]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - mm/vmscan.c
Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jikos/hid
[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>
1da177e4
LT
49
50#include <asm/tlbflush.h>
51#include <asm/div64.h>
52
53#include <linux/swapops.h>
117aad1e 54#include <linux/balloon_compaction.h>
1da177e4 55
0f8053a5
NP
56#include "internal.h"
57
33906bc5
MG
58#define CREATE_TRACE_POINTS
59#include <trace/events/vmscan.h>
60
1da177e4 61struct scan_control {
22fba335
KM
62 /* How many pages shrink_list() should reclaim */
63 unsigned long nr_to_reclaim;
64
1da177e4 65 /* This context's GFP mask */
6daa0e28 66 gfp_t gfp_mask;
1da177e4 67
ee814fe2 68 /* Allocation order */
5ad333eb 69 int order;
66e1707b 70
ee814fe2
JW
71 /*
72 * Nodemask of nodes allowed by the caller. If NULL, all nodes
73 * are scanned.
74 */
75 nodemask_t *nodemask;
9e3b2f8c 76
f16015fb
JW
77 /*
78 * The memory cgroup that hit its limit and as a result is the
79 * primary target of this reclaim invocation.
80 */
81 struct mem_cgroup *target_mem_cgroup;
66e1707b 82
ee814fe2
JW
83 /* Scan (total_size >> priority) pages at once */
84 int priority;
85
86 unsigned int may_writepage:1;
87
88 /* Can mapped pages be reclaimed? */
89 unsigned int may_unmap:1;
90
91 /* Can pages be swapped as part of reclaim? */
92 unsigned int may_swap:1;
93
241994ed
JW
94 /* Can cgroups be reclaimed below their normal consumption range? */
95 unsigned int may_thrash:1;
96
ee814fe2
JW
97 unsigned int hibernation_mode:1;
98
99 /* One of the zones is ready for compaction */
100 unsigned int compaction_ready:1;
101
102 /* Incremented by the number of inactive pages that were scanned */
103 unsigned long nr_scanned;
104
105 /* Number of pages freed so far during a call to shrink_zones() */
106 unsigned long nr_reclaimed;
1da177e4
LT
107};
108
1da177e4
LT
109#define lru_to_page(_head) (list_entry((_head)->prev, struct page, lru))
110
111#ifdef ARCH_HAS_PREFETCH
112#define prefetch_prev_lru_page(_page, _base, _field) \
113 do { \
114 if ((_page)->lru.prev != _base) { \
115 struct page *prev; \
116 \
117 prev = lru_to_page(&(_page->lru)); \
118 prefetch(&prev->_field); \
119 } \
120 } while (0)
121#else
122#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
123#endif
124
125#ifdef ARCH_HAS_PREFETCHW
126#define prefetchw_prev_lru_page(_page, _base, _field) \
127 do { \
128 if ((_page)->lru.prev != _base) { \
129 struct page *prev; \
130 \
131 prev = lru_to_page(&(_page->lru)); \
132 prefetchw(&prev->_field); \
133 } \
134 } while (0)
135#else
136#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
137#endif
138
139/*
140 * From 0 .. 100. Higher means more swappy.
141 */
142int vm_swappiness = 60;
d0480be4
WSH
143/*
144 * The total number of pages which are beyond the high watermark within all
145 * zones.
146 */
147unsigned long vm_total_pages;
1da177e4
LT
148
149static LIST_HEAD(shrinker_list);
150static DECLARE_RWSEM(shrinker_rwsem);
151
c255a458 152#ifdef CONFIG_MEMCG
89b5fae5
JW
153static bool global_reclaim(struct scan_control *sc)
154{
f16015fb 155 return !sc->target_mem_cgroup;
89b5fae5 156}
91a45470 157#else
89b5fae5
JW
158static bool global_reclaim(struct scan_control *sc)
159{
160 return true;
161}
91a45470
KH
162#endif
163
a1c3bfb2 164static unsigned long zone_reclaimable_pages(struct zone *zone)
6e543d57
LD
165{
166 int nr;
167
168 nr = zone_page_state(zone, NR_ACTIVE_FILE) +
169 zone_page_state(zone, NR_INACTIVE_FILE);
170
171 if (get_nr_swap_pages() > 0)
172 nr += zone_page_state(zone, NR_ACTIVE_ANON) +
173 zone_page_state(zone, NR_INACTIVE_ANON);
174
175 return nr;
176}
177
178bool zone_reclaimable(struct zone *zone)
179{
0d5d823a
MG
180 return zone_page_state(zone, NR_PAGES_SCANNED) <
181 zone_reclaimable_pages(zone) * 6;
6e543d57
LD
182}
183
4d7dcca2 184static unsigned long get_lru_size(struct lruvec *lruvec, enum lru_list lru)
c9f299d9 185{
c3c787e8 186 if (!mem_cgroup_disabled())
4d7dcca2 187 return mem_cgroup_get_lru_size(lruvec, lru);
a3d8e054 188
074291fe 189 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
c9f299d9
KM
190}
191
1da177e4 192/*
1d3d4437 193 * Add a shrinker callback to be called from the vm.
1da177e4 194 */
1d3d4437 195int register_shrinker(struct shrinker *shrinker)
1da177e4 196{
1d3d4437
GC
197 size_t size = sizeof(*shrinker->nr_deferred);
198
199 /*
200 * If we only have one possible node in the system anyway, save
201 * ourselves the trouble and disable NUMA aware behavior. This way we
202 * will save memory and some small loop time later.
203 */
204 if (nr_node_ids == 1)
205 shrinker->flags &= ~SHRINKER_NUMA_AWARE;
206
207 if (shrinker->flags & SHRINKER_NUMA_AWARE)
208 size *= nr_node_ids;
209
210 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
211 if (!shrinker->nr_deferred)
212 return -ENOMEM;
213
8e1f936b
RR
214 down_write(&shrinker_rwsem);
215 list_add_tail(&shrinker->list, &shrinker_list);
216 up_write(&shrinker_rwsem);
1d3d4437 217 return 0;
1da177e4 218}
8e1f936b 219EXPORT_SYMBOL(register_shrinker);
1da177e4
LT
220
221/*
222 * Remove one
223 */
8e1f936b 224void unregister_shrinker(struct shrinker *shrinker)
1da177e4
LT
225{
226 down_write(&shrinker_rwsem);
227 list_del(&shrinker->list);
228 up_write(&shrinker_rwsem);
ae393321 229 kfree(shrinker->nr_deferred);
1da177e4 230}
8e1f936b 231EXPORT_SYMBOL(unregister_shrinker);
1da177e4
LT
232
233#define SHRINK_BATCH 128
1d3d4437 234
cb731d6c
VD
235static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
236 struct shrinker *shrinker,
237 unsigned long nr_scanned,
238 unsigned long nr_eligible)
1d3d4437
GC
239{
240 unsigned long freed = 0;
241 unsigned long long delta;
242 long total_scan;
d5bc5fd3 243 long freeable;
1d3d4437
GC
244 long nr;
245 long new_nr;
246 int nid = shrinkctl->nid;
247 long batch_size = shrinker->batch ? shrinker->batch
248 : SHRINK_BATCH;
249
d5bc5fd3
VD
250 freeable = shrinker->count_objects(shrinker, shrinkctl);
251 if (freeable == 0)
1d3d4437
GC
252 return 0;
253
254 /*
255 * copy the current shrinker scan count into a local variable
256 * and zero it so that other concurrent shrinker invocations
257 * don't also do this scanning work.
258 */
259 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
260
261 total_scan = nr;
6b4f7799 262 delta = (4 * nr_scanned) / shrinker->seeks;
d5bc5fd3 263 delta *= freeable;
6b4f7799 264 do_div(delta, nr_eligible + 1);
1d3d4437
GC
265 total_scan += delta;
266 if (total_scan < 0) {
8612c663 267 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
a0b02131 268 shrinker->scan_objects, total_scan);
d5bc5fd3 269 total_scan = freeable;
1d3d4437
GC
270 }
271
272 /*
273 * We need to avoid excessive windup on filesystem shrinkers
274 * due to large numbers of GFP_NOFS allocations causing the
275 * shrinkers to return -1 all the time. This results in a large
276 * nr being built up so when a shrink that can do some work
277 * comes along it empties the entire cache due to nr >>>
d5bc5fd3 278 * freeable. This is bad for sustaining a working set in
1d3d4437
GC
279 * memory.
280 *
281 * Hence only allow the shrinker to scan the entire cache when
282 * a large delta change is calculated directly.
283 */
d5bc5fd3
VD
284 if (delta < freeable / 4)
285 total_scan = min(total_scan, freeable / 2);
1d3d4437
GC
286
287 /*
288 * Avoid risking looping forever due to too large nr value:
289 * never try to free more than twice the estimate number of
290 * freeable entries.
291 */
d5bc5fd3
VD
292 if (total_scan > freeable * 2)
293 total_scan = freeable * 2;
1d3d4437
GC
294
295 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
6b4f7799
JW
296 nr_scanned, nr_eligible,
297 freeable, delta, total_scan);
1d3d4437 298
0b1fb40a
VD
299 /*
300 * Normally, we should not scan less than batch_size objects in one
301 * pass to avoid too frequent shrinker calls, but if the slab has less
302 * than batch_size objects in total and we are really tight on memory,
303 * we will try to reclaim all available objects, otherwise we can end
304 * up failing allocations although there are plenty of reclaimable
305 * objects spread over several slabs with usage less than the
306 * batch_size.
307 *
308 * We detect the "tight on memory" situations by looking at the total
309 * number of objects we want to scan (total_scan). If it is greater
d5bc5fd3 310 * than the total number of objects on slab (freeable), we must be
0b1fb40a
VD
311 * scanning at high prio and therefore should try to reclaim as much as
312 * possible.
313 */
314 while (total_scan >= batch_size ||
d5bc5fd3 315 total_scan >= freeable) {
a0b02131 316 unsigned long ret;
0b1fb40a 317 unsigned long nr_to_scan = min(batch_size, total_scan);
1d3d4437 318
0b1fb40a 319 shrinkctl->nr_to_scan = nr_to_scan;
a0b02131
DC
320 ret = shrinker->scan_objects(shrinker, shrinkctl);
321 if (ret == SHRINK_STOP)
322 break;
323 freed += ret;
1d3d4437 324
0b1fb40a
VD
325 count_vm_events(SLABS_SCANNED, nr_to_scan);
326 total_scan -= nr_to_scan;
1d3d4437
GC
327
328 cond_resched();
329 }
330
331 /*
332 * move the unused scan count back into the shrinker in a
333 * manner that handles concurrent updates. If we exhausted the
334 * scan, there is no need to do an update.
335 */
336 if (total_scan > 0)
337 new_nr = atomic_long_add_return(total_scan,
338 &shrinker->nr_deferred[nid]);
339 else
340 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
341
df9024a8 342 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
1d3d4437 343 return freed;
1495f230
YH
344}
345
6b4f7799 346/**
cb731d6c 347 * shrink_slab - shrink slab caches
6b4f7799
JW
348 * @gfp_mask: allocation context
349 * @nid: node whose slab caches to target
cb731d6c 350 * @memcg: memory cgroup whose slab caches to target
6b4f7799
JW
351 * @nr_scanned: pressure numerator
352 * @nr_eligible: pressure denominator
1da177e4 353 *
6b4f7799 354 * Call the shrink functions to age shrinkable caches.
1da177e4 355 *
6b4f7799
JW
356 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
357 * unaware shrinkers will receive a node id of 0 instead.
1da177e4 358 *
cb731d6c
VD
359 * @memcg specifies the memory cgroup to target. If it is not NULL,
360 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
361 * objects from the memory cgroup specified. Otherwise all shrinkers
362 * are called, and memcg aware shrinkers are supposed to scan the
363 * global list then.
364 *
6b4f7799
JW
365 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
366 * the available objects should be scanned. Page reclaim for example
367 * passes the number of pages scanned and the number of pages on the
368 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
369 * when it encountered mapped pages. The ratio is further biased by
370 * the ->seeks setting of the shrink function, which indicates the
371 * cost to recreate an object relative to that of an LRU page.
b15e0905 372 *
6b4f7799 373 * Returns the number of reclaimed slab objects.
1da177e4 374 */
cb731d6c
VD
375static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
376 struct mem_cgroup *memcg,
377 unsigned long nr_scanned,
378 unsigned long nr_eligible)
1da177e4
LT
379{
380 struct shrinker *shrinker;
24f7c6b9 381 unsigned long freed = 0;
1da177e4 382
cb731d6c
VD
383 if (memcg && !memcg_kmem_is_active(memcg))
384 return 0;
385
6b4f7799
JW
386 if (nr_scanned == 0)
387 nr_scanned = SWAP_CLUSTER_MAX;
1da177e4 388
f06590bd 389 if (!down_read_trylock(&shrinker_rwsem)) {
24f7c6b9
DC
390 /*
391 * If we would return 0, our callers would understand that we
392 * have nothing else to shrink and give up trying. By returning
393 * 1 we keep it going and assume we'll be able to shrink next
394 * time.
395 */
396 freed = 1;
f06590bd
MK
397 goto out;
398 }
1da177e4
LT
399
400 list_for_each_entry(shrinker, &shrinker_list, list) {
6b4f7799
JW
401 struct shrink_control sc = {
402 .gfp_mask = gfp_mask,
403 .nid = nid,
cb731d6c 404 .memcg = memcg,
6b4f7799 405 };
ec97097b 406
cb731d6c
VD
407 if (memcg && !(shrinker->flags & SHRINKER_MEMCG_AWARE))
408 continue;
409
6b4f7799
JW
410 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
411 sc.nid = 0;
1da177e4 412
cb731d6c 413 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
1da177e4 414 }
6b4f7799 415
1da177e4 416 up_read(&shrinker_rwsem);
f06590bd
MK
417out:
418 cond_resched();
24f7c6b9 419 return freed;
1da177e4
LT
420}
421
cb731d6c
VD
422void drop_slab_node(int nid)
423{
424 unsigned long freed;
425
426 do {
427 struct mem_cgroup *memcg = NULL;
428
429 freed = 0;
430 do {
431 freed += shrink_slab(GFP_KERNEL, nid, memcg,
432 1000, 1000);
433 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
434 } while (freed > 10);
435}
436
437void drop_slab(void)
438{
439 int nid;
440
441 for_each_online_node(nid)
442 drop_slab_node(nid);
443}
444
1da177e4
LT
445static inline int is_page_cache_freeable(struct page *page)
446{
ceddc3a5
JW
447 /*
448 * A freeable page cache page is referenced only by the caller
449 * that isolated the page, the page cache radix tree and
450 * optional buffer heads at page->private.
451 */
edcf4748 452 return page_count(page) - page_has_private(page) == 2;
1da177e4
LT
453}
454
7d3579e8
KM
455static int may_write_to_queue(struct backing_dev_info *bdi,
456 struct scan_control *sc)
1da177e4 457{
930d9152 458 if (current->flags & PF_SWAPWRITE)
1da177e4
LT
459 return 1;
460 if (!bdi_write_congested(bdi))
461 return 1;
462 if (bdi == current->backing_dev_info)
463 return 1;
464 return 0;
465}
466
467/*
468 * We detected a synchronous write error writing a page out. Probably
469 * -ENOSPC. We need to propagate that into the address_space for a subsequent
470 * fsync(), msync() or close().
471 *
472 * The tricky part is that after writepage we cannot touch the mapping: nothing
473 * prevents it from being freed up. But we have a ref on the page and once
474 * that page is locked, the mapping is pinned.
475 *
476 * We're allowed to run sleeping lock_page() here because we know the caller has
477 * __GFP_FS.
478 */
479static void handle_write_error(struct address_space *mapping,
480 struct page *page, int error)
481{
7eaceacc 482 lock_page(page);
3e9f45bd
GC
483 if (page_mapping(page) == mapping)
484 mapping_set_error(mapping, error);
1da177e4
LT
485 unlock_page(page);
486}
487
04e62a29
CL
488/* possible outcome of pageout() */
489typedef enum {
490 /* failed to write page out, page is locked */
491 PAGE_KEEP,
492 /* move page to the active list, page is locked */
493 PAGE_ACTIVATE,
494 /* page has been sent to the disk successfully, page is unlocked */
495 PAGE_SUCCESS,
496 /* page is clean and locked */
497 PAGE_CLEAN,
498} pageout_t;
499
1da177e4 500/*
1742f19f
AM
501 * pageout is called by shrink_page_list() for each dirty page.
502 * Calls ->writepage().
1da177e4 503 */
c661b078 504static pageout_t pageout(struct page *page, struct address_space *mapping,
7d3579e8 505 struct scan_control *sc)
1da177e4
LT
506{
507 /*
508 * If the page is dirty, only perform writeback if that write
509 * will be non-blocking. To prevent this allocation from being
510 * stalled by pagecache activity. But note that there may be
511 * stalls if we need to run get_block(). We could test
512 * PagePrivate for that.
513 *
8174202b 514 * If this process is currently in __generic_file_write_iter() against
1da177e4
LT
515 * this page's queue, we can perform writeback even if that
516 * will block.
517 *
518 * If the page is swapcache, write it back even if that would
519 * block, for some throttling. This happens by accident, because
520 * swap_backing_dev_info is bust: it doesn't reflect the
521 * congestion state of the swapdevs. Easy to fix, if needed.
1da177e4
LT
522 */
523 if (!is_page_cache_freeable(page))
524 return PAGE_KEEP;
525 if (!mapping) {
526 /*
527 * Some data journaling orphaned pages can have
528 * page->mapping == NULL while being dirty with clean buffers.
529 */
266cf658 530 if (page_has_private(page)) {
1da177e4
LT
531 if (try_to_free_buffers(page)) {
532 ClearPageDirty(page);
b1de0d13 533 pr_info("%s: orphaned page\n", __func__);
1da177e4
LT
534 return PAGE_CLEAN;
535 }
536 }
537 return PAGE_KEEP;
538 }
539 if (mapping->a_ops->writepage == NULL)
540 return PAGE_ACTIVATE;
de1414a6 541 if (!may_write_to_queue(inode_to_bdi(mapping->host), sc))
1da177e4
LT
542 return PAGE_KEEP;
543
544 if (clear_page_dirty_for_io(page)) {
545 int res;
546 struct writeback_control wbc = {
547 .sync_mode = WB_SYNC_NONE,
548 .nr_to_write = SWAP_CLUSTER_MAX,
111ebb6e
OH
549 .range_start = 0,
550 .range_end = LLONG_MAX,
1da177e4
LT
551 .for_reclaim = 1,
552 };
553
554 SetPageReclaim(page);
555 res = mapping->a_ops->writepage(page, &wbc);
556 if (res < 0)
557 handle_write_error(mapping, page, res);
994fc28c 558 if (res == AOP_WRITEPAGE_ACTIVATE) {
1da177e4
LT
559 ClearPageReclaim(page);
560 return PAGE_ACTIVATE;
561 }
c661b078 562
1da177e4
LT
563 if (!PageWriteback(page)) {
564 /* synchronous write or broken a_ops? */
565 ClearPageReclaim(page);
566 }
23b9da55 567 trace_mm_vmscan_writepage(page, trace_reclaim_flags(page));
e129b5c2 568 inc_zone_page_state(page, NR_VMSCAN_WRITE);
1da177e4
LT
569 return PAGE_SUCCESS;
570 }
571
572 return PAGE_CLEAN;
573}
574
a649fd92 575/*
e286781d
NP
576 * Same as remove_mapping, but if the page is removed from the mapping, it
577 * gets returned with a refcount of 0.
a649fd92 578 */
a528910e
JW
579static int __remove_mapping(struct address_space *mapping, struct page *page,
580 bool reclaimed)
49d2e9cc 581{
28e4d965
NP
582 BUG_ON(!PageLocked(page));
583 BUG_ON(mapping != page_mapping(page));
49d2e9cc 584
19fd6231 585 spin_lock_irq(&mapping->tree_lock);
49d2e9cc 586 /*
0fd0e6b0
NP
587 * The non racy check for a busy page.
588 *
589 * Must be careful with the order of the tests. When someone has
590 * a ref to the page, it may be possible that they dirty it then
591 * drop the reference. So if PageDirty is tested before page_count
592 * here, then the following race may occur:
593 *
594 * get_user_pages(&page);
595 * [user mapping goes away]
596 * write_to(page);
597 * !PageDirty(page) [good]
598 * SetPageDirty(page);
599 * put_page(page);
600 * !page_count(page) [good, discard it]
601 *
602 * [oops, our write_to data is lost]
603 *
604 * Reversing the order of the tests ensures such a situation cannot
605 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
606 * load is not satisfied before that of page->_count.
607 *
608 * Note that if SetPageDirty is always performed via set_page_dirty,
609 * and thus under tree_lock, then this ordering is not required.
49d2e9cc 610 */
e286781d 611 if (!page_freeze_refs(page, 2))
49d2e9cc 612 goto cannot_free;
e286781d
NP
613 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
614 if (unlikely(PageDirty(page))) {
615 page_unfreeze_refs(page, 2);
49d2e9cc 616 goto cannot_free;
e286781d 617 }
49d2e9cc
CL
618
619 if (PageSwapCache(page)) {
620 swp_entry_t swap = { .val = page_private(page) };
0a31bc97 621 mem_cgroup_swapout(page, swap);
49d2e9cc 622 __delete_from_swap_cache(page);
19fd6231 623 spin_unlock_irq(&mapping->tree_lock);
0a31bc97 624 swapcache_free(swap);
e286781d 625 } else {
6072d13c 626 void (*freepage)(struct page *);
a528910e 627 void *shadow = NULL;
6072d13c
LT
628
629 freepage = mapping->a_ops->freepage;
a528910e
JW
630 /*
631 * Remember a shadow entry for reclaimed file cache in
632 * order to detect refaults, thus thrashing, later on.
633 *
634 * But don't store shadows in an address space that is
635 * already exiting. This is not just an optizimation,
636 * inode reclaim needs to empty out the radix tree or
637 * the nodes are lost. Don't plant shadows behind its
638 * back.
639 */
640 if (reclaimed && page_is_file_cache(page) &&
641 !mapping_exiting(mapping))
642 shadow = workingset_eviction(mapping, page);
643 __delete_from_page_cache(page, shadow);
19fd6231 644 spin_unlock_irq(&mapping->tree_lock);
6072d13c
LT
645
646 if (freepage != NULL)
647 freepage(page);
49d2e9cc
CL
648 }
649
49d2e9cc
CL
650 return 1;
651
652cannot_free:
19fd6231 653 spin_unlock_irq(&mapping->tree_lock);
49d2e9cc
CL
654 return 0;
655}
656
e286781d
NP
657/*
658 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
659 * someone else has a ref on the page, abort and return 0. If it was
660 * successfully detached, return 1. Assumes the caller has a single ref on
661 * this page.
662 */
663int remove_mapping(struct address_space *mapping, struct page *page)
664{
a528910e 665 if (__remove_mapping(mapping, page, false)) {
e286781d
NP
666 /*
667 * Unfreezing the refcount with 1 rather than 2 effectively
668 * drops the pagecache ref for us without requiring another
669 * atomic operation.
670 */
671 page_unfreeze_refs(page, 1);
672 return 1;
673 }
674 return 0;
675}
676
894bc310
LS
677/**
678 * putback_lru_page - put previously isolated page onto appropriate LRU list
679 * @page: page to be put back to appropriate lru list
680 *
681 * Add previously isolated @page to appropriate LRU list.
682 * Page may still be unevictable for other reasons.
683 *
684 * lru_lock must not be held, interrupts must be enabled.
685 */
894bc310
LS
686void putback_lru_page(struct page *page)
687{
0ec3b74c 688 bool is_unevictable;
bbfd28ee 689 int was_unevictable = PageUnevictable(page);
894bc310 690
309381fe 691 VM_BUG_ON_PAGE(PageLRU(page), page);
894bc310
LS
692
693redo:
694 ClearPageUnevictable(page);
695
39b5f29a 696 if (page_evictable(page)) {
894bc310
LS
697 /*
698 * For evictable pages, we can use the cache.
699 * In event of a race, worst case is we end up with an
700 * unevictable page on [in]active list.
701 * We know how to handle that.
702 */
0ec3b74c 703 is_unevictable = false;
c53954a0 704 lru_cache_add(page);
894bc310
LS
705 } else {
706 /*
707 * Put unevictable pages directly on zone's unevictable
708 * list.
709 */
0ec3b74c 710 is_unevictable = true;
894bc310 711 add_page_to_unevictable_list(page);
6a7b9548 712 /*
21ee9f39
MK
713 * When racing with an mlock or AS_UNEVICTABLE clearing
714 * (page is unlocked) make sure that if the other thread
715 * does not observe our setting of PG_lru and fails
24513264 716 * isolation/check_move_unevictable_pages,
21ee9f39 717 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
6a7b9548
JW
718 * the page back to the evictable list.
719 *
21ee9f39 720 * The other side is TestClearPageMlocked() or shmem_lock().
6a7b9548
JW
721 */
722 smp_mb();
894bc310 723 }
894bc310
LS
724
725 /*
726 * page's status can change while we move it among lru. If an evictable
727 * page is on unevictable list, it never be freed. To avoid that,
728 * check after we added it to the list, again.
729 */
0ec3b74c 730 if (is_unevictable && page_evictable(page)) {
894bc310
LS
731 if (!isolate_lru_page(page)) {
732 put_page(page);
733 goto redo;
734 }
735 /* This means someone else dropped this page from LRU
736 * So, it will be freed or putback to LRU again. There is
737 * nothing to do here.
738 */
739 }
740
0ec3b74c 741 if (was_unevictable && !is_unevictable)
bbfd28ee 742 count_vm_event(UNEVICTABLE_PGRESCUED);
0ec3b74c 743 else if (!was_unevictable && is_unevictable)
bbfd28ee
LS
744 count_vm_event(UNEVICTABLE_PGCULLED);
745
894bc310
LS
746 put_page(page); /* drop ref from isolate */
747}
748
dfc8d636
JW
749enum page_references {
750 PAGEREF_RECLAIM,
751 PAGEREF_RECLAIM_CLEAN,
64574746 752 PAGEREF_KEEP,
dfc8d636
JW
753 PAGEREF_ACTIVATE,
754};
755
756static enum page_references page_check_references(struct page *page,
757 struct scan_control *sc)
758{
64574746 759 int referenced_ptes, referenced_page;
dfc8d636 760 unsigned long vm_flags;
dfc8d636 761
c3ac9a8a
JW
762 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
763 &vm_flags);
64574746 764 referenced_page = TestClearPageReferenced(page);
dfc8d636 765
dfc8d636
JW
766 /*
767 * Mlock lost the isolation race with us. Let try_to_unmap()
768 * move the page to the unevictable list.
769 */
770 if (vm_flags & VM_LOCKED)
771 return PAGEREF_RECLAIM;
772
64574746 773 if (referenced_ptes) {
e4898273 774 if (PageSwapBacked(page))
64574746
JW
775 return PAGEREF_ACTIVATE;
776 /*
777 * All mapped pages start out with page table
778 * references from the instantiating fault, so we need
779 * to look twice if a mapped file page is used more
780 * than once.
781 *
782 * Mark it and spare it for another trip around the
783 * inactive list. Another page table reference will
784 * lead to its activation.
785 *
786 * Note: the mark is set for activated pages as well
787 * so that recently deactivated but used pages are
788 * quickly recovered.
789 */
790 SetPageReferenced(page);
791
34dbc67a 792 if (referenced_page || referenced_ptes > 1)
64574746
JW
793 return PAGEREF_ACTIVATE;
794
c909e993
KK
795 /*
796 * Activate file-backed executable pages after first usage.
797 */
798 if (vm_flags & VM_EXEC)
799 return PAGEREF_ACTIVATE;
800
64574746
JW
801 return PAGEREF_KEEP;
802 }
dfc8d636
JW
803
804 /* Reclaim if clean, defer dirty pages to writeback */
2e30244a 805 if (referenced_page && !PageSwapBacked(page))
64574746
JW
806 return PAGEREF_RECLAIM_CLEAN;
807
808 return PAGEREF_RECLAIM;
dfc8d636
JW
809}
810
e2be15f6
MG
811/* Check if a page is dirty or under writeback */
812static void page_check_dirty_writeback(struct page *page,
813 bool *dirty, bool *writeback)
814{
b4597226
MG
815 struct address_space *mapping;
816
e2be15f6
MG
817 /*
818 * Anonymous pages are not handled by flushers and must be written
819 * from reclaim context. Do not stall reclaim based on them
820 */
821 if (!page_is_file_cache(page)) {
822 *dirty = false;
823 *writeback = false;
824 return;
825 }
826
827 /* By default assume that the page flags are accurate */
828 *dirty = PageDirty(page);
829 *writeback = PageWriteback(page);
b4597226
MG
830
831 /* Verify dirty/writeback state if the filesystem supports it */
832 if (!page_has_private(page))
833 return;
834
835 mapping = page_mapping(page);
836 if (mapping && mapping->a_ops->is_dirty_writeback)
837 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
e2be15f6
MG
838}
839
1da177e4 840/*
1742f19f 841 * shrink_page_list() returns the number of reclaimed pages
1da177e4 842 */
1742f19f 843static unsigned long shrink_page_list(struct list_head *page_list,
6a18adb3 844 struct zone *zone,
f84f6e2b 845 struct scan_control *sc,
02c6de8d 846 enum ttu_flags ttu_flags,
8e950282 847 unsigned long *ret_nr_dirty,
d43006d5 848 unsigned long *ret_nr_unqueued_dirty,
8e950282 849 unsigned long *ret_nr_congested,
02c6de8d 850 unsigned long *ret_nr_writeback,
b1a6f21e 851 unsigned long *ret_nr_immediate,
02c6de8d 852 bool force_reclaim)
1da177e4
LT
853{
854 LIST_HEAD(ret_pages);
abe4c3b5 855 LIST_HEAD(free_pages);
1da177e4 856 int pgactivate = 0;
d43006d5 857 unsigned long nr_unqueued_dirty = 0;
0e093d99
MG
858 unsigned long nr_dirty = 0;
859 unsigned long nr_congested = 0;
05ff5137 860 unsigned long nr_reclaimed = 0;
92df3a72 861 unsigned long nr_writeback = 0;
b1a6f21e 862 unsigned long nr_immediate = 0;
1da177e4
LT
863
864 cond_resched();
865
1da177e4
LT
866 while (!list_empty(page_list)) {
867 struct address_space *mapping;
868 struct page *page;
869 int may_enter_fs;
02c6de8d 870 enum page_references references = PAGEREF_RECLAIM_CLEAN;
e2be15f6 871 bool dirty, writeback;
1da177e4
LT
872
873 cond_resched();
874
875 page = lru_to_page(page_list);
876 list_del(&page->lru);
877
529ae9aa 878 if (!trylock_page(page))
1da177e4
LT
879 goto keep;
880
309381fe
SL
881 VM_BUG_ON_PAGE(PageActive(page), page);
882 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1da177e4
LT
883
884 sc->nr_scanned++;
80e43426 885
39b5f29a 886 if (unlikely(!page_evictable(page)))
b291f000 887 goto cull_mlocked;
894bc310 888
a6dc60f8 889 if (!sc->may_unmap && page_mapped(page))
80e43426
CL
890 goto keep_locked;
891
1da177e4
LT
892 /* Double the slab pressure for mapped and swapcache pages */
893 if (page_mapped(page) || PageSwapCache(page))
894 sc->nr_scanned++;
895
c661b078
AW
896 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
897 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
898
e2be15f6
MG
899 /*
900 * The number of dirty pages determines if a zone is marked
901 * reclaim_congested which affects wait_iff_congested. kswapd
902 * will stall and start writing pages if the tail of the LRU
903 * is all dirty unqueued pages.
904 */
905 page_check_dirty_writeback(page, &dirty, &writeback);
906 if (dirty || writeback)
907 nr_dirty++;
908
909 if (dirty && !writeback)
910 nr_unqueued_dirty++;
911
d04e8acd
MG
912 /*
913 * Treat this page as congested if the underlying BDI is or if
914 * pages are cycling through the LRU so quickly that the
915 * pages marked for immediate reclaim are making it to the
916 * end of the LRU a second time.
917 */
e2be15f6 918 mapping = page_mapping(page);
1da58ee2 919 if (((dirty || writeback) && mapping &&
de1414a6 920 bdi_write_congested(inode_to_bdi(mapping->host))) ||
d04e8acd 921 (writeback && PageReclaim(page)))
e2be15f6
MG
922 nr_congested++;
923
283aba9f
MG
924 /*
925 * If a page at the tail of the LRU is under writeback, there
926 * are three cases to consider.
927 *
928 * 1) If reclaim is encountering an excessive number of pages
929 * under writeback and this page is both under writeback and
930 * PageReclaim then it indicates that pages are being queued
931 * for IO but are being recycled through the LRU before the
932 * IO can complete. Waiting on the page itself risks an
933 * indefinite stall if it is impossible to writeback the
934 * page due to IO error or disconnected storage so instead
b1a6f21e
MG
935 * note that the LRU is being scanned too quickly and the
936 * caller can stall after page list has been processed.
283aba9f
MG
937 *
938 * 2) Global reclaim encounters a page, memcg encounters a
939 * page that is not marked for immediate reclaim or
940 * the caller does not have __GFP_IO. In this case mark
941 * the page for immediate reclaim and continue scanning.
942 *
943 * __GFP_IO is checked because a loop driver thread might
944 * enter reclaim, and deadlock if it waits on a page for
945 * which it is needed to do the write (loop masks off
946 * __GFP_IO|__GFP_FS for this reason); but more thought
947 * would probably show more reasons.
948 *
949 * Don't require __GFP_FS, since we're not going into the
950 * FS, just waiting on its writeback completion. Worryingly,
951 * ext4 gfs2 and xfs allocate pages with
952 * grab_cache_page_write_begin(,,AOP_FLAG_NOFS), so testing
953 * may_enter_fs here is liable to OOM on them.
954 *
955 * 3) memcg encounters a page that is not already marked
956 * PageReclaim. memcg does not have any dirty pages
957 * throttling so we could easily OOM just because too many
958 * pages are in writeback and there is nothing else to
959 * reclaim. Wait for the writeback to complete.
960 */
c661b078 961 if (PageWriteback(page)) {
283aba9f
MG
962 /* Case 1 above */
963 if (current_is_kswapd() &&
964 PageReclaim(page) &&
57054651 965 test_bit(ZONE_WRITEBACK, &zone->flags)) {
b1a6f21e
MG
966 nr_immediate++;
967 goto keep_locked;
283aba9f
MG
968
969 /* Case 2 above */
970 } else if (global_reclaim(sc) ||
c3b94f44
HD
971 !PageReclaim(page) || !(sc->gfp_mask & __GFP_IO)) {
972 /*
973 * This is slightly racy - end_page_writeback()
974 * might have just cleared PageReclaim, then
975 * setting PageReclaim here end up interpreted
976 * as PageReadahead - but that does not matter
977 * enough to care. What we do want is for this
978 * page to have PageReclaim set next time memcg
979 * reclaim reaches the tests above, so it will
980 * then wait_on_page_writeback() to avoid OOM;
981 * and it's also appropriate in global reclaim.
982 */
983 SetPageReclaim(page);
e62e384e 984 nr_writeback++;
283aba9f 985
c3b94f44 986 goto keep_locked;
283aba9f
MG
987
988 /* Case 3 above */
989 } else {
990 wait_on_page_writeback(page);
e62e384e 991 }
c661b078 992 }
1da177e4 993
02c6de8d
MK
994 if (!force_reclaim)
995 references = page_check_references(page, sc);
996
dfc8d636
JW
997 switch (references) {
998 case PAGEREF_ACTIVATE:
1da177e4 999 goto activate_locked;
64574746
JW
1000 case PAGEREF_KEEP:
1001 goto keep_locked;
dfc8d636
JW
1002 case PAGEREF_RECLAIM:
1003 case PAGEREF_RECLAIM_CLEAN:
1004 ; /* try to reclaim the page below */
1005 }
1da177e4 1006
1da177e4
LT
1007 /*
1008 * Anonymous process memory has backing store?
1009 * Try to allocate it some swap space here.
1010 */
b291f000 1011 if (PageAnon(page) && !PageSwapCache(page)) {
63eb6b93
HD
1012 if (!(sc->gfp_mask & __GFP_IO))
1013 goto keep_locked;
5bc7b8ac 1014 if (!add_to_swap(page, page_list))
1da177e4 1015 goto activate_locked;
63eb6b93 1016 may_enter_fs = 1;
1da177e4 1017
e2be15f6
MG
1018 /* Adding to swap updated mapping */
1019 mapping = page_mapping(page);
1020 }
1da177e4
LT
1021
1022 /*
1023 * The page is mapped into the page tables of one or more
1024 * processes. Try to unmap it here.
1025 */
1026 if (page_mapped(page) && mapping) {
02c6de8d 1027 switch (try_to_unmap(page, ttu_flags)) {
1da177e4
LT
1028 case SWAP_FAIL:
1029 goto activate_locked;
1030 case SWAP_AGAIN:
1031 goto keep_locked;
b291f000
NP
1032 case SWAP_MLOCK:
1033 goto cull_mlocked;
1da177e4
LT
1034 case SWAP_SUCCESS:
1035 ; /* try to free the page below */
1036 }
1037 }
1038
1039 if (PageDirty(page)) {
ee72886d
MG
1040 /*
1041 * Only kswapd can writeback filesystem pages to
d43006d5
MG
1042 * avoid risk of stack overflow but only writeback
1043 * if many dirty pages have been encountered.
ee72886d 1044 */
f84f6e2b 1045 if (page_is_file_cache(page) &&
9e3b2f8c 1046 (!current_is_kswapd() ||
57054651 1047 !test_bit(ZONE_DIRTY, &zone->flags))) {
49ea7eb6
MG
1048 /*
1049 * Immediately reclaim when written back.
1050 * Similar in principal to deactivate_page()
1051 * except we already have the page isolated
1052 * and know it's dirty
1053 */
1054 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1055 SetPageReclaim(page);
1056
ee72886d
MG
1057 goto keep_locked;
1058 }
1059
dfc8d636 1060 if (references == PAGEREF_RECLAIM_CLEAN)
1da177e4 1061 goto keep_locked;
4dd4b920 1062 if (!may_enter_fs)
1da177e4 1063 goto keep_locked;
52a8363e 1064 if (!sc->may_writepage)
1da177e4
LT
1065 goto keep_locked;
1066
1067 /* Page is dirty, try to write it out here */
7d3579e8 1068 switch (pageout(page, mapping, sc)) {
1da177e4
LT
1069 case PAGE_KEEP:
1070 goto keep_locked;
1071 case PAGE_ACTIVATE:
1072 goto activate_locked;
1073 case PAGE_SUCCESS:
7d3579e8 1074 if (PageWriteback(page))
41ac1999 1075 goto keep;
7d3579e8 1076 if (PageDirty(page))
1da177e4 1077 goto keep;
7d3579e8 1078
1da177e4
LT
1079 /*
1080 * A synchronous write - probably a ramdisk. Go
1081 * ahead and try to reclaim the page.
1082 */
529ae9aa 1083 if (!trylock_page(page))
1da177e4
LT
1084 goto keep;
1085 if (PageDirty(page) || PageWriteback(page))
1086 goto keep_locked;
1087 mapping = page_mapping(page);
1088 case PAGE_CLEAN:
1089 ; /* try to free the page below */
1090 }
1091 }
1092
1093 /*
1094 * If the page has buffers, try to free the buffer mappings
1095 * associated with this page. If we succeed we try to free
1096 * the page as well.
1097 *
1098 * We do this even if the page is PageDirty().
1099 * try_to_release_page() does not perform I/O, but it is
1100 * possible for a page to have PageDirty set, but it is actually
1101 * clean (all its buffers are clean). This happens if the
1102 * buffers were written out directly, with submit_bh(). ext3
894bc310 1103 * will do this, as well as the blockdev mapping.
1da177e4
LT
1104 * try_to_release_page() will discover that cleanness and will
1105 * drop the buffers and mark the page clean - it can be freed.
1106 *
1107 * Rarely, pages can have buffers and no ->mapping. These are
1108 * the pages which were not successfully invalidated in
1109 * truncate_complete_page(). We try to drop those buffers here
1110 * and if that worked, and the page is no longer mapped into
1111 * process address space (page_count == 1) it can be freed.
1112 * Otherwise, leave the page on the LRU so it is swappable.
1113 */
266cf658 1114 if (page_has_private(page)) {
1da177e4
LT
1115 if (!try_to_release_page(page, sc->gfp_mask))
1116 goto activate_locked;
e286781d
NP
1117 if (!mapping && page_count(page) == 1) {
1118 unlock_page(page);
1119 if (put_page_testzero(page))
1120 goto free_it;
1121 else {
1122 /*
1123 * rare race with speculative reference.
1124 * the speculative reference will free
1125 * this page shortly, so we may
1126 * increment nr_reclaimed here (and
1127 * leave it off the LRU).
1128 */
1129 nr_reclaimed++;
1130 continue;
1131 }
1132 }
1da177e4
LT
1133 }
1134
a528910e 1135 if (!mapping || !__remove_mapping(mapping, page, true))
49d2e9cc 1136 goto keep_locked;
1da177e4 1137
a978d6f5
NP
1138 /*
1139 * At this point, we have no other references and there is
1140 * no way to pick any more up (removed from LRU, removed
1141 * from pagecache). Can use non-atomic bitops now (and
1142 * we obviously don't have to worry about waking up a process
1143 * waiting on the page lock, because there are no references.
1144 */
1145 __clear_page_locked(page);
e286781d 1146free_it:
05ff5137 1147 nr_reclaimed++;
abe4c3b5
MG
1148
1149 /*
1150 * Is there need to periodically free_page_list? It would
1151 * appear not as the counts should be low
1152 */
1153 list_add(&page->lru, &free_pages);
1da177e4
LT
1154 continue;
1155
b291f000 1156cull_mlocked:
63d6c5ad
HD
1157 if (PageSwapCache(page))
1158 try_to_free_swap(page);
b291f000
NP
1159 unlock_page(page);
1160 putback_lru_page(page);
1161 continue;
1162
1da177e4 1163activate_locked:
68a22394
RR
1164 /* Not a candidate for swapping, so reclaim swap space. */
1165 if (PageSwapCache(page) && vm_swap_full())
a2c43eed 1166 try_to_free_swap(page);
309381fe 1167 VM_BUG_ON_PAGE(PageActive(page), page);
1da177e4
LT
1168 SetPageActive(page);
1169 pgactivate++;
1170keep_locked:
1171 unlock_page(page);
1172keep:
1173 list_add(&page->lru, &ret_pages);
309381fe 1174 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1da177e4 1175 }
abe4c3b5 1176
747db954 1177 mem_cgroup_uncharge_list(&free_pages);
b745bc85 1178 free_hot_cold_page_list(&free_pages, true);
abe4c3b5 1179
1da177e4 1180 list_splice(&ret_pages, page_list);
f8891e5e 1181 count_vm_events(PGACTIVATE, pgactivate);
0a31bc97 1182
8e950282
MG
1183 *ret_nr_dirty += nr_dirty;
1184 *ret_nr_congested += nr_congested;
d43006d5 1185 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
92df3a72 1186 *ret_nr_writeback += nr_writeback;
b1a6f21e 1187 *ret_nr_immediate += nr_immediate;
05ff5137 1188 return nr_reclaimed;
1da177e4
LT
1189}
1190
02c6de8d
MK
1191unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1192 struct list_head *page_list)
1193{
1194 struct scan_control sc = {
1195 .gfp_mask = GFP_KERNEL,
1196 .priority = DEF_PRIORITY,
1197 .may_unmap = 1,
1198 };
8e950282 1199 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
02c6de8d
MK
1200 struct page *page, *next;
1201 LIST_HEAD(clean_pages);
1202
1203 list_for_each_entry_safe(page, next, page_list, lru) {
117aad1e
RA
1204 if (page_is_file_cache(page) && !PageDirty(page) &&
1205 !isolated_balloon_page(page)) {
02c6de8d
MK
1206 ClearPageActive(page);
1207 list_move(&page->lru, &clean_pages);
1208 }
1209 }
1210
1211 ret = shrink_page_list(&clean_pages, zone, &sc,
8e950282
MG
1212 TTU_UNMAP|TTU_IGNORE_ACCESS,
1213 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
02c6de8d 1214 list_splice(&clean_pages, page_list);
83da7510 1215 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
02c6de8d
MK
1216 return ret;
1217}
1218
5ad333eb
AW
1219/*
1220 * Attempt to remove the specified page from its LRU. Only take this page
1221 * if it is of the appropriate PageActive status. Pages which are being
1222 * freed elsewhere are also ignored.
1223 *
1224 * page: page to consider
1225 * mode: one of the LRU isolation modes defined above
1226 *
1227 * returns 0 on success, -ve errno on failure.
1228 */
f3fd4a61 1229int __isolate_lru_page(struct page *page, isolate_mode_t mode)
5ad333eb
AW
1230{
1231 int ret = -EINVAL;
1232
1233 /* Only take pages on the LRU. */
1234 if (!PageLRU(page))
1235 return ret;
1236
e46a2879
MK
1237 /* Compaction should not handle unevictable pages but CMA can do so */
1238 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
894bc310
LS
1239 return ret;
1240
5ad333eb 1241 ret = -EBUSY;
08e552c6 1242
c8244935
MG
1243 /*
1244 * To minimise LRU disruption, the caller can indicate that it only
1245 * wants to isolate pages it will be able to operate on without
1246 * blocking - clean pages for the most part.
1247 *
1248 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1249 * is used by reclaim when it is cannot write to backing storage
1250 *
1251 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1252 * that it is possible to migrate without blocking
1253 */
1254 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1255 /* All the caller can do on PageWriteback is block */
1256 if (PageWriteback(page))
1257 return ret;
1258
1259 if (PageDirty(page)) {
1260 struct address_space *mapping;
1261
1262 /* ISOLATE_CLEAN means only clean pages */
1263 if (mode & ISOLATE_CLEAN)
1264 return ret;
1265
1266 /*
1267 * Only pages without mappings or that have a
1268 * ->migratepage callback are possible to migrate
1269 * without blocking
1270 */
1271 mapping = page_mapping(page);
1272 if (mapping && !mapping->a_ops->migratepage)
1273 return ret;
1274 }
1275 }
39deaf85 1276
f80c0673
MK
1277 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1278 return ret;
1279
5ad333eb
AW
1280 if (likely(get_page_unless_zero(page))) {
1281 /*
1282 * Be careful not to clear PageLRU until after we're
1283 * sure the page is not being freed elsewhere -- the
1284 * page release code relies on it.
1285 */
1286 ClearPageLRU(page);
1287 ret = 0;
1288 }
1289
1290 return ret;
1291}
1292
1da177e4
LT
1293/*
1294 * zone->lru_lock is heavily contended. Some of the functions that
1295 * shrink the lists perform better by taking out a batch of pages
1296 * and working on them outside the LRU lock.
1297 *
1298 * For pagecache intensive workloads, this function is the hottest
1299 * spot in the kernel (apart from copy_*_user functions).
1300 *
1301 * Appropriate locks must be held before calling this function.
1302 *
1303 * @nr_to_scan: The number of pages to look through on the list.
5dc35979 1304 * @lruvec: The LRU vector to pull pages from.
1da177e4 1305 * @dst: The temp list to put pages on to.
f626012d 1306 * @nr_scanned: The number of pages that were scanned.
fe2c2a10 1307 * @sc: The scan_control struct for this reclaim session
5ad333eb 1308 * @mode: One of the LRU isolation modes
3cb99451 1309 * @lru: LRU list id for isolating
1da177e4
LT
1310 *
1311 * returns how many pages were moved onto *@dst.
1312 */
69e05944 1313static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
5dc35979 1314 struct lruvec *lruvec, struct list_head *dst,
fe2c2a10 1315 unsigned long *nr_scanned, struct scan_control *sc,
3cb99451 1316 isolate_mode_t mode, enum lru_list lru)
1da177e4 1317{
75b00af7 1318 struct list_head *src = &lruvec->lists[lru];
69e05944 1319 unsigned long nr_taken = 0;
c9b02d97 1320 unsigned long scan;
1da177e4 1321
c9b02d97 1322 for (scan = 0; scan < nr_to_scan && !list_empty(src); scan++) {
5ad333eb 1323 struct page *page;
fa9add64 1324 int nr_pages;
5ad333eb 1325
1da177e4
LT
1326 page = lru_to_page(src);
1327 prefetchw_prev_lru_page(page, src, flags);
1328
309381fe 1329 VM_BUG_ON_PAGE(!PageLRU(page), page);
8d438f96 1330
f3fd4a61 1331 switch (__isolate_lru_page(page, mode)) {
5ad333eb 1332 case 0:
fa9add64
HD
1333 nr_pages = hpage_nr_pages(page);
1334 mem_cgroup_update_lru_size(lruvec, lru, -nr_pages);
5ad333eb 1335 list_move(&page->lru, dst);
fa9add64 1336 nr_taken += nr_pages;
5ad333eb
AW
1337 break;
1338
1339 case -EBUSY:
1340 /* else it is being freed elsewhere */
1341 list_move(&page->lru, src);
1342 continue;
46453a6e 1343
5ad333eb
AW
1344 default:
1345 BUG();
1346 }
1da177e4
LT
1347 }
1348
f626012d 1349 *nr_scanned = scan;
75b00af7
HD
1350 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1351 nr_taken, mode, is_file_lru(lru));
1da177e4
LT
1352 return nr_taken;
1353}
1354
62695a84
NP
1355/**
1356 * isolate_lru_page - tries to isolate a page from its LRU list
1357 * @page: page to isolate from its LRU list
1358 *
1359 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1360 * vmstat statistic corresponding to whatever LRU list the page was on.
1361 *
1362 * Returns 0 if the page was removed from an LRU list.
1363 * Returns -EBUSY if the page was not on an LRU list.
1364 *
1365 * The returned page will have PageLRU() cleared. If it was found on
894bc310
LS
1366 * the active list, it will have PageActive set. If it was found on
1367 * the unevictable list, it will have the PageUnevictable bit set. That flag
1368 * may need to be cleared by the caller before letting the page go.
62695a84
NP
1369 *
1370 * The vmstat statistic corresponding to the list on which the page was
1371 * found will be decremented.
1372 *
1373 * Restrictions:
1374 * (1) Must be called with an elevated refcount on the page. This is a
1375 * fundamentnal difference from isolate_lru_pages (which is called
1376 * without a stable reference).
1377 * (2) the lru_lock must not be held.
1378 * (3) interrupts must be enabled.
1379 */
1380int isolate_lru_page(struct page *page)
1381{
1382 int ret = -EBUSY;
1383
309381fe 1384 VM_BUG_ON_PAGE(!page_count(page), page);
0c917313 1385
62695a84
NP
1386 if (PageLRU(page)) {
1387 struct zone *zone = page_zone(page);
fa9add64 1388 struct lruvec *lruvec;
62695a84
NP
1389
1390 spin_lock_irq(&zone->lru_lock);
fa9add64 1391 lruvec = mem_cgroup_page_lruvec(page, zone);
0c917313 1392 if (PageLRU(page)) {
894bc310 1393 int lru = page_lru(page);
0c917313 1394 get_page(page);
62695a84 1395 ClearPageLRU(page);
fa9add64
HD
1396 del_page_from_lru_list(page, lruvec, lru);
1397 ret = 0;
62695a84
NP
1398 }
1399 spin_unlock_irq(&zone->lru_lock);
1400 }
1401 return ret;
1402}
1403
35cd7815 1404/*
d37dd5dc
FW
1405 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1406 * then get resheduled. When there are massive number of tasks doing page
1407 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1408 * the LRU list will go small and be scanned faster than necessary, leading to
1409 * unnecessary swapping, thrashing and OOM.
35cd7815
RR
1410 */
1411static int too_many_isolated(struct zone *zone, int file,
1412 struct scan_control *sc)
1413{
1414 unsigned long inactive, isolated;
1415
1416 if (current_is_kswapd())
1417 return 0;
1418
89b5fae5 1419 if (!global_reclaim(sc))
35cd7815
RR
1420 return 0;
1421
1422 if (file) {
1423 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1424 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1425 } else {
1426 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1427 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1428 }
1429
3cf23841
FW
1430 /*
1431 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1432 * won't get blocked by normal direct-reclaimers, forming a circular
1433 * deadlock.
1434 */
1435 if ((sc->gfp_mask & GFP_IOFS) == GFP_IOFS)
1436 inactive >>= 3;
1437
35cd7815
RR
1438 return isolated > inactive;
1439}
1440
66635629 1441static noinline_for_stack void
75b00af7 1442putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
66635629 1443{
27ac81d8
KK
1444 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1445 struct zone *zone = lruvec_zone(lruvec);
3f79768f 1446 LIST_HEAD(pages_to_free);
66635629 1447
66635629
MG
1448 /*
1449 * Put back any unfreeable pages.
1450 */
66635629 1451 while (!list_empty(page_list)) {
3f79768f 1452 struct page *page = lru_to_page(page_list);
66635629 1453 int lru;
3f79768f 1454
309381fe 1455 VM_BUG_ON_PAGE(PageLRU(page), page);
66635629 1456 list_del(&page->lru);
39b5f29a 1457 if (unlikely(!page_evictable(page))) {
66635629
MG
1458 spin_unlock_irq(&zone->lru_lock);
1459 putback_lru_page(page);
1460 spin_lock_irq(&zone->lru_lock);
1461 continue;
1462 }
fa9add64
HD
1463
1464 lruvec = mem_cgroup_page_lruvec(page, zone);
1465
7a608572 1466 SetPageLRU(page);
66635629 1467 lru = page_lru(page);
fa9add64
HD
1468 add_page_to_lru_list(page, lruvec, lru);
1469
66635629
MG
1470 if (is_active_lru(lru)) {
1471 int file = is_file_lru(lru);
9992af10
RR
1472 int numpages = hpage_nr_pages(page);
1473 reclaim_stat->recent_rotated[file] += numpages;
66635629 1474 }
2bcf8879
HD
1475 if (put_page_testzero(page)) {
1476 __ClearPageLRU(page);
1477 __ClearPageActive(page);
fa9add64 1478 del_page_from_lru_list(page, lruvec, lru);
2bcf8879
HD
1479
1480 if (unlikely(PageCompound(page))) {
1481 spin_unlock_irq(&zone->lru_lock);
747db954 1482 mem_cgroup_uncharge(page);
2bcf8879
HD
1483 (*get_compound_page_dtor(page))(page);
1484 spin_lock_irq(&zone->lru_lock);
1485 } else
1486 list_add(&page->lru, &pages_to_free);
66635629
MG
1487 }
1488 }
66635629 1489
3f79768f
HD
1490 /*
1491 * To save our caller's stack, now use input list for pages to free.
1492 */
1493 list_splice(&pages_to_free, page_list);
66635629
MG
1494}
1495
399ba0b9
N
1496/*
1497 * If a kernel thread (such as nfsd for loop-back mounts) services
1498 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1499 * In that case we should only throttle if the backing device it is
1500 * writing to is congested. In other cases it is safe to throttle.
1501 */
1502static int current_may_throttle(void)
1503{
1504 return !(current->flags & PF_LESS_THROTTLE) ||
1505 current->backing_dev_info == NULL ||
1506 bdi_write_congested(current->backing_dev_info);
1507}
1508
1da177e4 1509/*
1742f19f
AM
1510 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1511 * of reclaimed pages
1da177e4 1512 */
66635629 1513static noinline_for_stack unsigned long
1a93be0e 1514shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
9e3b2f8c 1515 struct scan_control *sc, enum lru_list lru)
1da177e4
LT
1516{
1517 LIST_HEAD(page_list);
e247dbce 1518 unsigned long nr_scanned;
05ff5137 1519 unsigned long nr_reclaimed = 0;
e247dbce 1520 unsigned long nr_taken;
8e950282
MG
1521 unsigned long nr_dirty = 0;
1522 unsigned long nr_congested = 0;
e2be15f6 1523 unsigned long nr_unqueued_dirty = 0;
92df3a72 1524 unsigned long nr_writeback = 0;
b1a6f21e 1525 unsigned long nr_immediate = 0;
f3fd4a61 1526 isolate_mode_t isolate_mode = 0;
3cb99451 1527 int file = is_file_lru(lru);
1a93be0e
KK
1528 struct zone *zone = lruvec_zone(lruvec);
1529 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
78dc583d 1530
35cd7815 1531 while (unlikely(too_many_isolated(zone, file, sc))) {
58355c78 1532 congestion_wait(BLK_RW_ASYNC, HZ/10);
35cd7815
RR
1533
1534 /* We are about to die and free our memory. Return now. */
1535 if (fatal_signal_pending(current))
1536 return SWAP_CLUSTER_MAX;
1537 }
1538
1da177e4 1539 lru_add_drain();
f80c0673
MK
1540
1541 if (!sc->may_unmap)
61317289 1542 isolate_mode |= ISOLATE_UNMAPPED;
f80c0673 1543 if (!sc->may_writepage)
61317289 1544 isolate_mode |= ISOLATE_CLEAN;
f80c0673 1545
1da177e4 1546 spin_lock_irq(&zone->lru_lock);
b35ea17b 1547
5dc35979
KK
1548 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1549 &nr_scanned, sc, isolate_mode, lru);
95d918fc
KK
1550
1551 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
1552 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1553
89b5fae5 1554 if (global_reclaim(sc)) {
0d5d823a 1555 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
e247dbce 1556 if (current_is_kswapd())
75b00af7 1557 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
e247dbce 1558 else
75b00af7 1559 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
e247dbce 1560 }
d563c050 1561 spin_unlock_irq(&zone->lru_lock);
b35ea17b 1562
d563c050 1563 if (nr_taken == 0)
66635629 1564 return 0;
5ad333eb 1565
02c6de8d 1566 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
8e950282
MG
1567 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1568 &nr_writeback, &nr_immediate,
1569 false);
c661b078 1570
3f79768f
HD
1571 spin_lock_irq(&zone->lru_lock);
1572
95d918fc 1573 reclaim_stat->recent_scanned[file] += nr_taken;
d563c050 1574
904249aa
YH
1575 if (global_reclaim(sc)) {
1576 if (current_is_kswapd())
1577 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1578 nr_reclaimed);
1579 else
1580 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1581 nr_reclaimed);
1582 }
a74609fa 1583
27ac81d8 1584 putback_inactive_pages(lruvec, &page_list);
3f79768f 1585
95d918fc 1586 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
3f79768f
HD
1587
1588 spin_unlock_irq(&zone->lru_lock);
1589
747db954 1590 mem_cgroup_uncharge_list(&page_list);
b745bc85 1591 free_hot_cold_page_list(&page_list, true);
e11da5b4 1592
92df3a72
MG
1593 /*
1594 * If reclaim is isolating dirty pages under writeback, it implies
1595 * that the long-lived page allocation rate is exceeding the page
1596 * laundering rate. Either the global limits are not being effective
1597 * at throttling processes due to the page distribution throughout
1598 * zones or there is heavy usage of a slow backing device. The
1599 * only option is to throttle from reclaim context which is not ideal
1600 * as there is no guarantee the dirtying process is throttled in the
1601 * same way balance_dirty_pages() manages.
1602 *
8e950282
MG
1603 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1604 * of pages under pages flagged for immediate reclaim and stall if any
1605 * are encountered in the nr_immediate check below.
92df3a72 1606 */
918fc718 1607 if (nr_writeback && nr_writeback == nr_taken)
57054651 1608 set_bit(ZONE_WRITEBACK, &zone->flags);
92df3a72 1609
d43006d5 1610 /*
b1a6f21e
MG
1611 * memcg will stall in page writeback so only consider forcibly
1612 * stalling for global reclaim
d43006d5 1613 */
b1a6f21e 1614 if (global_reclaim(sc)) {
8e950282
MG
1615 /*
1616 * Tag a zone as congested if all the dirty pages scanned were
1617 * backed by a congested BDI and wait_iff_congested will stall.
1618 */
1619 if (nr_dirty && nr_dirty == nr_congested)
57054651 1620 set_bit(ZONE_CONGESTED, &zone->flags);
8e950282 1621
b1a6f21e
MG
1622 /*
1623 * If dirty pages are scanned that are not queued for IO, it
1624 * implies that flushers are not keeping up. In this case, flag
57054651
JW
1625 * the zone ZONE_DIRTY and kswapd will start writing pages from
1626 * reclaim context.
b1a6f21e
MG
1627 */
1628 if (nr_unqueued_dirty == nr_taken)
57054651 1629 set_bit(ZONE_DIRTY, &zone->flags);
b1a6f21e
MG
1630
1631 /*
b738d764
LT
1632 * If kswapd scans pages marked marked for immediate
1633 * reclaim and under writeback (nr_immediate), it implies
1634 * that pages are cycling through the LRU faster than
b1a6f21e
MG
1635 * they are written so also forcibly stall.
1636 */
b738d764 1637 if (nr_immediate && current_may_throttle())
b1a6f21e 1638 congestion_wait(BLK_RW_ASYNC, HZ/10);
e2be15f6 1639 }
d43006d5 1640
8e950282
MG
1641 /*
1642 * Stall direct reclaim for IO completions if underlying BDIs or zone
1643 * is congested. Allow kswapd to continue until it starts encountering
1644 * unqueued dirty pages or cycling through the LRU too quickly.
1645 */
399ba0b9
N
1646 if (!sc->hibernation_mode && !current_is_kswapd() &&
1647 current_may_throttle())
8e950282
MG
1648 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1649
e11da5b4
MG
1650 trace_mm_vmscan_lru_shrink_inactive(zone->zone_pgdat->node_id,
1651 zone_idx(zone),
1652 nr_scanned, nr_reclaimed,
9e3b2f8c 1653 sc->priority,
23b9da55 1654 trace_shrink_flags(file));
05ff5137 1655 return nr_reclaimed;
1da177e4
LT
1656}
1657
1658/*
1659 * This moves pages from the active list to the inactive list.
1660 *
1661 * We move them the other way if the page is referenced by one or more
1662 * processes, from rmap.
1663 *
1664 * If the pages are mostly unmapped, the processing is fast and it is
1665 * appropriate to hold zone->lru_lock across the whole operation. But if
1666 * the pages are mapped, the processing is slow (page_referenced()) so we
1667 * should drop zone->lru_lock around each page. It's impossible to balance
1668 * this, so instead we remove the pages from the LRU while processing them.
1669 * It is safe to rely on PG_active against the non-LRU pages in here because
1670 * nobody will play with that bit on a non-LRU page.
1671 *
1672 * The downside is that we have to touch page->_count against each page.
1673 * But we had to alter page->flags anyway.
1674 */
1cfb419b 1675
fa9add64 1676static void move_active_pages_to_lru(struct lruvec *lruvec,
3eb4140f 1677 struct list_head *list,
2bcf8879 1678 struct list_head *pages_to_free,
3eb4140f
WF
1679 enum lru_list lru)
1680{
fa9add64 1681 struct zone *zone = lruvec_zone(lruvec);
3eb4140f 1682 unsigned long pgmoved = 0;
3eb4140f 1683 struct page *page;
fa9add64 1684 int nr_pages;
3eb4140f 1685
3eb4140f
WF
1686 while (!list_empty(list)) {
1687 page = lru_to_page(list);
fa9add64 1688 lruvec = mem_cgroup_page_lruvec(page, zone);
3eb4140f 1689
309381fe 1690 VM_BUG_ON_PAGE(PageLRU(page), page);
3eb4140f
WF
1691 SetPageLRU(page);
1692
fa9add64
HD
1693 nr_pages = hpage_nr_pages(page);
1694 mem_cgroup_update_lru_size(lruvec, lru, nr_pages);
925b7673 1695 list_move(&page->lru, &lruvec->lists[lru]);
fa9add64 1696 pgmoved += nr_pages;
3eb4140f 1697
2bcf8879
HD
1698 if (put_page_testzero(page)) {
1699 __ClearPageLRU(page);
1700 __ClearPageActive(page);
fa9add64 1701 del_page_from_lru_list(page, lruvec, lru);
2bcf8879
HD
1702
1703 if (unlikely(PageCompound(page))) {
1704 spin_unlock_irq(&zone->lru_lock);
747db954 1705 mem_cgroup_uncharge(page);
2bcf8879
HD
1706 (*get_compound_page_dtor(page))(page);
1707 spin_lock_irq(&zone->lru_lock);
1708 } else
1709 list_add(&page->lru, pages_to_free);
3eb4140f
WF
1710 }
1711 }
1712 __mod_zone_page_state(zone, NR_LRU_BASE + lru, pgmoved);
1713 if (!is_active_lru(lru))
1714 __count_vm_events(PGDEACTIVATE, pgmoved);
1715}
1cfb419b 1716
f626012d 1717static void shrink_active_list(unsigned long nr_to_scan,
1a93be0e 1718 struct lruvec *lruvec,
f16015fb 1719 struct scan_control *sc,
9e3b2f8c 1720 enum lru_list lru)
1da177e4 1721{
44c241f1 1722 unsigned long nr_taken;
f626012d 1723 unsigned long nr_scanned;
6fe6b7e3 1724 unsigned long vm_flags;
1da177e4 1725 LIST_HEAD(l_hold); /* The pages which were snipped off */
8cab4754 1726 LIST_HEAD(l_active);
b69408e8 1727 LIST_HEAD(l_inactive);
1da177e4 1728 struct page *page;
1a93be0e 1729 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
44c241f1 1730 unsigned long nr_rotated = 0;
f3fd4a61 1731 isolate_mode_t isolate_mode = 0;
3cb99451 1732 int file = is_file_lru(lru);
1a93be0e 1733 struct zone *zone = lruvec_zone(lruvec);
1da177e4
LT
1734
1735 lru_add_drain();
f80c0673
MK
1736
1737 if (!sc->may_unmap)
61317289 1738 isolate_mode |= ISOLATE_UNMAPPED;
f80c0673 1739 if (!sc->may_writepage)
61317289 1740 isolate_mode |= ISOLATE_CLEAN;
f80c0673 1741
1da177e4 1742 spin_lock_irq(&zone->lru_lock);
925b7673 1743
5dc35979
KK
1744 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1745 &nr_scanned, sc, isolate_mode, lru);
89b5fae5 1746 if (global_reclaim(sc))
0d5d823a 1747 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
89b5fae5 1748
b7c46d15 1749 reclaim_stat->recent_scanned[file] += nr_taken;
1cfb419b 1750
f626012d 1751 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
3cb99451 1752 __mod_zone_page_state(zone, NR_LRU_BASE + lru, -nr_taken);
a731286d 1753 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
1da177e4
LT
1754 spin_unlock_irq(&zone->lru_lock);
1755
1da177e4
LT
1756 while (!list_empty(&l_hold)) {
1757 cond_resched();
1758 page = lru_to_page(&l_hold);
1759 list_del(&page->lru);
7e9cd484 1760
39b5f29a 1761 if (unlikely(!page_evictable(page))) {
894bc310
LS
1762 putback_lru_page(page);
1763 continue;
1764 }
1765
cc715d99
MG
1766 if (unlikely(buffer_heads_over_limit)) {
1767 if (page_has_private(page) && trylock_page(page)) {
1768 if (page_has_private(page))
1769 try_to_release_page(page, 0);
1770 unlock_page(page);
1771 }
1772 }
1773
c3ac9a8a
JW
1774 if (page_referenced(page, 0, sc->target_mem_cgroup,
1775 &vm_flags)) {
9992af10 1776 nr_rotated += hpage_nr_pages(page);
8cab4754
WF
1777 /*
1778 * Identify referenced, file-backed active pages and
1779 * give them one more trip around the active list. So
1780 * that executable code get better chances to stay in
1781 * memory under moderate memory pressure. Anon pages
1782 * are not likely to be evicted by use-once streaming
1783 * IO, plus JVM can create lots of anon VM_EXEC pages,
1784 * so we ignore them here.
1785 */
41e20983 1786 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
8cab4754
WF
1787 list_add(&page->lru, &l_active);
1788 continue;
1789 }
1790 }
7e9cd484 1791
5205e56e 1792 ClearPageActive(page); /* we are de-activating */
1da177e4
LT
1793 list_add(&page->lru, &l_inactive);
1794 }
1795
b555749a 1796 /*
8cab4754 1797 * Move pages back to the lru list.
b555749a 1798 */
2a1dc509 1799 spin_lock_irq(&zone->lru_lock);
556adecb 1800 /*
8cab4754
WF
1801 * Count referenced pages from currently used mappings as rotated,
1802 * even though only some of them are actually re-activated. This
1803 * helps balance scan pressure between file and anonymous pages in
7c0db9e9 1804 * get_scan_count.
7e9cd484 1805 */
b7c46d15 1806 reclaim_stat->recent_rotated[file] += nr_rotated;
556adecb 1807
fa9add64
HD
1808 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1809 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
a731286d 1810 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
f8891e5e 1811 spin_unlock_irq(&zone->lru_lock);
2bcf8879 1812
747db954 1813 mem_cgroup_uncharge_list(&l_hold);
b745bc85 1814 free_hot_cold_page_list(&l_hold, true);
1da177e4
LT
1815}
1816
74e3f3c3 1817#ifdef CONFIG_SWAP
14797e23 1818static int inactive_anon_is_low_global(struct zone *zone)
f89eb90e
KM
1819{
1820 unsigned long active, inactive;
1821
1822 active = zone_page_state(zone, NR_ACTIVE_ANON);
1823 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1824
1825 if (inactive * zone->inactive_ratio < active)
1826 return 1;
1827
1828 return 0;
1829}
1830
14797e23
KM
1831/**
1832 * inactive_anon_is_low - check if anonymous pages need to be deactivated
c56d5c7d 1833 * @lruvec: LRU vector to check
14797e23
KM
1834 *
1835 * Returns true if the zone does not have enough inactive anon pages,
1836 * meaning some active anon pages need to be deactivated.
1837 */
c56d5c7d 1838static int inactive_anon_is_low(struct lruvec *lruvec)
14797e23 1839{
74e3f3c3
MK
1840 /*
1841 * If we don't have swap space, anonymous page deactivation
1842 * is pointless.
1843 */
1844 if (!total_swap_pages)
1845 return 0;
1846
c3c787e8 1847 if (!mem_cgroup_disabled())
c56d5c7d 1848 return mem_cgroup_inactive_anon_is_low(lruvec);
f16015fb 1849
c56d5c7d 1850 return inactive_anon_is_low_global(lruvec_zone(lruvec));
14797e23 1851}
74e3f3c3 1852#else
c56d5c7d 1853static inline int inactive_anon_is_low(struct lruvec *lruvec)
74e3f3c3
MK
1854{
1855 return 0;
1856}
1857#endif
14797e23 1858
56e49d21
RR
1859/**
1860 * inactive_file_is_low - check if file pages need to be deactivated
c56d5c7d 1861 * @lruvec: LRU vector to check
56e49d21
RR
1862 *
1863 * When the system is doing streaming IO, memory pressure here
1864 * ensures that active file pages get deactivated, until more
1865 * than half of the file pages are on the inactive list.
1866 *
1867 * Once we get to that situation, protect the system's working
1868 * set from being evicted by disabling active file page aging.
1869 *
1870 * This uses a different ratio than the anonymous pages, because
1871 * the page cache uses a use-once replacement algorithm.
1872 */
c56d5c7d 1873static int inactive_file_is_low(struct lruvec *lruvec)
56e49d21 1874{
e3790144
JW
1875 unsigned long inactive;
1876 unsigned long active;
1877
1878 inactive = get_lru_size(lruvec, LRU_INACTIVE_FILE);
1879 active = get_lru_size(lruvec, LRU_ACTIVE_FILE);
56e49d21 1880
e3790144 1881 return active > inactive;
56e49d21
RR
1882}
1883
75b00af7 1884static int inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
b39415b2 1885{
75b00af7 1886 if (is_file_lru(lru))
c56d5c7d 1887 return inactive_file_is_low(lruvec);
b39415b2 1888 else
c56d5c7d 1889 return inactive_anon_is_low(lruvec);
b39415b2
RR
1890}
1891
4f98a2fe 1892static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1a93be0e 1893 struct lruvec *lruvec, struct scan_control *sc)
b69408e8 1894{
b39415b2 1895 if (is_active_lru(lru)) {
75b00af7 1896 if (inactive_list_is_low(lruvec, lru))
1a93be0e 1897 shrink_active_list(nr_to_scan, lruvec, sc, lru);
556adecb
RR
1898 return 0;
1899 }
1900
1a93be0e 1901 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
4f98a2fe
RR
1902}
1903
9a265114
JW
1904enum scan_balance {
1905 SCAN_EQUAL,
1906 SCAN_FRACT,
1907 SCAN_ANON,
1908 SCAN_FILE,
1909};
1910
4f98a2fe
RR
1911/*
1912 * Determine how aggressively the anon and file LRU lists should be
1913 * scanned. The relative value of each set of LRU lists is determined
1914 * by looking at the fraction of the pages scanned we did rotate back
1915 * onto the active list instead of evict.
1916 *
be7bd59d
WL
1917 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1918 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
4f98a2fe 1919 */
02695175 1920static void get_scan_count(struct lruvec *lruvec, int swappiness,
6b4f7799
JW
1921 struct scan_control *sc, unsigned long *nr,
1922 unsigned long *lru_pages)
4f98a2fe 1923{
9a265114
JW
1924 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1925 u64 fraction[2];
1926 u64 denominator = 0; /* gcc */
1927 struct zone *zone = lruvec_zone(lruvec);
4f98a2fe 1928 unsigned long anon_prio, file_prio;
9a265114 1929 enum scan_balance scan_balance;
0bf1457f 1930 unsigned long anon, file;
9a265114 1931 bool force_scan = false;
4f98a2fe 1932 unsigned long ap, fp;
4111304d 1933 enum lru_list lru;
6f04f48d
SS
1934 bool some_scanned;
1935 int pass;
246e87a9 1936
f11c0ca5
JW
1937 /*
1938 * If the zone or memcg is small, nr[l] can be 0. This
1939 * results in no scanning on this priority and a potential
1940 * priority drop. Global direct reclaim can go to the next
1941 * zone and tends to have no problems. Global kswapd is for
1942 * zone balancing and it needs to scan a minimum amount. When
1943 * reclaiming for a memcg, a priority drop can cause high
1944 * latencies, so it's better to scan a minimum amount there as
1945 * well.
1946 */
90cbc250
VD
1947 if (current_is_kswapd()) {
1948 if (!zone_reclaimable(zone))
1949 force_scan = true;
1950 if (!mem_cgroup_lruvec_online(lruvec))
1951 force_scan = true;
1952 }
89b5fae5 1953 if (!global_reclaim(sc))
a4d3e9e7 1954 force_scan = true;
76a33fc3
SL
1955
1956 /* If we have no swap space, do not bother scanning anon pages. */
ec8acf20 1957 if (!sc->may_swap || (get_nr_swap_pages() <= 0)) {
9a265114 1958 scan_balance = SCAN_FILE;
76a33fc3
SL
1959 goto out;
1960 }
4f98a2fe 1961
10316b31
JW
1962 /*
1963 * Global reclaim will swap to prevent OOM even with no
1964 * swappiness, but memcg users want to use this knob to
1965 * disable swapping for individual groups completely when
1966 * using the memory controller's swap limit feature would be
1967 * too expensive.
1968 */
02695175 1969 if (!global_reclaim(sc) && !swappiness) {
9a265114 1970 scan_balance = SCAN_FILE;
10316b31
JW
1971 goto out;
1972 }
1973
1974 /*
1975 * Do not apply any pressure balancing cleverness when the
1976 * system is close to OOM, scan both anon and file equally
1977 * (unless the swappiness setting disagrees with swapping).
1978 */
02695175 1979 if (!sc->priority && swappiness) {
9a265114 1980 scan_balance = SCAN_EQUAL;
10316b31
JW
1981 goto out;
1982 }
1983
62376251
JW
1984 /*
1985 * Prevent the reclaimer from falling into the cache trap: as
1986 * cache pages start out inactive, every cache fault will tip
1987 * the scan balance towards the file LRU. And as the file LRU
1988 * shrinks, so does the window for rotation from references.
1989 * This means we have a runaway feedback loop where a tiny
1990 * thrashing file LRU becomes infinitely more attractive than
1991 * anon pages. Try to detect this based on file LRU size.
1992 */
1993 if (global_reclaim(sc)) {
2ab051e1
JM
1994 unsigned long zonefile;
1995 unsigned long zonefree;
1996
1997 zonefree = zone_page_state(zone, NR_FREE_PAGES);
1998 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
1999 zone_page_state(zone, NR_INACTIVE_FILE);
62376251 2000
2ab051e1 2001 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
62376251
JW
2002 scan_balance = SCAN_ANON;
2003 goto out;
2004 }
2005 }
2006
7c5bd705
JW
2007 /*
2008 * There is enough inactive page cache, do not reclaim
2009 * anything from the anonymous working set right now.
2010 */
2011 if (!inactive_file_is_low(lruvec)) {
9a265114 2012 scan_balance = SCAN_FILE;
7c5bd705
JW
2013 goto out;
2014 }
2015
9a265114
JW
2016 scan_balance = SCAN_FRACT;
2017
58c37f6e
KM
2018 /*
2019 * With swappiness at 100, anonymous and file have the same priority.
2020 * This scanning priority is essentially the inverse of IO cost.
2021 */
02695175 2022 anon_prio = swappiness;
75b00af7 2023 file_prio = 200 - anon_prio;
58c37f6e 2024
4f98a2fe
RR
2025 /*
2026 * OK, so we have swap space and a fair amount of page cache
2027 * pages. We use the recently rotated / recently scanned
2028 * ratios to determine how valuable each cache is.
2029 *
2030 * Because workloads change over time (and to avoid overflow)
2031 * we keep these statistics as a floating average, which ends
2032 * up weighing recent references more than old ones.
2033 *
2034 * anon in [0], file in [1]
2035 */
2ab051e1
JM
2036
2037 anon = get_lru_size(lruvec, LRU_ACTIVE_ANON) +
2038 get_lru_size(lruvec, LRU_INACTIVE_ANON);
2039 file = get_lru_size(lruvec, LRU_ACTIVE_FILE) +
2040 get_lru_size(lruvec, LRU_INACTIVE_FILE);
2041
90126375 2042 spin_lock_irq(&zone->lru_lock);
6e901571 2043 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
6e901571
KM
2044 reclaim_stat->recent_scanned[0] /= 2;
2045 reclaim_stat->recent_rotated[0] /= 2;
4f98a2fe
RR
2046 }
2047
6e901571 2048 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
6e901571
KM
2049 reclaim_stat->recent_scanned[1] /= 2;
2050 reclaim_stat->recent_rotated[1] /= 2;
4f98a2fe
RR
2051 }
2052
4f98a2fe 2053 /*
00d8089c
RR
2054 * The amount of pressure on anon vs file pages is inversely
2055 * proportional to the fraction of recently scanned pages on
2056 * each list that were recently referenced and in active use.
4f98a2fe 2057 */
fe35004f 2058 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
6e901571 2059 ap /= reclaim_stat->recent_rotated[0] + 1;
4f98a2fe 2060
fe35004f 2061 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
6e901571 2062 fp /= reclaim_stat->recent_rotated[1] + 1;
90126375 2063 spin_unlock_irq(&zone->lru_lock);
4f98a2fe 2064
76a33fc3
SL
2065 fraction[0] = ap;
2066 fraction[1] = fp;
2067 denominator = ap + fp + 1;
2068out:
6f04f48d
SS
2069 some_scanned = false;
2070 /* Only use force_scan on second pass. */
2071 for (pass = 0; !some_scanned && pass < 2; pass++) {
6b4f7799 2072 *lru_pages = 0;
6f04f48d
SS
2073 for_each_evictable_lru(lru) {
2074 int file = is_file_lru(lru);
2075 unsigned long size;
2076 unsigned long scan;
6e08a369 2077
6f04f48d
SS
2078 size = get_lru_size(lruvec, lru);
2079 scan = size >> sc->priority;
9a265114 2080
6f04f48d
SS
2081 if (!scan && pass && force_scan)
2082 scan = min(size, SWAP_CLUSTER_MAX);
9a265114 2083
6f04f48d
SS
2084 switch (scan_balance) {
2085 case SCAN_EQUAL:
2086 /* Scan lists relative to size */
2087 break;
2088 case SCAN_FRACT:
2089 /*
2090 * Scan types proportional to swappiness and
2091 * their relative recent reclaim efficiency.
2092 */
2093 scan = div64_u64(scan * fraction[file],
2094 denominator);
2095 break;
2096 case SCAN_FILE:
2097 case SCAN_ANON:
2098 /* Scan one type exclusively */
6b4f7799
JW
2099 if ((scan_balance == SCAN_FILE) != file) {
2100 size = 0;
6f04f48d 2101 scan = 0;
6b4f7799 2102 }
6f04f48d
SS
2103 break;
2104 default:
2105 /* Look ma, no brain */
2106 BUG();
2107 }
6b4f7799
JW
2108
2109 *lru_pages += size;
6f04f48d 2110 nr[lru] = scan;
6b4f7799 2111
9a265114 2112 /*
6f04f48d
SS
2113 * Skip the second pass and don't force_scan,
2114 * if we found something to scan.
9a265114 2115 */
6f04f48d 2116 some_scanned |= !!scan;
9a265114 2117 }
76a33fc3 2118 }
6e08a369 2119}
4f98a2fe 2120
9b4f98cd
JW
2121/*
2122 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2123 */
02695175 2124static void shrink_lruvec(struct lruvec *lruvec, int swappiness,
6b4f7799 2125 struct scan_control *sc, unsigned long *lru_pages)
9b4f98cd
JW
2126{
2127 unsigned long nr[NR_LRU_LISTS];
e82e0561 2128 unsigned long targets[NR_LRU_LISTS];
9b4f98cd
JW
2129 unsigned long nr_to_scan;
2130 enum lru_list lru;
2131 unsigned long nr_reclaimed = 0;
2132 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2133 struct blk_plug plug;
1a501907 2134 bool scan_adjusted;
9b4f98cd 2135
6b4f7799 2136 get_scan_count(lruvec, swappiness, sc, nr, lru_pages);
9b4f98cd 2137
e82e0561
MG
2138 /* Record the original scan target for proportional adjustments later */
2139 memcpy(targets, nr, sizeof(nr));
2140
1a501907
MG
2141 /*
2142 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2143 * event that can occur when there is little memory pressure e.g.
2144 * multiple streaming readers/writers. Hence, we do not abort scanning
2145 * when the requested number of pages are reclaimed when scanning at
2146 * DEF_PRIORITY on the assumption that the fact we are direct
2147 * reclaiming implies that kswapd is not keeping up and it is best to
2148 * do a batch of work at once. For memcg reclaim one check is made to
2149 * abort proportional reclaim if either the file or anon lru has already
2150 * dropped to zero at the first pass.
2151 */
2152 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2153 sc->priority == DEF_PRIORITY);
2154
9b4f98cd
JW
2155 blk_start_plug(&plug);
2156 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2157 nr[LRU_INACTIVE_FILE]) {
e82e0561
MG
2158 unsigned long nr_anon, nr_file, percentage;
2159 unsigned long nr_scanned;
2160
9b4f98cd
JW
2161 for_each_evictable_lru(lru) {
2162 if (nr[lru]) {
2163 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2164 nr[lru] -= nr_to_scan;
2165
2166 nr_reclaimed += shrink_list(lru, nr_to_scan,
2167 lruvec, sc);
2168 }
2169 }
e82e0561
MG
2170
2171 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2172 continue;
2173
e82e0561
MG
2174 /*
2175 * For kswapd and memcg, reclaim at least the number of pages
1a501907 2176 * requested. Ensure that the anon and file LRUs are scanned
e82e0561
MG
2177 * proportionally what was requested by get_scan_count(). We
2178 * stop reclaiming one LRU and reduce the amount scanning
2179 * proportional to the original scan target.
2180 */
2181 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2182 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2183
1a501907
MG
2184 /*
2185 * It's just vindictive to attack the larger once the smaller
2186 * has gone to zero. And given the way we stop scanning the
2187 * smaller below, this makes sure that we only make one nudge
2188 * towards proportionality once we've got nr_to_reclaim.
2189 */
2190 if (!nr_file || !nr_anon)
2191 break;
2192
e82e0561
MG
2193 if (nr_file > nr_anon) {
2194 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2195 targets[LRU_ACTIVE_ANON] + 1;
2196 lru = LRU_BASE;
2197 percentage = nr_anon * 100 / scan_target;
2198 } else {
2199 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2200 targets[LRU_ACTIVE_FILE] + 1;
2201 lru = LRU_FILE;
2202 percentage = nr_file * 100 / scan_target;
2203 }
2204
2205 /* Stop scanning the smaller of the LRU */
2206 nr[lru] = 0;
2207 nr[lru + LRU_ACTIVE] = 0;
2208
2209 /*
2210 * Recalculate the other LRU scan count based on its original
2211 * scan target and the percentage scanning already complete
2212 */
2213 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2214 nr_scanned = targets[lru] - nr[lru];
2215 nr[lru] = targets[lru] * (100 - percentage) / 100;
2216 nr[lru] -= min(nr[lru], nr_scanned);
2217
2218 lru += LRU_ACTIVE;
2219 nr_scanned = targets[lru] - nr[lru];
2220 nr[lru] = targets[lru] * (100 - percentage) / 100;
2221 nr[lru] -= min(nr[lru], nr_scanned);
2222
2223 scan_adjusted = true;
9b4f98cd
JW
2224 }
2225 blk_finish_plug(&plug);
2226 sc->nr_reclaimed += nr_reclaimed;
2227
2228 /*
2229 * Even if we did not try to evict anon pages at all, we want to
2230 * rebalance the anon lru active/inactive ratio.
2231 */
2232 if (inactive_anon_is_low(lruvec))
2233 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2234 sc, LRU_ACTIVE_ANON);
2235
2236 throttle_vm_writeout(sc->gfp_mask);
2237}
2238
23b9da55 2239/* Use reclaim/compaction for costly allocs or under memory pressure */
9e3b2f8c 2240static bool in_reclaim_compaction(struct scan_control *sc)
23b9da55 2241{
d84da3f9 2242 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
23b9da55 2243 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
9e3b2f8c 2244 sc->priority < DEF_PRIORITY - 2))
23b9da55
MG
2245 return true;
2246
2247 return false;
2248}
2249
3e7d3449 2250/*
23b9da55
MG
2251 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2252 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2253 * true if more pages should be reclaimed such that when the page allocator
2254 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2255 * It will give up earlier than that if there is difficulty reclaiming pages.
3e7d3449 2256 */
9b4f98cd 2257static inline bool should_continue_reclaim(struct zone *zone,
3e7d3449
MG
2258 unsigned long nr_reclaimed,
2259 unsigned long nr_scanned,
2260 struct scan_control *sc)
2261{
2262 unsigned long pages_for_compaction;
2263 unsigned long inactive_lru_pages;
2264
2265 /* If not in reclaim/compaction mode, stop */
9e3b2f8c 2266 if (!in_reclaim_compaction(sc))
3e7d3449
MG
2267 return false;
2268
2876592f
MG
2269 /* Consider stopping depending on scan and reclaim activity */
2270 if (sc->gfp_mask & __GFP_REPEAT) {
2271 /*
2272 * For __GFP_REPEAT allocations, stop reclaiming if the
2273 * full LRU list has been scanned and we are still failing
2274 * to reclaim pages. This full LRU scan is potentially
2275 * expensive but a __GFP_REPEAT caller really wants to succeed
2276 */
2277 if (!nr_reclaimed && !nr_scanned)
2278 return false;
2279 } else {
2280 /*
2281 * For non-__GFP_REPEAT allocations which can presumably
2282 * fail without consequence, stop if we failed to reclaim
2283 * any pages from the last SWAP_CLUSTER_MAX number of
2284 * pages that were scanned. This will return to the
2285 * caller faster at the risk reclaim/compaction and
2286 * the resulting allocation attempt fails
2287 */
2288 if (!nr_reclaimed)
2289 return false;
2290 }
3e7d3449
MG
2291
2292 /*
2293 * If we have not reclaimed enough pages for compaction and the
2294 * inactive lists are large enough, continue reclaiming
2295 */
2296 pages_for_compaction = (2UL << sc->order);
9b4f98cd 2297 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
ec8acf20 2298 if (get_nr_swap_pages() > 0)
9b4f98cd 2299 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
3e7d3449
MG
2300 if (sc->nr_reclaimed < pages_for_compaction &&
2301 inactive_lru_pages > pages_for_compaction)
2302 return true;
2303
2304 /* If compaction would go ahead or the allocation would succeed, stop */
ebff3980 2305 switch (compaction_suitable(zone, sc->order, 0, 0)) {
3e7d3449
MG
2306 case COMPACT_PARTIAL:
2307 case COMPACT_CONTINUE:
2308 return false;
2309 default:
2310 return true;
2311 }
2312}
2313
6b4f7799
JW
2314static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2315 bool is_classzone)
1da177e4 2316{
cb731d6c 2317 struct reclaim_state *reclaim_state = current->reclaim_state;
f0fdc5e8 2318 unsigned long nr_reclaimed, nr_scanned;
2344d7e4 2319 bool reclaimable = false;
1da177e4 2320
9b4f98cd
JW
2321 do {
2322 struct mem_cgroup *root = sc->target_mem_cgroup;
2323 struct mem_cgroup_reclaim_cookie reclaim = {
2324 .zone = zone,
2325 .priority = sc->priority,
2326 };
6b4f7799 2327 unsigned long zone_lru_pages = 0;
694fbc0f 2328 struct mem_cgroup *memcg;
3e7d3449 2329
9b4f98cd
JW
2330 nr_reclaimed = sc->nr_reclaimed;
2331 nr_scanned = sc->nr_scanned;
1da177e4 2332
694fbc0f
AM
2333 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2334 do {
6b4f7799 2335 unsigned long lru_pages;
cb731d6c 2336 unsigned long scanned;
9b4f98cd 2337 struct lruvec *lruvec;
02695175 2338 int swappiness;
5660048c 2339
241994ed
JW
2340 if (mem_cgroup_low(root, memcg)) {
2341 if (!sc->may_thrash)
2342 continue;
2343 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2344 }
2345
9b4f98cd 2346 lruvec = mem_cgroup_zone_lruvec(zone, memcg);
02695175 2347 swappiness = mem_cgroup_swappiness(memcg);
cb731d6c 2348 scanned = sc->nr_scanned;
f9be23d6 2349
6b4f7799
JW
2350 shrink_lruvec(lruvec, swappiness, sc, &lru_pages);
2351 zone_lru_pages += lru_pages;
f16015fb 2352
cb731d6c
VD
2353 if (memcg && is_classzone)
2354 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2355 memcg, sc->nr_scanned - scanned,
2356 lru_pages);
2357
9b4f98cd 2358 /*
a394cb8e
MH
2359 * Direct reclaim and kswapd have to scan all memory
2360 * cgroups to fulfill the overall scan target for the
9b4f98cd 2361 * zone.
a394cb8e
MH
2362 *
2363 * Limit reclaim, on the other hand, only cares about
2364 * nr_to_reclaim pages to be reclaimed and it will
2365 * retry with decreasing priority if one round over the
2366 * whole hierarchy is not sufficient.
9b4f98cd 2367 */
a394cb8e
MH
2368 if (!global_reclaim(sc) &&
2369 sc->nr_reclaimed >= sc->nr_to_reclaim) {
9b4f98cd
JW
2370 mem_cgroup_iter_break(root, memcg);
2371 break;
2372 }
241994ed 2373 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
70ddf637 2374
6b4f7799
JW
2375 /*
2376 * Shrink the slab caches in the same proportion that
2377 * the eligible LRU pages were scanned.
2378 */
cb731d6c
VD
2379 if (global_reclaim(sc) && is_classzone)
2380 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2381 sc->nr_scanned - nr_scanned,
2382 zone_lru_pages);
2383
2384 if (reclaim_state) {
2385 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2386 reclaim_state->reclaimed_slab = 0;
6b4f7799
JW
2387 }
2388
70ddf637
AV
2389 vmpressure(sc->gfp_mask, sc->target_mem_cgroup,
2390 sc->nr_scanned - nr_scanned,
2391 sc->nr_reclaimed - nr_reclaimed);
2392
2344d7e4
JW
2393 if (sc->nr_reclaimed - nr_reclaimed)
2394 reclaimable = true;
2395
9b4f98cd
JW
2396 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2397 sc->nr_scanned - nr_scanned, sc));
2344d7e4
JW
2398
2399 return reclaimable;
f16015fb
JW
2400}
2401
53853e2d
VB
2402/*
2403 * Returns true if compaction should go ahead for a high-order request, or
2404 * the high-order allocation would succeed without compaction.
2405 */
0b06496a 2406static inline bool compaction_ready(struct zone *zone, int order)
fe4b1b24
MG
2407{
2408 unsigned long balance_gap, watermark;
2409 bool watermark_ok;
2410
fe4b1b24
MG
2411 /*
2412 * Compaction takes time to run and there are potentially other
2413 * callers using the pages just freed. Continue reclaiming until
2414 * there is a buffer of free pages available to give compaction
2415 * a reasonable chance of completing and allocating the page
2416 */
4be89a34
JZ
2417 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2418 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
0b06496a 2419 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
fe4b1b24
MG
2420 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0, 0);
2421
2422 /*
2423 * If compaction is deferred, reclaim up to a point where
2424 * compaction will have a chance of success when re-enabled
2425 */
0b06496a 2426 if (compaction_deferred(zone, order))
fe4b1b24
MG
2427 return watermark_ok;
2428
53853e2d
VB
2429 /*
2430 * If compaction is not ready to start and allocation is not likely
2431 * to succeed without it, then keep reclaiming.
2432 */
ebff3980 2433 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
fe4b1b24
MG
2434 return false;
2435
2436 return watermark_ok;
2437}
2438
1da177e4
LT
2439/*
2440 * This is the direct reclaim path, for page-allocating processes. We only
2441 * try to reclaim pages from zones which will satisfy the caller's allocation
2442 * request.
2443 *
41858966
MG
2444 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2445 * Because:
1da177e4
LT
2446 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2447 * allocation or
41858966
MG
2448 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2449 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2450 * zone defense algorithm.
1da177e4 2451 *
1da177e4
LT
2452 * If a zone is deemed to be full of pinned pages then just give it a light
2453 * scan then give up on it.
2344d7e4
JW
2454 *
2455 * Returns true if a zone was reclaimable.
1da177e4 2456 */
2344d7e4 2457static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
1da177e4 2458{
dd1a239f 2459 struct zoneref *z;
54a6eb5c 2460 struct zone *zone;
0608f43d
AM
2461 unsigned long nr_soft_reclaimed;
2462 unsigned long nr_soft_scanned;
619d0d76 2463 gfp_t orig_mask;
9bbc04ee 2464 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2344d7e4 2465 bool reclaimable = false;
1cfb419b 2466
cc715d99
MG
2467 /*
2468 * If the number of buffer_heads in the machine exceeds the maximum
2469 * allowed level, force direct reclaim to scan the highmem zone as
2470 * highmem pages could be pinning lowmem pages storing buffer_heads
2471 */
619d0d76 2472 orig_mask = sc->gfp_mask;
cc715d99
MG
2473 if (buffer_heads_over_limit)
2474 sc->gfp_mask |= __GFP_HIGHMEM;
2475
d4debc66 2476 for_each_zone_zonelist_nodemask(zone, z, zonelist,
6b4f7799
JW
2477 requested_highidx, sc->nodemask) {
2478 enum zone_type classzone_idx;
2479
f3fe6512 2480 if (!populated_zone(zone))
1da177e4 2481 continue;
6b4f7799
JW
2482
2483 classzone_idx = requested_highidx;
2484 while (!populated_zone(zone->zone_pgdat->node_zones +
2485 classzone_idx))
2486 classzone_idx--;
2487
1cfb419b
KH
2488 /*
2489 * Take care memory controller reclaiming has small influence
2490 * to global LRU.
2491 */
89b5fae5 2492 if (global_reclaim(sc)) {
344736f2
VD
2493 if (!cpuset_zone_allowed(zone,
2494 GFP_KERNEL | __GFP_HARDWALL))
1cfb419b 2495 continue;
65ec02cb 2496
6e543d57
LD
2497 if (sc->priority != DEF_PRIORITY &&
2498 !zone_reclaimable(zone))
1cfb419b 2499 continue; /* Let kswapd poll it */
0b06496a
JW
2500
2501 /*
2502 * If we already have plenty of memory free for
2503 * compaction in this zone, don't free any more.
2504 * Even though compaction is invoked for any
2505 * non-zero order, only frequent costly order
2506 * reclamation is disruptive enough to become a
2507 * noticeable problem, like transparent huge
2508 * page allocations.
2509 */
2510 if (IS_ENABLED(CONFIG_COMPACTION) &&
2511 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2512 zonelist_zone_idx(z) <= requested_highidx &&
2513 compaction_ready(zone, sc->order)) {
2514 sc->compaction_ready = true;
2515 continue;
e0887c19 2516 }
0b06496a 2517
0608f43d
AM
2518 /*
2519 * This steals pages from memory cgroups over softlimit
2520 * and returns the number of reclaimed pages and
2521 * scanned pages. This works for global memory pressure
2522 * and balancing, not for a memcg's limit.
2523 */
2524 nr_soft_scanned = 0;
2525 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2526 sc->order, sc->gfp_mask,
2527 &nr_soft_scanned);
2528 sc->nr_reclaimed += nr_soft_reclaimed;
2529 sc->nr_scanned += nr_soft_scanned;
2344d7e4
JW
2530 if (nr_soft_reclaimed)
2531 reclaimable = true;
ac34a1a3 2532 /* need some check for avoid more shrink_zone() */
1cfb419b 2533 }
408d8544 2534
6b4f7799 2535 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2344d7e4
JW
2536 reclaimable = true;
2537
2538 if (global_reclaim(sc) &&
2539 !reclaimable && zone_reclaimable(zone))
2540 reclaimable = true;
1da177e4 2541 }
e0c23279 2542
619d0d76
WY
2543 /*
2544 * Restore to original mask to avoid the impact on the caller if we
2545 * promoted it to __GFP_HIGHMEM.
2546 */
2547 sc->gfp_mask = orig_mask;
d1908362 2548
2344d7e4 2549 return reclaimable;
1da177e4 2550}
4f98a2fe 2551
1da177e4
LT
2552/*
2553 * This is the main entry point to direct page reclaim.
2554 *
2555 * If a full scan of the inactive list fails to free enough memory then we
2556 * are "out of memory" and something needs to be killed.
2557 *
2558 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2559 * high - the zone may be full of dirty or under-writeback pages, which this
5b0830cb
JA
2560 * caller can't do much about. We kick the writeback threads and take explicit
2561 * naps in the hope that some of these pages can be written. But if the
2562 * allocating task holds filesystem locks which prevent writeout this might not
2563 * work, and the allocation attempt will fail.
a41f24ea
NA
2564 *
2565 * returns: 0, if no pages reclaimed
2566 * else, the number of pages reclaimed
1da177e4 2567 */
dac1d27b 2568static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3115cd91 2569 struct scan_control *sc)
1da177e4 2570{
241994ed 2571 int initial_priority = sc->priority;
69e05944 2572 unsigned long total_scanned = 0;
22fba335 2573 unsigned long writeback_threshold;
2344d7e4 2574 bool zones_reclaimable;
241994ed 2575retry:
873b4771
KK
2576 delayacct_freepages_start();
2577
89b5fae5 2578 if (global_reclaim(sc))
1cfb419b 2579 count_vm_event(ALLOCSTALL);
1da177e4 2580
9e3b2f8c 2581 do {
70ddf637
AV
2582 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2583 sc->priority);
66e1707b 2584 sc->nr_scanned = 0;
2344d7e4 2585 zones_reclaimable = shrink_zones(zonelist, sc);
c6a8a8c5 2586
66e1707b 2587 total_scanned += sc->nr_scanned;
bb21c7ce 2588 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
0b06496a
JW
2589 break;
2590
2591 if (sc->compaction_ready)
2592 break;
1da177e4 2593
0e50ce3b
MK
2594 /*
2595 * If we're getting trouble reclaiming, start doing
2596 * writepage even in laptop mode.
2597 */
2598 if (sc->priority < DEF_PRIORITY - 2)
2599 sc->may_writepage = 1;
2600
1da177e4
LT
2601 /*
2602 * Try to write back as many pages as we just scanned. This
2603 * tends to cause slow streaming writers to write data to the
2604 * disk smoothly, at the dirtying rate, which is nice. But
2605 * that's undesirable in laptop mode, where we *want* lumpy
2606 * writeout. So in laptop mode, write out the whole world.
2607 */
22fba335
KM
2608 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2609 if (total_scanned > writeback_threshold) {
0e175a18
CW
2610 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2611 WB_REASON_TRY_TO_FREE_PAGES);
66e1707b 2612 sc->may_writepage = 1;
1da177e4 2613 }
0b06496a 2614 } while (--sc->priority >= 0);
bb21c7ce 2615
873b4771
KK
2616 delayacct_freepages_end();
2617
bb21c7ce
KM
2618 if (sc->nr_reclaimed)
2619 return sc->nr_reclaimed;
2620
0cee34fd 2621 /* Aborted reclaim to try compaction? don't OOM, then */
0b06496a 2622 if (sc->compaction_ready)
7335084d
MG
2623 return 1;
2624
241994ed
JW
2625 /* Untapped cgroup reserves? Don't OOM, retry. */
2626 if (!sc->may_thrash) {
2627 sc->priority = initial_priority;
2628 sc->may_thrash = 1;
2629 goto retry;
2630 }
2631
2344d7e4
JW
2632 /* Any of the zones still reclaimable? Don't OOM. */
2633 if (zones_reclaimable)
bb21c7ce
KM
2634 return 1;
2635
2636 return 0;
1da177e4
LT
2637}
2638
5515061d
MG
2639static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2640{
2641 struct zone *zone;
2642 unsigned long pfmemalloc_reserve = 0;
2643 unsigned long free_pages = 0;
2644 int i;
2645 bool wmark_ok;
2646
2647 for (i = 0; i <= ZONE_NORMAL; i++) {
2648 zone = &pgdat->node_zones[i];
675becce
MG
2649 if (!populated_zone(zone))
2650 continue;
2651
5515061d
MG
2652 pfmemalloc_reserve += min_wmark_pages(zone);
2653 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2654 }
2655
675becce
MG
2656 /* If there are no reserves (unexpected config) then do not throttle */
2657 if (!pfmemalloc_reserve)
2658 return true;
2659
5515061d
MG
2660 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2661
2662 /* kswapd must be awake if processes are being throttled */
2663 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2664 pgdat->classzone_idx = min(pgdat->classzone_idx,
2665 (enum zone_type)ZONE_NORMAL);
2666 wake_up_interruptible(&pgdat->kswapd_wait);
2667 }
2668
2669 return wmark_ok;
2670}
2671
2672/*
2673 * Throttle direct reclaimers if backing storage is backed by the network
2674 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2675 * depleted. kswapd will continue to make progress and wake the processes
50694c28
MG
2676 * when the low watermark is reached.
2677 *
2678 * Returns true if a fatal signal was delivered during throttling. If this
2679 * happens, the page allocator should not consider triggering the OOM killer.
5515061d 2680 */
50694c28 2681static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
5515061d
MG
2682 nodemask_t *nodemask)
2683{
675becce 2684 struct zoneref *z;
5515061d 2685 struct zone *zone;
675becce 2686 pg_data_t *pgdat = NULL;
5515061d
MG
2687
2688 /*
2689 * Kernel threads should not be throttled as they may be indirectly
2690 * responsible for cleaning pages necessary for reclaim to make forward
2691 * progress. kjournald for example may enter direct reclaim while
2692 * committing a transaction where throttling it could forcing other
2693 * processes to block on log_wait_commit().
2694 */
2695 if (current->flags & PF_KTHREAD)
50694c28
MG
2696 goto out;
2697
2698 /*
2699 * If a fatal signal is pending, this process should not throttle.
2700 * It should return quickly so it can exit and free its memory
2701 */
2702 if (fatal_signal_pending(current))
2703 goto out;
5515061d 2704
675becce
MG
2705 /*
2706 * Check if the pfmemalloc reserves are ok by finding the first node
2707 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2708 * GFP_KERNEL will be required for allocating network buffers when
2709 * swapping over the network so ZONE_HIGHMEM is unusable.
2710 *
2711 * Throttling is based on the first usable node and throttled processes
2712 * wait on a queue until kswapd makes progress and wakes them. There
2713 * is an affinity then between processes waking up and where reclaim
2714 * progress has been made assuming the process wakes on the same node.
2715 * More importantly, processes running on remote nodes will not compete
2716 * for remote pfmemalloc reserves and processes on different nodes
2717 * should make reasonable progress.
2718 */
2719 for_each_zone_zonelist_nodemask(zone, z, zonelist,
17636faa 2720 gfp_zone(gfp_mask), nodemask) {
675becce
MG
2721 if (zone_idx(zone) > ZONE_NORMAL)
2722 continue;
2723
2724 /* Throttle based on the first usable node */
2725 pgdat = zone->zone_pgdat;
2726 if (pfmemalloc_watermark_ok(pgdat))
2727 goto out;
2728 break;
2729 }
2730
2731 /* If no zone was usable by the allocation flags then do not throttle */
2732 if (!pgdat)
50694c28 2733 goto out;
5515061d 2734
68243e76
MG
2735 /* Account for the throttling */
2736 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2737
5515061d
MG
2738 /*
2739 * If the caller cannot enter the filesystem, it's possible that it
2740 * is due to the caller holding an FS lock or performing a journal
2741 * transaction in the case of a filesystem like ext[3|4]. In this case,
2742 * it is not safe to block on pfmemalloc_wait as kswapd could be
2743 * blocked waiting on the same lock. Instead, throttle for up to a
2744 * second before continuing.
2745 */
2746 if (!(gfp_mask & __GFP_FS)) {
2747 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2748 pfmemalloc_watermark_ok(pgdat), HZ);
50694c28
MG
2749
2750 goto check_pending;
5515061d
MG
2751 }
2752
2753 /* Throttle until kswapd wakes the process */
2754 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2755 pfmemalloc_watermark_ok(pgdat));
50694c28
MG
2756
2757check_pending:
2758 if (fatal_signal_pending(current))
2759 return true;
2760
2761out:
2762 return false;
5515061d
MG
2763}
2764
dac1d27b 2765unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
327c0e96 2766 gfp_t gfp_mask, nodemask_t *nodemask)
66e1707b 2767{
33906bc5 2768 unsigned long nr_reclaimed;
66e1707b 2769 struct scan_control sc = {
ee814fe2 2770 .nr_to_reclaim = SWAP_CLUSTER_MAX,
21caf2fc 2771 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
ee814fe2
JW
2772 .order = order,
2773 .nodemask = nodemask,
2774 .priority = DEF_PRIORITY,
66e1707b 2775 .may_writepage = !laptop_mode,
a6dc60f8 2776 .may_unmap = 1,
2e2e4259 2777 .may_swap = 1,
66e1707b
BS
2778 };
2779
5515061d 2780 /*
50694c28
MG
2781 * Do not enter reclaim if fatal signal was delivered while throttled.
2782 * 1 is returned so that the page allocator does not OOM kill at this
2783 * point.
5515061d 2784 */
50694c28 2785 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
5515061d
MG
2786 return 1;
2787
33906bc5
MG
2788 trace_mm_vmscan_direct_reclaim_begin(order,
2789 sc.may_writepage,
2790 gfp_mask);
2791
3115cd91 2792 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
33906bc5
MG
2793
2794 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2795
2796 return nr_reclaimed;
66e1707b
BS
2797}
2798
c255a458 2799#ifdef CONFIG_MEMCG
66e1707b 2800
72835c86 2801unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
4e416953 2802 gfp_t gfp_mask, bool noswap,
0ae5e89c
YH
2803 struct zone *zone,
2804 unsigned long *nr_scanned)
4e416953
BS
2805{
2806 struct scan_control sc = {
b8f5c566 2807 .nr_to_reclaim = SWAP_CLUSTER_MAX,
ee814fe2 2808 .target_mem_cgroup = memcg,
4e416953
BS
2809 .may_writepage = !laptop_mode,
2810 .may_unmap = 1,
2811 .may_swap = !noswap,
4e416953 2812 };
f9be23d6 2813 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
02695175 2814 int swappiness = mem_cgroup_swappiness(memcg);
6b4f7799 2815 unsigned long lru_pages;
0ae5e89c 2816
4e416953
BS
2817 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2818 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
bdce6d9e 2819
9e3b2f8c 2820 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
bdce6d9e
KM
2821 sc.may_writepage,
2822 sc.gfp_mask);
2823
4e416953
BS
2824 /*
2825 * NOTE: Although we can get the priority field, using it
2826 * here is not a good idea, since it limits the pages we can scan.
2827 * if we don't reclaim here, the shrink_zone from balance_pgdat
2828 * will pick up pages from other mem cgroup's as well. We hack
2829 * the priority and make it zero.
2830 */
6b4f7799 2831 shrink_lruvec(lruvec, swappiness, &sc, &lru_pages);
bdce6d9e
KM
2832
2833 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2834
0ae5e89c 2835 *nr_scanned = sc.nr_scanned;
4e416953
BS
2836 return sc.nr_reclaimed;
2837}
2838
72835c86 2839unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
b70a2a21 2840 unsigned long nr_pages,
a7885eb8 2841 gfp_t gfp_mask,
b70a2a21 2842 bool may_swap)
66e1707b 2843{
4e416953 2844 struct zonelist *zonelist;
bdce6d9e 2845 unsigned long nr_reclaimed;
889976db 2846 int nid;
66e1707b 2847 struct scan_control sc = {
b70a2a21 2848 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
a09ed5e0
YH
2849 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2850 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
ee814fe2
JW
2851 .target_mem_cgroup = memcg,
2852 .priority = DEF_PRIORITY,
2853 .may_writepage = !laptop_mode,
2854 .may_unmap = 1,
b70a2a21 2855 .may_swap = may_swap,
a09ed5e0 2856 };
66e1707b 2857
889976db
YH
2858 /*
2859 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2860 * take care of from where we get pages. So the node where we start the
2861 * scan does not need to be the current node.
2862 */
72835c86 2863 nid = mem_cgroup_select_victim_node(memcg);
889976db
YH
2864
2865 zonelist = NODE_DATA(nid)->node_zonelists;
bdce6d9e
KM
2866
2867 trace_mm_vmscan_memcg_reclaim_begin(0,
2868 sc.may_writepage,
2869 sc.gfp_mask);
2870
3115cd91 2871 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
bdce6d9e
KM
2872
2873 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2874
2875 return nr_reclaimed;
66e1707b
BS
2876}
2877#endif
2878
9e3b2f8c 2879static void age_active_anon(struct zone *zone, struct scan_control *sc)
f16015fb 2880{
b95a2f2d 2881 struct mem_cgroup *memcg;
f16015fb 2882
b95a2f2d
JW
2883 if (!total_swap_pages)
2884 return;
2885
2886 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2887 do {
c56d5c7d 2888 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
b95a2f2d 2889
c56d5c7d 2890 if (inactive_anon_is_low(lruvec))
1a93be0e 2891 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
9e3b2f8c 2892 sc, LRU_ACTIVE_ANON);
b95a2f2d
JW
2893
2894 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2895 } while (memcg);
f16015fb
JW
2896}
2897
60cefed4
JW
2898static bool zone_balanced(struct zone *zone, int order,
2899 unsigned long balance_gap, int classzone_idx)
2900{
2901 if (!zone_watermark_ok_safe(zone, order, high_wmark_pages(zone) +
2902 balance_gap, classzone_idx, 0))
2903 return false;
2904
ebff3980
VB
2905 if (IS_ENABLED(CONFIG_COMPACTION) && order && compaction_suitable(zone,
2906 order, 0, classzone_idx) == COMPACT_SKIPPED)
60cefed4
JW
2907 return false;
2908
2909 return true;
2910}
2911
1741c877 2912/*
4ae0a48b
ZC
2913 * pgdat_balanced() is used when checking if a node is balanced.
2914 *
2915 * For order-0, all zones must be balanced!
2916 *
2917 * For high-order allocations only zones that meet watermarks and are in a
2918 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2919 * total of balanced pages must be at least 25% of the zones allowed by
2920 * classzone_idx for the node to be considered balanced. Forcing all zones to
2921 * be balanced for high orders can cause excessive reclaim when there are
2922 * imbalanced zones.
1741c877
MG
2923 * The choice of 25% is due to
2924 * o a 16M DMA zone that is balanced will not balance a zone on any
2925 * reasonable sized machine
2926 * o On all other machines, the top zone must be at least a reasonable
25985edc 2927 * percentage of the middle zones. For example, on 32-bit x86, highmem
1741c877
MG
2928 * would need to be at least 256M for it to be balance a whole node.
2929 * Similarly, on x86-64 the Normal zone would need to be at least 1G
2930 * to balance a node on its own. These seemed like reasonable ratios.
2931 */
4ae0a48b 2932static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
1741c877 2933{
b40da049 2934 unsigned long managed_pages = 0;
4ae0a48b 2935 unsigned long balanced_pages = 0;
1741c877
MG
2936 int i;
2937
4ae0a48b
ZC
2938 /* Check the watermark levels */
2939 for (i = 0; i <= classzone_idx; i++) {
2940 struct zone *zone = pgdat->node_zones + i;
1741c877 2941
4ae0a48b
ZC
2942 if (!populated_zone(zone))
2943 continue;
2944
b40da049 2945 managed_pages += zone->managed_pages;
4ae0a48b
ZC
2946
2947 /*
2948 * A special case here:
2949 *
2950 * balance_pgdat() skips over all_unreclaimable after
2951 * DEF_PRIORITY. Effectively, it considers them balanced so
2952 * they must be considered balanced here as well!
2953 */
6e543d57 2954 if (!zone_reclaimable(zone)) {
b40da049 2955 balanced_pages += zone->managed_pages;
4ae0a48b
ZC
2956 continue;
2957 }
2958
2959 if (zone_balanced(zone, order, 0, i))
b40da049 2960 balanced_pages += zone->managed_pages;
4ae0a48b
ZC
2961 else if (!order)
2962 return false;
2963 }
2964
2965 if (order)
b40da049 2966 return balanced_pages >= (managed_pages >> 2);
4ae0a48b
ZC
2967 else
2968 return true;
1741c877
MG
2969}
2970
5515061d
MG
2971/*
2972 * Prepare kswapd for sleeping. This verifies that there are no processes
2973 * waiting in throttle_direct_reclaim() and that watermarks have been met.
2974 *
2975 * Returns true if kswapd is ready to sleep
2976 */
2977static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
dc83edd9 2978 int classzone_idx)
f50de2d3 2979{
f50de2d3
MG
2980 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
2981 if (remaining)
5515061d
MG
2982 return false;
2983
2984 /*
9e5e3661
VB
2985 * The throttled processes are normally woken up in balance_pgdat() as
2986 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
2987 * race between when kswapd checks the watermarks and a process gets
2988 * throttled. There is also a potential race if processes get
2989 * throttled, kswapd wakes, a large process exits thereby balancing the
2990 * zones, which causes kswapd to exit balance_pgdat() before reaching
2991 * the wake up checks. If kswapd is going to sleep, no process should
2992 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
2993 * the wake up is premature, processes will wake kswapd and get
2994 * throttled again. The difference from wake ups in balance_pgdat() is
2995 * that here we are under prepare_to_wait().
5515061d 2996 */
9e5e3661
VB
2997 if (waitqueue_active(&pgdat->pfmemalloc_wait))
2998 wake_up_all(&pgdat->pfmemalloc_wait);
f50de2d3 2999
4ae0a48b 3000 return pgdat_balanced(pgdat, order, classzone_idx);
f50de2d3
MG
3001}
3002
75485363
MG
3003/*
3004 * kswapd shrinks the zone by the number of pages required to reach
3005 * the high watermark.
b8e83b94
MG
3006 *
3007 * Returns true if kswapd scanned at least the requested number of pages to
283aba9f
MG
3008 * reclaim or if the lack of progress was due to pages under writeback.
3009 * This is used to determine if the scanning priority needs to be raised.
75485363 3010 */
b8e83b94 3011static bool kswapd_shrink_zone(struct zone *zone,
7c954f6d 3012 int classzone_idx,
75485363 3013 struct scan_control *sc,
2ab44f43 3014 unsigned long *nr_attempted)
75485363 3015{
7c954f6d
MG
3016 int testorder = sc->order;
3017 unsigned long balance_gap;
7c954f6d 3018 bool lowmem_pressure;
75485363
MG
3019
3020 /* Reclaim above the high watermark. */
3021 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
7c954f6d
MG
3022
3023 /*
3024 * Kswapd reclaims only single pages with compaction enabled. Trying
3025 * too hard to reclaim until contiguous free pages have become
3026 * available can hurt performance by evicting too much useful data
3027 * from memory. Do not reclaim more than needed for compaction.
3028 */
3029 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
ebff3980
VB
3030 compaction_suitable(zone, sc->order, 0, classzone_idx)
3031 != COMPACT_SKIPPED)
7c954f6d
MG
3032 testorder = 0;
3033
3034 /*
3035 * We put equal pressure on every zone, unless one zone has way too
3036 * many pages free already. The "too many pages" is defined as the
3037 * high wmark plus a "gap" where the gap is either the low
3038 * watermark or 1% of the zone, whichever is smaller.
3039 */
4be89a34
JZ
3040 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3041 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
7c954f6d
MG
3042
3043 /*
3044 * If there is no low memory pressure or the zone is balanced then no
3045 * reclaim is necessary
3046 */
3047 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
3048 if (!lowmem_pressure && zone_balanced(zone, testorder,
3049 balance_gap, classzone_idx))
3050 return true;
3051
6b4f7799 3052 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
75485363 3053
2ab44f43
MG
3054 /* Account for the number of pages attempted to reclaim */
3055 *nr_attempted += sc->nr_to_reclaim;
3056
57054651 3057 clear_bit(ZONE_WRITEBACK, &zone->flags);
283aba9f 3058
7c954f6d
MG
3059 /*
3060 * If a zone reaches its high watermark, consider it to be no longer
3061 * congested. It's possible there are dirty pages backed by congested
3062 * BDIs but as pressure is relieved, speculatively avoid congestion
3063 * waits.
3064 */
6e543d57 3065 if (zone_reclaimable(zone) &&
7c954f6d 3066 zone_balanced(zone, testorder, 0, classzone_idx)) {
57054651
JW
3067 clear_bit(ZONE_CONGESTED, &zone->flags);
3068 clear_bit(ZONE_DIRTY, &zone->flags);
7c954f6d
MG
3069 }
3070
b8e83b94 3071 return sc->nr_scanned >= sc->nr_to_reclaim;
75485363
MG
3072}
3073
1da177e4
LT
3074/*
3075 * For kswapd, balance_pgdat() will work across all this node's zones until
41858966 3076 * they are all at high_wmark_pages(zone).
1da177e4 3077 *
0abdee2b 3078 * Returns the final order kswapd was reclaiming at
1da177e4
LT
3079 *
3080 * There is special handling here for zones which are full of pinned pages.
3081 * This can happen if the pages are all mlocked, or if they are all used by
3082 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3083 * What we do is to detect the case where all pages in the zone have been
3084 * scanned twice and there has been zero successful reclaim. Mark the zone as
3085 * dead and from now on, only perform a short scan. Basically we're polling
3086 * the zone for when the problem goes away.
3087 *
3088 * kswapd scans the zones in the highmem->normal->dma direction. It skips
41858966
MG
3089 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3090 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3091 * lower zones regardless of the number of free pages in the lower zones. This
3092 * interoperates with the page allocator fallback scheme to ensure that aging
3093 * of pages is balanced across the zones.
1da177e4 3094 */
99504748 3095static unsigned long balance_pgdat(pg_data_t *pgdat, int order,
dc83edd9 3096 int *classzone_idx)
1da177e4 3097{
1da177e4 3098 int i;
99504748 3099 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
0608f43d
AM
3100 unsigned long nr_soft_reclaimed;
3101 unsigned long nr_soft_scanned;
179e9639
AM
3102 struct scan_control sc = {
3103 .gfp_mask = GFP_KERNEL,
ee814fe2 3104 .order = order,
b8e83b94 3105 .priority = DEF_PRIORITY,
ee814fe2 3106 .may_writepage = !laptop_mode,
a6dc60f8 3107 .may_unmap = 1,
2e2e4259 3108 .may_swap = 1,
179e9639 3109 };
f8891e5e 3110 count_vm_event(PAGEOUTRUN);
1da177e4 3111
9e3b2f8c 3112 do {
2ab44f43 3113 unsigned long nr_attempted = 0;
b8e83b94 3114 bool raise_priority = true;
2ab44f43 3115 bool pgdat_needs_compaction = (order > 0);
b8e83b94
MG
3116
3117 sc.nr_reclaimed = 0;
1da177e4 3118
d6277db4
RW
3119 /*
3120 * Scan in the highmem->dma direction for the highest
3121 * zone which needs scanning
3122 */
3123 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3124 struct zone *zone = pgdat->node_zones + i;
1da177e4 3125
d6277db4
RW
3126 if (!populated_zone(zone))
3127 continue;
1da177e4 3128
6e543d57
LD
3129 if (sc.priority != DEF_PRIORITY &&
3130 !zone_reclaimable(zone))
d6277db4 3131 continue;
1da177e4 3132
556adecb
RR
3133 /*
3134 * Do some background aging of the anon list, to give
3135 * pages a chance to be referenced before reclaiming.
3136 */
9e3b2f8c 3137 age_active_anon(zone, &sc);
556adecb 3138
cc715d99
MG
3139 /*
3140 * If the number of buffer_heads in the machine
3141 * exceeds the maximum allowed level and this node
3142 * has a highmem zone, force kswapd to reclaim from
3143 * it to relieve lowmem pressure.
3144 */
3145 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3146 end_zone = i;
3147 break;
3148 }
3149
60cefed4 3150 if (!zone_balanced(zone, order, 0, 0)) {
d6277db4 3151 end_zone = i;
e1dbeda6 3152 break;
439423f6 3153 } else {
d43006d5
MG
3154 /*
3155 * If balanced, clear the dirty and congested
3156 * flags
3157 */
57054651
JW
3158 clear_bit(ZONE_CONGESTED, &zone->flags);
3159 clear_bit(ZONE_DIRTY, &zone->flags);
1da177e4 3160 }
1da177e4 3161 }
dafcb73e 3162
b8e83b94 3163 if (i < 0)
e1dbeda6
AM
3164 goto out;
3165
1da177e4
LT
3166 for (i = 0; i <= end_zone; i++) {
3167 struct zone *zone = pgdat->node_zones + i;
3168
2ab44f43
MG
3169 if (!populated_zone(zone))
3170 continue;
3171
2ab44f43
MG
3172 /*
3173 * If any zone is currently balanced then kswapd will
3174 * not call compaction as it is expected that the
3175 * necessary pages are already available.
3176 */
3177 if (pgdat_needs_compaction &&
3178 zone_watermark_ok(zone, order,
3179 low_wmark_pages(zone),
3180 *classzone_idx, 0))
3181 pgdat_needs_compaction = false;
1da177e4
LT
3182 }
3183
b7ea3c41
MG
3184 /*
3185 * If we're getting trouble reclaiming, start doing writepage
3186 * even in laptop mode.
3187 */
3188 if (sc.priority < DEF_PRIORITY - 2)
3189 sc.may_writepage = 1;
3190
1da177e4
LT
3191 /*
3192 * Now scan the zone in the dma->highmem direction, stopping
3193 * at the last zone which needs scanning.
3194 *
3195 * We do this because the page allocator works in the opposite
3196 * direction. This prevents the page allocator from allocating
3197 * pages behind kswapd's direction of progress, which would
3198 * cause too much scanning of the lower zones.
3199 */
3200 for (i = 0; i <= end_zone; i++) {
3201 struct zone *zone = pgdat->node_zones + i;
3202
f3fe6512 3203 if (!populated_zone(zone))
1da177e4
LT
3204 continue;
3205
6e543d57
LD
3206 if (sc.priority != DEF_PRIORITY &&
3207 !zone_reclaimable(zone))
1da177e4
LT
3208 continue;
3209
1da177e4 3210 sc.nr_scanned = 0;
4e416953 3211
0608f43d
AM
3212 nr_soft_scanned = 0;
3213 /*
3214 * Call soft limit reclaim before calling shrink_zone.
3215 */
3216 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3217 order, sc.gfp_mask,
3218 &nr_soft_scanned);
3219 sc.nr_reclaimed += nr_soft_reclaimed;
3220
32a4330d 3221 /*
7c954f6d
MG
3222 * There should be no need to raise the scanning
3223 * priority if enough pages are already being scanned
3224 * that that high watermark would be met at 100%
3225 * efficiency.
fe2c2a10 3226 */
6b4f7799
JW
3227 if (kswapd_shrink_zone(zone, end_zone,
3228 &sc, &nr_attempted))
7c954f6d 3229 raise_priority = false;
1da177e4 3230 }
5515061d
MG
3231
3232 /*
3233 * If the low watermark is met there is no need for processes
3234 * to be throttled on pfmemalloc_wait as they should not be
3235 * able to safely make forward progress. Wake them
3236 */
3237 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3238 pfmemalloc_watermark_ok(pgdat))
cfc51155 3239 wake_up_all(&pgdat->pfmemalloc_wait);
5515061d 3240
1da177e4 3241 /*
b8e83b94
MG
3242 * Fragmentation may mean that the system cannot be rebalanced
3243 * for high-order allocations in all zones. If twice the
3244 * allocation size has been reclaimed and the zones are still
3245 * not balanced then recheck the watermarks at order-0 to
3246 * prevent kswapd reclaiming excessively. Assume that a
3247 * process requested a high-order can direct reclaim/compact.
1da177e4 3248 */
b8e83b94
MG
3249 if (order && sc.nr_reclaimed >= 2UL << order)
3250 order = sc.order = 0;
8357376d 3251
b8e83b94
MG
3252 /* Check if kswapd should be suspending */
3253 if (try_to_freeze() || kthread_should_stop())
3254 break;
8357376d 3255
2ab44f43
MG
3256 /*
3257 * Compact if necessary and kswapd is reclaiming at least the
3258 * high watermark number of pages as requsted
3259 */
3260 if (pgdat_needs_compaction && sc.nr_reclaimed > nr_attempted)
3261 compact_pgdat(pgdat, order);
3262
73ce02e9 3263 /*
b8e83b94
MG
3264 * Raise priority if scanning rate is too low or there was no
3265 * progress in reclaiming pages
73ce02e9 3266 */
b8e83b94
MG
3267 if (raise_priority || !sc.nr_reclaimed)
3268 sc.priority--;
9aa41348 3269 } while (sc.priority >= 1 &&
b8e83b94 3270 !pgdat_balanced(pgdat, order, *classzone_idx));
1da177e4 3271
b8e83b94 3272out:
0abdee2b 3273 /*
5515061d 3274 * Return the order we were reclaiming at so prepare_kswapd_sleep()
0abdee2b
MG
3275 * makes a decision on the order we were last reclaiming at. However,
3276 * if another caller entered the allocator slow path while kswapd
3277 * was awake, order will remain at the higher level
3278 */
dc83edd9 3279 *classzone_idx = end_zone;
0abdee2b 3280 return order;
1da177e4
LT
3281}
3282
dc83edd9 3283static void kswapd_try_to_sleep(pg_data_t *pgdat, int order, int classzone_idx)
f0bc0a60
KM
3284{
3285 long remaining = 0;
3286 DEFINE_WAIT(wait);
3287
3288 if (freezing(current) || kthread_should_stop())
3289 return;
3290
3291 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3292
3293 /* Try to sleep for a short interval */
5515061d 3294 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
f0bc0a60
KM
3295 remaining = schedule_timeout(HZ/10);
3296 finish_wait(&pgdat->kswapd_wait, &wait);
3297 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3298 }
3299
3300 /*
3301 * After a short sleep, check if it was a premature sleep. If not, then
3302 * go fully to sleep until explicitly woken up.
3303 */
5515061d 3304 if (prepare_kswapd_sleep(pgdat, order, remaining, classzone_idx)) {
f0bc0a60
KM
3305 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3306
3307 /*
3308 * vmstat counters are not perfectly accurate and the estimated
3309 * value for counters such as NR_FREE_PAGES can deviate from the
3310 * true value by nr_online_cpus * threshold. To avoid the zone
3311 * watermarks being breached while under pressure, we reduce the
3312 * per-cpu vmstat threshold while kswapd is awake and restore
3313 * them before going back to sleep.
3314 */
3315 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
1c7e7f6c 3316
62997027
MG
3317 /*
3318 * Compaction records what page blocks it recently failed to
3319 * isolate pages from and skips them in the future scanning.
3320 * When kswapd is going to sleep, it is reasonable to assume
3321 * that pages and compaction may succeed so reset the cache.
3322 */
3323 reset_isolation_suitable(pgdat);
3324
1c7e7f6c
AK
3325 if (!kthread_should_stop())
3326 schedule();
3327
f0bc0a60
KM
3328 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3329 } else {
3330 if (remaining)
3331 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3332 else
3333 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3334 }
3335 finish_wait(&pgdat->kswapd_wait, &wait);
3336}
3337
1da177e4
LT
3338/*
3339 * The background pageout daemon, started as a kernel thread
4f98a2fe 3340 * from the init process.
1da177e4
LT
3341 *
3342 * This basically trickles out pages so that we have _some_
3343 * free memory available even if there is no other activity
3344 * that frees anything up. This is needed for things like routing
3345 * etc, where we otherwise might have all activity going on in
3346 * asynchronous contexts that cannot page things out.
3347 *
3348 * If there are applications that are active memory-allocators
3349 * (most normal use), this basically shouldn't matter.
3350 */
3351static int kswapd(void *p)
3352{
215ddd66 3353 unsigned long order, new_order;
d2ebd0f6 3354 unsigned balanced_order;
215ddd66 3355 int classzone_idx, new_classzone_idx;
d2ebd0f6 3356 int balanced_classzone_idx;
1da177e4
LT
3357 pg_data_t *pgdat = (pg_data_t*)p;
3358 struct task_struct *tsk = current;
f0bc0a60 3359
1da177e4
LT
3360 struct reclaim_state reclaim_state = {
3361 .reclaimed_slab = 0,
3362 };
a70f7302 3363 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1da177e4 3364
cf40bd16
NP
3365 lockdep_set_current_reclaim_state(GFP_KERNEL);
3366
174596a0 3367 if (!cpumask_empty(cpumask))
c5f59f08 3368 set_cpus_allowed_ptr(tsk, cpumask);
1da177e4
LT
3369 current->reclaim_state = &reclaim_state;
3370
3371 /*
3372 * Tell the memory management that we're a "memory allocator",
3373 * and that if we need more memory we should get access to it
3374 * regardless (see "__alloc_pages()"). "kswapd" should
3375 * never get caught in the normal page freeing logic.
3376 *
3377 * (Kswapd normally doesn't need memory anyway, but sometimes
3378 * you need a small amount of memory in order to be able to
3379 * page out something else, and this flag essentially protects
3380 * us from recursively trying to free more memory as we're
3381 * trying to free the first piece of memory in the first place).
3382 */
930d9152 3383 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
83144186 3384 set_freezable();
1da177e4 3385
215ddd66 3386 order = new_order = 0;
d2ebd0f6 3387 balanced_order = 0;
215ddd66 3388 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
d2ebd0f6 3389 balanced_classzone_idx = classzone_idx;
1da177e4 3390 for ( ; ; ) {
6f6313d4 3391 bool ret;
3e1d1d28 3392
215ddd66
MG
3393 /*
3394 * If the last balance_pgdat was unsuccessful it's unlikely a
3395 * new request of a similar or harder type will succeed soon
3396 * so consider going to sleep on the basis we reclaimed at
3397 */
d2ebd0f6
AS
3398 if (balanced_classzone_idx >= new_classzone_idx &&
3399 balanced_order == new_order) {
215ddd66
MG
3400 new_order = pgdat->kswapd_max_order;
3401 new_classzone_idx = pgdat->classzone_idx;
3402 pgdat->kswapd_max_order = 0;
3403 pgdat->classzone_idx = pgdat->nr_zones - 1;
3404 }
3405
99504748 3406 if (order < new_order || classzone_idx > new_classzone_idx) {
1da177e4
LT
3407 /*
3408 * Don't sleep if someone wants a larger 'order'
99504748 3409 * allocation or has tigher zone constraints
1da177e4
LT
3410 */
3411 order = new_order;
99504748 3412 classzone_idx = new_classzone_idx;
1da177e4 3413 } else {
d2ebd0f6
AS
3414 kswapd_try_to_sleep(pgdat, balanced_order,
3415 balanced_classzone_idx);
1da177e4 3416 order = pgdat->kswapd_max_order;
99504748 3417 classzone_idx = pgdat->classzone_idx;
f0dfcde0
AS
3418 new_order = order;
3419 new_classzone_idx = classzone_idx;
4d40502e 3420 pgdat->kswapd_max_order = 0;
215ddd66 3421 pgdat->classzone_idx = pgdat->nr_zones - 1;
1da177e4 3422 }
1da177e4 3423
8fe23e05
DR
3424 ret = try_to_freeze();
3425 if (kthread_should_stop())
3426 break;
3427
3428 /*
3429 * We can speed up thawing tasks if we don't call balance_pgdat
3430 * after returning from the refrigerator
3431 */
33906bc5
MG
3432 if (!ret) {
3433 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
d2ebd0f6
AS
3434 balanced_classzone_idx = classzone_idx;
3435 balanced_order = balance_pgdat(pgdat, order,
3436 &balanced_classzone_idx);
33906bc5 3437 }
1da177e4 3438 }
b0a8cc58 3439
71abdc15 3440 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
b0a8cc58 3441 current->reclaim_state = NULL;
71abdc15
JW
3442 lockdep_clear_current_reclaim_state();
3443
1da177e4
LT
3444 return 0;
3445}
3446
3447/*
3448 * A zone is low on free memory, so wake its kswapd task to service it.
3449 */
99504748 3450void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
1da177e4
LT
3451{
3452 pg_data_t *pgdat;
3453
f3fe6512 3454 if (!populated_zone(zone))
1da177e4
LT
3455 return;
3456
344736f2 3457 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
1da177e4 3458 return;
88f5acf8 3459 pgdat = zone->zone_pgdat;
99504748 3460 if (pgdat->kswapd_max_order < order) {
1da177e4 3461 pgdat->kswapd_max_order = order;
99504748
MG
3462 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3463 }
8d0986e2 3464 if (!waitqueue_active(&pgdat->kswapd_wait))
1da177e4 3465 return;
892f795d 3466 if (zone_balanced(zone, order, 0, 0))
88f5acf8
MG
3467 return;
3468
3469 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
8d0986e2 3470 wake_up_interruptible(&pgdat->kswapd_wait);
1da177e4
LT
3471}
3472
c6f37f12 3473#ifdef CONFIG_HIBERNATION
1da177e4 3474/*
7b51755c 3475 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
d6277db4
RW
3476 * freed pages.
3477 *
3478 * Rather than trying to age LRUs the aim is to preserve the overall
3479 * LRU order by reclaiming preferentially
3480 * inactive > active > active referenced > active mapped
1da177e4 3481 */
7b51755c 3482unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
1da177e4 3483{
d6277db4 3484 struct reclaim_state reclaim_state;
d6277db4 3485 struct scan_control sc = {
ee814fe2 3486 .nr_to_reclaim = nr_to_reclaim,
7b51755c 3487 .gfp_mask = GFP_HIGHUSER_MOVABLE,
ee814fe2 3488 .priority = DEF_PRIORITY,
d6277db4 3489 .may_writepage = 1,
ee814fe2
JW
3490 .may_unmap = 1,
3491 .may_swap = 1,
7b51755c 3492 .hibernation_mode = 1,
1da177e4 3493 };
a09ed5e0 3494 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7b51755c
KM
3495 struct task_struct *p = current;
3496 unsigned long nr_reclaimed;
1da177e4 3497
7b51755c
KM
3498 p->flags |= PF_MEMALLOC;
3499 lockdep_set_current_reclaim_state(sc.gfp_mask);
3500 reclaim_state.reclaimed_slab = 0;
3501 p->reclaim_state = &reclaim_state;
d6277db4 3502
3115cd91 3503 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
d979677c 3504
7b51755c
KM
3505 p->reclaim_state = NULL;
3506 lockdep_clear_current_reclaim_state();
3507 p->flags &= ~PF_MEMALLOC;
d6277db4 3508
7b51755c 3509 return nr_reclaimed;
1da177e4 3510}
c6f37f12 3511#endif /* CONFIG_HIBERNATION */
1da177e4 3512
1da177e4
LT
3513/* It's optimal to keep kswapds on the same CPUs as their memory, but
3514 not required for correctness. So if the last cpu in a node goes
3515 away, we get changed to run anywhere: as the first one comes back,
3516 restore their cpu bindings. */
fcb35a9b
GKH
3517static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3518 void *hcpu)
1da177e4 3519{
58c0a4a7 3520 int nid;
1da177e4 3521
8bb78442 3522 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
48fb2e24 3523 for_each_node_state(nid, N_MEMORY) {
c5f59f08 3524 pg_data_t *pgdat = NODE_DATA(nid);
a70f7302
RR
3525 const struct cpumask *mask;
3526
3527 mask = cpumask_of_node(pgdat->node_id);
c5f59f08 3528
3e597945 3529 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
1da177e4 3530 /* One of our CPUs online: restore mask */
c5f59f08 3531 set_cpus_allowed_ptr(pgdat->kswapd, mask);
1da177e4
LT
3532 }
3533 }
3534 return NOTIFY_OK;
3535}
1da177e4 3536
3218ae14
YG
3537/*
3538 * This kswapd start function will be called by init and node-hot-add.
3539 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3540 */
3541int kswapd_run(int nid)
3542{
3543 pg_data_t *pgdat = NODE_DATA(nid);
3544 int ret = 0;
3545
3546 if (pgdat->kswapd)
3547 return 0;
3548
3549 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3550 if (IS_ERR(pgdat->kswapd)) {
3551 /* failure at boot is fatal */
3552 BUG_ON(system_state == SYSTEM_BOOTING);
d5dc0ad9
GS
3553 pr_err("Failed to start kswapd on node %d\n", nid);
3554 ret = PTR_ERR(pgdat->kswapd);
d72515b8 3555 pgdat->kswapd = NULL;
3218ae14
YG
3556 }
3557 return ret;
3558}
3559
8fe23e05 3560/*
d8adde17 3561 * Called by memory hotplug when all memory in a node is offlined. Caller must
bfc8c901 3562 * hold mem_hotplug_begin/end().
8fe23e05
DR
3563 */
3564void kswapd_stop(int nid)
3565{
3566 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3567
d8adde17 3568 if (kswapd) {
8fe23e05 3569 kthread_stop(kswapd);
d8adde17
JL
3570 NODE_DATA(nid)->kswapd = NULL;
3571 }
8fe23e05
DR
3572}
3573
1da177e4
LT
3574static int __init kswapd_init(void)
3575{
3218ae14 3576 int nid;
69e05944 3577
1da177e4 3578 swap_setup();
48fb2e24 3579 for_each_node_state(nid, N_MEMORY)
3218ae14 3580 kswapd_run(nid);
1da177e4
LT
3581 hotcpu_notifier(cpu_callback, 0);
3582 return 0;
3583}
3584
3585module_init(kswapd_init)
9eeff239
CL
3586
3587#ifdef CONFIG_NUMA
3588/*
3589 * Zone reclaim mode
3590 *
3591 * If non-zero call zone_reclaim when the number of free pages falls below
3592 * the watermarks.
9eeff239
CL
3593 */
3594int zone_reclaim_mode __read_mostly;
3595
1b2ffb78 3596#define RECLAIM_OFF 0
7d03431c 3597#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
1b2ffb78
CL
3598#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
3599#define RECLAIM_SWAP (1<<2) /* Swap pages out during reclaim */
3600
a92f7126
CL
3601/*
3602 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3603 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3604 * a zone.
3605 */
3606#define ZONE_RECLAIM_PRIORITY 4
3607
9614634f
CL
3608/*
3609 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3610 * occur.
3611 */
3612int sysctl_min_unmapped_ratio = 1;
3613
0ff38490
CL
3614/*
3615 * If the number of slab pages in a zone grows beyond this percentage then
3616 * slab reclaim needs to occur.
3617 */
3618int sysctl_min_slab_ratio = 5;
3619
90afa5de
MG
3620static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3621{
3622 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3623 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3624 zone_page_state(zone, NR_ACTIVE_FILE);
3625
3626 /*
3627 * It's possible for there to be more file mapped pages than
3628 * accounted for by the pages on the file LRU lists because
3629 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3630 */
3631 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3632}
3633
3634/* Work out how many page cache pages we can reclaim in this reclaim_mode */
3635static long zone_pagecache_reclaimable(struct zone *zone)
3636{
3637 long nr_pagecache_reclaimable;
3638 long delta = 0;
3639
3640 /*
3641 * If RECLAIM_SWAP is set, then all file pages are considered
3642 * potentially reclaimable. Otherwise, we have to worry about
3643 * pages like swapcache and zone_unmapped_file_pages() provides
3644 * a better estimate
3645 */
3646 if (zone_reclaim_mode & RECLAIM_SWAP)
3647 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3648 else
3649 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3650
3651 /* If we can't clean pages, remove dirty pages from consideration */
3652 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3653 delta += zone_page_state(zone, NR_FILE_DIRTY);
3654
3655 /* Watch for any possible underflows due to delta */
3656 if (unlikely(delta > nr_pagecache_reclaimable))
3657 delta = nr_pagecache_reclaimable;
3658
3659 return nr_pagecache_reclaimable - delta;
3660}
3661
9eeff239
CL
3662/*
3663 * Try to free up some pages from this zone through reclaim.
3664 */
179e9639 3665static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
9eeff239 3666{
7fb2d46d 3667 /* Minimum pages needed in order to stay on node */
69e05944 3668 const unsigned long nr_pages = 1 << order;
9eeff239
CL
3669 struct task_struct *p = current;
3670 struct reclaim_state reclaim_state;
179e9639 3671 struct scan_control sc = {
62b726c1 3672 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
21caf2fc 3673 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
bd2f6199 3674 .order = order,
9e3b2f8c 3675 .priority = ZONE_RECLAIM_PRIORITY,
ee814fe2
JW
3676 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
3677 .may_unmap = !!(zone_reclaim_mode & RECLAIM_SWAP),
3678 .may_swap = 1,
179e9639 3679 };
9eeff239 3680
9eeff239 3681 cond_resched();
d4f7796e
CL
3682 /*
3683 * We need to be able to allocate from the reserves for RECLAIM_SWAP
3684 * and we also need to be able to write out pages for RECLAIM_WRITE
3685 * and RECLAIM_SWAP.
3686 */
3687 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
76ca542d 3688 lockdep_set_current_reclaim_state(gfp_mask);
9eeff239
CL
3689 reclaim_state.reclaimed_slab = 0;
3690 p->reclaim_state = &reclaim_state;
c84db23c 3691
90afa5de 3692 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
0ff38490
CL
3693 /*
3694 * Free memory by calling shrink zone with increasing
3695 * priorities until we have enough memory freed.
3696 */
0ff38490 3697 do {
6b4f7799 3698 shrink_zone(zone, &sc, true);
9e3b2f8c 3699 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
0ff38490 3700 }
c84db23c 3701
9eeff239 3702 p->reclaim_state = NULL;
d4f7796e 3703 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
76ca542d 3704 lockdep_clear_current_reclaim_state();
a79311c1 3705 return sc.nr_reclaimed >= nr_pages;
9eeff239 3706}
179e9639
AM
3707
3708int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3709{
179e9639 3710 int node_id;
d773ed6b 3711 int ret;
179e9639
AM
3712
3713 /*
0ff38490
CL
3714 * Zone reclaim reclaims unmapped file backed pages and
3715 * slab pages if we are over the defined limits.
34aa1330 3716 *
9614634f
CL
3717 * A small portion of unmapped file backed pages is needed for
3718 * file I/O otherwise pages read by file I/O will be immediately
3719 * thrown out if the zone is overallocated. So we do not reclaim
3720 * if less than a specified percentage of the zone is used by
3721 * unmapped file backed pages.
179e9639 3722 */
90afa5de
MG
3723 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3724 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
fa5e084e 3725 return ZONE_RECLAIM_FULL;
179e9639 3726
6e543d57 3727 if (!zone_reclaimable(zone))
fa5e084e 3728 return ZONE_RECLAIM_FULL;
d773ed6b 3729
179e9639 3730 /*
d773ed6b 3731 * Do not scan if the allocation should not be delayed.
179e9639 3732 */
d773ed6b 3733 if (!(gfp_mask & __GFP_WAIT) || (current->flags & PF_MEMALLOC))
fa5e084e 3734 return ZONE_RECLAIM_NOSCAN;
179e9639
AM
3735
3736 /*
3737 * Only run zone reclaim on the local zone or on zones that do not
3738 * have associated processors. This will favor the local processor
3739 * over remote processors and spread off node memory allocations
3740 * as wide as possible.
3741 */
89fa3024 3742 node_id = zone_to_nid(zone);
37c0708d 3743 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
fa5e084e 3744 return ZONE_RECLAIM_NOSCAN;
d773ed6b 3745
57054651 3746 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
fa5e084e
MG
3747 return ZONE_RECLAIM_NOSCAN;
3748
d773ed6b 3749 ret = __zone_reclaim(zone, gfp_mask, order);
57054651 3750 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
d773ed6b 3751
24cf7251
MG
3752 if (!ret)
3753 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3754
d773ed6b 3755 return ret;
179e9639 3756}
9eeff239 3757#endif
894bc310 3758
894bc310
LS
3759/*
3760 * page_evictable - test whether a page is evictable
3761 * @page: the page to test
894bc310
LS
3762 *
3763 * Test whether page is evictable--i.e., should be placed on active/inactive
39b5f29a 3764 * lists vs unevictable list.
894bc310
LS
3765 *
3766 * Reasons page might not be evictable:
ba9ddf49 3767 * (1) page's mapping marked unevictable
b291f000 3768 * (2) page is part of an mlocked VMA
ba9ddf49 3769 *
894bc310 3770 */
39b5f29a 3771int page_evictable(struct page *page)
894bc310 3772{
39b5f29a 3773 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
894bc310 3774}
89e004ea 3775
85046579 3776#ifdef CONFIG_SHMEM
89e004ea 3777/**
24513264
HD
3778 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3779 * @pages: array of pages to check
3780 * @nr_pages: number of pages to check
89e004ea 3781 *
24513264 3782 * Checks pages for evictability and moves them to the appropriate lru list.
85046579
HD
3783 *
3784 * This function is only used for SysV IPC SHM_UNLOCK.
89e004ea 3785 */
24513264 3786void check_move_unevictable_pages(struct page **pages, int nr_pages)
89e004ea 3787{
925b7673 3788 struct lruvec *lruvec;
24513264
HD
3789 struct zone *zone = NULL;
3790 int pgscanned = 0;
3791 int pgrescued = 0;
3792 int i;
89e004ea 3793
24513264
HD
3794 for (i = 0; i < nr_pages; i++) {
3795 struct page *page = pages[i];
3796 struct zone *pagezone;
89e004ea 3797
24513264
HD
3798 pgscanned++;
3799 pagezone = page_zone(page);
3800 if (pagezone != zone) {
3801 if (zone)
3802 spin_unlock_irq(&zone->lru_lock);
3803 zone = pagezone;
3804 spin_lock_irq(&zone->lru_lock);
3805 }
fa9add64 3806 lruvec = mem_cgroup_page_lruvec(page, zone);
89e004ea 3807
24513264
HD
3808 if (!PageLRU(page) || !PageUnevictable(page))
3809 continue;
89e004ea 3810
39b5f29a 3811 if (page_evictable(page)) {
24513264
HD
3812 enum lru_list lru = page_lru_base_type(page);
3813
309381fe 3814 VM_BUG_ON_PAGE(PageActive(page), page);
24513264 3815 ClearPageUnevictable(page);
fa9add64
HD
3816 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3817 add_page_to_lru_list(page, lruvec, lru);
24513264 3818 pgrescued++;
89e004ea 3819 }
24513264 3820 }
89e004ea 3821
24513264
HD
3822 if (zone) {
3823 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3824 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3825 spin_unlock_irq(&zone->lru_lock);
89e004ea 3826 }
89e004ea 3827}
85046579 3828#endif /* CONFIG_SHMEM */