]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - mm/vmscan.c
Merge remote-tracking branches 'asoc/topic/inntel', 'asoc/topic/input', 'asoc/topic...
[mirror_ubuntu-bionic-kernel.git] / mm / vmscan.c
CommitLineData
1da177e4
LT
1/*
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
b1de0d13
MH
14#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
1da177e4
LT
16#include <linux/mm.h>
17#include <linux/module.h>
5a0e3ad6 18#include <linux/gfp.h>
1da177e4
LT
19#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/pagemap.h>
22#include <linux/init.h>
23#include <linux/highmem.h>
70ddf637 24#include <linux/vmpressure.h>
e129b5c2 25#include <linux/vmstat.h>
1da177e4
LT
26#include <linux/file.h>
27#include <linux/writeback.h>
28#include <linux/blkdev.h>
29#include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
1da177e4
LT
32#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
3e7d3449 37#include <linux/compaction.h>
1da177e4
LT
38#include <linux/notifier.h>
39#include <linux/rwsem.h>
248a0301 40#include <linux/delay.h>
3218ae14 41#include <linux/kthread.h>
7dfb7103 42#include <linux/freezer.h>
66e1707b 43#include <linux/memcontrol.h>
873b4771 44#include <linux/delayacct.h>
af936a16 45#include <linux/sysctl.h>
929bea7c 46#include <linux/oom.h>
268bb0ce 47#include <linux/prefetch.h>
b1de0d13 48#include <linux/printk.h>
f9fe48be 49#include <linux/dax.h>
1da177e4
LT
50
51#include <asm/tlbflush.h>
52#include <asm/div64.h>
53
54#include <linux/swapops.h>
117aad1e 55#include <linux/balloon_compaction.h>
1da177e4 56
0f8053a5
NP
57#include "internal.h"
58
33906bc5
MG
59#define CREATE_TRACE_POINTS
60#include <trace/events/vmscan.h>
61
1da177e4 62struct scan_control {
22fba335
KM
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
65
1da177e4 66 /* This context's GFP mask */
6daa0e28 67 gfp_t gfp_mask;
1da177e4 68
ee814fe2 69 /* Allocation order */
5ad333eb 70 int order;
66e1707b 71
ee814fe2
JW
72 /*
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
74 * are scanned.
75 */
76 nodemask_t *nodemask;
9e3b2f8c 77
f16015fb
JW
78 /*
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
81 */
82 struct mem_cgroup *target_mem_cgroup;
66e1707b 83
ee814fe2
JW
84 /* Scan (total_size >> priority) pages at once */
85 int priority;
86
b2e18757
MG
87 /* The highest zone to isolate pages for reclaim from */
88 enum zone_type reclaim_idx;
89
ee814fe2
JW
90 unsigned int may_writepage:1;
91
92 /* Can mapped pages be reclaimed? */
93 unsigned int may_unmap:1;
94
95 /* Can pages be swapped as part of reclaim? */
96 unsigned int may_swap:1;
97
241994ed
JW
98 /* Can cgroups be reclaimed below their normal consumption range? */
99 unsigned int may_thrash:1;
100
ee814fe2
JW
101 unsigned int hibernation_mode:1;
102
103 /* One of the zones is ready for compaction */
104 unsigned int compaction_ready:1;
105
106 /* Incremented by the number of inactive pages that were scanned */
107 unsigned long nr_scanned;
108
109 /* Number of pages freed so far during a call to shrink_zones() */
110 unsigned long nr_reclaimed;
1da177e4
LT
111};
112
1da177e4
LT
113#ifdef ARCH_HAS_PREFETCH
114#define prefetch_prev_lru_page(_page, _base, _field) \
115 do { \
116 if ((_page)->lru.prev != _base) { \
117 struct page *prev; \
118 \
119 prev = lru_to_page(&(_page->lru)); \
120 prefetch(&prev->_field); \
121 } \
122 } while (0)
123#else
124#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
125#endif
126
127#ifdef ARCH_HAS_PREFETCHW
128#define prefetchw_prev_lru_page(_page, _base, _field) \
129 do { \
130 if ((_page)->lru.prev != _base) { \
131 struct page *prev; \
132 \
133 prev = lru_to_page(&(_page->lru)); \
134 prefetchw(&prev->_field); \
135 } \
136 } while (0)
137#else
138#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
139#endif
140
141/*
142 * From 0 .. 100. Higher means more swappy.
143 */
144int vm_swappiness = 60;
d0480be4
WSH
145/*
146 * The total number of pages which are beyond the high watermark within all
147 * zones.
148 */
149unsigned long vm_total_pages;
1da177e4
LT
150
151static LIST_HEAD(shrinker_list);
152static DECLARE_RWSEM(shrinker_rwsem);
153
c255a458 154#ifdef CONFIG_MEMCG
89b5fae5
JW
155static bool global_reclaim(struct scan_control *sc)
156{
f16015fb 157 return !sc->target_mem_cgroup;
89b5fae5 158}
97c9341f
TH
159
160/**
161 * sane_reclaim - is the usual dirty throttling mechanism operational?
162 * @sc: scan_control in question
163 *
164 * The normal page dirty throttling mechanism in balance_dirty_pages() is
165 * completely broken with the legacy memcg and direct stalling in
166 * shrink_page_list() is used for throttling instead, which lacks all the
167 * niceties such as fairness, adaptive pausing, bandwidth proportional
168 * allocation and configurability.
169 *
170 * This function tests whether the vmscan currently in progress can assume
171 * that the normal dirty throttling mechanism is operational.
172 */
173static bool sane_reclaim(struct scan_control *sc)
174{
175 struct mem_cgroup *memcg = sc->target_mem_cgroup;
176
177 if (!memcg)
178 return true;
179#ifdef CONFIG_CGROUP_WRITEBACK
69234ace 180 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
97c9341f
TH
181 return true;
182#endif
183 return false;
184}
91a45470 185#else
89b5fae5
JW
186static bool global_reclaim(struct scan_control *sc)
187{
188 return true;
189}
97c9341f
TH
190
191static bool sane_reclaim(struct scan_control *sc)
192{
193 return true;
194}
91a45470
KH
195#endif
196
5a1c84b4
MG
197/*
198 * This misses isolated pages which are not accounted for to save counters.
199 * As the data only determines if reclaim or compaction continues, it is
200 * not expected that isolated pages will be a dominating factor.
201 */
202unsigned long zone_reclaimable_pages(struct zone *zone)
203{
204 unsigned long nr;
205
206 nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) +
207 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE);
208 if (get_nr_swap_pages() > 0)
209 nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) +
210 zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON);
211
212 return nr;
213}
214
599d0c95
MG
215unsigned long pgdat_reclaimable_pages(struct pglist_data *pgdat)
216{
217 unsigned long nr;
218
219 nr = node_page_state_snapshot(pgdat, NR_ACTIVE_FILE) +
220 node_page_state_snapshot(pgdat, NR_INACTIVE_FILE) +
221 node_page_state_snapshot(pgdat, NR_ISOLATED_FILE);
6e543d57
LD
222
223 if (get_nr_swap_pages() > 0)
599d0c95
MG
224 nr += node_page_state_snapshot(pgdat, NR_ACTIVE_ANON) +
225 node_page_state_snapshot(pgdat, NR_INACTIVE_ANON) +
226 node_page_state_snapshot(pgdat, NR_ISOLATED_ANON);
6e543d57
LD
227
228 return nr;
229}
230
599d0c95 231bool pgdat_reclaimable(struct pglist_data *pgdat)
6e543d57 232{
599d0c95
MG
233 return node_page_state_snapshot(pgdat, NR_PAGES_SCANNED) <
234 pgdat_reclaimable_pages(pgdat) * 6;
6e543d57
LD
235}
236
23047a96 237unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru)
c9f299d9 238{
c3c787e8 239 if (!mem_cgroup_disabled())
4d7dcca2 240 return mem_cgroup_get_lru_size(lruvec, lru);
a3d8e054 241
599d0c95 242 return node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru);
c9f299d9
KM
243}
244
1da177e4 245/*
1d3d4437 246 * Add a shrinker callback to be called from the vm.
1da177e4 247 */
1d3d4437 248int register_shrinker(struct shrinker *shrinker)
1da177e4 249{
1d3d4437
GC
250 size_t size = sizeof(*shrinker->nr_deferred);
251
1d3d4437
GC
252 if (shrinker->flags & SHRINKER_NUMA_AWARE)
253 size *= nr_node_ids;
254
255 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
256 if (!shrinker->nr_deferred)
257 return -ENOMEM;
258
8e1f936b
RR
259 down_write(&shrinker_rwsem);
260 list_add_tail(&shrinker->list, &shrinker_list);
261 up_write(&shrinker_rwsem);
1d3d4437 262 return 0;
1da177e4 263}
8e1f936b 264EXPORT_SYMBOL(register_shrinker);
1da177e4
LT
265
266/*
267 * Remove one
268 */
8e1f936b 269void unregister_shrinker(struct shrinker *shrinker)
1da177e4
LT
270{
271 down_write(&shrinker_rwsem);
272 list_del(&shrinker->list);
273 up_write(&shrinker_rwsem);
ae393321 274 kfree(shrinker->nr_deferred);
1da177e4 275}
8e1f936b 276EXPORT_SYMBOL(unregister_shrinker);
1da177e4
LT
277
278#define SHRINK_BATCH 128
1d3d4437 279
cb731d6c
VD
280static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
281 struct shrinker *shrinker,
282 unsigned long nr_scanned,
283 unsigned long nr_eligible)
1d3d4437
GC
284{
285 unsigned long freed = 0;
286 unsigned long long delta;
287 long total_scan;
d5bc5fd3 288 long freeable;
1d3d4437
GC
289 long nr;
290 long new_nr;
291 int nid = shrinkctl->nid;
292 long batch_size = shrinker->batch ? shrinker->batch
293 : SHRINK_BATCH;
294
d5bc5fd3
VD
295 freeable = shrinker->count_objects(shrinker, shrinkctl);
296 if (freeable == 0)
1d3d4437
GC
297 return 0;
298
299 /*
300 * copy the current shrinker scan count into a local variable
301 * and zero it so that other concurrent shrinker invocations
302 * don't also do this scanning work.
303 */
304 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
305
306 total_scan = nr;
6b4f7799 307 delta = (4 * nr_scanned) / shrinker->seeks;
d5bc5fd3 308 delta *= freeable;
6b4f7799 309 do_div(delta, nr_eligible + 1);
1d3d4437
GC
310 total_scan += delta;
311 if (total_scan < 0) {
8612c663 312 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
a0b02131 313 shrinker->scan_objects, total_scan);
d5bc5fd3 314 total_scan = freeable;
1d3d4437
GC
315 }
316
317 /*
318 * We need to avoid excessive windup on filesystem shrinkers
319 * due to large numbers of GFP_NOFS allocations causing the
320 * shrinkers to return -1 all the time. This results in a large
321 * nr being built up so when a shrink that can do some work
322 * comes along it empties the entire cache due to nr >>>
d5bc5fd3 323 * freeable. This is bad for sustaining a working set in
1d3d4437
GC
324 * memory.
325 *
326 * Hence only allow the shrinker to scan the entire cache when
327 * a large delta change is calculated directly.
328 */
d5bc5fd3
VD
329 if (delta < freeable / 4)
330 total_scan = min(total_scan, freeable / 2);
1d3d4437
GC
331
332 /*
333 * Avoid risking looping forever due to too large nr value:
334 * never try to free more than twice the estimate number of
335 * freeable entries.
336 */
d5bc5fd3
VD
337 if (total_scan > freeable * 2)
338 total_scan = freeable * 2;
1d3d4437
GC
339
340 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
6b4f7799
JW
341 nr_scanned, nr_eligible,
342 freeable, delta, total_scan);
1d3d4437 343
0b1fb40a
VD
344 /*
345 * Normally, we should not scan less than batch_size objects in one
346 * pass to avoid too frequent shrinker calls, but if the slab has less
347 * than batch_size objects in total and we are really tight on memory,
348 * we will try to reclaim all available objects, otherwise we can end
349 * up failing allocations although there are plenty of reclaimable
350 * objects spread over several slabs with usage less than the
351 * batch_size.
352 *
353 * We detect the "tight on memory" situations by looking at the total
354 * number of objects we want to scan (total_scan). If it is greater
d5bc5fd3 355 * than the total number of objects on slab (freeable), we must be
0b1fb40a
VD
356 * scanning at high prio and therefore should try to reclaim as much as
357 * possible.
358 */
359 while (total_scan >= batch_size ||
d5bc5fd3 360 total_scan >= freeable) {
a0b02131 361 unsigned long ret;
0b1fb40a 362 unsigned long nr_to_scan = min(batch_size, total_scan);
1d3d4437 363
0b1fb40a 364 shrinkctl->nr_to_scan = nr_to_scan;
a0b02131
DC
365 ret = shrinker->scan_objects(shrinker, shrinkctl);
366 if (ret == SHRINK_STOP)
367 break;
368 freed += ret;
1d3d4437 369
0b1fb40a
VD
370 count_vm_events(SLABS_SCANNED, nr_to_scan);
371 total_scan -= nr_to_scan;
1d3d4437
GC
372
373 cond_resched();
374 }
375
376 /*
377 * move the unused scan count back into the shrinker in a
378 * manner that handles concurrent updates. If we exhausted the
379 * scan, there is no need to do an update.
380 */
381 if (total_scan > 0)
382 new_nr = atomic_long_add_return(total_scan,
383 &shrinker->nr_deferred[nid]);
384 else
385 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
386
df9024a8 387 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
1d3d4437 388 return freed;
1495f230
YH
389}
390
6b4f7799 391/**
cb731d6c 392 * shrink_slab - shrink slab caches
6b4f7799
JW
393 * @gfp_mask: allocation context
394 * @nid: node whose slab caches to target
cb731d6c 395 * @memcg: memory cgroup whose slab caches to target
6b4f7799
JW
396 * @nr_scanned: pressure numerator
397 * @nr_eligible: pressure denominator
1da177e4 398 *
6b4f7799 399 * Call the shrink functions to age shrinkable caches.
1da177e4 400 *
6b4f7799
JW
401 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
402 * unaware shrinkers will receive a node id of 0 instead.
1da177e4 403 *
cb731d6c
VD
404 * @memcg specifies the memory cgroup to target. If it is not NULL,
405 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
0fc9f58a
VD
406 * objects from the memory cgroup specified. Otherwise, only unaware
407 * shrinkers are called.
cb731d6c 408 *
6b4f7799
JW
409 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
410 * the available objects should be scanned. Page reclaim for example
411 * passes the number of pages scanned and the number of pages on the
412 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
413 * when it encountered mapped pages. The ratio is further biased by
414 * the ->seeks setting of the shrink function, which indicates the
415 * cost to recreate an object relative to that of an LRU page.
b15e0905 416 *
6b4f7799 417 * Returns the number of reclaimed slab objects.
1da177e4 418 */
cb731d6c
VD
419static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
420 struct mem_cgroup *memcg,
421 unsigned long nr_scanned,
422 unsigned long nr_eligible)
1da177e4
LT
423{
424 struct shrinker *shrinker;
24f7c6b9 425 unsigned long freed = 0;
1da177e4 426
0fc9f58a 427 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
cb731d6c
VD
428 return 0;
429
6b4f7799
JW
430 if (nr_scanned == 0)
431 nr_scanned = SWAP_CLUSTER_MAX;
1da177e4 432
f06590bd 433 if (!down_read_trylock(&shrinker_rwsem)) {
24f7c6b9
DC
434 /*
435 * If we would return 0, our callers would understand that we
436 * have nothing else to shrink and give up trying. By returning
437 * 1 we keep it going and assume we'll be able to shrink next
438 * time.
439 */
440 freed = 1;
f06590bd
MK
441 goto out;
442 }
1da177e4
LT
443
444 list_for_each_entry(shrinker, &shrinker_list, list) {
6b4f7799
JW
445 struct shrink_control sc = {
446 .gfp_mask = gfp_mask,
447 .nid = nid,
cb731d6c 448 .memcg = memcg,
6b4f7799 449 };
ec97097b 450
0fc9f58a
VD
451 /*
452 * If kernel memory accounting is disabled, we ignore
453 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
454 * passing NULL for memcg.
455 */
456 if (memcg_kmem_enabled() &&
457 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
cb731d6c
VD
458 continue;
459
6b4f7799
JW
460 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
461 sc.nid = 0;
1da177e4 462
cb731d6c 463 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
1da177e4 464 }
6b4f7799 465
1da177e4 466 up_read(&shrinker_rwsem);
f06590bd
MK
467out:
468 cond_resched();
24f7c6b9 469 return freed;
1da177e4
LT
470}
471
cb731d6c
VD
472void drop_slab_node(int nid)
473{
474 unsigned long freed;
475
476 do {
477 struct mem_cgroup *memcg = NULL;
478
479 freed = 0;
480 do {
481 freed += shrink_slab(GFP_KERNEL, nid, memcg,
482 1000, 1000);
483 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
484 } while (freed > 10);
485}
486
487void drop_slab(void)
488{
489 int nid;
490
491 for_each_online_node(nid)
492 drop_slab_node(nid);
493}
494
1da177e4
LT
495static inline int is_page_cache_freeable(struct page *page)
496{
ceddc3a5
JW
497 /*
498 * A freeable page cache page is referenced only by the caller
499 * that isolated the page, the page cache radix tree and
500 * optional buffer heads at page->private.
501 */
edcf4748 502 return page_count(page) - page_has_private(page) == 2;
1da177e4
LT
503}
504
703c2708 505static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
1da177e4 506{
930d9152 507 if (current->flags & PF_SWAPWRITE)
1da177e4 508 return 1;
703c2708 509 if (!inode_write_congested(inode))
1da177e4 510 return 1;
703c2708 511 if (inode_to_bdi(inode) == current->backing_dev_info)
1da177e4
LT
512 return 1;
513 return 0;
514}
515
516/*
517 * We detected a synchronous write error writing a page out. Probably
518 * -ENOSPC. We need to propagate that into the address_space for a subsequent
519 * fsync(), msync() or close().
520 *
521 * The tricky part is that after writepage we cannot touch the mapping: nothing
522 * prevents it from being freed up. But we have a ref on the page and once
523 * that page is locked, the mapping is pinned.
524 *
525 * We're allowed to run sleeping lock_page() here because we know the caller has
526 * __GFP_FS.
527 */
528static void handle_write_error(struct address_space *mapping,
529 struct page *page, int error)
530{
7eaceacc 531 lock_page(page);
3e9f45bd
GC
532 if (page_mapping(page) == mapping)
533 mapping_set_error(mapping, error);
1da177e4
LT
534 unlock_page(page);
535}
536
04e62a29
CL
537/* possible outcome of pageout() */
538typedef enum {
539 /* failed to write page out, page is locked */
540 PAGE_KEEP,
541 /* move page to the active list, page is locked */
542 PAGE_ACTIVATE,
543 /* page has been sent to the disk successfully, page is unlocked */
544 PAGE_SUCCESS,
545 /* page is clean and locked */
546 PAGE_CLEAN,
547} pageout_t;
548
1da177e4 549/*
1742f19f
AM
550 * pageout is called by shrink_page_list() for each dirty page.
551 * Calls ->writepage().
1da177e4 552 */
c661b078 553static pageout_t pageout(struct page *page, struct address_space *mapping,
7d3579e8 554 struct scan_control *sc)
1da177e4
LT
555{
556 /*
557 * If the page is dirty, only perform writeback if that write
558 * will be non-blocking. To prevent this allocation from being
559 * stalled by pagecache activity. But note that there may be
560 * stalls if we need to run get_block(). We could test
561 * PagePrivate for that.
562 *
8174202b 563 * If this process is currently in __generic_file_write_iter() against
1da177e4
LT
564 * this page's queue, we can perform writeback even if that
565 * will block.
566 *
567 * If the page is swapcache, write it back even if that would
568 * block, for some throttling. This happens by accident, because
569 * swap_backing_dev_info is bust: it doesn't reflect the
570 * congestion state of the swapdevs. Easy to fix, if needed.
1da177e4
LT
571 */
572 if (!is_page_cache_freeable(page))
573 return PAGE_KEEP;
574 if (!mapping) {
575 /*
576 * Some data journaling orphaned pages can have
577 * page->mapping == NULL while being dirty with clean buffers.
578 */
266cf658 579 if (page_has_private(page)) {
1da177e4
LT
580 if (try_to_free_buffers(page)) {
581 ClearPageDirty(page);
b1de0d13 582 pr_info("%s: orphaned page\n", __func__);
1da177e4
LT
583 return PAGE_CLEAN;
584 }
585 }
586 return PAGE_KEEP;
587 }
588 if (mapping->a_ops->writepage == NULL)
589 return PAGE_ACTIVATE;
703c2708 590 if (!may_write_to_inode(mapping->host, sc))
1da177e4
LT
591 return PAGE_KEEP;
592
593 if (clear_page_dirty_for_io(page)) {
594 int res;
595 struct writeback_control wbc = {
596 .sync_mode = WB_SYNC_NONE,
597 .nr_to_write = SWAP_CLUSTER_MAX,
111ebb6e
OH
598 .range_start = 0,
599 .range_end = LLONG_MAX,
1da177e4
LT
600 .for_reclaim = 1,
601 };
602
603 SetPageReclaim(page);
604 res = mapping->a_ops->writepage(page, &wbc);
605 if (res < 0)
606 handle_write_error(mapping, page, res);
994fc28c 607 if (res == AOP_WRITEPAGE_ACTIVATE) {
1da177e4
LT
608 ClearPageReclaim(page);
609 return PAGE_ACTIVATE;
610 }
c661b078 611
1da177e4
LT
612 if (!PageWriteback(page)) {
613 /* synchronous write or broken a_ops? */
614 ClearPageReclaim(page);
615 }
3aa23851 616 trace_mm_vmscan_writepage(page);
c4a25635 617 inc_node_page_state(page, NR_VMSCAN_WRITE);
1da177e4
LT
618 return PAGE_SUCCESS;
619 }
620
621 return PAGE_CLEAN;
622}
623
a649fd92 624/*
e286781d
NP
625 * Same as remove_mapping, but if the page is removed from the mapping, it
626 * gets returned with a refcount of 0.
a649fd92 627 */
a528910e
JW
628static int __remove_mapping(struct address_space *mapping, struct page *page,
629 bool reclaimed)
49d2e9cc 630{
c4843a75 631 unsigned long flags;
c4843a75 632
28e4d965
NP
633 BUG_ON(!PageLocked(page));
634 BUG_ON(mapping != page_mapping(page));
49d2e9cc 635
c4843a75 636 spin_lock_irqsave(&mapping->tree_lock, flags);
49d2e9cc 637 /*
0fd0e6b0
NP
638 * The non racy check for a busy page.
639 *
640 * Must be careful with the order of the tests. When someone has
641 * a ref to the page, it may be possible that they dirty it then
642 * drop the reference. So if PageDirty is tested before page_count
643 * here, then the following race may occur:
644 *
645 * get_user_pages(&page);
646 * [user mapping goes away]
647 * write_to(page);
648 * !PageDirty(page) [good]
649 * SetPageDirty(page);
650 * put_page(page);
651 * !page_count(page) [good, discard it]
652 *
653 * [oops, our write_to data is lost]
654 *
655 * Reversing the order of the tests ensures such a situation cannot
656 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
0139aa7b 657 * load is not satisfied before that of page->_refcount.
0fd0e6b0
NP
658 *
659 * Note that if SetPageDirty is always performed via set_page_dirty,
660 * and thus under tree_lock, then this ordering is not required.
49d2e9cc 661 */
fe896d18 662 if (!page_ref_freeze(page, 2))
49d2e9cc 663 goto cannot_free;
e286781d
NP
664 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
665 if (unlikely(PageDirty(page))) {
fe896d18 666 page_ref_unfreeze(page, 2);
49d2e9cc 667 goto cannot_free;
e286781d 668 }
49d2e9cc
CL
669
670 if (PageSwapCache(page)) {
671 swp_entry_t swap = { .val = page_private(page) };
0a31bc97 672 mem_cgroup_swapout(page, swap);
49d2e9cc 673 __delete_from_swap_cache(page);
c4843a75 674 spin_unlock_irqrestore(&mapping->tree_lock, flags);
0a31bc97 675 swapcache_free(swap);
e286781d 676 } else {
6072d13c 677 void (*freepage)(struct page *);
a528910e 678 void *shadow = NULL;
6072d13c
LT
679
680 freepage = mapping->a_ops->freepage;
a528910e
JW
681 /*
682 * Remember a shadow entry for reclaimed file cache in
683 * order to detect refaults, thus thrashing, later on.
684 *
685 * But don't store shadows in an address space that is
686 * already exiting. This is not just an optizimation,
687 * inode reclaim needs to empty out the radix tree or
688 * the nodes are lost. Don't plant shadows behind its
689 * back.
f9fe48be
RZ
690 *
691 * We also don't store shadows for DAX mappings because the
692 * only page cache pages found in these are zero pages
693 * covering holes, and because we don't want to mix DAX
694 * exceptional entries and shadow exceptional entries in the
695 * same page_tree.
a528910e
JW
696 */
697 if (reclaimed && page_is_file_cache(page) &&
f9fe48be 698 !mapping_exiting(mapping) && !dax_mapping(mapping))
a528910e 699 shadow = workingset_eviction(mapping, page);
62cccb8c 700 __delete_from_page_cache(page, shadow);
c4843a75 701 spin_unlock_irqrestore(&mapping->tree_lock, flags);
6072d13c
LT
702
703 if (freepage != NULL)
704 freepage(page);
49d2e9cc
CL
705 }
706
49d2e9cc
CL
707 return 1;
708
709cannot_free:
c4843a75 710 spin_unlock_irqrestore(&mapping->tree_lock, flags);
49d2e9cc
CL
711 return 0;
712}
713
e286781d
NP
714/*
715 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
716 * someone else has a ref on the page, abort and return 0. If it was
717 * successfully detached, return 1. Assumes the caller has a single ref on
718 * this page.
719 */
720int remove_mapping(struct address_space *mapping, struct page *page)
721{
a528910e 722 if (__remove_mapping(mapping, page, false)) {
e286781d
NP
723 /*
724 * Unfreezing the refcount with 1 rather than 2 effectively
725 * drops the pagecache ref for us without requiring another
726 * atomic operation.
727 */
fe896d18 728 page_ref_unfreeze(page, 1);
e286781d
NP
729 return 1;
730 }
731 return 0;
732}
733
894bc310
LS
734/**
735 * putback_lru_page - put previously isolated page onto appropriate LRU list
736 * @page: page to be put back to appropriate lru list
737 *
738 * Add previously isolated @page to appropriate LRU list.
739 * Page may still be unevictable for other reasons.
740 *
741 * lru_lock must not be held, interrupts must be enabled.
742 */
894bc310
LS
743void putback_lru_page(struct page *page)
744{
0ec3b74c 745 bool is_unevictable;
bbfd28ee 746 int was_unevictable = PageUnevictable(page);
894bc310 747
309381fe 748 VM_BUG_ON_PAGE(PageLRU(page), page);
894bc310
LS
749
750redo:
751 ClearPageUnevictable(page);
752
39b5f29a 753 if (page_evictable(page)) {
894bc310
LS
754 /*
755 * For evictable pages, we can use the cache.
756 * In event of a race, worst case is we end up with an
757 * unevictable page on [in]active list.
758 * We know how to handle that.
759 */
0ec3b74c 760 is_unevictable = false;
c53954a0 761 lru_cache_add(page);
894bc310
LS
762 } else {
763 /*
764 * Put unevictable pages directly on zone's unevictable
765 * list.
766 */
0ec3b74c 767 is_unevictable = true;
894bc310 768 add_page_to_unevictable_list(page);
6a7b9548 769 /*
21ee9f39
MK
770 * When racing with an mlock or AS_UNEVICTABLE clearing
771 * (page is unlocked) make sure that if the other thread
772 * does not observe our setting of PG_lru and fails
24513264 773 * isolation/check_move_unevictable_pages,
21ee9f39 774 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
6a7b9548
JW
775 * the page back to the evictable list.
776 *
21ee9f39 777 * The other side is TestClearPageMlocked() or shmem_lock().
6a7b9548
JW
778 */
779 smp_mb();
894bc310 780 }
894bc310
LS
781
782 /*
783 * page's status can change while we move it among lru. If an evictable
784 * page is on unevictable list, it never be freed. To avoid that,
785 * check after we added it to the list, again.
786 */
0ec3b74c 787 if (is_unevictable && page_evictable(page)) {
894bc310
LS
788 if (!isolate_lru_page(page)) {
789 put_page(page);
790 goto redo;
791 }
792 /* This means someone else dropped this page from LRU
793 * So, it will be freed or putback to LRU again. There is
794 * nothing to do here.
795 */
796 }
797
0ec3b74c 798 if (was_unevictable && !is_unevictable)
bbfd28ee 799 count_vm_event(UNEVICTABLE_PGRESCUED);
0ec3b74c 800 else if (!was_unevictable && is_unevictable)
bbfd28ee
LS
801 count_vm_event(UNEVICTABLE_PGCULLED);
802
894bc310
LS
803 put_page(page); /* drop ref from isolate */
804}
805
dfc8d636
JW
806enum page_references {
807 PAGEREF_RECLAIM,
808 PAGEREF_RECLAIM_CLEAN,
64574746 809 PAGEREF_KEEP,
dfc8d636
JW
810 PAGEREF_ACTIVATE,
811};
812
813static enum page_references page_check_references(struct page *page,
814 struct scan_control *sc)
815{
64574746 816 int referenced_ptes, referenced_page;
dfc8d636 817 unsigned long vm_flags;
dfc8d636 818
c3ac9a8a
JW
819 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
820 &vm_flags);
64574746 821 referenced_page = TestClearPageReferenced(page);
dfc8d636 822
dfc8d636
JW
823 /*
824 * Mlock lost the isolation race with us. Let try_to_unmap()
825 * move the page to the unevictable list.
826 */
827 if (vm_flags & VM_LOCKED)
828 return PAGEREF_RECLAIM;
829
64574746 830 if (referenced_ptes) {
e4898273 831 if (PageSwapBacked(page))
64574746
JW
832 return PAGEREF_ACTIVATE;
833 /*
834 * All mapped pages start out with page table
835 * references from the instantiating fault, so we need
836 * to look twice if a mapped file page is used more
837 * than once.
838 *
839 * Mark it and spare it for another trip around the
840 * inactive list. Another page table reference will
841 * lead to its activation.
842 *
843 * Note: the mark is set for activated pages as well
844 * so that recently deactivated but used pages are
845 * quickly recovered.
846 */
847 SetPageReferenced(page);
848
34dbc67a 849 if (referenced_page || referenced_ptes > 1)
64574746
JW
850 return PAGEREF_ACTIVATE;
851
c909e993
KK
852 /*
853 * Activate file-backed executable pages after first usage.
854 */
855 if (vm_flags & VM_EXEC)
856 return PAGEREF_ACTIVATE;
857
64574746
JW
858 return PAGEREF_KEEP;
859 }
dfc8d636
JW
860
861 /* Reclaim if clean, defer dirty pages to writeback */
2e30244a 862 if (referenced_page && !PageSwapBacked(page))
64574746
JW
863 return PAGEREF_RECLAIM_CLEAN;
864
865 return PAGEREF_RECLAIM;
dfc8d636
JW
866}
867
e2be15f6
MG
868/* Check if a page is dirty or under writeback */
869static void page_check_dirty_writeback(struct page *page,
870 bool *dirty, bool *writeback)
871{
b4597226
MG
872 struct address_space *mapping;
873
e2be15f6
MG
874 /*
875 * Anonymous pages are not handled by flushers and must be written
876 * from reclaim context. Do not stall reclaim based on them
877 */
878 if (!page_is_file_cache(page)) {
879 *dirty = false;
880 *writeback = false;
881 return;
882 }
883
884 /* By default assume that the page flags are accurate */
885 *dirty = PageDirty(page);
886 *writeback = PageWriteback(page);
b4597226
MG
887
888 /* Verify dirty/writeback state if the filesystem supports it */
889 if (!page_has_private(page))
890 return;
891
892 mapping = page_mapping(page);
893 if (mapping && mapping->a_ops->is_dirty_writeback)
894 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
e2be15f6
MG
895}
896
1da177e4 897/*
1742f19f 898 * shrink_page_list() returns the number of reclaimed pages
1da177e4 899 */
1742f19f 900static unsigned long shrink_page_list(struct list_head *page_list,
599d0c95 901 struct pglist_data *pgdat,
f84f6e2b 902 struct scan_control *sc,
02c6de8d 903 enum ttu_flags ttu_flags,
8e950282 904 unsigned long *ret_nr_dirty,
d43006d5 905 unsigned long *ret_nr_unqueued_dirty,
8e950282 906 unsigned long *ret_nr_congested,
02c6de8d 907 unsigned long *ret_nr_writeback,
b1a6f21e 908 unsigned long *ret_nr_immediate,
02c6de8d 909 bool force_reclaim)
1da177e4
LT
910{
911 LIST_HEAD(ret_pages);
abe4c3b5 912 LIST_HEAD(free_pages);
1da177e4 913 int pgactivate = 0;
d43006d5 914 unsigned long nr_unqueued_dirty = 0;
0e093d99
MG
915 unsigned long nr_dirty = 0;
916 unsigned long nr_congested = 0;
05ff5137 917 unsigned long nr_reclaimed = 0;
92df3a72 918 unsigned long nr_writeback = 0;
b1a6f21e 919 unsigned long nr_immediate = 0;
1da177e4
LT
920
921 cond_resched();
922
1da177e4
LT
923 while (!list_empty(page_list)) {
924 struct address_space *mapping;
925 struct page *page;
926 int may_enter_fs;
02c6de8d 927 enum page_references references = PAGEREF_RECLAIM_CLEAN;
e2be15f6 928 bool dirty, writeback;
854e9ed0
MK
929 bool lazyfree = false;
930 int ret = SWAP_SUCCESS;
1da177e4
LT
931
932 cond_resched();
933
934 page = lru_to_page(page_list);
935 list_del(&page->lru);
936
529ae9aa 937 if (!trylock_page(page))
1da177e4
LT
938 goto keep;
939
309381fe 940 VM_BUG_ON_PAGE(PageActive(page), page);
1da177e4
LT
941
942 sc->nr_scanned++;
80e43426 943
39b5f29a 944 if (unlikely(!page_evictable(page)))
b291f000 945 goto cull_mlocked;
894bc310 946
a6dc60f8 947 if (!sc->may_unmap && page_mapped(page))
80e43426
CL
948 goto keep_locked;
949
1da177e4
LT
950 /* Double the slab pressure for mapped and swapcache pages */
951 if (page_mapped(page) || PageSwapCache(page))
952 sc->nr_scanned++;
953
c661b078
AW
954 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
955 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
956
e2be15f6
MG
957 /*
958 * The number of dirty pages determines if a zone is marked
959 * reclaim_congested which affects wait_iff_congested. kswapd
960 * will stall and start writing pages if the tail of the LRU
961 * is all dirty unqueued pages.
962 */
963 page_check_dirty_writeback(page, &dirty, &writeback);
964 if (dirty || writeback)
965 nr_dirty++;
966
967 if (dirty && !writeback)
968 nr_unqueued_dirty++;
969
d04e8acd
MG
970 /*
971 * Treat this page as congested if the underlying BDI is or if
972 * pages are cycling through the LRU so quickly that the
973 * pages marked for immediate reclaim are making it to the
974 * end of the LRU a second time.
975 */
e2be15f6 976 mapping = page_mapping(page);
1da58ee2 977 if (((dirty || writeback) && mapping &&
703c2708 978 inode_write_congested(mapping->host)) ||
d04e8acd 979 (writeback && PageReclaim(page)))
e2be15f6
MG
980 nr_congested++;
981
283aba9f
MG
982 /*
983 * If a page at the tail of the LRU is under writeback, there
984 * are three cases to consider.
985 *
986 * 1) If reclaim is encountering an excessive number of pages
987 * under writeback and this page is both under writeback and
988 * PageReclaim then it indicates that pages are being queued
989 * for IO but are being recycled through the LRU before the
990 * IO can complete. Waiting on the page itself risks an
991 * indefinite stall if it is impossible to writeback the
992 * page due to IO error or disconnected storage so instead
b1a6f21e
MG
993 * note that the LRU is being scanned too quickly and the
994 * caller can stall after page list has been processed.
283aba9f 995 *
97c9341f 996 * 2) Global or new memcg reclaim encounters a page that is
ecf5fc6e
MH
997 * not marked for immediate reclaim, or the caller does not
998 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
999 * not to fs). In this case mark the page for immediate
97c9341f 1000 * reclaim and continue scanning.
283aba9f 1001 *
ecf5fc6e
MH
1002 * Require may_enter_fs because we would wait on fs, which
1003 * may not have submitted IO yet. And the loop driver might
283aba9f
MG
1004 * enter reclaim, and deadlock if it waits on a page for
1005 * which it is needed to do the write (loop masks off
1006 * __GFP_IO|__GFP_FS for this reason); but more thought
1007 * would probably show more reasons.
1008 *
7fadc820 1009 * 3) Legacy memcg encounters a page that is already marked
283aba9f
MG
1010 * PageReclaim. memcg does not have any dirty pages
1011 * throttling so we could easily OOM just because too many
1012 * pages are in writeback and there is nothing else to
1013 * reclaim. Wait for the writeback to complete.
1014 */
c661b078 1015 if (PageWriteback(page)) {
283aba9f
MG
1016 /* Case 1 above */
1017 if (current_is_kswapd() &&
1018 PageReclaim(page) &&
599d0c95 1019 test_bit(PGDAT_WRITEBACK, &pgdat->flags)) {
b1a6f21e
MG
1020 nr_immediate++;
1021 goto keep_locked;
283aba9f
MG
1022
1023 /* Case 2 above */
97c9341f 1024 } else if (sane_reclaim(sc) ||
ecf5fc6e 1025 !PageReclaim(page) || !may_enter_fs) {
c3b94f44
HD
1026 /*
1027 * This is slightly racy - end_page_writeback()
1028 * might have just cleared PageReclaim, then
1029 * setting PageReclaim here end up interpreted
1030 * as PageReadahead - but that does not matter
1031 * enough to care. What we do want is for this
1032 * page to have PageReclaim set next time memcg
1033 * reclaim reaches the tests above, so it will
1034 * then wait_on_page_writeback() to avoid OOM;
1035 * and it's also appropriate in global reclaim.
1036 */
1037 SetPageReclaim(page);
e62e384e 1038 nr_writeback++;
c3b94f44 1039 goto keep_locked;
283aba9f
MG
1040
1041 /* Case 3 above */
1042 } else {
7fadc820 1043 unlock_page(page);
283aba9f 1044 wait_on_page_writeback(page);
7fadc820
HD
1045 /* then go back and try same page again */
1046 list_add_tail(&page->lru, page_list);
1047 continue;
e62e384e 1048 }
c661b078 1049 }
1da177e4 1050
02c6de8d
MK
1051 if (!force_reclaim)
1052 references = page_check_references(page, sc);
1053
dfc8d636
JW
1054 switch (references) {
1055 case PAGEREF_ACTIVATE:
1da177e4 1056 goto activate_locked;
64574746
JW
1057 case PAGEREF_KEEP:
1058 goto keep_locked;
dfc8d636
JW
1059 case PAGEREF_RECLAIM:
1060 case PAGEREF_RECLAIM_CLEAN:
1061 ; /* try to reclaim the page below */
1062 }
1da177e4 1063
1da177e4
LT
1064 /*
1065 * Anonymous process memory has backing store?
1066 * Try to allocate it some swap space here.
1067 */
b291f000 1068 if (PageAnon(page) && !PageSwapCache(page)) {
63eb6b93
HD
1069 if (!(sc->gfp_mask & __GFP_IO))
1070 goto keep_locked;
5bc7b8ac 1071 if (!add_to_swap(page, page_list))
1da177e4 1072 goto activate_locked;
854e9ed0 1073 lazyfree = true;
63eb6b93 1074 may_enter_fs = 1;
1da177e4 1075
e2be15f6
MG
1076 /* Adding to swap updated mapping */
1077 mapping = page_mapping(page);
7751b2da
KS
1078 } else if (unlikely(PageTransHuge(page))) {
1079 /* Split file THP */
1080 if (split_huge_page_to_list(page, page_list))
1081 goto keep_locked;
e2be15f6 1082 }
1da177e4 1083
7751b2da
KS
1084 VM_BUG_ON_PAGE(PageTransHuge(page), page);
1085
1da177e4
LT
1086 /*
1087 * The page is mapped into the page tables of one or more
1088 * processes. Try to unmap it here.
1089 */
1090 if (page_mapped(page) && mapping) {
854e9ed0
MK
1091 switch (ret = try_to_unmap(page, lazyfree ?
1092 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1093 (ttu_flags | TTU_BATCH_FLUSH))) {
1da177e4
LT
1094 case SWAP_FAIL:
1095 goto activate_locked;
1096 case SWAP_AGAIN:
1097 goto keep_locked;
b291f000
NP
1098 case SWAP_MLOCK:
1099 goto cull_mlocked;
854e9ed0
MK
1100 case SWAP_LZFREE:
1101 goto lazyfree;
1da177e4
LT
1102 case SWAP_SUCCESS:
1103 ; /* try to free the page below */
1104 }
1105 }
1106
1107 if (PageDirty(page)) {
ee72886d
MG
1108 /*
1109 * Only kswapd can writeback filesystem pages to
d43006d5
MG
1110 * avoid risk of stack overflow but only writeback
1111 * if many dirty pages have been encountered.
ee72886d 1112 */
f84f6e2b 1113 if (page_is_file_cache(page) &&
9e3b2f8c 1114 (!current_is_kswapd() ||
599d0c95 1115 !test_bit(PGDAT_DIRTY, &pgdat->flags))) {
49ea7eb6
MG
1116 /*
1117 * Immediately reclaim when written back.
1118 * Similar in principal to deactivate_page()
1119 * except we already have the page isolated
1120 * and know it's dirty
1121 */
c4a25635 1122 inc_node_page_state(page, NR_VMSCAN_IMMEDIATE);
49ea7eb6
MG
1123 SetPageReclaim(page);
1124
ee72886d
MG
1125 goto keep_locked;
1126 }
1127
dfc8d636 1128 if (references == PAGEREF_RECLAIM_CLEAN)
1da177e4 1129 goto keep_locked;
4dd4b920 1130 if (!may_enter_fs)
1da177e4 1131 goto keep_locked;
52a8363e 1132 if (!sc->may_writepage)
1da177e4
LT
1133 goto keep_locked;
1134
d950c947
MG
1135 /*
1136 * Page is dirty. Flush the TLB if a writable entry
1137 * potentially exists to avoid CPU writes after IO
1138 * starts and then write it out here.
1139 */
1140 try_to_unmap_flush_dirty();
7d3579e8 1141 switch (pageout(page, mapping, sc)) {
1da177e4
LT
1142 case PAGE_KEEP:
1143 goto keep_locked;
1144 case PAGE_ACTIVATE:
1145 goto activate_locked;
1146 case PAGE_SUCCESS:
7d3579e8 1147 if (PageWriteback(page))
41ac1999 1148 goto keep;
7d3579e8 1149 if (PageDirty(page))
1da177e4 1150 goto keep;
7d3579e8 1151
1da177e4
LT
1152 /*
1153 * A synchronous write - probably a ramdisk. Go
1154 * ahead and try to reclaim the page.
1155 */
529ae9aa 1156 if (!trylock_page(page))
1da177e4
LT
1157 goto keep;
1158 if (PageDirty(page) || PageWriteback(page))
1159 goto keep_locked;
1160 mapping = page_mapping(page);
1161 case PAGE_CLEAN:
1162 ; /* try to free the page below */
1163 }
1164 }
1165
1166 /*
1167 * If the page has buffers, try to free the buffer mappings
1168 * associated with this page. If we succeed we try to free
1169 * the page as well.
1170 *
1171 * We do this even if the page is PageDirty().
1172 * try_to_release_page() does not perform I/O, but it is
1173 * possible for a page to have PageDirty set, but it is actually
1174 * clean (all its buffers are clean). This happens if the
1175 * buffers were written out directly, with submit_bh(). ext3
894bc310 1176 * will do this, as well as the blockdev mapping.
1da177e4
LT
1177 * try_to_release_page() will discover that cleanness and will
1178 * drop the buffers and mark the page clean - it can be freed.
1179 *
1180 * Rarely, pages can have buffers and no ->mapping. These are
1181 * the pages which were not successfully invalidated in
1182 * truncate_complete_page(). We try to drop those buffers here
1183 * and if that worked, and the page is no longer mapped into
1184 * process address space (page_count == 1) it can be freed.
1185 * Otherwise, leave the page on the LRU so it is swappable.
1186 */
266cf658 1187 if (page_has_private(page)) {
1da177e4
LT
1188 if (!try_to_release_page(page, sc->gfp_mask))
1189 goto activate_locked;
e286781d
NP
1190 if (!mapping && page_count(page) == 1) {
1191 unlock_page(page);
1192 if (put_page_testzero(page))
1193 goto free_it;
1194 else {
1195 /*
1196 * rare race with speculative reference.
1197 * the speculative reference will free
1198 * this page shortly, so we may
1199 * increment nr_reclaimed here (and
1200 * leave it off the LRU).
1201 */
1202 nr_reclaimed++;
1203 continue;
1204 }
1205 }
1da177e4
LT
1206 }
1207
854e9ed0 1208lazyfree:
a528910e 1209 if (!mapping || !__remove_mapping(mapping, page, true))
49d2e9cc 1210 goto keep_locked;
1da177e4 1211
a978d6f5
NP
1212 /*
1213 * At this point, we have no other references and there is
1214 * no way to pick any more up (removed from LRU, removed
1215 * from pagecache). Can use non-atomic bitops now (and
1216 * we obviously don't have to worry about waking up a process
1217 * waiting on the page lock, because there are no references.
1218 */
48c935ad 1219 __ClearPageLocked(page);
e286781d 1220free_it:
854e9ed0
MK
1221 if (ret == SWAP_LZFREE)
1222 count_vm_event(PGLAZYFREED);
1223
05ff5137 1224 nr_reclaimed++;
abe4c3b5
MG
1225
1226 /*
1227 * Is there need to periodically free_page_list? It would
1228 * appear not as the counts should be low
1229 */
1230 list_add(&page->lru, &free_pages);
1da177e4
LT
1231 continue;
1232
b291f000 1233cull_mlocked:
63d6c5ad
HD
1234 if (PageSwapCache(page))
1235 try_to_free_swap(page);
b291f000 1236 unlock_page(page);
c54839a7 1237 list_add(&page->lru, &ret_pages);
b291f000
NP
1238 continue;
1239
1da177e4 1240activate_locked:
68a22394 1241 /* Not a candidate for swapping, so reclaim swap space. */
5ccc5aba 1242 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
a2c43eed 1243 try_to_free_swap(page);
309381fe 1244 VM_BUG_ON_PAGE(PageActive(page), page);
1da177e4
LT
1245 SetPageActive(page);
1246 pgactivate++;
1247keep_locked:
1248 unlock_page(page);
1249keep:
1250 list_add(&page->lru, &ret_pages);
309381fe 1251 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1da177e4 1252 }
abe4c3b5 1253
747db954 1254 mem_cgroup_uncharge_list(&free_pages);
72b252ae 1255 try_to_unmap_flush();
b745bc85 1256 free_hot_cold_page_list(&free_pages, true);
abe4c3b5 1257
1da177e4 1258 list_splice(&ret_pages, page_list);
f8891e5e 1259 count_vm_events(PGACTIVATE, pgactivate);
0a31bc97 1260
8e950282
MG
1261 *ret_nr_dirty += nr_dirty;
1262 *ret_nr_congested += nr_congested;
d43006d5 1263 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
92df3a72 1264 *ret_nr_writeback += nr_writeback;
b1a6f21e 1265 *ret_nr_immediate += nr_immediate;
05ff5137 1266 return nr_reclaimed;
1da177e4
LT
1267}
1268
02c6de8d
MK
1269unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1270 struct list_head *page_list)
1271{
1272 struct scan_control sc = {
1273 .gfp_mask = GFP_KERNEL,
1274 .priority = DEF_PRIORITY,
1275 .may_unmap = 1,
1276 };
8e950282 1277 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
02c6de8d
MK
1278 struct page *page, *next;
1279 LIST_HEAD(clean_pages);
1280
1281 list_for_each_entry_safe(page, next, page_list, lru) {
117aad1e 1282 if (page_is_file_cache(page) && !PageDirty(page) &&
b1123ea6 1283 !__PageMovable(page)) {
02c6de8d
MK
1284 ClearPageActive(page);
1285 list_move(&page->lru, &clean_pages);
1286 }
1287 }
1288
599d0c95 1289 ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc,
8e950282
MG
1290 TTU_UNMAP|TTU_IGNORE_ACCESS,
1291 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
02c6de8d 1292 list_splice(&clean_pages, page_list);
599d0c95 1293 mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret);
02c6de8d
MK
1294 return ret;
1295}
1296
5ad333eb
AW
1297/*
1298 * Attempt to remove the specified page from its LRU. Only take this page
1299 * if it is of the appropriate PageActive status. Pages which are being
1300 * freed elsewhere are also ignored.
1301 *
1302 * page: page to consider
1303 * mode: one of the LRU isolation modes defined above
1304 *
1305 * returns 0 on success, -ve errno on failure.
1306 */
f3fd4a61 1307int __isolate_lru_page(struct page *page, isolate_mode_t mode)
5ad333eb
AW
1308{
1309 int ret = -EINVAL;
1310
1311 /* Only take pages on the LRU. */
1312 if (!PageLRU(page))
1313 return ret;
1314
e46a2879
MK
1315 /* Compaction should not handle unevictable pages but CMA can do so */
1316 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
894bc310
LS
1317 return ret;
1318
5ad333eb 1319 ret = -EBUSY;
08e552c6 1320
c8244935
MG
1321 /*
1322 * To minimise LRU disruption, the caller can indicate that it only
1323 * wants to isolate pages it will be able to operate on without
1324 * blocking - clean pages for the most part.
1325 *
1326 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1327 * is used by reclaim when it is cannot write to backing storage
1328 *
1329 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1330 * that it is possible to migrate without blocking
1331 */
1332 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1333 /* All the caller can do on PageWriteback is block */
1334 if (PageWriteback(page))
1335 return ret;
1336
1337 if (PageDirty(page)) {
1338 struct address_space *mapping;
1339
1340 /* ISOLATE_CLEAN means only clean pages */
1341 if (mode & ISOLATE_CLEAN)
1342 return ret;
1343
1344 /*
1345 * Only pages without mappings or that have a
1346 * ->migratepage callback are possible to migrate
1347 * without blocking
1348 */
1349 mapping = page_mapping(page);
1350 if (mapping && !mapping->a_ops->migratepage)
1351 return ret;
1352 }
1353 }
39deaf85 1354
f80c0673
MK
1355 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1356 return ret;
1357
5ad333eb
AW
1358 if (likely(get_page_unless_zero(page))) {
1359 /*
1360 * Be careful not to clear PageLRU until after we're
1361 * sure the page is not being freed elsewhere -- the
1362 * page release code relies on it.
1363 */
1364 ClearPageLRU(page);
1365 ret = 0;
1366 }
1367
1368 return ret;
1369}
1370
7ee36a14
MG
1371
1372/*
1373 * Update LRU sizes after isolating pages. The LRU size updates must
1374 * be complete before mem_cgroup_update_lru_size due to a santity check.
1375 */
1376static __always_inline void update_lru_sizes(struct lruvec *lruvec,
1377 enum lru_list lru, unsigned long *nr_zone_taken,
1378 unsigned long nr_taken)
1379{
7ee36a14
MG
1380 int zid;
1381
7ee36a14
MG
1382 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1383 if (!nr_zone_taken[zid])
1384 continue;
1385
1386 __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]);
1387 }
7ee36a14
MG
1388
1389#ifdef CONFIG_MEMCG
1390 mem_cgroup_update_lru_size(lruvec, lru, -nr_taken);
1391#endif
1392}
1393
1da177e4 1394/*
a52633d8 1395 * zone_lru_lock is heavily contended. Some of the functions that
1da177e4
LT
1396 * shrink the lists perform better by taking out a batch of pages
1397 * and working on them outside the LRU lock.
1398 *
1399 * For pagecache intensive workloads, this function is the hottest
1400 * spot in the kernel (apart from copy_*_user functions).
1401 *
1402 * Appropriate locks must be held before calling this function.
1403 *
1404 * @nr_to_scan: The number of pages to look through on the list.
5dc35979 1405 * @lruvec: The LRU vector to pull pages from.
1da177e4 1406 * @dst: The temp list to put pages on to.
f626012d 1407 * @nr_scanned: The number of pages that were scanned.
fe2c2a10 1408 * @sc: The scan_control struct for this reclaim session
5ad333eb 1409 * @mode: One of the LRU isolation modes
3cb99451 1410 * @lru: LRU list id for isolating
1da177e4
LT
1411 *
1412 * returns how many pages were moved onto *@dst.
1413 */
69e05944 1414static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
5dc35979 1415 struct lruvec *lruvec, struct list_head *dst,
fe2c2a10 1416 unsigned long *nr_scanned, struct scan_control *sc,
3cb99451 1417 isolate_mode_t mode, enum lru_list lru)
1da177e4 1418{
75b00af7 1419 struct list_head *src = &lruvec->lists[lru];
69e05944 1420 unsigned long nr_taken = 0;
599d0c95 1421 unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 };
7cc30fcf 1422 unsigned long nr_skipped[MAX_NR_ZONES] = { 0, };
599d0c95 1423 unsigned long scan, nr_pages;
b2e18757 1424 LIST_HEAD(pages_skipped);
1da177e4 1425
0b802f10 1426 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
d7f05528 1427 !list_empty(src);) {
5ad333eb 1428 struct page *page;
5ad333eb 1429
1da177e4
LT
1430 page = lru_to_page(src);
1431 prefetchw_prev_lru_page(page, src, flags);
1432
309381fe 1433 VM_BUG_ON_PAGE(!PageLRU(page), page);
8d438f96 1434
b2e18757
MG
1435 if (page_zonenum(page) > sc->reclaim_idx) {
1436 list_move(&page->lru, &pages_skipped);
7cc30fcf 1437 nr_skipped[page_zonenum(page)]++;
b2e18757
MG
1438 continue;
1439 }
1440
d7f05528
MG
1441 /*
1442 * Account for scanned and skipped separetly to avoid the pgdat
1443 * being prematurely marked unreclaimable by pgdat_reclaimable.
1444 */
1445 scan++;
1446
f3fd4a61 1447 switch (__isolate_lru_page(page, mode)) {
5ad333eb 1448 case 0:
599d0c95
MG
1449 nr_pages = hpage_nr_pages(page);
1450 nr_taken += nr_pages;
1451 nr_zone_taken[page_zonenum(page)] += nr_pages;
5ad333eb 1452 list_move(&page->lru, dst);
5ad333eb
AW
1453 break;
1454
1455 case -EBUSY:
1456 /* else it is being freed elsewhere */
1457 list_move(&page->lru, src);
1458 continue;
46453a6e 1459
5ad333eb
AW
1460 default:
1461 BUG();
1462 }
1da177e4
LT
1463 }
1464
b2e18757
MG
1465 /*
1466 * Splice any skipped pages to the start of the LRU list. Note that
1467 * this disrupts the LRU order when reclaiming for lower zones but
1468 * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX
1469 * scanning would soon rescan the same pages to skip and put the
1470 * system at risk of premature OOM.
1471 */
7cc30fcf
MG
1472 if (!list_empty(&pages_skipped)) {
1473 int zid;
d7f05528 1474 unsigned long total_skipped = 0;
7cc30fcf 1475
7cc30fcf
MG
1476 for (zid = 0; zid < MAX_NR_ZONES; zid++) {
1477 if (!nr_skipped[zid])
1478 continue;
1479
1480 __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]);
d7f05528 1481 total_skipped += nr_skipped[zid];
7cc30fcf 1482 }
d7f05528
MG
1483
1484 /*
1485 * Account skipped pages as a partial scan as the pgdat may be
1486 * close to unreclaimable. If the LRU list is empty, account
1487 * skipped pages as a full scan.
1488 */
1489 scan += list_empty(src) ? total_skipped : total_skipped >> 2;
1490
1491 list_splice(&pages_skipped, src);
7cc30fcf 1492 }
f626012d 1493 *nr_scanned = scan;
e5146b12 1494 trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, scan,
75b00af7 1495 nr_taken, mode, is_file_lru(lru));
7ee36a14 1496 update_lru_sizes(lruvec, lru, nr_zone_taken, nr_taken);
1da177e4
LT
1497 return nr_taken;
1498}
1499
62695a84
NP
1500/**
1501 * isolate_lru_page - tries to isolate a page from its LRU list
1502 * @page: page to isolate from its LRU list
1503 *
1504 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1505 * vmstat statistic corresponding to whatever LRU list the page was on.
1506 *
1507 * Returns 0 if the page was removed from an LRU list.
1508 * Returns -EBUSY if the page was not on an LRU list.
1509 *
1510 * The returned page will have PageLRU() cleared. If it was found on
894bc310
LS
1511 * the active list, it will have PageActive set. If it was found on
1512 * the unevictable list, it will have the PageUnevictable bit set. That flag
1513 * may need to be cleared by the caller before letting the page go.
62695a84
NP
1514 *
1515 * The vmstat statistic corresponding to the list on which the page was
1516 * found will be decremented.
1517 *
1518 * Restrictions:
1519 * (1) Must be called with an elevated refcount on the page. This is a
1520 * fundamentnal difference from isolate_lru_pages (which is called
1521 * without a stable reference).
1522 * (2) the lru_lock must not be held.
1523 * (3) interrupts must be enabled.
1524 */
1525int isolate_lru_page(struct page *page)
1526{
1527 int ret = -EBUSY;
1528
309381fe 1529 VM_BUG_ON_PAGE(!page_count(page), page);
cf2a82ee 1530 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
0c917313 1531
62695a84
NP
1532 if (PageLRU(page)) {
1533 struct zone *zone = page_zone(page);
fa9add64 1534 struct lruvec *lruvec;
62695a84 1535
a52633d8 1536 spin_lock_irq(zone_lru_lock(zone));
599d0c95 1537 lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat);
0c917313 1538 if (PageLRU(page)) {
894bc310 1539 int lru = page_lru(page);
0c917313 1540 get_page(page);
62695a84 1541 ClearPageLRU(page);
fa9add64
HD
1542 del_page_from_lru_list(page, lruvec, lru);
1543 ret = 0;
62695a84 1544 }
a52633d8 1545 spin_unlock_irq(zone_lru_lock(zone));
62695a84
NP
1546 }
1547 return ret;
1548}
1549
35cd7815 1550/*
d37dd5dc
FW
1551 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1552 * then get resheduled. When there are massive number of tasks doing page
1553 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1554 * the LRU list will go small and be scanned faster than necessary, leading to
1555 * unnecessary swapping, thrashing and OOM.
35cd7815 1556 */
599d0c95 1557static int too_many_isolated(struct pglist_data *pgdat, int file,
35cd7815
RR
1558 struct scan_control *sc)
1559{
1560 unsigned long inactive, isolated;
1561
1562 if (current_is_kswapd())
1563 return 0;
1564
97c9341f 1565 if (!sane_reclaim(sc))
35cd7815
RR
1566 return 0;
1567
1568 if (file) {
599d0c95
MG
1569 inactive = node_page_state(pgdat, NR_INACTIVE_FILE);
1570 isolated = node_page_state(pgdat, NR_ISOLATED_FILE);
35cd7815 1571 } else {
599d0c95
MG
1572 inactive = node_page_state(pgdat, NR_INACTIVE_ANON);
1573 isolated = node_page_state(pgdat, NR_ISOLATED_ANON);
35cd7815
RR
1574 }
1575
3cf23841
FW
1576 /*
1577 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1578 * won't get blocked by normal direct-reclaimers, forming a circular
1579 * deadlock.
1580 */
d0164adc 1581 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
3cf23841
FW
1582 inactive >>= 3;
1583
35cd7815
RR
1584 return isolated > inactive;
1585}
1586
66635629 1587static noinline_for_stack void
75b00af7 1588putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
66635629 1589{
27ac81d8 1590 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
599d0c95 1591 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
3f79768f 1592 LIST_HEAD(pages_to_free);
66635629 1593
66635629
MG
1594 /*
1595 * Put back any unfreeable pages.
1596 */
66635629 1597 while (!list_empty(page_list)) {
3f79768f 1598 struct page *page = lru_to_page(page_list);
66635629 1599 int lru;
3f79768f 1600
309381fe 1601 VM_BUG_ON_PAGE(PageLRU(page), page);
66635629 1602 list_del(&page->lru);
39b5f29a 1603 if (unlikely(!page_evictable(page))) {
599d0c95 1604 spin_unlock_irq(&pgdat->lru_lock);
66635629 1605 putback_lru_page(page);
599d0c95 1606 spin_lock_irq(&pgdat->lru_lock);
66635629
MG
1607 continue;
1608 }
fa9add64 1609
599d0c95 1610 lruvec = mem_cgroup_page_lruvec(page, pgdat);
fa9add64 1611
7a608572 1612 SetPageLRU(page);
66635629 1613 lru = page_lru(page);
fa9add64
HD
1614 add_page_to_lru_list(page, lruvec, lru);
1615
66635629
MG
1616 if (is_active_lru(lru)) {
1617 int file = is_file_lru(lru);
9992af10
RR
1618 int numpages = hpage_nr_pages(page);
1619 reclaim_stat->recent_rotated[file] += numpages;
66635629 1620 }
2bcf8879
HD
1621 if (put_page_testzero(page)) {
1622 __ClearPageLRU(page);
1623 __ClearPageActive(page);
fa9add64 1624 del_page_from_lru_list(page, lruvec, lru);
2bcf8879
HD
1625
1626 if (unlikely(PageCompound(page))) {
599d0c95 1627 spin_unlock_irq(&pgdat->lru_lock);
747db954 1628 mem_cgroup_uncharge(page);
2bcf8879 1629 (*get_compound_page_dtor(page))(page);
599d0c95 1630 spin_lock_irq(&pgdat->lru_lock);
2bcf8879
HD
1631 } else
1632 list_add(&page->lru, &pages_to_free);
66635629
MG
1633 }
1634 }
66635629 1635
3f79768f
HD
1636 /*
1637 * To save our caller's stack, now use input list for pages to free.
1638 */
1639 list_splice(&pages_to_free, page_list);
66635629
MG
1640}
1641
399ba0b9
N
1642/*
1643 * If a kernel thread (such as nfsd for loop-back mounts) services
1644 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1645 * In that case we should only throttle if the backing device it is
1646 * writing to is congested. In other cases it is safe to throttle.
1647 */
1648static int current_may_throttle(void)
1649{
1650 return !(current->flags & PF_LESS_THROTTLE) ||
1651 current->backing_dev_info == NULL ||
1652 bdi_write_congested(current->backing_dev_info);
1653}
1654
91dcade4
MK
1655static bool inactive_reclaimable_pages(struct lruvec *lruvec,
1656 struct scan_control *sc, enum lru_list lru)
1657{
1658 int zid;
1659 struct zone *zone;
1660 int file = is_file_lru(lru);
1661 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1662
1663 if (!global_reclaim(sc))
1664 return true;
1665
1666 for (zid = sc->reclaim_idx; zid >= 0; zid--) {
1667 zone = &pgdat->node_zones[zid];
6aa303de 1668 if (!managed_zone(zone))
91dcade4
MK
1669 continue;
1670
1671 if (zone_page_state_snapshot(zone, NR_ZONE_LRU_BASE +
1672 LRU_FILE * file) >= SWAP_CLUSTER_MAX)
1673 return true;
1674 }
1675
1676 return false;
1677}
1678
1da177e4 1679/*
b2e18757 1680 * shrink_inactive_list() is a helper for shrink_node(). It returns the number
1742f19f 1681 * of reclaimed pages
1da177e4 1682 */
66635629 1683static noinline_for_stack unsigned long
1a93be0e 1684shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
9e3b2f8c 1685 struct scan_control *sc, enum lru_list lru)
1da177e4
LT
1686{
1687 LIST_HEAD(page_list);
e247dbce 1688 unsigned long nr_scanned;
05ff5137 1689 unsigned long nr_reclaimed = 0;
e247dbce 1690 unsigned long nr_taken;
8e950282
MG
1691 unsigned long nr_dirty = 0;
1692 unsigned long nr_congested = 0;
e2be15f6 1693 unsigned long nr_unqueued_dirty = 0;
92df3a72 1694 unsigned long nr_writeback = 0;
b1a6f21e 1695 unsigned long nr_immediate = 0;
f3fd4a61 1696 isolate_mode_t isolate_mode = 0;
3cb99451 1697 int file = is_file_lru(lru);
599d0c95 1698 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1a93be0e 1699 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
78dc583d 1700
91dcade4
MK
1701 if (!inactive_reclaimable_pages(lruvec, sc, lru))
1702 return 0;
1703
599d0c95 1704 while (unlikely(too_many_isolated(pgdat, file, sc))) {
58355c78 1705 congestion_wait(BLK_RW_ASYNC, HZ/10);
35cd7815
RR
1706
1707 /* We are about to die and free our memory. Return now. */
1708 if (fatal_signal_pending(current))
1709 return SWAP_CLUSTER_MAX;
1710 }
1711
1da177e4 1712 lru_add_drain();
f80c0673
MK
1713
1714 if (!sc->may_unmap)
61317289 1715 isolate_mode |= ISOLATE_UNMAPPED;
f80c0673 1716 if (!sc->may_writepage)
61317289 1717 isolate_mode |= ISOLATE_CLEAN;
f80c0673 1718
599d0c95 1719 spin_lock_irq(&pgdat->lru_lock);
b35ea17b 1720
5dc35979
KK
1721 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1722 &nr_scanned, sc, isolate_mode, lru);
95d918fc 1723
599d0c95 1724 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
9d5e6a9f 1725 reclaim_stat->recent_scanned[file] += nr_taken;
95d918fc 1726
89b5fae5 1727 if (global_reclaim(sc)) {
599d0c95 1728 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
e247dbce 1729 if (current_is_kswapd())
599d0c95 1730 __count_vm_events(PGSCAN_KSWAPD, nr_scanned);
e247dbce 1731 else
599d0c95 1732 __count_vm_events(PGSCAN_DIRECT, nr_scanned);
e247dbce 1733 }
599d0c95 1734 spin_unlock_irq(&pgdat->lru_lock);
b35ea17b 1735
d563c050 1736 if (nr_taken == 0)
66635629 1737 return 0;
5ad333eb 1738
599d0c95 1739 nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, TTU_UNMAP,
8e950282
MG
1740 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1741 &nr_writeback, &nr_immediate,
1742 false);
c661b078 1743
599d0c95 1744 spin_lock_irq(&pgdat->lru_lock);
3f79768f 1745
904249aa
YH
1746 if (global_reclaim(sc)) {
1747 if (current_is_kswapd())
599d0c95 1748 __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed);
904249aa 1749 else
599d0c95 1750 __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed);
904249aa 1751 }
a74609fa 1752
27ac81d8 1753 putback_inactive_pages(lruvec, &page_list);
3f79768f 1754
599d0c95 1755 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
3f79768f 1756
599d0c95 1757 spin_unlock_irq(&pgdat->lru_lock);
3f79768f 1758
747db954 1759 mem_cgroup_uncharge_list(&page_list);
b745bc85 1760 free_hot_cold_page_list(&page_list, true);
e11da5b4 1761
92df3a72
MG
1762 /*
1763 * If reclaim is isolating dirty pages under writeback, it implies
1764 * that the long-lived page allocation rate is exceeding the page
1765 * laundering rate. Either the global limits are not being effective
1766 * at throttling processes due to the page distribution throughout
1767 * zones or there is heavy usage of a slow backing device. The
1768 * only option is to throttle from reclaim context which is not ideal
1769 * as there is no guarantee the dirtying process is throttled in the
1770 * same way balance_dirty_pages() manages.
1771 *
8e950282
MG
1772 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1773 * of pages under pages flagged for immediate reclaim and stall if any
1774 * are encountered in the nr_immediate check below.
92df3a72 1775 */
918fc718 1776 if (nr_writeback && nr_writeback == nr_taken)
599d0c95 1777 set_bit(PGDAT_WRITEBACK, &pgdat->flags);
92df3a72 1778
d43006d5 1779 /*
97c9341f
TH
1780 * Legacy memcg will stall in page writeback so avoid forcibly
1781 * stalling here.
d43006d5 1782 */
97c9341f 1783 if (sane_reclaim(sc)) {
8e950282
MG
1784 /*
1785 * Tag a zone as congested if all the dirty pages scanned were
1786 * backed by a congested BDI and wait_iff_congested will stall.
1787 */
1788 if (nr_dirty && nr_dirty == nr_congested)
599d0c95 1789 set_bit(PGDAT_CONGESTED, &pgdat->flags);
8e950282 1790
b1a6f21e
MG
1791 /*
1792 * If dirty pages are scanned that are not queued for IO, it
1793 * implies that flushers are not keeping up. In this case, flag
599d0c95 1794 * the pgdat PGDAT_DIRTY and kswapd will start writing pages from
57054651 1795 * reclaim context.
b1a6f21e
MG
1796 */
1797 if (nr_unqueued_dirty == nr_taken)
599d0c95 1798 set_bit(PGDAT_DIRTY, &pgdat->flags);
b1a6f21e
MG
1799
1800 /*
b738d764
LT
1801 * If kswapd scans pages marked marked for immediate
1802 * reclaim and under writeback (nr_immediate), it implies
1803 * that pages are cycling through the LRU faster than
b1a6f21e
MG
1804 * they are written so also forcibly stall.
1805 */
b738d764 1806 if (nr_immediate && current_may_throttle())
b1a6f21e 1807 congestion_wait(BLK_RW_ASYNC, HZ/10);
e2be15f6 1808 }
d43006d5 1809
8e950282
MG
1810 /*
1811 * Stall direct reclaim for IO completions if underlying BDIs or zone
1812 * is congested. Allow kswapd to continue until it starts encountering
1813 * unqueued dirty pages or cycling through the LRU too quickly.
1814 */
399ba0b9
N
1815 if (!sc->hibernation_mode && !current_is_kswapd() &&
1816 current_may_throttle())
599d0c95 1817 wait_iff_congested(pgdat, BLK_RW_ASYNC, HZ/10);
8e950282 1818
599d0c95
MG
1819 trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id,
1820 nr_scanned, nr_reclaimed,
ba5e9579 1821 sc->priority, file);
05ff5137 1822 return nr_reclaimed;
1da177e4
LT
1823}
1824
1825/*
1826 * This moves pages from the active list to the inactive list.
1827 *
1828 * We move them the other way if the page is referenced by one or more
1829 * processes, from rmap.
1830 *
1831 * If the pages are mostly unmapped, the processing is fast and it is
a52633d8 1832 * appropriate to hold zone_lru_lock across the whole operation. But if
1da177e4 1833 * the pages are mapped, the processing is slow (page_referenced()) so we
a52633d8 1834 * should drop zone_lru_lock around each page. It's impossible to balance
1da177e4
LT
1835 * this, so instead we remove the pages from the LRU while processing them.
1836 * It is safe to rely on PG_active against the non-LRU pages in here because
1837 * nobody will play with that bit on a non-LRU page.
1838 *
0139aa7b 1839 * The downside is that we have to touch page->_refcount against each page.
1da177e4
LT
1840 * But we had to alter page->flags anyway.
1841 */
1cfb419b 1842
fa9add64 1843static void move_active_pages_to_lru(struct lruvec *lruvec,
3eb4140f 1844 struct list_head *list,
2bcf8879 1845 struct list_head *pages_to_free,
3eb4140f
WF
1846 enum lru_list lru)
1847{
599d0c95 1848 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
3eb4140f 1849 unsigned long pgmoved = 0;
3eb4140f 1850 struct page *page;
fa9add64 1851 int nr_pages;
3eb4140f 1852
3eb4140f
WF
1853 while (!list_empty(list)) {
1854 page = lru_to_page(list);
599d0c95 1855 lruvec = mem_cgroup_page_lruvec(page, pgdat);
3eb4140f 1856
309381fe 1857 VM_BUG_ON_PAGE(PageLRU(page), page);
3eb4140f
WF
1858 SetPageLRU(page);
1859
fa9add64 1860 nr_pages = hpage_nr_pages(page);
599d0c95 1861 update_lru_size(lruvec, lru, page_zonenum(page), nr_pages);
925b7673 1862 list_move(&page->lru, &lruvec->lists[lru]);
fa9add64 1863 pgmoved += nr_pages;
3eb4140f 1864
2bcf8879
HD
1865 if (put_page_testzero(page)) {
1866 __ClearPageLRU(page);
1867 __ClearPageActive(page);
fa9add64 1868 del_page_from_lru_list(page, lruvec, lru);
2bcf8879
HD
1869
1870 if (unlikely(PageCompound(page))) {
599d0c95 1871 spin_unlock_irq(&pgdat->lru_lock);
747db954 1872 mem_cgroup_uncharge(page);
2bcf8879 1873 (*get_compound_page_dtor(page))(page);
599d0c95 1874 spin_lock_irq(&pgdat->lru_lock);
2bcf8879
HD
1875 } else
1876 list_add(&page->lru, pages_to_free);
3eb4140f
WF
1877 }
1878 }
9d5e6a9f 1879
3eb4140f
WF
1880 if (!is_active_lru(lru))
1881 __count_vm_events(PGDEACTIVATE, pgmoved);
1882}
1cfb419b 1883
f626012d 1884static void shrink_active_list(unsigned long nr_to_scan,
1a93be0e 1885 struct lruvec *lruvec,
f16015fb 1886 struct scan_control *sc,
9e3b2f8c 1887 enum lru_list lru)
1da177e4 1888{
44c241f1 1889 unsigned long nr_taken;
f626012d 1890 unsigned long nr_scanned;
6fe6b7e3 1891 unsigned long vm_flags;
1da177e4 1892 LIST_HEAD(l_hold); /* The pages which were snipped off */
8cab4754 1893 LIST_HEAD(l_active);
b69408e8 1894 LIST_HEAD(l_inactive);
1da177e4 1895 struct page *page;
1a93be0e 1896 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
44c241f1 1897 unsigned long nr_rotated = 0;
f3fd4a61 1898 isolate_mode_t isolate_mode = 0;
3cb99451 1899 int file = is_file_lru(lru);
599d0c95 1900 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
1da177e4
LT
1901
1902 lru_add_drain();
f80c0673
MK
1903
1904 if (!sc->may_unmap)
61317289 1905 isolate_mode |= ISOLATE_UNMAPPED;
f80c0673 1906 if (!sc->may_writepage)
61317289 1907 isolate_mode |= ISOLATE_CLEAN;
f80c0673 1908
599d0c95 1909 spin_lock_irq(&pgdat->lru_lock);
925b7673 1910
5dc35979
KK
1911 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1912 &nr_scanned, sc, isolate_mode, lru);
89b5fae5 1913
599d0c95 1914 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken);
b7c46d15 1915 reclaim_stat->recent_scanned[file] += nr_taken;
1cfb419b 1916
9d5e6a9f 1917 if (global_reclaim(sc))
599d0c95
MG
1918 __mod_node_page_state(pgdat, NR_PAGES_SCANNED, nr_scanned);
1919 __count_vm_events(PGREFILL, nr_scanned);
9d5e6a9f 1920
599d0c95 1921 spin_unlock_irq(&pgdat->lru_lock);
1da177e4 1922
1da177e4
LT
1923 while (!list_empty(&l_hold)) {
1924 cond_resched();
1925 page = lru_to_page(&l_hold);
1926 list_del(&page->lru);
7e9cd484 1927
39b5f29a 1928 if (unlikely(!page_evictable(page))) {
894bc310
LS
1929 putback_lru_page(page);
1930 continue;
1931 }
1932
cc715d99
MG
1933 if (unlikely(buffer_heads_over_limit)) {
1934 if (page_has_private(page) && trylock_page(page)) {
1935 if (page_has_private(page))
1936 try_to_release_page(page, 0);
1937 unlock_page(page);
1938 }
1939 }
1940
c3ac9a8a
JW
1941 if (page_referenced(page, 0, sc->target_mem_cgroup,
1942 &vm_flags)) {
9992af10 1943 nr_rotated += hpage_nr_pages(page);
8cab4754
WF
1944 /*
1945 * Identify referenced, file-backed active pages and
1946 * give them one more trip around the active list. So
1947 * that executable code get better chances to stay in
1948 * memory under moderate memory pressure. Anon pages
1949 * are not likely to be evicted by use-once streaming
1950 * IO, plus JVM can create lots of anon VM_EXEC pages,
1951 * so we ignore them here.
1952 */
41e20983 1953 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
8cab4754
WF
1954 list_add(&page->lru, &l_active);
1955 continue;
1956 }
1957 }
7e9cd484 1958
5205e56e 1959 ClearPageActive(page); /* we are de-activating */
1da177e4
LT
1960 list_add(&page->lru, &l_inactive);
1961 }
1962
b555749a 1963 /*
8cab4754 1964 * Move pages back to the lru list.
b555749a 1965 */
599d0c95 1966 spin_lock_irq(&pgdat->lru_lock);
556adecb 1967 /*
8cab4754
WF
1968 * Count referenced pages from currently used mappings as rotated,
1969 * even though only some of them are actually re-activated. This
1970 * helps balance scan pressure between file and anonymous pages in
7c0db9e9 1971 * get_scan_count.
7e9cd484 1972 */
b7c46d15 1973 reclaim_stat->recent_rotated[file] += nr_rotated;
556adecb 1974
fa9add64
HD
1975 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1976 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
599d0c95
MG
1977 __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken);
1978 spin_unlock_irq(&pgdat->lru_lock);
2bcf8879 1979
747db954 1980 mem_cgroup_uncharge_list(&l_hold);
b745bc85 1981 free_hot_cold_page_list(&l_hold, true);
1da177e4
LT
1982}
1983
59dc76b0
RR
1984/*
1985 * The inactive anon list should be small enough that the VM never has
1986 * to do too much work.
14797e23 1987 *
59dc76b0
RR
1988 * The inactive file list should be small enough to leave most memory
1989 * to the established workingset on the scan-resistant active list,
1990 * but large enough to avoid thrashing the aggregate readahead window.
56e49d21 1991 *
59dc76b0
RR
1992 * Both inactive lists should also be large enough that each inactive
1993 * page has a chance to be referenced again before it is reclaimed.
56e49d21 1994 *
59dc76b0
RR
1995 * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages
1996 * on this LRU, maintained by the pageout code. A zone->inactive_ratio
1997 * of 3 means 3:1 or 25% of the pages are kept on the inactive list.
56e49d21 1998 *
59dc76b0
RR
1999 * total target max
2000 * memory ratio inactive
2001 * -------------------------------------
2002 * 10MB 1 5MB
2003 * 100MB 1 50MB
2004 * 1GB 3 250MB
2005 * 10GB 10 0.9GB
2006 * 100GB 31 3GB
2007 * 1TB 101 10GB
2008 * 10TB 320 32GB
56e49d21 2009 */
f8d1a311
MG
2010static bool inactive_list_is_low(struct lruvec *lruvec, bool file,
2011 struct scan_control *sc)
56e49d21 2012{
59dc76b0 2013 unsigned long inactive_ratio;
e3790144
JW
2014 unsigned long inactive;
2015 unsigned long active;
59dc76b0 2016 unsigned long gb;
f8d1a311
MG
2017 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
2018 int zid;
e3790144 2019
59dc76b0
RR
2020 /*
2021 * If we don't have swap space, anonymous page deactivation
2022 * is pointless.
2023 */
2024 if (!file && !total_swap_pages)
2025 return false;
56e49d21 2026
59dc76b0
RR
2027 inactive = lruvec_lru_size(lruvec, file * LRU_FILE);
2028 active = lruvec_lru_size(lruvec, file * LRU_FILE + LRU_ACTIVE);
56e49d21 2029
f8d1a311
MG
2030 /*
2031 * For zone-constrained allocations, it is necessary to check if
2032 * deactivations are required for lowmem to be reclaimed. This
2033 * calculates the inactive/active pages available in eligible zones.
2034 */
2035 for (zid = sc->reclaim_idx + 1; zid < MAX_NR_ZONES; zid++) {
2036 struct zone *zone = &pgdat->node_zones[zid];
2037 unsigned long inactive_zone, active_zone;
2038
6aa303de 2039 if (!managed_zone(zone))
f8d1a311
MG
2040 continue;
2041
2042 inactive_zone = zone_page_state(zone,
2043 NR_ZONE_LRU_BASE + (file * LRU_FILE));
2044 active_zone = zone_page_state(zone,
2045 NR_ZONE_LRU_BASE + (file * LRU_FILE) + LRU_ACTIVE);
2046
2047 inactive -= min(inactive, inactive_zone);
2048 active -= min(active, active_zone);
2049 }
2050
59dc76b0
RR
2051 gb = (inactive + active) >> (30 - PAGE_SHIFT);
2052 if (gb)
2053 inactive_ratio = int_sqrt(10 * gb);
b39415b2 2054 else
59dc76b0
RR
2055 inactive_ratio = 1;
2056
2057 return inactive * inactive_ratio < active;
b39415b2
RR
2058}
2059
4f98a2fe 2060static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1a93be0e 2061 struct lruvec *lruvec, struct scan_control *sc)
b69408e8 2062{
b39415b2 2063 if (is_active_lru(lru)) {
f8d1a311 2064 if (inactive_list_is_low(lruvec, is_file_lru(lru), sc))
1a93be0e 2065 shrink_active_list(nr_to_scan, lruvec, sc, lru);
556adecb
RR
2066 return 0;
2067 }
2068
1a93be0e 2069 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
4f98a2fe
RR
2070}
2071
9a265114
JW
2072enum scan_balance {
2073 SCAN_EQUAL,
2074 SCAN_FRACT,
2075 SCAN_ANON,
2076 SCAN_FILE,
2077};
2078
4f98a2fe
RR
2079/*
2080 * Determine how aggressively the anon and file LRU lists should be
2081 * scanned. The relative value of each set of LRU lists is determined
2082 * by looking at the fraction of the pages scanned we did rotate back
2083 * onto the active list instead of evict.
2084 *
be7bd59d
WL
2085 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
2086 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
4f98a2fe 2087 */
33377678 2088static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
6b4f7799
JW
2089 struct scan_control *sc, unsigned long *nr,
2090 unsigned long *lru_pages)
4f98a2fe 2091{
33377678 2092 int swappiness = mem_cgroup_swappiness(memcg);
9a265114
JW
2093 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
2094 u64 fraction[2];
2095 u64 denominator = 0; /* gcc */
599d0c95 2096 struct pglist_data *pgdat = lruvec_pgdat(lruvec);
4f98a2fe 2097 unsigned long anon_prio, file_prio;
9a265114 2098 enum scan_balance scan_balance;
0bf1457f 2099 unsigned long anon, file;
9a265114 2100 bool force_scan = false;
4f98a2fe 2101 unsigned long ap, fp;
4111304d 2102 enum lru_list lru;
6f04f48d
SS
2103 bool some_scanned;
2104 int pass;
246e87a9 2105
f11c0ca5
JW
2106 /*
2107 * If the zone or memcg is small, nr[l] can be 0. This
2108 * results in no scanning on this priority and a potential
2109 * priority drop. Global direct reclaim can go to the next
2110 * zone and tends to have no problems. Global kswapd is for
2111 * zone balancing and it needs to scan a minimum amount. When
2112 * reclaiming for a memcg, a priority drop can cause high
2113 * latencies, so it's better to scan a minimum amount there as
2114 * well.
2115 */
90cbc250 2116 if (current_is_kswapd()) {
599d0c95 2117 if (!pgdat_reclaimable(pgdat))
90cbc250 2118 force_scan = true;
eb01aaab 2119 if (!mem_cgroup_online(memcg))
90cbc250
VD
2120 force_scan = true;
2121 }
89b5fae5 2122 if (!global_reclaim(sc))
a4d3e9e7 2123 force_scan = true;
76a33fc3
SL
2124
2125 /* If we have no swap space, do not bother scanning anon pages. */
d8b38438 2126 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
9a265114 2127 scan_balance = SCAN_FILE;
76a33fc3
SL
2128 goto out;
2129 }
4f98a2fe 2130
10316b31
JW
2131 /*
2132 * Global reclaim will swap to prevent OOM even with no
2133 * swappiness, but memcg users want to use this knob to
2134 * disable swapping for individual groups completely when
2135 * using the memory controller's swap limit feature would be
2136 * too expensive.
2137 */
02695175 2138 if (!global_reclaim(sc) && !swappiness) {
9a265114 2139 scan_balance = SCAN_FILE;
10316b31
JW
2140 goto out;
2141 }
2142
2143 /*
2144 * Do not apply any pressure balancing cleverness when the
2145 * system is close to OOM, scan both anon and file equally
2146 * (unless the swappiness setting disagrees with swapping).
2147 */
02695175 2148 if (!sc->priority && swappiness) {
9a265114 2149 scan_balance = SCAN_EQUAL;
10316b31
JW
2150 goto out;
2151 }
2152
62376251
JW
2153 /*
2154 * Prevent the reclaimer from falling into the cache trap: as
2155 * cache pages start out inactive, every cache fault will tip
2156 * the scan balance towards the file LRU. And as the file LRU
2157 * shrinks, so does the window for rotation from references.
2158 * This means we have a runaway feedback loop where a tiny
2159 * thrashing file LRU becomes infinitely more attractive than
2160 * anon pages. Try to detect this based on file LRU size.
2161 */
2162 if (global_reclaim(sc)) {
599d0c95
MG
2163 unsigned long pgdatfile;
2164 unsigned long pgdatfree;
2165 int z;
2166 unsigned long total_high_wmark = 0;
2ab051e1 2167
599d0c95
MG
2168 pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES);
2169 pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) +
2170 node_page_state(pgdat, NR_INACTIVE_FILE);
2171
2172 for (z = 0; z < MAX_NR_ZONES; z++) {
2173 struct zone *zone = &pgdat->node_zones[z];
6aa303de 2174 if (!managed_zone(zone))
599d0c95
MG
2175 continue;
2176
2177 total_high_wmark += high_wmark_pages(zone);
2178 }
62376251 2179
599d0c95 2180 if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) {
62376251
JW
2181 scan_balance = SCAN_ANON;
2182 goto out;
2183 }
2184 }
2185
7c5bd705 2186 /*
316bda0e
VD
2187 * If there is enough inactive page cache, i.e. if the size of the
2188 * inactive list is greater than that of the active list *and* the
2189 * inactive list actually has some pages to scan on this priority, we
2190 * do not reclaim anything from the anonymous working set right now.
2191 * Without the second condition we could end up never scanning an
2192 * lruvec even if it has plenty of old anonymous pages unless the
2193 * system is under heavy pressure.
7c5bd705 2194 */
f8d1a311 2195 if (!inactive_list_is_low(lruvec, true, sc) &&
23047a96 2196 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
9a265114 2197 scan_balance = SCAN_FILE;
7c5bd705
JW
2198 goto out;
2199 }
2200
9a265114
JW
2201 scan_balance = SCAN_FRACT;
2202
58c37f6e
KM
2203 /*
2204 * With swappiness at 100, anonymous and file have the same priority.
2205 * This scanning priority is essentially the inverse of IO cost.
2206 */
02695175 2207 anon_prio = swappiness;
75b00af7 2208 file_prio = 200 - anon_prio;
58c37f6e 2209
4f98a2fe
RR
2210 /*
2211 * OK, so we have swap space and a fair amount of page cache
2212 * pages. We use the recently rotated / recently scanned
2213 * ratios to determine how valuable each cache is.
2214 *
2215 * Because workloads change over time (and to avoid overflow)
2216 * we keep these statistics as a floating average, which ends
2217 * up weighing recent references more than old ones.
2218 *
2219 * anon in [0], file in [1]
2220 */
2ab051e1 2221
23047a96
JW
2222 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2223 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2224 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2225 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
2ab051e1 2226
599d0c95 2227 spin_lock_irq(&pgdat->lru_lock);
6e901571 2228 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
6e901571
KM
2229 reclaim_stat->recent_scanned[0] /= 2;
2230 reclaim_stat->recent_rotated[0] /= 2;
4f98a2fe
RR
2231 }
2232
6e901571 2233 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
6e901571
KM
2234 reclaim_stat->recent_scanned[1] /= 2;
2235 reclaim_stat->recent_rotated[1] /= 2;
4f98a2fe
RR
2236 }
2237
4f98a2fe 2238 /*
00d8089c
RR
2239 * The amount of pressure on anon vs file pages is inversely
2240 * proportional to the fraction of recently scanned pages on
2241 * each list that were recently referenced and in active use.
4f98a2fe 2242 */
fe35004f 2243 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
6e901571 2244 ap /= reclaim_stat->recent_rotated[0] + 1;
4f98a2fe 2245
fe35004f 2246 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
6e901571 2247 fp /= reclaim_stat->recent_rotated[1] + 1;
599d0c95 2248 spin_unlock_irq(&pgdat->lru_lock);
4f98a2fe 2249
76a33fc3
SL
2250 fraction[0] = ap;
2251 fraction[1] = fp;
2252 denominator = ap + fp + 1;
2253out:
6f04f48d
SS
2254 some_scanned = false;
2255 /* Only use force_scan on second pass. */
2256 for (pass = 0; !some_scanned && pass < 2; pass++) {
6b4f7799 2257 *lru_pages = 0;
6f04f48d
SS
2258 for_each_evictable_lru(lru) {
2259 int file = is_file_lru(lru);
2260 unsigned long size;
2261 unsigned long scan;
6e08a369 2262
23047a96 2263 size = lruvec_lru_size(lruvec, lru);
6f04f48d 2264 scan = size >> sc->priority;
9a265114 2265
6f04f48d
SS
2266 if (!scan && pass && force_scan)
2267 scan = min(size, SWAP_CLUSTER_MAX);
9a265114 2268
6f04f48d
SS
2269 switch (scan_balance) {
2270 case SCAN_EQUAL:
2271 /* Scan lists relative to size */
2272 break;
2273 case SCAN_FRACT:
2274 /*
2275 * Scan types proportional to swappiness and
2276 * their relative recent reclaim efficiency.
2277 */
2278 scan = div64_u64(scan * fraction[file],
2279 denominator);
2280 break;
2281 case SCAN_FILE:
2282 case SCAN_ANON:
2283 /* Scan one type exclusively */
6b4f7799
JW
2284 if ((scan_balance == SCAN_FILE) != file) {
2285 size = 0;
6f04f48d 2286 scan = 0;
6b4f7799 2287 }
6f04f48d
SS
2288 break;
2289 default:
2290 /* Look ma, no brain */
2291 BUG();
2292 }
6b4f7799
JW
2293
2294 *lru_pages += size;
6f04f48d 2295 nr[lru] = scan;
6b4f7799 2296
9a265114 2297 /*
6f04f48d
SS
2298 * Skip the second pass and don't force_scan,
2299 * if we found something to scan.
9a265114 2300 */
6f04f48d 2301 some_scanned |= !!scan;
9a265114 2302 }
76a33fc3 2303 }
6e08a369 2304}
4f98a2fe 2305
9b4f98cd 2306/*
a9dd0a83 2307 * This is a basic per-node page freer. Used by both kswapd and direct reclaim.
9b4f98cd 2308 */
a9dd0a83 2309static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg,
33377678 2310 struct scan_control *sc, unsigned long *lru_pages)
9b4f98cd 2311{
ef8f2327 2312 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
9b4f98cd 2313 unsigned long nr[NR_LRU_LISTS];
e82e0561 2314 unsigned long targets[NR_LRU_LISTS];
9b4f98cd
JW
2315 unsigned long nr_to_scan;
2316 enum lru_list lru;
2317 unsigned long nr_reclaimed = 0;
2318 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2319 struct blk_plug plug;
1a501907 2320 bool scan_adjusted;
9b4f98cd 2321
33377678 2322 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
9b4f98cd 2323
e82e0561
MG
2324 /* Record the original scan target for proportional adjustments later */
2325 memcpy(targets, nr, sizeof(nr));
2326
1a501907
MG
2327 /*
2328 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2329 * event that can occur when there is little memory pressure e.g.
2330 * multiple streaming readers/writers. Hence, we do not abort scanning
2331 * when the requested number of pages are reclaimed when scanning at
2332 * DEF_PRIORITY on the assumption that the fact we are direct
2333 * reclaiming implies that kswapd is not keeping up and it is best to
2334 * do a batch of work at once. For memcg reclaim one check is made to
2335 * abort proportional reclaim if either the file or anon lru has already
2336 * dropped to zero at the first pass.
2337 */
2338 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2339 sc->priority == DEF_PRIORITY);
2340
9b4f98cd
JW
2341 blk_start_plug(&plug);
2342 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2343 nr[LRU_INACTIVE_FILE]) {
e82e0561
MG
2344 unsigned long nr_anon, nr_file, percentage;
2345 unsigned long nr_scanned;
2346
9b4f98cd
JW
2347 for_each_evictable_lru(lru) {
2348 if (nr[lru]) {
2349 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2350 nr[lru] -= nr_to_scan;
2351
2352 nr_reclaimed += shrink_list(lru, nr_to_scan,
2353 lruvec, sc);
2354 }
2355 }
e82e0561 2356
bd041733
MH
2357 cond_resched();
2358
e82e0561
MG
2359 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2360 continue;
2361
e82e0561
MG
2362 /*
2363 * For kswapd and memcg, reclaim at least the number of pages
1a501907 2364 * requested. Ensure that the anon and file LRUs are scanned
e82e0561
MG
2365 * proportionally what was requested by get_scan_count(). We
2366 * stop reclaiming one LRU and reduce the amount scanning
2367 * proportional to the original scan target.
2368 */
2369 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2370 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2371
1a501907
MG
2372 /*
2373 * It's just vindictive to attack the larger once the smaller
2374 * has gone to zero. And given the way we stop scanning the
2375 * smaller below, this makes sure that we only make one nudge
2376 * towards proportionality once we've got nr_to_reclaim.
2377 */
2378 if (!nr_file || !nr_anon)
2379 break;
2380
e82e0561
MG
2381 if (nr_file > nr_anon) {
2382 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2383 targets[LRU_ACTIVE_ANON] + 1;
2384 lru = LRU_BASE;
2385 percentage = nr_anon * 100 / scan_target;
2386 } else {
2387 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2388 targets[LRU_ACTIVE_FILE] + 1;
2389 lru = LRU_FILE;
2390 percentage = nr_file * 100 / scan_target;
2391 }
2392
2393 /* Stop scanning the smaller of the LRU */
2394 nr[lru] = 0;
2395 nr[lru + LRU_ACTIVE] = 0;
2396
2397 /*
2398 * Recalculate the other LRU scan count based on its original
2399 * scan target and the percentage scanning already complete
2400 */
2401 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2402 nr_scanned = targets[lru] - nr[lru];
2403 nr[lru] = targets[lru] * (100 - percentage) / 100;
2404 nr[lru] -= min(nr[lru], nr_scanned);
2405
2406 lru += LRU_ACTIVE;
2407 nr_scanned = targets[lru] - nr[lru];
2408 nr[lru] = targets[lru] * (100 - percentage) / 100;
2409 nr[lru] -= min(nr[lru], nr_scanned);
2410
2411 scan_adjusted = true;
9b4f98cd
JW
2412 }
2413 blk_finish_plug(&plug);
2414 sc->nr_reclaimed += nr_reclaimed;
2415
2416 /*
2417 * Even if we did not try to evict anon pages at all, we want to
2418 * rebalance the anon lru active/inactive ratio.
2419 */
f8d1a311 2420 if (inactive_list_is_low(lruvec, false, sc))
9b4f98cd
JW
2421 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2422 sc, LRU_ACTIVE_ANON);
9b4f98cd
JW
2423}
2424
23b9da55 2425/* Use reclaim/compaction for costly allocs or under memory pressure */
9e3b2f8c 2426static bool in_reclaim_compaction(struct scan_control *sc)
23b9da55 2427{
d84da3f9 2428 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
23b9da55 2429 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
9e3b2f8c 2430 sc->priority < DEF_PRIORITY - 2))
23b9da55
MG
2431 return true;
2432
2433 return false;
2434}
2435
3e7d3449 2436/*
23b9da55
MG
2437 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2438 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2439 * true if more pages should be reclaimed such that when the page allocator
2440 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2441 * It will give up earlier than that if there is difficulty reclaiming pages.
3e7d3449 2442 */
a9dd0a83 2443static inline bool should_continue_reclaim(struct pglist_data *pgdat,
3e7d3449
MG
2444 unsigned long nr_reclaimed,
2445 unsigned long nr_scanned,
2446 struct scan_control *sc)
2447{
2448 unsigned long pages_for_compaction;
2449 unsigned long inactive_lru_pages;
a9dd0a83 2450 int z;
3e7d3449
MG
2451
2452 /* If not in reclaim/compaction mode, stop */
9e3b2f8c 2453 if (!in_reclaim_compaction(sc))
3e7d3449
MG
2454 return false;
2455
2876592f
MG
2456 /* Consider stopping depending on scan and reclaim activity */
2457 if (sc->gfp_mask & __GFP_REPEAT) {
2458 /*
2459 * For __GFP_REPEAT allocations, stop reclaiming if the
2460 * full LRU list has been scanned and we are still failing
2461 * to reclaim pages. This full LRU scan is potentially
2462 * expensive but a __GFP_REPEAT caller really wants to succeed
2463 */
2464 if (!nr_reclaimed && !nr_scanned)
2465 return false;
2466 } else {
2467 /*
2468 * For non-__GFP_REPEAT allocations which can presumably
2469 * fail without consequence, stop if we failed to reclaim
2470 * any pages from the last SWAP_CLUSTER_MAX number of
2471 * pages that were scanned. This will return to the
2472 * caller faster at the risk reclaim/compaction and
2473 * the resulting allocation attempt fails
2474 */
2475 if (!nr_reclaimed)
2476 return false;
2477 }
3e7d3449
MG
2478
2479 /*
2480 * If we have not reclaimed enough pages for compaction and the
2481 * inactive lists are large enough, continue reclaiming
2482 */
9861a62c 2483 pages_for_compaction = compact_gap(sc->order);
a9dd0a83 2484 inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE);
ec8acf20 2485 if (get_nr_swap_pages() > 0)
a9dd0a83 2486 inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON);
3e7d3449
MG
2487 if (sc->nr_reclaimed < pages_for_compaction &&
2488 inactive_lru_pages > pages_for_compaction)
2489 return true;
2490
2491 /* If compaction would go ahead or the allocation would succeed, stop */
a9dd0a83
MG
2492 for (z = 0; z <= sc->reclaim_idx; z++) {
2493 struct zone *zone = &pgdat->node_zones[z];
6aa303de 2494 if (!managed_zone(zone))
a9dd0a83
MG
2495 continue;
2496
2497 switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) {
cf378319 2498 case COMPACT_SUCCESS:
a9dd0a83
MG
2499 case COMPACT_CONTINUE:
2500 return false;
2501 default:
2502 /* check next zone */
2503 ;
2504 }
3e7d3449 2505 }
a9dd0a83 2506 return true;
3e7d3449
MG
2507}
2508
970a39a3 2509static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc)
1da177e4 2510{
cb731d6c 2511 struct reclaim_state *reclaim_state = current->reclaim_state;
f0fdc5e8 2512 unsigned long nr_reclaimed, nr_scanned;
2344d7e4 2513 bool reclaimable = false;
1da177e4 2514
9b4f98cd
JW
2515 do {
2516 struct mem_cgroup *root = sc->target_mem_cgroup;
2517 struct mem_cgroup_reclaim_cookie reclaim = {
ef8f2327 2518 .pgdat = pgdat,
9b4f98cd
JW
2519 .priority = sc->priority,
2520 };
a9dd0a83 2521 unsigned long node_lru_pages = 0;
694fbc0f 2522 struct mem_cgroup *memcg;
3e7d3449 2523
9b4f98cd
JW
2524 nr_reclaimed = sc->nr_reclaimed;
2525 nr_scanned = sc->nr_scanned;
1da177e4 2526
694fbc0f
AM
2527 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2528 do {
6b4f7799 2529 unsigned long lru_pages;
8e8ae645 2530 unsigned long reclaimed;
cb731d6c 2531 unsigned long scanned;
5660048c 2532
241994ed
JW
2533 if (mem_cgroup_low(root, memcg)) {
2534 if (!sc->may_thrash)
2535 continue;
2536 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2537 }
2538
8e8ae645 2539 reclaimed = sc->nr_reclaimed;
cb731d6c 2540 scanned = sc->nr_scanned;
f9be23d6 2541
a9dd0a83
MG
2542 shrink_node_memcg(pgdat, memcg, sc, &lru_pages);
2543 node_lru_pages += lru_pages;
f16015fb 2544
b5afba29 2545 if (memcg)
a9dd0a83 2546 shrink_slab(sc->gfp_mask, pgdat->node_id,
cb731d6c
VD
2547 memcg, sc->nr_scanned - scanned,
2548 lru_pages);
2549
8e8ae645
JW
2550 /* Record the group's reclaim efficiency */
2551 vmpressure(sc->gfp_mask, memcg, false,
2552 sc->nr_scanned - scanned,
2553 sc->nr_reclaimed - reclaimed);
2554
9b4f98cd 2555 /*
a394cb8e
MH
2556 * Direct reclaim and kswapd have to scan all memory
2557 * cgroups to fulfill the overall scan target for the
a9dd0a83 2558 * node.
a394cb8e
MH
2559 *
2560 * Limit reclaim, on the other hand, only cares about
2561 * nr_to_reclaim pages to be reclaimed and it will
2562 * retry with decreasing priority if one round over the
2563 * whole hierarchy is not sufficient.
9b4f98cd 2564 */
a394cb8e
MH
2565 if (!global_reclaim(sc) &&
2566 sc->nr_reclaimed >= sc->nr_to_reclaim) {
9b4f98cd
JW
2567 mem_cgroup_iter_break(root, memcg);
2568 break;
2569 }
241994ed 2570 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
70ddf637 2571
6b4f7799
JW
2572 /*
2573 * Shrink the slab caches in the same proportion that
2574 * the eligible LRU pages were scanned.
2575 */
b2e18757 2576 if (global_reclaim(sc))
a9dd0a83 2577 shrink_slab(sc->gfp_mask, pgdat->node_id, NULL,
cb731d6c 2578 sc->nr_scanned - nr_scanned,
a9dd0a83 2579 node_lru_pages);
cb731d6c
VD
2580
2581 if (reclaim_state) {
2582 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2583 reclaim_state->reclaimed_slab = 0;
6b4f7799
JW
2584 }
2585
8e8ae645
JW
2586 /* Record the subtree's reclaim efficiency */
2587 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
70ddf637
AV
2588 sc->nr_scanned - nr_scanned,
2589 sc->nr_reclaimed - nr_reclaimed);
2590
2344d7e4
JW
2591 if (sc->nr_reclaimed - nr_reclaimed)
2592 reclaimable = true;
2593
a9dd0a83 2594 } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed,
9b4f98cd 2595 sc->nr_scanned - nr_scanned, sc));
2344d7e4
JW
2596
2597 return reclaimable;
f16015fb
JW
2598}
2599
53853e2d 2600/*
fdd4c614
VB
2601 * Returns true if compaction should go ahead for a costly-order request, or
2602 * the allocation would already succeed without compaction. Return false if we
2603 * should reclaim first.
53853e2d 2604 */
4f588331 2605static inline bool compaction_ready(struct zone *zone, struct scan_control *sc)
fe4b1b24 2606{
31483b6a 2607 unsigned long watermark;
fdd4c614 2608 enum compact_result suitable;
fe4b1b24 2609
fdd4c614
VB
2610 suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx);
2611 if (suitable == COMPACT_SUCCESS)
2612 /* Allocation should succeed already. Don't reclaim. */
2613 return true;
2614 if (suitable == COMPACT_SKIPPED)
2615 /* Compaction cannot yet proceed. Do reclaim. */
2616 return false;
fe4b1b24 2617
53853e2d 2618 /*
fdd4c614
VB
2619 * Compaction is already possible, but it takes time to run and there
2620 * are potentially other callers using the pages just freed. So proceed
2621 * with reclaim to make a buffer of free pages available to give
2622 * compaction a reasonable chance of completing and allocating the page.
2623 * Note that we won't actually reclaim the whole buffer in one attempt
2624 * as the target watermark in should_continue_reclaim() is lower. But if
2625 * we are already above the high+gap watermark, don't reclaim at all.
53853e2d 2626 */
fdd4c614 2627 watermark = high_wmark_pages(zone) + compact_gap(sc->order);
fe4b1b24 2628
fdd4c614 2629 return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx);
fe4b1b24
MG
2630}
2631
1da177e4
LT
2632/*
2633 * This is the direct reclaim path, for page-allocating processes. We only
2634 * try to reclaim pages from zones which will satisfy the caller's allocation
2635 * request.
2636 *
1da177e4
LT
2637 * If a zone is deemed to be full of pinned pages then just give it a light
2638 * scan then give up on it.
2639 */
0a0337e0 2640static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
1da177e4 2641{
dd1a239f 2642 struct zoneref *z;
54a6eb5c 2643 struct zone *zone;
0608f43d
AM
2644 unsigned long nr_soft_reclaimed;
2645 unsigned long nr_soft_scanned;
619d0d76 2646 gfp_t orig_mask;
79dafcdc 2647 pg_data_t *last_pgdat = NULL;
1cfb419b 2648
cc715d99
MG
2649 /*
2650 * If the number of buffer_heads in the machine exceeds the maximum
2651 * allowed level, force direct reclaim to scan the highmem zone as
2652 * highmem pages could be pinning lowmem pages storing buffer_heads
2653 */
619d0d76 2654 orig_mask = sc->gfp_mask;
b2e18757 2655 if (buffer_heads_over_limit) {
cc715d99 2656 sc->gfp_mask |= __GFP_HIGHMEM;
4f588331 2657 sc->reclaim_idx = gfp_zone(sc->gfp_mask);
b2e18757 2658 }
cc715d99 2659
d4debc66 2660 for_each_zone_zonelist_nodemask(zone, z, zonelist,
b2e18757 2661 sc->reclaim_idx, sc->nodemask) {
1cfb419b
KH
2662 /*
2663 * Take care memory controller reclaiming has small influence
2664 * to global LRU.
2665 */
89b5fae5 2666 if (global_reclaim(sc)) {
344736f2
VD
2667 if (!cpuset_zone_allowed(zone,
2668 GFP_KERNEL | __GFP_HARDWALL))
1cfb419b 2669 continue;
65ec02cb 2670
6e543d57 2671 if (sc->priority != DEF_PRIORITY &&
599d0c95 2672 !pgdat_reclaimable(zone->zone_pgdat))
1cfb419b 2673 continue; /* Let kswapd poll it */
0b06496a
JW
2674
2675 /*
2676 * If we already have plenty of memory free for
2677 * compaction in this zone, don't free any more.
2678 * Even though compaction is invoked for any
2679 * non-zero order, only frequent costly order
2680 * reclamation is disruptive enough to become a
2681 * noticeable problem, like transparent huge
2682 * page allocations.
2683 */
2684 if (IS_ENABLED(CONFIG_COMPACTION) &&
2685 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
4f588331 2686 compaction_ready(zone, sc)) {
0b06496a
JW
2687 sc->compaction_ready = true;
2688 continue;
e0887c19 2689 }
0b06496a 2690
79dafcdc
MG
2691 /*
2692 * Shrink each node in the zonelist once. If the
2693 * zonelist is ordered by zone (not the default) then a
2694 * node may be shrunk multiple times but in that case
2695 * the user prefers lower zones being preserved.
2696 */
2697 if (zone->zone_pgdat == last_pgdat)
2698 continue;
2699
0608f43d
AM
2700 /*
2701 * This steals pages from memory cgroups over softlimit
2702 * and returns the number of reclaimed pages and
2703 * scanned pages. This works for global memory pressure
2704 * and balancing, not for a memcg's limit.
2705 */
2706 nr_soft_scanned = 0;
ef8f2327 2707 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat,
0608f43d
AM
2708 sc->order, sc->gfp_mask,
2709 &nr_soft_scanned);
2710 sc->nr_reclaimed += nr_soft_reclaimed;
2711 sc->nr_scanned += nr_soft_scanned;
ac34a1a3 2712 /* need some check for avoid more shrink_zone() */
1cfb419b 2713 }
408d8544 2714
79dafcdc
MG
2715 /* See comment about same check for global reclaim above */
2716 if (zone->zone_pgdat == last_pgdat)
2717 continue;
2718 last_pgdat = zone->zone_pgdat;
970a39a3 2719 shrink_node(zone->zone_pgdat, sc);
1da177e4 2720 }
e0c23279 2721
619d0d76
WY
2722 /*
2723 * Restore to original mask to avoid the impact on the caller if we
2724 * promoted it to __GFP_HIGHMEM.
2725 */
2726 sc->gfp_mask = orig_mask;
1da177e4 2727}
4f98a2fe 2728
1da177e4
LT
2729/*
2730 * This is the main entry point to direct page reclaim.
2731 *
2732 * If a full scan of the inactive list fails to free enough memory then we
2733 * are "out of memory" and something needs to be killed.
2734 *
2735 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2736 * high - the zone may be full of dirty or under-writeback pages, which this
5b0830cb
JA
2737 * caller can't do much about. We kick the writeback threads and take explicit
2738 * naps in the hope that some of these pages can be written. But if the
2739 * allocating task holds filesystem locks which prevent writeout this might not
2740 * work, and the allocation attempt will fail.
a41f24ea
NA
2741 *
2742 * returns: 0, if no pages reclaimed
2743 * else, the number of pages reclaimed
1da177e4 2744 */
dac1d27b 2745static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3115cd91 2746 struct scan_control *sc)
1da177e4 2747{
241994ed 2748 int initial_priority = sc->priority;
69e05944 2749 unsigned long total_scanned = 0;
22fba335 2750 unsigned long writeback_threshold;
241994ed 2751retry:
873b4771
KK
2752 delayacct_freepages_start();
2753
89b5fae5 2754 if (global_reclaim(sc))
7cc30fcf 2755 __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1);
1da177e4 2756
9e3b2f8c 2757 do {
70ddf637
AV
2758 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2759 sc->priority);
66e1707b 2760 sc->nr_scanned = 0;
0a0337e0 2761 shrink_zones(zonelist, sc);
c6a8a8c5 2762
66e1707b 2763 total_scanned += sc->nr_scanned;
bb21c7ce 2764 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
0b06496a
JW
2765 break;
2766
2767 if (sc->compaction_ready)
2768 break;
1da177e4 2769
0e50ce3b
MK
2770 /*
2771 * If we're getting trouble reclaiming, start doing
2772 * writepage even in laptop mode.
2773 */
2774 if (sc->priority < DEF_PRIORITY - 2)
2775 sc->may_writepage = 1;
2776
1da177e4
LT
2777 /*
2778 * Try to write back as many pages as we just scanned. This
2779 * tends to cause slow streaming writers to write data to the
2780 * disk smoothly, at the dirtying rate, which is nice. But
2781 * that's undesirable in laptop mode, where we *want* lumpy
2782 * writeout. So in laptop mode, write out the whole world.
2783 */
22fba335
KM
2784 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2785 if (total_scanned > writeback_threshold) {
0e175a18
CW
2786 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2787 WB_REASON_TRY_TO_FREE_PAGES);
66e1707b 2788 sc->may_writepage = 1;
1da177e4 2789 }
0b06496a 2790 } while (--sc->priority >= 0);
bb21c7ce 2791
873b4771
KK
2792 delayacct_freepages_end();
2793
bb21c7ce
KM
2794 if (sc->nr_reclaimed)
2795 return sc->nr_reclaimed;
2796
0cee34fd 2797 /* Aborted reclaim to try compaction? don't OOM, then */
0b06496a 2798 if (sc->compaction_ready)
7335084d
MG
2799 return 1;
2800
241994ed
JW
2801 /* Untapped cgroup reserves? Don't OOM, retry. */
2802 if (!sc->may_thrash) {
2803 sc->priority = initial_priority;
2804 sc->may_thrash = 1;
2805 goto retry;
2806 }
2807
bb21c7ce 2808 return 0;
1da177e4
LT
2809}
2810
5515061d
MG
2811static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2812{
2813 struct zone *zone;
2814 unsigned long pfmemalloc_reserve = 0;
2815 unsigned long free_pages = 0;
2816 int i;
2817 bool wmark_ok;
2818
2819 for (i = 0; i <= ZONE_NORMAL; i++) {
2820 zone = &pgdat->node_zones[i];
6aa303de 2821 if (!managed_zone(zone) ||
599d0c95 2822 pgdat_reclaimable_pages(pgdat) == 0)
675becce
MG
2823 continue;
2824
5515061d
MG
2825 pfmemalloc_reserve += min_wmark_pages(zone);
2826 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2827 }
2828
675becce
MG
2829 /* If there are no reserves (unexpected config) then do not throttle */
2830 if (!pfmemalloc_reserve)
2831 return true;
2832
5515061d
MG
2833 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2834
2835 /* kswapd must be awake if processes are being throttled */
2836 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
38087d9b 2837 pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx,
5515061d
MG
2838 (enum zone_type)ZONE_NORMAL);
2839 wake_up_interruptible(&pgdat->kswapd_wait);
2840 }
2841
2842 return wmark_ok;
2843}
2844
2845/*
2846 * Throttle direct reclaimers if backing storage is backed by the network
2847 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2848 * depleted. kswapd will continue to make progress and wake the processes
50694c28
MG
2849 * when the low watermark is reached.
2850 *
2851 * Returns true if a fatal signal was delivered during throttling. If this
2852 * happens, the page allocator should not consider triggering the OOM killer.
5515061d 2853 */
50694c28 2854static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
5515061d
MG
2855 nodemask_t *nodemask)
2856{
675becce 2857 struct zoneref *z;
5515061d 2858 struct zone *zone;
675becce 2859 pg_data_t *pgdat = NULL;
5515061d
MG
2860
2861 /*
2862 * Kernel threads should not be throttled as they may be indirectly
2863 * responsible for cleaning pages necessary for reclaim to make forward
2864 * progress. kjournald for example may enter direct reclaim while
2865 * committing a transaction where throttling it could forcing other
2866 * processes to block on log_wait_commit().
2867 */
2868 if (current->flags & PF_KTHREAD)
50694c28
MG
2869 goto out;
2870
2871 /*
2872 * If a fatal signal is pending, this process should not throttle.
2873 * It should return quickly so it can exit and free its memory
2874 */
2875 if (fatal_signal_pending(current))
2876 goto out;
5515061d 2877
675becce
MG
2878 /*
2879 * Check if the pfmemalloc reserves are ok by finding the first node
2880 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2881 * GFP_KERNEL will be required for allocating network buffers when
2882 * swapping over the network so ZONE_HIGHMEM is unusable.
2883 *
2884 * Throttling is based on the first usable node and throttled processes
2885 * wait on a queue until kswapd makes progress and wakes them. There
2886 * is an affinity then between processes waking up and where reclaim
2887 * progress has been made assuming the process wakes on the same node.
2888 * More importantly, processes running on remote nodes will not compete
2889 * for remote pfmemalloc reserves and processes on different nodes
2890 * should make reasonable progress.
2891 */
2892 for_each_zone_zonelist_nodemask(zone, z, zonelist,
17636faa 2893 gfp_zone(gfp_mask), nodemask) {
675becce
MG
2894 if (zone_idx(zone) > ZONE_NORMAL)
2895 continue;
2896
2897 /* Throttle based on the first usable node */
2898 pgdat = zone->zone_pgdat;
2899 if (pfmemalloc_watermark_ok(pgdat))
2900 goto out;
2901 break;
2902 }
2903
2904 /* If no zone was usable by the allocation flags then do not throttle */
2905 if (!pgdat)
50694c28 2906 goto out;
5515061d 2907
68243e76
MG
2908 /* Account for the throttling */
2909 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2910
5515061d
MG
2911 /*
2912 * If the caller cannot enter the filesystem, it's possible that it
2913 * is due to the caller holding an FS lock or performing a journal
2914 * transaction in the case of a filesystem like ext[3|4]. In this case,
2915 * it is not safe to block on pfmemalloc_wait as kswapd could be
2916 * blocked waiting on the same lock. Instead, throttle for up to a
2917 * second before continuing.
2918 */
2919 if (!(gfp_mask & __GFP_FS)) {
2920 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2921 pfmemalloc_watermark_ok(pgdat), HZ);
50694c28
MG
2922
2923 goto check_pending;
5515061d
MG
2924 }
2925
2926 /* Throttle until kswapd wakes the process */
2927 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2928 pfmemalloc_watermark_ok(pgdat));
50694c28
MG
2929
2930check_pending:
2931 if (fatal_signal_pending(current))
2932 return true;
2933
2934out:
2935 return false;
5515061d
MG
2936}
2937
dac1d27b 2938unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
327c0e96 2939 gfp_t gfp_mask, nodemask_t *nodemask)
66e1707b 2940{
33906bc5 2941 unsigned long nr_reclaimed;
66e1707b 2942 struct scan_control sc = {
ee814fe2 2943 .nr_to_reclaim = SWAP_CLUSTER_MAX,
21caf2fc 2944 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
b2e18757 2945 .reclaim_idx = gfp_zone(gfp_mask),
ee814fe2
JW
2946 .order = order,
2947 .nodemask = nodemask,
2948 .priority = DEF_PRIORITY,
66e1707b 2949 .may_writepage = !laptop_mode,
a6dc60f8 2950 .may_unmap = 1,
2e2e4259 2951 .may_swap = 1,
66e1707b
BS
2952 };
2953
5515061d 2954 /*
50694c28
MG
2955 * Do not enter reclaim if fatal signal was delivered while throttled.
2956 * 1 is returned so that the page allocator does not OOM kill at this
2957 * point.
5515061d 2958 */
50694c28 2959 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
5515061d
MG
2960 return 1;
2961
33906bc5
MG
2962 trace_mm_vmscan_direct_reclaim_begin(order,
2963 sc.may_writepage,
e5146b12
MG
2964 gfp_mask,
2965 sc.reclaim_idx);
33906bc5 2966
3115cd91 2967 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
33906bc5
MG
2968
2969 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2970
2971 return nr_reclaimed;
66e1707b
BS
2972}
2973
c255a458 2974#ifdef CONFIG_MEMCG
66e1707b 2975
a9dd0a83 2976unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg,
4e416953 2977 gfp_t gfp_mask, bool noswap,
ef8f2327 2978 pg_data_t *pgdat,
0ae5e89c 2979 unsigned long *nr_scanned)
4e416953
BS
2980{
2981 struct scan_control sc = {
b8f5c566 2982 .nr_to_reclaim = SWAP_CLUSTER_MAX,
ee814fe2 2983 .target_mem_cgroup = memcg,
4e416953
BS
2984 .may_writepage = !laptop_mode,
2985 .may_unmap = 1,
b2e18757 2986 .reclaim_idx = MAX_NR_ZONES - 1,
4e416953 2987 .may_swap = !noswap,
4e416953 2988 };
6b4f7799 2989 unsigned long lru_pages;
0ae5e89c 2990
4e416953
BS
2991 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2992 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
bdce6d9e 2993
9e3b2f8c 2994 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
bdce6d9e 2995 sc.may_writepage,
e5146b12
MG
2996 sc.gfp_mask,
2997 sc.reclaim_idx);
bdce6d9e 2998
4e416953
BS
2999 /*
3000 * NOTE: Although we can get the priority field, using it
3001 * here is not a good idea, since it limits the pages we can scan.
a9dd0a83 3002 * if we don't reclaim here, the shrink_node from balance_pgdat
4e416953
BS
3003 * will pick up pages from other mem cgroup's as well. We hack
3004 * the priority and make it zero.
3005 */
ef8f2327 3006 shrink_node_memcg(pgdat, memcg, &sc, &lru_pages);
bdce6d9e
KM
3007
3008 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
3009
0ae5e89c 3010 *nr_scanned = sc.nr_scanned;
4e416953
BS
3011 return sc.nr_reclaimed;
3012}
3013
72835c86 3014unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
b70a2a21 3015 unsigned long nr_pages,
a7885eb8 3016 gfp_t gfp_mask,
b70a2a21 3017 bool may_swap)
66e1707b 3018{
4e416953 3019 struct zonelist *zonelist;
bdce6d9e 3020 unsigned long nr_reclaimed;
889976db 3021 int nid;
66e1707b 3022 struct scan_control sc = {
b70a2a21 3023 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
a09ed5e0
YH
3024 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
3025 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
b2e18757 3026 .reclaim_idx = MAX_NR_ZONES - 1,
ee814fe2
JW
3027 .target_mem_cgroup = memcg,
3028 .priority = DEF_PRIORITY,
3029 .may_writepage = !laptop_mode,
3030 .may_unmap = 1,
b70a2a21 3031 .may_swap = may_swap,
a09ed5e0 3032 };
66e1707b 3033
889976db
YH
3034 /*
3035 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
3036 * take care of from where we get pages. So the node where we start the
3037 * scan does not need to be the current node.
3038 */
72835c86 3039 nid = mem_cgroup_select_victim_node(memcg);
889976db 3040
c9634cf0 3041 zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK];
bdce6d9e
KM
3042
3043 trace_mm_vmscan_memcg_reclaim_begin(0,
3044 sc.may_writepage,
e5146b12
MG
3045 sc.gfp_mask,
3046 sc.reclaim_idx);
bdce6d9e 3047
89a28483 3048 current->flags |= PF_MEMALLOC;
3115cd91 3049 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
89a28483 3050 current->flags &= ~PF_MEMALLOC;
bdce6d9e
KM
3051
3052 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
3053
3054 return nr_reclaimed;
66e1707b
BS
3055}
3056#endif
3057
1d82de61 3058static void age_active_anon(struct pglist_data *pgdat,
ef8f2327 3059 struct scan_control *sc)
f16015fb 3060{
b95a2f2d 3061 struct mem_cgroup *memcg;
f16015fb 3062
b95a2f2d
JW
3063 if (!total_swap_pages)
3064 return;
3065
3066 memcg = mem_cgroup_iter(NULL, NULL, NULL);
3067 do {
ef8f2327 3068 struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg);
b95a2f2d 3069
f8d1a311 3070 if (inactive_list_is_low(lruvec, false, sc))
1a93be0e 3071 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
9e3b2f8c 3072 sc, LRU_ACTIVE_ANON);
b95a2f2d
JW
3073
3074 memcg = mem_cgroup_iter(NULL, memcg, NULL);
3075 } while (memcg);
f16015fb
JW
3076}
3077
31483b6a 3078static bool zone_balanced(struct zone *zone, int order, int classzone_idx)
60cefed4 3079{
31483b6a 3080 unsigned long mark = high_wmark_pages(zone);
60cefed4 3081
6256c6b4
MG
3082 if (!zone_watermark_ok_safe(zone, order, mark, classzone_idx))
3083 return false;
3084
3085 /*
3086 * If any eligible zone is balanced then the node is not considered
3087 * to be congested or dirty
3088 */
3089 clear_bit(PGDAT_CONGESTED, &zone->zone_pgdat->flags);
3090 clear_bit(PGDAT_DIRTY, &zone->zone_pgdat->flags);
3091
3092 return true;
60cefed4
JW
3093}
3094
5515061d
MG
3095/*
3096 * Prepare kswapd for sleeping. This verifies that there are no processes
3097 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3098 *
3099 * Returns true if kswapd is ready to sleep
3100 */
d9f21d42 3101static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx)
f50de2d3 3102{
1d82de61
MG
3103 int i;
3104
5515061d 3105 /*
9e5e3661
VB
3106 * The throttled processes are normally woken up in balance_pgdat() as
3107 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3108 * race between when kswapd checks the watermarks and a process gets
3109 * throttled. There is also a potential race if processes get
3110 * throttled, kswapd wakes, a large process exits thereby balancing the
3111 * zones, which causes kswapd to exit balance_pgdat() before reaching
3112 * the wake up checks. If kswapd is going to sleep, no process should
3113 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3114 * the wake up is premature, processes will wake kswapd and get
3115 * throttled again. The difference from wake ups in balance_pgdat() is
3116 * that here we are under prepare_to_wait().
5515061d 3117 */
9e5e3661
VB
3118 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3119 wake_up_all(&pgdat->pfmemalloc_wait);
f50de2d3 3120
1d82de61
MG
3121 for (i = 0; i <= classzone_idx; i++) {
3122 struct zone *zone = pgdat->node_zones + i;
3123
6aa303de 3124 if (!managed_zone(zone))
1d82de61
MG
3125 continue;
3126
38087d9b
MG
3127 if (!zone_balanced(zone, order, classzone_idx))
3128 return false;
1d82de61
MG
3129 }
3130
38087d9b 3131 return true;
f50de2d3
MG
3132}
3133
75485363 3134/*
1d82de61
MG
3135 * kswapd shrinks a node of pages that are at or below the highest usable
3136 * zone that is currently unbalanced.
b8e83b94
MG
3137 *
3138 * Returns true if kswapd scanned at least the requested number of pages to
283aba9f
MG
3139 * reclaim or if the lack of progress was due to pages under writeback.
3140 * This is used to determine if the scanning priority needs to be raised.
75485363 3141 */
1d82de61 3142static bool kswapd_shrink_node(pg_data_t *pgdat,
accf6242 3143 struct scan_control *sc)
75485363 3144{
1d82de61
MG
3145 struct zone *zone;
3146 int z;
75485363 3147
1d82de61
MG
3148 /* Reclaim a number of pages proportional to the number of zones */
3149 sc->nr_to_reclaim = 0;
970a39a3 3150 for (z = 0; z <= sc->reclaim_idx; z++) {
1d82de61 3151 zone = pgdat->node_zones + z;
6aa303de 3152 if (!managed_zone(zone))
1d82de61 3153 continue;
7c954f6d 3154
1d82de61
MG
3155 sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX);
3156 }
7c954f6d
MG
3157
3158 /*
1d82de61
MG
3159 * Historically care was taken to put equal pressure on all zones but
3160 * now pressure is applied based on node LRU order.
7c954f6d 3161 */
970a39a3 3162 shrink_node(pgdat, sc);
283aba9f 3163
7c954f6d 3164 /*
1d82de61
MG
3165 * Fragmentation may mean that the system cannot be rebalanced for
3166 * high-order allocations. If twice the allocation size has been
3167 * reclaimed then recheck watermarks only at order-0 to prevent
3168 * excessive reclaim. Assume that a process requested a high-order
3169 * can direct reclaim/compact.
7c954f6d 3170 */
9861a62c 3171 if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
1d82de61 3172 sc->order = 0;
7c954f6d 3173
b8e83b94 3174 return sc->nr_scanned >= sc->nr_to_reclaim;
75485363
MG
3175}
3176
1da177e4 3177/*
1d82de61
MG
3178 * For kswapd, balance_pgdat() will reclaim pages across a node from zones
3179 * that are eligible for use by the caller until at least one zone is
3180 * balanced.
1da177e4 3181 *
1d82de61 3182 * Returns the order kswapd finished reclaiming at.
1da177e4
LT
3183 *
3184 * kswapd scans the zones in the highmem->normal->dma direction. It skips
41858966 3185 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
1d82de61
MG
3186 * found to have free_pages <= high_wmark_pages(zone), any page is that zone
3187 * or lower is eligible for reclaim until at least one usable zone is
3188 * balanced.
1da177e4 3189 */
accf6242 3190static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
1da177e4 3191{
1da177e4 3192 int i;
0608f43d
AM
3193 unsigned long nr_soft_reclaimed;
3194 unsigned long nr_soft_scanned;
1d82de61 3195 struct zone *zone;
179e9639
AM
3196 struct scan_control sc = {
3197 .gfp_mask = GFP_KERNEL,
ee814fe2 3198 .order = order,
b8e83b94 3199 .priority = DEF_PRIORITY,
ee814fe2 3200 .may_writepage = !laptop_mode,
a6dc60f8 3201 .may_unmap = 1,
2e2e4259 3202 .may_swap = 1,
179e9639 3203 };
f8891e5e 3204 count_vm_event(PAGEOUTRUN);
1da177e4 3205
9e3b2f8c 3206 do {
b8e83b94
MG
3207 bool raise_priority = true;
3208
3209 sc.nr_reclaimed = 0;
84c7a777 3210 sc.reclaim_idx = classzone_idx;
1da177e4 3211
86c79f6b 3212 /*
84c7a777
MG
3213 * If the number of buffer_heads exceeds the maximum allowed
3214 * then consider reclaiming from all zones. This has a dual
3215 * purpose -- on 64-bit systems it is expected that
3216 * buffer_heads are stripped during active rotation. On 32-bit
3217 * systems, highmem pages can pin lowmem memory and shrinking
3218 * buffers can relieve lowmem pressure. Reclaim may still not
3219 * go ahead if all eligible zones for the original allocation
3220 * request are balanced to avoid excessive reclaim from kswapd.
86c79f6b
MG
3221 */
3222 if (buffer_heads_over_limit) {
3223 for (i = MAX_NR_ZONES - 1; i >= 0; i--) {
3224 zone = pgdat->node_zones + i;
6aa303de 3225 if (!managed_zone(zone))
86c79f6b 3226 continue;
cc715d99 3227
970a39a3 3228 sc.reclaim_idx = i;
e1dbeda6 3229 break;
1da177e4 3230 }
1da177e4 3231 }
dafcb73e 3232
86c79f6b
MG
3233 /*
3234 * Only reclaim if there are no eligible zones. Check from
3235 * high to low zone as allocations prefer higher zones.
3236 * Scanning from low to high zone would allow congestion to be
3237 * cleared during a very small window when a small low
3238 * zone was balanced even under extreme pressure when the
84c7a777
MG
3239 * overall node may be congested. Note that sc.reclaim_idx
3240 * is not used as buffer_heads_over_limit may have adjusted
3241 * it.
86c79f6b 3242 */
84c7a777 3243 for (i = classzone_idx; i >= 0; i--) {
86c79f6b 3244 zone = pgdat->node_zones + i;
6aa303de 3245 if (!managed_zone(zone))
86c79f6b
MG
3246 continue;
3247
84c7a777 3248 if (zone_balanced(zone, sc.order, classzone_idx))
86c79f6b
MG
3249 goto out;
3250 }
e1dbeda6 3251
1d82de61
MG
3252 /*
3253 * Do some background aging of the anon list, to give
3254 * pages a chance to be referenced before reclaiming. All
3255 * pages are rotated regardless of classzone as this is
3256 * about consistent aging.
3257 */
ef8f2327 3258 age_active_anon(pgdat, &sc);
1d82de61 3259
b7ea3c41
MG
3260 /*
3261 * If we're getting trouble reclaiming, start doing writepage
3262 * even in laptop mode.
3263 */
1d82de61 3264 if (sc.priority < DEF_PRIORITY - 2 || !pgdat_reclaimable(pgdat))
b7ea3c41
MG
3265 sc.may_writepage = 1;
3266
1d82de61
MG
3267 /* Call soft limit reclaim before calling shrink_node. */
3268 sc.nr_scanned = 0;
3269 nr_soft_scanned = 0;
ef8f2327 3270 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order,
1d82de61
MG
3271 sc.gfp_mask, &nr_soft_scanned);
3272 sc.nr_reclaimed += nr_soft_reclaimed;
3273
1da177e4 3274 /*
1d82de61
MG
3275 * There should be no need to raise the scanning priority if
3276 * enough pages are already being scanned that that high
3277 * watermark would be met at 100% efficiency.
1da177e4 3278 */
970a39a3 3279 if (kswapd_shrink_node(pgdat, &sc))
1d82de61 3280 raise_priority = false;
5515061d
MG
3281
3282 /*
3283 * If the low watermark is met there is no need for processes
3284 * to be throttled on pfmemalloc_wait as they should not be
3285 * able to safely make forward progress. Wake them
3286 */
3287 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3288 pfmemalloc_watermark_ok(pgdat))
cfc51155 3289 wake_up_all(&pgdat->pfmemalloc_wait);
5515061d 3290
b8e83b94
MG
3291 /* Check if kswapd should be suspending */
3292 if (try_to_freeze() || kthread_should_stop())
3293 break;
8357376d 3294
73ce02e9 3295 /*
b8e83b94
MG
3296 * Raise priority if scanning rate is too low or there was no
3297 * progress in reclaiming pages
73ce02e9 3298 */
b8e83b94
MG
3299 if (raise_priority || !sc.nr_reclaimed)
3300 sc.priority--;
1d82de61 3301 } while (sc.priority >= 1);
1da177e4 3302
b8e83b94 3303out:
0abdee2b 3304 /*
1d82de61
MG
3305 * Return the order kswapd stopped reclaiming at as
3306 * prepare_kswapd_sleep() takes it into account. If another caller
3307 * entered the allocator slow path while kswapd was awake, order will
3308 * remain at the higher level.
0abdee2b 3309 */
1d82de61 3310 return sc.order;
1da177e4
LT
3311}
3312
38087d9b
MG
3313static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order,
3314 unsigned int classzone_idx)
f0bc0a60
KM
3315{
3316 long remaining = 0;
3317 DEFINE_WAIT(wait);
3318
3319 if (freezing(current) || kthread_should_stop())
3320 return;
3321
3322 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3323
3324 /* Try to sleep for a short interval */
d9f21d42 3325 if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
fd901c95
VB
3326 /*
3327 * Compaction records what page blocks it recently failed to
3328 * isolate pages from and skips them in the future scanning.
3329 * When kswapd is going to sleep, it is reasonable to assume
3330 * that pages and compaction may succeed so reset the cache.
3331 */
3332 reset_isolation_suitable(pgdat);
3333
3334 /*
3335 * We have freed the memory, now we should compact it to make
3336 * allocation of the requested order possible.
3337 */
38087d9b 3338 wakeup_kcompactd(pgdat, alloc_order, classzone_idx);
fd901c95 3339
f0bc0a60 3340 remaining = schedule_timeout(HZ/10);
38087d9b
MG
3341
3342 /*
3343 * If woken prematurely then reset kswapd_classzone_idx and
3344 * order. The values will either be from a wakeup request or
3345 * the previous request that slept prematurely.
3346 */
3347 if (remaining) {
3348 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3349 pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order);
3350 }
3351
f0bc0a60
KM
3352 finish_wait(&pgdat->kswapd_wait, &wait);
3353 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3354 }
3355
3356 /*
3357 * After a short sleep, check if it was a premature sleep. If not, then
3358 * go fully to sleep until explicitly woken up.
3359 */
d9f21d42
MG
3360 if (!remaining &&
3361 prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) {
f0bc0a60
KM
3362 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3363
3364 /*
3365 * vmstat counters are not perfectly accurate and the estimated
3366 * value for counters such as NR_FREE_PAGES can deviate from the
3367 * true value by nr_online_cpus * threshold. To avoid the zone
3368 * watermarks being breached while under pressure, we reduce the
3369 * per-cpu vmstat threshold while kswapd is awake and restore
3370 * them before going back to sleep.
3371 */
3372 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
1c7e7f6c
AK
3373
3374 if (!kthread_should_stop())
3375 schedule();
3376
f0bc0a60
KM
3377 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3378 } else {
3379 if (remaining)
3380 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3381 else
3382 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3383 }
3384 finish_wait(&pgdat->kswapd_wait, &wait);
3385}
3386
1da177e4
LT
3387/*
3388 * The background pageout daemon, started as a kernel thread
4f98a2fe 3389 * from the init process.
1da177e4
LT
3390 *
3391 * This basically trickles out pages so that we have _some_
3392 * free memory available even if there is no other activity
3393 * that frees anything up. This is needed for things like routing
3394 * etc, where we otherwise might have all activity going on in
3395 * asynchronous contexts that cannot page things out.
3396 *
3397 * If there are applications that are active memory-allocators
3398 * (most normal use), this basically shouldn't matter.
3399 */
3400static int kswapd(void *p)
3401{
38087d9b 3402 unsigned int alloc_order, reclaim_order, classzone_idx;
1da177e4
LT
3403 pg_data_t *pgdat = (pg_data_t*)p;
3404 struct task_struct *tsk = current;
f0bc0a60 3405
1da177e4
LT
3406 struct reclaim_state reclaim_state = {
3407 .reclaimed_slab = 0,
3408 };
a70f7302 3409 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1da177e4 3410
cf40bd16
NP
3411 lockdep_set_current_reclaim_state(GFP_KERNEL);
3412
174596a0 3413 if (!cpumask_empty(cpumask))
c5f59f08 3414 set_cpus_allowed_ptr(tsk, cpumask);
1da177e4
LT
3415 current->reclaim_state = &reclaim_state;
3416
3417 /*
3418 * Tell the memory management that we're a "memory allocator",
3419 * and that if we need more memory we should get access to it
3420 * regardless (see "__alloc_pages()"). "kswapd" should
3421 * never get caught in the normal page freeing logic.
3422 *
3423 * (Kswapd normally doesn't need memory anyway, but sometimes
3424 * you need a small amount of memory in order to be able to
3425 * page out something else, and this flag essentially protects
3426 * us from recursively trying to free more memory as we're
3427 * trying to free the first piece of memory in the first place).
3428 */
930d9152 3429 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
83144186 3430 set_freezable();
1da177e4 3431
38087d9b
MG
3432 pgdat->kswapd_order = alloc_order = reclaim_order = 0;
3433 pgdat->kswapd_classzone_idx = classzone_idx = 0;
1da177e4 3434 for ( ; ; ) {
6f6313d4 3435 bool ret;
3e1d1d28 3436
38087d9b
MG
3437kswapd_try_sleep:
3438 kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order,
3439 classzone_idx);
215ddd66 3440
38087d9b
MG
3441 /* Read the new order and classzone_idx */
3442 alloc_order = reclaim_order = pgdat->kswapd_order;
3443 classzone_idx = pgdat->kswapd_classzone_idx;
3444 pgdat->kswapd_order = 0;
3445 pgdat->kswapd_classzone_idx = 0;
1da177e4 3446
8fe23e05
DR
3447 ret = try_to_freeze();
3448 if (kthread_should_stop())
3449 break;
3450
3451 /*
3452 * We can speed up thawing tasks if we don't call balance_pgdat
3453 * after returning from the refrigerator
3454 */
38087d9b
MG
3455 if (ret)
3456 continue;
3457
3458 /*
3459 * Reclaim begins at the requested order but if a high-order
3460 * reclaim fails then kswapd falls back to reclaiming for
3461 * order-0. If that happens, kswapd will consider sleeping
3462 * for the order it finished reclaiming at (reclaim_order)
3463 * but kcompactd is woken to compact for the original
3464 * request (alloc_order).
3465 */
e5146b12
MG
3466 trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx,
3467 alloc_order);
38087d9b
MG
3468 reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx);
3469 if (reclaim_order < alloc_order)
3470 goto kswapd_try_sleep;
1d82de61 3471
38087d9b
MG
3472 alloc_order = reclaim_order = pgdat->kswapd_order;
3473 classzone_idx = pgdat->kswapd_classzone_idx;
1da177e4 3474 }
b0a8cc58 3475
71abdc15 3476 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
b0a8cc58 3477 current->reclaim_state = NULL;
71abdc15
JW
3478 lockdep_clear_current_reclaim_state();
3479
1da177e4
LT
3480 return 0;
3481}
3482
3483/*
3484 * A zone is low on free memory, so wake its kswapd task to service it.
3485 */
99504748 3486void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
1da177e4
LT
3487{
3488 pg_data_t *pgdat;
e1a55637 3489 int z;
1da177e4 3490
6aa303de 3491 if (!managed_zone(zone))
1da177e4
LT
3492 return;
3493
344736f2 3494 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
1da177e4 3495 return;
88f5acf8 3496 pgdat = zone->zone_pgdat;
38087d9b
MG
3497 pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, classzone_idx);
3498 pgdat->kswapd_order = max(pgdat->kswapd_order, order);
8d0986e2 3499 if (!waitqueue_active(&pgdat->kswapd_wait))
1da177e4 3500 return;
e1a55637
MG
3501
3502 /* Only wake kswapd if all zones are unbalanced */
3503 for (z = 0; z <= classzone_idx; z++) {
3504 zone = pgdat->node_zones + z;
6aa303de 3505 if (!managed_zone(zone))
e1a55637
MG
3506 continue;
3507
3508 if (zone_balanced(zone, order, classzone_idx))
3509 return;
3510 }
88f5acf8
MG
3511
3512 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
8d0986e2 3513 wake_up_interruptible(&pgdat->kswapd_wait);
1da177e4
LT
3514}
3515
c6f37f12 3516#ifdef CONFIG_HIBERNATION
1da177e4 3517/*
7b51755c 3518 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
d6277db4
RW
3519 * freed pages.
3520 *
3521 * Rather than trying to age LRUs the aim is to preserve the overall
3522 * LRU order by reclaiming preferentially
3523 * inactive > active > active referenced > active mapped
1da177e4 3524 */
7b51755c 3525unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
1da177e4 3526{
d6277db4 3527 struct reclaim_state reclaim_state;
d6277db4 3528 struct scan_control sc = {
ee814fe2 3529 .nr_to_reclaim = nr_to_reclaim,
7b51755c 3530 .gfp_mask = GFP_HIGHUSER_MOVABLE,
b2e18757 3531 .reclaim_idx = MAX_NR_ZONES - 1,
ee814fe2 3532 .priority = DEF_PRIORITY,
d6277db4 3533 .may_writepage = 1,
ee814fe2
JW
3534 .may_unmap = 1,
3535 .may_swap = 1,
7b51755c 3536 .hibernation_mode = 1,
1da177e4 3537 };
a09ed5e0 3538 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7b51755c
KM
3539 struct task_struct *p = current;
3540 unsigned long nr_reclaimed;
1da177e4 3541
7b51755c
KM
3542 p->flags |= PF_MEMALLOC;
3543 lockdep_set_current_reclaim_state(sc.gfp_mask);
3544 reclaim_state.reclaimed_slab = 0;
3545 p->reclaim_state = &reclaim_state;
d6277db4 3546
3115cd91 3547 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
d979677c 3548
7b51755c
KM
3549 p->reclaim_state = NULL;
3550 lockdep_clear_current_reclaim_state();
3551 p->flags &= ~PF_MEMALLOC;
d6277db4 3552
7b51755c 3553 return nr_reclaimed;
1da177e4 3554}
c6f37f12 3555#endif /* CONFIG_HIBERNATION */
1da177e4 3556
1da177e4
LT
3557/* It's optimal to keep kswapds on the same CPUs as their memory, but
3558 not required for correctness. So if the last cpu in a node goes
3559 away, we get changed to run anywhere: as the first one comes back,
3560 restore their cpu bindings. */
fcb35a9b
GKH
3561static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3562 void *hcpu)
1da177e4 3563{
58c0a4a7 3564 int nid;
1da177e4 3565
8bb78442 3566 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
48fb2e24 3567 for_each_node_state(nid, N_MEMORY) {
c5f59f08 3568 pg_data_t *pgdat = NODE_DATA(nid);
a70f7302
RR
3569 const struct cpumask *mask;
3570
3571 mask = cpumask_of_node(pgdat->node_id);
c5f59f08 3572
3e597945 3573 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
1da177e4 3574 /* One of our CPUs online: restore mask */
c5f59f08 3575 set_cpus_allowed_ptr(pgdat->kswapd, mask);
1da177e4
LT
3576 }
3577 }
3578 return NOTIFY_OK;
3579}
1da177e4 3580
3218ae14
YG
3581/*
3582 * This kswapd start function will be called by init and node-hot-add.
3583 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3584 */
3585int kswapd_run(int nid)
3586{
3587 pg_data_t *pgdat = NODE_DATA(nid);
3588 int ret = 0;
3589
3590 if (pgdat->kswapd)
3591 return 0;
3592
3593 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3594 if (IS_ERR(pgdat->kswapd)) {
3595 /* failure at boot is fatal */
3596 BUG_ON(system_state == SYSTEM_BOOTING);
d5dc0ad9
GS
3597 pr_err("Failed to start kswapd on node %d\n", nid);
3598 ret = PTR_ERR(pgdat->kswapd);
d72515b8 3599 pgdat->kswapd = NULL;
3218ae14
YG
3600 }
3601 return ret;
3602}
3603
8fe23e05 3604/*
d8adde17 3605 * Called by memory hotplug when all memory in a node is offlined. Caller must
bfc8c901 3606 * hold mem_hotplug_begin/end().
8fe23e05
DR
3607 */
3608void kswapd_stop(int nid)
3609{
3610 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3611
d8adde17 3612 if (kswapd) {
8fe23e05 3613 kthread_stop(kswapd);
d8adde17
JL
3614 NODE_DATA(nid)->kswapd = NULL;
3615 }
8fe23e05
DR
3616}
3617
1da177e4
LT
3618static int __init kswapd_init(void)
3619{
3218ae14 3620 int nid;
69e05944 3621
1da177e4 3622 swap_setup();
48fb2e24 3623 for_each_node_state(nid, N_MEMORY)
3218ae14 3624 kswapd_run(nid);
1da177e4
LT
3625 hotcpu_notifier(cpu_callback, 0);
3626 return 0;
3627}
3628
3629module_init(kswapd_init)
9eeff239
CL
3630
3631#ifdef CONFIG_NUMA
3632/*
a5f5f91d 3633 * Node reclaim mode
9eeff239 3634 *
a5f5f91d 3635 * If non-zero call node_reclaim when the number of free pages falls below
9eeff239 3636 * the watermarks.
9eeff239 3637 */
a5f5f91d 3638int node_reclaim_mode __read_mostly;
9eeff239 3639
1b2ffb78 3640#define RECLAIM_OFF 0
7d03431c 3641#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
1b2ffb78 3642#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
95bbc0c7 3643#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
1b2ffb78 3644
a92f7126 3645/*
a5f5f91d 3646 * Priority for NODE_RECLAIM. This determines the fraction of pages
a92f7126
CL
3647 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3648 * a zone.
3649 */
a5f5f91d 3650#define NODE_RECLAIM_PRIORITY 4
a92f7126 3651
9614634f 3652/*
a5f5f91d 3653 * Percentage of pages in a zone that must be unmapped for node_reclaim to
9614634f
CL
3654 * occur.
3655 */
3656int sysctl_min_unmapped_ratio = 1;
3657
0ff38490
CL
3658/*
3659 * If the number of slab pages in a zone grows beyond this percentage then
3660 * slab reclaim needs to occur.
3661 */
3662int sysctl_min_slab_ratio = 5;
3663
11fb9989 3664static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat)
90afa5de 3665{
11fb9989
MG
3666 unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED);
3667 unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) +
3668 node_page_state(pgdat, NR_ACTIVE_FILE);
90afa5de
MG
3669
3670 /*
3671 * It's possible for there to be more file mapped pages than
3672 * accounted for by the pages on the file LRU lists because
3673 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3674 */
3675 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3676}
3677
3678/* Work out how many page cache pages we can reclaim in this reclaim_mode */
a5f5f91d 3679static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat)
90afa5de 3680{
d031a157
AM
3681 unsigned long nr_pagecache_reclaimable;
3682 unsigned long delta = 0;
90afa5de
MG
3683
3684 /*
95bbc0c7 3685 * If RECLAIM_UNMAP is set, then all file pages are considered
90afa5de 3686 * potentially reclaimable. Otherwise, we have to worry about
11fb9989 3687 * pages like swapcache and node_unmapped_file_pages() provides
90afa5de
MG
3688 * a better estimate
3689 */
a5f5f91d
MG
3690 if (node_reclaim_mode & RECLAIM_UNMAP)
3691 nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES);
90afa5de 3692 else
a5f5f91d 3693 nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat);
90afa5de
MG
3694
3695 /* If we can't clean pages, remove dirty pages from consideration */
a5f5f91d
MG
3696 if (!(node_reclaim_mode & RECLAIM_WRITE))
3697 delta += node_page_state(pgdat, NR_FILE_DIRTY);
90afa5de
MG
3698
3699 /* Watch for any possible underflows due to delta */
3700 if (unlikely(delta > nr_pagecache_reclaimable))
3701 delta = nr_pagecache_reclaimable;
3702
3703 return nr_pagecache_reclaimable - delta;
3704}
3705
9eeff239 3706/*
a5f5f91d 3707 * Try to free up some pages from this node through reclaim.
9eeff239 3708 */
a5f5f91d 3709static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
9eeff239 3710{
7fb2d46d 3711 /* Minimum pages needed in order to stay on node */
69e05944 3712 const unsigned long nr_pages = 1 << order;
9eeff239
CL
3713 struct task_struct *p = current;
3714 struct reclaim_state reclaim_state;
a5f5f91d 3715 int classzone_idx = gfp_zone(gfp_mask);
179e9639 3716 struct scan_control sc = {
62b726c1 3717 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
21caf2fc 3718 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
bd2f6199 3719 .order = order,
a5f5f91d
MG
3720 .priority = NODE_RECLAIM_PRIORITY,
3721 .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE),
3722 .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP),
ee814fe2 3723 .may_swap = 1,
a5f5f91d 3724 .reclaim_idx = classzone_idx,
179e9639 3725 };
9eeff239 3726
9eeff239 3727 cond_resched();
d4f7796e 3728 /*
95bbc0c7 3729 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
d4f7796e 3730 * and we also need to be able to write out pages for RECLAIM_WRITE
95bbc0c7 3731 * and RECLAIM_UNMAP.
d4f7796e
CL
3732 */
3733 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
76ca542d 3734 lockdep_set_current_reclaim_state(gfp_mask);
9eeff239
CL
3735 reclaim_state.reclaimed_slab = 0;
3736 p->reclaim_state = &reclaim_state;
c84db23c 3737
a5f5f91d 3738 if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) {
0ff38490
CL
3739 /*
3740 * Free memory by calling shrink zone with increasing
3741 * priorities until we have enough memory freed.
3742 */
0ff38490 3743 do {
970a39a3 3744 shrink_node(pgdat, &sc);
9e3b2f8c 3745 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
0ff38490 3746 }
c84db23c 3747
9eeff239 3748 p->reclaim_state = NULL;
d4f7796e 3749 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
76ca542d 3750 lockdep_clear_current_reclaim_state();
a79311c1 3751 return sc.nr_reclaimed >= nr_pages;
9eeff239 3752}
179e9639 3753
a5f5f91d 3754int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order)
179e9639 3755{
d773ed6b 3756 int ret;
179e9639
AM
3757
3758 /*
a5f5f91d 3759 * Node reclaim reclaims unmapped file backed pages and
0ff38490 3760 * slab pages if we are over the defined limits.
34aa1330 3761 *
9614634f
CL
3762 * A small portion of unmapped file backed pages is needed for
3763 * file I/O otherwise pages read by file I/O will be immediately
a5f5f91d
MG
3764 * thrown out if the node is overallocated. So we do not reclaim
3765 * if less than a specified percentage of the node is used by
9614634f 3766 * unmapped file backed pages.
179e9639 3767 */
a5f5f91d
MG
3768 if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages &&
3769 sum_zone_node_page_state(pgdat->node_id, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages)
3770 return NODE_RECLAIM_FULL;
179e9639 3771
a5f5f91d
MG
3772 if (!pgdat_reclaimable(pgdat))
3773 return NODE_RECLAIM_FULL;
d773ed6b 3774
179e9639 3775 /*
d773ed6b 3776 * Do not scan if the allocation should not be delayed.
179e9639 3777 */
d0164adc 3778 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
a5f5f91d 3779 return NODE_RECLAIM_NOSCAN;
179e9639
AM
3780
3781 /*
a5f5f91d 3782 * Only run node reclaim on the local node or on nodes that do not
179e9639
AM
3783 * have associated processors. This will favor the local processor
3784 * over remote processors and spread off node memory allocations
3785 * as wide as possible.
3786 */
a5f5f91d
MG
3787 if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id())
3788 return NODE_RECLAIM_NOSCAN;
d773ed6b 3789
a5f5f91d
MG
3790 if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags))
3791 return NODE_RECLAIM_NOSCAN;
fa5e084e 3792
a5f5f91d
MG
3793 ret = __node_reclaim(pgdat, gfp_mask, order);
3794 clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags);
d773ed6b 3795
24cf7251
MG
3796 if (!ret)
3797 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3798
d773ed6b 3799 return ret;
179e9639 3800}
9eeff239 3801#endif
894bc310 3802
894bc310
LS
3803/*
3804 * page_evictable - test whether a page is evictable
3805 * @page: the page to test
894bc310
LS
3806 *
3807 * Test whether page is evictable--i.e., should be placed on active/inactive
39b5f29a 3808 * lists vs unevictable list.
894bc310
LS
3809 *
3810 * Reasons page might not be evictable:
ba9ddf49 3811 * (1) page's mapping marked unevictable
b291f000 3812 * (2) page is part of an mlocked VMA
ba9ddf49 3813 *
894bc310 3814 */
39b5f29a 3815int page_evictable(struct page *page)
894bc310 3816{
39b5f29a 3817 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
894bc310 3818}
89e004ea 3819
85046579 3820#ifdef CONFIG_SHMEM
89e004ea 3821/**
24513264
HD
3822 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3823 * @pages: array of pages to check
3824 * @nr_pages: number of pages to check
89e004ea 3825 *
24513264 3826 * Checks pages for evictability and moves them to the appropriate lru list.
85046579
HD
3827 *
3828 * This function is only used for SysV IPC SHM_UNLOCK.
89e004ea 3829 */
24513264 3830void check_move_unevictable_pages(struct page **pages, int nr_pages)
89e004ea 3831{
925b7673 3832 struct lruvec *lruvec;
785b99fe 3833 struct pglist_data *pgdat = NULL;
24513264
HD
3834 int pgscanned = 0;
3835 int pgrescued = 0;
3836 int i;
89e004ea 3837
24513264
HD
3838 for (i = 0; i < nr_pages; i++) {
3839 struct page *page = pages[i];
785b99fe 3840 struct pglist_data *pagepgdat = page_pgdat(page);
89e004ea 3841
24513264 3842 pgscanned++;
785b99fe
MG
3843 if (pagepgdat != pgdat) {
3844 if (pgdat)
3845 spin_unlock_irq(&pgdat->lru_lock);
3846 pgdat = pagepgdat;
3847 spin_lock_irq(&pgdat->lru_lock);
24513264 3848 }
785b99fe 3849 lruvec = mem_cgroup_page_lruvec(page, pgdat);
89e004ea 3850
24513264
HD
3851 if (!PageLRU(page) || !PageUnevictable(page))
3852 continue;
89e004ea 3853
39b5f29a 3854 if (page_evictable(page)) {
24513264
HD
3855 enum lru_list lru = page_lru_base_type(page);
3856
309381fe 3857 VM_BUG_ON_PAGE(PageActive(page), page);
24513264 3858 ClearPageUnevictable(page);
fa9add64
HD
3859 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3860 add_page_to_lru_list(page, lruvec, lru);
24513264 3861 pgrescued++;
89e004ea 3862 }
24513264 3863 }
89e004ea 3864
785b99fe 3865 if (pgdat) {
24513264
HD
3866 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3867 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
785b99fe 3868 spin_unlock_irq(&pgdat->lru_lock);
89e004ea 3869 }
89e004ea 3870}
85046579 3871#endif /* CONFIG_SHMEM */