blob: a9c74b409681a460f2c2ef5f45b3f2283a81a4b8 [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 Weinera6dc60f2009-03-31 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
Johannes Weinera1c3bfb2014-01-29 14:05:41 -0800150static unsigned long zone_reclaimable_pages(struct zone *zone)
Lisa Du6e543d52013-09-11 14:22:36 -0700151{
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
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800284 /*
285 * Normally, we should not scan less than batch_size objects in one
286 * pass to avoid too frequent shrinker calls, but if the slab has less
287 * than batch_size objects in total and we are really tight on memory,
288 * we will try to reclaim all available objects, otherwise we can end
289 * up failing allocations although there are plenty of reclaimable
290 * objects spread over several slabs with usage less than the
291 * batch_size.
292 *
293 * We detect the "tight on memory" situations by looking at the total
294 * number of objects we want to scan (total_scan). If it is greater
295 * than the total number of objects on slab (max_pass), we must be
296 * scanning at high prio and therefore should try to reclaim as much as
297 * possible.
298 */
299 while (total_scan >= batch_size ||
300 total_scan >= max_pass) {
Dave Chinnera0b02132013-08-28 10:18:16 +1000301 unsigned long ret;
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800302 unsigned long nr_to_scan = min(batch_size, total_scan);
Glauber Costa1d3d4432013-08-28 10:18:04 +1000303
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800304 shrinkctl->nr_to_scan = nr_to_scan;
Dave Chinnera0b02132013-08-28 10:18:16 +1000305 ret = shrinker->scan_objects(shrinker, shrinkctl);
306 if (ret == SHRINK_STOP)
307 break;
308 freed += ret;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000309
Vladimir Davydov0b1fb402014-01-23 15:53:22 -0800310 count_vm_events(SLABS_SCANNED, nr_to_scan);
311 total_scan -= nr_to_scan;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000312
313 cond_resched();
314 }
315
316 /*
317 * move the unused scan count back into the shrinker in a
318 * manner that handles concurrent updates. If we exhausted the
319 * scan, there is no need to do an update.
320 */
321 if (total_scan > 0)
322 new_nr = atomic_long_add_return(total_scan,
323 &shrinker->nr_deferred[nid]);
324 else
325 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
326
327 trace_mm_shrink_slab_end(shrinker, freed, nr, new_nr);
328 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700329}
330
Linus Torvalds1da177e2005-04-16 15:20:36 -0700331/*
332 * Call the shrink functions to age shrinkable caches
333 *
334 * Here we assume it costs one seek to replace a lru page and that it also
335 * takes a seek to recreate a cache object. With this in mind we age equal
336 * percentages of the lru and ageable caches. This should balance the seeks
337 * generated by these structures.
338 *
Simon Arlott183ff222007-10-20 01:27:18 +0200339 * If the vm encountered mapped pages on the LRU it increase the pressure on
Linus Torvalds1da177e2005-04-16 15:20:36 -0700340 * slab to avoid swapping.
341 *
342 * We do weird things to avoid (scanned*seeks*entries) overflowing 32 bits.
343 *
344 * `lru_pages' represents the number of on-LRU pages in all the zones which
345 * are eligible for the caller's allocation attempt. It is used for balancing
346 * slab reclaim versus page reclaim.
akpm@osdl.orgb15e0902005-06-21 17:14:35 -0700347 *
348 * Returns the number of slab objects which we shrunk.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700349 */
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000350unsigned long shrink_slab(struct shrink_control *shrinkctl,
Ying Han1495f232011-05-24 17:12:27 -0700351 unsigned long nr_pages_scanned,
Ying Hana09ed5e2011-05-24 17:12:26 -0700352 unsigned long lru_pages)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700353{
354 struct shrinker *shrinker;
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000355 unsigned long freed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700356
Ying Han1495f232011-05-24 17:12:27 -0700357 if (nr_pages_scanned == 0)
358 nr_pages_scanned = SWAP_CLUSTER_MAX;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700359
Minchan Kimf06590b2011-05-24 17:11:11 -0700360 if (!down_read_trylock(&shrinker_rwsem)) {
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000361 /*
362 * If we would return 0, our callers would understand that we
363 * have nothing else to shrink and give up trying. By returning
364 * 1 we keep it going and assume we'll be able to shrink next
365 * time.
366 */
367 freed = 1;
Minchan Kimf06590b2011-05-24 17:11:11 -0700368 goto out;
369 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700370
371 list_for_each_entry(shrinker, &shrinker_list, list) {
Vladimir Davydovec970972014-01-23 15:53:23 -0800372 if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) {
373 shrinkctl->nid = 0;
Glauber Costa1d3d4432013-08-28 10:18:04 +1000374 freed += shrink_slab_node(shrinkctl, shrinker,
Vladimir Davydovec970972014-01-23 15:53:23 -0800375 nr_pages_scanned, lru_pages);
376 continue;
377 }
378
379 for_each_node_mask(shrinkctl->nid, shrinkctl->nodes_to_scan) {
380 if (node_online(shrinkctl->nid))
381 freed += shrink_slab_node(shrinkctl, shrinker,
382 nr_pages_scanned, lru_pages);
Dave Chinneracf92b42011-07-08 14:14:35 +1000383
Linus Torvalds1da177e2005-04-16 15:20:36 -0700384 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700385 }
386 up_read(&shrinker_rwsem);
Minchan Kimf06590b2011-05-24 17:11:11 -0700387out:
388 cond_resched();
Dave Chinner24f7c6b2013-08-28 10:17:56 +1000389 return freed;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700390}
391
Linus Torvalds1da177e2005-04-16 15:20:36 -0700392static inline int is_page_cache_freeable(struct page *page)
393{
Johannes Weinerceddc3a2009-09-21 17:03:00 -0700394 /*
395 * A freeable page cache page is referenced only by the caller
396 * that isolated the page, the page cache radix tree and
397 * optional buffer heads at page->private.
398 */
Johannes Weineredcf4742009-09-21 17:02:59 -0700399 return page_count(page) - page_has_private(page) == 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700400}
401
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700402static int may_write_to_queue(struct backing_dev_info *bdi,
403 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700404{
Christoph Lameter930d9152006-01-08 01:00:47 -0800405 if (current->flags & PF_SWAPWRITE)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700406 return 1;
407 if (!bdi_write_congested(bdi))
408 return 1;
409 if (bdi == current->backing_dev_info)
410 return 1;
411 return 0;
412}
413
414/*
415 * We detected a synchronous write error writing a page out. Probably
416 * -ENOSPC. We need to propagate that into the address_space for a subsequent
417 * fsync(), msync() or close().
418 *
419 * The tricky part is that after writepage we cannot touch the mapping: nothing
420 * prevents it from being freed up. But we have a ref on the page and once
421 * that page is locked, the mapping is pinned.
422 *
423 * We're allowed to run sleeping lock_page() here because we know the caller has
424 * __GFP_FS.
425 */
426static void handle_write_error(struct address_space *mapping,
427 struct page *page, int error)
428{
Jens Axboe7eaceac2011-03-10 08:52:07 +0100429 lock_page(page);
Guillaume Chazarain3e9f45b2007-05-08 00:23:25 -0700430 if (page_mapping(page) == mapping)
431 mapping_set_error(mapping, error);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700432 unlock_page(page);
433}
434
Christoph Lameter04e62a22006-06-23 02:03:38 -0700435/* possible outcome of pageout() */
436typedef enum {
437 /* failed to write page out, page is locked */
438 PAGE_KEEP,
439 /* move page to the active list, page is locked */
440 PAGE_ACTIVATE,
441 /* page has been sent to the disk successfully, page is unlocked */
442 PAGE_SUCCESS,
443 /* page is clean and locked */
444 PAGE_CLEAN,
445} pageout_t;
446
Linus Torvalds1da177e2005-04-16 15:20:36 -0700447/*
Andrew Morton1742f192006-03-22 00:08:21 -0800448 * pageout is called by shrink_page_list() for each dirty page.
449 * Calls ->writepage().
Linus Torvalds1da177e2005-04-16 15:20:36 -0700450 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700451static pageout_t pageout(struct page *page, struct address_space *mapping,
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -0700452 struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700453{
454 /*
455 * If the page is dirty, only perform writeback if that write
456 * will be non-blocking. To prevent this allocation from being
457 * stalled by pagecache activity. But note that there may be
458 * stalls if we need to run get_block(). We could test
459 * PagePrivate for that.
460 *
Vincent Li6aceb532009-12-14 17:58:49 -0800461 * If this process is currently in __generic_file_aio_write() against
Linus Torvalds1da177e2005-04-16 15:20:36 -0700462 * this page's queue, we can perform writeback even if that
463 * will block.
464 *
465 * If the page is swapcache, write it back even if that would
466 * block, for some throttling. This happens by accident, because
467 * swap_backing_dev_info is bust: it doesn't reflect the
468 * congestion state of the swapdevs. Easy to fix, if needed.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700469 */
470 if (!is_page_cache_freeable(page))
471 return PAGE_KEEP;
472 if (!mapping) {
473 /*
474 * Some data journaling orphaned pages can have
475 * page->mapping == NULL while being dirty with clean buffers.
476 */
David Howells266cf652009-04-03 16:42:36 +0100477 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700478 if (try_to_free_buffers(page)) {
479 ClearPageDirty(page);
Harvey Harrisond40cee22008-04-30 00:55:07 -0700480 printk("%s: orphaned page\n", __func__);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700481 return PAGE_CLEAN;
482 }
483 }
484 return PAGE_KEEP;
485 }
486 if (mapping->a_ops->writepage == NULL)
487 return PAGE_ACTIVATE;
Mel Gorman0e093d992010-10-26 14:21:45 -0700488 if (!may_write_to_queue(mapping->backing_dev_info, sc))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700489 return PAGE_KEEP;
490
491 if (clear_page_dirty_for_io(page)) {
492 int res;
493 struct writeback_control wbc = {
494 .sync_mode = WB_SYNC_NONE,
495 .nr_to_write = SWAP_CLUSTER_MAX,
OGAWA Hirofumi111ebb62006-06-23 02:03:26 -0700496 .range_start = 0,
497 .range_end = LLONG_MAX,
Linus Torvalds1da177e2005-04-16 15:20:36 -0700498 .for_reclaim = 1,
499 };
500
501 SetPageReclaim(page);
502 res = mapping->a_ops->writepage(page, &wbc);
503 if (res < 0)
504 handle_write_error(mapping, page, res);
Zach Brown994fc28c2005-12-15 14:28:17 -0800505 if (res == AOP_WRITEPAGE_ACTIVATE) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700506 ClearPageReclaim(page);
507 return PAGE_ACTIVATE;
508 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700509
Linus Torvalds1da177e2005-04-16 15:20:36 -0700510 if (!PageWriteback(page)) {
511 /* synchronous write or broken a_ops? */
512 ClearPageReclaim(page);
513 }
Mel Gorman23b9da52012-05-29 15:06:20 -0700514 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
Andrew Mortone129b5c2006-09-27 01:50:00 -0700515 inc_zone_page_state(page, NR_VMSCAN_WRITE);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516 return PAGE_SUCCESS;
517 }
518
519 return PAGE_CLEAN;
520}
521
Andrew Mortona649fd92006-10-17 00:09:36 -0700522/*
Nick Piggine2867812008-07-25 19:45:30 -0700523 * Same as remove_mapping, but if the page is removed from the mapping, it
524 * gets returned with a refcount of 0.
Andrew Mortona649fd92006-10-17 00:09:36 -0700525 */
Nick Piggine2867812008-07-25 19:45:30 -0700526static int __remove_mapping(struct address_space *mapping, struct page *page)
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800527{
Nick Piggin28e4d962006-09-25 23:31:23 -0700528 BUG_ON(!PageLocked(page));
529 BUG_ON(mapping != page_mapping(page));
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800530
Nick Piggin19fd6232008-07-25 19:45:32 -0700531 spin_lock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800532 /*
Nick Piggin0fd0e6b2006-09-27 01:50:02 -0700533 * The non racy check for a busy page.
534 *
535 * Must be careful with the order of the tests. When someone has
536 * a ref to the page, it may be possible that they dirty it then
537 * drop the reference. So if PageDirty is tested before page_count
538 * here, then the following race may occur:
539 *
540 * get_user_pages(&page);
541 * [user mapping goes away]
542 * write_to(page);
543 * !PageDirty(page) [good]
544 * SetPageDirty(page);
545 * put_page(page);
546 * !page_count(page) [good, discard it]
547 *
548 * [oops, our write_to data is lost]
549 *
550 * Reversing the order of the tests ensures such a situation cannot
551 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
552 * load is not satisfied before that of page->_count.
553 *
554 * Note that if SetPageDirty is always performed via set_page_dirty,
555 * and thus under tree_lock, then this ordering is not required.
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800556 */
Nick Piggine2867812008-07-25 19:45:30 -0700557 if (!page_freeze_refs(page, 2))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800558 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700559 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
560 if (unlikely(PageDirty(page))) {
561 page_unfreeze_refs(page, 2);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800562 goto cannot_free;
Nick Piggine2867812008-07-25 19:45:30 -0700563 }
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800564
565 if (PageSwapCache(page)) {
566 swp_entry_t swap = { .val = page_private(page) };
567 __delete_from_swap_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700568 spin_unlock_irq(&mapping->tree_lock);
KAMEZAWA Hiroyukicb4b86b2009-06-16 15:32:52 -0700569 swapcache_free(swap, page);
Nick Piggine2867812008-07-25 19:45:30 -0700570 } else {
Linus Torvalds6072d132010-12-01 13:35:19 -0500571 void (*freepage)(struct page *);
572
573 freepage = mapping->a_ops->freepage;
574
Minchan Kime64a7822011-03-22 16:32:44 -0700575 __delete_from_page_cache(page);
Nick Piggin19fd6232008-07-25 19:45:32 -0700576 spin_unlock_irq(&mapping->tree_lock);
Daisuke Nishimurae767e052009-05-28 14:34:28 -0700577 mem_cgroup_uncharge_cache_page(page);
Linus Torvalds6072d132010-12-01 13:35:19 -0500578
579 if (freepage != NULL)
580 freepage(page);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800581 }
582
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800583 return 1;
584
585cannot_free:
Nick Piggin19fd6232008-07-25 19:45:32 -0700586 spin_unlock_irq(&mapping->tree_lock);
Christoph Lameter49d2e9c2006-01-08 01:00:48 -0800587 return 0;
588}
589
Linus Torvalds1da177e2005-04-16 15:20:36 -0700590/*
Nick Piggine2867812008-07-25 19:45:30 -0700591 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
592 * someone else has a ref on the page, abort and return 0. If it was
593 * successfully detached, return 1. Assumes the caller has a single ref on
594 * this page.
595 */
596int remove_mapping(struct address_space *mapping, struct page *page)
597{
598 if (__remove_mapping(mapping, page)) {
599 /*
600 * Unfreezing the refcount with 1 rather than 2 effectively
601 * drops the pagecache ref for us without requiring another
602 * atomic operation.
603 */
604 page_unfreeze_refs(page, 1);
605 return 1;
606 }
607 return 0;
608}
609
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700610/**
611 * putback_lru_page - put previously isolated page onto appropriate LRU list
612 * @page: page to be put back to appropriate lru list
613 *
614 * Add previously isolated @page to appropriate LRU list.
615 * Page may still be unevictable for other reasons.
616 *
617 * lru_lock must not be held, interrupts must be enabled.
618 */
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700619void putback_lru_page(struct page *page)
620{
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700621 bool is_unevictable;
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700622 int was_unevictable = PageUnevictable(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700623
Sasha Levin309381fea2014-01-23 15:52:54 -0800624 VM_BUG_ON_PAGE(PageLRU(page), page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700625
626redo:
627 ClearPageUnevictable(page);
628
Hugh Dickins39b5f292012-10-08 16:33:18 -0700629 if (page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700630 /*
631 * For evictable pages, we can use the cache.
632 * In event of a race, worst case is we end up with an
633 * unevictable page on [in]active list.
634 * We know how to handle that.
635 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700636 is_unevictable = false;
Mel Gormanc53954a2013-07-03 15:02:34 -0700637 lru_cache_add(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700638 } else {
639 /*
640 * Put unevictable pages directly on zone's unevictable
641 * list.
642 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700643 is_unevictable = true;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700644 add_page_to_unevictable_list(page);
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700645 /*
Minchan Kim21ee9f32011-10-31 17:09:28 -0700646 * When racing with an mlock or AS_UNEVICTABLE clearing
647 * (page is unlocked) make sure that if the other thread
648 * does not observe our setting of PG_lru and fails
Hugh Dickins24513262012-01-20 14:34:21 -0800649 * isolation/check_move_unevictable_pages,
Minchan Kim21ee9f32011-10-31 17:09:28 -0700650 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700651 * the page back to the evictable list.
652 *
Minchan Kim21ee9f32011-10-31 17:09:28 -0700653 * The other side is TestClearPageMlocked() or shmem_lock().
Johannes Weiner6a7b9542009-10-26 16:50:00 -0700654 */
655 smp_mb();
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700656 }
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700657
658 /*
659 * page's status can change while we move it among lru. If an evictable
660 * page is on unevictable list, it never be freed. To avoid that,
661 * check after we added it to the list, again.
662 */
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700663 if (is_unevictable && page_evictable(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700664 if (!isolate_lru_page(page)) {
665 put_page(page);
666 goto redo;
667 }
668 /* This means someone else dropped this page from LRU
669 * So, it will be freed or putback to LRU again. There is
670 * nothing to do here.
671 */
672 }
673
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700674 if (was_unevictable && !is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700675 count_vm_event(UNEVICTABLE_PGRESCUED);
Vlastimil Babka0ec3b742013-09-11 14:22:26 -0700676 else if (!was_unevictable && is_unevictable)
Lee Schermerhornbbfd28e2008-10-18 20:26:40 -0700677 count_vm_event(UNEVICTABLE_PGCULLED);
678
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700679 put_page(page); /* drop ref from isolate */
680}
681
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800682enum page_references {
683 PAGEREF_RECLAIM,
684 PAGEREF_RECLAIM_CLEAN,
Johannes Weiner645747462010-03-05 13:42:22 -0800685 PAGEREF_KEEP,
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800686 PAGEREF_ACTIVATE,
687};
688
689static enum page_references page_check_references(struct page *page,
690 struct scan_control *sc)
691{
Johannes Weiner645747462010-03-05 13:42:22 -0800692 int referenced_ptes, referenced_page;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800693 unsigned long vm_flags;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800694
Johannes Weinerc3ac9a82012-05-29 15:06:25 -0700695 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
696 &vm_flags);
Johannes Weiner645747462010-03-05 13:42:22 -0800697 referenced_page = TestClearPageReferenced(page);
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800698
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800699 /*
700 * Mlock lost the isolation race with us. Let try_to_unmap()
701 * move the page to the unevictable list.
702 */
703 if (vm_flags & VM_LOCKED)
704 return PAGEREF_RECLAIM;
705
Johannes Weiner645747462010-03-05 13:42:22 -0800706 if (referenced_ptes) {
Michal Hockoe48982732012-05-29 15:06:45 -0700707 if (PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800708 return PAGEREF_ACTIVATE;
709 /*
710 * All mapped pages start out with page table
711 * references from the instantiating fault, so we need
712 * to look twice if a mapped file page is used more
713 * than once.
714 *
715 * Mark it and spare it for another trip around the
716 * inactive list. Another page table reference will
717 * lead to its activation.
718 *
719 * Note: the mark is set for activated pages as well
720 * so that recently deactivated but used pages are
721 * quickly recovered.
722 */
723 SetPageReferenced(page);
724
Konstantin Khlebnikov34dbc672012-01-10 15:06:59 -0800725 if (referenced_page || referenced_ptes > 1)
Johannes Weiner645747462010-03-05 13:42:22 -0800726 return PAGEREF_ACTIVATE;
727
Konstantin Khlebnikovc909e992012-01-10 15:07:03 -0800728 /*
729 * Activate file-backed executable pages after first usage.
730 */
731 if (vm_flags & VM_EXEC)
732 return PAGEREF_ACTIVATE;
733
Johannes Weiner645747462010-03-05 13:42:22 -0800734 return PAGEREF_KEEP;
735 }
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800736
737 /* Reclaim if clean, defer dirty pages to writeback */
KOSAKI Motohiro2e302442010-10-26 14:21:46 -0700738 if (referenced_page && !PageSwapBacked(page))
Johannes Weiner645747462010-03-05 13:42:22 -0800739 return PAGEREF_RECLAIM_CLEAN;
740
741 return PAGEREF_RECLAIM;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800742}
743
Mel Gormane2be15f2013-07-03 15:01:57 -0700744/* Check if a page is dirty or under writeback */
745static void page_check_dirty_writeback(struct page *page,
746 bool *dirty, bool *writeback)
747{
Mel Gormanb4597222013-07-03 15:02:05 -0700748 struct address_space *mapping;
749
Mel Gormane2be15f2013-07-03 15:01:57 -0700750 /*
751 * Anonymous pages are not handled by flushers and must be written
752 * from reclaim context. Do not stall reclaim based on them
753 */
754 if (!page_is_file_cache(page)) {
755 *dirty = false;
756 *writeback = false;
757 return;
758 }
759
760 /* By default assume that the page flags are accurate */
761 *dirty = PageDirty(page);
762 *writeback = PageWriteback(page);
Mel Gormanb4597222013-07-03 15:02:05 -0700763
764 /* Verify dirty/writeback state if the filesystem supports it */
765 if (!page_has_private(page))
766 return;
767
768 mapping = page_mapping(page);
769 if (mapping && mapping->a_ops->is_dirty_writeback)
770 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
Mel Gormane2be15f2013-07-03 15:01:57 -0700771}
772
Nick Piggine2867812008-07-25 19:45:30 -0700773/*
Andrew Morton1742f192006-03-22 00:08:21 -0800774 * shrink_page_list() returns the number of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -0700775 */
Andrew Morton1742f192006-03-22 00:08:21 -0800776static unsigned long shrink_page_list(struct list_head *page_list,
Konstantin Khlebnikov6a18adb2012-05-29 15:06:59 -0700777 struct zone *zone,
Mel Gormanf84f6e22011-10-31 17:07:51 -0700778 struct scan_control *sc,
Minchan Kim02c6de82012-10-08 16:31:55 -0700779 enum ttu_flags ttu_flags,
Mel Gorman8e950282013-07-03 15:02:02 -0700780 unsigned long *ret_nr_dirty,
Mel Gormand43006d2013-07-03 15:01:50 -0700781 unsigned long *ret_nr_unqueued_dirty,
Mel Gorman8e950282013-07-03 15:02:02 -0700782 unsigned long *ret_nr_congested,
Minchan Kim02c6de82012-10-08 16:31:55 -0700783 unsigned long *ret_nr_writeback,
Mel Gormanb1a6f212013-07-03 15:01:58 -0700784 unsigned long *ret_nr_immediate,
Minchan Kim02c6de82012-10-08 16:31:55 -0700785 bool force_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700786{
787 LIST_HEAD(ret_pages);
Mel Gormanabe4c3b2010-08-09 17:19:31 -0700788 LIST_HEAD(free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700789 int pgactivate = 0;
Mel Gormand43006d2013-07-03 15:01:50 -0700790 unsigned long nr_unqueued_dirty = 0;
Mel Gorman0e093d992010-10-26 14:21:45 -0700791 unsigned long nr_dirty = 0;
792 unsigned long nr_congested = 0;
Andrew Morton05ff5132006-03-22 00:08:20 -0800793 unsigned long nr_reclaimed = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -0700794 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -0700795 unsigned long nr_immediate = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700796
797 cond_resched();
798
Tim Chen69980e32012-07-31 16:46:08 -0700799 mem_cgroup_uncharge_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -0700800 while (!list_empty(page_list)) {
801 struct address_space *mapping;
802 struct page *page;
803 int may_enter_fs;
Minchan Kim02c6de82012-10-08 16:31:55 -0700804 enum page_references references = PAGEREF_RECLAIM_CLEAN;
Mel Gormane2be15f2013-07-03 15:01:57 -0700805 bool dirty, writeback;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700806
807 cond_resched();
808
809 page = lru_to_page(page_list);
810 list_del(&page->lru);
811
Nick Piggin529ae9a2008-08-02 12:01:03 +0200812 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700813 goto keep;
814
Sasha Levin309381fea2014-01-23 15:52:54 -0800815 VM_BUG_ON_PAGE(PageActive(page), page);
816 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700817
818 sc->nr_scanned++;
Christoph Lameter80e43422006-02-11 17:55:53 -0800819
Hugh Dickins39b5f292012-10-08 16:33:18 -0700820 if (unlikely(!page_evictable(page)))
Nick Pigginb291f002008-10-18 20:26:44 -0700821 goto cull_mlocked;
Lee Schermerhorn894bc312008-10-18 20:26:39 -0700822
Johannes Weinera6dc60f2009-03-31 15:19:30 -0700823 if (!sc->may_unmap && page_mapped(page))
Christoph Lameter80e43422006-02-11 17:55:53 -0800824 goto keep_locked;
825
Linus Torvalds1da177e2005-04-16 15:20:36 -0700826 /* Double the slab pressure for mapped and swapcache pages */
827 if (page_mapped(page) || PageSwapCache(page))
828 sc->nr_scanned++;
829
Andy Whitcroftc661b072007-08-22 14:01:26 -0700830 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
831 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
832
Mel Gorman283aba92013-07-03 15:01:51 -0700833 /*
Mel Gormane2be15f2013-07-03 15:01:57 -0700834 * The number of dirty pages determines if a zone is marked
835 * reclaim_congested which affects wait_iff_congested. kswapd
836 * will stall and start writing pages if the tail of the LRU
837 * is all dirty unqueued pages.
838 */
839 page_check_dirty_writeback(page, &dirty, &writeback);
840 if (dirty || writeback)
841 nr_dirty++;
842
843 if (dirty && !writeback)
844 nr_unqueued_dirty++;
845
Mel Gormand04e8ac2013-07-03 15:02:03 -0700846 /*
847 * Treat this page as congested if the underlying BDI is or if
848 * pages are cycling through the LRU so quickly that the
849 * pages marked for immediate reclaim are making it to the
850 * end of the LRU a second time.
851 */
Mel Gormane2be15f2013-07-03 15:01:57 -0700852 mapping = page_mapping(page);
Mel Gormand04e8ac2013-07-03 15:02:03 -0700853 if ((mapping && bdi_write_congested(mapping->backing_dev_info)) ||
854 (writeback && PageReclaim(page)))
Mel Gormane2be15f2013-07-03 15:01:57 -0700855 nr_congested++;
856
857 /*
Mel Gorman283aba92013-07-03 15:01:51 -0700858 * If a page at the tail of the LRU is under writeback, there
859 * are three cases to consider.
860 *
861 * 1) If reclaim is encountering an excessive number of pages
862 * under writeback and this page is both under writeback and
863 * PageReclaim then it indicates that pages are being queued
864 * for IO but are being recycled through the LRU before the
865 * IO can complete. Waiting on the page itself risks an
866 * indefinite stall if it is impossible to writeback the
867 * page due to IO error or disconnected storage so instead
Mel Gormanb1a6f212013-07-03 15:01:58 -0700868 * note that the LRU is being scanned too quickly and the
869 * caller can stall after page list has been processed.
Mel Gorman283aba92013-07-03 15:01:51 -0700870 *
871 * 2) Global reclaim encounters a page, memcg encounters a
872 * page that is not marked for immediate reclaim or
873 * the caller does not have __GFP_IO. In this case mark
874 * the page for immediate reclaim and continue scanning.
875 *
876 * __GFP_IO is checked because a loop driver thread might
877 * enter reclaim, and deadlock if it waits on a page for
878 * which it is needed to do the write (loop masks off
879 * __GFP_IO|__GFP_FS for this reason); but more thought
880 * would probably show more reasons.
881 *
882 * Don't require __GFP_FS, since we're not going into the
883 * FS, just waiting on its writeback completion. Worryingly,
884 * ext4 gfs2 and xfs allocate pages with
885 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
886 * may_enter_fs here is liable to OOM on them.
887 *
888 * 3) memcg encounters a page that is not already marked
889 * PageReclaim. memcg does not have any dirty pages
890 * throttling so we could easily OOM just because too many
891 * pages are in writeback and there is nothing else to
892 * reclaim. Wait for the writeback to complete.
893 */
Andy Whitcroftc661b072007-08-22 14:01:26 -0700894 if (PageWriteback(page)) {
Mel Gorman283aba92013-07-03 15:01:51 -0700895 /* Case 1 above */
896 if (current_is_kswapd() &&
897 PageReclaim(page) &&
898 zone_is_reclaim_writeback(zone)) {
Mel Gormanb1a6f212013-07-03 15:01:58 -0700899 nr_immediate++;
900 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700901
902 /* Case 2 above */
903 } else if (global_reclaim(sc) ||
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700904 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
905 /*
906 * This is slightly racy - end_page_writeback()
907 * might have just cleared PageReclaim, then
908 * setting PageReclaim here end up interpreted
909 * as PageReadahead - but that does not matter
910 * enough to care. What we do want is for this
911 * page to have PageReclaim set next time memcg
912 * reclaim reaches the tests above, so it will
913 * then wait_on_page_writeback() to avoid OOM;
914 * and it's also appropriate in global reclaim.
915 */
916 SetPageReclaim(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700917 nr_writeback++;
Mel Gorman283aba92013-07-03 15:01:51 -0700918
Hugh Dickinsc3b94f42012-07-31 16:45:59 -0700919 goto keep_locked;
Mel Gorman283aba92013-07-03 15:01:51 -0700920
921 /* Case 3 above */
922 } else {
923 wait_on_page_writeback(page);
Michal Hockoe62e3842012-07-31 16:45:55 -0700924 }
Andy Whitcroftc661b072007-08-22 14:01:26 -0700925 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700926
Minchan Kim02c6de82012-10-08 16:31:55 -0700927 if (!force_reclaim)
928 references = page_check_references(page, sc);
929
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800930 switch (references) {
931 case PAGEREF_ACTIVATE:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700932 goto activate_locked;
Johannes Weiner645747462010-03-05 13:42:22 -0800933 case PAGEREF_KEEP:
934 goto keep_locked;
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800935 case PAGEREF_RECLAIM:
936 case PAGEREF_RECLAIM_CLEAN:
937 ; /* try to reclaim the page below */
938 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700939
Linus Torvalds1da177e2005-04-16 15:20:36 -0700940 /*
941 * Anonymous process memory has backing store?
942 * Try to allocate it some swap space here.
943 */
Nick Pigginb291f002008-10-18 20:26:44 -0700944 if (PageAnon(page) && !PageSwapCache(page)) {
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800945 if (!(sc->gfp_mask & __GFP_IO))
946 goto keep_locked;
Shaohua Li5bc7b8a2013-04-29 15:08:36 -0700947 if (!add_to_swap(page, page_list))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700948 goto activate_locked;
Hugh Dickins63eb6b92008-11-19 15:36:37 -0800949 may_enter_fs = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700950
Mel Gormane2be15f2013-07-03 15:01:57 -0700951 /* Adding to swap updated mapping */
952 mapping = page_mapping(page);
953 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700954
955 /*
956 * The page is mapped into the page tables of one or more
957 * processes. Try to unmap it here.
958 */
959 if (page_mapped(page) && mapping) {
Minchan Kim02c6de82012-10-08 16:31:55 -0700960 switch (try_to_unmap(page, ttu_flags)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 case SWAP_FAIL:
962 goto activate_locked;
963 case SWAP_AGAIN:
964 goto keep_locked;
Nick Pigginb291f002008-10-18 20:26:44 -0700965 case SWAP_MLOCK:
966 goto cull_mlocked;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967 case SWAP_SUCCESS:
968 ; /* try to free the page below */
969 }
970 }
971
972 if (PageDirty(page)) {
Mel Gormanee728862011-10-31 17:07:38 -0700973 /*
974 * Only kswapd can writeback filesystem pages to
Mel Gormand43006d2013-07-03 15:01:50 -0700975 * avoid risk of stack overflow but only writeback
976 * if many dirty pages have been encountered.
Mel Gormanee728862011-10-31 17:07:38 -0700977 */
Mel Gormanf84f6e22011-10-31 17:07:51 -0700978 if (page_is_file_cache(page) &&
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -0700979 (!current_is_kswapd() ||
Mel Gormand43006d2013-07-03 15:01:50 -0700980 !zone_is_reclaim_dirty(zone))) {
Mel Gorman49ea7eb2011-10-31 17:07:59 -0700981 /*
982 * Immediately reclaim when written back.
983 * Similar in principal to deactivate_page()
984 * except we already have the page isolated
985 * and know it's dirty
986 */
987 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
988 SetPageReclaim(page);
989
Mel Gormanee728862011-10-31 17:07:38 -0700990 goto keep_locked;
991 }
992
Johannes Weinerdfc8d632010-03-05 13:42:19 -0800993 if (references == PAGEREF_RECLAIM_CLEAN)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700994 goto keep_locked;
Andrew Morton4dd4b922008-03-24 12:29:52 -0700995 if (!may_enter_fs)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996 goto keep_locked;
Christoph Lameter52a83632006-02-01 03:05:28 -0800997 if (!sc->may_writepage)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998 goto keep_locked;
999
1000 /* Page is dirty, try to write it out here */
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001001 switch (pageout(page, mapping, sc)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 case PAGE_KEEP:
1003 goto keep_locked;
1004 case PAGE_ACTIVATE:
1005 goto activate_locked;
1006 case PAGE_SUCCESS:
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001007 if (PageWriteback(page))
Mel Gorman41ac1992012-05-29 15:06:19 -07001008 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001009 if (PageDirty(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001010 goto keep;
KOSAKI Motohiro7d3579e2010-10-26 14:21:42 -07001011
Linus Torvalds1da177e2005-04-16 15:20:36 -07001012 /*
1013 * A synchronous write - probably a ramdisk. Go
1014 * ahead and try to reclaim the page.
1015 */
Nick Piggin529ae9a2008-08-02 12:01:03 +02001016 if (!trylock_page(page))
Linus Torvalds1da177e2005-04-16 15:20:36 -07001017 goto keep;
1018 if (PageDirty(page) || PageWriteback(page))
1019 goto keep_locked;
1020 mapping = page_mapping(page);
1021 case PAGE_CLEAN:
1022 ; /* try to free the page below */
1023 }
1024 }
1025
1026 /*
1027 * If the page has buffers, try to free the buffer mappings
1028 * associated with this page. If we succeed we try to free
1029 * the page as well.
1030 *
1031 * We do this even if the page is PageDirty().
1032 * try_to_release_page() does not perform I/O, but it is
1033 * possible for a page to have PageDirty set, but it is actually
1034 * clean (all its buffers are clean). This happens if the
1035 * buffers were written out directly, with submit_bh(). ext3
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001036 * will do this, as well as the blockdev mapping.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001037 * try_to_release_page() will discover that cleanness and will
1038 * drop the buffers and mark the page clean - it can be freed.
1039 *
1040 * Rarely, pages can have buffers and no ->mapping. These are
1041 * the pages which were not successfully invalidated in
1042 * truncate_complete_page(). We try to drop those buffers here
1043 * and if that worked, and the page is no longer mapped into
1044 * process address space (page_count == 1) it can be freed.
1045 * Otherwise, leave the page on the LRU so it is swappable.
1046 */
David Howells266cf652009-04-03 16:42:36 +01001047 if (page_has_private(page)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001048 if (!try_to_release_page(page, sc->gfp_mask))
1049 goto activate_locked;
Nick Piggine2867812008-07-25 19:45:30 -07001050 if (!mapping && page_count(page) == 1) {
1051 unlock_page(page);
1052 if (put_page_testzero(page))
1053 goto free_it;
1054 else {
1055 /*
1056 * rare race with speculative reference.
1057 * the speculative reference will free
1058 * this page shortly, so we may
1059 * increment nr_reclaimed here (and
1060 * leave it off the LRU).
1061 */
1062 nr_reclaimed++;
1063 continue;
1064 }
1065 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001066 }
1067
Nick Piggine2867812008-07-25 19:45:30 -07001068 if (!mapping || !__remove_mapping(mapping, page))
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001069 goto keep_locked;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001070
Nick Piggina978d6f2008-10-18 20:26:58 -07001071 /*
1072 * At this point, we have no other references and there is
1073 * no way to pick any more up (removed from LRU, removed
1074 * from pagecache). Can use non-atomic bitops now (and
1075 * we obviously don't have to worry about waking up a process
1076 * waiting on the page lock, because there are no references.
1077 */
1078 __clear_page_locked(page);
Nick Piggine2867812008-07-25 19:45:30 -07001079free_it:
Andrew Morton05ff5132006-03-22 00:08:20 -08001080 nr_reclaimed++;
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001081
1082 /*
1083 * Is there need to periodically free_page_list? It would
1084 * appear not as the counts should be low
1085 */
1086 list_add(&page->lru, &free_pages);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001087 continue;
1088
Nick Pigginb291f002008-10-18 20:26:44 -07001089cull_mlocked:
Hugh Dickins63d6c5a2009-01-06 14:39:38 -08001090 if (PageSwapCache(page))
1091 try_to_free_swap(page);
Nick Pigginb291f002008-10-18 20:26:44 -07001092 unlock_page(page);
1093 putback_lru_page(page);
1094 continue;
1095
Linus Torvalds1da177e2005-04-16 15:20:36 -07001096activate_locked:
Rik van Riel68a223942008-10-18 20:26:23 -07001097 /* Not a candidate for swapping, so reclaim swap space. */
1098 if (PageSwapCache(page) && vm_swap_full())
Hugh Dickinsa2c43ee2009-01-06 14:39:36 -08001099 try_to_free_swap(page);
Sasha Levin309381fea2014-01-23 15:52:54 -08001100 VM_BUG_ON_PAGE(PageActive(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001101 SetPageActive(page);
1102 pgactivate++;
1103keep_locked:
1104 unlock_page(page);
1105keep:
1106 list_add(&page->lru, &ret_pages);
Sasha Levin309381fea2014-01-23 15:52:54 -08001107 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001108 }
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001109
Konstantin Khlebnikovcc598502012-01-10 15:07:04 -08001110 free_hot_cold_page_list(&free_pages, 1);
Mel Gormanabe4c3b2010-08-09 17:19:31 -07001111
Linus Torvalds1da177e2005-04-16 15:20:36 -07001112 list_splice(&ret_pages, page_list);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001113 count_vm_events(PGACTIVATE, pgactivate);
Tim Chen69980e32012-07-31 16:46:08 -07001114 mem_cgroup_uncharge_end();
Mel Gorman8e950282013-07-03 15:02:02 -07001115 *ret_nr_dirty += nr_dirty;
1116 *ret_nr_congested += nr_congested;
Mel Gormand43006d2013-07-03 15:01:50 -07001117 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
Mel Gorman92df3a72011-10-31 17:07:56 -07001118 *ret_nr_writeback += nr_writeback;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001119 *ret_nr_immediate += nr_immediate;
Andrew Morton05ff5132006-03-22 00:08:20 -08001120 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001121}
1122
Minchan Kim02c6de82012-10-08 16:31:55 -07001123unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1124 struct list_head *page_list)
1125{
1126 struct scan_control sc = {
1127 .gfp_mask = GFP_KERNEL,
1128 .priority = DEF_PRIORITY,
1129 .may_unmap = 1,
1130 };
Mel Gorman8e950282013-07-03 15:02:02 -07001131 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
Minchan Kim02c6de82012-10-08 16:31:55 -07001132 struct page *page, *next;
1133 LIST_HEAD(clean_pages);
1134
1135 list_for_each_entry_safe(page, next, page_list, lru) {
Rafael Aquini117aad12013-09-30 13:45:16 -07001136 if (page_is_file_cache(page) && !PageDirty(page) &&
1137 !isolated_balloon_page(page)) {
Minchan Kim02c6de82012-10-08 16:31:55 -07001138 ClearPageActive(page);
1139 list_move(&page->lru, &clean_pages);
1140 }
1141 }
1142
1143 ret = shrink_page_list(&clean_pages, zone, &sc,
Mel Gorman8e950282013-07-03 15:02:02 -07001144 TTU_UNMAP|TTU_IGNORE_ACCESS,
1145 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
Minchan Kim02c6de82012-10-08 16:31:55 -07001146 list_splice(&clean_pages, page_list);
1147 __mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
1148 return ret;
1149}
1150
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001151/*
1152 * Attempt to remove the specified page from its LRU. Only take this page
1153 * if it is of the appropriate PageActive status. Pages which are being
1154 * freed elsewhere are also ignored.
1155 *
1156 * page: page to consider
1157 * mode: one of the LRU isolation modes defined above
1158 *
1159 * returns 0 on success, -ve errno on failure.
1160 */
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001161int __isolate_lru_page(struct page *page, isolate_mode_t mode)
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001162{
1163 int ret = -EINVAL;
1164
1165 /* Only take pages on the LRU. */
1166 if (!PageLRU(page))
1167 return ret;
1168
Minchan Kime46a2872012-10-08 16:33:48 -07001169 /* Compaction should not handle unevictable pages but CMA can do so */
1170 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001171 return ret;
1172
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001173 ret = -EBUSY;
KAMEZAWA Hiroyuki08e552c2009-01-07 18:08:01 -08001174
Mel Gormanc8244932012-01-12 17:19:38 -08001175 /*
1176 * To minimise LRU disruption, the caller can indicate that it only
1177 * wants to isolate pages it will be able to operate on without
1178 * blocking - clean pages for the most part.
1179 *
1180 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1181 * is used by reclaim when it is cannot write to backing storage
1182 *
1183 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1184 * that it is possible to migrate without blocking
1185 */
1186 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1187 /* All the caller can do on PageWriteback is block */
1188 if (PageWriteback(page))
1189 return ret;
1190
1191 if (PageDirty(page)) {
1192 struct address_space *mapping;
1193
1194 /* ISOLATE_CLEAN means only clean pages */
1195 if (mode & ISOLATE_CLEAN)
1196 return ret;
1197
1198 /*
1199 * Only pages without mappings or that have a
1200 * ->migratepage callback are possible to migrate
1201 * without blocking
1202 */
1203 mapping = page_mapping(page);
1204 if (mapping && !mapping->a_ops->migratepage)
1205 return ret;
1206 }
1207 }
Minchan Kim39deaf82011-10-31 17:06:51 -07001208
Minchan Kimf80c0672011-10-31 17:06:55 -07001209 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1210 return ret;
1211
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001212 if (likely(get_page_unless_zero(page))) {
1213 /*
1214 * Be careful not to clear PageLRU until after we're
1215 * sure the page is not being freed elsewhere -- the
1216 * page release code relies on it.
1217 */
1218 ClearPageLRU(page);
1219 ret = 0;
1220 }
1221
1222 return ret;
1223}
1224
Christoph Lameter49d2e9c2006-01-08 01:00:48 -08001225/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001226 * zone->lru_lock is heavily contended. Some of the functions that
1227 * shrink the lists perform better by taking out a batch of pages
1228 * and working on them outside the LRU lock.
1229 *
1230 * For pagecache intensive workloads, this function is the hottest
1231 * spot in the kernel (apart from copy_*_user functions).
1232 *
1233 * Appropriate locks must be held before calling this function.
1234 *
1235 * @nr_to_scan: The number of pages to look through on the list.
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001236 * @lruvec: The LRU vector to pull pages from.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001237 * @dst: The temp list to put pages on to.
Hugh Dickinsf6260122012-01-12 17:20:06 -08001238 * @nr_scanned: The number of pages that were scanned.
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001239 * @sc: The scan_control struct for this reclaim session
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001240 * @mode: One of the LRU isolation modes
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001241 * @lru: LRU list id for isolating
Linus Torvalds1da177e2005-04-16 15:20:36 -07001242 *
1243 * returns how many pages were moved onto *@dst.
1244 */
Andrew Morton69e05942006-03-22 00:08:19 -08001245static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001246 struct lruvec *lruvec, struct list_head *dst,
Rik van Rielfe2c2a12012-03-21 16:33:51 -07001247 unsigned long *nr_scanned, struct scan_control *sc,
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001248 isolate_mode_t mode, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001249{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001250 struct list_head *src = &lruvec->lists[lru];
Andrew Morton69e05942006-03-22 00:08:19 -08001251 unsigned long nr_taken = 0;
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001252 unsigned long scan;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001253
Wu Fengguangc9b02d92006-03-22 00:08:23 -08001254 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001255 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001256 int nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001257
Linus Torvalds1da177e2005-04-16 15:20:36 -07001258 page = lru_to_page(src);
1259 prefetchw_prev_lru_page(page, src, flags);
1260
Sasha Levin309381fea2014-01-23 15:52:54 -08001261 VM_BUG_ON_PAGE(!PageLRU(page), page);
Nick Piggin8d438f92006-03-22 00:07:59 -08001262
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001263 switch (__isolate_lru_page(page, mode)) {
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001264 case 0:
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001265 nr_pages = hpage_nr_pages(page);
1266 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001267 list_move(&page->lru, dst);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001268 nr_taken += nr_pages;
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001269 break;
Nick Piggin46453a62006-03-22 00:07:58 -08001270
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001271 case -EBUSY:
1272 /* else it is being freed elsewhere */
1273 list_move(&page->lru, src);
1274 continue;
1275
1276 default:
1277 BUG();
1278 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001279 }
1280
Hugh Dickinsf6260122012-01-12 17:20:06 -08001281 *nr_scanned = scan;
Hugh Dickins75b00af2012-05-29 15:07:09 -07001282 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1283 nr_taken, mode, is_file_lru(lru));
Linus Torvalds1da177e2005-04-16 15:20:36 -07001284 return nr_taken;
1285}
1286
Nick Piggin62695a82008-10-18 20:26:09 -07001287/**
1288 * isolate_lru_page - tries to isolate a page from its LRU list
1289 * @page: page to isolate from its LRU list
1290 *
1291 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1292 * vmstat statistic corresponding to whatever LRU list the page was on.
1293 *
1294 * Returns 0 if the page was removed from an LRU list.
1295 * Returns -EBUSY if the page was not on an LRU list.
1296 *
1297 * The returned page will have PageLRU() cleared. If it was found on
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001298 * the active list, it will have PageActive set. If it was found on
1299 * the unevictable list, it will have the PageUnevictable bit set. That flag
1300 * may need to be cleared by the caller before letting the page go.
Nick Piggin62695a82008-10-18 20:26:09 -07001301 *
1302 * The vmstat statistic corresponding to the list on which the page was
1303 * found will be decremented.
1304 *
1305 * Restrictions:
1306 * (1) Must be called with an elevated refcount on the page. This is a
1307 * fundamentnal difference from isolate_lru_pages (which is called
1308 * without a stable reference).
1309 * (2) the lru_lock must not be held.
1310 * (3) interrupts must be enabled.
1311 */
1312int isolate_lru_page(struct page *page)
1313{
1314 int ret = -EBUSY;
1315
Sasha Levin309381fea2014-01-23 15:52:54 -08001316 VM_BUG_ON_PAGE(!page_count(page), page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001317
Nick Piggin62695a82008-10-18 20:26:09 -07001318 if (PageLRU(page)) {
1319 struct zone *zone = page_zone(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001320 struct lruvec *lruvec;
Nick Piggin62695a82008-10-18 20:26:09 -07001321
1322 spin_lock_irq(&zone->lru_lock);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001323 lruvec = mem_cgroup_page_lruvec(page, zone);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001324 if (PageLRU(page)) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001325 int lru = page_lru(page);
Konstantin Khlebnikov0c917312011-05-24 17:12:21 -07001326 get_page(page);
Nick Piggin62695a82008-10-18 20:26:09 -07001327 ClearPageLRU(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001328 del_page_from_lru_list(page, lruvec, lru);
1329 ret = 0;
Nick Piggin62695a82008-10-18 20:26:09 -07001330 }
1331 spin_unlock_irq(&zone->lru_lock);
1332 }
1333 return ret;
1334}
1335
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07001336/*
Fengguang Wud37dd5d2012-12-18 14:23:28 -08001337 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1338 * then get resheduled. When there are massive number of tasks doing page
1339 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1340 * the LRU list will go small and be scanned faster than necessary, leading to
1341 * unnecessary swapping, thrashing and OOM.
Rik van Riel35cd7812009-09-21 17:01:38 -07001342 */
1343static int too_many_isolated(struct zone *zone, int file,
1344 struct scan_control *sc)
1345{
1346 unsigned long inactive, isolated;
1347
1348 if (current_is_kswapd())
1349 return 0;
1350
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001351 if (!global_reclaim(sc))
Rik van Riel35cd7812009-09-21 17:01:38 -07001352 return 0;
1353
1354 if (file) {
1355 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1356 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1357 } else {
1358 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1359 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1360 }
1361
Fengguang Wu3cf23842012-12-18 14:23:31 -08001362 /*
1363 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1364 * won't get blocked by normal direct-reclaimers, forming a circular
1365 * deadlock.
1366 */
1367 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1368 inactive >>= 3;
1369
Rik van Riel35cd7812009-09-21 17:01:38 -07001370 return isolated > inactive;
1371}
1372
Mel Gorman66635622010-08-09 17:19:30 -07001373static noinline_for_stack void
Hugh Dickins75b00af2012-05-29 15:07:09 -07001374putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
Mel Gorman66635622010-08-09 17:19:30 -07001375{
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001376 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1377 struct zone *zone = lruvec_zone(lruvec);
Hugh Dickins3f797682012-01-12 17:20:07 -08001378 LIST_HEAD(pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001379
Mel Gorman66635622010-08-09 17:19:30 -07001380 /*
1381 * Put back any unfreeable pages.
1382 */
Mel Gorman66635622010-08-09 17:19:30 -07001383 while (!list_empty(page_list)) {
Hugh Dickins3f797682012-01-12 17:20:07 -08001384 struct page *page = lru_to_page(page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001385 int lru;
Hugh Dickins3f797682012-01-12 17:20:07 -08001386
Sasha Levin309381fea2014-01-23 15:52:54 -08001387 VM_BUG_ON_PAGE(PageLRU(page), page);
Mel Gorman66635622010-08-09 17:19:30 -07001388 list_del(&page->lru);
Hugh Dickins39b5f292012-10-08 16:33:18 -07001389 if (unlikely(!page_evictable(page))) {
Mel Gorman66635622010-08-09 17:19:30 -07001390 spin_unlock_irq(&zone->lru_lock);
1391 putback_lru_page(page);
1392 spin_lock_irq(&zone->lru_lock);
1393 continue;
1394 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001395
1396 lruvec = mem_cgroup_page_lruvec(page, zone);
1397
Linus Torvalds7a608572011-01-17 14:42:19 -08001398 SetPageLRU(page);
Mel Gorman66635622010-08-09 17:19:30 -07001399 lru = page_lru(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001400 add_page_to_lru_list(page, lruvec, lru);
1401
Mel Gorman66635622010-08-09 17:19:30 -07001402 if (is_active_lru(lru)) {
1403 int file = is_file_lru(lru);
Rik van Riel9992af12011-01-13 15:47:13 -08001404 int numpages = hpage_nr_pages(page);
1405 reclaim_stat->recent_rotated[file] += numpages;
Mel Gorman66635622010-08-09 17:19:30 -07001406 }
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001407 if (put_page_testzero(page)) {
1408 __ClearPageLRU(page);
1409 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001410 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001411
1412 if (unlikely(PageCompound(page))) {
1413 spin_unlock_irq(&zone->lru_lock);
1414 (*get_compound_page_dtor(page))(page);
1415 spin_lock_irq(&zone->lru_lock);
1416 } else
1417 list_add(&page->lru, &pages_to_free);
Mel Gorman66635622010-08-09 17:19:30 -07001418 }
1419 }
Mel Gorman66635622010-08-09 17:19:30 -07001420
Hugh Dickins3f797682012-01-12 17:20:07 -08001421 /*
1422 * To save our caller's stack, now use input list for pages to free.
1423 */
1424 list_splice(&pages_to_free, page_list);
Mel Gorman66635622010-08-09 17:19:30 -07001425}
1426
1427/*
Andrew Morton1742f192006-03-22 00:08:21 -08001428 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1429 * of reclaimed pages
Linus Torvalds1da177e2005-04-16 15:20:36 -07001430 */
Mel Gorman66635622010-08-09 17:19:30 -07001431static noinline_for_stack unsigned long
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001432shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001433 struct scan_control *sc, enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001434{
1435 LIST_HEAD(page_list);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001436 unsigned long nr_scanned;
Andrew Morton05ff5132006-03-22 00:08:20 -08001437 unsigned long nr_reclaimed = 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001438 unsigned long nr_taken;
Mel Gorman8e950282013-07-03 15:02:02 -07001439 unsigned long nr_dirty = 0;
1440 unsigned long nr_congested = 0;
Mel Gormane2be15f2013-07-03 15:01:57 -07001441 unsigned long nr_unqueued_dirty = 0;
Mel Gorman92df3a72011-10-31 17:07:56 -07001442 unsigned long nr_writeback = 0;
Mel Gormanb1a6f212013-07-03 15:01:58 -07001443 unsigned long nr_immediate = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001444 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001445 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001446 struct zone *zone = lruvec_zone(lruvec);
1447 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro78dc5832009-06-16 15:31:40 -07001448
Rik van Riel35cd7812009-09-21 17:01:38 -07001449 while (unlikely(too_many_isolated(zone, file, sc))) {
KOSAKI Motohiro58355c72009-10-26 16:49:35 -07001450 congestion_wait(BLK_RW_ASYNC, HZ/10);
Rik van Riel35cd7812009-09-21 17:01:38 -07001451
1452 /* We are about to die and free our memory. Return now. */
1453 if (fatal_signal_pending(current))
1454 return SWAP_CLUSTER_MAX;
1455 }
1456
Linus Torvalds1da177e2005-04-16 15:20:36 -07001457 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001458
1459 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001460 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001461 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001462 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001463
Linus Torvalds1da177e2005-04-16 15:20:36 -07001464 spin_lock_irq(&zone->lru_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001465
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001466 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1467 &nr_scanned, sc, isolate_mode, lru);
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001468
1469 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1470 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1471
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001472 if (global_reclaim(sc)) {
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001473 zone->pages_scanned += nr_scanned;
KOSAKI Motohirob35ea172009-09-21 17:01:36 -07001474 if (current_is_kswapd())
Hugh Dickins75b00af2012-05-29 15:07:09 -07001475 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001476 else
Hugh Dickins75b00af2012-05-29 15:07:09 -07001477 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001478 }
Hillf Dantond563c052012-03-21 16:34:02 -07001479 spin_unlock_irq(&zone->lru_lock);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001480
Hillf Dantond563c052012-03-21 16:34:02 -07001481 if (nr_taken == 0)
Mel Gorman66635622010-08-09 17:19:30 -07001482 return 0;
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001483
Minchan Kim02c6de82012-10-08 16:31:55 -07001484 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
Mel Gorman8e950282013-07-03 15:02:02 -07001485 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1486 &nr_writeback, &nr_immediate,
1487 false);
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001488
Hugh Dickins3f797682012-01-12 17:20:07 -08001489 spin_lock_irq(&zone->lru_lock);
1490
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001491 reclaim_stat->recent_scanned[file] += nr_taken;
Hillf Dantond563c052012-03-21 16:34:02 -07001492
Ying Han904249a2012-04-25 16:01:48 -07001493 if (global_reclaim(sc)) {
1494 if (current_is_kswapd())
1495 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1496 nr_reclaimed);
1497 else
1498 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1499 nr_reclaimed);
1500 }
KOSAKI Motohiroe247dbc2010-08-09 17:19:28 -07001501
Konstantin Khlebnikov27ac81d2012-05-29 15:07:00 -07001502 putback_inactive_pages(lruvec, &page_list);
Hugh Dickins3f797682012-01-12 17:20:07 -08001503
Konstantin Khlebnikov95d918f2012-05-29 15:06:59 -07001504 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Hugh Dickins3f797682012-01-12 17:20:07 -08001505
1506 spin_unlock_irq(&zone->lru_lock);
1507
1508 free_hot_cold_page_list(&page_list, 1);
Mel Gormane11da5b2010-10-26 14:21:40 -07001509
Mel Gorman92df3a72011-10-31 17:07:56 -07001510 /*
1511 * If reclaim is isolating dirty pages under writeback, it implies
1512 * that the long-lived page allocation rate is exceeding the page
1513 * laundering rate. Either the global limits are not being effective
1514 * at throttling processes due to the page distribution throughout
1515 * zones or there is heavy usage of a slow backing device. The
1516 * only option is to throttle from reclaim context which is not ideal
1517 * as there is no guarantee the dirtying process is throttled in the
1518 * same way balance_dirty_pages() manages.
1519 *
Mel Gorman8e950282013-07-03 15:02:02 -07001520 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1521 * of pages under pages flagged for immediate reclaim and stall if any
1522 * are encountered in the nr_immediate check below.
Mel Gorman92df3a72011-10-31 17:07:56 -07001523 */
Mel Gorman918fc712013-07-08 16:00:25 -07001524 if (nr_writeback && nr_writeback == nr_taken)
Mel Gorman283aba92013-07-03 15:01:51 -07001525 zone_set_flag(zone, ZONE_WRITEBACK);
Mel Gorman92df3a72011-10-31 17:07:56 -07001526
Mel Gormand43006d2013-07-03 15:01:50 -07001527 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001528 * memcg will stall in page writeback so only consider forcibly
1529 * stalling for global reclaim
Mel Gormand43006d2013-07-03 15:01:50 -07001530 */
Mel Gormanb1a6f212013-07-03 15:01:58 -07001531 if (global_reclaim(sc)) {
1532 /*
Mel Gorman8e950282013-07-03 15:02:02 -07001533 * Tag a zone as congested if all the dirty pages scanned were
1534 * backed by a congested BDI and wait_iff_congested will stall.
1535 */
1536 if (nr_dirty && nr_dirty == nr_congested)
1537 zone_set_flag(zone, ZONE_CONGESTED);
1538
1539 /*
Mel Gormanb1a6f212013-07-03 15:01:58 -07001540 * If dirty pages are scanned that are not queued for IO, it
1541 * implies that flushers are not keeping up. In this case, flag
1542 * the zone ZONE_TAIL_LRU_DIRTY and kswapd will start writing
1543 * pages from reclaim context. It will forcibly stall in the
1544 * next check.
1545 */
1546 if (nr_unqueued_dirty == nr_taken)
1547 zone_set_flag(zone, ZONE_TAIL_LRU_DIRTY);
1548
1549 /*
1550 * In addition, if kswapd scans pages marked marked for
1551 * immediate reclaim and under writeback (nr_immediate), it
1552 * implies that pages are cycling through the LRU faster than
1553 * they are written so also forcibly stall.
1554 */
1555 if (nr_unqueued_dirty == nr_taken || nr_immediate)
1556 congestion_wait(BLK_RW_ASYNC, HZ/10);
Mel Gormane2be15f2013-07-03 15:01:57 -07001557 }
Mel Gormand43006d2013-07-03 15:01:50 -07001558
Mel Gorman8e950282013-07-03 15:02:02 -07001559 /*
1560 * Stall direct reclaim for IO completions if underlying BDIs or zone
1561 * is congested. Allow kswapd to continue until it starts encountering
1562 * unqueued dirty pages or cycling through the LRU too quickly.
1563 */
1564 if (!sc->hibernation_mode && !current_is_kswapd())
1565 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1566
Mel Gormane11da5b2010-10-26 14:21:40 -07001567 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1568 zone_idx(zone),
1569 nr_scanned, nr_reclaimed,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001570 sc->priority,
Mel Gorman23b9da52012-05-29 15:06:20 -07001571 trace_shrink_flags(file));
Andrew Morton05ff5132006-03-22 00:08:20 -08001572 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001573}
1574
Martin Bligh3bb1a8522006-10-28 10:38:24 -07001575/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001576 * This moves pages from the active list to the inactive list.
1577 *
1578 * We move them the other way if the page is referenced by one or more
1579 * processes, from rmap.
1580 *
1581 * If the pages are mostly unmapped, the processing is fast and it is
1582 * appropriate to hold zone->lru_lock across the whole operation. But if
1583 * the pages are mapped, the processing is slow (page_referenced()) so we
1584 * should drop zone->lru_lock around each page. It's impossible to balance
1585 * this, so instead we remove the pages from the LRU while processing them.
1586 * It is safe to rely on PG_active against the non-LRU pages in here because
1587 * nobody will play with that bit on a non-LRU page.
1588 *
1589 * The downside is that we have to touch page->_count against each page.
1590 * But we had to alter page->flags anyway.
1591 */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001592
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001593static void move_active_pages_to_lru(struct lruvec *lruvec,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001594 struct list_head *list,
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001595 struct list_head *pages_to_free,
Wu Fengguang3eb41402009-06-16 15:33:13 -07001596 enum lru_list lru)
1597{
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001598 struct zone *zone = lruvec_zone(lruvec);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001599 unsigned long pgmoved = 0;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001600 struct page *page;
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001601 int nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001602
Wu Fengguang3eb41402009-06-16 15:33:13 -07001603 while (!list_empty(list)) {
1604 page = lru_to_page(list);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001605 lruvec = mem_cgroup_page_lruvec(page, zone);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001606
Sasha Levin309381fea2014-01-23 15:52:54 -08001607 VM_BUG_ON_PAGE(PageLRU(page), page);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001608 SetPageLRU(page);
1609
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001610 nr_pages = hpage_nr_pages(page);
1611 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
Johannes Weiner925b7672012-01-12 17:18:15 -08001612 list_move(&page->lru, &lruvec->lists[lru]);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001613 pgmoved += nr_pages;
Wu Fengguang3eb41402009-06-16 15:33:13 -07001614
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001615 if (put_page_testzero(page)) {
1616 __ClearPageLRU(page);
1617 __ClearPageActive(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001618 del_page_from_lru_list(page, lruvec, lru);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001619
1620 if (unlikely(PageCompound(page))) {
1621 spin_unlock_irq(&zone->lru_lock);
1622 (*get_compound_page_dtor(page))(page);
1623 spin_lock_irq(&zone->lru_lock);
1624 } else
1625 list_add(&page->lru, pages_to_free);
Wu Fengguang3eb41402009-06-16 15:33:13 -07001626 }
1627 }
1628 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1629 if (!is_active_lru(lru))
1630 __count_vm_events(PGDEACTIVATE, pgmoved);
1631}
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001632
Hugh Dickinsf6260122012-01-12 17:20:06 -08001633static void shrink_active_list(unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001634 struct lruvec *lruvec,
Johannes Weinerf16015f2012-01-12 17:17:52 -08001635 struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001636 enum lru_list lru)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001637{
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001638 unsigned long nr_taken;
Hugh Dickinsf6260122012-01-12 17:20:06 -08001639 unsigned long nr_scanned;
Wu Fengguang6fe6b7e2009-06-16 15:33:05 -07001640 unsigned long vm_flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001641 LIST_HEAD(l_hold); /* The pages which were snipped off */
Wu Fengguang8cab4752009-06-16 15:33:12 -07001642 LIST_HEAD(l_active);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001643 LIST_HEAD(l_inactive);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001644 struct page *page;
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001645 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
KOSAKI Motohiro44c241f2009-09-21 17:01:35 -07001646 unsigned long nr_rotated = 0;
Konstantin Khlebnikovf3fd4a62012-05-29 15:06:54 -07001647 isolate_mode_t isolate_mode = 0;
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001648 int file = is_file_lru(lru);
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001649 struct zone *zone = lruvec_zone(lruvec);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001650
1651 lru_add_drain();
Minchan Kimf80c0672011-10-31 17:06:55 -07001652
1653 if (!sc->may_unmap)
Hillf Danton61317282012-03-21 16:33:48 -07001654 isolate_mode |= ISOLATE_UNMAPPED;
Minchan Kimf80c0672011-10-31 17:06:55 -07001655 if (!sc->may_writepage)
Hillf Danton61317282012-03-21 16:33:48 -07001656 isolate_mode |= ISOLATE_CLEAN;
Minchan Kimf80c0672011-10-31 17:06:55 -07001657
Linus Torvalds1da177e2005-04-16 15:20:36 -07001658 spin_lock_irq(&zone->lru_lock);
Johannes Weiner925b7672012-01-12 17:18:15 -08001659
Konstantin Khlebnikov5dc35972012-05-29 15:06:58 -07001660 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1661 &nr_scanned, sc, isolate_mode, lru);
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001662 if (global_reclaim(sc))
Hugh Dickinsf6260122012-01-12 17:20:06 -08001663 zone->pages_scanned += nr_scanned;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001664
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001665 reclaim_stat->recent_scanned[file] += nr_taken;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08001666
Hugh Dickinsf6260122012-01-12 17:20:06 -08001667 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
Konstantin Khlebnikov3cb99452012-05-29 15:06:53 -07001668 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001669 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001670 spin_unlock_irq(&zone->lru_lock);
1671
Linus Torvalds1da177e2005-04-16 15:20:36 -07001672 while (!list_empty(&l_hold)) {
1673 cond_resched();
1674 page = lru_to_page(&l_hold);
1675 list_del(&page->lru);
Rik van Riel7e9cd482008-10-18 20:26:35 -07001676
Hugh Dickins39b5f292012-10-08 16:33:18 -07001677 if (unlikely(!page_evictable(page))) {
Lee Schermerhorn894bc312008-10-18 20:26:39 -07001678 putback_lru_page(page);
1679 continue;
1680 }
1681
Mel Gormancc715d92012-03-21 16:34:00 -07001682 if (unlikely(buffer_heads_over_limit)) {
1683 if (page_has_private(page) && trylock_page(page)) {
1684 if (page_has_private(page))
1685 try_to_release_page(page, 0);
1686 unlock_page(page);
1687 }
1688 }
1689
Johannes Weinerc3ac9a82012-05-29 15:06:25 -07001690 if (page_referenced(page, 0, sc->target_mem_cgroup,
1691 &vm_flags)) {
Rik van Riel9992af12011-01-13 15:47:13 -08001692 nr_rotated += hpage_nr_pages(page);
Wu Fengguang8cab4752009-06-16 15:33:12 -07001693 /*
1694 * Identify referenced, file-backed active pages and
1695 * give them one more trip around the active list. So
1696 * that executable code get better chances to stay in
1697 * memory under moderate memory pressure. Anon pages
1698 * are not likely to be evicted by use-once streaming
1699 * IO, plus JVM can create lots of anon VM_EXEC pages,
1700 * so we ignore them here.
1701 */
Wu Fengguang41e20982009-10-26 16:49:53 -07001702 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
Wu Fengguang8cab4752009-06-16 15:33:12 -07001703 list_add(&page->lru, &l_active);
1704 continue;
1705 }
1706 }
Rik van Riel7e9cd482008-10-18 20:26:35 -07001707
KOSAKI Motohiro5205e562009-09-21 17:01:44 -07001708 ClearPageActive(page); /* we are de-activating */
Linus Torvalds1da177e2005-04-16 15:20:36 -07001709 list_add(&page->lru, &l_inactive);
1710 }
1711
Andrew Mortonb5557492009-01-06 14:40:13 -08001712 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001713 * Move pages back to the lru list.
Andrew Mortonb5557492009-01-06 14:40:13 -08001714 */
Johannes Weiner2a1dc502008-12-01 03:00:35 +01001715 spin_lock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001716 /*
Wu Fengguang8cab4752009-06-16 15:33:12 -07001717 * Count referenced pages from currently used mappings as rotated,
1718 * even though only some of them are actually re-activated. This
1719 * helps balance scan pressure between file and anonymous pages in
1720 * get_scan_ratio.
Rik van Riel7e9cd482008-10-18 20:26:35 -07001721 */
Johannes Weinerb7c46d12009-09-21 17:02:56 -07001722 reclaim_stat->recent_rotated[file] += nr_rotated;
Rik van Riel556adec2008-10-18 20:26:34 -07001723
Hugh Dickinsfa9add62012-05-29 15:07:09 -07001724 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1725 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
KOSAKI Motohiroa7312862009-09-21 17:01:37 -07001726 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
Christoph Lameterf8891e52006-06-30 01:55:45 -07001727 spin_unlock_irq(&zone->lru_lock);
Hugh Dickins2bcf8872012-01-12 17:19:56 -08001728
1729 free_hot_cold_page_list(&l_hold, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001730}
1731
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001732#ifdef CONFIG_SWAP
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001733static int inactive_anon_is_low_global(struct zone *zone)
KOSAKI Motohirof89eb902009-01-07 18:08:14 -08001734{
1735 unsigned long active, inactive;
1736
1737 active = zone_page_state(zone, NR_ACTIVE_ANON);
1738 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1739
1740 if (inactive * zone->inactive_ratio < active)
1741 return 1;
1742
1743 return 0;
1744}
1745
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001746/**
1747 * inactive_anon_is_low - check if anonymous pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001748 * @lruvec: LRU vector to check
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001749 *
1750 * Returns true if the zone does not have enough inactive anon pages,
1751 * meaning some active anon pages need to be deactivated.
1752 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001753static int inactive_anon_is_low(struct lruvec *lruvec)
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001754{
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001755 /*
1756 * If we don't have swap space, anonymous page deactivation
1757 * is pointless.
1758 */
1759 if (!total_swap_pages)
1760 return 0;
1761
Hugh Dickinsc3c787e2012-05-29 15:06:52 -07001762 if (!mem_cgroup_disabled())
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001763 return mem_cgroup_inactive_anon_is_low(lruvec);
Johannes Weinerf16015f2012-01-12 17:17:52 -08001764
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001765 return inactive_anon_is_low_global(lruvec_zone(lruvec));
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001766}
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001767#else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001768static inline int inactive_anon_is_low(struct lruvec *lruvec)
Minchan Kim74e3f3c2010-10-26 14:21:31 -07001769{
1770 return 0;
1771}
1772#endif
KOSAKI Motohiro14797e22009-01-07 18:08:18 -08001773
Rik van Riel56e49d22009-06-16 15:32:28 -07001774/**
1775 * inactive_file_is_low - check if file pages need to be deactivated
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001776 * @lruvec: LRU vector to check
Rik van Riel56e49d22009-06-16 15:32:28 -07001777 *
1778 * When the system is doing streaming IO, memory pressure here
1779 * ensures that active file pages get deactivated, until more
1780 * than half of the file pages are on the inactive list.
1781 *
1782 * Once we get to that situation, protect the system's working
1783 * set from being evicted by disabling active file page aging.
1784 *
1785 * This uses a different ratio than the anonymous pages, because
1786 * the page cache uses a use-once replacement algorithm.
1787 */
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001788static int inactive_file_is_low(struct lruvec *lruvec)
Rik van Riel56e49d22009-06-16 15:32:28 -07001789{
Johannes Weinere3790142013-02-22 16:35:19 -08001790 unsigned long inactive;
1791 unsigned long active;
Rik van Riel56e49d22009-06-16 15:32:28 -07001792
Johannes Weinere3790142013-02-22 16:35:19 -08001793 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1794 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
1795
1796 return active > inactive;
Rik van Riel56e49d22009-06-16 15:32:28 -07001797}
1798
Hugh Dickins75b00af2012-05-29 15:07:09 -07001799static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
Rik van Rielb39415b2009-12-14 17:59:48 -08001800{
Hugh Dickins75b00af2012-05-29 15:07:09 -07001801 if (is_file_lru(lru))
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001802 return inactive_file_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001803 else
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07001804 return inactive_anon_is_low(lruvec);
Rik van Rielb39415b2009-12-14 17:59:48 -08001805}
1806
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001807static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001808 struct lruvec *lruvec, struct scan_control *sc)
Christoph Lameterb69408e2008-10-18 20:26:14 -07001809{
Rik van Rielb39415b2009-12-14 17:59:48 -08001810 if (is_active_lru(lru)) {
Hugh Dickins75b00af2012-05-29 15:07:09 -07001811 if (inactive_list_is_low(lruvec, lru))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001812 shrink_active_list(nr_to_scan, lruvec, sc, lru);
Rik van Riel556adec2008-10-18 20:26:34 -07001813 return 0;
1814 }
1815
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07001816 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
Christoph Lameterb69408e2008-10-18 20:26:14 -07001817}
1818
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001819static int vmscan_swappiness(struct scan_control *sc)
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001820{
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001821 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001822 return vm_swappiness;
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001823 return mem_cgroup_swappiness(sc->target_mem_cgroup);
KAMEZAWA Hiroyuki1f4c0252011-07-26 16:08:21 -07001824}
1825
Johannes Weiner9a265112013-02-22 16:32:17 -08001826enum scan_balance {
1827 SCAN_EQUAL,
1828 SCAN_FRACT,
1829 SCAN_ANON,
1830 SCAN_FILE,
1831};
1832
Linus Torvalds1da177e2005-04-16 15:20:36 -07001833/*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001834 * Determine how aggressively the anon and file LRU lists should be
1835 * scanned. The relative value of each set of LRU lists is determined
1836 * by looking at the fraction of the pages scanned we did rotate back
1837 * onto the active list instead of evict.
1838 *
Wanpeng Libe7bd592012-06-14 20:41:02 +08001839 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1840 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001841 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001842static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07001843 unsigned long *nr)
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001844{
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001845 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
Johannes Weiner9a265112013-02-22 16:32:17 -08001846 u64 fraction[2];
1847 u64 denominator = 0; /* gcc */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001848 struct zone *zone = lruvec_zone(lruvec);
Johannes Weiner9a265112013-02-22 16:32:17 -08001849 unsigned long anon_prio, file_prio;
1850 enum scan_balance scan_balance;
1851 unsigned long anon, file, free;
1852 bool force_scan = false;
1853 unsigned long ap, fp;
1854 enum lru_list lru;
KAMEZAWA Hiroyuki246e87a2011-05-26 16:25:34 -07001855
Johannes Weinerf11c0ca2011-10-31 17:07:27 -07001856 /*
1857 * If the zone or memcg is small, nr[l] can be 0. This
1858 * results in no scanning on this priority and a potential
1859 * priority drop. Global direct reclaim can go to the next
1860 * zone and tends to have no problems. Global kswapd is for
1861 * zone balancing and it needs to scan a minimum amount. When
1862 * reclaiming for a memcg, a priority drop can cause high
1863 * latencies, so it's better to scan a minimum amount there as
1864 * well.
1865 */
Lisa Du6e543d52013-09-11 14:22:36 -07001866 if (current_is_kswapd() && !zone_reclaimable(zone))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001867 force_scan = true;
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001868 if (!global_reclaim(sc))
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001869 force_scan = true;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001870
1871 /* If we have no swap space, do not bother scanning anon pages. */
Shaohua Liec8acf22013-02-22 16:34:38 -08001872 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001873 scan_balance = SCAN_FILE;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001874 goto out;
1875 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001876
Johannes Weiner10316b32013-02-22 16:32:14 -08001877 /*
1878 * Global reclaim will swap to prevent OOM even with no
1879 * swappiness, but memcg users want to use this knob to
1880 * disable swapping for individual groups completely when
1881 * using the memory controller's swap limit feature would be
1882 * too expensive.
1883 */
1884 if (!global_reclaim(sc) && !vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001885 scan_balance = SCAN_FILE;
Johannes Weiner10316b32013-02-22 16:32:14 -08001886 goto out;
1887 }
1888
1889 /*
1890 * Do not apply any pressure balancing cleverness when the
1891 * system is close to OOM, scan both anon and file equally
1892 * (unless the swappiness setting disagrees with swapping).
1893 */
1894 if (!sc->priority && vmscan_swappiness(sc)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001895 scan_balance = SCAN_EQUAL;
Johannes Weiner10316b32013-02-22 16:32:14 -08001896 goto out;
1897 }
1898
Hugh Dickins4d7dcca2012-05-29 15:07:08 -07001899 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
1900 get_lru_size(lruvec, LRU_INACTIVE_ANON);
1901 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
1902 get_lru_size(lruvec, LRU_INACTIVE_FILE);
Johannes Weinera4d3e9e2011-09-14 16:21:52 -07001903
Johannes Weiner11d16c22013-02-22 16:32:15 -08001904 /*
1905 * If it's foreseeable that reclaiming the file cache won't be
1906 * enough to get the zone back into a desirable shape, we have
1907 * to swap. Better start now and leave the - probably heavily
1908 * thrashing - remaining file pages alone.
1909 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08001910 if (global_reclaim(sc)) {
Johannes Weiner11d16c22013-02-22 16:32:15 -08001911 free = zone_page_state(zone, NR_FREE_PAGES);
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001912 if (unlikely(file + free <= high_wmark_pages(zone))) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001913 scan_balance = SCAN_ANON;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001914 goto out;
KOSAKI Motohiroeeee9a82009-01-07 18:08:17 -08001915 }
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001916 }
1917
1918 /*
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001919 * There is enough inactive page cache, do not reclaim
1920 * anything from the anonymous working set right now.
1921 */
1922 if (!inactive_file_is_low(lruvec)) {
Johannes Weiner9a265112013-02-22 16:32:17 -08001923 scan_balance = SCAN_FILE;
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001924 goto out;
1925 }
1926
Johannes Weiner9a265112013-02-22 16:32:17 -08001927 scan_balance = SCAN_FRACT;
1928
Johannes Weiner7c5bd702013-02-22 16:32:10 -08001929 /*
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001930 * With swappiness at 100, anonymous and file have the same priority.
1931 * This scanning priority is essentially the inverse of IO cost.
1932 */
Konstantin Khlebnikov3d58ab52012-05-29 15:06:57 -07001933 anon_prio = vmscan_swappiness(sc);
Hugh Dickins75b00af2012-05-29 15:07:09 -07001934 file_prio = 200 - anon_prio;
KOSAKI Motohiro58c37f62010-08-09 17:19:51 -07001935
1936 /*
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001937 * OK, so we have swap space and a fair amount of page cache
1938 * pages. We use the recently rotated / recently scanned
1939 * ratios to determine how valuable each cache is.
1940 *
1941 * Because workloads change over time (and to avoid overflow)
1942 * we keep these statistics as a floating average, which ends
1943 * up weighing recent references more than old ones.
1944 *
1945 * anon in [0], file in [1]
1946 */
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001947 spin_lock_irq(&zone->lru_lock);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001948 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001949 reclaim_stat->recent_scanned[0] /= 2;
1950 reclaim_stat->recent_rotated[0] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001951 }
1952
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001953 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001954 reclaim_stat->recent_scanned[1] /= 2;
1955 reclaim_stat->recent_rotated[1] /= 2;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001956 }
1957
1958 /*
Rik van Riel00d80892008-11-19 15:36:44 -08001959 * The amount of pressure on anon vs file pages is inversely
1960 * proportional to the fraction of recently scanned pages on
1961 * each list that were recently referenced and in active use.
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001962 */
Satoru Moriyafe350042012-05-29 15:06:47 -07001963 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001964 ap /= reclaim_stat->recent_rotated[0] + 1;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001965
Satoru Moriyafe350042012-05-29 15:06:47 -07001966 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
KOSAKI Motohiro6e901572009-01-07 18:08:15 -08001967 fp /= reclaim_stat->recent_rotated[1] + 1;
Konstantin Khlebnikov90126372012-05-29 15:07:01 -07001968 spin_unlock_irq(&zone->lru_lock);
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001969
Shaohua Li76a33fc2010-05-24 14:32:36 -07001970 fraction[0] = ap;
1971 fraction[1] = fp;
1972 denominator = ap + fp + 1;
1973out:
Hugh Dickins41113042012-01-12 17:20:01 -08001974 for_each_evictable_lru(lru) {
1975 int file = is_file_lru(lru);
Johannes Weinerd778df52013-02-22 16:32:12 -08001976 unsigned long size;
Shaohua Li76a33fc2010-05-24 14:32:36 -07001977 unsigned long scan;
Rik van Riel4f98a2f2008-10-18 20:26:32 -07001978
Johannes Weinerd778df52013-02-22 16:32:12 -08001979 size = get_lru_size(lruvec, lru);
Johannes Weiner10316b32013-02-22 16:32:14 -08001980 scan = size >> sc->priority;
Johannes Weiner9a265112013-02-22 16:32:17 -08001981
Johannes Weiner10316b32013-02-22 16:32:14 -08001982 if (!scan && force_scan)
1983 scan = min(size, SWAP_CLUSTER_MAX);
Johannes Weiner9a265112013-02-22 16:32:17 -08001984
1985 switch (scan_balance) {
1986 case SCAN_EQUAL:
1987 /* Scan lists relative to size */
1988 break;
1989 case SCAN_FRACT:
1990 /*
1991 * Scan types proportional to swappiness and
1992 * their relative recent reclaim efficiency.
1993 */
1994 scan = div64_u64(scan * fraction[file], denominator);
1995 break;
1996 case SCAN_FILE:
1997 case SCAN_ANON:
1998 /* Scan one type exclusively */
1999 if ((scan_balance == SCAN_FILE) != file)
2000 scan = 0;
2001 break;
2002 default:
2003 /* Look ma, no brain */
2004 BUG();
2005 }
Hugh Dickins41113042012-01-12 17:20:01 -08002006 nr[lru] = scan;
Shaohua Li76a33fc2010-05-24 14:32:36 -07002007 }
Wu Fengguang6e08a362009-06-16 15:32:29 -07002008}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002009
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002010/*
2011 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2012 */
2013static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc)
2014{
2015 unsigned long nr[NR_LRU_LISTS];
Mel Gormane82e0562013-07-03 15:01:44 -07002016 unsigned long targets[NR_LRU_LISTS];
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002017 unsigned long nr_to_scan;
2018 enum lru_list lru;
2019 unsigned long nr_reclaimed = 0;
2020 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2021 struct blk_plug plug;
Mel Gormane82e0562013-07-03 15:01:44 -07002022 bool scan_adjusted = false;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002023
2024 get_scan_count(lruvec, sc, nr);
2025
Mel Gormane82e0562013-07-03 15:01:44 -07002026 /* Record the original scan target for proportional adjustments later */
2027 memcpy(targets, nr, sizeof(nr));
2028
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002029 blk_start_plug(&plug);
2030 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2031 nr[LRU_INACTIVE_FILE]) {
Mel Gormane82e0562013-07-03 15:01:44 -07002032 unsigned long nr_anon, nr_file, percentage;
2033 unsigned long nr_scanned;
2034
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002035 for_each_evictable_lru(lru) {
2036 if (nr[lru]) {
2037 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2038 nr[lru] -= nr_to_scan;
2039
2040 nr_reclaimed += shrink_list(lru, nr_to_scan,
2041 lruvec, sc);
2042 }
2043 }
Mel Gormane82e0562013-07-03 15:01:44 -07002044
2045 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2046 continue;
2047
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002048 /*
Mel Gormane82e0562013-07-03 15:01:44 -07002049 * For global direct reclaim, reclaim only the number of pages
2050 * requested. Less care is taken to scan proportionally as it
2051 * is more important to minimise direct reclaim stall latency
2052 * than it is to properly age the LRU lists.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002053 */
Mel Gormane82e0562013-07-03 15:01:44 -07002054 if (global_reclaim(sc) && !current_is_kswapd())
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002055 break;
Mel Gormane82e0562013-07-03 15:01:44 -07002056
2057 /*
2058 * For kswapd and memcg, reclaim at least the number of pages
2059 * requested. Ensure that the anon and file LRUs shrink
2060 * proportionally what was requested by get_scan_count(). We
2061 * stop reclaiming one LRU and reduce the amount scanning
2062 * proportional to the original scan target.
2063 */
2064 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2065 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2066
2067 if (nr_file > nr_anon) {
2068 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2069 targets[LRU_ACTIVE_ANON] + 1;
2070 lru = LRU_BASE;
2071 percentage = nr_anon * 100 / scan_target;
2072 } else {
2073 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2074 targets[LRU_ACTIVE_FILE] + 1;
2075 lru = LRU_FILE;
2076 percentage = nr_file * 100 / scan_target;
2077 }
2078
2079 /* Stop scanning the smaller of the LRU */
2080 nr[lru] = 0;
2081 nr[lru + LRU_ACTIVE] = 0;
2082
2083 /*
2084 * Recalculate the other LRU scan count based on its original
2085 * scan target and the percentage scanning already complete
2086 */
2087 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2088 nr_scanned = targets[lru] - nr[lru];
2089 nr[lru] = targets[lru] * (100 - percentage) / 100;
2090 nr[lru] -= min(nr[lru], nr_scanned);
2091
2092 lru += LRU_ACTIVE;
2093 nr_scanned = targets[lru] - nr[lru];
2094 nr[lru] = targets[lru] * (100 - percentage) / 100;
2095 nr[lru] -= min(nr[lru], nr_scanned);
2096
2097 scan_adjusted = true;
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002098 }
2099 blk_finish_plug(&plug);
2100 sc->nr_reclaimed += nr_reclaimed;
2101
2102 /*
2103 * Even if we did not try to evict anon pages at all, we want to
2104 * rebalance the anon lru active/inactive ratio.
2105 */
2106 if (inactive_anon_is_low(lruvec))
2107 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2108 sc, LRU_ACTIVE_ANON);
2109
2110 throttle_vm_writeout(sc->gfp_mask);
2111}
2112
Mel Gorman23b9da52012-05-29 15:06:20 -07002113/* Use reclaim/compaction for costly allocs or under memory pressure */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002114static bool in_reclaim_compaction(struct scan_control *sc)
Mel Gorman23b9da52012-05-29 15:06:20 -07002115{
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002116 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
Mel Gorman23b9da52012-05-29 15:06:20 -07002117 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002118 sc->priority < DEF_PRIORITY - 2))
Mel Gorman23b9da52012-05-29 15:06:20 -07002119 return true;
2120
2121 return false;
2122}
2123
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002124/*
Mel Gorman23b9da52012-05-29 15:06:20 -07002125 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2126 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2127 * true if more pages should be reclaimed such that when the page allocator
2128 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2129 * It will give up earlier than that if there is difficulty reclaiming pages.
Mel Gorman3e7d3442011-01-13 15:45:56 -08002130 */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002131static inline bool should_continue_reclaim(struct zone *zone,
Mel Gorman3e7d3442011-01-13 15:45:56 -08002132 unsigned long nr_reclaimed,
2133 unsigned long nr_scanned,
2134 struct scan_control *sc)
2135{
2136 unsigned long pages_for_compaction;
2137 unsigned long inactive_lru_pages;
2138
2139 /* If not in reclaim/compaction mode, stop */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002140 if (!in_reclaim_compaction(sc))
Mel Gorman3e7d3442011-01-13 15:45:56 -08002141 return false;
2142
Mel Gorman28765922011-02-25 14:44:20 -08002143 /* Consider stopping depending on scan and reclaim activity */
2144 if (sc->gfp_mask & __GFP_REPEAT) {
2145 /*
2146 * For __GFP_REPEAT allocations, stop reclaiming if the
2147 * full LRU list has been scanned and we are still failing
2148 * to reclaim pages. This full LRU scan is potentially
2149 * expensive but a __GFP_REPEAT caller really wants to succeed
2150 */
2151 if (!nr_reclaimed && !nr_scanned)
2152 return false;
2153 } else {
2154 /*
2155 * For non-__GFP_REPEAT allocations which can presumably
2156 * fail without consequence, stop if we failed to reclaim
2157 * any pages from the last SWAP_CLUSTER_MAX number of
2158 * pages that were scanned. This will return to the
2159 * caller faster at the risk reclaim/compaction and
2160 * the resulting allocation attempt fails
2161 */
2162 if (!nr_reclaimed)
2163 return false;
2164 }
Mel Gorman3e7d3442011-01-13 15:45:56 -08002165
2166 /*
2167 * If we have not reclaimed enough pages for compaction and the
2168 * inactive lists are large enough, continue reclaiming
2169 */
2170 pages_for_compaction = (2UL << sc->order);
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002171 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
Shaohua Liec8acf22013-02-22 16:34:38 -08002172 if (get_nr_swap_pages() > 0)
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002173 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
Mel Gorman3e7d3442011-01-13 15:45:56 -08002174 if (sc->nr_reclaimed < pages_for_compaction &&
2175 inactive_lru_pages > pages_for_compaction)
2176 return true;
2177
2178 /* If compaction would go ahead or the allocation would succeed, stop */
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002179 switch (compaction_suitable(zone, sc->order)) {
Mel Gorman3e7d3442011-01-13 15:45:56 -08002180 case COMPACT_PARTIAL:
2181 case COMPACT_CONTINUE:
2182 return false;
2183 default:
2184 return true;
2185 }
2186}
2187
Andrew Morton0608f432013-09-24 15:27:41 -07002188static void shrink_zone(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002189{
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002190 unsigned long nr_reclaimed, nr_scanned;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002191
Johannes Weiner56600482012-01-12 17:17:59 -08002192 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002193 struct mem_cgroup *root = sc->target_mem_cgroup;
2194 struct mem_cgroup_reclaim_cookie reclaim = {
2195 .zone = zone,
2196 .priority = sc->priority,
2197 };
Andrew Morton694fbc02013-09-24 15:27:37 -07002198 struct mem_cgroup *memcg;
Johannes Weiner56600482012-01-12 17:17:59 -08002199
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002200 nr_reclaimed = sc->nr_reclaimed;
2201 nr_scanned = sc->nr_scanned;
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002202
Andrew Morton694fbc02013-09-24 15:27:37 -07002203 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2204 do {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002205 struct lruvec *lruvec;
2206
2207 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
2208
2209 shrink_lruvec(lruvec, sc);
2210
2211 /*
Michal Hockoa394cb82013-02-22 16:32:30 -08002212 * Direct reclaim and kswapd have to scan all memory
2213 * cgroups to fulfill the overall scan target for the
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002214 * zone.
Michal Hockoa394cb82013-02-22 16:32:30 -08002215 *
2216 * Limit reclaim, on the other hand, only cares about
2217 * nr_to_reclaim pages to be reclaimed and it will
2218 * retry with decreasing priority if one round over the
2219 * whole hierarchy is not sufficient.
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002220 */
Michal Hockoa394cb82013-02-22 16:32:30 -08002221 if (!global_reclaim(sc) &&
2222 sc->nr_reclaimed >= sc->nr_to_reclaim) {
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002223 mem_cgroup_iter_break(root, memcg);
2224 break;
2225 }
Andrew Morton694fbc02013-09-24 15:27:37 -07002226 memcg = mem_cgroup_iter(root, memcg, &reclaim);
2227 } while (memcg);
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002228
2229 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2230 sc->nr_scanned - nr_scanned,
2231 sc->nr_reclaimed - nr_reclaimed);
2232
Johannes Weiner9b4f98c2013-02-22 16:32:19 -08002233 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2234 sc->nr_scanned - nr_scanned, sc));
Johannes Weinerf16015f2012-01-12 17:17:52 -08002235}
2236
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002237/* Returns true if compaction should go ahead for a high-order request */
2238static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
2239{
2240 unsigned long balance_gap, watermark;
2241 bool watermark_ok;
2242
2243 /* Do not consider compaction for orders reclaim is meant to satisfy */
2244 if (sc->order <= PAGE_ALLOC_COSTLY_ORDER)
2245 return false;
2246
2247 /*
2248 * Compaction takes time to run and there are potentially other
2249 * callers using the pages just freed. Continue reclaiming until
2250 * there is a buffer of free pages available to give compaction
2251 * a reasonable chance of completing and allocating the page
2252 */
2253 balance_gap = min(low_wmark_pages(zone),
Jiang Liub40da042013-02-22 16:33:52 -08002254 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002255 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2256 watermark = high_wmark_pages(zone) + balance_gap + (2UL << sc->order);
2257 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2258
2259 /*
2260 * If compaction is deferred, reclaim up to a point where
2261 * compaction will have a chance of success when re-enabled
2262 */
Rik van Rielaff62242012-03-21 16:33:52 -07002263 if (compaction_deferred(zone, sc->order))
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002264 return watermark_ok;
2265
2266 /* If compaction is not ready to start, keep reclaiming */
2267 if (!compaction_suitable(zone, sc->order))
2268 return false;
2269
2270 return watermark_ok;
2271}
2272
Linus Torvalds1da177e2005-04-16 15:20:36 -07002273/*
2274 * This is the direct reclaim path, for page-allocating processes. We only
2275 * try to reclaim pages from zones which will satisfy the caller's allocation
2276 * request.
2277 *
Mel Gorman41858962009-06-16 15:32:12 -07002278 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2279 * Because:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002280 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2281 * allocation or
Mel Gorman41858962009-06-16 15:32:12 -07002282 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2283 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2284 * zone defense algorithm.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002285 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002286 * If a zone is deemed to be full of pinned pages then just give it a light
2287 * scan then give up on it.
Mel Gormane0c23272011-10-31 17:09:33 -07002288 *
2289 * This function returns true if a zone is being reclaimed for a costly
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002290 * high-order allocation and compaction is ready to begin. This indicates to
Mel Gorman0cee34f2012-01-12 17:19:49 -08002291 * the caller that it should consider retrying the allocation instead of
2292 * further reclaim.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002294static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002295{
Mel Gormandd1a2392008-04-28 02:12:17 -07002296 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002297 struct zone *zone;
Andrew Morton0608f432013-09-24 15:27:41 -07002298 unsigned long nr_soft_reclaimed;
2299 unsigned long nr_soft_scanned;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002300 bool aborted_reclaim = false;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002301
Mel Gormancc715d92012-03-21 16:34:00 -07002302 /*
2303 * If the number of buffer_heads in the machine exceeds the maximum
2304 * allowed level, force direct reclaim to scan the highmem zone as
2305 * highmem pages could be pinning lowmem pages storing buffer_heads
2306 */
2307 if (buffer_heads_over_limit)
2308 sc->gfp_mask |= __GFP_HIGHMEM;
2309
Mel Gormand4debc62010-08-09 17:19:29 -07002310 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2311 gfp_zone(sc->gfp_mask), sc->nodemask) {
Con Kolivasf3fe6512006-01-06 00:11:15 -08002312 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002313 continue;
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002314 /*
2315 * Take care memory controller reclaiming has small influence
2316 * to global LRU.
2317 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002318 if (global_reclaim(sc)) {
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002319 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2320 continue;
Lisa Du6e543d52013-09-11 14:22:36 -07002321 if (sc->priority != DEF_PRIORITY &&
2322 !zone_reclaimable(zone))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002323 continue; /* Let kswapd poll it */
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002324 if (IS_ENABLED(CONFIG_COMPACTION)) {
Rik van Riele0887c12011-10-31 17:09:31 -07002325 /*
Mel Gormane0c23272011-10-31 17:09:33 -07002326 * If we already have plenty of memory free for
2327 * compaction in this zone, don't free any more.
2328 * Even though compaction is invoked for any
2329 * non-zero order, only frequent costly order
2330 * reclamation is disruptive enough to become a
Copot Alexandruc7cfa372012-03-21 16:34:10 -07002331 * noticeable problem, like transparent huge
2332 * page allocations.
Rik van Riele0887c12011-10-31 17:09:31 -07002333 */
Mel Gormanfe4b1b22012-01-12 17:19:45 -08002334 if (compaction_ready(zone, sc)) {
Mel Gorman0cee34f2012-01-12 17:19:49 -08002335 aborted_reclaim = true;
Rik van Riele0887c12011-10-31 17:09:31 -07002336 continue;
Mel Gormane0c23272011-10-31 17:09:33 -07002337 }
Rik van Riele0887c12011-10-31 17:09:31 -07002338 }
Andrew Morton0608f432013-09-24 15:27:41 -07002339 /*
2340 * This steals pages from memory cgroups over softlimit
2341 * and returns the number of reclaimed pages and
2342 * scanned pages. This works for global memory pressure
2343 * and balancing, not for a memcg's limit.
2344 */
2345 nr_soft_scanned = 0;
2346 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2347 sc->order, sc->gfp_mask,
2348 &nr_soft_scanned);
2349 sc->nr_reclaimed += nr_soft_reclaimed;
2350 sc->nr_scanned += nr_soft_scanned;
KAMEZAWA Hiroyukiac34a1a2011-06-27 16:18:12 -07002351 /* need some check for avoid more shrink_zone() */
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002352 }
Nick Piggin408d8542006-09-25 23:31:27 -07002353
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002354 shrink_zone(zone, sc);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002355 }
Mel Gormane0c23272011-10-31 17:09:33 -07002356
Mel Gorman0cee34f2012-01-12 17:19:49 -08002357 return aborted_reclaim;
Minchan Kimd1908362010-09-22 13:05:01 -07002358}
2359
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002360/* All zones in zonelist are unreclaimable? */
Minchan Kimd1908362010-09-22 13:05:01 -07002361static bool all_unreclaimable(struct zonelist *zonelist,
2362 struct scan_control *sc)
2363{
2364 struct zoneref *z;
2365 struct zone *zone;
Minchan Kimd1908362010-09-22 13:05:01 -07002366
2367 for_each_zone_zonelist_nodemask(zone, z, zonelist,
2368 gfp_zone(sc->gfp_mask), sc->nodemask) {
2369 if (!populated_zone(zone))
2370 continue;
2371 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2372 continue;
Lisa Du6e543d52013-09-11 14:22:36 -07002373 if (zone_reclaimable(zone))
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002374 return false;
Minchan Kimd1908362010-09-22 13:05:01 -07002375 }
2376
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002377 return true;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378}
Rik van Riel4f98a2f2008-10-18 20:26:32 -07002379
Linus Torvalds1da177e2005-04-16 15:20:36 -07002380/*
2381 * This is the main entry point to direct page reclaim.
2382 *
2383 * If a full scan of the inactive list fails to free enough memory then we
2384 * are "out of memory" and something needs to be killed.
2385 *
2386 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2387 * high - the zone may be full of dirty or under-writeback pages, which this
Jens Axboe5b0830c2009-09-23 19:37:09 +02002388 * caller can't do much about. We kick the writeback threads and take explicit
2389 * naps in the hope that some of these pages can be written. But if the
2390 * allocating task holds filesystem locks which prevent writeout this might not
2391 * work, and the allocation attempt will fail.
Nishanth Aravamudana41f24e2008-04-29 00:58:25 -07002392 *
2393 * returns: 0, if no pages reclaimed
2394 * else, the number of pages reclaimed
Linus Torvalds1da177e2005-04-16 15:20:36 -07002395 */
Mel Gormandac1d272008-04-28 02:12:12 -07002396static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
Ying Hana09ed5e2011-05-24 17:12:26 -07002397 struct scan_control *sc,
2398 struct shrink_control *shrink)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002399{
Andrew Morton69e05942006-03-22 00:08:19 -08002400 unsigned long total_scanned = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002401 struct reclaim_state *reclaim_state = current->reclaim_state;
Mel Gormandd1a2392008-04-28 02:12:17 -07002402 struct zoneref *z;
Mel Gorman54a6eb52008-04-28 02:12:16 -07002403 struct zone *zone;
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002404 unsigned long writeback_threshold;
Mel Gorman0cee34f2012-01-12 17:19:49 -08002405 bool aborted_reclaim;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002406
Keika Kobayashi873b4772008-07-25 01:48:52 -07002407 delayacct_freepages_start();
2408
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002409 if (global_reclaim(sc))
KAMEZAWA Hiroyuki1cfb4192008-02-07 00:14:37 -08002410 count_vm_event(ALLOCSTALL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002411
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002412 do {
Anton Vorontsov70ddf632013-04-29 15:08:31 -07002413 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2414 sc->priority);
Balbir Singh66e17072008-02-07 00:13:56 -08002415 sc->nr_scanned = 0;
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002416 aborted_reclaim = shrink_zones(zonelist, sc);
Mel Gormane0c23272011-10-31 17:09:33 -07002417
Balbir Singh66e17072008-02-07 00:13:56 -08002418 /*
Mel Gorman5a1c9cb2013-07-08 16:00:24 -07002419 * Don't shrink slabs when reclaiming memory from over limit
2420 * cgroups but do shrink slab at least once when aborting
2421 * reclaim for compaction to avoid unevenly scanning file/anon
2422 * LRU pages over slab pages.
Balbir Singh66e17072008-02-07 00:13:56 -08002423 */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002424 if (global_reclaim(sc)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002425 unsigned long lru_pages = 0;
Dave Chinner0ce3d742013-08-28 10:18:03 +10002426
2427 nodes_clear(shrink->nodes_to_scan);
Mel Gormand4debc62010-08-09 17:19:29 -07002428 for_each_zone_zonelist(zone, z, zonelist,
2429 gfp_zone(sc->gfp_mask)) {
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002430 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
2431 continue;
2432
2433 lru_pages += zone_reclaimable_pages(zone);
Dave Chinner0ce3d742013-08-28 10:18:03 +10002434 node_set(zone_to_nid(zone),
2435 shrink->nodes_to_scan);
KOSAKI Motohiroc6a8a8c2010-08-09 17:19:14 -07002436 }
2437
Ying Han1495f232011-05-24 17:12:27 -07002438 shrink_slab(shrink, sc->nr_scanned, lru_pages);
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002439 if (reclaim_state) {
Rik van Riela79311c2009-01-06 14:40:01 -08002440 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
KAMEZAWA Hiroyuki91a45472008-02-07 00:14:29 -08002441 reclaim_state->reclaimed_slab = 0;
2442 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443 }
Balbir Singh66e17072008-02-07 00:13:56 -08002444 total_scanned += sc->nr_scanned;
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002445 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002447
2448 /*
Minchan Kim0e50ce32013-02-22 16:35:37 -08002449 * If we're getting trouble reclaiming, start doing
2450 * writepage even in laptop mode.
2451 */
2452 if (sc->priority < DEF_PRIORITY - 2)
2453 sc->may_writepage = 1;
2454
2455 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002456 * Try to write back as many pages as we just scanned. This
2457 * tends to cause slow streaming writers to write data to the
2458 * disk smoothly, at the dirtying rate, which is nice. But
2459 * that's undesirable in laptop mode, where we *want* lumpy
2460 * writeout. So in laptop mode, write out the whole world.
2461 */
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002462 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2463 if (total_scanned > writeback_threshold) {
Curt Wohlgemuth0e175a12011-10-07 21:54:10 -06002464 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2465 WB_REASON_TRY_TO_FREE_PAGES);
Balbir Singh66e17072008-02-07 00:13:56 -08002466 sc->may_writepage = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002467 }
Mel Gorman5a1c9cb2013-07-08 16:00:24 -07002468 } while (--sc->priority >= 0 && !aborted_reclaim);
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002469
Linus Torvalds1da177e2005-04-16 15:20:36 -07002470out:
Keika Kobayashi873b4772008-07-25 01:48:52 -07002471 delayacct_freepages_end();
2472
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002473 if (sc->nr_reclaimed)
2474 return sc->nr_reclaimed;
2475
KOSAKI Motohiro929bea72011-04-14 15:22:12 -07002476 /*
2477 * As hibernation is going on, kswapd is freezed so that it can't mark
2478 * the zone into all_unreclaimable. Thus bypassing all_unreclaimable
2479 * check.
2480 */
2481 if (oom_killer_disabled)
2482 return 0;
2483
Mel Gorman0cee34f2012-01-12 17:19:49 -08002484 /* Aborted reclaim to try compaction? don't OOM, then */
2485 if (aborted_reclaim)
Mel Gorman73350842012-01-12 17:19:33 -08002486 return 1;
2487
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002488 /* top priority shrink_zones still had more to do? don't OOM, then */
Johannes Weiner89b5fae2012-01-12 17:17:50 -08002489 if (global_reclaim(sc) && !all_unreclaimable(zonelist, sc))
KOSAKI Motohirobb21c7c2010-06-04 14:15:05 -07002490 return 1;
2491
2492 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493}
2494
Mel Gorman55150612012-07-31 16:44:35 -07002495static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2496{
2497 struct zone *zone;
2498 unsigned long pfmemalloc_reserve = 0;
2499 unsigned long free_pages = 0;
2500 int i;
2501 bool wmark_ok;
2502
2503 for (i = 0; i <= ZONE_NORMAL; i++) {
2504 zone = &pgdat->node_zones[i];
2505 pfmemalloc_reserve += min_wmark_pages(zone);
2506 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2507 }
2508
2509 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2510
2511 /* kswapd must be awake if processes are being throttled */
2512 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2513 pgdat->classzone_idx = min(pgdat->classzone_idx,
2514 (enum zone_type)ZONE_NORMAL);
2515 wake_up_interruptible(&pgdat->kswapd_wait);
2516 }
2517
2518 return wmark_ok;
2519}
2520
2521/*
2522 * Throttle direct reclaimers if backing storage is backed by the network
2523 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2524 * depleted. kswapd will continue to make progress and wake the processes
Mel Gorman50694c22012-11-26 16:29:48 -08002525 * when the low watermark is reached.
2526 *
2527 * Returns true if a fatal signal was delivered during throttling. If this
2528 * happens, the page allocator should not consider triggering the OOM killer.
Mel Gorman55150612012-07-31 16:44:35 -07002529 */
Mel Gorman50694c22012-11-26 16:29:48 -08002530static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
Mel Gorman55150612012-07-31 16:44:35 -07002531 nodemask_t *nodemask)
2532{
2533 struct zone *zone;
2534 int high_zoneidx = gfp_zone(gfp_mask);
2535 pg_data_t *pgdat;
2536
2537 /*
2538 * Kernel threads should not be throttled as they may be indirectly
2539 * responsible for cleaning pages necessary for reclaim to make forward
2540 * progress. kjournald for example may enter direct reclaim while
2541 * committing a transaction where throttling it could forcing other
2542 * processes to block on log_wait_commit().
2543 */
2544 if (current->flags & PF_KTHREAD)
Mel Gorman50694c22012-11-26 16:29:48 -08002545 goto out;
2546
2547 /*
2548 * If a fatal signal is pending, this process should not throttle.
2549 * It should return quickly so it can exit and free its memory
2550 */
2551 if (fatal_signal_pending(current))
2552 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002553
2554 /* Check if the pfmemalloc reserves are ok */
2555 first_zones_zonelist(zonelist, high_zoneidx, NULL, &zone);
2556 pgdat = zone->zone_pgdat;
2557 if (pfmemalloc_watermark_ok(pgdat))
Mel Gorman50694c22012-11-26 16:29:48 -08002558 goto out;
Mel Gorman55150612012-07-31 16:44:35 -07002559
Mel Gorman68243e72012-07-31 16:44:39 -07002560 /* Account for the throttling */
2561 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2562
Mel Gorman55150612012-07-31 16:44:35 -07002563 /*
2564 * If the caller cannot enter the filesystem, it's possible that it
2565 * is due to the caller holding an FS lock or performing a journal
2566 * transaction in the case of a filesystem like ext[3|4]. In this case,
2567 * it is not safe to block on pfmemalloc_wait as kswapd could be
2568 * blocked waiting on the same lock. Instead, throttle for up to a
2569 * second before continuing.
2570 */
2571 if (!(gfp_mask & __GFP_FS)) {
2572 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2573 pfmemalloc_watermark_ok(pgdat), HZ);
Mel Gorman50694c22012-11-26 16:29:48 -08002574
2575 goto check_pending;
Mel Gorman55150612012-07-31 16:44:35 -07002576 }
2577
2578 /* Throttle until kswapd wakes the process */
2579 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2580 pfmemalloc_watermark_ok(pgdat));
Mel Gorman50694c22012-11-26 16:29:48 -08002581
2582check_pending:
2583 if (fatal_signal_pending(current))
2584 return true;
2585
2586out:
2587 return false;
Mel Gorman55150612012-07-31 16:44:35 -07002588}
2589
Mel Gormandac1d272008-04-28 02:12:12 -07002590unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002591 gfp_t gfp_mask, nodemask_t *nodemask)
Balbir Singh66e17072008-02-07 00:13:56 -08002592{
Mel Gorman33906bc2010-08-09 17:19:16 -07002593 unsigned long nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002594 struct scan_control sc = {
Ming Lei21caf2f2013-02-22 16:34:08 -08002595 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Balbir Singh66e17072008-02-07 00:13:56 -08002596 .may_writepage = !laptop_mode,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002597 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002598 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002599 .may_swap = 1,
Balbir Singh66e17072008-02-07 00:13:56 -08002600 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002601 .priority = DEF_PRIORITY,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002602 .target_mem_cgroup = NULL,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002603 .nodemask = nodemask,
Balbir Singh66e17072008-02-07 00:13:56 -08002604 };
Ying Hana09ed5e2011-05-24 17:12:26 -07002605 struct shrink_control shrink = {
2606 .gfp_mask = sc.gfp_mask,
2607 };
Balbir Singh66e17072008-02-07 00:13:56 -08002608
Mel Gorman55150612012-07-31 16:44:35 -07002609 /*
Mel Gorman50694c22012-11-26 16:29:48 -08002610 * Do not enter reclaim if fatal signal was delivered while throttled.
2611 * 1 is returned so that the page allocator does not OOM kill at this
2612 * point.
Mel Gorman55150612012-07-31 16:44:35 -07002613 */
Mel Gorman50694c22012-11-26 16:29:48 -08002614 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
Mel Gorman55150612012-07-31 16:44:35 -07002615 return 1;
2616
Mel Gorman33906bc2010-08-09 17:19:16 -07002617 trace_mm_vmscan_direct_reclaim_begin(order,
2618 sc.may_writepage,
2619 gfp_mask);
2620
Ying Hana09ed5e2011-05-24 17:12:26 -07002621 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Mel Gorman33906bc2010-08-09 17:19:16 -07002622
2623 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2624
2625 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002626}
2627
Andrew Mortonc255a452012-07-31 16:43:02 -07002628#ifdef CONFIG_MEMCG
Balbir Singh66e17072008-02-07 00:13:56 -08002629
Johannes Weiner72835c82012-01-12 17:18:32 -08002630unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002631 gfp_t gfp_mask, bool noswap,
Ying Han0ae5e892011-05-26 16:25:25 -07002632 struct zone *zone,
2633 unsigned long *nr_scanned)
Balbir Singh4e416952009-09-23 15:56:39 -07002634{
2635 struct scan_control sc = {
Ying Han0ae5e892011-05-26 16:25:25 -07002636 .nr_scanned = 0,
KOSAKI Motohirob8f5c562010-08-10 18:03:02 -07002637 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh4e416952009-09-23 15:56:39 -07002638 .may_writepage = !laptop_mode,
2639 .may_unmap = 1,
2640 .may_swap = !noswap,
Balbir Singh4e416952009-09-23 15:56:39 -07002641 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002642 .priority = 0,
Johannes Weiner72835c82012-01-12 17:18:32 -08002643 .target_mem_cgroup = memcg,
Balbir Singh4e416952009-09-23 15:56:39 -07002644 };
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002645 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Ying Han0ae5e892011-05-26 16:25:25 -07002646
Balbir Singh4e416952009-09-23 15:56:39 -07002647 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2648 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002649
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002650 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002651 sc.may_writepage,
2652 sc.gfp_mask);
2653
Balbir Singh4e416952009-09-23 15:56:39 -07002654 /*
2655 * NOTE: Although we can get the priority field, using it
2656 * here is not a good idea, since it limits the pages we can scan.
2657 * if we don't reclaim here, the shrink_zone from balance_pgdat
2658 * will pick up pages from other mem cgroup's as well. We hack
2659 * the priority and make it zero.
2660 */
Konstantin Khlebnikovf9be23d2012-05-29 15:07:02 -07002661 shrink_lruvec(lruvec, &sc);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002662
2663 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2664
Ying Han0ae5e892011-05-26 16:25:25 -07002665 *nr_scanned = sc.nr_scanned;
Balbir Singh4e416952009-09-23 15:56:39 -07002666 return sc.nr_reclaimed;
2667}
2668
Johannes Weiner72835c82012-01-12 17:18:32 -08002669unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
KOSAKI Motohiroa7885eb2009-01-07 18:08:24 -08002670 gfp_t gfp_mask,
Johannes Weiner185efc02011-09-14 16:21:58 -07002671 bool noswap)
Balbir Singh66e17072008-02-07 00:13:56 -08002672{
Balbir Singh4e416952009-09-23 15:56:39 -07002673 struct zonelist *zonelist;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002674 unsigned long nr_reclaimed;
Ying Han889976d2011-05-26 16:25:33 -07002675 int nid;
Balbir Singh66e17072008-02-07 00:13:56 -08002676 struct scan_control sc = {
Balbir Singh66e17072008-02-07 00:13:56 -08002677 .may_writepage = !laptop_mode,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002678 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002679 .may_swap = !noswap,
KOSAKI Motohiro22fba332009-12-14 17:59:10 -08002680 .nr_to_reclaim = SWAP_CLUSTER_MAX,
Balbir Singh66e17072008-02-07 00:13:56 -08002681 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002682 .priority = DEF_PRIORITY,
Johannes Weiner72835c82012-01-12 17:18:32 -08002683 .target_mem_cgroup = memcg,
KAMEZAWA Hiroyuki327c0e92009-03-31 15:23:31 -07002684 .nodemask = NULL, /* we don't care the placement */
Ying Hana09ed5e2011-05-24 17:12:26 -07002685 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2686 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
2687 };
2688 struct shrink_control shrink = {
2689 .gfp_mask = sc.gfp_mask,
Balbir Singh66e17072008-02-07 00:13:56 -08002690 };
Balbir Singh66e17072008-02-07 00:13:56 -08002691
Ying Han889976d2011-05-26 16:25:33 -07002692 /*
2693 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2694 * take care of from where we get pages. So the node where we start the
2695 * scan does not need to be the current node.
2696 */
Johannes Weiner72835c82012-01-12 17:18:32 -08002697 nid = mem_cgroup_select_victim_node(memcg);
Ying Han889976d2011-05-26 16:25:33 -07002698
2699 zonelist = NODE_DATA(nid)->node_zonelists;
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002700
2701 trace_mm_vmscan_memcg_reclaim_begin(0,
2702 sc.may_writepage,
2703 sc.gfp_mask);
2704
Ying Hana09ed5e2011-05-24 17:12:26 -07002705 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
KOSAKI Motohirobdce6d92010-08-09 17:19:56 -07002706
2707 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2708
2709 return nr_reclaimed;
Balbir Singh66e17072008-02-07 00:13:56 -08002710}
2711#endif
2712
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002713static void age_active_anon(struct zone *zone, struct scan_control *sc)
Johannes Weinerf16015f2012-01-12 17:17:52 -08002714{
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002715 struct mem_cgroup *memcg;
Johannes Weinerf16015f2012-01-12 17:17:52 -08002716
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002717 if (!total_swap_pages)
2718 return;
2719
2720 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2721 do {
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002722 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002723
Konstantin Khlebnikovc56d5c72012-05-29 15:07:00 -07002724 if (inactive_anon_is_low(lruvec))
Konstantin Khlebnikov1a93be02012-05-29 15:07:01 -07002725 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002726 sc, LRU_ACTIVE_ANON);
Johannes Weinerb95a2f22012-01-12 17:18:06 -08002727
2728 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2729 } while (memcg);
Johannes Weinerf16015f2012-01-12 17:17:52 -08002730}
2731
Johannes Weiner60cefed2012-11-29 13:54:23 -08002732static bool zone_balanced(struct zone *zone, int order,
2733 unsigned long balance_gap, int classzone_idx)
2734{
2735 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2736 balance_gap, classzone_idx, 0))
2737 return false;
2738
Kirill A. Shutemovd84da3f2012-12-11 16:00:31 -08002739 if (IS_ENABLED(CONFIG_COMPACTION) && order &&
2740 !compaction_suitable(zone, order))
Johannes Weiner60cefed2012-11-29 13:54:23 -08002741 return false;
2742
2743 return true;
2744}
2745
Mel Gorman1741c872011-01-13 15:46:21 -08002746/*
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002747 * pgdat_balanced() is used when checking if a node is balanced.
2748 *
2749 * For order-0, all zones must be balanced!
2750 *
2751 * For high-order allocations only zones that meet watermarks and are in a
2752 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2753 * total of balanced pages must be at least 25% of the zones allowed by
2754 * classzone_idx for the node to be considered balanced. Forcing all zones to
2755 * be balanced for high orders can cause excessive reclaim when there are
2756 * imbalanced zones.
Mel Gorman1741c872011-01-13 15:46:21 -08002757 * The choice of 25% is due to
2758 * o a 16M DMA zone that is balanced will not balance a zone on any
2759 * reasonable sized machine
2760 * o On all other machines, the top zone must be at least a reasonable
Lucas De Marchi25985ed2011-03-30 22:57:33 -03002761 * percentage of the middle zones. For example, on 32-bit x86, highmem
Mel Gorman1741c872011-01-13 15:46:21 -08002762 * would need to be at least 256M for it to be balance a whole node.
2763 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2764 * to balance a node on its own. These seemed like reasonable ratios.
2765 */
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002766static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
Mel Gorman1741c872011-01-13 15:46:21 -08002767{
Jiang Liub40da042013-02-22 16:33:52 -08002768 unsigned long managed_pages = 0;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002769 unsigned long balanced_pages = 0;
Mel Gorman1741c872011-01-13 15:46:21 -08002770 int i;
2771
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002772 /* Check the watermark levels */
2773 for (i = 0; i <= classzone_idx; i++) {
2774 struct zone *zone = pgdat->node_zones + i;
Mel Gorman1741c872011-01-13 15:46:21 -08002775
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002776 if (!populated_zone(zone))
2777 continue;
2778
Jiang Liub40da042013-02-22 16:33:52 -08002779 managed_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002780
2781 /*
2782 * A special case here:
2783 *
2784 * balance_pgdat() skips over all_unreclaimable after
2785 * DEF_PRIORITY. Effectively, it considers them balanced so
2786 * they must be considered balanced here as well!
2787 */
Lisa Du6e543d52013-09-11 14:22:36 -07002788 if (!zone_reclaimable(zone)) {
Jiang Liub40da042013-02-22 16:33:52 -08002789 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002790 continue;
2791 }
2792
2793 if (zone_balanced(zone, order, 0, i))
Jiang Liub40da042013-02-22 16:33:52 -08002794 balanced_pages += zone->managed_pages;
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002795 else if (!order)
2796 return false;
2797 }
2798
2799 if (order)
Jiang Liub40da042013-02-22 16:33:52 -08002800 return balanced_pages >= (managed_pages >> 2);
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002801 else
2802 return true;
Mel Gorman1741c872011-01-13 15:46:21 -08002803}
2804
Mel Gorman55150612012-07-31 16:44:35 -07002805/*
2806 * Prepare kswapd for sleeping. This verifies that there are no processes
2807 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2808 *
2809 * Returns true if kswapd is ready to sleep
2810 */
2811static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
Mel Gormandc83edd2011-01-13 15:46:26 -08002812 int classzone_idx)
Mel Gormanf50de2d2009-12-14 17:58:53 -08002813{
Mel Gormanf50de2d2009-12-14 17:58:53 -08002814 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2815 if (remaining)
Mel Gorman55150612012-07-31 16:44:35 -07002816 return false;
2817
2818 /*
2819 * There is a potential race between when kswapd checks its watermarks
2820 * and a process gets throttled. There is also a potential race if
2821 * processes get throttled, kswapd wakes, a large process exits therby
2822 * balancing the zones that causes kswapd to miss a wakeup. If kswapd
2823 * is going to sleep, no process should be sleeping on pfmemalloc_wait
2824 * so wake them now if necessary. If necessary, processes will wake
2825 * kswapd and get throttled again
2826 */
2827 if (waitqueue_active(&pgdat->pfmemalloc_wait)) {
2828 wake_up(&pgdat->pfmemalloc_wait);
2829 return false;
2830 }
Mel Gormanf50de2d2009-12-14 17:58:53 -08002831
Zlatko Calusic4ae0a482012-12-23 15:12:54 +01002832 return pgdat_balanced(pgdat, order, classzone_idx);
Mel Gormanf50de2d2009-12-14 17:58:53 -08002833}
2834
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835/*
Mel Gorman75485362013-07-03 15:01:42 -07002836 * kswapd shrinks the zone by the number of pages required to reach
2837 * the high watermark.
Mel Gormanb8e83b92013-07-03 15:01:45 -07002838 *
2839 * Returns true if kswapd scanned at least the requested number of pages to
Mel Gorman283aba92013-07-03 15:01:51 -07002840 * reclaim or if the lack of progress was due to pages under writeback.
2841 * This is used to determine if the scanning priority needs to be raised.
Mel Gorman75485362013-07-03 15:01:42 -07002842 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07002843static bool kswapd_shrink_zone(struct zone *zone,
Mel Gorman7c954f62013-07-03 15:01:54 -07002844 int classzone_idx,
Mel Gorman75485362013-07-03 15:01:42 -07002845 struct scan_control *sc,
Mel Gorman2ab44f42013-07-03 15:01:47 -07002846 unsigned long lru_pages,
2847 unsigned long *nr_attempted)
Mel Gorman75485362013-07-03 15:01:42 -07002848{
Mel Gorman7c954f62013-07-03 15:01:54 -07002849 int testorder = sc->order;
2850 unsigned long balance_gap;
Mel Gorman75485362013-07-03 15:01:42 -07002851 struct reclaim_state *reclaim_state = current->reclaim_state;
2852 struct shrink_control shrink = {
2853 .gfp_mask = sc->gfp_mask,
2854 };
Mel Gorman7c954f62013-07-03 15:01:54 -07002855 bool lowmem_pressure;
Mel Gorman75485362013-07-03 15:01:42 -07002856
2857 /* Reclaim above the high watermark. */
2858 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
Mel Gorman7c954f62013-07-03 15:01:54 -07002859
2860 /*
2861 * Kswapd reclaims only single pages with compaction enabled. Trying
2862 * too hard to reclaim until contiguous free pages have become
2863 * available can hurt performance by evicting too much useful data
2864 * from memory. Do not reclaim more than needed for compaction.
2865 */
2866 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
2867 compaction_suitable(zone, sc->order) !=
2868 COMPACT_SKIPPED)
2869 testorder = 0;
2870
2871 /*
2872 * We put equal pressure on every zone, unless one zone has way too
2873 * many pages free already. The "too many pages" is defined as the
2874 * high wmark plus a "gap" where the gap is either the low
2875 * watermark or 1% of the zone, whichever is smaller.
2876 */
2877 balance_gap = min(low_wmark_pages(zone),
2878 (zone->managed_pages + KSWAPD_ZONE_BALANCE_GAP_RATIO-1) /
2879 KSWAPD_ZONE_BALANCE_GAP_RATIO);
2880
2881 /*
2882 * If there is no low memory pressure or the zone is balanced then no
2883 * reclaim is necessary
2884 */
2885 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
2886 if (!lowmem_pressure && zone_balanced(zone, testorder,
2887 balance_gap, classzone_idx))
2888 return true;
2889
Mel Gorman75485362013-07-03 15:01:42 -07002890 shrink_zone(zone, sc);
Dave Chinner0ce3d742013-08-28 10:18:03 +10002891 nodes_clear(shrink.nodes_to_scan);
2892 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
Mel Gorman75485362013-07-03 15:01:42 -07002893
2894 reclaim_state->reclaimed_slab = 0;
Lisa Du6e543d52013-09-11 14:22:36 -07002895 shrink_slab(&shrink, sc->nr_scanned, lru_pages);
Mel Gorman75485362013-07-03 15:01:42 -07002896 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2897
Mel Gorman2ab44f42013-07-03 15:01:47 -07002898 /* Account for the number of pages attempted to reclaim */
2899 *nr_attempted += sc->nr_to_reclaim;
2900
Mel Gorman283aba92013-07-03 15:01:51 -07002901 zone_clear_flag(zone, ZONE_WRITEBACK);
2902
Mel Gorman7c954f62013-07-03 15:01:54 -07002903 /*
2904 * If a zone reaches its high watermark, consider it to be no longer
2905 * congested. It's possible there are dirty pages backed by congested
2906 * BDIs but as pressure is relieved, speculatively avoid congestion
2907 * waits.
2908 */
Lisa Du6e543d52013-09-11 14:22:36 -07002909 if (zone_reclaimable(zone) &&
Mel Gorman7c954f62013-07-03 15:01:54 -07002910 zone_balanced(zone, testorder, 0, classzone_idx)) {
2911 zone_clear_flag(zone, ZONE_CONGESTED);
2912 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
2913 }
2914
Mel Gormanb8e83b92013-07-03 15:01:45 -07002915 return sc->nr_scanned >= sc->nr_to_reclaim;
Mel Gorman75485362013-07-03 15:01:42 -07002916}
2917
2918/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002919 * For kswapd, balance_pgdat() will work across all this node's zones until
Mel Gorman41858962009-06-16 15:32:12 -07002920 * they are all at high_wmark_pages(zone).
Linus Torvalds1da177e2005-04-16 15:20:36 -07002921 *
Mel Gorman0abdee22011-01-13 15:46:22 -08002922 * Returns the final order kswapd was reclaiming at
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923 *
2924 * There is special handling here for zones which are full of pinned pages.
2925 * This can happen if the pages are all mlocked, or if they are all used by
2926 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
2927 * What we do is to detect the case where all pages in the zone have been
2928 * scanned twice and there has been zero successful reclaim. Mark the zone as
2929 * dead and from now on, only perform a short scan. Basically we're polling
2930 * the zone for when the problem goes away.
2931 *
2932 * kswapd scans the zones in the highmem->normal->dma direction. It skips
Mel Gorman41858962009-06-16 15:32:12 -07002933 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
2934 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
2935 * lower zones regardless of the number of free pages in the lower zones. This
2936 * interoperates with the page allocator fallback scheme to ensure that aging
2937 * of pages is balanced across the zones.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002938 */
Mel Gorman99504742011-01-13 15:46:20 -08002939static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
Mel Gormandc83edd2011-01-13 15:46:26 -08002940 int *classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002942 int i;
Mel Gorman99504742011-01-13 15:46:20 -08002943 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
Andrew Morton0608f432013-09-24 15:27:41 -07002944 unsigned long nr_soft_reclaimed;
2945 unsigned long nr_soft_scanned;
Andrew Morton179e9632006-03-22 00:08:18 -08002946 struct scan_control sc = {
2947 .gfp_mask = GFP_KERNEL,
Mel Gormanb8e83b92013-07-03 15:01:45 -07002948 .priority = DEF_PRIORITY,
Johannes Weinera6dc60f2009-03-31 15:19:30 -07002949 .may_unmap = 1,
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07002950 .may_swap = 1,
Mel Gormanb8e83b92013-07-03 15:01:45 -07002951 .may_writepage = !laptop_mode,
Andy Whitcroft5ad333e2007-07-17 04:03:16 -07002952 .order = order,
Johannes Weinerf16015f2012-01-12 17:17:52 -08002953 .target_mem_cgroup = NULL,
Andrew Morton179e9632006-03-22 00:08:18 -08002954 };
Christoph Lameterf8891e52006-06-30 01:55:45 -07002955 count_vm_event(PAGEOUTRUN);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002956
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002957 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002958 unsigned long lru_pages = 0;
Mel Gorman2ab44f42013-07-03 15:01:47 -07002959 unsigned long nr_attempted = 0;
Mel Gormanb8e83b92013-07-03 15:01:45 -07002960 bool raise_priority = true;
Mel Gorman2ab44f42013-07-03 15:01:47 -07002961 bool pgdat_needs_compaction = (order > 0);
Mel Gormanb8e83b92013-07-03 15:01:45 -07002962
2963 sc.nr_reclaimed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002964
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002965 /*
2966 * Scan in the highmem->dma direction for the highest
2967 * zone which needs scanning
2968 */
2969 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
2970 struct zone *zone = pgdat->node_zones + i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002971
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002972 if (!populated_zone(zone))
2973 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002974
Lisa Du6e543d52013-09-11 14:22:36 -07002975 if (sc.priority != DEF_PRIORITY &&
2976 !zone_reclaimable(zone))
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002977 continue;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002978
Rik van Riel556adec2008-10-18 20:26:34 -07002979 /*
2980 * Do some background aging of the anon list, to give
2981 * pages a chance to be referenced before reclaiming.
2982 */
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07002983 age_active_anon(zone, &sc);
Rik van Riel556adec2008-10-18 20:26:34 -07002984
Mel Gormancc715d92012-03-21 16:34:00 -07002985 /*
2986 * If the number of buffer_heads in the machine
2987 * exceeds the maximum allowed level and this node
2988 * has a highmem zone, force kswapd to reclaim from
2989 * it to relieve lowmem pressure.
2990 */
2991 if (buffer_heads_over_limit && is_highmem_idx(i)) {
2992 end_zone = i;
2993 break;
2994 }
2995
Johannes Weiner60cefed2012-11-29 13:54:23 -08002996 if (!zone_balanced(zone, order, 0, 0)) {
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07002997 end_zone = i;
Andrew Mortone1dbeda2006-12-06 20:32:01 -08002998 break;
Shaohua Li439423f2011-08-25 15:59:12 -07002999 } else {
Mel Gormand43006d2013-07-03 15:01:50 -07003000 /*
3001 * If balanced, clear the dirty and congested
3002 * flags
3003 */
Shaohua Li439423f2011-08-25 15:59:12 -07003004 zone_clear_flag(zone, ZONE_CONGESTED);
Mel Gormand43006d2013-07-03 15:01:50 -07003005 zone_clear_flag(zone, ZONE_TAIL_LRU_DIRTY);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003006 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003007 }
Zlatko Calusicdafcb732013-02-22 16:32:34 -08003008
Mel Gormanb8e83b92013-07-03 15:01:45 -07003009 if (i < 0)
Andrew Mortone1dbeda2006-12-06 20:32:01 -08003010 goto out;
3011
Linus Torvalds1da177e2005-04-16 15:20:36 -07003012 for (i = 0; i <= end_zone; i++) {
3013 struct zone *zone = pgdat->node_zones + i;
3014
Mel Gorman2ab44f42013-07-03 15:01:47 -07003015 if (!populated_zone(zone))
3016 continue;
3017
Wu Fengguangadea02a2009-09-21 17:01:42 -07003018 lru_pages += zone_reclaimable_pages(zone);
Mel Gorman2ab44f42013-07-03 15:01:47 -07003019
3020 /*
3021 * If any zone is currently balanced then kswapd will
3022 * not call compaction as it is expected that the
3023 * necessary pages are already available.
3024 */
3025 if (pgdat_needs_compaction &&
3026 zone_watermark_ok(zone, order,
3027 low_wmark_pages(zone),
3028 *classzone_idx, 0))
3029 pgdat_needs_compaction = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003030 }
3031
3032 /*
Mel Gormanb7ea3c42013-07-03 15:01:53 -07003033 * If we're getting trouble reclaiming, start doing writepage
3034 * even in laptop mode.
3035 */
3036 if (sc.priority < DEF_PRIORITY - 2)
3037 sc.may_writepage = 1;
3038
3039 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003040 * Now scan the zone in the dma->highmem direction, stopping
3041 * at the last zone which needs scanning.
3042 *
3043 * We do this because the page allocator works in the opposite
3044 * direction. This prevents the page allocator from allocating
3045 * pages behind kswapd's direction of progress, which would
3046 * cause too much scanning of the lower zones.
3047 */
3048 for (i = 0; i <= end_zone; i++) {
3049 struct zone *zone = pgdat->node_zones + i;
3050
Con Kolivasf3fe6512006-01-06 00:11:15 -08003051 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003052 continue;
3053
Lisa Du6e543d52013-09-11 14:22:36 -07003054 if (sc.priority != DEF_PRIORITY &&
3055 !zone_reclaimable(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003056 continue;
3057
Linus Torvalds1da177e2005-04-16 15:20:36 -07003058 sc.nr_scanned = 0;
Balbir Singh4e416952009-09-23 15:56:39 -07003059
Andrew Morton0608f432013-09-24 15:27:41 -07003060 nr_soft_scanned = 0;
3061 /*
3062 * Call soft limit reclaim before calling shrink_zone.
3063 */
3064 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3065 order, sc.gfp_mask,
3066 &nr_soft_scanned);
3067 sc.nr_reclaimed += nr_soft_reclaimed;
3068
Rik van Riel32a43302007-10-16 01:24:50 -07003069 /*
Mel Gorman7c954f62013-07-03 15:01:54 -07003070 * There should be no need to raise the scanning
3071 * priority if enough pages are already being scanned
3072 * that that high watermark would be met at 100%
3073 * efficiency.
Rik van Riel32a43302007-10-16 01:24:50 -07003074 */
Mel Gorman7c954f62013-07-03 15:01:54 -07003075 if (kswapd_shrink_zone(zone, end_zone, &sc,
3076 lru_pages, &nr_attempted))
3077 raise_priority = false;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003078 }
Mel Gorman55150612012-07-31 16:44:35 -07003079
3080 /*
3081 * If the low watermark is met there is no need for processes
3082 * to be throttled on pfmemalloc_wait as they should not be
3083 * able to safely make forward progress. Wake them
3084 */
3085 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3086 pfmemalloc_watermark_ok(pgdat))
3087 wake_up(&pgdat->pfmemalloc_wait);
3088
Linus Torvalds1da177e2005-04-16 15:20:36 -07003089 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003090 * Fragmentation may mean that the system cannot be rebalanced
3091 * for high-order allocations in all zones. If twice the
3092 * allocation size has been reclaimed and the zones are still
3093 * not balanced then recheck the watermarks at order-0 to
3094 * prevent kswapd reclaiming excessively. Assume that a
3095 * process requested a high-order can direct reclaim/compact.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003096 */
Mel Gormanb8e83b92013-07-03 15:01:45 -07003097 if (order && sc.nr_reclaimed >= 2UL << order)
KOSAKI Motohiro73ce02e2009-01-06 14:40:33 -08003098 order = sc.order = 0;
3099
Mel Gormanb8e83b92013-07-03 15:01:45 -07003100 /* Check if kswapd should be suspending */
3101 if (try_to_freeze() || kthread_should_stop())
3102 break;
3103
3104 /*
Mel Gorman2ab44f42013-07-03 15:01:47 -07003105 * Compact if necessary and kswapd is reclaiming at least the
3106 * high watermark number of pages as requsted
3107 */
3108 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3109 compact_pgdat(pgdat, order);
3110
3111 /*
Mel Gormanb8e83b92013-07-03 15:01:45 -07003112 * Raise priority if scanning rate is too low or there was no
3113 * progress in reclaiming pages
3114 */
3115 if (raise_priority || !sc.nr_reclaimed)
3116 sc.priority--;
Mel Gorman9aa41342013-07-03 15:01:48 -07003117 } while (sc.priority >= 1 &&
Mel Gormanb8e83b92013-07-03 15:01:45 -07003118 !pgdat_balanced(pgdat, order, *classzone_idx));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119
Mel Gormanb8e83b92013-07-03 15:01:45 -07003120out:
Mel Gorman0abdee22011-01-13 15:46:22 -08003121 /*
Mel Gorman55150612012-07-31 16:44:35 -07003122 * Return the order we were reclaiming at so prepare_kswapd_sleep()
Mel Gorman0abdee22011-01-13 15:46:22 -08003123 * makes a decision on the order we were last reclaiming at. However,
3124 * if another caller entered the allocator slow path while kswapd
3125 * was awake, order will remain at the higher level
3126 */
Mel Gormandc83edd2011-01-13 15:46:26 -08003127 *classzone_idx = end_zone;
Mel Gorman0abdee22011-01-13 15:46:22 -08003128 return order;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129}
3130
Mel Gormandc83edd2011-01-13 15:46:26 -08003131static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003132{
3133 long remaining = 0;
3134 DEFINE_WAIT(wait);
3135
3136 if (freezing(current) || kthread_should_stop())
3137 return;
3138
3139 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3140
3141 /* Try to sleep for a short interval */
Mel Gorman55150612012-07-31 16:44:35 -07003142 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003143 remaining = schedule_timeout(HZ/10);
3144 finish_wait(&pgdat->kswapd_wait, &wait);
3145 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3146 }
3147
3148 /*
3149 * After a short sleep, check if it was a premature sleep. If not, then
3150 * go fully to sleep until explicitly woken up.
3151 */
Mel Gorman55150612012-07-31 16:44:35 -07003152 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003153 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3154
3155 /*
3156 * vmstat counters are not perfectly accurate and the estimated
3157 * value for counters such as NR_FREE_PAGES can deviate from the
3158 * true value by nr_online_cpus * threshold. To avoid the zone
3159 * watermarks being breached while under pressure, we reduce the
3160 * per-cpu vmstat threshold while kswapd is awake and restore
3161 * them before going back to sleep.
3162 */
3163 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003164
Mel Gorman62997022012-10-08 16:32:47 -07003165 /*
3166 * Compaction records what page blocks it recently failed to
3167 * isolate pages from and skips them in the future scanning.
3168 * When kswapd is going to sleep, it is reasonable to assume
3169 * that pages and compaction may succeed so reset the cache.
3170 */
3171 reset_isolation_suitable(pgdat);
3172
Aaditya Kumar1c7e7f62012-07-17 15:48:07 -07003173 if (!kthread_should_stop())
3174 schedule();
3175
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003176 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3177 } else {
3178 if (remaining)
3179 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3180 else
3181 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3182 }
3183 finish_wait(&pgdat->kswapd_wait, &wait);
3184}
3185
Linus Torvalds1da177e2005-04-16 15:20:36 -07003186/*
3187 * The background pageout daemon, started as a kernel thread
Rik van Riel4f98a2f2008-10-18 20:26:32 -07003188 * from the init process.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189 *
3190 * This basically trickles out pages so that we have _some_
3191 * free memory available even if there is no other activity
3192 * that frees anything up. This is needed for things like routing
3193 * etc, where we otherwise might have all activity going on in
3194 * asynchronous contexts that cannot page things out.
3195 *
3196 * If there are applications that are active memory-allocators
3197 * (most normal use), this basically shouldn't matter.
3198 */
3199static int kswapd(void *p)
3200{
Mel Gorman215ddd62011-07-08 15:39:40 -07003201 unsigned long order, new_order;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003202 unsigned balanced_order;
Mel Gorman215ddd62011-07-08 15:39:40 -07003203 int classzone_idx, new_classzone_idx;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003204 int balanced_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003205 pg_data_t *pgdat = (pg_data_t*)p;
3206 struct task_struct *tsk = current;
KOSAKI Motohirof0bc0a62011-01-13 15:45:50 -08003207
Linus Torvalds1da177e2005-04-16 15:20:36 -07003208 struct reclaim_state reclaim_state = {
3209 .reclaimed_slab = 0,
3210 };
Rusty Russella70f7302009-03-13 14:49:46 +10303211 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003212
Nick Piggincf40bd12009-01-21 08:12:39 +01003213 lockdep_set_current_reclaim_state(GFP_KERNEL);
3214
Rusty Russell174596a2009-01-01 10:12:29 +10303215 if (!cpumask_empty(cpumask))
Mike Travisc5f59f02008-04-04 18:11:10 -07003216 set_cpus_allowed_ptr(tsk, cpumask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003217 current->reclaim_state = &reclaim_state;
3218
3219 /*
3220 * Tell the memory management that we're a "memory allocator",
3221 * and that if we need more memory we should get access to it
3222 * regardless (see "__alloc_pages()"). "kswapd" should
3223 * never get caught in the normal page freeing logic.
3224 *
3225 * (Kswapd normally doesn't need memory anyway, but sometimes
3226 * you need a small amount of memory in order to be able to
3227 * page out something else, and this flag essentially protects
3228 * us from recursively trying to free more memory as we're
3229 * trying to free the first piece of memory in the first place).
3230 */
Christoph Lameter930d9152006-01-08 01:00:47 -08003231 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
Rafael J. Wysocki83144182007-07-17 04:03:35 -07003232 set_freezable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07003233
Mel Gorman215ddd62011-07-08 15:39:40 -07003234 order = new_order = 0;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003235 balanced_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003236 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003237 balanced_classzone_idx = classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003238 for ( ; ; ) {
Jeff Liu6f6313d2012-12-11 16:02:48 -08003239 bool ret;
Christoph Lameter3e1d1d22005-06-24 23:13:50 -07003240
Mel Gorman215ddd62011-07-08 15:39:40 -07003241 /*
3242 * If the last balance_pgdat was unsuccessful it's unlikely a
3243 * new request of a similar or harder type will succeed soon
3244 * so consider going to sleep on the basis we reclaimed at
3245 */
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003246 if (balanced_classzone_idx >= new_classzone_idx &&
3247 balanced_order == new_order) {
Mel Gorman215ddd62011-07-08 15:39:40 -07003248 new_order = pgdat->kswapd_max_order;
3249 new_classzone_idx = pgdat->classzone_idx;
3250 pgdat->kswapd_max_order = 0;
3251 pgdat->classzone_idx = pgdat->nr_zones - 1;
3252 }
3253
Mel Gorman99504742011-01-13 15:46:20 -08003254 if (order < new_order || classzone_idx > new_classzone_idx) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003255 /*
3256 * Don't sleep if someone wants a larger 'order'
Mel Gorman99504742011-01-13 15:46:20 -08003257 * allocation or has tigher zone constraints
Linus Torvalds1da177e2005-04-16 15:20:36 -07003258 */
3259 order = new_order;
Mel Gorman99504742011-01-13 15:46:20 -08003260 classzone_idx = new_classzone_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003261 } else {
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003262 kswapd_try_to_sleep(pgdat, balanced_order,
3263 balanced_classzone_idx);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003264 order = pgdat->kswapd_max_order;
Mel Gorman99504742011-01-13 15:46:20 -08003265 classzone_idx = pgdat->classzone_idx;
Alex,Shif0dfcde2011-10-31 17:08:45 -07003266 new_order = order;
3267 new_classzone_idx = classzone_idx;
Mel Gorman4d405022011-01-13 15:46:23 -08003268 pgdat->kswapd_max_order = 0;
Mel Gorman215ddd62011-07-08 15:39:40 -07003269 pgdat->classzone_idx = pgdat->nr_zones - 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003270 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003271
David Rientjes8fe23e02009-12-14 17:58:33 -08003272 ret = try_to_freeze();
3273 if (kthread_should_stop())
3274 break;
3275
3276 /*
3277 * We can speed up thawing tasks if we don't call balance_pgdat
3278 * after returning from the refrigerator
3279 */
Mel Gorman33906bc2010-08-09 17:19:16 -07003280 if (!ret) {
3281 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
Alex,Shid2ebd0f62011-10-31 17:08:39 -07003282 balanced_classzone_idx = classzone_idx;
3283 balanced_order = balance_pgdat(pgdat, order,
3284 &balanced_classzone_idx);
Mel Gorman33906bc2010-08-09 17:19:16 -07003285 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003286 }
Takamori Yamaguchib0a8cc52012-11-08 15:53:39 -08003287
3288 current->reclaim_state = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003289 return 0;
3290}
3291
3292/*
3293 * A zone is low on free memory, so wake its kswapd task to service it.
3294 */
Mel Gorman99504742011-01-13 15:46:20 -08003295void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003296{
3297 pg_data_t *pgdat;
3298
Con Kolivasf3fe6512006-01-06 00:11:15 -08003299 if (!populated_zone(zone))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003300 return;
3301
Paul Jackson02a0e532006-12-13 00:34:25 -08003302 if (!cpuset_zone_allowed_hardwall(zone, GFP_KERNEL))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303 return;
Mel Gorman88f5acf2011-01-13 15:45:41 -08003304 pgdat = zone->zone_pgdat;
Mel Gorman99504742011-01-13 15:46:20 -08003305 if (pgdat->kswapd_max_order < order) {
Mel Gorman88f5acf2011-01-13 15:45:41 -08003306 pgdat->kswapd_max_order = order;
Mel Gorman99504742011-01-13 15:46:20 -08003307 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3308 }
Con Kolivas8d0986e2005-09-13 01:25:07 -07003309 if (!waitqueue_active(&pgdat->kswapd_wait))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310 return;
Johannes Weiner892f7952013-09-11 14:20:39 -07003311 if (zone_balanced(zone, order, 0, 0))
Mel Gorman88f5acf2011-01-13 15:45:41 -08003312 return;
3313
3314 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
Con Kolivas8d0986e2005-09-13 01:25:07 -07003315 wake_up_interruptible(&pgdat->kswapd_wait);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003316}
3317
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003318#ifdef CONFIG_HIBERNATION
Linus Torvalds1da177e2005-04-16 15:20:36 -07003319/*
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003320 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003321 * freed pages.
3322 *
3323 * Rather than trying to age LRUs the aim is to preserve the overall
3324 * LRU order by reclaiming preferentially
3325 * inactive > active > active referenced > active mapped
Linus Torvalds1da177e2005-04-16 15:20:36 -07003326 */
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003327unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003328{
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003329 struct reclaim_state reclaim_state;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003330 struct scan_control sc = {
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003331 .gfp_mask = GFP_HIGHUSER_MOVABLE,
3332 .may_swap = 1,
3333 .may_unmap = 1,
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003334 .may_writepage = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003335 .nr_to_reclaim = nr_to_reclaim,
3336 .hibernation_mode = 1,
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003337 .order = 0,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003338 .priority = DEF_PRIORITY,
Linus Torvalds1da177e2005-04-16 15:20:36 -07003339 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003340 struct shrink_control shrink = {
3341 .gfp_mask = sc.gfp_mask,
3342 };
3343 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003344 struct task_struct *p = current;
3345 unsigned long nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003346
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003347 p->flags |= PF_MEMALLOC;
3348 lockdep_set_current_reclaim_state(sc.gfp_mask);
3349 reclaim_state.reclaimed_slab = 0;
3350 p->reclaim_state = &reclaim_state;
Andrew Morton69e05942006-03-22 00:08:19 -08003351
Ying Hana09ed5e2011-05-24 17:12:26 -07003352 nr_reclaimed = do_try_to_free_pages(zonelist, &sc, &shrink);
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003353
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003354 p->reclaim_state = NULL;
3355 lockdep_clear_current_reclaim_state();
3356 p->flags &= ~PF_MEMALLOC;
Rafael J. Wysockid6277db2006-06-23 02:03:18 -07003357
KOSAKI Motohiro7b517552009-12-14 17:59:12 -08003358 return nr_reclaimed;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003359}
Rafael J. Wysockic6f37f12009-05-24 22:16:31 +02003360#endif /* CONFIG_HIBERNATION */
Linus Torvalds1da177e2005-04-16 15:20:36 -07003361
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362/* It's optimal to keep kswapds on the same CPUs as their memory, but
3363 not required for correctness. So if the last cpu in a node goes
3364 away, we get changed to run anywhere: as the first one comes back,
3365 restore their cpu bindings. */
Greg Kroah-Hartmanfcb35a92012-12-21 15:01:06 -08003366static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3367 void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003368{
Yasunori Goto58c0a4a2007-10-16 01:25:40 -07003369 int nid;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07003371 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003372 for_each_node_state(nid, N_MEMORY) {
Mike Travisc5f59f02008-04-04 18:11:10 -07003373 pg_data_t *pgdat = NODE_DATA(nid);
Rusty Russella70f7302009-03-13 14:49:46 +10303374 const struct cpumask *mask;
3375
3376 mask = cpumask_of_node(pgdat->node_id);
Mike Travisc5f59f02008-04-04 18:11:10 -07003377
Rusty Russell3e597942009-01-01 10:12:24 +10303378 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003379 /* One of our CPUs online: restore mask */
Mike Travisc5f59f02008-04-04 18:11:10 -07003380 set_cpus_allowed_ptr(pgdat->kswapd, mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381 }
3382 }
3383 return NOTIFY_OK;
3384}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003385
Yasunori Goto3218ae12006-06-27 02:53:33 -07003386/*
3387 * This kswapd start function will be called by init and node-hot-add.
3388 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3389 */
3390int kswapd_run(int nid)
3391{
3392 pg_data_t *pgdat = NODE_DATA(nid);
3393 int ret = 0;
3394
3395 if (pgdat->kswapd)
3396 return 0;
3397
3398 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3399 if (IS_ERR(pgdat->kswapd)) {
3400 /* failure at boot is fatal */
3401 BUG_ON(system_state == SYSTEM_BOOTING);
Gavin Shand5dc0ad2012-10-08 16:29:27 -07003402 pr_err("Failed to start kswapd on node %d\n", nid);
3403 ret = PTR_ERR(pgdat->kswapd);
Xishi Qiud72515b2013-04-17 15:58:34 -07003404 pgdat->kswapd = NULL;
Yasunori Goto3218ae12006-06-27 02:53:33 -07003405 }
3406 return ret;
3407}
3408
David Rientjes8fe23e02009-12-14 17:58:33 -08003409/*
Jiang Liud8adde12012-07-11 14:01:52 -07003410 * Called by memory hotplug when all memory in a node is offlined. Caller must
3411 * hold lock_memory_hotplug().
David Rientjes8fe23e02009-12-14 17:58:33 -08003412 */
3413void kswapd_stop(int nid)
3414{
3415 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3416
Jiang Liud8adde12012-07-11 14:01:52 -07003417 if (kswapd) {
David Rientjes8fe23e02009-12-14 17:58:33 -08003418 kthread_stop(kswapd);
Jiang Liud8adde12012-07-11 14:01:52 -07003419 NODE_DATA(nid)->kswapd = NULL;
3420 }
David Rientjes8fe23e02009-12-14 17:58:33 -08003421}
3422
Linus Torvalds1da177e2005-04-16 15:20:36 -07003423static int __init kswapd_init(void)
3424{
Yasunori Goto3218ae12006-06-27 02:53:33 -07003425 int nid;
Andrew Morton69e05942006-03-22 00:08:19 -08003426
Linus Torvalds1da177e2005-04-16 15:20:36 -07003427 swap_setup();
Lai Jiangshan48fb2e22012-12-12 13:51:43 -08003428 for_each_node_state(nid, N_MEMORY)
Yasunori Goto3218ae12006-06-27 02:53:33 -07003429 kswapd_run(nid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003430 hotcpu_notifier(cpu_callback, 0);
3431 return 0;
3432}
3433
3434module_init(kswapd_init)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003435
3436#ifdef CONFIG_NUMA
3437/*
3438 * Zone reclaim mode
3439 *
3440 * If non-zero call zone_reclaim when the number of free pages falls below
3441 * the watermarks.
Christoph Lameter9eeff232006-01-18 17:42:31 -08003442 */
3443int zone_reclaim_mode __read_mostly;
3444
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003445#define RECLAIM_OFF 0
Fernando Luis Vazquez Cao7d034312008-07-29 22:33:41 -07003446#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
Christoph Lameter1b2ffb72006-02-01 03:05:34 -08003447#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3448#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3449
Christoph Lameter9eeff232006-01-18 17:42:31 -08003450/*
Christoph Lametera92f7122006-02-01 03:05:32 -08003451 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3452 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3453 * a zone.
3454 */
3455#define ZONE_RECLAIM_PRIORITY 4
3456
Christoph Lameter9eeff232006-01-18 17:42:31 -08003457/*
Christoph Lameter96146342006-07-03 00:24:13 -07003458 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3459 * occur.
3460 */
3461int sysctl_min_unmapped_ratio = 1;
3462
3463/*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003464 * If the number of slab pages in a zone grows beyond this percentage then
3465 * slab reclaim needs to occur.
3466 */
3467int sysctl_min_slab_ratio = 5;
3468
Mel Gorman90afa5d2009-06-16 15:33:20 -07003469static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3470{
3471 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3472 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3473 zone_page_state(zone, NR_ACTIVE_FILE);
3474
3475 /*
3476 * It's possible for there to be more file mapped pages than
3477 * accounted for by the pages on the file LRU lists because
3478 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3479 */
3480 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3481}
3482
3483/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3484static long zone_pagecache_reclaimable(struct zone *zone)
3485{
3486 long nr_pagecache_reclaimable;
3487 long delta = 0;
3488
3489 /*
3490 * If RECLAIM_SWAP is set, then all file pages are considered
3491 * potentially reclaimable. Otherwise, we have to worry about
3492 * pages like swapcache and zone_unmapped_file_pages() provides
3493 * a better estimate
3494 */
3495 if (zone_reclaim_mode & RECLAIM_SWAP)
3496 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3497 else
3498 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3499
3500 /* If we can't clean pages, remove dirty pages from consideration */
3501 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3502 delta += zone_page_state(zone, NR_FILE_DIRTY);
3503
3504 /* Watch for any possible underflows due to delta */
3505 if (unlikely(delta > nr_pagecache_reclaimable))
3506 delta = nr_pagecache_reclaimable;
3507
3508 return nr_pagecache_reclaimable - delta;
3509}
3510
Christoph Lameter0ff38492006-09-25 23:31:52 -07003511/*
Christoph Lameter9eeff232006-01-18 17:42:31 -08003512 * Try to free up some pages from this zone through reclaim.
3513 */
Andrew Morton179e9632006-03-22 00:08:18 -08003514static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
Christoph Lameter9eeff232006-01-18 17:42:31 -08003515{
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003516 /* Minimum pages needed in order to stay on node */
Andrew Morton69e05942006-03-22 00:08:19 -08003517 const unsigned long nr_pages = 1 << order;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003518 struct task_struct *p = current;
3519 struct reclaim_state reclaim_state;
Andrew Morton179e9632006-03-22 00:08:18 -08003520 struct scan_control sc = {
3521 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
Johannes Weinera6dc60f2009-03-31 15:19:30 -07003522 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
KOSAKI Motohiro2e2e4252009-04-21 12:24:57 -07003523 .may_swap = 1,
Andrew Morton62b726c2013-02-22 16:32:24 -08003524 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
Ming Lei21caf2f2013-02-22 16:34:08 -08003525 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
Johannes Weinerbd2f6192009-03-31 15:19:38 -07003526 .order = order,
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003527 .priority = ZONE_RECLAIM_PRIORITY,
Andrew Morton179e9632006-03-22 00:08:18 -08003528 };
Ying Hana09ed5e2011-05-24 17:12:26 -07003529 struct shrink_control shrink = {
3530 .gfp_mask = sc.gfp_mask,
3531 };
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003532 unsigned long nr_slab_pages0, nr_slab_pages1;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003533
Christoph Lameter9eeff232006-01-18 17:42:31 -08003534 cond_resched();
Christoph Lameterd4f77962006-02-24 13:04:22 -08003535 /*
3536 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3537 * and we also need to be able to write out pages for RECLAIM_WRITE
3538 * and RECLAIM_SWAP.
3539 */
3540 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003541 lockdep_set_current_reclaim_state(gfp_mask);
Christoph Lameter9eeff232006-01-18 17:42:31 -08003542 reclaim_state.reclaimed_slab = 0;
3543 p->reclaim_state = &reclaim_state;
Christoph Lameterc84db232006-02-01 03:05:29 -08003544
Mel Gorman90afa5d2009-06-16 15:33:20 -07003545 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
Christoph Lameter0ff38492006-09-25 23:31:52 -07003546 /*
3547 * Free memory by calling shrink zone with increasing
3548 * priorities until we have enough memory freed.
3549 */
Christoph Lameter0ff38492006-09-25 23:31:52 -07003550 do {
Konstantin Khlebnikov9e3b2f82012-05-29 15:06:57 -07003551 shrink_zone(zone, &sc);
3552 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
Christoph Lameter0ff38492006-09-25 23:31:52 -07003553 }
Christoph Lameterc84db232006-02-01 03:05:29 -08003554
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003555 nr_slab_pages0 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3556 if (nr_slab_pages0 > zone->min_slab_pages) {
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003557 /*
Christoph Lameter7fb2d462006-03-22 00:08:22 -08003558 * shrink_slab() does not currently allow us to determine how
Christoph Lameter0ff38492006-09-25 23:31:52 -07003559 * many pages were freed in this zone. So we take the current
3560 * number of slab pages and shake the slab until it is reduced
3561 * by the same nr_pages that we used for reclaiming unmapped
3562 * pages.
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003563 */
Dave Chinner0ce3d742013-08-28 10:18:03 +10003564 nodes_clear(shrink.nodes_to_scan);
3565 node_set(zone_to_nid(zone), shrink.nodes_to_scan);
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003566 for (;;) {
3567 unsigned long lru_pages = zone_reclaimable_pages(zone);
3568
3569 /* No reclaimable slab or very low memory pressure */
Ying Han1495f232011-05-24 17:12:27 -07003570 if (!shrink_slab(&shrink, sc.nr_scanned, lru_pages))
KOSAKI Motohiro4dc4b3d2010-08-09 17:19:54 -07003571 break;
3572
3573 /* Freed enough memory */
3574 nr_slab_pages1 = zone_page_state(zone,
3575 NR_SLAB_RECLAIMABLE);
3576 if (nr_slab_pages1 + nr_pages <= nr_slab_pages0)
3577 break;
3578 }
Christoph Lameter83e33a42006-09-25 23:31:53 -07003579
3580 /*
3581 * Update nr_reclaimed by the number of slab pages we
3582 * reclaimed from this zone.
3583 */
KOSAKI Motohiro15748042010-08-09 17:19:50 -07003584 nr_slab_pages1 = zone_page_state(zone, NR_SLAB_RECLAIMABLE);
3585 if (nr_slab_pages1 < nr_slab_pages0)
3586 sc.nr_reclaimed += nr_slab_pages0 - nr_slab_pages1;
Christoph Lameter2a16e3f2006-02-01 03:05:35 -08003587 }
3588
Christoph Lameter9eeff232006-01-18 17:42:31 -08003589 p->reclaim_state = NULL;
Christoph Lameterd4f77962006-02-24 13:04:22 -08003590 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
KOSAKI Motohiro76ca5422010-03-05 13:41:47 -08003591 lockdep_clear_current_reclaim_state();
Rik van Riela79311c2009-01-06 14:40:01 -08003592 return sc.nr_reclaimed >= nr_pages;
Christoph Lameter9eeff232006-01-18 17:42:31 -08003593}
Andrew Morton179e9632006-03-22 00:08:18 -08003594
3595int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3596{
Andrew Morton179e9632006-03-22 00:08:18 -08003597 int node_id;
David Rientjesd773ed62007-10-16 23:26:01 -07003598 int ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003599
3600 /*
Christoph Lameter0ff38492006-09-25 23:31:52 -07003601 * Zone reclaim reclaims unmapped file backed pages and
3602 * slab pages if we are over the defined limits.
Christoph Lameter34aa1332006-06-30 01:55:37 -07003603 *
Christoph Lameter96146342006-07-03 00:24:13 -07003604 * A small portion of unmapped file backed pages is needed for
3605 * file I/O otherwise pages read by file I/O will be immediately
3606 * thrown out if the zone is overallocated. So we do not reclaim
3607 * if less than a specified percentage of the zone is used by
3608 * unmapped file backed pages.
Andrew Morton179e9632006-03-22 00:08:18 -08003609 */
Mel Gorman90afa5d2009-06-16 15:33:20 -07003610 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3611 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
Mel Gormanfa5e0842009-06-16 15:33:22 -07003612 return ZONE_RECLAIM_FULL;
Andrew Morton179e9632006-03-22 00:08:18 -08003613
Lisa Du6e543d52013-09-11 14:22:36 -07003614 if (!zone_reclaimable(zone))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003615 return ZONE_RECLAIM_FULL;
David Rientjesd773ed62007-10-16 23:26:01 -07003616
Andrew Morton179e9632006-03-22 00:08:18 -08003617 /*
David Rientjesd773ed62007-10-16 23:26:01 -07003618 * Do not scan if the allocation should not be delayed.
Andrew Morton179e9632006-03-22 00:08:18 -08003619 */
David Rientjesd773ed62007-10-16 23:26:01 -07003620 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003621 return ZONE_RECLAIM_NOSCAN;
Andrew Morton179e9632006-03-22 00:08:18 -08003622
3623 /*
3624 * Only run zone reclaim on the local zone or on zones that do not
3625 * have associated processors. This will favor the local processor
3626 * over remote processors and spread off node memory allocations
3627 * as wide as possible.
3628 */
Christoph Lameter89fa3022006-09-25 23:31:55 -07003629 node_id = zone_to_nid(zone);
Christoph Lameter37c07082007-10-16 01:25:36 -07003630 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
Mel Gormanfa5e0842009-06-16 15:33:22 -07003631 return ZONE_RECLAIM_NOSCAN;
David Rientjesd773ed62007-10-16 23:26:01 -07003632
3633 if (zone_test_and_set_flag(zone, ZONE_RECLAIM_LOCKED))
Mel Gormanfa5e0842009-06-16 15:33:22 -07003634 return ZONE_RECLAIM_NOSCAN;
3635
David Rientjesd773ed62007-10-16 23:26:01 -07003636 ret = __zone_reclaim(zone, gfp_mask, order);
3637 zone_clear_flag(zone, ZONE_RECLAIM_LOCKED);
3638
Mel Gorman24cf725182009-06-16 15:33:23 -07003639 if (!ret)
3640 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3641
David Rientjesd773ed62007-10-16 23:26:01 -07003642 return ret;
Andrew Morton179e9632006-03-22 00:08:18 -08003643}
Christoph Lameter9eeff232006-01-18 17:42:31 -08003644#endif
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003645
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003646/*
3647 * page_evictable - test whether a page is evictable
3648 * @page: the page to test
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003649 *
3650 * Test whether page is evictable--i.e., should be placed on active/inactive
Hugh Dickins39b5f292012-10-08 16:33:18 -07003651 * lists vs unevictable list.
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003652 *
3653 * Reasons page might not be evictable:
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003654 * (1) page's mapping marked unevictable
Nick Pigginb291f002008-10-18 20:26:44 -07003655 * (2) page is part of an mlocked VMA
Lee Schermerhornba9ddf42008-10-18 20:26:42 -07003656 *
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003657 */
Hugh Dickins39b5f292012-10-08 16:33:18 -07003658int page_evictable(struct page *page)
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003659{
Hugh Dickins39b5f292012-10-08 16:33:18 -07003660 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
Lee Schermerhorn894bc312008-10-18 20:26:39 -07003661}
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003662
Hugh Dickins850465792012-01-20 14:34:19 -08003663#ifdef CONFIG_SHMEM
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003664/**
Hugh Dickins24513262012-01-20 14:34:21 -08003665 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3666 * @pages: array of pages to check
3667 * @nr_pages: number of pages to check
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003668 *
Hugh Dickins24513262012-01-20 14:34:21 -08003669 * Checks pages for evictability and moves them to the appropriate lru list.
Hugh Dickins850465792012-01-20 14:34:19 -08003670 *
3671 * This function is only used for SysV IPC SHM_UNLOCK.
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003672 */
Hugh Dickins24513262012-01-20 14:34:21 -08003673void check_move_unevictable_pages(struct page **pages, int nr_pages)
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003674{
Johannes Weiner925b7672012-01-12 17:18:15 -08003675 struct lruvec *lruvec;
Hugh Dickins24513262012-01-20 14:34:21 -08003676 struct zone *zone = NULL;
3677 int pgscanned = 0;
3678 int pgrescued = 0;
3679 int i;
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003680
Hugh Dickins24513262012-01-20 14:34:21 -08003681 for (i = 0; i < nr_pages; i++) {
3682 struct page *page = pages[i];
3683 struct zone *pagezone;
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003684
Hugh Dickins24513262012-01-20 14:34:21 -08003685 pgscanned++;
3686 pagezone = page_zone(page);
3687 if (pagezone != zone) {
3688 if (zone)
3689 spin_unlock_irq(&zone->lru_lock);
3690 zone = pagezone;
3691 spin_lock_irq(&zone->lru_lock);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003692 }
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003693 lruvec = mem_cgroup_page_lruvec(page, zone);
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003694
Hugh Dickins24513262012-01-20 14:34:21 -08003695 if (!PageLRU(page) || !PageUnevictable(page))
3696 continue;
3697
Hugh Dickins39b5f292012-10-08 16:33:18 -07003698 if (page_evictable(page)) {
Hugh Dickins24513262012-01-20 14:34:21 -08003699 enum lru_list lru = page_lru_base_type(page);
3700
Sasha Levin309381fea2014-01-23 15:52:54 -08003701 VM_BUG_ON_PAGE(PageActive(page), page);
Hugh Dickins24513262012-01-20 14:34:21 -08003702 ClearPageUnevictable(page);
Hugh Dickinsfa9add62012-05-29 15:07:09 -07003703 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3704 add_page_to_lru_list(page, lruvec, lru);
Hugh Dickins24513262012-01-20 14:34:21 -08003705 pgrescued++;
3706 }
Lee Schermerhorn89e004ea2008-10-18 20:26:43 -07003707 }
Hugh Dickins24513262012-01-20 14:34:21 -08003708
3709 if (zone) {
3710 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3711 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3712 spin_unlock_irq(&zone->lru_lock);
3713 }
Hugh Dickins850465792012-01-20 14:34:19 -08003714}
3715#endif /* CONFIG_SHMEM */
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003716
Johannes Weiner264e56d2011-10-31 17:09:13 -07003717static void warn_scan_unevictable_pages(void)
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003718{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003719 printk_once(KERN_WARNING
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003720 "%s: The scan_unevictable_pages sysctl/node-interface has been "
Johannes Weiner264e56d2011-10-31 17:09:13 -07003721 "disabled for lack of a legitimate use case. If you have "
KOSAKI Motohiro25bd91b2012-01-10 15:07:40 -08003722 "one, please send an email to linux-mm@kvack.org.\n",
3723 current->comm);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003724}
3725
3726/*
3727 * scan_unevictable_pages [vm] sysctl handler. On demand re-scan of
3728 * all nodes' unevictable lists for evictable pages
3729 */
3730unsigned long scan_unevictable_pages;
3731
3732int scan_unevictable_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003733 void __user *buffer,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003734 size_t *length, loff_t *ppos)
3735{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003736 warn_scan_unevictable_pages();
Alexey Dobriyan8d65af72009-09-23 15:57:19 -07003737 proc_doulongvec_minmax(table, write, buffer, length, ppos);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003738 scan_unevictable_pages = 0;
3739 return 0;
3740}
3741
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003742#ifdef CONFIG_NUMA
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003743/*
3744 * per node 'scan_unevictable_pages' attribute. On demand re-scan of
3745 * a specified node's per zone unevictable lists for evictable pages.
3746 */
3747
Kay Sievers10fbcf42011-12-21 14:48:43 -08003748static ssize_t read_scan_unevictable_node(struct device *dev,
3749 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003750 char *buf)
3751{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003752 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003753 return sprintf(buf, "0\n"); /* always zero; should fit... */
3754}
3755
Kay Sievers10fbcf42011-12-21 14:48:43 -08003756static ssize_t write_scan_unevictable_node(struct device *dev,
3757 struct device_attribute *attr,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003758 const char *buf, size_t count)
3759{
Johannes Weiner264e56d2011-10-31 17:09:13 -07003760 warn_scan_unevictable_pages();
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003761 return 1;
3762}
3763
3764
Kay Sievers10fbcf42011-12-21 14:48:43 -08003765static DEVICE_ATTR(scan_unevictable_pages, S_IRUGO | S_IWUSR,
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003766 read_scan_unevictable_node,
3767 write_scan_unevictable_node);
3768
3769int scan_unevictable_register_node(struct node *node)
3770{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003771 return device_create_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003772}
3773
3774void scan_unevictable_unregister_node(struct node *node)
3775{
Kay Sievers10fbcf42011-12-21 14:48:43 -08003776 device_remove_file(&node->dev, &dev_attr_scan_unevictable_pages);
Lee Schermerhornaf936a12008-10-18 20:26:53 -07003777}
Thadeu Lima de Souza Cascardoe4455ab2010-10-26 14:21:28 -07003778#endif