
347 return PAGE_KEEP;
348 }
349 if (mapping->a_ops->writepage == NULL)
350 return PAGE_ACTIVATE;
351 if (!may_write_to_queue(mapping->backing_dev_info))
352 return PAGE_KEEP;
353
354 if (clear_page_dirty_for_io(page)) {
355 int res;
356 struct writeback_control wbc = {
357 .sync_mode = WB_SYNC_NONE,
358 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700359 .range_start = 0,
360 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 .nonblocking = 1,
362 .for_reclaim = 1,
363 };
364
365 SetPageReclaim(page);
366 res = mapping->a_ops->writepage(page, &wbc);
367 if (res < 0)
368 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800369 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370 ClearPageReclaim(page);
371 return PAGE_ACTIVATE;
372 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700373
374 /*
375 * Wait on writeback if requested to. This happens when
376 * direct reclaiming a large contiguous area and the
377 * first attempt to free a range of pages fails.
378 */
379 if (PageWriteback(page) && sync_writeback == PAGEOUT_IO_SYNC)
380 wait_on_page_writeback(page);
381
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382 if (!PageWriteback(page)) {
383 /* synchronous write or broken a_ops? */
384 ClearPageReclaim(page);
385 }
Andrew Mortone129b5c2006-09-27 01:50:00 -0700386 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700387 return PAGE_SUCCESS;
388 }
389
390 return PAGE_CLEAN;
391}
392
Andrew Mortona649fd92006-10-17 00:09:36 -0700393/*
Nick Piggine2867812008-07-25 19:45:30 -0700394 * Same as remove_mapping, but if the page is removed from the mapping, it
395 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700396 */
Nick Piggine2867812008-07-25 19:45:30 -0700397static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800398{
Nick Piggin28e4d962006-09-25 23:31:23 -0700399 BUG_ON(!PageLocked(page));
400 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800401
Nick Piggin19fd6232008-07-25 19:45:32 -0700402 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800403 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700404 * The non racy check for a busy page.
405 *
406 * Must be careful with the order of the tests. When someone has
407 * a ref to the page, it may be possible that they dirty it then
408 * drop the reference. So if PageDirty is tested before page_count
409 * here, then the following race may occur:
410 *
411 * get_user_pages(&page);
412 * [user mapping goes away]
413 * write_to(page);
414 * !PageDirty(page) [good]
415 * SetPageDirty(page);
416 * put_page(page);
417 * !page_count(page) [good, discard it]
418 *
419 * [oops, our write_to data is lost]
420 *
421 * Reversing the order of the tests ensures such a situation cannot
422 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
423 * load is not satisfied before that of page->_count.
424 *
425 * Note that if SetPageDirty is always performed via set_page_dirty,
426 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800427 */
Nick Piggine2867812008-07-25 19:45:30 -0700428 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800429 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700430 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
431 if (unlikely(PageDirty(page))) {
432 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800433 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700434 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800435
436 if (PageSwapCache(page)) {
437 swp_entry_t swap = { .val = page_private(page) };
438 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700439 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800440 swap_free(swap);
Nick Piggine2867812008-07-25 19:45:30 -0700441 } else {
442 __remove_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700443 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800444 }
445
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800446 return 1;
447
448cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700449 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800450 return 0;
451}
452
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453/*
Nick Piggine2867812008-07-25 19:45:30 -0700454 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
455 * someone else has a ref on the page, abort and return 0. If it was
456 * successfully detached, return 1. Assumes the caller has a single ref on
457 * this page.
458 */
459int remove_mapping(struct address_space *mapping, struct page *page)
460{
461 if (__remove_mapping(mapping, page)) {
462 /*
463 * Unfreezing the refcount with 1 rather than 2 effectively
464 * drops the pagecache ref for us without requiring another
465 * atomic operation.
466 */
467 page_unfreeze_refs(page, 1);
468 return 1;
469 }
470 return 0;
471}
472
473/*
Andrew Morton1742f192006-03-22 00:08:21 -0800474 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700475 */
Andrew Morton1742f192006-03-22 00:08:21 -0800476static unsigned long shrink_page_list(struct list_head *page_list,
Andy Whitcroftc661b072007-08-22 14:01:26 -0700477 struct scan_control *sc,
478 enum pageout_io sync_writeback)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700479{
480 LIST_HEAD(ret_pages);
481 struct pagevec freed_pvec;
482 int pgactivate = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800483 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700484
485 cond_resched();
486
487 pagevec_init(&freed_pvec, 1);
488 while (!list_empty(page_list)) {
489 struct address_space *mapping;
490 struct page *page;
491 int may_enter_fs;
492 int referenced;
493
494 cond_resched();
495
496 page = lru_to_page(page_list);
497 list_del(&page->lru);
498
Nick Piggin529ae9a2008-08-02 12:01:03 +0200499 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700500 goto keep;
501
Nick Piggin725d7042006-09-25 23:30:55 -0700502 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700503
504 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800505
506 if (!sc->may_swap && page_mapped(page))
507 goto keep_locked;
508
Linus Torvalds1da177e2005-04-16 15:20:36 -0700509 /* Double the slab pressure for mapped and swapcache pages */
510 if (page_mapped(page) || PageSwapCache(page))
511 sc->nr_scanned++;
512
Andy Whitcroftc661b072007-08-22 14:01:26 -0700513 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
514 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
515
516 if (PageWriteback(page)) {
517 /*
518 * Synchronous reclaim is performed in two passes,
519 * first an asynchronous pass over the list to
520 * start parallel writeback, and a second synchronous
521 * pass to wait for the IO to complete. Wait here
522 * for any page for which writeback has already
523 * started.
524 */
525 if (sync_writeback == PAGEOUT_IO_SYNC && may_enter_fs)
526 wait_on_page_writeback(page);
Andrew Morton4dd4b922008-03-24 12:29:52 -0700527 else
Andy Whitcroftc661b072007-08-22 14:01:26 -0700528 goto keep_locked;
529 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700530
Balbir Singhbed71612008-02-07 00:14:01 -0800531 referenced = page_referenced(page, 1, sc->mem_cgroup);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700532 /* In active use or really unfreeable? Activate it. */
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700533 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER &&
534 referenced && page_mapping_inuse(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700535 goto activate_locked;
536
537#ifdef CONFIG_SWAP
538 /*
539 * Anonymous process memory has backing store?
540 * Try to allocate it some swap space here.
541 */
Christoph Lameter6e5ef1a2006-03-22 00:08:45 -0800542 if (PageAnon(page) && !PageSwapCache(page))
Christoph Lameter1480a542006-01-08 01:00:53 -0800543 if (!add_to_swap(page, GFP_ATOMIC))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544 goto activate_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700545#endif /* CONFIG_SWAP */
546
547 mapping = page_mapping(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700548
549 /*
550 * The page is mapped into the page tables of one or more
551 * processes. Try to unmap it here.
552 */
553 if (page_mapped(page) && mapping) {
Christoph Lametera48d07a2006-02-01 03:05:38 -0800554 switch (try_to_unmap(page, 0)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555 case SWAP_FAIL:
556 goto activate_locked;
557 case SWAP_AGAIN:
558 goto keep_locked;
559 case SWAP_SUCCESS:
560 ; /* try to free the page below */
561 }
562 }
563
564 if (PageDirty(page)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700565 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER && referenced)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700567 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700568 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800569 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 goto keep_locked;
571
572 /* Page is dirty, try to write it out here */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700573 switch (pageout(page, mapping, sync_writeback)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 case PAGE_KEEP:
575 goto keep_locked;
576 case PAGE_ACTIVATE:
577 goto activate_locked;
578 case PAGE_SUCCESS:
Andrew Morton4dd4b922008-03-24 12:29:52 -0700579 if (PageWriteback(page) || PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700580 goto keep;
581 /*
582 * A synchronous write - probably a ramdisk. Go
583 * ahead and try to reclaim the page.
584 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200585 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586 goto keep;
587 if (PageDirty(page) || PageWriteback(page))
588 goto keep_locked;
589 mapping = page_mapping(page);
590 case PAGE_CLEAN:
591 ; /* try to free the page below */
592 }
593 }
594
595 /*
596 * If the page has buffers, try to free the buffer mappings
597 * associated with this page. If we succeed we try to free
598 * the page as well.
599 *
600 * We do this even if the page is PageDirty().
601 * try_to_release_page() does not perform I/O, but it is
602 * possible for a page to have PageDirty set, but it is actually
603 * clean (all its buffers are clean). This happens if the
604 * buffers were written out directly, with submit_bh(). ext3
605 * will do this, as well as the blockdev mapping.
606 * try_to_release_page() will discover that cleanness and will
607 * drop the buffers and mark the page clean - it can be freed.
608 *
609 * Rarely, pages can have buffers and no ->mapping. These are
610 * the pages which were not successfully invalidated in
611 * truncate_complete_page(). We try to drop those buffers here
612 * and if that worked, and the page is no longer mapped into
613 * process address space (page_count == 1) it can be freed.
614 * Otherwise, leave the page on the LRU so it is swappable.
615 */
616 if (PagePrivate(page)) {
617 if (!try_to_release_page(page, sc->gfp_mask))
618 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -0700619 if (!mapping && page_count(page) == 1) {
620 unlock_page(page);
621 if (put_page_testzero(page))
622 goto free_it;
623 else {
624 /*
625 * rare race with speculative reference.
626 * the speculative reference will free
627 * this page shortly, so we may
628 * increment nr_reclaimed here (and
629 * leave it off the LRU).
630 */
631 nr_reclaimed++;
632 continue;
633 }
634 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635 }
636
Nick Piggine2867812008-07-25 19:45:30 -0700637 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800638 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700639
Linus Torvalds1da177e2005-04-16 15:20:36 -0700640 unlock_page(page);
Nick Piggine2867812008-07-25 19:45:30 -0700641free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -0800642 nr_reclaimed++;
Nick Piggine2867812008-07-25 19:45:30 -0700643 if (!pagevec_add(&freed_pvec, page)) {
644 __pagevec_free(&freed_pvec);
645 pagevec_reinit(&freed_pvec);
646 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647 continue;
648
649activate_locked:
650 SetPageActive(page);
651 pgactivate++;
652keep_locked:
653 unlock_page(page);
654keep:
655 list_add(&page->lru, &ret_pages);
Nick Piggin725d7042006-09-25 23:30:55 -0700656 VM_BUG_ON(PageLRU(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657 }
658 list_splice(&ret_pages, page_list);
659 if (pagevec_count(&freed_pvec))
Nick Piggine2867812008-07-25 19:45:30 -0700660 __pagevec_free(&freed_pvec);
Christoph Lameterf8891e52006-06-30 01:55:45 -0700661 count_vm_events(PGACTIVATE, pgactivate);
Andrew Morton05ff5132006-03-22 00:08:20 -0800662 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663}
664
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700665/* LRU Isolation modes. */
666#define ISOLATE_INACTIVE 0 /* Isolate inactive pages. */
667#define ISOLATE_ACTIVE 1 /* Isolate active pages. */
668#define ISOLATE_BOTH 2 /* Isolate both active and inactive pages. */
669
670/*
671 * Attempt to remove the specified page from its LRU. Only take this page
672 * if it is of the appropriate PageActive status. Pages which are being
673 * freed elsewhere are also ignored.
674 *
675 * page: page to consider
676 * mode: one of the LRU isolation modes defined above
677 *
678 * returns 0 on success, -ve errno on failure.
679 */
Balbir Singh66e17072008-02-07 00:13:56 -0800680int __isolate_lru_page(struct page *page, int mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700681{
682 int ret = -EINVAL;
683
684 /* Only take pages on the LRU. */
685 if (!PageLRU(page))
686 return ret;
687
688 /*
689 * When checking the active state, we need to be sure we are
690 * dealing with comparible boolean values. Take the logical not
691 * of each.
692 */
693 if (mode != ISOLATE_BOTH && (!PageActive(page) != !mode))
694 return ret;
695
696 ret = -EBUSY;
697 if (likely(get_page_unless_zero(page))) {
698 /*
699 * Be careful not to clear PageLRU until after we're
700 * sure the page is not being freed elsewhere -- the
701 * page release code relies on it.
702 */
703 ClearPageLRU(page);
704 ret = 0;
705 }
706
707 return ret;
708}
709
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800710/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700711 * zone->lru_lock is heavily contended. Some of the functions that
712 * shrink the lists perform better by taking out a batch of pages
713 * and working on them outside the LRU lock.
714 *
715 * For pagecache intensive workloads, this function is the hottest
716 * spot in the kernel (apart from copy_*_user functions).
717 *
718 * Appropriate locks must be held before calling this function.
719 *
720 * @nr_to_scan: The number of pages to look through on the list.
721 * @src: The LRU list to pull pages off.
722 * @dst: The temp list to put pages on to.
723 * @scanned: The number of pages that were scanned.
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700724 * @order: The caller's attempted allocation order
725 * @mode: One of the LRU isolation modes
Linus Torvalds1da177e2005-04-16 15:20:36 -0700726 *
727 * returns how many pages were moved onto *@dst.
728 */
Andrew Morton69e05942006-03-22 00:08:19 -0800729static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
730 struct list_head *src, struct list_head *dst,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700731 unsigned long *scanned, int order, int mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700732{
Andrew Morton69e05942006-03-22 00:08:19 -0800733 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800734 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700735
Wu Fengguangc9b02d92006-03-22 00:08:23 -0800736 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700737 struct page *page;
738 unsigned long pfn;
739 unsigned long end_pfn;
740 unsigned long page_pfn;
741 int zone_id;
742
Linus Torvalds1da177e2005-04-16 15:20:36 -0700743 page = lru_to_page(src);
744 prefetchw_prev_lru_page(page, src, flags);
745
Nick Piggin725d7042006-09-25 23:30:55 -0700746 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800747
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700748 switch (__isolate_lru_page(page, mode)) {
749 case 0:
750 list_move(&page->lru, dst);
Nick Piggin7c8ee9a2006-03-22 00:08:03 -0800751 nr_taken++;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700752 break;
Nick Piggin46453a62006-03-22 00:07:58 -0800753
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700754 case -EBUSY:
755 /* else it is being freed elsewhere */
756 list_move(&page->lru, src);
757 continue;
758
759 default:
760 BUG();
761 }
762
763 if (!order)
764 continue;
765
766 /*
767 * Attempt to take all pages in the order aligned region
768 * surrounding the tag page. Only take those pages of
769 * the same active state as that tag page. We may safely
770 * round the target page pfn down to the requested order
771 * as the mem_map is guarenteed valid out to MAX_ORDER,
772 * where that page is in a different zone we will detect
773 * it from its zone id and abort this block scan.
774 */
775 zone_id = page_zone_id(page);
776 page_pfn = page_to_pfn(page);
777 pfn = page_pfn & ~((1 << order) - 1);
778 end_pfn = pfn + (1 << order);
779 for (; pfn < end_pfn; pfn++) {
780 struct page *cursor_page;
781
782 /* The target page is in the block, ignore it. */
783 if (unlikely(pfn == page_pfn))
784 continue;
785
786 /* Avoid holes within the zone. */
787 if (unlikely(!pfn_valid_within(pfn)))
788 break;
789
790 cursor_page = pfn_to_page(pfn);
791 /* Check that we have not crossed a zone boundary. */
792 if (unlikely(page_zone_id(cursor_page) != zone_id))
793 continue;
794 switch (__isolate_lru_page(cursor_page, mode)) {
795 case 0:
796 list_move(&cursor_page->lru, dst);
797 nr_taken++;
798 scan++;
799 break;
800
801 case -EBUSY:
802 /* else it is being freed elsewhere */
803 list_move(&cursor_page->lru, src);
804 default:
805 break;
806 }
807 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 }
809
810 *scanned = scan;
811 return nr_taken;
812}
813
Balbir Singh66e17072008-02-07 00:13:56 -0800814static unsigned long isolate_pages_global(unsigned long nr,
815 struct list_head *dst,
816 unsigned long *scanned, int order,
817 int mode, struct zone *z,
818 struct mem_cgroup *mem_cont,
819 int active)
820{
821 if (active)
Christoph Lameterb69408e2008-10-18 20:26:14 -0700822 return isolate_lru_pages(nr, &z->lru[LRU_ACTIVE].list, dst,
Balbir Singh66e17072008-02-07 00:13:56 -0800823 scanned, order, mode);
824 else
Christoph Lameterb69408e2008-10-18 20:26:14 -0700825 return isolate_lru_pages(nr, &z->lru[LRU_INACTIVE].list, dst,
Balbir Singh66e17072008-02-07 00:13:56 -0800826 scanned, order, mode);
827}
828
Linus Torvalds1da177e2005-04-16 15:20:36 -0700829/*
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700830 * clear_active_flags() is a helper for shrink_active_list(), clearing
831 * any active bits from the pages in the list.
832 */
833static unsigned long clear_active_flags(struct list_head *page_list)
834{
835 int nr_active = 0;
836 struct page *page;
837
838 list_for_each_entry(page, page_list, lru)
839 if (PageActive(page)) {
840 ClearPageActive(page);
841 nr_active++;
842 }
843
844 return nr_active;
845}
846
Nick Piggin62695a82008-10-18 20:26:09 -0700847/**
848 * isolate_lru_page - tries to isolate a page from its LRU list
849 * @page: page to isolate from its LRU list
850 *
851 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
852 * vmstat statistic corresponding to whatever LRU list the page was on.
853 *
854 * Returns 0 if the page was removed from an LRU list.
855 * Returns -EBUSY if the page was not on an LRU list.
856 *
857 * The returned page will have PageLRU() cleared. If it was found on
858 * the active list, it will have PageActive set. That flag may need
859 * to be cleared by the caller before letting the page go.
860 *
861 * The vmstat statistic corresponding to the list on which the page was
862 * found will be decremented.
863 *
864 * Restrictions:
865 * (1) Must be called with an elevated refcount on the page. This is a
866 * fundamentnal difference from isolate_lru_pages (which is called
867 * without a stable reference).
868 * (2) the lru_lock must not be held.
869 * (3) interrupts must be enabled.
870 */
871int isolate_lru_page(struct page *page)
872{
873 int ret = -EBUSY;
874
875 if (PageLRU(page)) {
876 struct zone *zone = page_zone(page);
877
878 spin_lock_irq(&zone->lru_lock);
879 if (PageLRU(page) && get_page_unless_zero(page)) {
880 ret = 0;
881 ClearPageLRU(page);
882 if (PageActive(page))
883 del_page_from_active_list(zone, page);
884 else
885 del_page_from_inactive_list(zone, page);
886 }
887 spin_unlock_irq(&zone->lru_lock);
888 }
889 return ret;
890}
891
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700892/*
Andrew Morton1742f192006-03-22 00:08:21 -0800893 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
894 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700895 */
Andrew Morton1742f192006-03-22 00:08:21 -0800896static unsigned long shrink_inactive_list(unsigned long max_scan,
897 struct zone *zone, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700898{
899 LIST_HEAD(page_list);
900 struct pagevec pvec;
Andrew Morton69e05942006-03-22 00:08:19 -0800901 unsigned long nr_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800902 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700903
904 pagevec_init(&pvec, 1);
905
906 lru_add_drain();
907 spin_lock_irq(&zone->lru_lock);
Andrew Morton69e05942006-03-22 00:08:19 -0800908 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700909 struct page *page;
Andrew Morton69e05942006-03-22 00:08:19 -0800910 unsigned long nr_taken;
911 unsigned long nr_scan;
912 unsigned long nr_freed;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700913 unsigned long nr_active;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914
Balbir Singh66e17072008-02-07 00:13:56 -0800915 nr_taken = sc->isolate_pages(sc->swap_cluster_max,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700916 &page_list, &nr_scan, sc->order,
917 (sc->order > PAGE_ALLOC_COSTLY_ORDER)?
Balbir Singh66e17072008-02-07 00:13:56 -0800918 ISOLATE_BOTH : ISOLATE_INACTIVE,
919 zone, sc->mem_cgroup, 0);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700920 nr_active = clear_active_flags(&page_list);
Andy Whitcrofte9187bd2007-08-22 14:01:25 -0700921 __count_vm_events(PGDEACTIVATE, nr_active);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -0700922
923 __mod_zone_page_state(zone, NR_ACTIVE, -nr_active);
924 __mod_zone_page_state(zone, NR_INACTIVE,
925 -(nr_taken - nr_active));
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -0800926 if (scan_global_lru(sc))
927 zone->pages_scanned += nr_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700928 spin_unlock_irq(&zone->lru_lock);
929
Andrew Morton69e05942006-03-22 00:08:19 -0800930 nr_scanned += nr_scan;
Andy Whitcroftc661b072007-08-22 14:01:26 -0700931 nr_freed = shrink_page_list(&page_list, sc, PAGEOUT_IO_ASYNC);
932
933 /*
934 * If we are direct reclaiming for contiguous pages and we do
935 * not reclaim everything in the list, try again and wait
936 * for IO to complete. This will stall high-order allocations
937 * but that should be acceptable to the caller
938 */
939 if (nr_freed < nr_taken && !current_is_kswapd() &&
940 sc->order > PAGE_ALLOC_COSTLY_ORDER) {
941 congestion_wait(WRITE, HZ/10);
942
943 /*
944 * The attempt at page out may have made some
945 * of the pages active, mark them inactive again.
946 */
947 nr_active = clear_active_flags(&page_list);
948 count_vm_events(PGDEACTIVATE, nr_active);
949
950 nr_freed += shrink_page_list(&page_list, sc,
951 PAGEOUT_IO_SYNC);
952 }
953
Andrew Morton05ff5132006-03-22 00:08:20 -0800954 nr_reclaimed += nr_freed;
Nick Piggina74609f2006-01-06 00:11:20 -0800955 local_irq_disable();
956 if (current_is_kswapd()) {
Christoph Lameterf8891e52006-06-30 01:55:45 -0700957 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scan);
958 __count_vm_events(KSWAPD_STEAL, nr_freed);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -0800959 } else if (scan_global_lru(sc))
Christoph Lameterf8891e52006-06-30 01:55:45 -0700960 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scan);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -0800961
Shantanu Goel918d3f92006-12-29 16:48:59 -0800962 __count_zone_vm_events(PGSTEAL, zone, nr_freed);
Nick Piggina74609f2006-01-06 00:11:20 -0800963
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800964 if (nr_taken == 0)
965 goto done;
966
Nick Piggina74609f2006-01-06 00:11:20 -0800967 spin_lock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 /*
969 * Put back any unfreeable pages.
970 */
971 while (!list_empty(&page_list)) {
972 page = lru_to_page(&page_list);
Nick Piggin725d7042006-09-25 23:30:55 -0700973 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -0800974 SetPageLRU(page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 list_del(&page->lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -0700976 add_page_to_lru_list(zone, page, page_lru(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -0700977 if (!pagevec_add(&pvec, page)) {
978 spin_unlock_irq(&zone->lru_lock);
979 __pagevec_release(&pvec);
980 spin_lock_irq(&zone->lru_lock);
981 }
982 }
Andrew Morton69e05942006-03-22 00:08:19 -0800983 } while (nr_scanned < max_scan);
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800984 spin_unlock(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700985done:
Wu Fengguangfb8d14e2006-03-22 00:08:28 -0800986 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700987 pagevec_release(&pvec);
Andrew Morton05ff5132006-03-22 00:08:20 -0800988 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700989}
990
Martin Bligh3bb1a8522006-10-28 10:38:24 -0700991/*
992 * We are about to scan this zone at a certain priority level. If that priority
993 * level is smaller (ie: more urgent) than the previous priority, then note
994 * that priority level within the zone. This is done so that when the next
995 * process comes in to scan this zone, it will immediately start out at this
996 * priority level rather than having to build up its own scanning priority.
997 * Here, this priority affects only the reclaim-mapped threshold.
998 */
999static inline void note_zone_scanning_priority(struct zone *zone, int priority)
1000{
1001 if (priority < zone->prev_priority)
1002 zone->prev_priority = priority;
1003}
1004
Nick Piggin4ff1ffb2006-09-25 23:31:28 -07001005static inline int zone_is_near_oom(struct zone *zone)
1006{
Christoph Lameterc8785382007-02-10 01:43:01 -08001007 return zone->pages_scanned >= (zone_page_state(zone, NR_ACTIVE)
1008 + zone_page_state(zone, NR_INACTIVE))*3;
Nick Piggin4ff1ffb2006-09-25 23:31:28 -07001009}
1010
Linus Torvalds1da177e2005-04-16 15:20:36 -07001011/*
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001012 * Determine we should try to reclaim mapped pages.
1013 * This is called only when sc->mem_cgroup is NULL.
1014 */
1015static int calc_reclaim_mapped(struct scan_control *sc, struct zone *zone,
1016 int priority)
1017{
1018 long mapped_ratio;
1019 long distress;
1020 long swap_tendency;
1021 long imbalance;
1022 int reclaim_mapped = 0;
1023 int prev_priority;
1024
1025 if (scan_global_lru(sc) && zone_is_near_oom(zone))
1026 return 1;
1027 /*
1028 * `distress' is a measure of how much trouble we're having
1029 * reclaiming pages. 0 -> no problems. 100 -> great trouble.
1030 */
1031 if (scan_global_lru(sc))
1032 prev_priority = zone->prev_priority;
1033 else
1034 prev_priority = mem_cgroup_get_reclaim_priority(sc->mem_cgroup);
1035
1036 distress = 100 >> min(prev_priority, priority);
1037
1038 /*
1039 * The point of this algorithm is to decide when to start
1040 * reclaiming mapped memory instead of just pagecache. Work out
1041 * how much memory
1042 * is mapped.
1043 */
1044 if (scan_global_lru(sc))
1045 mapped_ratio = ((global_page_state(NR_FILE_MAPPED) +
1046 global_page_state(NR_ANON_PAGES)) * 100) /
1047 vm_total_pages;
1048 else
1049 mapped_ratio = mem_cgroup_calc_mapped_ratio(sc->mem_cgroup);
1050
1051 /*
1052 * Now decide how much we really want to unmap some pages. The
1053 * mapped ratio is downgraded - just because there's a lot of
1054 * mapped memory doesn't necessarily mean that page reclaim
1055 * isn't succeeding.
1056 *
1057 * The distress ratio is important - we don't want to start
1058 * going oom.
1059 *
1060 * A 100% value of vm_swappiness overrides this algorithm
1061 * altogether.
1062 */
1063 swap_tendency = mapped_ratio / 2 + distress + sc->swappiness;
1064
1065 /*
1066 * If there's huge imbalance between active and inactive
1067 * (think active 100 times larger than inactive) we should
1068 * become more permissive, or the system will take too much
1069 * cpu before it start swapping during memory pressure.
1070 * Distress is about avoiding early-oom, this is about
1071 * making swappiness graceful despite setting it to low
1072 * values.
1073 *
1074 * Avoid div by zero with nr_inactive+1, and max resulting
1075 * value is vm_total_pages.
1076 */
1077 if (scan_global_lru(sc)) {
1078 imbalance = zone_page_state(zone, NR_ACTIVE);
1079 imbalance /= zone_page_state(zone, NR_INACTIVE) + 1;
1080 } else
1081 imbalance = mem_cgroup_reclaim_imbalance(sc->mem_cgroup);
1082
1083 /*
1084 * Reduce the effect of imbalance if swappiness is low,
1085 * this means for a swappiness very low, the imbalance
1086 * must be much higher than 100 for this logic to make
1087 * the difference.
1088 *
1089 * Max temporary value is vm_total_pages*100.
1090 */
1091 imbalance *= (vm_swappiness + 1);
1092 imbalance /= 100;
1093
1094 /*
1095 * If not much of the ram is mapped, makes the imbalance
1096 * less relevant, it's high priority we refill the inactive
1097 * list with mapped pages only in presence of high ratio of
1098 * mapped pages.
1099 *
1100 * Max temporary value is vm_total_pages*100.
1101 */
1102 imbalance *= mapped_ratio;
1103 imbalance /= 100;
1104
1105 /* apply imbalance feedback to swap_tendency */
1106 swap_tendency += imbalance;
1107
1108 /*
1109 * Now use this metric to decide whether to start moving mapped
1110 * memory onto the inactive list.
1111 */
1112 if (swap_tendency >= 100)
1113 reclaim_mapped = 1;
1114
1115 return reclaim_mapped;
1116}
1117
1118/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001119 * This moves pages from the active list to the inactive list.
1120 *
1121 * We move them the other way if the page is referenced by one or more
1122 * processes, from rmap.
1123 *
1124 * If the pages are mostly unmapped, the processing is fast and it is
1125 * appropriate to hold zone->lru_lock across the whole operation. But if
1126 * the pages are mapped, the processing is slow (page_referenced()) so we
1127 * should drop zone->lru_lock around each page. It's impossible to balance
1128 * this, so instead we remove the pages from the LRU while processing them.
1129 * It is safe to rely on PG_active against the non-LRU pages in here because
1130 * nobody will play with that bit on a non-LRU page.
1131 *
1132 * The downside is that we have to touch page->_count against each page.
1133 * But we had to alter page->flags anyway.
1134 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001135
1136
Andrew Morton1742f192006-03-22 00:08:21 -08001137static void shrink_active_list(unsigned long nr_pages, struct zone *zone,
Martin Blighbbdb3962006-10-28 10:38:25 -07001138 struct scan_control *sc, int priority)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001139{
Andrew Morton69e05942006-03-22 00:08:19 -08001140 unsigned long pgmoved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001141 int pgdeactivate = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001142 unsigned long pgscanned;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001143 LIST_HEAD(l_hold); /* The pages which were snipped off */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001144 LIST_HEAD(l_active);
1145 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001146 struct page *page;
1147 struct pagevec pvec;
1148 int reclaim_mapped = 0;
Christoph Lameter2903fb12006-02-11 17:55:55 -08001149
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001150 if (sc->may_swap)
1151 reclaim_mapped = calc_reclaim_mapped(sc, zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001152
1153 lru_add_drain();
1154 spin_lock_irq(&zone->lru_lock);
Balbir Singh66e17072008-02-07 00:13:56 -08001155 pgmoved = sc->isolate_pages(nr_pages, &l_hold, &pgscanned, sc->order,
1156 ISOLATE_ACTIVE, zone,
1157 sc->mem_cgroup, 1);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001158 /*
1159 * zone->pages_scanned is used for detect zone's oom
1160 * mem_cgroup remembers nr_scan by itself.
1161 */
1162 if (scan_global_lru(sc))
1163 zone->pages_scanned += pgscanned;
1164
Christoph Lameterc8785382007-02-10 01:43:01 -08001165 __mod_zone_page_state(zone, NR_ACTIVE, -pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001166 spin_unlock_irq(&zone->lru_lock);
1167
Linus Torvalds1da177e2005-04-16 15:20:36 -07001168 while (!list_empty(&l_hold)) {
1169 cond_resched();
1170 page = lru_to_page(&l_hold);
1171 list_del(&page->lru);
1172 if (page_mapped(page)) {
1173 if (!reclaim_mapped ||
1174 (total_swap_pages == 0 && PageAnon(page)) ||
Balbir Singhbed71612008-02-07 00:14:01 -08001175 page_referenced(page, 0, sc->mem_cgroup)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001176 list_add(&page->lru, &l_active);
1177 continue;
1178 }
1179 }
1180 list_add(&page->lru, &l_inactive);
1181 }
1182
1183 pagevec_init(&pvec, 1);
1184 pgmoved = 0;
1185 spin_lock_irq(&zone->lru_lock);
1186 while (!list_empty(&l_inactive)) {
1187 page = lru_to_page(&l_inactive);
1188 prefetchw_prev_lru_page(page, &l_inactive, flags);
Nick Piggin725d7042006-09-25 23:30:55 -07001189 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001190 SetPageLRU(page);
Nick Piggin725d7042006-09-25 23:30:55 -07001191 VM_BUG_ON(!PageActive(page));
Nick Piggin4c84cac2006-03-22 00:08:00 -08001192 ClearPageActive(page);
1193
Christoph Lameterb69408e2008-10-18 20:26:14 -07001194 list_move(&page->lru, &zone->lru[LRU_INACTIVE].list);
Hugh Dickins427d5412008-03-04 14:29:03 -08001195 mem_cgroup_move_lists(page, false);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001196 pgmoved++;
1197 if (!pagevec_add(&pvec, page)) {
Christoph Lameterc8785382007-02-10 01:43:01 -08001198 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001199 spin_unlock_irq(&zone->lru_lock);
1200 pgdeactivate += pgmoved;
1201 pgmoved = 0;
1202 if (buffer_heads_over_limit)
1203 pagevec_strip(&pvec);
1204 __pagevec_release(&pvec);
1205 spin_lock_irq(&zone->lru_lock);
1206 }
1207 }
Christoph Lameterc8785382007-02-10 01:43:01 -08001208 __mod_zone_page_state(zone, NR_INACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001209 pgdeactivate += pgmoved;
1210 if (buffer_heads_over_limit) {
1211 spin_unlock_irq(&zone->lru_lock);
1212 pagevec_strip(&pvec);
1213 spin_lock_irq(&zone->lru_lock);
1214 }
1215
1216 pgmoved = 0;
1217 while (!list_empty(&l_active)) {
1218 page = lru_to_page(&l_active);
1219 prefetchw_prev_lru_page(page, &l_active, flags);
Nick Piggin725d7042006-09-25 23:30:55 -07001220 VM_BUG_ON(PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001221 SetPageLRU(page);
Nick Piggin725d7042006-09-25 23:30:55 -07001222 VM_BUG_ON(!PageActive(page));
Hugh Dickins427d5412008-03-04 14:29:03 -08001223
Christoph Lameterb69408e2008-10-18 20:26:14 -07001224 list_move(&page->lru, &zone->lru[LRU_ACTIVE].list);
Hugh Dickins427d5412008-03-04 14:29:03 -08001225 mem_cgroup_move_lists(page, true);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 pgmoved++;
1227 if (!pagevec_add(&pvec, page)) {
Christoph Lameterc8785382007-02-10 01:43:01 -08001228 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001229 pgmoved = 0;
1230 spin_unlock_irq(&zone->lru_lock);
1231 __pagevec_release(&pvec);
1232 spin_lock_irq(&zone->lru_lock);
1233 }
1234 }
Christoph Lameterc8785382007-02-10 01:43:01 -08001235 __mod_zone_page_state(zone, NR_ACTIVE, pgmoved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001236
Christoph Lameterf8891e52006-06-30 01:55:45 -07001237 __count_zone_vm_events(PGREFILL, zone, pgscanned);
1238 __count_vm_events(PGDEACTIVATE, pgdeactivate);
1239 spin_unlock_irq(&zone->lru_lock);
Nick Piggina74609f2006-01-06 00:11:20 -08001240
1241 pagevec_release(&pvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242}
1243
Christoph Lameterb69408e2008-10-18 20:26:14 -07001244static unsigned long shrink_list(enum lru_list l, unsigned long nr_to_scan,
1245 struct zone *zone, struct scan_control *sc, int priority)
1246{
1247 if (l == LRU_ACTIVE) {
1248 shrink_active_list(nr_to_scan, zone, sc, priority);
1249 return 0;
1250 }
1251 return shrink_inactive_list(nr_to_scan, zone, sc);
1252}
1253
Linus Torvalds1da177e2005-04-16 15:20:36 -07001254/*
1255 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1256 */
Andrew Morton05ff5132006-03-22 00:08:20 -08001257static unsigned long shrink_zone(int priority, struct zone *zone,
1258 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001259{
Christoph Lameterb69408e2008-10-18 20:26:14 -07001260 unsigned long nr[NR_LRU_LISTS];
Christoph Lameter86959492006-03-22 00:08:18 -08001261 unsigned long nr_to_scan;
Andrew Morton05ff5132006-03-22 00:08:20 -08001262 unsigned long nr_reclaimed = 0;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001263 enum lru_list l;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001264
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001265 if (scan_global_lru(sc)) {
1266 /*
1267 * Add one to nr_to_scan just to make sure that the kernel
1268 * will slowly sift through the active list.
1269 */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001270 for_each_lru(l) {
1271 zone->lru[l].nr_scan += (zone_page_state(zone,
1272 NR_LRU_BASE + l) >> priority) + 1;
1273 nr[l] = zone->lru[l].nr_scan;
1274 if (nr[l] >= sc->swap_cluster_max)
1275 zone->lru[l].nr_scan = 0;
1276 else
1277 nr[l] = 0;
1278 }
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001279 } else {
1280 /*
1281 * This reclaim occurs not because zone memory shortage but
1282 * because memory controller hits its limit.
1283 * Then, don't modify zone reclaim related data.
1284 */
Christoph Lameterb69408e2008-10-18 20:26:14 -07001285 nr[LRU_ACTIVE] = mem_cgroup_calc_reclaim(sc->mem_cgroup,
1286 zone, priority, LRU_ACTIVE);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001287
Christoph Lameterb69408e2008-10-18 20:26:14 -07001288 nr[LRU_INACTIVE] = mem_cgroup_calc_reclaim(sc->mem_cgroup,
1289 zone, priority, LRU_INACTIVE);
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001290 }
1291
Christoph Lameterb69408e2008-10-18 20:26:14 -07001292 while (nr[LRU_ACTIVE] || nr[LRU_INACTIVE]) {
1293 for_each_lru(l) {
1294 if (nr[l]) {
1295 nr_to_scan = min(nr[l],
Linus Torvalds1da177e2005-04-16 15:20:36 -07001296 (unsigned long)sc->swap_cluster_max);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001297 nr[l] -= nr_to_scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001298
Christoph Lameterb69408e2008-10-18 20:26:14 -07001299 nr_reclaimed += shrink_list(l, nr_to_scan,
1300 zone, sc, priority);
1301 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001302 }
1303 }
1304
Andrew Morton232ea4d2007-02-28 20:13:21 -08001305 throttle_vm_writeout(sc->gfp_mask);
Andrew Morton05ff5132006-03-22 00:08:20 -08001306 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001307}
1308
1309/*
1310 * This is the direct reclaim path, for page-allocating processes. We only
1311 * try to reclaim pages from zones which will satisfy the caller's allocation
1312 * request.
1313 *
1314 * We reclaim from a zone even if that zone is over pages_high. Because:
1315 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
1316 * allocation or
1317 * b) The zones may be over pages_high but they must go *over* pages_high to
1318 * satisfy the `incremental min' zone defense algorithm.
1319 *
1320 * Returns the number of reclaimed pages.
1321 *
1322 * If a zone is deemed to be full of pinned pages then just give it a light
1323 * scan then give up on it.
1324 */
Mel Gormandac1d272008-04-28 02:12:12 -07001325static unsigned long shrink_zones(int priority, struct zonelist *zonelist,
Andrew Morton05ff5132006-03-22 00:08:20 -08001326 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001327{
Mel Gorman54a6eb52008-04-28 02:12:16 -07001328 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Andrew Morton05ff5132006-03-22 00:08:20 -08001329 unsigned long nr_reclaimed = 0;
Mel Gormandd1a2392008-04-28 02:12:17 -07001330 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001331 struct zone *zone;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001332
Nick Piggin408d8542006-09-25 23:31:27 -07001333 sc->all_unreclaimable = 1;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001334 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08001335 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001336 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001337 /*
1338 * Take care memory controller reclaiming has small influence
1339 * to global LRU.
1340 */
1341 if (scan_global_lru(sc)) {
1342 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1343 continue;
1344 note_zone_scanning_priority(zone, priority);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001345
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001346 if (zone_is_all_unreclaimable(zone) &&
1347 priority != DEF_PRIORITY)
1348 continue; /* Let kswapd poll it */
1349 sc->all_unreclaimable = 0;
1350 } else {
1351 /*
1352 * Ignore cpuset limitation here. We just want to reduce
1353 * # of used pages by us regardless of memory shortage.
1354 */
1355 sc->all_unreclaimable = 0;
1356 mem_cgroup_note_reclaim_priority(sc->mem_cgroup,
1357 priority);
1358 }
Nick Piggin408d8542006-09-25 23:31:27 -07001359
Andrew Morton05ff5132006-03-22 00:08:20 -08001360 nr_reclaimed += shrink_zone(priority, zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001361 }
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001362
Andrew Morton05ff5132006-03-22 00:08:20 -08001363 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001364}
1365
1366/*
1367 * This is the main entry point to direct page reclaim.
1368 *
1369 * If a full scan of the inactive list fails to free enough memory then we
1370 * are "out of memory" and something needs to be killed.
1371 *
1372 * If the caller is !__GFP_FS then the probability of a failure is reasonably
1373 * high - the zone may be full of dirty or under-writeback pages, which this
1374 * caller can't do much about. We kick pdflush and take explicit naps in the
1375 * hope that some of these pages can be written. But if the allocating task
1376 * holds filesystem locks which prevent writeout this might not work, and the
1377 * allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001378 *
1379 * returns: 0, if no pages reclaimed
1380 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07001381 */
Mel Gormandac1d272008-04-28 02:12:12 -07001382static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Mel Gormandd1a2392008-04-28 02:12:17 -07001383 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001384{
1385 int priority;
kosaki.motohiro@jp.fujitsu.comc700be32008-06-12 15:21:27 -07001386 unsigned long ret = 0;
Andrew Morton69e05942006-03-22 00:08:19 -08001387 unsigned long total_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001388 unsigned long nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001389 struct reclaim_state *reclaim_state = current->reclaim_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001390 unsigned long lru_pages = 0;
Mel Gormandd1a2392008-04-28 02:12:17 -07001391 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07001392 struct zone *zone;
Mel Gormandd1a2392008-04-28 02:12:17 -07001393 enum zone_type high_zoneidx = gfp_zone(sc->gfp_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001394
Keika Kobayashi873b4772008-07-25 01:48:52 -07001395 delayacct_freepages_start();
1396
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001397 if (scan_global_lru(sc))
1398 count_vm_event(ALLOCSTALL);
1399 /*
1400 * mem_cgroup will not do shrink_slab.
1401 */
1402 if (scan_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001403 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001404
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001405 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1406 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001407
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001408 lru_pages += zone_page_state(zone, NR_ACTIVE)
1409 + zone_page_state(zone, NR_INACTIVE);
1410 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001411 }
1412
1413 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
Balbir Singh66e17072008-02-07 00:13:56 -08001414 sc->nr_scanned = 0;
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001415 if (!priority)
1416 disable_swap_token();
Mel Gormandac1d272008-04-28 02:12:12 -07001417 nr_reclaimed += shrink_zones(priority, zonelist, sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001418 /*
1419 * Don't shrink slabs when reclaiming memory from
1420 * over limit cgroups
1421 */
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001422 if (scan_global_lru(sc)) {
Mel Gormandd1a2392008-04-28 02:12:17 -07001423 shrink_slab(sc->nr_scanned, sc->gfp_mask, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001424 if (reclaim_state) {
1425 nr_reclaimed += reclaim_state->reclaimed_slab;
1426 reclaim_state->reclaimed_slab = 0;
1427 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001428 }
Balbir Singh66e17072008-02-07 00:13:56 -08001429 total_scanned += sc->nr_scanned;
1430 if (nr_reclaimed >= sc->swap_cluster_max) {
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001431 ret = nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001432 goto out;
1433 }
1434
1435 /*
1436 * Try to write back as many pages as we just scanned. This
1437 * tends to cause slow streaming writers to write data to the
1438 * disk smoothly, at the dirtying rate, which is nice. But
1439 * that's undesirable in laptop mode, where we *want* lumpy
1440 * writeout. So in laptop mode, write out the whole world.
1441 */
Balbir Singh66e17072008-02-07 00:13:56 -08001442 if (total_scanned > sc->swap_cluster_max +
1443 sc->swap_cluster_max / 2) {
Pekka J Enberg687a21c2005-06-28 20:44:55 -07001444 wakeup_pdflush(laptop_mode ? 0 : total_scanned);
Balbir Singh66e17072008-02-07 00:13:56 -08001445 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446 }
1447
1448 /* Take a nap, wait for some writeback to complete */
Andrew Morton4dd4b922008-03-24 12:29:52 -07001449 if (sc->nr_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001450 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001451 }
Fernando Luis Vazquez Cao87547ee2008-07-29 22:33:42 -07001452 /* top priority shrink_zones still had more to do? don't OOM, then */
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08001453 if (!sc->all_unreclaimable && scan_global_lru(sc))
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07001454 ret = nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001455out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001456 /*
1457 * Now that we've scanned all the zones at this priority level, note
1458 * that level within the zone so that the next thread which performs
1459 * scanning of this zone will immediately start out at this priority
1460 * level. This affects only the decision whether or not to bring
1461 * mapped pages onto the inactive list.
1462 */
1463 if (priority < 0)
1464 priority = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001466 if (scan_global_lru(sc)) {
Mel Gorman54a6eb52008-04-28 02:12:16 -07001467 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001468
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001469 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
1470 continue;
1471
1472 zone->prev_priority = priority;
1473 }
1474 } else
1475 mem_cgroup_record_reclaim_priority(sc->mem_cgroup, priority);
1476
Keika Kobayashi873b4772008-07-25 01:48:52 -07001477 delayacct_freepages_end();
1478
Linus Torvalds1da177e2005-04-16 15:20:36 -07001479 return ret;
1480}
1481
Mel Gormandac1d272008-04-28 02:12:12 -07001482unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
1483 gfp_t gfp_mask)
Balbir Singh66e17072008-02-07 00:13:56 -08001484{
1485 struct scan_control sc = {
1486 .gfp_mask = gfp_mask,
1487 .may_writepage = !laptop_mode,
1488 .swap_cluster_max = SWAP_CLUSTER_MAX,
1489 .may_swap = 1,
1490 .swappiness = vm_swappiness,
1491 .order = order,
1492 .mem_cgroup = NULL,
1493 .isolate_pages = isolate_pages_global,
1494 };
1495
Mel Gormandd1a2392008-04-28 02:12:17 -07001496 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001497}
1498
Balbir Singh00f0b822008-03-04 14:28:39 -08001499#ifdef CONFIG_CGROUP_MEM_RES_CTLR
Balbir Singh66e17072008-02-07 00:13:56 -08001500
Balbir Singhe1a1cd52008-02-07 00:14:02 -08001501unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *mem_cont,
1502 gfp_t gfp_mask)
Balbir Singh66e17072008-02-07 00:13:56 -08001503{
1504 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08001505 .may_writepage = !laptop_mode,
1506 .may_swap = 1,
1507 .swap_cluster_max = SWAP_CLUSTER_MAX,
1508 .swappiness = vm_swappiness,
1509 .order = 0,
1510 .mem_cgroup = mem_cont,
1511 .isolate_pages = mem_cgroup_isolate_pages,
1512 };
Mel Gormandac1d272008-04-28 02:12:12 -07001513 struct zonelist *zonelist;
Balbir Singh66e17072008-02-07 00:13:56 -08001514
Mel Gormandd1a2392008-04-28 02:12:17 -07001515 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
1516 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
1517 zonelist = NODE_DATA(numa_node_id())->node_zonelists;
1518 return do_try_to_free_pages(zonelist, &sc);
Balbir Singh66e17072008-02-07 00:13:56 -08001519}
1520#endif
1521
Linus Torvalds1da177e2005-04-16 15:20:36 -07001522/*
1523 * For kswapd, balance_pgdat() will work across all this node's zones until
1524 * they are all at pages_high.
1525 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001526 * Returns the number of pages which were actually freed.
1527 *
1528 * There is special handling here for zones which are full of pinned pages.
1529 * This can happen if the pages are all mlocked, or if they are all used by
1530 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
1531 * What we do is to detect the case where all pages in the zone have been
1532 * scanned twice and there has been zero successful reclaim. Mark the zone as
1533 * dead and from now on, only perform a short scan. Basically we're polling
1534 * the zone for when the problem goes away.
1535 *
1536 * kswapd scans the zones in the highmem->normal->dma direction. It skips
1537 * zones which have free_pages > pages_high, but once a zone is found to have
1538 * free_pages <= pages_high, we scan that zone and the lower zones regardless
1539 * of the number of free pages in the lower zones. This interoperates with
1540 * the page allocator fallback scheme to ensure that aging of pages is balanced
1541 * across the zones.
1542 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001543static unsigned long balance_pgdat(pg_data_t *pgdat, int order)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001544{
Linus Torvalds1da177e2005-04-16 15:20:36 -07001545 int all_zones_ok;
1546 int priority;
1547 int i;
Andrew Morton69e05942006-03-22 00:08:19 -08001548 unsigned long total_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001549 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001550 struct reclaim_state *reclaim_state = current->reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08001551 struct scan_control sc = {
1552 .gfp_mask = GFP_KERNEL,
1553 .may_swap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001554 .swap_cluster_max = SWAP_CLUSTER_MAX,
1555 .swappiness = vm_swappiness,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001556 .order = order,
Balbir Singh66e17072008-02-07 00:13:56 -08001557 .mem_cgroup = NULL,
1558 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08001559 };
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001560 /*
1561 * temp_priority is used to remember the scanning priority at which
1562 * this zone was successfully refilled to free_pages == pages_high.
1563 */
1564 int temp_priority[MAX_NR_ZONES];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001565
1566loop_again:
1567 total_scanned = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -08001568 nr_reclaimed = 0;
Christoph Lameterc0bbbc72006-06-11 15:22:26 -07001569 sc.may_writepage = !laptop_mode;
Christoph Lameterf8891e52006-06-30 01:55:45 -07001570 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001571
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001572 for (i = 0; i < pgdat->nr_zones; i++)
1573 temp_priority[i] = DEF_PRIORITY;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001574
1575 for (priority = DEF_PRIORITY; priority >= 0; priority--) {
1576 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
1577 unsigned long lru_pages = 0;
1578
Rik van Rielf7b7fd82005-11-28 13:44:07 -08001579 /* The swap token gets in the way of swapout... */
1580 if (!priority)
1581 disable_swap_token();
1582
Linus Torvalds1da177e2005-04-16 15:20:36 -07001583 all_zones_ok = 1;
1584
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001585 /*
1586 * Scan in the highmem->dma direction for the highest
1587 * zone which needs scanning
1588 */
1589 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
1590 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001591
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001592 if (!populated_zone(zone))
1593 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001594
David Rientjese815af92007-10-16 23:25:54 -07001595 if (zone_is_all_unreclaimable(zone) &&
1596 priority != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001597 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001598
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001599 if (!zone_watermark_ok(zone, order, zone->pages_high,
1600 0, 0)) {
1601 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001602 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001603 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001604 }
Andrew Mortone1dbeda2006-12-06 20:32:01 -08001605 if (i < 0)
1606 goto out;
1607
Linus Torvalds1da177e2005-04-16 15:20:36 -07001608 for (i = 0; i <= end_zone; i++) {
1609 struct zone *zone = pgdat->node_zones + i;
1610
Christoph Lameterc8785382007-02-10 01:43:01 -08001611 lru_pages += zone_page_state(zone, NR_ACTIVE)
1612 + zone_page_state(zone, NR_INACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001613 }
1614
1615 /*
1616 * Now scan the zone in the dma->highmem direction, stopping
1617 * at the last zone which needs scanning.
1618 *
1619 * We do this because the page allocator works in the opposite
1620 * direction. This prevents the page allocator from allocating
1621 * pages behind kswapd's direction of progress, which would
1622 * cause too much scanning of the lower zones.
1623 */
1624 for (i = 0; i <= end_zone; i++) {
1625 struct zone *zone = pgdat->node_zones + i;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001626 int nr_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001627
Con Kolivasf3fe6512006-01-06 00:11:15 -08001628 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001629 continue;
1630
David Rientjese815af92007-10-16 23:25:54 -07001631 if (zone_is_all_unreclaimable(zone) &&
1632 priority != DEF_PRIORITY)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001633 continue;
1634
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001635 if (!zone_watermark_ok(zone, order, zone->pages_high,
1636 end_zone, 0))
1637 all_zones_ok = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001638 temp_priority[i] = priority;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001639 sc.nr_scanned = 0;
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001640 note_zone_scanning_priority(zone, priority);
Rik van Riel32a43302007-10-16 01:24:50 -07001641 /*
1642 * We put equal pressure on every zone, unless one
1643 * zone has way too many pages free already.
1644 */
1645 if (!zone_watermark_ok(zone, order, 8*zone->pages_high,
1646 end_zone, 0))
1647 nr_reclaimed += shrink_zone(priority, zone, &sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001648 reclaim_state->reclaimed_slab = 0;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001649 nr_slab = shrink_slab(sc.nr_scanned, GFP_KERNEL,
1650 lru_pages);
Andrew Morton05ff5132006-03-22 00:08:20 -08001651 nr_reclaimed += reclaim_state->reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652 total_scanned += sc.nr_scanned;
David Rientjese815af92007-10-16 23:25:54 -07001653 if (zone_is_all_unreclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 continue;
akpm@osdl.orgb15e0902005-06-21 17:14:35 -07001655 if (nr_slab == 0 && zone->pages_scanned >=
Christoph Lameterc8785382007-02-10 01:43:01 -08001656 (zone_page_state(zone, NR_ACTIVE)
1657 + zone_page_state(zone, NR_INACTIVE)) * 6)
David Rientjese815af92007-10-16 23:25:54 -07001658 zone_set_flag(zone,
1659 ZONE_ALL_UNRECLAIMABLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001660 /*
1661 * If we've done a decent amount of scanning and
1662 * the reclaim ratio is low, start doing writepage
1663 * even in laptop mode
1664 */
1665 if (total_scanned > SWAP_CLUSTER_MAX * 2 &&
Andrew Morton05ff5132006-03-22 00:08:20 -08001666 total_scanned > nr_reclaimed + nr_reclaimed / 2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001667 sc.may_writepage = 1;
1668 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001669 if (all_zones_ok)
1670 break; /* kswapd: all done */
1671 /*
1672 * OK, kswapd is getting into trouble. Take a nap, then take
1673 * another pass across the zones.
1674 */
Andrew Morton4dd4b922008-03-24 12:29:52 -07001675 if (total_scanned && priority < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001676 congestion_wait(WRITE, HZ/10);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001677
1678 /*
1679 * We do this so kswapd doesn't build up large priorities for
1680 * example when it is freeing in parallel with allocators. It
1681 * matches the direct reclaim path behaviour in terms of impact
1682 * on zone->*_priority.
1683 */
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001684 if (nr_reclaimed >= SWAP_CLUSTER_MAX)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001685 break;
1686 }
1687out:
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001688 /*
1689 * Note within each zone the priority level at which this zone was
1690 * brought into a happy state. So that the next thread which scans this
1691 * zone will start out at that priority level.
1692 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001693 for (i = 0; i < pgdat->nr_zones; i++) {
1694 struct zone *zone = pgdat->node_zones + i;
1695
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001696 zone->prev_priority = temp_priority[i];
Linus Torvalds1da177e2005-04-16 15:20:36 -07001697 }
1698 if (!all_zones_ok) {
1699 cond_resched();
Rafael J. Wysocki83573762006-12-06 20:34:18 -08001700
1701 try_to_freeze();
1702
Linus Torvalds1da177e2005-04-16 15:20:36 -07001703 goto loop_again;
1704 }
1705
Andrew Morton05ff5132006-03-22 00:08:20 -08001706 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001707}
1708
1709/*
1710 * The background pageout daemon, started as a kernel thread
1711 * from the init process.
1712 *
1713 * This basically trickles out pages so that we have _some_
1714 * free memory available even if there is no other activity
1715 * that frees anything up. This is needed for things like routing
1716 * etc, where we otherwise might have all activity going on in
1717 * asynchronous contexts that cannot page things out.
1718 *
1719 * If there are applications that are active memory-allocators
1720 * (most normal use), this basically shouldn't matter.
1721 */
1722static int kswapd(void *p)
1723{
1724 unsigned long order;
1725 pg_data_t *pgdat = (pg_data_t*)p;
1726 struct task_struct *tsk = current;
1727 DEFINE_WAIT(wait);
1728 struct reclaim_state reclaim_state = {
1729 .reclaimed_slab = 0,
1730 };
Mike Travisc5f59f02008-04-04 18:11:10 -07001731 node_to_cpumask_ptr(cpumask, pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732
Mike Travisc5f59f02008-04-04 18:11:10 -07001733 if (!cpus_empty(*cpumask))
1734 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001735 current->reclaim_state = &reclaim_state;
1736
1737 /*
1738 * Tell the memory management that we're a "memory allocator",
1739 * and that if we need more memory we should get access to it
1740 * regardless (see "__alloc_pages()"). "kswapd" should
1741 * never get caught in the normal page freeing logic.
1742 *
1743 * (Kswapd normally doesn't need memory anyway, but sometimes
1744 * you need a small amount of memory in order to be able to
1745 * page out something else, and this flag essentially protects
1746 * us from recursively trying to free more memory as we're
1747 * trying to free the first piece of memory in the first place).
1748 */
Christoph Lameter930d9152006-01-08 01:00:47 -08001749 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07001750 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07001751
1752 order = 0;
1753 for ( ; ; ) {
1754 unsigned long new_order;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07001755
Linus Torvalds1da177e2005-04-16 15:20:36 -07001756 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
1757 new_order = pgdat->kswapd_max_order;
1758 pgdat->kswapd_max_order = 0;
1759 if (order < new_order) {
1760 /*
1761 * Don't sleep if someone wants a larger 'order'
1762 * allocation
1763 */
1764 order = new_order;
1765 } else {
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07001766 if (!freezing(current))
1767 schedule();
1768
Linus Torvalds1da177e2005-04-16 15:20:36 -07001769 order = pgdat->kswapd_max_order;
1770 }
1771 finish_wait(&pgdat->kswapd_wait, &wait);
1772
Rafael J. Wysockib1296cc2007-05-06 14:50:48 -07001773 if (!try_to_freeze()) {
1774 /* We can speed up thawing tasks if we don't call
1775 * balance_pgdat after returning from the refrigerator
1776 */
1777 balance_pgdat(pgdat, order);
1778 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001779 }
1780 return 0;
1781}
1782
1783/*
1784 * A zone is low on free memory, so wake its kswapd task to service it.
1785 */
1786void wakeup_kswapd(struct zone *zone, int order)
1787{
1788 pg_data_t *pgdat;
1789
Con Kolivasf3fe6512006-01-06 00:11:15 -08001790 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001791 return;
1792
1793 pgdat = zone->zone_pgdat;
Rohit Seth7fb1d9f2005-11-13 16:06:43 -08001794 if (zone_watermark_ok(zone, order, zone->pages_low, 0, 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001795 return;
1796 if (pgdat->kswapd_max_order < order)
1797 pgdat->kswapd_max_order = order;
Paul Jackson02a0e532006-12-13 00:34:25 -08001798 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001799 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001800 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001801 return;
Con Kolivas8d0986e2005-09-13 01:25:07 -07001802 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001803}
1804
1805#ifdef CONFIG_PM
1806/*
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001807 * Helper function for shrink_all_memory(). Tries to reclaim 'nr_pages' pages
1808 * from LRU lists system-wide, for given pass and priority, and returns the
1809 * number of reclaimed pages
1810 *
1811 * For pass > 3 we also try to shrink the LRU lists that contain a few pages
1812 */
Nigel Cunninghame07aa052006-12-22 01:07:21 -08001813static unsigned long shrink_all_zones(unsigned long nr_pages, int prio,
1814 int pass, struct scan_control *sc)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001815{
1816 struct zone *zone;
1817 unsigned long nr_to_scan, ret = 0;
Christoph Lameterb69408e2008-10-18 20:26:14 -07001818 enum lru_list l;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001819
1820 for_each_zone(zone) {
1821
1822 if (!populated_zone(zone))
1823 continue;
1824
David Rientjese815af92007-10-16 23:25:54 -07001825 if (zone_is_all_unreclaimable(zone) && prio != DEF_PRIORITY)
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001826 continue;
1827
Christoph Lameterb69408e2008-10-18 20:26:14 -07001828 for_each_lru(l) {
1829 /* For pass = 0 we don't shrink the active list */
1830 if (pass == 0 && l == LRU_ACTIVE)
1831 continue;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001832
Christoph Lameterb69408e2008-10-18 20:26:14 -07001833 zone->lru[l].nr_scan +=
1834 (zone_page_state(zone, NR_LRU_BASE + l)
1835 >> prio) + 1;
1836 if (zone->lru[l].nr_scan >= nr_pages || pass > 3) {
1837 zone->lru[l].nr_scan = 0;
1838 nr_to_scan = min(nr_pages,
1839 zone_page_state(zone,
1840 NR_LRU_BASE + l));
1841 ret += shrink_list(l, nr_to_scan, zone,
1842 sc, prio);
1843 if (ret >= nr_pages)
1844 return ret;
1845 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001846 }
1847 }
1848
1849 return ret;
1850}
1851
Andrew Morton76395d32007-01-05 16:37:05 -08001852static unsigned long count_lru_pages(void)
1853{
Christoph Lameterc8785382007-02-10 01:43:01 -08001854 return global_page_state(NR_ACTIVE) + global_page_state(NR_INACTIVE);
Andrew Morton76395d32007-01-05 16:37:05 -08001855}
1856
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001857/*
1858 * Try to free `nr_pages' of memory, system-wide, and return the number of
1859 * freed pages.
1860 *
1861 * Rather than trying to age LRUs the aim is to preserve the overall
1862 * LRU order by reclaiming preferentially
1863 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07001864 */
Andrew Morton69e05942006-03-22 00:08:19 -08001865unsigned long shrink_all_memory(unsigned long nr_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001866{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001867 unsigned long lru_pages, nr_slab;
Andrew Morton69e05942006-03-22 00:08:19 -08001868 unsigned long ret = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001869 int pass;
1870 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001871 struct scan_control sc = {
1872 .gfp_mask = GFP_KERNEL,
1873 .may_swap = 0,
1874 .swap_cluster_max = nr_pages,
1875 .may_writepage = 1,
1876 .swappiness = vm_swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08001877 .isolate_pages = isolate_pages_global,
Linus Torvalds1da177e2005-04-16 15:20:36 -07001878 };
1879
1880 current->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08001881
Andrew Morton76395d32007-01-05 16:37:05 -08001882 lru_pages = count_lru_pages();
Christoph Lameter972d1a72006-09-25 23:31:51 -07001883 nr_slab = global_page_state(NR_SLAB_RECLAIMABLE);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001884 /* If slab caches are huge, it's better to hit them first */
1885 while (nr_slab >= lru_pages) {
1886 reclaim_state.reclaimed_slab = 0;
1887 shrink_slab(nr_pages, sc.gfp_mask, lru_pages);
1888 if (!reclaim_state.reclaimed_slab)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001889 break;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001890
1891 ret += reclaim_state.reclaimed_slab;
1892 if (ret >= nr_pages)
1893 goto out;
1894
1895 nr_slab -= reclaim_state.reclaimed_slab;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001896 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001897
1898 /*
1899 * We try to shrink LRUs in 5 passes:
1900 * 0 = Reclaim from inactive_list only
1901 * 1 = Reclaim from active list but don't reclaim mapped
1902 * 2 = 2nd pass of type 1
1903 * 3 = Reclaim mapped (normal reclaim)
1904 * 4 = 2nd pass of type 3
1905 */
1906 for (pass = 0; pass < 5; pass++) {
1907 int prio;
1908
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001909 /* Force reclaiming mapped pages in the passes #3 and #4 */
1910 if (pass > 2) {
1911 sc.may_swap = 1;
1912 sc.swappiness = 100;
1913 }
1914
1915 for (prio = DEF_PRIORITY; prio >= 0; prio--) {
1916 unsigned long nr_to_scan = nr_pages - ret;
1917
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001918 sc.nr_scanned = 0;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001919 ret += shrink_all_zones(nr_to_scan, prio, pass, &sc);
1920 if (ret >= nr_pages)
1921 goto out;
1922
1923 reclaim_state.reclaimed_slab = 0;
Andrew Morton76395d32007-01-05 16:37:05 -08001924 shrink_slab(sc.nr_scanned, sc.gfp_mask,
1925 count_lru_pages());
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001926 ret += reclaim_state.reclaimed_slab;
1927 if (ret >= nr_pages)
1928 goto out;
1929
1930 if (sc.nr_scanned && prio < DEF_PRIORITY - 2)
Andrew Morton3fcfab12006-10-19 23:28:16 -07001931 congestion_wait(WRITE, HZ / 10);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001932 }
Rafael J. Wysocki248a0302006-03-22 00:09:04 -08001933 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001934
1935 /*
1936 * If ret = 0, we could not shrink LRUs, but there may be something
1937 * in slab caches
1938 */
Andrew Morton76395d32007-01-05 16:37:05 -08001939 if (!ret) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001940 do {
1941 reclaim_state.reclaimed_slab = 0;
Andrew Morton76395d32007-01-05 16:37:05 -08001942 shrink_slab(nr_pages, sc.gfp_mask, count_lru_pages());
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001943 ret += reclaim_state.reclaimed_slab;
1944 } while (ret < nr_pages && reclaim_state.reclaimed_slab > 0);
Andrew Morton76395d32007-01-05 16:37:05 -08001945 }
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001946
1947out:
Linus Torvalds1da177e2005-04-16 15:20:36 -07001948 current->reclaim_state = NULL;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07001949
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950 return ret;
1951}
1952#endif
1953
Linus Torvalds1da177e2005-04-16 15:20:36 -07001954/* It's optimal to keep kswapds on the same CPUs as their memory, but
1955 not required for correctness. So if the last cpu in a node goes
1956 away, we get changed to run anywhere: as the first one comes back,
1957 restore their cpu bindings. */
Chandra Seetharaman9c7b2162006-06-27 02:54:07 -07001958static int __devinit cpu_callback(struct notifier_block *nfb,
Andrew Morton69e05942006-03-22 00:08:19 -08001959 unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001960{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07001961 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001962
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07001963 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07001964 for_each_node_state(nid, N_HIGH_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07001965 pg_data_t *pgdat = NODE_DATA(nid);
1966 node_to_cpumask_ptr(mask, pgdat->node_id);
1967
1968 if (any_online_cpu(*mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001969 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07001970 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001971 }
1972 }
1973 return NOTIFY_OK;
1974}
Linus Torvalds1da177e2005-04-16 15:20:36 -07001975
Yasunori Goto3218ae12006-06-27 02:53:33 -07001976/*
1977 * This kswapd start function will be called by init and node-hot-add.
1978 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
1979 */
1980int kswapd_run(int nid)
1981{
1982 pg_data_t *pgdat = NODE_DATA(nid);
1983 int ret = 0;
1984
1985 if (pgdat->kswapd)
1986 return 0;
1987
1988 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
1989 if (IS_ERR(pgdat->kswapd)) {
1990 /* failure at boot is fatal */
1991 BUG_ON(system_state == SYSTEM_BOOTING);
1992 printk("Failed to start kswapd on node %d\n",nid);
1993 ret = -1;
1994 }
1995 return ret;
1996}
1997
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998static int __init kswapd_init(void)
1999{
Yasunori Goto3218ae12006-06-27 02:53:33 -07002000 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08002001
Linus Torvalds1da177e2005-04-16 15:20:36 -07002002 swap_setup();
Christoph Lameter9422ffb2007-10-16 01:25:31 -07002003 for_each_node_state(nid, N_HIGH_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07002004 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005 hotcpu_notifier(cpu_callback, 0);
2006 return 0;
2007}
2008
2009module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002010
2011#ifdef CONFIG_NUMA
2012/*
2013 * Zone reclaim mode
2014 *
2015 * If non-zero call zone_reclaim when the number of free pages falls below
2016 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08002017 */
2018int zone_reclaim_mode __read_mostly;
2019
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002020#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07002021#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08002022#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
2023#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
2024
Christoph Lameter9eeff232006-01-18 17:42:31 -08002025/*
Christoph Lametera92f7122006-02-01 03:05:32 -08002026 * Priority for ZONE_RECLAIM. This determines the fraction of pages
2027 * of a node considered for each zone_reclaim. 4 scans 1/16th of
2028 * a zone.
2029 */
2030#define ZONE_RECLAIM_PRIORITY 4
2031
Christoph Lameter9eeff232006-01-18 17:42:31 -08002032/*
Christoph Lameter96146342006-07-03 00:24:13 -07002033 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
2034 * occur.
2035 */
2036int sysctl_min_unmapped_ratio = 1;
2037
2038/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002039 * If the number of slab pages in a zone grows beyond this percentage then
2040 * slab reclaim needs to occur.
2041 */
2042int sysctl_min_slab_ratio = 5;
2043
2044/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08002045 * Try to free up some pages from this zone through reclaim.
2046 */
Andrew Morton179e9632006-03-22 00:08:18 -08002047static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08002048{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002049 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08002050 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002051 struct task_struct *p = current;
2052 struct reclaim_state reclaim_state;
Christoph Lameter86959492006-03-22 00:08:18 -08002053 int priority;
Andrew Morton05ff5132006-03-22 00:08:20 -08002054 unsigned long nr_reclaimed = 0;
Andrew Morton179e9632006-03-22 00:08:18 -08002055 struct scan_control sc = {
2056 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
2057 .may_swap = !!(zone_reclaim_mode & RECLAIM_SWAP),
Andrew Morton69e05942006-03-22 00:08:19 -08002058 .swap_cluster_max = max_t(unsigned long, nr_pages,
2059 SWAP_CLUSTER_MAX),
Andrew Morton179e9632006-03-22 00:08:18 -08002060 .gfp_mask = gfp_mask,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002061 .swappiness = vm_swappiness,
Balbir Singh66e17072008-02-07 00:13:56 -08002062 .isolate_pages = isolate_pages_global,
Andrew Morton179e9632006-03-22 00:08:18 -08002063 };
Christoph Lameter83e33a42006-09-25 23:31:53 -07002064 unsigned long slab_reclaimable;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002065
2066 disable_swap_token();
Christoph Lameter9eeff232006-01-18 17:42:31 -08002067 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08002068 /*
2069 * We need to be able to allocate from the reserves for RECLAIM_SWAP
2070 * and we also need to be able to write out pages for RECLAIM_WRITE
2071 * and RECLAIM_SWAP.
2072 */
2073 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002074 reclaim_state.reclaimed_slab = 0;
2075 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08002076
Christoph Lameter0ff38492006-09-25 23:31:52 -07002077 if (zone_page_state(zone, NR_FILE_PAGES) -
2078 zone_page_state(zone, NR_FILE_MAPPED) >
2079 zone->min_unmapped_pages) {
2080 /*
2081 * Free memory by calling shrink zone with increasing
2082 * priorities until we have enough memory freed.
2083 */
2084 priority = ZONE_RECLAIM_PRIORITY;
2085 do {
Martin Bligh3bb1a8522006-10-28 10:38:24 -07002086 note_zone_scanning_priority(zone, priority);
Christoph Lameter0ff38492006-09-25 23:31:52 -07002087 nr_reclaimed += shrink_zone(priority, zone, &sc);
2088 priority--;
2089 } while (priority >= 0 && nr_reclaimed < nr_pages);
2090 }
Christoph Lameterc84db232006-02-01 03:05:29 -08002091
Christoph Lameter83e33a42006-09-25 23:31:53 -07002092 slab_reclaimable = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
2093 if (slab_reclaimable > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002094 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08002095 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07002096 * many pages were freed in this zone. So we take the current
2097 * number of slab pages and shake the slab until it is reduced
2098 * by the same nr_pages that we used for reclaiming unmapped
2099 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002100 *
Christoph Lameter0ff38492006-09-25 23:31:52 -07002101 * Note that shrink_slab will free memory on all zones and may
2102 * take a long time.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002103 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07002104 while (shrink_slab(sc.nr_scanned, gfp_mask, order) &&
Christoph Lameter83e33a42006-09-25 23:31:53 -07002105 zone_page_state(zone, NR_SLAB_RECLAIMABLE) >
2106 slab_reclaimable - nr_pages)
Christoph Lameter0ff38492006-09-25 23:31:52 -07002107 ;
Christoph Lameter83e33a42006-09-25 23:31:53 -07002108
2109 /*
2110 * Update nr_reclaimed by the number of slab pages we
2111 * reclaimed from this zone.
2112 */
2113 nr_reclaimed += slab_reclaimable -
2114 zone_page_state(zone, NR_SLAB_RECLAIMABLE);
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08002115 }
2116
Christoph Lameter9eeff232006-01-18 17:42:31 -08002117 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08002118 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
Andrew Morton05ff5132006-03-22 00:08:20 -08002119 return nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08002120}
Andrew Morton179e9632006-03-22 00:08:18 -08002121
2122int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
2123{
Andrew Morton179e9632006-03-22 00:08:18 -08002124 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07002125 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002126
2127 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07002128 * Zone reclaim reclaims unmapped file backed pages and
2129 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07002130 *
Christoph Lameter96146342006-07-03 00:24:13 -07002131 * A small portion of unmapped file backed pages is needed for
2132 * file I/O otherwise pages read by file I/O will be immediately
2133 * thrown out if the zone is overallocated. So we do not reclaim
2134 * if less than a specified percentage of the zone is used by
2135 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08002136 */
Christoph Lameter34aa1332006-06-30 01:55:37 -07002137 if (zone_page_state(zone, NR_FILE_PAGES) -
Christoph Lameter0ff38492006-09-25 23:31:52 -07002138 zone_page_state(zone, NR_FILE_MAPPED) <= zone->min_unmapped_pages
2139 && zone_page_state(zone, NR_SLAB_RECLAIMABLE)
2140 <= zone->min_slab_pages)
Christoph Lameter96146342006-07-03 00:24:13 -07002141 return 0;
Andrew Morton179e9632006-03-22 00:08:18 -08002142
David Rientjesd773ed62007-10-16 23:26:01 -07002143 if (zone_is_all_unreclaimable(zone))
2144 return 0;
2145
Andrew Morton179e9632006-03-22 00:08:18 -08002146 /*
David Rientjesd773ed62007-10-16 23:26:01 -07002147 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08002148 */
David Rientjesd773ed62007-10-16 23:26:01 -07002149 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Andrew Morton179e9632006-03-22 00:08:18 -08002150 return 0;
2151
2152 /*
2153 * Only run zone reclaim on the local zone or on zones that do not
2154 * have associated processors. This will favor the local processor
2155 * over remote processors and spread off node memory allocations
2156 * as wide as possible.
2157 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07002158 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07002159 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Andrew Morton179e9632006-03-22 00:08:18 -08002160 return 0;
David Rientjesd773ed62007-10-16 23:26:01 -07002161
2162 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
2163 return 0;
2164 ret = __zone_reclaim(zone, gfp_mask, order);
2165 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
2166
2167 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08002168}
Christoph Lameter9eeff232006-01-18 17:42:31 -08002169#endif