blob: eea668d9cff6c578ada0cf6c02eca5e22de5598d [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
14#include <linux/mm.h>
15#include <linux/module.h>
Tejun Heo5a0e3ad2010-03-24 17:04:11 +090016#include <linux/gfp.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070017#include <linux/kernel_stat.h>
18#include <linux/swap.h>
19#include <linux/pagemap.h>
20#include <linux/init.h>
21#include <linux/highmem.h>
Anton Vorontsov70ddf632013-04-29 15:08:31 -070022#include <linux/vmpressure.h>
Andrew Mortone129b5c2006-09-27 01:50:00 -070023#include <linux/vmstat.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070024#include <linux/file.h>
25#include <linux/writeback.h>
26#include <linux/blkdev.h>
27#include <linux/buffer_head.h> /* for try_to_release_page(),
28 buffer_heads_over_limit */
29#include <linux/mm_inline.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070030#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>
Mel Gorman3e7d3442011-01-13 15:45:56 -080035#include <linux/compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070036#include <linux/notifier.h>
37#include <linux/rwsem.h>
Rafael J. Wysocki248a0302006-03-22 00:09:04 -080038#include <linux/delay.h>
Yasunori Goto3218ae12006-06-27 02:53:33 -070039#include <linux/kthread.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080040#include <linux/freezer.h>
Balbir Singh66e17072008-02-07 00:13:56 -080041#include <linux/memcontrol.h>
Keika Kobayashi873b4772008-07-25 01:48:52 -070042#include <linux/delayacct.h>
Lee Schermerhornaf936a12008-10-18 20:26:53 -070043#include <linux/sysctl.h>
KOSAKI Motohiro929bea72011-04-14 15:22:12 -070044#include <linux/oom.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070045#include <linux/prefetch.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070046
47#include <asm/tlbflush.h>
48#include <asm/div64.h>
49
50#include <linux/swapops.h>
Rafael Aquini117aad12013-09-30 13:45:16 -070051#include <linux/balloon_compaction.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070052
Nick Piggin0f8053a2006-03-22 00:08:33 -080053#include "internal.h"
54
Mel Gorman33906bc2010-08-09 17:19:16 -070055#define CREATE_TRACE_POINTS
56#include <trace/events/vmscan.h>
57
Linus Torvalds1da177e2005-04-16 15:20:36 -070058struct scan_control {
Linus Torvalds1da177e2005-04-16 15:20:36 -070059 /* Incremented by the number of inactive pages that were scanned */
60 unsigned long nr_scanned;
61
Rik van Riela79311c2009-01-06 14:40:01 -080062 /* Number of pages freed so far during a call to shrink_zones() */
63 unsigned long nr_reclaimed;
64
KOSAKI Motohiro22fba332009-12-14 17:59:10 -080065 /* How many pages shrink_list() should reclaim */
66 unsigned long nr_to_reclaim;
67
KOSAKI Motohiro7b517552009-12-14 17:59:12 -080068 unsigned long hibernation_mode;
69
Linus Torvalds1da177e2005-04-16 15:20:36 -070070 /* This context's GFP mask */
Al Viro6daa0e22005-10-21 03:18:50 -040071 gfp_t gfp_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070072
73 int may_writepage;
74
Johannes Weinera6dc60f82009-03-31 15:19:30 -070075 /* Can mapped pages be reclaimed? */
76 int may_unmap;
Christoph Lameterf1fd1062006-01-18 17:42:30 -080077
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -070078 /* Can pages be swapped as part of reclaim? */
79 int may_swap;
80
Andy Whitcroft5ad333e2007-07-17 04:03:16 -070081 int order;
Balbir Singh66e17072008-02-07 00:13:56 -080082
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -070083 /* Scan (total_size >> priority) pages at once */
84 int priority;
85
KOSAKI Motohiro5f53e762010-05-24 14:32:37 -070086 /*
Johannes Weinerf16015f2012-01-12 17:17:52 -080087 * The memory cgroup that hit its limit and as a result is the
88 * primary target of this reclaim invocation.
89 */
90 struct mem_cgroup *target_mem_cgroup;
Balbir Singh66e17072008-02-07 00:13:56 -080091
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -070092 /*
93 * Nodemask of nodes allowed by the caller. If NULL, all nodes
94 * are scanned.
95 */
96 nodemask_t *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -070097};
98
Linus Torvalds1da177e2005-04-16 15:20:36 -070099#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
100
101#ifdef ARCH_HAS_PREFETCH
102#define prefetch_prev_lru_page(_page, _base, _field) \
103 do { \
104 if ((_page)->lru.prev != _base) { \
105 struct page *prev; \
106 \
107 prev = lru_to_page(&(_page->lru)); \
108 prefetch(&prev->_field); \
109 } \
110 } while (0)
111#else
112#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
113#endif
114
115#ifdef ARCH_HAS_PREFETCHW
116#define prefetchw_prev_lru_page(_page, _base, _field) \
117 do { \
118 if ((_page)->lru.prev != _base) { \
119 struct page *prev; \
120 \
121 prev = lru_to_page(&(_page->lru)); \
122 prefetchw(&prev->_field); \
123 } \
124 } while (0)
125#else
126#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
127#endif
128
129/*
130 * From 0 .. 100. Higher means more swappy.
131 */
132int vm_swappiness = 60;
Zhang Yanfeib21e0b92013-02-22 16:35:48 -0800133unsigned long vm_total_pages; /* The total number of pages which the VM controls */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134
135static LIST_HEAD(shrinker_list);
136static DECLARE_RWSEM(shrinker_rwsem);
137
Andrew Mortonc255a452012-07-31 16:43:02 -0700138#ifdef CONFIG_MEMCG
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800139static bool global_reclaim(struct scan_control *sc)
140{
Johannes Weinerf16015f2012-01-12 17:17:52 -0800141 return !sc->target_mem_cgroup;
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800142}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800143#else
Johannes Weiner89b5fae2012-01-12 17:17:50 -0800144static bool global_reclaim(struct scan_control *sc)
145{
146 return true;
147}
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -0800148#endif
149
Lisa Du6e543d52013-09-11 14:22:36 -0700150unsigned long zone_reclaimable_pages(struct zone *zone)
151{
152 int nr;
153
154 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
155 zone_page_state(zone, NR_INACTIVE_FILE);
156
157 if (get_nr_swap_pages() > 0)
158 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
159 zone_page_state(zone, NR_INACTIVE_ANON);
160
161 return nr;
162}
163
164bool zone_reclaimable(struct zone *zone)
165{
166 return zone->pages_scanned < zone_reclaimable_pages(zone) * 6;
167}
168
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700169static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800170{
Hugh Dickinsc3c787e2012-05-29 15:06:52 -0700171 if (!mem_cgroup_disabled())
Hugh Dickins4d7dcca2012-05-29 15:07:08 -0700172 return mem_cgroup_get_lru_size(lruvec, lru);
KOSAKI Motohiroa3d8e052009-01-07 18:08:19 -0800173
Konstantin Khlebnikov074291f2012-05-29 15:07:00 -0700174 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
KOSAKI Motohiroc9f299d2009-01-07 18:08:16 -0800175}
176
Linus Torvalds1da177e2005-04-16 15:20:36 -0700177/*
Glauber Costa1d3d4432013-08-28 10:18:04 +1000178 * Add a shrinker callback to be called from the vm.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700179 */
Glauber Costa1d3d4432013-08-28 10:18:04 +1000180int register_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700181{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000182 size_t size = sizeof(*shrinker->nr_deferred);
183
184 /*
185 * If we only have one possible node in the system anyway, save
186 * ourselves the trouble and disable NUMA aware behavior. This way we
187 * will save memory and some small loop time later.
188 */
189 if (nr_node_ids == 1)
190 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
191
192 if (shrinker->flags & SHRINKER_NUMA_AWARE)
193 size *= nr_node_ids;
194
195 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
196 if (!shrinker->nr_deferred)
197 return -ENOMEM;
198
Rusty Russell8e1f9362007-07-17 04:03:17 -0700199 down_write(&shrinker_rwsem);
200 list_add_tail(&shrinker->list, &shrinker_list);
201 up_write(&shrinker_rwsem);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000202 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700203}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700204EXPORT_SYMBOL(register_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700205
206/*
207 * Remove one
208 */
Rusty Russell8e1f9362007-07-17 04:03:17 -0700209void unregister_shrinker(struct shrinker *shrinker)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700210{
211 down_write(&shrinker_rwsem);
212 list_del(&shrinker->list);
213 up_write(&shrinker_rwsem);
Andrew Vaginae393322013-10-16 13:46:46 -0700214 kfree(shrinker->nr_deferred);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700215}
Rusty Russell8e1f9362007-07-17 04:03:17 -0700216EXPORT_SYMBOL(unregister_shrinker);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700217
218#define SHRINK_BATCH 128
Glauber Costa1d3d4432013-08-28 10:18:04 +1000219
220static unsigned long
221shrink_slab_node(struct shrink_control *shrinkctl, struct shrinker *shrinker,
222 unsigned long nr_pages_scanned, unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700223{
Glauber Costa1d3d4432013-08-28 10:18:04 +1000224 unsigned long freed = 0;
225 unsigned long long delta;
226 long total_scan;
227 long max_pass;
228 long nr;
229 long new_nr;
230 int nid = shrinkctl->nid;
231 long batch_size = shrinker->batch ? shrinker->batch
232 : SHRINK_BATCH;
233
Dave Chinnera0b02132013-08-28 10:18:16 +1000234 max_pass = shrinker->count_objects(shrinker, shrinkctl);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000235 if (max_pass == 0)
236 return 0;
237
238 /*
239 * copy the current shrinker scan count into a local variable
240 * and zero it so that other concurrent shrinker invocations
241 * don't also do this scanning work.
242 */
243 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
244
245 total_scan = nr;
246 delta = (4 * nr_pages_scanned) / shrinker->seeks;
247 delta *= max_pass;
248 do_div(delta, lru_pages + 1);
249 total_scan += delta;
250 if (total_scan < 0) {
251 printk(KERN_ERR
252 "shrink_slab: %pF negative objects to delete nr=%ld\n",
Dave Chinnera0b02132013-08-28 10:18:16 +1000253 shrinker->scan_objects, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000254 total_scan = max_pass;
255 }
256
257 /*
258 * We need to avoid excessive windup on filesystem shrinkers
259 * due to large numbers of GFP_NOFS allocations causing the
260 * shrinkers to return -1 all the time. This results in a large
261 * nr being built up so when a shrink that can do some work
262 * comes along it empties the entire cache due to nr >>>
263 * max_pass. This is bad for sustaining a working set in
264 * memory.
265 *
266 * Hence only allow the shrinker to scan the entire cache when
267 * a large delta change is calculated directly.
268 */
269 if (delta < max_pass / 4)
270 total_scan = min(total_scan, max_pass / 2);
271
272 /*
273 * Avoid risking looping forever due to too large nr value:
274 * never try to free more than twice the estimate number of
275 * freeable entries.
276 */
277 if (total_scan > max_pass * 2)
278 total_scan = max_pass * 2;
279
280 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
281 nr_pages_scanned, lru_pages,
282 max_pass, delta, total_scan);
283
284 while (total_scan >= batch_size) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000285 unsigned long ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000286
Dave Chinnera0b02132013-08-28 10:18:16 +1000287 shrinkctl->nr_to_scan = batch_size;
288 ret = shrinker->scan_objects(shrinker, shrinkctl);
289 if (ret == SHRINK_STOP)
290 break;
291 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000292
293 count_vm_events(SLABS_SCANNED, batch_size);
294 total_scan -= batch_size;
295
296 cond_resched();
297 }
298
299 /*
300 * move the unused scan count back into the shrinker in a
301 * manner that handles concurrent updates. If we exhausted the
302 * scan, there is no need to do an update.
303 */
304 if (total_scan > 0)
305 new_nr = atomic_long_add_return(total_scan,
306 &shrinker->nr_deferred[nid]);
307 else
308 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
309
310 trace_mm_shrink_slab_end(shrinker, freed, nr, new_nr);
311 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700312}
313
Linus Torvalds1da177e2005-04-16 15:20:36 -0700314/*
315 * Call the shrink functions to age shrinkable caches
316 *
317 * Here we assume it costs one seek to replace a lru page and that it also
318 * takes a seek to recreate a cache object. With this in mind we age equal
319 * percentages of the lru and ageable caches. This should balance the seeks
320 * generated by these structures.
321 *
Simon Arlott183ff222007-10-20 01:27:18 +0200322 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700323 * slab to avoid swapping.
324 *
325 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
326 *
327 * `lru_pages' represents the number of on-LRU pages in all the zones which
328 * are eligible for the caller's allocation attempt. It is used for balancing
329 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700330 *
331 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700332 */
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000333unsigned long shrink_slab(struct shrink_control *shrinkctl,
Ying Han1495f232011-05-24 17:12:27 -0700334 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700335 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700336{
337 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000338 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700339
Ying Han1495f232011-05-24 17:12:27 -0700340 if (nr_pages_scanned == 0)
341 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700342
Minchan Kimf06590b2011-05-24 17:11:11 -0700343 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000344 /*
345 * If we would return 0, our callers would understand that we
346 * have nothing else to shrink and give up trying. By returning
347 * 1 we keep it going and assume we'll be able to shrink next
348 * time.
349 */
350 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700351 goto out;
352 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353
354 list_for_each_entry(shrinker, &shrinker_list, list) {
Glauber Costa1d3d4432013-08-28 10:18:04 +1000355 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
356 if (!node_online(shrinkctl->nid))
357 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700358
Glauber Costa1d3d4432013-08-28 10:18:04 +1000359 if (!(shrinker->flags & SHRINKER_NUMA_AWARE) &&
360 (shrinkctl->nid != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700361 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700362
Glauber Costa1d3d4432013-08-28 10:18:04 +1000363 freed += shrink_slab_node(shrinkctl, shrinker,
364 nr_pages_scanned, lru_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700365
Linus Torvalds1da177e2005-04-16 15:20:36 -0700366 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700367 }
368 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700369out:
370 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000371 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700372}
373
Linus Torvalds1da177e2005-04-16 15:20:36 -0700374static inline int is_page_cache_freeable(struct page *page)
375{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700376 /*
377 * A freeable page cache page is referenced only by the caller
378 * that isolated the page, the page cache radix tree and
379 * optional buffer heads at page->private.
380 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700381 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700382}
383
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700384static int may_write_to_queue(struct backing_dev_info *bdi,
385 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700386{
Christoph Lameter930d9152006-01-08 01:00:47 -0800387 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700388 return 1;
389 if (!bdi_write_congested(bdi))
390 return 1;
391 if (bdi == current->backing_dev_info)
392 return 1;
393 return 0;
394}
395
396/*
397 * We detected a synchronous write error writing a page out. Probably
398 * -ENOSPC. We need to propagate that into the address_space for a subsequent
399 * fsync(), msync() or close().
400 *
401 * The tricky part is that after writepage we cannot touch the mapping: nothing
402 * prevents it from being freed up. But we have a ref on the page and once
403 * that page is locked, the mapping is pinned.
404 *
405 * We're allowed to run sleeping lock_page() here because we know the caller has
406 * __GFP_FS.
407 */
408static void handle_write_error(struct address_space *mapping,
409 struct page *page, int error)
410{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100411 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700412 if (page_mapping(page) == mapping)
413 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700414 unlock_page(page);
415}
416
Christoph Lameter04e62a22006-06-23 02:03:38 -0700417/* possible outcome of pageout() */
418typedef enum {
419 /* failed to write page out, page is locked */
420 PAGE_KEEP,
421 /* move page to the active list, page is locked */
422 PAGE_ACTIVATE,
423 /* page has been sent to the disk successfully, page is unlocked */
424 PAGE_SUCCESS,
425 /* page is clean and locked */
426 PAGE_CLEAN,
427} pageout_t;
428
Linus Torvalds1da177e2005-04-16 15:20:36 -0700429/*
Andrew Morton1742f192006-03-22 00:08:21 -0800430 * pageout is called by shrink_page_list() for each dirty page.
431 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700433static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700434 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700435{
436 /*
437 * If the page is dirty, only perform writeback if that write
438 * will be non-blocking. To prevent this allocation from being
439 * stalled by pagecache activity. But note that there may be
440 * stalls if we need to run get_block(). We could test
441 * PagePrivate for that.
442 *
Vincent Li6aceb532009-12-14 17:58:49 -0800443 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700444 * this page's queue, we can perform writeback even if that
445 * will block.
446 *
447 * If the page is swapcache, write it back even if that would
448 * block, for some throttling. This happens by accident, because
449 * swap_backing_dev_info is bust: it doesn't reflect the
450 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700451 */
452 if (!is_page_cache_freeable(page))
453 return PAGE_KEEP;
454 if (!mapping) {
455 /*
456 * Some data journaling orphaned pages can have
457 * page->mapping == NULL while being dirty with clean buffers.
458 */
David Howells266cf652009-04-03 16:42:36 +0100459 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700460 if (try_to_free_buffers(page)) {
461 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700462 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700463 return PAGE_CLEAN;
464 }
465 }
466 return PAGE_KEEP;
467 }
468 if (mapping->a_ops->writepage == NULL)
469 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700470 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700471 return PAGE_KEEP;
472
473 if (clear_page_dirty_for_io(page)) {
474 int res;
475 struct writeback_control wbc = {
476 .sync_mode = WB_SYNC_NONE,
477 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700478 .range_start = 0,
479 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700480 .for_reclaim = 1,
481 };
482
483 SetPageReclaim(page);
484 res = mapping->a_ops->writepage(page, &wbc);
485 if (res < 0)
486 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800487 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700488 ClearPageReclaim(page);
489 return PAGE_ACTIVATE;
490 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700491
Linus Torvalds1da177e2005-04-16 15:20:36 -0700492 if (!PageWriteback(page)) {
493 /* synchronous write or broken a_ops? */
494 ClearPageReclaim(page);
495 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700496 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700497 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 return PAGE_SUCCESS;
499 }
500
501 return PAGE_CLEAN;
502}
503
Andrew Mortona649fd92006-10-17 00:09:36 -0700504/*
Nick Piggine2867812008-07-25 19:45:30 -0700505 * Same as remove_mapping, but if the page is removed from the mapping, it
506 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700507 */
Nick Piggine2867812008-07-25 19:45:30 -0700508static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800509{
Nick Piggin28e4d962006-09-25 23:31:23 -0700510 BUG_ON(!PageLocked(page));
511 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800512
Nick Piggin19fd6232008-07-25 19:45:32 -0700513 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800514 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700515 * The non racy check for a busy page.
516 *
517 * Must be careful with the order of the tests. When someone has
518 * a ref to the page, it may be possible that they dirty it then
519 * drop the reference. So if PageDirty is tested before page_count
520 * here, then the following race may occur:
521 *
522 * get_user_pages(&page);
523 * [user mapping goes away]
524 * write_to(page);
525 * !PageDirty(page) [good]
526 * SetPageDirty(page);
527 * put_page(page);
528 * !page_count(page) [good, discard it]
529 *
530 * [oops, our write_to data is lost]
531 *
532 * Reversing the order of the tests ensures such a situation cannot
533 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
534 * load is not satisfied before that of page->_count.
535 *
536 * Note that if SetPageDirty is always performed via set_page_dirty,
537 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800538 */
Nick Piggine2867812008-07-25 19:45:30 -0700539 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800540 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700541 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
542 if (unlikely(PageDirty(page))) {
543 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800544 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700545 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800546
547 if (PageSwapCache(page)) {
548 swp_entry_t swap = { .val = page_private(page) };
549 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700550 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700551 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700552 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500553 void (*freepage)(struct page *);
554
555 freepage = mapping->a_ops->freepage;
556
Minchan Kime64a7822011-03-22 16:32:44 -0700557 __delete_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700558 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700559 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500560
561 if (freepage != NULL)
562 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800563 }
564
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800565 return 1;
566
567cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700568 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800569 return 0;
570}
571
Linus Torvalds1da177e2005-04-16 15:20:36 -0700572/*
Nick Piggine2867812008-07-25 19:45:30 -0700573 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
574 * someone else has a ref on the page, abort and return 0. If it was
575 * successfully detached, return 1. Assumes the caller has a single ref on
576 * this page.
577 */
578int remove_mapping(struct address_space *mapping, struct page *page)
579{
580 if (__remove_mapping(mapping, page)) {
581 /*
582 * Unfreezing the refcount with 1 rather than 2 effectively
583 * drops the pagecache ref for us without requiring another
584 * atomic operation.
585 */
586 page_unfreeze_refs(page, 1);
587 return 1;
588 }
589 return 0;
590}
591
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700592/**
593 * putback_lru_page - put previously isolated page onto appropriate LRU list
594 * @page: page to be put back to appropriate lru list
595 *
596 * Add previously isolated @page to appropriate LRU list.
597 * Page may still be unevictable for other reasons.
598 *
599 * lru_lock must not be held, interrupts must be enabled.
600 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700601void putback_lru_page(struct page *page)
602{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700603 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700604 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700605
606 VM_BUG_ON(PageLRU(page));
607
608redo:
609 ClearPageUnevictable(page);
610
Hugh Dickins39b5f292012-10-08 16:33:18 -0700611 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700612 /*
613 * For evictable pages, we can use the cache.
614 * In event of a race, worst case is we end up with an
615 * unevictable page on [in]active list.
616 * We know how to handle that.
617 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700618 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700619 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700620 } else {
621 /*
622 * Put unevictable pages directly on zone's unevictable
623 * list.
624 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700625 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700626 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700627 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700628 * When racing with an mlock or AS_UNEVICTABLE clearing
629 * (page is unlocked) make sure that if the other thread
630 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800631 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700632 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700633 * the page back to the evictable list.
634 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700635 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700636 */
637 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700638 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700639
640 /*
641 * page's status can change while we move it among lru. If an evictable
642 * page is on unevictable list, it never be freed. To avoid that,
643 * check after we added it to the list, again.
644 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700645 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700646 if (!isolate_lru_page(page)) {
647 put_page(page);
648 goto redo;
649 }
650 /* This means someone else dropped this page from LRU
651 * So, it will be freed or putback to LRU again. There is
652 * nothing to do here.
653 */
654 }
655
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700656 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700657 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700658 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700659 count_vm_event(UNEVICTABLE_PGCULLED);
660
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700661 put_page(page); /* drop ref from isolate */
662}
663
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800664enum page_references {
665 PAGEREF_RECLAIM,
666 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800667 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800668 PAGEREF_ACTIVATE,
669};
670
671static enum page_references page_check_references(struct page *page,
672 struct scan_control *sc)
673{
Johannes Weiner645747462010-03-05 13:42:22 -0800674 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800675 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800676
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700677 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
678 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800679 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800680
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800681 /*
682 * Mlock lost the isolation race with us. Let try_to_unmap()
683 * move the page to the unevictable list.
684 */
685 if (vm_flags & VM_LOCKED)
686 return PAGEREF_RECLAIM;
687
Johannes Weiner645747462010-03-05 13:42:22 -0800688 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700689 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800690 return PAGEREF_ACTIVATE;
691 /*
692 * All mapped pages start out with page table
693 * references from the instantiating fault, so we need
694 * to look twice if a mapped file page is used more
695 * than once.
696 *
697 * Mark it and spare it for another trip around the
698 * inactive list. Another page table reference will
699 * lead to its activation.
700 *
701 * Note: the mark is set for activated pages as well
702 * so that recently deactivated but used pages are
703 * quickly recovered.
704 */
705 SetPageReferenced(page);
706
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800707 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800708 return PAGEREF_ACTIVATE;
709
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800710 /*
711 * Activate file-backed executable pages after first usage.
712 */
713 if (vm_flags & VM_EXEC)
714 return PAGEREF_ACTIVATE;
715
Johannes Weiner645747462010-03-05 13:42:22 -0800716 return PAGEREF_KEEP;
717 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800718
719 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700720 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800721 return PAGEREF_RECLAIM_CLEAN;
722
723 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800724}
725
Mel Gormane2be15f2013-07-03 15:01:57 -0700726/* Check if a page is dirty or under writeback */
727static void page_check_dirty_writeback(struct page *page,
728 bool *dirty, bool *writeback)
729{
Mel Gormanb4597222013-07-03 15:02:05 -0700730 struct address_space *mapping;
731
Mel Gormane2be15f2013-07-03 15:01:57 -0700732 /*
733 * Anonymous pages are not handled by flushers and must be written
734 * from reclaim context. Do not stall reclaim based on them
735 */
736 if (!page_is_file_cache(page)) {
737 *dirty = false;
738 *writeback = false;
739 return;
740 }
741
742 /* By default assume that the page flags are accurate */
743 *dirty = PageDirty(page);
744 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700745
746 /* Verify dirty/writeback state if the filesystem supports it */
747 if (!page_has_private(page))
748 return;
749
750 mapping = page_mapping(page);
751 if (mapping && mapping->a_ops->is_dirty_writeback)
752 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700753}
754
Nick Piggine2867812008-07-25 19:45:30 -0700755/*
Andrew Morton1742f192006-03-22 00:08:21 -0800756 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700757 */
Andrew Morton1742f192006-03-22 00:08:21 -0800758static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700759 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700760 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700761 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700762 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700763 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700764 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700765 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700766 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700767 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700768{
769 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700770 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700771 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700772 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700773 unsigned long nr_dirty = 0;
774 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800775 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700776 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700777 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700778
779 cond_resched();
780
Tim Chen69980e32012-07-31 16:46:08 -0700781 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700782 while (!list_empty(page_list)) {
783 struct address_space *mapping;
784 struct page *page;
785 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700786 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700787 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700788
789 cond_resched();
790
791 page = lru_to_page(page_list);
792 list_del(&page->lru);
793
Nick Piggin529ae9a2008-08-02 12:01:03 +0200794 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700795 goto keep;
796
Nick Piggin725d7042006-09-25 23:30:55 -0700797 VM_BUG_ON(PageActive(page));
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700798 VM_BUG_ON(page_zone(page) != zone);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700799
800 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800801
Hugh Dickins39b5f292012-10-08 16:33:18 -0700802 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700803 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700804
Johannes Weinera6dc60f82009-03-31 15:19:30 -0700805 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800806 goto keep_locked;
807
Linus Torvalds1da177e2005-04-16 15:20:36 -0700808 /* Double the slab pressure for mapped and swapcache pages */
809 if (page_mapped(page) || PageSwapCache(page))
810 sc->nr_scanned++;
811
Andy Whitcroftc661b072007-08-22 14:01:26 -0700812 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
813 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
814
Mel Gorman283aba92013-07-03 15:01:51 -0700815 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700816 * The number of dirty pages determines if a zone is marked
817 * reclaim_congested which affects wait_iff_congested. kswapd
818 * will stall and start writing pages if the tail of the LRU
819 * is all dirty unqueued pages.
820 */
821 page_check_dirty_writeback(page, &dirty, &writeback);
822 if (dirty || writeback)
823 nr_dirty++;
824
825 if (dirty && !writeback)
826 nr_unqueued_dirty++;
827
Mel Gormand04e8ac2013-07-03 15:02:03 -0700828 /*
829 * Treat this page as congested if the underlying BDI is or if
830 * pages are cycling through the LRU so quickly that the
831 * pages marked for immediate reclaim are making it to the
832 * end of the LRU a second time.
833 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700834 mapping = page_mapping(page);
Mel Gormand04e8ac2013-07-03 15:02:03 -0700835 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
836 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700837 nr_congested++;
838
839 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700840 * If a page at the tail of the LRU is under writeback, there
841 * are three cases to consider.
842 *
843 * 1) If reclaim is encountering an excessive number of pages
844 * under writeback and this page is both under writeback and
845 * PageReclaim then it indicates that pages are being queued
846 * for IO but are being recycled through the LRU before the
847 * IO can complete. Waiting on the page itself risks an
848 * indefinite stall if it is impossible to writeback the
849 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700850 * note that the LRU is being scanned too quickly and the
851 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700852 *
853 * 2) Global reclaim encounters a page, memcg encounters a
854 * page that is not marked for immediate reclaim or
855 * the caller does not have __GFP_IO. In this case mark
856 * the page for immediate reclaim and continue scanning.
857 *
858 * __GFP_IO is checked because a loop driver thread might
859 * enter reclaim, and deadlock if it waits on a page for
860 * which it is needed to do the write (loop masks off
861 * __GFP_IO|__GFP_FS for this reason); but more thought
862 * would probably show more reasons.
863 *
864 * Don't require __GFP_FS, since we're not going into the
865 * FS, just waiting on its writeback completion. Worryingly,
866 * ext4 gfs2 and xfs allocate pages with
867 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
868 * may_enter_fs here is liable to OOM on them.
869 *
870 * 3) memcg encounters a page that is not already marked
871 * PageReclaim. memcg does not have any dirty pages
872 * throttling so we could easily OOM just because too many
873 * pages are in writeback and there is nothing else to
874 * reclaim. Wait for the writeback to complete.
875 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700876 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700877 /* Case 1 above */
878 if (current_is_kswapd() &&
879 PageReclaim(page) &&
880 zone_is_reclaim_writeback(zone)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700881 nr_immediate++;
882 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700883
884 /* Case 2 above */
885 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700886 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
887 /*
888 * This is slightly racy - end_page_writeback()
889 * might have just cleared PageReclaim, then
890 * setting PageReclaim here end up interpreted
891 * as PageReadahead - but that does not matter
892 * enough to care. What we do want is for this
893 * page to have PageReclaim set next time memcg
894 * reclaim reaches the tests above, so it will
895 * then wait_on_page_writeback() to avoid OOM;
896 * and it's also appropriate in global reclaim.
897 */
898 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700899 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700900
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700901 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700902
903 /* Case 3 above */
904 } else {
905 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700906 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700907 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700908
Minchan Kim02c6de82012-10-08 16:31:55 -0700909 if (!force_reclaim)
910 references = page_check_references(page, sc);
911
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800912 switch (references) {
913 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700914 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800915 case PAGEREF_KEEP:
916 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800917 case PAGEREF_RECLAIM:
918 case PAGEREF_RECLAIM_CLEAN:
919 ; /* try to reclaim the page below */
920 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700921
Linus Torvalds1da177e2005-04-16 15:20:36 -0700922 /*
923 * Anonymous process memory has backing store?
924 * Try to allocate it some swap space here.
925 */
Nick Pigginb291f002008-10-18 20:26:44 -0700926 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800927 if (!(sc->gfp_mask & __GFP_IO))
928 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -0700929 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700930 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800931 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932
Mel Gormane2be15f2013-07-03 15:01:57 -0700933 /* Adding to swap updated mapping */
934 mapping = page_mapping(page);
935 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700936
937 /*
938 * The page is mapped into the page tables of one or more
939 * processes. Try to unmap it here.
940 */
941 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700942 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700943 case SWAP_FAIL:
944 goto activate_locked;
945 case SWAP_AGAIN:
946 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700947 case SWAP_MLOCK:
948 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700949 case SWAP_SUCCESS:
950 ; /* try to free the page below */
951 }
952 }
953
954 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -0700955 /*
956 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -0700957 * avoid risk of stack overflow but only writeback
958 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -0700959 */
Mel Gormanf84f6e22011-10-31 17:07:51 -0700960 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -0700961 (!current_is_kswapd() ||
Mel Gormand43006d2013-07-03 15:01:50 -0700962 !zone_is_reclaim_dirty(zone))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -0700963 /*
964 * Immediately reclaim when written back.
965 * Similar in principal to deactivate_page()
966 * except we already have the page isolated
967 * and know it's dirty
968 */
969 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
970 SetPageReclaim(page);
971
Mel Gormanee728862011-10-31 17:07:38 -0700972 goto keep_locked;
973 }
974
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800975 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700977 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700978 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800979 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700980 goto keep_locked;
981
982 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700983 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984 case PAGE_KEEP:
985 goto keep_locked;
986 case PAGE_ACTIVATE:
987 goto activate_locked;
988 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700989 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -0700990 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700991 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700992 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700993
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 /*
995 * A synchronous write - probably a ramdisk. Go
996 * ahead and try to reclaim the page.
997 */
Nick Piggin529ae9a2008-08-02 12:01:03 +0200998 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700999 goto keep;
1000 if (PageDirty(page) || PageWriteback(page))
1001 goto keep_locked;
1002 mapping = page_mapping(page);
1003 case PAGE_CLEAN:
1004 ; /* try to free the page below */
1005 }
1006 }
1007
1008 /*
1009 * If the page has buffers, try to free the buffer mappings
1010 * associated with this page. If we succeed we try to free
1011 * the page as well.
1012 *
1013 * We do this even if the page is PageDirty().
1014 * try_to_release_page() does not perform I/O, but it is
1015 * possible for a page to have PageDirty set, but it is actually
1016 * clean (all its buffers are clean). This happens if the
1017 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001018 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001019 * try_to_release_page() will discover that cleanness and will
1020 * drop the buffers and mark the page clean - it can be freed.
1021 *
1022 * Rarely, pages can have buffers and no ->mapping. These are
1023 * the pages which were not successfully invalidated in
1024 * truncate_complete_page(). We try to drop those buffers here
1025 * and if that worked, and the page is no longer mapped into
1026 * process address space (page_count == 1) it can be freed.
1027 * Otherwise, leave the page on the LRU so it is swappable.
1028 */
David Howells266cf652009-04-03 16:42:36 +01001029 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001030 if (!try_to_release_page(page, sc->gfp_mask))
1031 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001032 if (!mapping && page_count(page) == 1) {
1033 unlock_page(page);
1034 if (put_page_testzero(page))
1035 goto free_it;
1036 else {
1037 /*
1038 * rare race with speculative reference.
1039 * the speculative reference will free
1040 * this page shortly, so we may
1041 * increment nr_reclaimed here (and
1042 * leave it off the LRU).
1043 */
1044 nr_reclaimed++;
1045 continue;
1046 }
1047 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 }
1049
Nick Piggine2867812008-07-25 19:45:30 -07001050 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001051 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001052
Nick Piggina978d6f2008-10-18 20:26:58 -07001053 /*
1054 * At this point, we have no other references and there is
1055 * no way to pick any more up (removed from LRU, removed
1056 * from pagecache). Can use non-atomic bitops now (and
1057 * we obviously don't have to worry about waking up a process
1058 * waiting on the page lock, because there are no references.
1059 */
1060 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001061free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001062 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001063
1064 /*
1065 * Is there need to periodically free_page_list? It would
1066 * appear not as the counts should be low
1067 */
1068 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001069 continue;
1070
Nick Pigginb291f002008-10-18 20:26:44 -07001071cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001072 if (PageSwapCache(page))
1073 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001074 unlock_page(page);
1075 putback_lru_page(page);
1076 continue;
1077
Linus Torvalds1da177e2005-04-16 15:20:36 -07001078activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001079 /* Not a candidate for swapping, so reclaim swap space. */
1080 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001081 try_to_free_swap(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001082 VM_BUG_ON(PageActive(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001083 SetPageActive(page);
1084 pgactivate++;
1085keep_locked:
1086 unlock_page(page);
1087keep:
1088 list_add(&page->lru, &ret_pages);
Nick Pigginb291f002008-10-18 20:26:44 -07001089 VM_BUG_ON(PageLRU(page) || PageUnevictable(page));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001090 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001091
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -08001092 free_hot_cold_page_list(&free_pages, 1);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001093
Linus Torvalds1da177e2005-04-16 15:20:36 -07001094 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001095 count_vm_events(PGACTIVATE, pgactivate);
Tim Chen69980e32012-07-31 16:46:08 -07001096 mem_cgroup_uncharge_end();
Mel Gorman8e950282013-07-03 15:02:02 -07001097 *ret_nr_dirty += nr_dirty;
1098 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001099 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001100 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001101 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001102 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001103}
1104
Minchan Kim02c6de82012-10-08 16:31:55 -07001105unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1106 struct list_head *page_list)
1107{
1108 struct scan_control sc = {
1109 .gfp_mask = GFP_KERNEL,
1110 .priority = DEF_PRIORITY,
1111 .may_unmap = 1,
1112 };
Mel Gorman8e950282013-07-03 15:02:02 -07001113 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001114 struct page *page, *next;
1115 LIST_HEAD(clean_pages);
1116
1117 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001118 if (page_is_file_cache(page) && !PageDirty(page) &&
1119 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001120 ClearPageActive(page);
1121 list_move(&page->lru, &clean_pages);
1122 }
1123 }
1124
1125 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001126 TTU_UNMAP|TTU_IGNORE_ACCESS,
1127 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001128 list_splice(&clean_pages, page_list);
1129 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1130 return ret;
1131}
1132
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001133/*
1134 * Attempt to remove the specified page from its LRU. Only take this page
1135 * if it is of the appropriate PageActive status. Pages which are being
1136 * freed elsewhere are also ignored.
1137 *
1138 * page: page to consider
1139 * mode: one of the LRU isolation modes defined above
1140 *
1141 * returns 0 on success, -ve errno on failure.
1142 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001143int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001144{
1145 int ret = -EINVAL;
1146
1147 /* Only take pages on the LRU. */
1148 if (!PageLRU(page))
1149 return ret;
1150
Minchan Kime46a2872012-10-08 16:33:48 -07001151 /* Compaction should not handle unevictable pages but CMA can do so */
1152 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001153 return ret;
1154
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001155 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001156
Mel Gormanc8244932012-01-12 17:19:38 -08001157 /*
1158 * To minimise LRU disruption, the caller can indicate that it only
1159 * wants to isolate pages it will be able to operate on without
1160 * blocking - clean pages for the most part.
1161 *
1162 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1163 * is used by reclaim when it is cannot write to backing storage
1164 *
1165 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1166 * that it is possible to migrate without blocking
1167 */
1168 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1169 /* All the caller can do on PageWriteback is block */
1170 if (PageWriteback(page))
1171 return ret;
1172
1173 if (PageDirty(page)) {
1174 struct address_space *mapping;
1175
1176 /* ISOLATE_CLEAN means only clean pages */
1177 if (mode & ISOLATE_CLEAN)
1178 return ret;
1179
1180 /*
1181 * Only pages without mappings or that have a
1182 * ->migratepage callback are possible to migrate
1183 * without blocking
1184 */
1185 mapping = page_mapping(page);
1186 if (mapping && !mapping->a_ops->migratepage)
1187 return ret;
1188 }
1189 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001190
Minchan Kimf80c0672011-10-31 17:06:55 -07001191 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1192 return ret;
1193
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001194 if (likely(get_page_unless_zero(page))) {
1195 /*
1196 * Be careful not to clear PageLRU until after we're
1197 * sure the page is not being freed elsewhere -- the
1198 * page release code relies on it.
1199 */
1200 ClearPageLRU(page);
1201 ret = 0;
1202 }
1203
1204 return ret;
1205}
1206
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001207/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001208 * zone->lru_lock is heavily contended. Some of the functions that
1209 * shrink the lists perform better by taking out a batch of pages
1210 * and working on them outside the LRU lock.
1211 *
1212 * For pagecache intensive workloads, this function is the hottest
1213 * spot in the kernel (apart from copy_*_user functions).
1214 *
1215 * Appropriate locks must be held before calling this function.
1216 *
1217 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001218 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001219 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001220 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001221 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001222 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001223 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001224 *
1225 * returns how many pages were moved onto *@dst.
1226 */
Andrew Morton69e05942006-03-22 00:08:19 -08001227static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001228 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001229 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001230 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001231{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001232 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001233 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001234 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001235
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001236 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001237 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001238 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001239
Linus Torvalds1da177e2005-04-16 15:20:36 -07001240 page = lru_to_page(src);
1241 prefetchw_prev_lru_page(page, src, flags);
1242
Nick Piggin725d7042006-09-25 23:30:55 -07001243 VM_BUG_ON(!PageLRU(page));
Nick Piggin8d438f92006-03-22 00:07:59 -08001244
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001245 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001246 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001247 nr_pages = hpage_nr_pages(page);
1248 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001249 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001250 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001251 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001252
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001253 case -EBUSY:
1254 /* else it is being freed elsewhere */
1255 list_move(&page->lru, src);
1256 continue;
1257
1258 default:
1259 BUG();
1260 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001261 }
1262
Hugh Dickinsf6260122012-01-12 17:20:06 -08001263 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001264 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1265 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001266 return nr_taken;
1267}
1268
Nick Piggin62695a82008-10-18 20:26:09 -07001269/**
1270 * isolate_lru_page - tries to isolate a page from its LRU list
1271 * @page: page to isolate from its LRU list
1272 *
1273 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1274 * vmstat statistic corresponding to whatever LRU list the page was on.
1275 *
1276 * Returns 0 if the page was removed from an LRU list.
1277 * Returns -EBUSY if the page was not on an LRU list.
1278 *
1279 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001280 * the active list, it will have PageActive set. If it was found on
1281 * the unevictable list, it will have the PageUnevictable bit set. That flag
1282 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001283 *
1284 * The vmstat statistic corresponding to the list on which the page was
1285 * found will be decremented.
1286 *
1287 * Restrictions:
1288 * (1) Must be called with an elevated refcount on the page. This is a
1289 * fundamentnal difference from isolate_lru_pages (which is called
1290 * without a stable reference).
1291 * (2) the lru_lock must not be held.
1292 * (3) interrupts must be enabled.
1293 */
1294int isolate_lru_page(struct page *page)
1295{
1296 int ret = -EBUSY;
1297
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001298 VM_BUG_ON(!page_count(page));
1299
Nick Piggin62695a82008-10-18 20:26:09 -07001300 if (PageLRU(page)) {
1301 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001302 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001303
1304 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001305 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001306 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001307 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001308 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001309 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001310 del_page_from_lru_list(page, lruvec, lru);
1311 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001312 }
1313 spin_unlock_irq(&zone->lru_lock);
1314 }
1315 return ret;
1316}
1317
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001318/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001319 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1320 * then get resheduled. When there are massive number of tasks doing page
1321 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1322 * the LRU list will go small and be scanned faster than necessary, leading to
1323 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001324 */
1325static int too_many_isolated(struct zone *zone, int file,
1326 struct scan_control *sc)
1327{
1328 unsigned long inactive, isolated;
1329
1330 if (current_is_kswapd())
1331 return 0;
1332
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001333 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001334 return 0;
1335
1336 if (file) {
1337 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1338 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1339 } else {
1340 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1341 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1342 }
1343
Fengguang Wu3cf23842012-12-18 14:23:31 -08001344 /*
1345 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1346 * won't get blocked by normal direct-reclaimers, forming a circular
1347 * deadlock.
1348 */
1349 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1350 inactive >>= 3;
1351
Rik van Riel35cd7812009-09-21 17:01:38 -07001352 return isolated > inactive;
1353}
1354
Mel Gorman66635622010-08-09 17:19:30 -07001355static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001356putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001357{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001358 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1359 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001360 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001361
Mel Gorman66635622010-08-09 17:19:30 -07001362 /*
1363 * Put back any unfreeable pages.
1364 */
Mel Gorman66635622010-08-09 17:19:30 -07001365 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001366 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001367 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001368
Mel Gorman66635622010-08-09 17:19:30 -07001369 VM_BUG_ON(PageLRU(page));
1370 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001371 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001372 spin_unlock_irq(&zone->lru_lock);
1373 putback_lru_page(page);
1374 spin_lock_irq(&zone->lru_lock);
1375 continue;
1376 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001377
1378 lruvec = mem_cgroup_page_lruvec(page, zone);
1379
Linus Torvalds7a608572011-01-17 14:42:19 -08001380 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001381 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001382 add_page_to_lru_list(page, lruvec, lru);
1383
Mel Gorman66635622010-08-09 17:19:30 -07001384 if (is_active_lru(lru)) {
1385 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001386 int numpages = hpage_nr_pages(page);
1387 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001388 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001389 if (put_page_testzero(page)) {
1390 __ClearPageLRU(page);
1391 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001392 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001393
1394 if (unlikely(PageCompound(page))) {
1395 spin_unlock_irq(&zone->lru_lock);
1396 (*get_compound_page_dtor(page))(page);
1397 spin_lock_irq(&zone->lru_lock);
1398 } else
1399 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001400 }
1401 }
Mel Gorman66635622010-08-09 17:19:30 -07001402
Hugh Dickins3f797682012-01-12 17:20:07 -08001403 /*
1404 * To save our caller's stack, now use input list for pages to free.
1405 */
1406 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001407}
1408
1409/*
Andrew Morton1742f192006-03-22 00:08:21 -08001410 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1411 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001412 */
Mel Gorman66635622010-08-09 17:19:30 -07001413static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001414shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001415 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001416{
1417 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001418 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001419 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001420 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001421 unsigned long nr_dirty = 0;
1422 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001423 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001424 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001425 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001426 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001427 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001428 struct zone *zone = lruvec_zone(lruvec);
1429 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001430
Rik van Riel35cd7812009-09-21 17:01:38 -07001431 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001432 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001433
1434 /* We are about to die and free our memory. Return now. */
1435 if (fatal_signal_pending(current))
1436 return SWAP_CLUSTER_MAX;
1437 }
1438
Linus Torvalds1da177e2005-04-16 15:20:36 -07001439 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001440
1441 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001442 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001443 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001444 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001445
Linus Torvalds1da177e2005-04-16 15:20:36 -07001446 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001447
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001448 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1449 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001450
1451 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1452 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1453
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001454 if (global_reclaim(sc)) {
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001455 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001456 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001457 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001458 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001459 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001460 }
Hillf Dantond563c052012-03-21 16:34:02 -07001461 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001462
Hillf Dantond563c052012-03-21 16:34:02 -07001463 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001464 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001465
Minchan Kim02c6de82012-10-08 16:31:55 -07001466 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001467 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1468 &nr_writeback, &nr_immediate,
1469 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001470
Hugh Dickins3f797682012-01-12 17:20:07 -08001471 spin_lock_irq(&zone->lru_lock);
1472
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001473 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001474
Ying Han904249a2012-04-25 16:01:48 -07001475 if (global_reclaim(sc)) {
1476 if (current_is_kswapd())
1477 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1478 nr_reclaimed);
1479 else
1480 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1481 nr_reclaimed);
1482 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001483
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001484 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001485
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001486 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001487
1488 spin_unlock_irq(&zone->lru_lock);
1489
1490 free_hot_cold_page_list(&page_list, 1);
Mel Gormane11da5b2010-10-26 14:21:40 -07001491
Mel Gorman92df3a72011-10-31 17:07:56 -07001492 /*
1493 * If reclaim is isolating dirty pages under writeback, it implies
1494 * that the long-lived page allocation rate is exceeding the page
1495 * laundering rate. Either the global limits are not being effective
1496 * at throttling processes due to the page distribution throughout
1497 * zones or there is heavy usage of a slow backing device. The
1498 * only option is to throttle from reclaim context which is not ideal
1499 * as there is no guarantee the dirtying process is throttled in the
1500 * same way balance_dirty_pages() manages.
1501 *
Mel Gorman8e950282013-07-03 15:02:02 -07001502 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1503 * of pages under pages flagged for immediate reclaim and stall if any
1504 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001505 */
Mel Gorman918fc712013-07-08 16:00:25 -07001506 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman283aba92013-07-03 15:01:51 -07001507 zone_set_flag(zone, ZONE_WRITEBACK);
Mel Gorman92df3a72011-10-31 17:07:56 -07001508
Mel Gormand43006d2013-07-03 15:01:50 -07001509 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001510 * memcg will stall in page writeback so only consider forcibly
1511 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001512 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001513 if (global_reclaim(sc)) {
1514 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001515 * Tag a zone as congested if all the dirty pages scanned were
1516 * backed by a congested BDI and wait_iff_congested will stall.
1517 */
1518 if (nr_dirty && nr_dirty == nr_congested)
1519 zone_set_flag(zone, ZONE_CONGESTED);
1520
1521 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001522 * If dirty pages are scanned that are not queued for IO, it
1523 * implies that flushers are not keeping up. In this case, flag
1524 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1525 * pages from reclaim context. It will forcibly stall in the
1526 * next check.
1527 */
1528 if (nr_unqueued_dirty == nr_taken)
1529 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1530
1531 /*
1532 * In addition, if kswapd scans pages marked marked for
1533 * immediate reclaim and under writeback (nr_immediate), it
1534 * implies that pages are cycling through the LRU faster than
1535 * they are written so also forcibly stall.
1536 */
1537 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1538 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001539 }
Mel Gormand43006d2013-07-03 15:01:50 -07001540
Mel Gorman8e950282013-07-03 15:02:02 -07001541 /*
1542 * Stall direct reclaim for IO completions if underlying BDIs or zone
1543 * is congested. Allow kswapd to continue until it starts encountering
1544 * unqueued dirty pages or cycling through the LRU too quickly.
1545 */
1546 if (!sc->hibernation_mode && !current_is_kswapd())
1547 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1548
Mel Gormane11da5b2010-10-26 14:21:40 -07001549 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1550 zone_idx(zone),
1551 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001552 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001553 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001554 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001555}
1556
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001557/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001558 * This moves pages from the active list to the inactive list.
1559 *
1560 * We move them the other way if the page is referenced by one or more
1561 * processes, from rmap.
1562 *
1563 * If the pages are mostly unmapped, the processing is fast and it is
1564 * appropriate to hold zone->lru_lock across the whole operation. But if
1565 * the pages are mapped, the processing is slow (page_referenced()) so we
1566 * should drop zone->lru_lock around each page. It's impossible to balance
1567 * this, so instead we remove the pages from the LRU while processing them.
1568 * It is safe to rely on PG_active against the non-LRU pages in here because
1569 * nobody will play with that bit on a non-LRU page.
1570 *
1571 * The downside is that we have to touch page->_count against each page.
1572 * But we had to alter page->flags anyway.
1573 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001574
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001575static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001576 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001577 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001578 enum lru_list lru)
1579{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001580 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001581 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001582 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001583 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001584
Wu Fengguang3eb41402009-06-16 15:33:13 -07001585 while (!list_empty(list)) {
1586 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001587 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001588
1589 VM_BUG_ON(PageLRU(page));
1590 SetPageLRU(page);
1591
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001592 nr_pages = hpage_nr_pages(page);
1593 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001594 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001595 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001596
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001597 if (put_page_testzero(page)) {
1598 __ClearPageLRU(page);
1599 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001600 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001601
1602 if (unlikely(PageCompound(page))) {
1603 spin_unlock_irq(&zone->lru_lock);
1604 (*get_compound_page_dtor(page))(page);
1605 spin_lock_irq(&zone->lru_lock);
1606 } else
1607 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001608 }
1609 }
1610 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1611 if (!is_active_lru(lru))
1612 __count_vm_events(PGDEACTIVATE, pgmoved);
1613}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001614
Hugh Dickinsf6260122012-01-12 17:20:06 -08001615static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001616 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001617 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001618 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001619{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001620 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001621 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001622 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001623 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001624 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001625 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001626 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001627 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001628 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001629 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001630 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001631 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001632
1633 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001634
1635 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001636 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001637 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001638 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001639
Linus Torvalds1da177e2005-04-16 15:20:36 -07001640 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001641
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001642 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1643 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001644 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001645 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001646
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001647 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001648
Hugh Dickinsf6260122012-01-12 17:20:06 -08001649 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001650 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001651 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001652 spin_unlock_irq(&zone->lru_lock);
1653
Linus Torvalds1da177e2005-04-16 15:20:36 -07001654 while (!list_empty(&l_hold)) {
1655 cond_resched();
1656 page = lru_to_page(&l_hold);
1657 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001658
Hugh Dickins39b5f292012-10-08 16:33:18 -07001659 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001660 putback_lru_page(page);
1661 continue;
1662 }
1663
Mel Gormancc715d92012-03-21 16:34:00 -07001664 if (unlikely(buffer_heads_over_limit)) {
1665 if (page_has_private(page) && trylock_page(page)) {
1666 if (page_has_private(page))
1667 try_to_release_page(page, 0);
1668 unlock_page(page);
1669 }
1670 }
1671
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001672 if (page_referenced(page, 0, sc->target_mem_cgroup,
1673 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001674 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001675 /*
1676 * Identify referenced, file-backed active pages and
1677 * give them one more trip around the active list. So
1678 * that executable code get better chances to stay in
1679 * memory under moderate memory pressure. Anon pages
1680 * are not likely to be evicted by use-once streaming
1681 * IO, plus JVM can create lots of anon VM_EXEC pages,
1682 * so we ignore them here.
1683 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001684 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001685 list_add(&page->lru, &l_active);
1686 continue;
1687 }
1688 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001689
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001690 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001691 list_add(&page->lru, &l_inactive);
1692 }
1693
Andrew Mortonb5557492009-01-06 14:40:13 -08001694 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001695 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001696 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001697 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001698 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001699 * Count referenced pages from currently used mappings as rotated,
1700 * even though only some of them are actually re-activated. This
1701 * helps balance scan pressure between file and anonymous pages in
1702 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001703 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001704 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001705
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001706 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1707 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001708 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001709 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001710
1711 free_hot_cold_page_list(&l_hold, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001712}
1713
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001714#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001715static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001716{
1717 unsigned long active, inactive;
1718
1719 active = zone_page_state(zone, NR_ACTIVE_ANON);
1720 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1721
1722 if (inactive * zone->inactive_ratio < active)
1723 return 1;
1724
1725 return 0;
1726}
1727
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001728/**
1729 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001730 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001731 *
1732 * Returns true if the zone does not have enough inactive anon pages,
1733 * meaning some active anon pages need to be deactivated.
1734 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001735static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001736{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001737 /*
1738 * If we don't have swap space, anonymous page deactivation
1739 * is pointless.
1740 */
1741 if (!total_swap_pages)
1742 return 0;
1743
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001744 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001745 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001746
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001747 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001748}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001749#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001750static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001751{
1752 return 0;
1753}
1754#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001755
Rik van Riel56e49d22009-06-16 15:32:28 -07001756/**
1757 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001758 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001759 *
1760 * When the system is doing streaming IO, memory pressure here
1761 * ensures that active file pages get deactivated, until more
1762 * than half of the file pages are on the inactive list.
1763 *
1764 * Once we get to that situation, protect the system's working
1765 * set from being evicted by disabling active file page aging.
1766 *
1767 * This uses a different ratio than the anonymous pages, because
1768 * the page cache uses a use-once replacement algorithm.
1769 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001770static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001771{
Johannes Weinere3790142013-02-22 16:35:19 -08001772 unsigned long inactive;
1773 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001774
Johannes Weinere3790142013-02-22 16:35:19 -08001775 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1776 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1777
1778 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001779}
1780
Hugh Dickins75b00af2012-05-29 15:07:09 -07001781static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001782{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001783 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001784 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001785 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001786 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001787}
1788
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001789static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001790 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001791{
Rik van Rielb39415b2009-12-14 17:59:48 -08001792 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001793 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001794 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001795 return 0;
1796 }
1797
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001798 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001799}
1800
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001801static int vmscan_swappiness(struct scan_control *sc)
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001802{
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001803 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001804 return vm_swappiness;
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001805 return mem_cgroup_swappiness(sc->target_mem_cgroup);
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001806}
1807
Johannes Weiner9a265112013-02-22 16:32:17 -08001808enum scan_balance {
1809 SCAN_EQUAL,
1810 SCAN_FRACT,
1811 SCAN_ANON,
1812 SCAN_FILE,
1813};
1814
Linus Torvalds1da177e2005-04-16 15:20:36 -07001815/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001816 * Determine how aggressively the anon and file LRU lists should be
1817 * scanned. The relative value of each set of LRU lists is determined
1818 * by looking at the fraction of the pages scanned we did rotate back
1819 * onto the active list instead of evict.
1820 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001821 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1822 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001823 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001824static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001825 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001826{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001827 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001828 u64 fraction[2];
1829 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001830 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001831 unsigned long anon_prio, file_prio;
1832 enum scan_balance scan_balance;
1833 unsigned long anon, file, free;
1834 bool force_scan = false;
1835 unsigned long ap, fp;
1836 enum lru_list lru;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001837
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001838 /*
1839 * If the zone or memcg is small, nr[l] can be 0. This
1840 * results in no scanning on this priority and a potential
1841 * priority drop. Global direct reclaim can go to the next
1842 * zone and tends to have no problems. Global kswapd is for
1843 * zone balancing and it needs to scan a minimum amount. When
1844 * reclaiming for a memcg, a priority drop can cause high
1845 * latencies, so it's better to scan a minimum amount there as
1846 * well.
1847 */
Lisa Du6e543d52013-09-11 14:22:36 -07001848 if (current_is_kswapd() && !zone_reclaimable(zone))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001849 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001850 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001851 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001852
1853 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001854 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001855 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001856 goto out;
1857 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001858
Johannes Weiner10316b32013-02-22 16:32:14 -08001859 /*
1860 * Global reclaim will swap to prevent OOM even with no
1861 * swappiness, but memcg users want to use this knob to
1862 * disable swapping for individual groups completely when
1863 * using the memory controller's swap limit feature would be
1864 * too expensive.
1865 */
1866 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001867 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001868 goto out;
1869 }
1870
1871 /*
1872 * Do not apply any pressure balancing cleverness when the
1873 * system is close to OOM, scan both anon and file equally
1874 * (unless the swappiness setting disagrees with swapping).
1875 */
1876 if (!sc->priority && vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001877 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001878 goto out;
1879 }
1880
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001881 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1882 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1883 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1884 get_lru_size(lruvec, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001885
Johannes Weiner11d16c22013-02-22 16:32:15 -08001886 /*
1887 * If it's foreseeable that reclaiming the file cache won't be
1888 * enough to get the zone back into a desirable shape, we have
1889 * to swap. Better start now and leave the - probably heavily
1890 * thrashing - remaining file pages alone.
1891 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001892 if (global_reclaim(sc)) {
Johannes Weiner11d16c22013-02-22 16:32:15 -08001893 free = zone_page_state(zone, NR_FREE_PAGES);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001894 if (unlikely(file + free <= high_wmark_pages(zone))) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001895 scan_balance = SCAN_ANON;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001896 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001897 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001898 }
1899
1900 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001901 * There is enough inactive page cache, do not reclaim
1902 * anything from the anonymous working set right now.
1903 */
1904 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001905 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001906 goto out;
1907 }
1908
Johannes Weiner9a265112013-02-22 16:32:17 -08001909 scan_balance = SCAN_FRACT;
1910
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001911 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001912 * With swappiness at 100, anonymous and file have the same priority.
1913 * This scanning priority is essentially the inverse of IO cost.
1914 */
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001915 anon_prio = vmscan_swappiness(sc);
Hugh Dickins75b00af2012-05-29 15:07:09 -07001916 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001917
1918 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001919 * OK, so we have swap space and a fair amount of page cache
1920 * pages. We use the recently rotated / recently scanned
1921 * ratios to determine how valuable each cache is.
1922 *
1923 * Because workloads change over time (and to avoid overflow)
1924 * we keep these statistics as a floating average, which ends
1925 * up weighing recent references more than old ones.
1926 *
1927 * anon in [0], file in [1]
1928 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001929 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001930 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001931 reclaim_stat->recent_scanned[0] /= 2;
1932 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001933 }
1934
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001935 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001936 reclaim_stat->recent_scanned[1] /= 2;
1937 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001938 }
1939
1940 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001941 * The amount of pressure on anon vs file pages is inversely
1942 * proportional to the fraction of recently scanned pages on
1943 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001944 */
Satoru Moriyafe350042012-05-29 15:06:47 -07001945 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001946 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001947
Satoru Moriyafe350042012-05-29 15:06:47 -07001948 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001949 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001950 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001951
Shaohua Li76a33fc2010-05-24 14:32:36 -07001952 fraction[0] = ap;
1953 fraction[1] = fp;
1954 denominator = ap + fp + 1;
1955out:
Hugh Dickins41113042012-01-12 17:20:01 -08001956 for_each_evictable_lru(lru) {
1957 int file = is_file_lru(lru);
Johannes Weinerd778df52013-02-22 16:32:12 -08001958 unsigned long size;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001959 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001960
Johannes Weinerd778df52013-02-22 16:32:12 -08001961 size = get_lru_size(lruvec, lru);
Johannes Weiner10316b32013-02-22 16:32:14 -08001962 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08001963
Johannes Weiner10316b32013-02-22 16:32:14 -08001964 if (!scan && force_scan)
1965 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08001966
1967 switch (scan_balance) {
1968 case SCAN_EQUAL:
1969 /* Scan lists relative to size */
1970 break;
1971 case SCAN_FRACT:
1972 /*
1973 * Scan types proportional to swappiness and
1974 * their relative recent reclaim efficiency.
1975 */
1976 scan = div64_u64(scan * fraction[file], denominator);
1977 break;
1978 case SCAN_FILE:
1979 case SCAN_ANON:
1980 /* Scan one type exclusively */
1981 if ((scan_balance == SCAN_FILE) != file)
1982 scan = 0;
1983 break;
1984 default:
1985 /* Look ma, no brain */
1986 BUG();
1987 }
Hugh Dickins41113042012-01-12 17:20:01 -08001988 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001989 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07001990}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001991
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001992/*
1993 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
1994 */
1995static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
1996{
1997 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07001998 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08001999 unsigned long nr_to_scan;
2000 enum lru_list lru;
2001 unsigned long nr_reclaimed = 0;
2002 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2003 struct blk_plug plug;
Mel Gormane82e0562013-07-03 15:01:44 -07002004 bool scan_adjusted = false;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002005
2006 get_scan_count(lruvec, sc, nr);
2007
Mel Gormane82e0562013-07-03 15:01:44 -07002008 /* Record the original scan target for proportional adjustments later */
2009 memcpy(targets, nr, sizeof(nr));
2010
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002011 blk_start_plug(&plug);
2012 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2013 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002014 unsigned long nr_anon, nr_file, percentage;
2015 unsigned long nr_scanned;
2016
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002017 for_each_evictable_lru(lru) {
2018 if (nr[lru]) {
2019 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2020 nr[lru] -= nr_to_scan;
2021
2022 nr_reclaimed += shrink_list(lru, nr_to_scan,
2023 lruvec, sc);
2024 }
2025 }
Mel Gormane82e0562013-07-03 15:01:44 -07002026
2027 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2028 continue;
2029
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002030 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002031 * For global direct reclaim, reclaim only the number of pages
2032 * requested. Less care is taken to scan proportionally as it
2033 * is more important to minimise direct reclaim stall latency
2034 * than it is to properly age the LRU lists.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002035 */
Mel Gormane82e0562013-07-03 15:01:44 -07002036 if (global_reclaim(sc) && !current_is_kswapd())
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002037 break;
Mel Gormane82e0562013-07-03 15:01:44 -07002038
2039 /*
2040 * For kswapd and memcg, reclaim at least the number of pages
2041 * requested. Ensure that the anon and file LRUs shrink
2042 * proportionally what was requested by get_scan_count(). We
2043 * stop reclaiming one LRU and reduce the amount scanning
2044 * proportional to the original scan target.
2045 */
2046 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2047 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2048
2049 if (nr_file > nr_anon) {
2050 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2051 targets[LRU_ACTIVE_ANON] + 1;
2052 lru = LRU_BASE;
2053 percentage = nr_anon * 100 / scan_target;
2054 } else {
2055 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2056 targets[LRU_ACTIVE_FILE] + 1;
2057 lru = LRU_FILE;
2058 percentage = nr_file * 100 / scan_target;
2059 }
2060
2061 /* Stop scanning the smaller of the LRU */
2062 nr[lru] = 0;
2063 nr[lru + LRU_ACTIVE] = 0;
2064
2065 /*
2066 * Recalculate the other LRU scan count based on its original
2067 * scan target and the percentage scanning already complete
2068 */
2069 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2070 nr_scanned = targets[lru] - nr[lru];
2071 nr[lru] = targets[lru] * (100 - percentage) / 100;
2072 nr[lru] -= min(nr[lru], nr_scanned);
2073
2074 lru += LRU_ACTIVE;
2075 nr_scanned = targets[lru] - nr[lru];
2076 nr[lru] = targets[lru] * (100 - percentage) / 100;
2077 nr[lru] -= min(nr[lru], nr_scanned);
2078
2079 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002080 }
2081 blk_finish_plug(&plug);
2082 sc->nr_reclaimed += nr_reclaimed;
2083
2084 /*
2085 * Even if we did not try to evict anon pages at all, we want to
2086 * rebalance the anon lru active/inactive ratio.
2087 */
2088 if (inactive_anon_is_low(lruvec))
2089 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2090 sc, LRU_ACTIVE_ANON);
2091
2092 throttle_vm_writeout(sc->gfp_mask);
2093}
2094
Mel Gorman23b9da52012-05-29 15:06:20 -07002095/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002096static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002097{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002098 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002099 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002100 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002101 return true;
2102
2103 return false;
2104}
2105
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002106/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002107 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2108 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2109 * true if more pages should be reclaimed such that when the page allocator
2110 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2111 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002112 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002113static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002114 unsigned long nr_reclaimed,
2115 unsigned long nr_scanned,
2116 struct scan_control *sc)
2117{
2118 unsigned long pages_for_compaction;
2119 unsigned long inactive_lru_pages;
2120
2121 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002122 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002123 return false;
2124
Mel Gorman28765922011-02-25 14:44:20 -08002125 /* Consider stopping depending on scan and reclaim activity */
2126 if (sc->gfp_mask & __GFP_REPEAT) {
2127 /*
2128 * For __GFP_REPEAT allocations, stop reclaiming if the
2129 * full LRU list has been scanned and we are still failing
2130 * to reclaim pages. This full LRU scan is potentially
2131 * expensive but a __GFP_REPEAT caller really wants to succeed
2132 */
2133 if (!nr_reclaimed && !nr_scanned)
2134 return false;
2135 } else {
2136 /*
2137 * For non-__GFP_REPEAT allocations which can presumably
2138 * fail without consequence, stop if we failed to reclaim
2139 * any pages from the last SWAP_CLUSTER_MAX number of
2140 * pages that were scanned. This will return to the
2141 * caller faster at the risk reclaim/compaction and
2142 * the resulting allocation attempt fails
2143 */
2144 if (!nr_reclaimed)
2145 return false;
2146 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002147
2148 /*
2149 * If we have not reclaimed enough pages for compaction and the
2150 * inactive lists are large enough, continue reclaiming
2151 */
2152 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002153 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002154 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002155 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002156 if (sc->nr_reclaimed < pages_for_compaction &&
2157 inactive_lru_pages > pages_for_compaction)
2158 return true;
2159
2160 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002161 switch (compaction_suitable(zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002162 case COMPACT_PARTIAL:
2163 case COMPACT_CONTINUE:
2164 return false;
2165 default:
2166 return true;
2167 }
2168}
2169
Andrew Morton0608f432013-09-24 15:27:41 -07002170static void shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002171{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002172 unsigned long nr_reclaimed, nr_scanned;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002173
Johannes Weiner56600482012-01-12 17:17:59 -08002174 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002175 struct mem_cgroup *root = sc->target_mem_cgroup;
2176 struct mem_cgroup_reclaim_cookie reclaim = {
2177 .zone = zone,
2178 .priority = sc->priority,
2179 };
Andrew Morton694fbc02013-09-24 15:27:37 -07002180 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002181
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002182 nr_reclaimed = sc->nr_reclaimed;
2183 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002184
Andrew Morton694fbc02013-09-24 15:27:37 -07002185 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2186 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002187 struct lruvec *lruvec;
2188
2189 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2190
2191 shrink_lruvec(lruvec, sc);
2192
2193 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002194 * Direct reclaim and kswapd have to scan all memory
2195 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002196 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002197 *
2198 * Limit reclaim, on the other hand, only cares about
2199 * nr_to_reclaim pages to be reclaimed and it will
2200 * retry with decreasing priority if one round over the
2201 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002202 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002203 if (!global_reclaim(sc) &&
2204 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002205 mem_cgroup_iter_break(root, memcg);
2206 break;
2207 }
Andrew Morton694fbc02013-09-24 15:27:37 -07002208 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2209 } while (memcg);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002210
2211 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2212 sc->nr_scanned - nr_scanned,
2213 sc->nr_reclaimed - nr_reclaimed);
2214
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002215 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2216 sc->nr_scanned - nr_scanned, sc));
Johannes Weinerf16015f2012-01-12 17:17:52 -08002217}
2218
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002219/* Returns true if compaction should go ahead for a high-order request */
2220static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2221{
2222 unsigned long balance_gap, watermark;
2223 bool watermark_ok;
2224
2225 /* Do not consider compaction for orders reclaim is meant to satisfy */
2226 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2227 return false;
2228
2229 /*
2230 * Compaction takes time to run and there are potentially other
2231 * callers using the pages just freed. Continue reclaiming until
2232 * there is a buffer of free pages available to give compaction
2233 * a reasonable chance of completing and allocating the page
2234 */
2235 balance_gap = min(low_wmark_pages(zone),
Jiang Liub40da042013-02-22 16:33:52 -08002236 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002237 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2238 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2239 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2240
2241 /*
2242 * If compaction is deferred, reclaim up to a point where
2243 * compaction will have a chance of success when re-enabled
2244 */
Rik van Rielaff62242012-03-21 16:33:52 -07002245 if (compaction_deferred(zone, sc->order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002246 return watermark_ok;
2247
2248 /* If compaction is not ready to start, keep reclaiming */
2249 if (!compaction_suitable(zone, sc->order))
2250 return false;
2251
2252 return watermark_ok;
2253}
2254
Linus Torvalds1da177e2005-04-16 15:20:36 -07002255/*
2256 * This is the direct reclaim path, for page-allocating processes. We only
2257 * try to reclaim pages from zones which will satisfy the caller's allocation
2258 * request.
2259 *
Mel Gorman41858962009-06-16 15:32:12 -07002260 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2261 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2263 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002264 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2265 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2266 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002267 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002268 * If a zone is deemed to be full of pinned pages then just give it a light
2269 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07002270 *
2271 * This function returns true if a zone is being reclaimed for a costly
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002272 * high-order allocation and compaction is ready to begin. This indicates to
Mel Gorman0cee34f2012-01-12 17:19:49 -08002273 * the caller that it should consider retrying the allocation instead of
2274 * further reclaim.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002275 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002276static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277{
Mel Gormandd1a2392008-04-28 02:12:17 -07002278 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002279 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002280 unsigned long nr_soft_reclaimed;
2281 unsigned long nr_soft_scanned;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002282 bool aborted_reclaim = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002283
Mel Gormancc715d92012-03-21 16:34:00 -07002284 /*
2285 * If the number of buffer_heads in the machine exceeds the maximum
2286 * allowed level, force direct reclaim to scan the highmem zone as
2287 * highmem pages could be pinning lowmem pages storing buffer_heads
2288 */
2289 if (buffer_heads_over_limit)
2290 sc->gfp_mask |= __GFP_HIGHMEM;
2291
Mel Gormand4debc62010-08-09 17:19:29 -07002292 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2293 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002294 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002296 /*
2297 * Take care memory controller reclaiming has small influence
2298 * to global LRU.
2299 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002300 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002301 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2302 continue;
Lisa Du6e543d52013-09-11 14:22:36 -07002303 if (sc->priority != DEF_PRIORITY &&
2304 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002305 continue; /* Let kswapd poll it */
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002306 if (IS_ENABLED(CONFIG_COMPACTION)) {
Rik van Riele0887c12011-10-31 17:09:31 -07002307 /*
Mel Gormane0c23272011-10-31 17:09:33 -07002308 * If we already have plenty of memory free for
2309 * compaction in this zone, don't free any more.
2310 * Even though compaction is invoked for any
2311 * non-zero order, only frequent costly order
2312 * reclamation is disruptive enough to become a
Copot Alexandruc7cfa372012-03-21 16:34:10 -07002313 * noticeable problem, like transparent huge
2314 * page allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07002315 */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002316 if (compaction_ready(zone, sc)) {
Mel Gorman0cee34f2012-01-12 17:19:49 -08002317 aborted_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07002318 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07002319 }
Rik van Riele0887c12011-10-31 17:09:31 -07002320 }
Andrew Morton0608f432013-09-24 15:27:41 -07002321 /*
2322 * This steals pages from memory cgroups over softlimit
2323 * and returns the number of reclaimed pages and
2324 * scanned pages. This works for global memory pressure
2325 * and balancing, not for a memcg's limit.
2326 */
2327 nr_soft_scanned = 0;
2328 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2329 sc->order, sc->gfp_mask,
2330 &nr_soft_scanned);
2331 sc->nr_reclaimed += nr_soft_reclaimed;
2332 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002333 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002334 }
Nick Piggin408d8542006-09-25 23:31:27 -07002335
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002336 shrink_zone(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002337 }
Mel Gormane0c23272011-10-31 17:09:33 -07002338
Mel Gorman0cee34f2012-01-12 17:19:49 -08002339 return aborted_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002340}
2341
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002342/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002343static bool all_unreclaimable(struct zonelist *zonelist,
2344 struct scan_control *sc)
2345{
2346 struct zoneref *z;
2347 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002348
2349 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2350 gfp_zone(sc->gfp_mask), sc->nodemask) {
2351 if (!populated_zone(zone))
2352 continue;
2353 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2354 continue;
Lisa Du6e543d52013-09-11 14:22:36 -07002355 if (zone_reclaimable(zone))
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002356 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002357 }
2358
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002359 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002360}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002361
Linus Torvalds1da177e2005-04-16 15:20:36 -07002362/*
2363 * This is the main entry point to direct page reclaim.
2364 *
2365 * If a full scan of the inactive list fails to free enough memory then we
2366 * are "out of memory" and something needs to be killed.
2367 *
2368 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2369 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002370 * caller can't do much about. We kick the writeback threads and take explicit
2371 * naps in the hope that some of these pages can be written. But if the
2372 * allocating task holds filesystem locks which prevent writeout this might not
2373 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002374 *
2375 * returns: 0, if no pages reclaimed
2376 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002377 */
Mel Gormandac1d272008-04-28 02:12:12 -07002378static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002379 struct scan_control *sc,
2380 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002381{
Andrew Morton69e05942006-03-22 00:08:19 -08002382 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002383 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002384 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002385 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002386 unsigned long writeback_threshold;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002387 bool aborted_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002388
Keika Kobayashi873b4772008-07-25 01:48:52 -07002389 delayacct_freepages_start();
2390
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002391 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002392 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002393
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002394 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002395 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2396 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002397 sc->nr_scanned = 0;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002398 aborted_reclaim = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002399
Balbir Singh66e17072008-02-07 00:13:56 -08002400 /*
Mel Gorman5a1c9cb2013-07-08 16:00:24 -07002401 * Don't shrink slabs when reclaiming memory from over limit
2402 * cgroups but do shrink slab at least once when aborting
2403 * reclaim for compaction to avoid unevenly scanning file/anon
2404 * LRU pages over slab pages.
Balbir Singh66e17072008-02-07 00:13:56 -08002405 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002406 if (global_reclaim(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002407 unsigned long lru_pages = 0;
Dave Chinner0ce3d742013-08-28 10:18:03 +10002408
2409 nodes_clear(shrink->nodes_to_scan);
Mel Gormand4debc62010-08-09 17:19:29 -07002410 for_each_zone_zonelist(zone, z, zonelist,
2411 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002412 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2413 continue;
2414
2415 lru_pages += zone_reclaimable_pages(zone);
Dave Chinner0ce3d742013-08-28 10:18:03 +10002416 node_set(zone_to_nid(zone),
2417 shrink->nodes_to_scan);
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002418 }
2419
Ying Han1495f232011-05-24 17:12:27 -07002420 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002421 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002422 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002423 reclaim_state->reclaimed_slab = 0;
2424 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002425 }
Balbir Singh66e17072008-02-07 00:13:56 -08002426 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002427 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002428 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002429
2430 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002431 * If we're getting trouble reclaiming, start doing
2432 * writepage even in laptop mode.
2433 */
2434 if (sc->priority < DEF_PRIORITY - 2)
2435 sc->may_writepage = 1;
2436
2437 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438 * Try to write back as many pages as we just scanned. This
2439 * tends to cause slow streaming writers to write data to the
2440 * disk smoothly, at the dirtying rate, which is nice. But
2441 * that's undesirable in laptop mode, where we *want* lumpy
2442 * writeout. So in laptop mode, write out the whole world.
2443 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002444 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2445 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002446 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2447 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002448 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002449 }
Mel Gorman5a1c9cb2013-07-08 16:00:24 -07002450 } while (--sc->priority >= 0 && !aborted_reclaim);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002451
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002453 delayacct_freepages_end();
2454
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002455 if (sc->nr_reclaimed)
2456 return sc->nr_reclaimed;
2457
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002458 /*
2459 * As hibernation is going on, kswapd is freezed so that it can't mark
2460 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2461 * check.
2462 */
2463 if (oom_killer_disabled)
2464 return 0;
2465
Mel Gorman0cee34f2012-01-12 17:19:49 -08002466 /* Aborted reclaim to try compaction? don't OOM, then */
2467 if (aborted_reclaim)
Mel Gorman73350842012-01-12 17:19:33 -08002468 return 1;
2469
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002470 /* top priority shrink_zones still had more to do? don't OOM, then */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002471 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002472 return 1;
2473
2474 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002475}
2476
Mel Gorman55150612012-07-31 16:44:35 -07002477static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2478{
2479 struct zone *zone;
2480 unsigned long pfmemalloc_reserve = 0;
2481 unsigned long free_pages = 0;
2482 int i;
2483 bool wmark_ok;
2484
2485 for (i = 0; i <= ZONE_NORMAL; i++) {
2486 zone = &pgdat->node_zones[i];
2487 pfmemalloc_reserve += min_wmark_pages(zone);
2488 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2489 }
2490
2491 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2492
2493 /* kswapd must be awake if processes are being throttled */
2494 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2495 pgdat->classzone_idx = min(pgdat->classzone_idx,
2496 (enum zone_type)ZONE_NORMAL);
2497 wake_up_interruptible(&pgdat->kswapd_wait);
2498 }
2499
2500 return wmark_ok;
2501}
2502
2503/*
2504 * Throttle direct reclaimers if backing storage is backed by the network
2505 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2506 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002507 * when the low watermark is reached.
2508 *
2509 * Returns true if a fatal signal was delivered during throttling. If this
2510 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002511 */
Mel Gorman50694c22012-11-26 16:29:48 -08002512static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002513 nodemask_t *nodemask)
2514{
2515 struct zone *zone;
2516 int high_zoneidx = gfp_zone(gfp_mask);
2517 pg_data_t *pgdat;
2518
2519 /*
2520 * Kernel threads should not be throttled as they may be indirectly
2521 * responsible for cleaning pages necessary for reclaim to make forward
2522 * progress. kjournald for example may enter direct reclaim while
2523 * committing a transaction where throttling it could forcing other
2524 * processes to block on log_wait_commit().
2525 */
2526 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002527 goto out;
2528
2529 /*
2530 * If a fatal signal is pending, this process should not throttle.
2531 * It should return quickly so it can exit and free its memory
2532 */
2533 if (fatal_signal_pending(current))
2534 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002535
2536 /* Check if the pfmemalloc reserves are ok */
2537 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2538 pgdat = zone->zone_pgdat;
2539 if (pfmemalloc_watermark_ok(pgdat))
Mel Gorman50694c22012-11-26 16:29:48 -08002540 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002541
Mel Gorman68243e72012-07-31 16:44:39 -07002542 /* Account for the throttling */
2543 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2544
Mel Gorman55150612012-07-31 16:44:35 -07002545 /*
2546 * If the caller cannot enter the filesystem, it's possible that it
2547 * is due to the caller holding an FS lock or performing a journal
2548 * transaction in the case of a filesystem like ext[3|4]. In this case,
2549 * it is not safe to block on pfmemalloc_wait as kswapd could be
2550 * blocked waiting on the same lock. Instead, throttle for up to a
2551 * second before continuing.
2552 */
2553 if (!(gfp_mask & __GFP_FS)) {
2554 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2555 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002556
2557 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002558 }
2559
2560 /* Throttle until kswapd wakes the process */
2561 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2562 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002563
2564check_pending:
2565 if (fatal_signal_pending(current))
2566 return true;
2567
2568out:
2569 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002570}
2571
Mel Gormandac1d272008-04-28 02:12:12 -07002572unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002573 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002574{
Mel Gorman33906bc2010-08-09 17:19:16 -07002575 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002576 struct scan_control sc = {
Ming Lei21caf2f2013-02-22 16:34:08 -08002577 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Balbir Singh66e17072008-02-07 00:13:56 -08002578 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002579 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002580 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002581 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002582 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002583 .priority = DEF_PRIORITY,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002584 .target_mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002585 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002586 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002587 struct shrink_control shrink = {
2588 .gfp_mask = sc.gfp_mask,
2589 };
Balbir Singh66e17072008-02-07 00:13:56 -08002590
Mel Gorman55150612012-07-31 16:44:35 -07002591 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002592 * Do not enter reclaim if fatal signal was delivered while throttled.
2593 * 1 is returned so that the page allocator does not OOM kill at this
2594 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002595 */
Mel Gorman50694c22012-11-26 16:29:48 -08002596 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002597 return 1;
2598
Mel Gorman33906bc2010-08-09 17:19:16 -07002599 trace_mm_vmscan_direct_reclaim_begin(order,
2600 sc.may_writepage,
2601 gfp_mask);
2602
Ying Hana09ed5e2011-05-24 17:12:26 -07002603 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002604
2605 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2606
2607 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002608}
2609
Andrew Mortonc255a452012-07-31 16:43:02 -07002610#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002611
Johannes Weiner72835c82012-01-12 17:18:32 -08002612unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002613 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002614 struct zone *zone,
2615 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002616{
2617 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002618 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002619 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002620 .may_writepage = !laptop_mode,
2621 .may_unmap = 1,
2622 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002623 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002624 .priority = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002625 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002626 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002627 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002628
Balbir Singh4e416952009-09-23 15:56:39 -07002629 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2630 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002631
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002632 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002633 sc.may_writepage,
2634 sc.gfp_mask);
2635
Balbir Singh4e416952009-09-23 15:56:39 -07002636 /*
2637 * NOTE: Although we can get the priority field, using it
2638 * here is not a good idea, since it limits the pages we can scan.
2639 * if we don't reclaim here, the shrink_zone from balance_pgdat
2640 * will pick up pages from other mem cgroup's as well. We hack
2641 * the priority and make it zero.
2642 */
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002643 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002644
2645 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2646
Ying Han0ae5e892011-05-26 16:25:25 -07002647 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002648 return sc.nr_reclaimed;
2649}
2650
Johannes Weiner72835c82012-01-12 17:18:32 -08002651unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002652 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002653 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002654{
Balbir Singh4e416952009-09-23 15:56:39 -07002655 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002656 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002657 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002658 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002659 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002660 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002661 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002662 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002663 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002664 .priority = DEF_PRIORITY,
Johannes Weiner72835c82012-01-12 17:18:32 -08002665 .target_mem_cgroup = memcg,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002666 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002667 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2668 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2669 };
2670 struct shrink_control shrink = {
2671 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002672 };
Balbir Singh66e17072008-02-07 00:13:56 -08002673
Ying Han889976d2011-05-26 16:25:33 -07002674 /*
2675 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2676 * take care of from where we get pages. So the node where we start the
2677 * scan does not need to be the current node.
2678 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002679 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002680
2681 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002682
2683 trace_mm_vmscan_memcg_reclaim_begin(0,
2684 sc.may_writepage,
2685 sc.gfp_mask);
2686
Ying Hana09ed5e2011-05-24 17:12:26 -07002687 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002688
2689 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2690
2691 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002692}
2693#endif
2694
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002695static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002696{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002697 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002698
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002699 if (!total_swap_pages)
2700 return;
2701
2702 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2703 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002704 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002705
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002706 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002707 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002708 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002709
2710 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2711 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002712}
2713
Johannes Weiner60cefed2012-11-29 13:54:23 -08002714static bool zone_balanced(struct zone *zone, int order,
2715 unsigned long balance_gap, int classzone_idx)
2716{
2717 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2718 balance_gap, classzone_idx, 0))
2719 return false;
2720
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002721 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2722 !compaction_suitable(zone, order))
Johannes Weiner60cefed2012-11-29 13:54:23 -08002723 return false;
2724
2725 return true;
2726}
2727
Mel Gorman1741c872011-01-13 15:46:21 -08002728/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002729 * pgdat_balanced() is used when checking if a node is balanced.
2730 *
2731 * For order-0, all zones must be balanced!
2732 *
2733 * For high-order allocations only zones that meet watermarks and are in a
2734 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2735 * total of balanced pages must be at least 25% of the zones allowed by
2736 * classzone_idx for the node to be considered balanced. Forcing all zones to
2737 * be balanced for high orders can cause excessive reclaim when there are
2738 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002739 * The choice of 25% is due to
2740 * o a 16M DMA zone that is balanced will not balance a zone on any
2741 * reasonable sized machine
2742 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002743 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002744 * would need to be at least 256M for it to be balance a whole node.
2745 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2746 * to balance a node on its own. These seemed like reasonable ratios.
2747 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002748static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002749{
Jiang Liub40da042013-02-22 16:33:52 -08002750 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002751 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002752 int i;
2753
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002754 /* Check the watermark levels */
2755 for (i = 0; i <= classzone_idx; i++) {
2756 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002757
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002758 if (!populated_zone(zone))
2759 continue;
2760
Jiang Liub40da042013-02-22 16:33:52 -08002761 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002762
2763 /*
2764 * A special case here:
2765 *
2766 * balance_pgdat() skips over all_unreclaimable after
2767 * DEF_PRIORITY. Effectively, it considers them balanced so
2768 * they must be considered balanced here as well!
2769 */
Lisa Du6e543d52013-09-11 14:22:36 -07002770 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08002771 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002772 continue;
2773 }
2774
2775 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002776 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002777 else if (!order)
2778 return false;
2779 }
2780
2781 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002782 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002783 else
2784 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002785}
2786
Mel Gorman55150612012-07-31 16:44:35 -07002787/*
2788 * Prepare kswapd for sleeping. This verifies that there are no processes
2789 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2790 *
2791 * Returns true if kswapd is ready to sleep
2792 */
2793static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002794 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002795{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002796 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2797 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002798 return false;
2799
2800 /*
2801 * There is a potential race between when kswapd checks its watermarks
2802 * and a process gets throttled. There is also a potential race if
2803 * processes get throttled, kswapd wakes, a large process exits therby
2804 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2805 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2806 * so wake them now if necessary. If necessary, processes will wake
2807 * kswapd and get throttled again
2808 */
2809 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2810 wake_up(&pgdat->pfmemalloc_wait);
2811 return false;
2812 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002813
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002814 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002815}
2816
Linus Torvalds1da177e2005-04-16 15:20:36 -07002817/*
Mel Gorman75485362013-07-03 15:01:42 -07002818 * kswapd shrinks the zone by the number of pages required to reach
2819 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07002820 *
2821 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07002822 * reclaim or if the lack of progress was due to pages under writeback.
2823 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07002824 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07002825static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07002826 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07002827 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07002828 unsigned long lru_pages,
2829 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07002830{
Mel Gorman7c954f62013-07-03 15:01:54 -07002831 int testorder = sc->order;
2832 unsigned long balance_gap;
Mel Gorman75485362013-07-03 15:01:42 -07002833 struct reclaim_state *reclaim_state = current->reclaim_state;
2834 struct shrink_control shrink = {
2835 .gfp_mask = sc->gfp_mask,
2836 };
Mel Gorman7c954f62013-07-03 15:01:54 -07002837 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07002838
2839 /* Reclaim above the high watermark. */
2840 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07002841
2842 /*
2843 * Kswapd reclaims only single pages with compaction enabled. Trying
2844 * too hard to reclaim until contiguous free pages have become
2845 * available can hurt performance by evicting too much useful data
2846 * from memory. Do not reclaim more than needed for compaction.
2847 */
2848 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2849 compaction_suitable(zone, sc->order) !=
2850 COMPACT_SKIPPED)
2851 testorder = 0;
2852
2853 /*
2854 * We put equal pressure on every zone, unless one zone has way too
2855 * many pages free already. The "too many pages" is defined as the
2856 * high wmark plus a "gap" where the gap is either the low
2857 * watermark or 1% of the zone, whichever is smaller.
2858 */
2859 balance_gap = min(low_wmark_pages(zone),
2860 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2861 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2862
2863 /*
2864 * If there is no low memory pressure or the zone is balanced then no
2865 * reclaim is necessary
2866 */
2867 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2868 if (!lowmem_pressure && zone_balanced(zone, testorder,
2869 balance_gap, classzone_idx))
2870 return true;
2871
Mel Gorman75485362013-07-03 15:01:42 -07002872 shrink_zone(zone, sc);
Dave Chinner0ce3d742013-08-28 10:18:03 +10002873 nodes_clear(shrink.nodes_to_scan);
2874 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Mel Gorman75485362013-07-03 15:01:42 -07002875
2876 reclaim_state->reclaimed_slab = 0;
Lisa Du6e543d52013-09-11 14:22:36 -07002877 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Mel Gorman75485362013-07-03 15:01:42 -07002878 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2879
Mel Gorman2ab44f42013-07-03 15:01:47 -07002880 /* Account for the number of pages attempted to reclaim */
2881 *nr_attempted += sc->nr_to_reclaim;
2882
Mel Gorman283aba92013-07-03 15:01:51 -07002883 zone_clear_flag(zone, ZONE_WRITEBACK);
2884
Mel Gorman7c954f62013-07-03 15:01:54 -07002885 /*
2886 * If a zone reaches its high watermark, consider it to be no longer
2887 * congested. It's possible there are dirty pages backed by congested
2888 * BDIs but as pressure is relieved, speculatively avoid congestion
2889 * waits.
2890 */
Lisa Du6e543d52013-09-11 14:22:36 -07002891 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07002892 zone_balanced(zone, testorder, 0, classzone_idx)) {
2893 zone_clear_flag(zone, ZONE_CONGESTED);
2894 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2895 }
2896
Mel Gormanb8e83b92013-07-03 15:01:45 -07002897 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07002898}
2899
2900/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002901 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002902 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002903 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002904 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002905 *
2906 * There is special handling here for zones which are full of pinned pages.
2907 * This can happen if the pages are all mlocked, or if they are all used by
2908 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2909 * What we do is to detect the case where all pages in the zone have been
2910 * scanned twice and there has been zero successful reclaim. Mark the zone as
2911 * dead and from now on, only perform a short scan. Basically we're polling
2912 * the zone for when the problem goes away.
2913 *
2914 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002915 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2916 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2917 * lower zones regardless of the number of free pages in the lower zones. This
2918 * interoperates with the page allocator fallback scheme to ensure that aging
2919 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002920 */
Mel Gorman99504742011-01-13 15:46:20 -08002921static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002922 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002925 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07002926 unsigned long nr_soft_reclaimed;
2927 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002928 struct scan_control sc = {
2929 .gfp_mask = GFP_KERNEL,
Mel Gormanb8e83b92013-07-03 15:01:45 -07002930 .priority = DEF_PRIORITY,
Johannes Weinera6dc60f82009-03-31 15:19:30 -07002931 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002932 .may_swap = 1,
Mel Gormanb8e83b92013-07-03 15:01:45 -07002933 .may_writepage = !laptop_mode,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002934 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002935 .target_mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002936 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07002937 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002939 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002940 unsigned long lru_pages = 0;
Mel Gorman2ab44f42013-07-03 15:01:47 -07002941 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07002942 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07002943 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07002944
2945 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002947 /*
2948 * Scan in the highmem->dma direction for the highest
2949 * zone which needs scanning
2950 */
2951 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2952 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002953
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002954 if (!populated_zone(zone))
2955 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956
Lisa Du6e543d52013-09-11 14:22:36 -07002957 if (sc.priority != DEF_PRIORITY &&
2958 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002959 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002960
Rik van Riel556adec2008-10-18 20:26:34 -07002961 /*
2962 * Do some background aging of the anon list, to give
2963 * pages a chance to be referenced before reclaiming.
2964 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002965 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07002966
Mel Gormancc715d92012-03-21 16:34:00 -07002967 /*
2968 * If the number of buffer_heads in the machine
2969 * exceeds the maximum allowed level and this node
2970 * has a highmem zone, force kswapd to reclaim from
2971 * it to relieve lowmem pressure.
2972 */
2973 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2974 end_zone = i;
2975 break;
2976 }
2977
Johannes Weiner60cefed2012-11-29 13:54:23 -08002978 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002979 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002980 break;
Shaohua Li439423f2011-08-25 15:59:12 -07002981 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07002982 /*
2983 * If balanced, clear the dirty and congested
2984 * flags
2985 */
Shaohua Li439423f2011-08-25 15:59:12 -07002986 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormand43006d2013-07-03 15:01:50 -07002987 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002988 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002989 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08002990
Mel Gormanb8e83b92013-07-03 15:01:45 -07002991 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002992 goto out;
2993
Linus Torvalds1da177e2005-04-16 15:20:36 -07002994 for (i = 0; i <= end_zone; i++) {
2995 struct zone *zone = pgdat->node_zones + i;
2996
Mel Gorman2ab44f42013-07-03 15:01:47 -07002997 if (!populated_zone(zone))
2998 continue;
2999
Wu Fengguangadea02a2009-09-21 17:01:42 -07003000 lru_pages += zone_reclaimable_pages(zone);
Mel Gorman2ab44f42013-07-03 15:01:47 -07003001
3002 /*
3003 * If any zone is currently balanced then kswapd will
3004 * not call compaction as it is expected that the
3005 * necessary pages are already available.
3006 */
3007 if (pgdat_needs_compaction &&
3008 zone_watermark_ok(zone, order,
3009 low_wmark_pages(zone),
3010 *classzone_idx, 0))
3011 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012 }
3013
3014 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003015 * If we're getting trouble reclaiming, start doing writepage
3016 * even in laptop mode.
3017 */
3018 if (sc.priority < DEF_PRIORITY - 2)
3019 sc.may_writepage = 1;
3020
3021 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003022 * Now scan the zone in the dma->highmem direction, stopping
3023 * at the last zone which needs scanning.
3024 *
3025 * We do this because the page allocator works in the opposite
3026 * direction. This prevents the page allocator from allocating
3027 * pages behind kswapd's direction of progress, which would
3028 * cause too much scanning of the lower zones.
3029 */
3030 for (i = 0; i <= end_zone; i++) {
3031 struct zone *zone = pgdat->node_zones + i;
3032
Con Kolivasf3fe6512006-01-06 00:11:15 -08003033 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003034 continue;
3035
Lisa Du6e543d52013-09-11 14:22:36 -07003036 if (sc.priority != DEF_PRIORITY &&
3037 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003038 continue;
3039
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003041
Andrew Morton0608f432013-09-24 15:27:41 -07003042 nr_soft_scanned = 0;
3043 /*
3044 * Call soft limit reclaim before calling shrink_zone.
3045 */
3046 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3047 order, sc.gfp_mask,
3048 &nr_soft_scanned);
3049 sc.nr_reclaimed += nr_soft_reclaimed;
3050
Rik van Riel32a43302007-10-16 01:24:50 -07003051 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003052 * There should be no need to raise the scanning
3053 * priority if enough pages are already being scanned
3054 * that that high watermark would be met at 100%
3055 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003056 */
Mel Gorman7c954f62013-07-03 15:01:54 -07003057 if (kswapd_shrink_zone(zone, end_zone, &sc,
3058 lru_pages, &nr_attempted))
3059 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 }
Mel Gorman55150612012-07-31 16:44:35 -07003061
3062 /*
3063 * If the low watermark is met there is no need for processes
3064 * to be throttled on pfmemalloc_wait as they should not be
3065 * able to safely make forward progress. Wake them
3066 */
3067 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3068 pfmemalloc_watermark_ok(pgdat))
3069 wake_up(&pgdat->pfmemalloc_wait);
3070
Linus Torvalds1da177e2005-04-16 15:20:36 -07003071 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003072 * Fragmentation may mean that the system cannot be rebalanced
3073 * for high-order allocations in all zones. If twice the
3074 * allocation size has been reclaimed and the zones are still
3075 * not balanced then recheck the watermarks at order-0 to
3076 * prevent kswapd reclaiming excessively. Assume that a
3077 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003079 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003080 order = sc.order = 0;
3081
Mel Gormanb8e83b92013-07-03 15:01:45 -07003082 /* Check if kswapd should be suspending */
3083 if (try_to_freeze() || kthread_should_stop())
3084 break;
3085
3086 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003087 * Compact if necessary and kswapd is reclaiming at least the
3088 * high watermark number of pages as requsted
3089 */
3090 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3091 compact_pgdat(pgdat, order);
3092
3093 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003094 * Raise priority if scanning rate is too low or there was no
3095 * progress in reclaiming pages
3096 */
3097 if (raise_priority || !sc.nr_reclaimed)
3098 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003099 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003100 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003101
Mel Gormanb8e83b92013-07-03 15:01:45 -07003102out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003103 /*
Mel Gorman55150612012-07-31 16:44:35 -07003104 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003105 * makes a decision on the order we were last reclaiming at. However,
3106 * if another caller entered the allocator slow path while kswapd
3107 * was awake, order will remain at the higher level
3108 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003109 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003110 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003111}
3112
Mel Gormandc83edd2011-01-13 15:46:26 -08003113static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003114{
3115 long remaining = 0;
3116 DEFINE_WAIT(wait);
3117
3118 if (freezing(current) || kthread_should_stop())
3119 return;
3120
3121 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3122
3123 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003124 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003125 remaining = schedule_timeout(HZ/10);
3126 finish_wait(&pgdat->kswapd_wait, &wait);
3127 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3128 }
3129
3130 /*
3131 * After a short sleep, check if it was a premature sleep. If not, then
3132 * go fully to sleep until explicitly woken up.
3133 */
Mel Gorman55150612012-07-31 16:44:35 -07003134 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003135 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3136
3137 /*
3138 * vmstat counters are not perfectly accurate and the estimated
3139 * value for counters such as NR_FREE_PAGES can deviate from the
3140 * true value by nr_online_cpus * threshold. To avoid the zone
3141 * watermarks being breached while under pressure, we reduce the
3142 * per-cpu vmstat threshold while kswapd is awake and restore
3143 * them before going back to sleep.
3144 */
3145 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003146
Mel Gorman62997022012-10-08 16:32:47 -07003147 /*
3148 * Compaction records what page blocks it recently failed to
3149 * isolate pages from and skips them in the future scanning.
3150 * When kswapd is going to sleep, it is reasonable to assume
3151 * that pages and compaction may succeed so reset the cache.
3152 */
3153 reset_isolation_suitable(pgdat);
3154
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003155 if (!kthread_should_stop())
3156 schedule();
3157
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003158 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3159 } else {
3160 if (remaining)
3161 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3162 else
3163 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3164 }
3165 finish_wait(&pgdat->kswapd_wait, &wait);
3166}
3167
Linus Torvalds1da177e2005-04-16 15:20:36 -07003168/*
3169 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003170 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003171 *
3172 * This basically trickles out pages so that we have _some_
3173 * free memory available even if there is no other activity
3174 * that frees anything up. This is needed for things like routing
3175 * etc, where we otherwise might have all activity going on in
3176 * asynchronous contexts that cannot page things out.
3177 *
3178 * If there are applications that are active memory-allocators
3179 * (most normal use), this basically shouldn't matter.
3180 */
3181static int kswapd(void *p)
3182{
Mel Gorman215ddd62011-07-08 15:39:40 -07003183 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003184 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003185 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003186 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003187 pg_data_t *pgdat = (pg_data_t*)p;
3188 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003189
Linus Torvalds1da177e2005-04-16 15:20:36 -07003190 struct reclaim_state reclaim_state = {
3191 .reclaimed_slab = 0,
3192 };
Rusty Russella70f7302009-03-13 14:49:46 +10303193 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003194
Nick Piggincf40bd12009-01-21 08:12:39 +01003195 lockdep_set_current_reclaim_state(GFP_KERNEL);
3196
Rusty Russell174596a2009-01-01 10:12:29 +10303197 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003198 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003199 current->reclaim_state = &reclaim_state;
3200
3201 /*
3202 * Tell the memory management that we're a "memory allocator",
3203 * and that if we need more memory we should get access to it
3204 * regardless (see "__alloc_pages()"). "kswapd" should
3205 * never get caught in the normal page freeing logic.
3206 *
3207 * (Kswapd normally doesn't need memory anyway, but sometimes
3208 * you need a small amount of memory in order to be able to
3209 * page out something else, and this flag essentially protects
3210 * us from recursively trying to free more memory as we're
3211 * trying to free the first piece of memory in the first place).
3212 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003213 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003214 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215
Mel Gorman215ddd62011-07-08 15:39:40 -07003216 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003217 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003218 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003219 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003221 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003222
Mel Gorman215ddd62011-07-08 15:39:40 -07003223 /*
3224 * If the last balance_pgdat was unsuccessful it's unlikely a
3225 * new request of a similar or harder type will succeed soon
3226 * so consider going to sleep on the basis we reclaimed at
3227 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003228 if (balanced_classzone_idx >= new_classzone_idx &&
3229 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003230 new_order = pgdat->kswapd_max_order;
3231 new_classzone_idx = pgdat->classzone_idx;
3232 pgdat->kswapd_max_order = 0;
3233 pgdat->classzone_idx = pgdat->nr_zones - 1;
3234 }
3235
Mel Gorman99504742011-01-13 15:46:20 -08003236 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003237 /*
3238 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003239 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003240 */
3241 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003242 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003243 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003244 kswapd_try_to_sleep(pgdat, balanced_order,
3245 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003246 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003247 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003248 new_order = order;
3249 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003250 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003251 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003252 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003253
David Rientjes8fe23e02009-12-14 17:58:33 -08003254 ret = try_to_freeze();
3255 if (kthread_should_stop())
3256 break;
3257
3258 /*
3259 * We can speed up thawing tasks if we don't call balance_pgdat
3260 * after returning from the refrigerator
3261 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003262 if (!ret) {
3263 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003264 balanced_classzone_idx = classzone_idx;
3265 balanced_order = balance_pgdat(pgdat, order,
3266 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003267 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003268 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003269
3270 current->reclaim_state = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271 return 0;
3272}
3273
3274/*
3275 * A zone is low on free memory, so wake its kswapd task to service it.
3276 */
Mel Gorman99504742011-01-13 15:46:20 -08003277void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278{
3279 pg_data_t *pgdat;
3280
Con Kolivasf3fe6512006-01-06 00:11:15 -08003281 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282 return;
3283
Paul Jackson02a0e532006-12-13 00:34:25 -08003284 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003285 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003286 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003287 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003288 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003289 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3290 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003291 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003292 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003293 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003294 return;
3295
3296 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003297 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003298}
3299
Wu Fengguangadea02a2009-09-21 17:01:42 -07003300/*
3301 * The reclaimable count would be mostly accurate.
3302 * The less reclaimable pages may be
3303 * - mlocked pages, which will be moved to unevictable list when encountered
3304 * - mapped pages, which may require several travels to be reclaimed
3305 * - dirty pages, which is not "instantly" reclaimable
3306 */
3307unsigned long global_reclaimable_pages(void)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003308{
Wu Fengguangadea02a2009-09-21 17:01:42 -07003309 int nr;
3310
3311 nr = global_page_state(NR_ACTIVE_FILE) +
3312 global_page_state(NR_INACTIVE_FILE);
3313
Shaohua Liec8acf22013-02-22 16:34:38 -08003314 if (get_nr_swap_pages() > 0)
Wu Fengguangadea02a2009-09-21 17:01:42 -07003315 nr += global_page_state(NR_ACTIVE_ANON) +
3316 global_page_state(NR_INACTIVE_ANON);
3317
3318 return nr;
3319}
3320
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003321#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003322/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003323 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003324 * freed pages.
3325 *
3326 * Rather than trying to age LRUs the aim is to preserve the overall
3327 * LRU order by reclaiming preferentially
3328 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003329 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003330unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003331{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003332 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003333 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003334 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3335 .may_swap = 1,
3336 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003337 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003338 .nr_to_reclaim = nr_to_reclaim,
3339 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003340 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003341 .priority = DEF_PRIORITY,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003342 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003343 struct shrink_control shrink = {
3344 .gfp_mask = sc.gfp_mask,
3345 };
3346 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003347 struct task_struct *p = current;
3348 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003350 p->flags |= PF_MEMALLOC;
3351 lockdep_set_current_reclaim_state(sc.gfp_mask);
3352 reclaim_state.reclaimed_slab = 0;
3353 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003354
Ying Hana09ed5e2011-05-24 17:12:26 -07003355 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003356
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003357 p->reclaim_state = NULL;
3358 lockdep_clear_current_reclaim_state();
3359 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003360
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003361 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003363#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364
Linus Torvalds1da177e2005-04-16 15:20:36 -07003365/* It's optimal to keep kswapds on the same CPUs as their memory, but
3366 not required for correctness. So if the last cpu in a node goes
3367 away, we get changed to run anywhere: as the first one comes back,
3368 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003369static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3370 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003371{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003372 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003373
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003374 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003375 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003376 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303377 const struct cpumask *mask;
3378
3379 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003380
Rusty Russell3e597942009-01-01 10:12:24 +10303381 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003382 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003383 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003384 }
3385 }
3386 return NOTIFY_OK;
3387}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388
Yasunori Goto3218ae12006-06-27 02:53:33 -07003389/*
3390 * This kswapd start function will be called by init and node-hot-add.
3391 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3392 */
3393int kswapd_run(int nid)
3394{
3395 pg_data_t *pgdat = NODE_DATA(nid);
3396 int ret = 0;
3397
3398 if (pgdat->kswapd)
3399 return 0;
3400
3401 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3402 if (IS_ERR(pgdat->kswapd)) {
3403 /* failure at boot is fatal */
3404 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003405 pr_err("Failed to start kswapd on node %d\n", nid);
3406 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003407 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003408 }
3409 return ret;
3410}
3411
David Rientjes8fe23e02009-12-14 17:58:33 -08003412/*
Jiang Liud8adde12012-07-11 14:01:52 -07003413 * Called by memory hotplug when all memory in a node is offlined. Caller must
3414 * hold lock_memory_hotplug().
David Rientjes8fe23e02009-12-14 17:58:33 -08003415 */
3416void kswapd_stop(int nid)
3417{
3418 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3419
Jiang Liud8adde12012-07-11 14:01:52 -07003420 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003421 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003422 NODE_DATA(nid)->kswapd = NULL;
3423 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003424}
3425
Linus Torvalds1da177e2005-04-16 15:20:36 -07003426static int __init kswapd_init(void)
3427{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003428 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003429
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003431 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003432 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003433 hotcpu_notifier(cpu_callback, 0);
3434 return 0;
3435}
3436
3437module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003438
3439#ifdef CONFIG_NUMA
3440/*
3441 * Zone reclaim mode
3442 *
3443 * If non-zero call zone_reclaim when the number of free pages falls below
3444 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003445 */
3446int zone_reclaim_mode __read_mostly;
3447
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003448#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003449#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003450#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3451#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3452
Christoph Lameter9eeff232006-01-18 17:42:31 -08003453/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003454 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3455 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3456 * a zone.
3457 */
3458#define ZONE_RECLAIM_PRIORITY 4
3459
Christoph Lameter9eeff232006-01-18 17:42:31 -08003460/*
Christoph Lameter96146342006-07-03 00:24:13 -07003461 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3462 * occur.
3463 */
3464int sysctl_min_unmapped_ratio = 1;
3465
3466/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003467 * If the number of slab pages in a zone grows beyond this percentage then
3468 * slab reclaim needs to occur.
3469 */
3470int sysctl_min_slab_ratio = 5;
3471
Mel Gorman90afa5d2009-06-16 15:33:20 -07003472static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3473{
3474 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3475 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3476 zone_page_state(zone, NR_ACTIVE_FILE);
3477
3478 /*
3479 * It's possible for there to be more file mapped pages than
3480 * accounted for by the pages on the file LRU lists because
3481 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3482 */
3483 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3484}
3485
3486/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3487static long zone_pagecache_reclaimable(struct zone *zone)
3488{
3489 long nr_pagecache_reclaimable;
3490 long delta = 0;
3491
3492 /*
3493 * If RECLAIM_SWAP is set, then all file pages are considered
3494 * potentially reclaimable. Otherwise, we have to worry about
3495 * pages like swapcache and zone_unmapped_file_pages() provides
3496 * a better estimate
3497 */
3498 if (zone_reclaim_mode & RECLAIM_SWAP)
3499 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3500 else
3501 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3502
3503 /* If we can't clean pages, remove dirty pages from consideration */
3504 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3505 delta += zone_page_state(zone, NR_FILE_DIRTY);
3506
3507 /* Watch for any possible underflows due to delta */
3508 if (unlikely(delta > nr_pagecache_reclaimable))
3509 delta = nr_pagecache_reclaimable;
3510
3511 return nr_pagecache_reclaimable - delta;
3512}
3513
Christoph Lameter0ff38492006-09-25 23:31:52 -07003514/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003515 * Try to free up some pages from this zone through reclaim.
3516 */
Andrew Morton179e9632006-03-22 00:08:18 -08003517static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003518{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003519 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003520 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003521 struct task_struct *p = current;
3522 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003523 struct scan_control sc = {
3524 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f82009-03-31 15:19:30 -07003525 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003526 .may_swap = 1,
Andrew Morton62b726c2013-02-22 16:32:24 -08003527 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003528 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003529 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003530 .priority = ZONE_RECLAIM_PRIORITY,
Andrew Morton179e9632006-03-22 00:08:18 -08003531 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003532 struct shrink_control shrink = {
3533 .gfp_mask = sc.gfp_mask,
3534 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003535 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003536
Christoph Lameter9eeff232006-01-18 17:42:31 -08003537 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003538 /*
3539 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3540 * and we also need to be able to write out pages for RECLAIM_WRITE
3541 * and RECLAIM_SWAP.
3542 */
3543 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003544 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003545 reclaim_state.reclaimed_slab = 0;
3546 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003547
Mel Gorman90afa5d2009-06-16 15:33:20 -07003548 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003549 /*
3550 * Free memory by calling shrink zone with increasing
3551 * priorities until we have enough memory freed.
3552 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003553 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003554 shrink_zone(zone, &sc);
3555 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003556 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003557
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003558 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3559 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003560 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003561 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003562 * many pages were freed in this zone. So we take the current
3563 * number of slab pages and shake the slab until it is reduced
3564 * by the same nr_pages that we used for reclaiming unmapped
3565 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003566 */
Dave Chinner0ce3d742013-08-28 10:18:03 +10003567 nodes_clear(shrink.nodes_to_scan);
3568 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003569 for (;;) {
3570 unsigned long lru_pages = zone_reclaimable_pages(zone);
3571
3572 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003573 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003574 break;
3575
3576 /* Freed enough memory */
3577 nr_slab_pages1 = zone_page_state(zone,
3578 NR_SLAB_RECLAIMABLE);
3579 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3580 break;
3581 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003582
3583 /*
3584 * Update nr_reclaimed by the number of slab pages we
3585 * reclaimed from this zone.
3586 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003587 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3588 if (nr_slab_pages1 < nr_slab_pages0)
3589 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003590 }
3591
Christoph Lameter9eeff232006-01-18 17:42:31 -08003592 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003593 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003594 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003595 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003596}
Andrew Morton179e9632006-03-22 00:08:18 -08003597
3598int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3599{
Andrew Morton179e9632006-03-22 00:08:18 -08003600 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003601 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003602
3603 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003604 * Zone reclaim reclaims unmapped file backed pages and
3605 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003606 *
Christoph Lameter96146342006-07-03 00:24:13 -07003607 * A small portion of unmapped file backed pages is needed for
3608 * file I/O otherwise pages read by file I/O will be immediately
3609 * thrown out if the zone is overallocated. So we do not reclaim
3610 * if less than a specified percentage of the zone is used by
3611 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003612 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003613 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3614 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003615 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003616
Lisa Du6e543d52013-09-11 14:22:36 -07003617 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003618 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003619
Andrew Morton179e9632006-03-22 00:08:18 -08003620 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003621 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003622 */
David Rientjesd773ed62007-10-16 23:26:01 -07003623 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003624 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003625
3626 /*
3627 * Only run zone reclaim on the local zone or on zones that do not
3628 * have associated processors. This will favor the local processor
3629 * over remote processors and spread off node memory allocations
3630 * as wide as possible.
3631 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003632 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003633 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003634 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003635
3636 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003637 return ZONE_RECLAIM_NOSCAN;
3638
David Rientjesd773ed62007-10-16 23:26:01 -07003639 ret = __zone_reclaim(zone, gfp_mask, order);
3640 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3641
Mel Gorman24cf725182009-06-16 15:33:23 -07003642 if (!ret)
3643 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3644
David Rientjesd773ed62007-10-16 23:26:01 -07003645 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003646}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003647#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003648
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003649/*
3650 * page_evictable - test whether a page is evictable
3651 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003652 *
3653 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003654 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003655 *
3656 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003657 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003658 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003659 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003660 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003661int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003662{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003663 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003664}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003665
Hugh Dickins850465792012-01-20 14:34:19 -08003666#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003667/**
Hugh Dickins24513262012-01-20 14:34:21 -08003668 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3669 * @pages: array of pages to check
3670 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003671 *
Hugh Dickins24513262012-01-20 14:34:21 -08003672 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003673 *
3674 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003675 */
Hugh Dickins24513262012-01-20 14:34:21 -08003676void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003677{
Johannes Weiner925b7672012-01-12 17:18:15 -08003678 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003679 struct zone *zone = NULL;
3680 int pgscanned = 0;
3681 int pgrescued = 0;
3682 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003683
Hugh Dickins24513262012-01-20 14:34:21 -08003684 for (i = 0; i < nr_pages; i++) {
3685 struct page *page = pages[i];
3686 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003687
Hugh Dickins24513262012-01-20 14:34:21 -08003688 pgscanned++;
3689 pagezone = page_zone(page);
3690 if (pagezone != zone) {
3691 if (zone)
3692 spin_unlock_irq(&zone->lru_lock);
3693 zone = pagezone;
3694 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003695 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003696 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003697
Hugh Dickins24513262012-01-20 14:34:21 -08003698 if (!PageLRU(page) || !PageUnevictable(page))
3699 continue;
3700
Hugh Dickins39b5f292012-10-08 16:33:18 -07003701 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003702 enum lru_list lru = page_lru_base_type(page);
3703
3704 VM_BUG_ON(PageActive(page));
3705 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003706 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3707 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003708 pgrescued++;
3709 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003710 }
Hugh Dickins24513262012-01-20 14:34:21 -08003711
3712 if (zone) {
3713 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3714 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3715 spin_unlock_irq(&zone->lru_lock);
3716 }
Hugh Dickins850465792012-01-20 14:34:19 -08003717}
3718#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003719
Johannes Weiner264e56d2011-10-31 17:09:13 -07003720static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003721{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003722 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003723 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003724 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003725 "one, please send an email to linux-mm@kvack.org.\n",
3726 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003727}
3728
3729/*
3730 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3731 * all nodes' unevictable lists for evictable pages
3732 */
3733unsigned long scan_unevictable_pages;
3734
3735int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003736 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003737 size_t *length, loff_t *ppos)
3738{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003739 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003740 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003741 scan_unevictable_pages = 0;
3742 return 0;
3743}
3744
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003745#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003746/*
3747 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3748 * a specified node's per zone unevictable lists for evictable pages.
3749 */
3750
Kay Sievers10fbcf42011-12-21 14:48:43 -08003751static ssize_t read_scan_unevictable_node(struct device *dev,
3752 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003753 char *buf)
3754{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003755 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003756 return sprintf(buf, "0\n"); /* always zero; should fit... */
3757}
3758
Kay Sievers10fbcf42011-12-21 14:48:43 -08003759static ssize_t write_scan_unevictable_node(struct device *dev,
3760 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003761 const char *buf, size_t count)
3762{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003763 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003764 return 1;
3765}
3766
3767
Kay Sievers10fbcf42011-12-21 14:48:43 -08003768static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003769 read_scan_unevictable_node,
3770 write_scan_unevictable_node);
3771
3772int scan_unevictable_register_node(struct node *node)
3773{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003774 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003775}
3776
3777void scan_unevictable_unregister_node(struct node *node)
3778{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003779 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003780}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003781#endif