]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blame - mm/vmscan.c
mm: filemap: only do access activations on reads
[mirror_ubuntu-artful-kernel.git] / mm / vmscan.c
CommitLineData
1da177e4
LT
1/*
2 * linux/mm/vmscan.c
3 *
4 * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds
5 *
6 * Swap reorganised 29.12.95, Stephen Tweedie.
7 * kswapd added: 7.1.96 sct
8 * Removed kswapd_ctl limits, and swap out as many pages as needed
9 * to bring the system back to freepages.high: 2.4.97, Rik van Riel.
10 * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com).
11 * Multiqueue VM started 5.8.00, Rik van Riel.
12 */
13
b1de0d13
MH
14#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15
1da177e4
LT
16#include <linux/mm.h>
17#include <linux/module.h>
5a0e3ad6 18#include <linux/gfp.h>
1da177e4
LT
19#include <linux/kernel_stat.h>
20#include <linux/swap.h>
21#include <linux/pagemap.h>
22#include <linux/init.h>
23#include <linux/highmem.h>
70ddf637 24#include <linux/vmpressure.h>
e129b5c2 25#include <linux/vmstat.h>
1da177e4
LT
26#include <linux/file.h>
27#include <linux/writeback.h>
28#include <linux/blkdev.h>
29#include <linux/buffer_head.h> /* for try_to_release_page(),
30 buffer_heads_over_limit */
31#include <linux/mm_inline.h>
1da177e4
LT
32#include <linux/backing-dev.h>
33#include <linux/rmap.h>
34#include <linux/topology.h>
35#include <linux/cpu.h>
36#include <linux/cpuset.h>
3e7d3449 37#include <linux/compaction.h>
1da177e4
LT
38#include <linux/notifier.h>
39#include <linux/rwsem.h>
248a0301 40#include <linux/delay.h>
3218ae14 41#include <linux/kthread.h>
7dfb7103 42#include <linux/freezer.h>
66e1707b 43#include <linux/memcontrol.h>
873b4771 44#include <linux/delayacct.h>
af936a16 45#include <linux/sysctl.h>
929bea7c 46#include <linux/oom.h>
268bb0ce 47#include <linux/prefetch.h>
b1de0d13 48#include <linux/printk.h>
f9fe48be 49#include <linux/dax.h>
1da177e4
LT
50
51#include <asm/tlbflush.h>
52#include <asm/div64.h>
53
54#include <linux/swapops.h>
117aad1e 55#include <linux/balloon_compaction.h>
1da177e4 56
0f8053a5
NP
57#include "internal.h"
58
33906bc5
MG
59#define CREATE_TRACE_POINTS
60#include <trace/events/vmscan.h>
61
1da177e4 62struct scan_control {
22fba335
KM
63 /* How many pages shrink_list() should reclaim */
64 unsigned long nr_to_reclaim;
65
1da177e4 66 /* This context's GFP mask */
6daa0e28 67 gfp_t gfp_mask;
1da177e4 68
ee814fe2 69 /* Allocation order */
5ad333eb 70 int order;
66e1707b 71
ee814fe2
JW
72 /*
73 * Nodemask of nodes allowed by the caller. If NULL, all nodes
74 * are scanned.
75 */
76 nodemask_t *nodemask;
9e3b2f8c 77
f16015fb
JW
78 /*
79 * The memory cgroup that hit its limit and as a result is the
80 * primary target of this reclaim invocation.
81 */
82 struct mem_cgroup *target_mem_cgroup;
66e1707b 83
ee814fe2
JW
84 /* Scan (total_size >> priority) pages at once */
85 int priority;
86
87 unsigned int may_writepage:1;
88
89 /* Can mapped pages be reclaimed? */
90 unsigned int may_unmap:1;
91
92 /* Can pages be swapped as part of reclaim? */
93 unsigned int may_swap:1;
94
241994ed
JW
95 /* Can cgroups be reclaimed below their normal consumption range? */
96 unsigned int may_thrash:1;
97
ee814fe2
JW
98 unsigned int hibernation_mode:1;
99
100 /* One of the zones is ready for compaction */
101 unsigned int compaction_ready:1;
102
103 /* Incremented by the number of inactive pages that were scanned */
104 unsigned long nr_scanned;
105
106 /* Number of pages freed so far during a call to shrink_zones() */
107 unsigned long nr_reclaimed;
1da177e4
LT
108};
109
1da177e4
LT
110#ifdef ARCH_HAS_PREFETCH
111#define prefetch_prev_lru_page(_page, _base, _field) \
112 do { \
113 if ((_page)->lru.prev != _base) { \
114 struct page *prev; \
115 \
116 prev = lru_to_page(&(_page->lru)); \
117 prefetch(&prev->_field); \
118 } \
119 } while (0)
120#else
121#define prefetch_prev_lru_page(_page, _base, _field) do { } while (0)
122#endif
123
124#ifdef ARCH_HAS_PREFETCHW
125#define prefetchw_prev_lru_page(_page, _base, _field) \
126 do { \
127 if ((_page)->lru.prev != _base) { \
128 struct page *prev; \
129 \
130 prev = lru_to_page(&(_page->lru)); \
131 prefetchw(&prev->_field); \
132 } \
133 } while (0)
134#else
135#define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0)
136#endif
137
138/*
139 * From 0 .. 100. Higher means more swappy.
140 */
141int vm_swappiness = 60;
d0480be4
WSH
142/*
143 * The total number of pages which are beyond the high watermark within all
144 * zones.
145 */
146unsigned long vm_total_pages;
1da177e4
LT
147
148static LIST_HEAD(shrinker_list);
149static DECLARE_RWSEM(shrinker_rwsem);
150
c255a458 151#ifdef CONFIG_MEMCG
89b5fae5
JW
152static bool global_reclaim(struct scan_control *sc)
153{
f16015fb 154 return !sc->target_mem_cgroup;
89b5fae5 155}
97c9341f
TH
156
157/**
158 * sane_reclaim - is the usual dirty throttling mechanism operational?
159 * @sc: scan_control in question
160 *
161 * The normal page dirty throttling mechanism in balance_dirty_pages() is
162 * completely broken with the legacy memcg and direct stalling in
163 * shrink_page_list() is used for throttling instead, which lacks all the
164 * niceties such as fairness, adaptive pausing, bandwidth proportional
165 * allocation and configurability.
166 *
167 * This function tests whether the vmscan currently in progress can assume
168 * that the normal dirty throttling mechanism is operational.
169 */
170static bool sane_reclaim(struct scan_control *sc)
171{
172 struct mem_cgroup *memcg = sc->target_mem_cgroup;
173
174 if (!memcg)
175 return true;
176#ifdef CONFIG_CGROUP_WRITEBACK
69234ace 177 if (cgroup_subsys_on_dfl(memory_cgrp_subsys))
97c9341f
TH
178 return true;
179#endif
180 return false;
181}
91a45470 182#else
89b5fae5
JW
183static bool global_reclaim(struct scan_control *sc)
184{
185 return true;
186}
97c9341f
TH
187
188static bool sane_reclaim(struct scan_control *sc)
189{
190 return true;
191}
91a45470
KH
192#endif
193
a1c3bfb2 194static unsigned long zone_reclaimable_pages(struct zone *zone)
6e543d57 195{
d031a157 196 unsigned long nr;
6e543d57 197
0db2cb8d
MH
198 nr = zone_page_state_snapshot(zone, NR_ACTIVE_FILE) +
199 zone_page_state_snapshot(zone, NR_INACTIVE_FILE) +
200 zone_page_state_snapshot(zone, NR_ISOLATED_FILE);
6e543d57
LD
201
202 if (get_nr_swap_pages() > 0)
0db2cb8d
MH
203 nr += zone_page_state_snapshot(zone, NR_ACTIVE_ANON) +
204 zone_page_state_snapshot(zone, NR_INACTIVE_ANON) +
205 zone_page_state_snapshot(zone, NR_ISOLATED_ANON);
6e543d57
LD
206
207 return nr;
208}
209
210bool zone_reclaimable(struct zone *zone)
211{
0db2cb8d 212 return zone_page_state_snapshot(zone, NR_PAGES_SCANNED) <
0d5d823a 213 zone_reclaimable_pages(zone) * 6;
6e543d57
LD
214}
215
23047a96 216unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru)
c9f299d9 217{
c3c787e8 218 if (!mem_cgroup_disabled())
4d7dcca2 219 return mem_cgroup_get_lru_size(lruvec, lru);
a3d8e054 220
074291fe 221 return zone_page_state(lruvec_zone(lruvec), NR_LRU_BASE + lru);
c9f299d9
KM
222}
223
1da177e4 224/*
1d3d4437 225 * Add a shrinker callback to be called from the vm.
1da177e4 226 */
1d3d4437 227int register_shrinker(struct shrinker *shrinker)
1da177e4 228{
1d3d4437
GC
229 size_t size = sizeof(*shrinker->nr_deferred);
230
1d3d4437
GC
231 if (shrinker->flags & SHRINKER_NUMA_AWARE)
232 size *= nr_node_ids;
233
234 shrinker->nr_deferred = kzalloc(size, GFP_KERNEL);
235 if (!shrinker->nr_deferred)
236 return -ENOMEM;
237
8e1f936b
RR
238 down_write(&shrinker_rwsem);
239 list_add_tail(&shrinker->list, &shrinker_list);
240 up_write(&shrinker_rwsem);
1d3d4437 241 return 0;
1da177e4 242}
8e1f936b 243EXPORT_SYMBOL(register_shrinker);
1da177e4
LT
244
245/*
246 * Remove one
247 */
8e1f936b 248void unregister_shrinker(struct shrinker *shrinker)
1da177e4
LT
249{
250 down_write(&shrinker_rwsem);
251 list_del(&shrinker->list);
252 up_write(&shrinker_rwsem);
ae393321 253 kfree(shrinker->nr_deferred);
1da177e4 254}
8e1f936b 255EXPORT_SYMBOL(unregister_shrinker);
1da177e4
LT
256
257#define SHRINK_BATCH 128
1d3d4437 258
cb731d6c
VD
259static unsigned long do_shrink_slab(struct shrink_control *shrinkctl,
260 struct shrinker *shrinker,
261 unsigned long nr_scanned,
262 unsigned long nr_eligible)
1d3d4437
GC
263{
264 unsigned long freed = 0;
265 unsigned long long delta;
266 long total_scan;
d5bc5fd3 267 long freeable;
1d3d4437
GC
268 long nr;
269 long new_nr;
270 int nid = shrinkctl->nid;
271 long batch_size = shrinker->batch ? shrinker->batch
272 : SHRINK_BATCH;
273
d5bc5fd3
VD
274 freeable = shrinker->count_objects(shrinker, shrinkctl);
275 if (freeable == 0)
1d3d4437
GC
276 return 0;
277
278 /*
279 * copy the current shrinker scan count into a local variable
280 * and zero it so that other concurrent shrinker invocations
281 * don't also do this scanning work.
282 */
283 nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0);
284
285 total_scan = nr;
6b4f7799 286 delta = (4 * nr_scanned) / shrinker->seeks;
d5bc5fd3 287 delta *= freeable;
6b4f7799 288 do_div(delta, nr_eligible + 1);
1d3d4437
GC
289 total_scan += delta;
290 if (total_scan < 0) {
8612c663 291 pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n",
a0b02131 292 shrinker->scan_objects, total_scan);
d5bc5fd3 293 total_scan = freeable;
1d3d4437
GC
294 }
295
296 /*
297 * We need to avoid excessive windup on filesystem shrinkers
298 * due to large numbers of GFP_NOFS allocations causing the
299 * shrinkers to return -1 all the time. This results in a large
300 * nr being built up so when a shrink that can do some work
301 * comes along it empties the entire cache due to nr >>>
d5bc5fd3 302 * freeable. This is bad for sustaining a working set in
1d3d4437
GC
303 * memory.
304 *
305 * Hence only allow the shrinker to scan the entire cache when
306 * a large delta change is calculated directly.
307 */
d5bc5fd3
VD
308 if (delta < freeable / 4)
309 total_scan = min(total_scan, freeable / 2);
1d3d4437
GC
310
311 /*
312 * Avoid risking looping forever due to too large nr value:
313 * never try to free more than twice the estimate number of
314 * freeable entries.
315 */
d5bc5fd3
VD
316 if (total_scan > freeable * 2)
317 total_scan = freeable * 2;
1d3d4437
GC
318
319 trace_mm_shrink_slab_start(shrinker, shrinkctl, nr,
6b4f7799
JW
320 nr_scanned, nr_eligible,
321 freeable, delta, total_scan);
1d3d4437 322
0b1fb40a
VD
323 /*
324 * Normally, we should not scan less than batch_size objects in one
325 * pass to avoid too frequent shrinker calls, but if the slab has less
326 * than batch_size objects in total and we are really tight on memory,
327 * we will try to reclaim all available objects, otherwise we can end
328 * up failing allocations although there are plenty of reclaimable
329 * objects spread over several slabs with usage less than the
330 * batch_size.
331 *
332 * We detect the "tight on memory" situations by looking at the total
333 * number of objects we want to scan (total_scan). If it is greater
d5bc5fd3 334 * than the total number of objects on slab (freeable), we must be
0b1fb40a
VD
335 * scanning at high prio and therefore should try to reclaim as much as
336 * possible.
337 */
338 while (total_scan >= batch_size ||
d5bc5fd3 339 total_scan >= freeable) {
a0b02131 340 unsigned long ret;
0b1fb40a 341 unsigned long nr_to_scan = min(batch_size, total_scan);
1d3d4437 342
0b1fb40a 343 shrinkctl->nr_to_scan = nr_to_scan;
a0b02131
DC
344 ret = shrinker->scan_objects(shrinker, shrinkctl);
345 if (ret == SHRINK_STOP)
346 break;
347 freed += ret;
1d3d4437 348
0b1fb40a
VD
349 count_vm_events(SLABS_SCANNED, nr_to_scan);
350 total_scan -= nr_to_scan;
1d3d4437
GC
351
352 cond_resched();
353 }
354
355 /*
356 * move the unused scan count back into the shrinker in a
357 * manner that handles concurrent updates. If we exhausted the
358 * scan, there is no need to do an update.
359 */
360 if (total_scan > 0)
361 new_nr = atomic_long_add_return(total_scan,
362 &shrinker->nr_deferred[nid]);
363 else
364 new_nr = atomic_long_read(&shrinker->nr_deferred[nid]);
365
df9024a8 366 trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan);
1d3d4437 367 return freed;
1495f230
YH
368}
369
6b4f7799 370/**
cb731d6c 371 * shrink_slab - shrink slab caches
6b4f7799
JW
372 * @gfp_mask: allocation context
373 * @nid: node whose slab caches to target
cb731d6c 374 * @memcg: memory cgroup whose slab caches to target
6b4f7799
JW
375 * @nr_scanned: pressure numerator
376 * @nr_eligible: pressure denominator
1da177e4 377 *
6b4f7799 378 * Call the shrink functions to age shrinkable caches.
1da177e4 379 *
6b4f7799
JW
380 * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set,
381 * unaware shrinkers will receive a node id of 0 instead.
1da177e4 382 *
cb731d6c
VD
383 * @memcg specifies the memory cgroup to target. If it is not NULL,
384 * only shrinkers with SHRINKER_MEMCG_AWARE set will be called to scan
0fc9f58a
VD
385 * objects from the memory cgroup specified. Otherwise, only unaware
386 * shrinkers are called.
cb731d6c 387 *
6b4f7799
JW
388 * @nr_scanned and @nr_eligible form a ratio that indicate how much of
389 * the available objects should be scanned. Page reclaim for example
390 * passes the number of pages scanned and the number of pages on the
391 * LRU lists that it considered on @nid, plus a bias in @nr_scanned
392 * when it encountered mapped pages. The ratio is further biased by
393 * the ->seeks setting of the shrink function, which indicates the
394 * cost to recreate an object relative to that of an LRU page.
b15e0905 395 *
6b4f7799 396 * Returns the number of reclaimed slab objects.
1da177e4 397 */
cb731d6c
VD
398static unsigned long shrink_slab(gfp_t gfp_mask, int nid,
399 struct mem_cgroup *memcg,
400 unsigned long nr_scanned,
401 unsigned long nr_eligible)
1da177e4
LT
402{
403 struct shrinker *shrinker;
24f7c6b9 404 unsigned long freed = 0;
1da177e4 405
0fc9f58a 406 if (memcg && (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)))
cb731d6c
VD
407 return 0;
408
6b4f7799
JW
409 if (nr_scanned == 0)
410 nr_scanned = SWAP_CLUSTER_MAX;
1da177e4 411
f06590bd 412 if (!down_read_trylock(&shrinker_rwsem)) {
24f7c6b9
DC
413 /*
414 * If we would return 0, our callers would understand that we
415 * have nothing else to shrink and give up trying. By returning
416 * 1 we keep it going and assume we'll be able to shrink next
417 * time.
418 */
419 freed = 1;
f06590bd
MK
420 goto out;
421 }
1da177e4
LT
422
423 list_for_each_entry(shrinker, &shrinker_list, list) {
6b4f7799
JW
424 struct shrink_control sc = {
425 .gfp_mask = gfp_mask,
426 .nid = nid,
cb731d6c 427 .memcg = memcg,
6b4f7799 428 };
ec97097b 429
0fc9f58a
VD
430 /*
431 * If kernel memory accounting is disabled, we ignore
432 * SHRINKER_MEMCG_AWARE flag and call all shrinkers
433 * passing NULL for memcg.
434 */
435 if (memcg_kmem_enabled() &&
436 !!memcg != !!(shrinker->flags & SHRINKER_MEMCG_AWARE))
cb731d6c
VD
437 continue;
438
6b4f7799
JW
439 if (!(shrinker->flags & SHRINKER_NUMA_AWARE))
440 sc.nid = 0;
1da177e4 441
cb731d6c 442 freed += do_shrink_slab(&sc, shrinker, nr_scanned, nr_eligible);
1da177e4 443 }
6b4f7799 444
1da177e4 445 up_read(&shrinker_rwsem);
f06590bd
MK
446out:
447 cond_resched();
24f7c6b9 448 return freed;
1da177e4
LT
449}
450
cb731d6c
VD
451void drop_slab_node(int nid)
452{
453 unsigned long freed;
454
455 do {
456 struct mem_cgroup *memcg = NULL;
457
458 freed = 0;
459 do {
460 freed += shrink_slab(GFP_KERNEL, nid, memcg,
461 1000, 1000);
462 } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL);
463 } while (freed > 10);
464}
465
466void drop_slab(void)
467{
468 int nid;
469
470 for_each_online_node(nid)
471 drop_slab_node(nid);
472}
473
1da177e4
LT
474static inline int is_page_cache_freeable(struct page *page)
475{
ceddc3a5
JW
476 /*
477 * A freeable page cache page is referenced only by the caller
478 * that isolated the page, the page cache radix tree and
479 * optional buffer heads at page->private.
480 */
edcf4748 481 return page_count(page) - page_has_private(page) == 2;
1da177e4
LT
482}
483
703c2708 484static int may_write_to_inode(struct inode *inode, struct scan_control *sc)
1da177e4 485{
930d9152 486 if (current->flags & PF_SWAPWRITE)
1da177e4 487 return 1;
703c2708 488 if (!inode_write_congested(inode))
1da177e4 489 return 1;
703c2708 490 if (inode_to_bdi(inode) == current->backing_dev_info)
1da177e4
LT
491 return 1;
492 return 0;
493}
494
495/*
496 * We detected a synchronous write error writing a page out. Probably
497 * -ENOSPC. We need to propagate that into the address_space for a subsequent
498 * fsync(), msync() or close().
499 *
500 * The tricky part is that after writepage we cannot touch the mapping: nothing
501 * prevents it from being freed up. But we have a ref on the page and once
502 * that page is locked, the mapping is pinned.
503 *
504 * We're allowed to run sleeping lock_page() here because we know the caller has
505 * __GFP_FS.
506 */
507static void handle_write_error(struct address_space *mapping,
508 struct page *page, int error)
509{
7eaceacc 510 lock_page(page);
3e9f45bd
GC
511 if (page_mapping(page) == mapping)
512 mapping_set_error(mapping, error);
1da177e4
LT
513 unlock_page(page);
514}
515
04e62a29
CL
516/* possible outcome of pageout() */
517typedef enum {
518 /* failed to write page out, page is locked */
519 PAGE_KEEP,
520 /* move page to the active list, page is locked */
521 PAGE_ACTIVATE,
522 /* page has been sent to the disk successfully, page is unlocked */
523 PAGE_SUCCESS,
524 /* page is clean and locked */
525 PAGE_CLEAN,
526} pageout_t;
527
1da177e4 528/*
1742f19f
AM
529 * pageout is called by shrink_page_list() for each dirty page.
530 * Calls ->writepage().
1da177e4 531 */
c661b078 532static pageout_t pageout(struct page *page, struct address_space *mapping,
7d3579e8 533 struct scan_control *sc)
1da177e4
LT
534{
535 /*
536 * If the page is dirty, only perform writeback if that write
537 * will be non-blocking. To prevent this allocation from being
538 * stalled by pagecache activity. But note that there may be
539 * stalls if we need to run get_block(). We could test
540 * PagePrivate for that.
541 *
8174202b 542 * If this process is currently in __generic_file_write_iter() against
1da177e4
LT
543 * this page's queue, we can perform writeback even if that
544 * will block.
545 *
546 * If the page is swapcache, write it back even if that would
547 * block, for some throttling. This happens by accident, because
548 * swap_backing_dev_info is bust: it doesn't reflect the
549 * congestion state of the swapdevs. Easy to fix, if needed.
1da177e4
LT
550 */
551 if (!is_page_cache_freeable(page))
552 return PAGE_KEEP;
553 if (!mapping) {
554 /*
555 * Some data journaling orphaned pages can have
556 * page->mapping == NULL while being dirty with clean buffers.
557 */
266cf658 558 if (page_has_private(page)) {
1da177e4
LT
559 if (try_to_free_buffers(page)) {
560 ClearPageDirty(page);
b1de0d13 561 pr_info("%s: orphaned page\n", __func__);
1da177e4
LT
562 return PAGE_CLEAN;
563 }
564 }
565 return PAGE_KEEP;
566 }
567 if (mapping->a_ops->writepage == NULL)
568 return PAGE_ACTIVATE;
703c2708 569 if (!may_write_to_inode(mapping->host, sc))
1da177e4
LT
570 return PAGE_KEEP;
571
572 if (clear_page_dirty_for_io(page)) {
573 int res;
574 struct writeback_control wbc = {
575 .sync_mode = WB_SYNC_NONE,
576 .nr_to_write = SWAP_CLUSTER_MAX,
111ebb6e
OH
577 .range_start = 0,
578 .range_end = LLONG_MAX,
1da177e4
LT
579 .for_reclaim = 1,
580 };
581
582 SetPageReclaim(page);
583 res = mapping->a_ops->writepage(page, &wbc);
584 if (res < 0)
585 handle_write_error(mapping, page, res);
994fc28c 586 if (res == AOP_WRITEPAGE_ACTIVATE) {
1da177e4
LT
587 ClearPageReclaim(page);
588 return PAGE_ACTIVATE;
589 }
c661b078 590
1da177e4
LT
591 if (!PageWriteback(page)) {
592 /* synchronous write or broken a_ops? */
593 ClearPageReclaim(page);
594 }
3aa23851 595 trace_mm_vmscan_writepage(page);
e129b5c2 596 inc_zone_page_state(page, NR_VMSCAN_WRITE);
1da177e4
LT
597 return PAGE_SUCCESS;
598 }
599
600 return PAGE_CLEAN;
601}
602
a649fd92 603/*
e286781d
NP
604 * Same as remove_mapping, but if the page is removed from the mapping, it
605 * gets returned with a refcount of 0.
a649fd92 606 */
a528910e
JW
607static int __remove_mapping(struct address_space *mapping, struct page *page,
608 bool reclaimed)
49d2e9cc 609{
c4843a75 610 unsigned long flags;
c4843a75 611
28e4d965
NP
612 BUG_ON(!PageLocked(page));
613 BUG_ON(mapping != page_mapping(page));
49d2e9cc 614
c4843a75 615 spin_lock_irqsave(&mapping->tree_lock, flags);
49d2e9cc 616 /*
0fd0e6b0
NP
617 * The non racy check for a busy page.
618 *
619 * Must be careful with the order of the tests. When someone has
620 * a ref to the page, it may be possible that they dirty it then
621 * drop the reference. So if PageDirty is tested before page_count
622 * here, then the following race may occur:
623 *
624 * get_user_pages(&page);
625 * [user mapping goes away]
626 * write_to(page);
627 * !PageDirty(page) [good]
628 * SetPageDirty(page);
629 * put_page(page);
630 * !page_count(page) [good, discard it]
631 *
632 * [oops, our write_to data is lost]
633 *
634 * Reversing the order of the tests ensures such a situation cannot
635 * escape unnoticed. The smp_rmb is needed to ensure the page->flags
0139aa7b 636 * load is not satisfied before that of page->_refcount.
0fd0e6b0
NP
637 *
638 * Note that if SetPageDirty is always performed via set_page_dirty,
639 * and thus under tree_lock, then this ordering is not required.
49d2e9cc 640 */
fe896d18 641 if (!page_ref_freeze(page, 2))
49d2e9cc 642 goto cannot_free;
e286781d
NP
643 /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */
644 if (unlikely(PageDirty(page))) {
fe896d18 645 page_ref_unfreeze(page, 2);
49d2e9cc 646 goto cannot_free;
e286781d 647 }
49d2e9cc
CL
648
649 if (PageSwapCache(page)) {
650 swp_entry_t swap = { .val = page_private(page) };
0a31bc97 651 mem_cgroup_swapout(page, swap);
49d2e9cc 652 __delete_from_swap_cache(page);
c4843a75 653 spin_unlock_irqrestore(&mapping->tree_lock, flags);
0a31bc97 654 swapcache_free(swap);
e286781d 655 } else {
6072d13c 656 void (*freepage)(struct page *);
a528910e 657 void *shadow = NULL;
6072d13c
LT
658
659 freepage = mapping->a_ops->freepage;
a528910e
JW
660 /*
661 * Remember a shadow entry for reclaimed file cache in
662 * order to detect refaults, thus thrashing, later on.
663 *
664 * But don't store shadows in an address space that is
665 * already exiting. This is not just an optizimation,
666 * inode reclaim needs to empty out the radix tree or
667 * the nodes are lost. Don't plant shadows behind its
668 * back.
f9fe48be
RZ
669 *
670 * We also don't store shadows for DAX mappings because the
671 * only page cache pages found in these are zero pages
672 * covering holes, and because we don't want to mix DAX
673 * exceptional entries and shadow exceptional entries in the
674 * same page_tree.
a528910e
JW
675 */
676 if (reclaimed && page_is_file_cache(page) &&
f9fe48be 677 !mapping_exiting(mapping) && !dax_mapping(mapping))
a528910e 678 shadow = workingset_eviction(mapping, page);
62cccb8c 679 __delete_from_page_cache(page, shadow);
c4843a75 680 spin_unlock_irqrestore(&mapping->tree_lock, flags);
6072d13c
LT
681
682 if (freepage != NULL)
683 freepage(page);
49d2e9cc
CL
684 }
685
49d2e9cc
CL
686 return 1;
687
688cannot_free:
c4843a75 689 spin_unlock_irqrestore(&mapping->tree_lock, flags);
49d2e9cc
CL
690 return 0;
691}
692
e286781d
NP
693/*
694 * Attempt to detach a locked page from its ->mapping. If it is dirty or if
695 * someone else has a ref on the page, abort and return 0. If it was
696 * successfully detached, return 1. Assumes the caller has a single ref on
697 * this page.
698 */
699int remove_mapping(struct address_space *mapping, struct page *page)
700{
a528910e 701 if (__remove_mapping(mapping, page, false)) {
e286781d
NP
702 /*
703 * Unfreezing the refcount with 1 rather than 2 effectively
704 * drops the pagecache ref for us without requiring another
705 * atomic operation.
706 */
fe896d18 707 page_ref_unfreeze(page, 1);
e286781d
NP
708 return 1;
709 }
710 return 0;
711}
712
894bc310
LS
713/**
714 * putback_lru_page - put previously isolated page onto appropriate LRU list
715 * @page: page to be put back to appropriate lru list
716 *
717 * Add previously isolated @page to appropriate LRU list.
718 * Page may still be unevictable for other reasons.
719 *
720 * lru_lock must not be held, interrupts must be enabled.
721 */
894bc310
LS
722void putback_lru_page(struct page *page)
723{
0ec3b74c 724 bool is_unevictable;
bbfd28ee 725 int was_unevictable = PageUnevictable(page);
894bc310 726
309381fe 727 VM_BUG_ON_PAGE(PageLRU(page), page);
894bc310
LS
728
729redo:
730 ClearPageUnevictable(page);
731
39b5f29a 732 if (page_evictable(page)) {
894bc310
LS
733 /*
734 * For evictable pages, we can use the cache.
735 * In event of a race, worst case is we end up with an
736 * unevictable page on [in]active list.
737 * We know how to handle that.
738 */
0ec3b74c 739 is_unevictable = false;
c53954a0 740 lru_cache_add(page);
894bc310
LS
741 } else {
742 /*
743 * Put unevictable pages directly on zone's unevictable
744 * list.
745 */
0ec3b74c 746 is_unevictable = true;
894bc310 747 add_page_to_unevictable_list(page);
6a7b9548 748 /*
21ee9f39
MK
749 * When racing with an mlock or AS_UNEVICTABLE clearing
750 * (page is unlocked) make sure that if the other thread
751 * does not observe our setting of PG_lru and fails
24513264 752 * isolation/check_move_unevictable_pages,
21ee9f39 753 * we see PG_mlocked/AS_UNEVICTABLE cleared below and move
6a7b9548
JW
754 * the page back to the evictable list.
755 *
21ee9f39 756 * The other side is TestClearPageMlocked() or shmem_lock().
6a7b9548
JW
757 */
758 smp_mb();
894bc310 759 }
894bc310
LS
760
761 /*
762 * page's status can change while we move it among lru. If an evictable
763 * page is on unevictable list, it never be freed. To avoid that,
764 * check after we added it to the list, again.
765 */
0ec3b74c 766 if (is_unevictable && page_evictable(page)) {
894bc310
LS
767 if (!isolate_lru_page(page)) {
768 put_page(page);
769 goto redo;
770 }
771 /* This means someone else dropped this page from LRU
772 * So, it will be freed or putback to LRU again. There is
773 * nothing to do here.
774 */
775 }
776
0ec3b74c 777 if (was_unevictable && !is_unevictable)
bbfd28ee 778 count_vm_event(UNEVICTABLE_PGRESCUED);
0ec3b74c 779 else if (!was_unevictable && is_unevictable)
bbfd28ee
LS
780 count_vm_event(UNEVICTABLE_PGCULLED);
781
894bc310
LS
782 put_page(page); /* drop ref from isolate */
783}
784
dfc8d636
JW
785enum page_references {
786 PAGEREF_RECLAIM,
787 PAGEREF_RECLAIM_CLEAN,
64574746 788 PAGEREF_KEEP,
dfc8d636
JW
789 PAGEREF_ACTIVATE,
790};
791
792static enum page_references page_check_references(struct page *page,
793 struct scan_control *sc)
794{
64574746 795 int referenced_ptes, referenced_page;
dfc8d636 796 unsigned long vm_flags;
dfc8d636 797
c3ac9a8a
JW
798 referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup,
799 &vm_flags);
64574746 800 referenced_page = TestClearPageReferenced(page);
dfc8d636 801
dfc8d636
JW
802 /*
803 * Mlock lost the isolation race with us. Let try_to_unmap()
804 * move the page to the unevictable list.
805 */
806 if (vm_flags & VM_LOCKED)
807 return PAGEREF_RECLAIM;
808
64574746 809 if (referenced_ptes) {
e4898273 810 if (PageSwapBacked(page))
64574746
JW
811 return PAGEREF_ACTIVATE;
812 /*
813 * All mapped pages start out with page table
814 * references from the instantiating fault, so we need
815 * to look twice if a mapped file page is used more
816 * than once.
817 *
818 * Mark it and spare it for another trip around the
819 * inactive list. Another page table reference will
820 * lead to its activation.
821 *
822 * Note: the mark is set for activated pages as well
823 * so that recently deactivated but used pages are
824 * quickly recovered.
825 */
826 SetPageReferenced(page);
827
34dbc67a 828 if (referenced_page || referenced_ptes > 1)
64574746
JW
829 return PAGEREF_ACTIVATE;
830
c909e993
KK
831 /*
832 * Activate file-backed executable pages after first usage.
833 */
834 if (vm_flags & VM_EXEC)
835 return PAGEREF_ACTIVATE;
836
64574746
JW
837 return PAGEREF_KEEP;
838 }
dfc8d636
JW
839
840 /* Reclaim if clean, defer dirty pages to writeback */
2e30244a 841 if (referenced_page && !PageSwapBacked(page))
64574746
JW
842 return PAGEREF_RECLAIM_CLEAN;
843
844 return PAGEREF_RECLAIM;
dfc8d636
JW
845}
846
e2be15f6
MG
847/* Check if a page is dirty or under writeback */
848static void page_check_dirty_writeback(struct page *page,
849 bool *dirty, bool *writeback)
850{
b4597226
MG
851 struct address_space *mapping;
852
e2be15f6
MG
853 /*
854 * Anonymous pages are not handled by flushers and must be written
855 * from reclaim context. Do not stall reclaim based on them
856 */
857 if (!page_is_file_cache(page)) {
858 *dirty = false;
859 *writeback = false;
860 return;
861 }
862
863 /* By default assume that the page flags are accurate */
864 *dirty = PageDirty(page);
865 *writeback = PageWriteback(page);
b4597226
MG
866
867 /* Verify dirty/writeback state if the filesystem supports it */
868 if (!page_has_private(page))
869 return;
870
871 mapping = page_mapping(page);
872 if (mapping && mapping->a_ops->is_dirty_writeback)
873 mapping->a_ops->is_dirty_writeback(page, dirty, writeback);
e2be15f6
MG
874}
875
1da177e4 876/*
1742f19f 877 * shrink_page_list() returns the number of reclaimed pages
1da177e4 878 */
1742f19f 879static unsigned long shrink_page_list(struct list_head *page_list,
6a18adb3 880 struct zone *zone,
f84f6e2b 881 struct scan_control *sc,
02c6de8d 882 enum ttu_flags ttu_flags,
8e950282 883 unsigned long *ret_nr_dirty,
d43006d5 884 unsigned long *ret_nr_unqueued_dirty,
8e950282 885 unsigned long *ret_nr_congested,
02c6de8d 886 unsigned long *ret_nr_writeback,
b1a6f21e 887 unsigned long *ret_nr_immediate,
02c6de8d 888 bool force_reclaim)
1da177e4
LT
889{
890 LIST_HEAD(ret_pages);
abe4c3b5 891 LIST_HEAD(free_pages);
1da177e4 892 int pgactivate = 0;
d43006d5 893 unsigned long nr_unqueued_dirty = 0;
0e093d99
MG
894 unsigned long nr_dirty = 0;
895 unsigned long nr_congested = 0;
05ff5137 896 unsigned long nr_reclaimed = 0;
92df3a72 897 unsigned long nr_writeback = 0;
b1a6f21e 898 unsigned long nr_immediate = 0;
1da177e4
LT
899
900 cond_resched();
901
1da177e4
LT
902 while (!list_empty(page_list)) {
903 struct address_space *mapping;
904 struct page *page;
905 int may_enter_fs;
02c6de8d 906 enum page_references references = PAGEREF_RECLAIM_CLEAN;
e2be15f6 907 bool dirty, writeback;
854e9ed0
MK
908 bool lazyfree = false;
909 int ret = SWAP_SUCCESS;
1da177e4
LT
910
911 cond_resched();
912
913 page = lru_to_page(page_list);
914 list_del(&page->lru);
915
529ae9aa 916 if (!trylock_page(page))
1da177e4
LT
917 goto keep;
918
309381fe
SL
919 VM_BUG_ON_PAGE(PageActive(page), page);
920 VM_BUG_ON_PAGE(page_zone(page) != zone, page);
1da177e4
LT
921
922 sc->nr_scanned++;
80e43426 923
39b5f29a 924 if (unlikely(!page_evictable(page)))
b291f000 925 goto cull_mlocked;
894bc310 926
a6dc60f8 927 if (!sc->may_unmap && page_mapped(page))
80e43426
CL
928 goto keep_locked;
929
1da177e4
LT
930 /* Double the slab pressure for mapped and swapcache pages */
931 if (page_mapped(page) || PageSwapCache(page))
932 sc->nr_scanned++;
933
c661b078
AW
934 may_enter_fs = (sc->gfp_mask & __GFP_FS) ||
935 (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO));
936
e2be15f6
MG
937 /*
938 * The number of dirty pages determines if a zone is marked
939 * reclaim_congested which affects wait_iff_congested. kswapd
940 * will stall and start writing pages if the tail of the LRU
941 * is all dirty unqueued pages.
942 */
943 page_check_dirty_writeback(page, &dirty, &writeback);
944 if (dirty || writeback)
945 nr_dirty++;
946
947 if (dirty && !writeback)
948 nr_unqueued_dirty++;
949
d04e8acd
MG
950 /*
951 * Treat this page as congested if the underlying BDI is or if
952 * pages are cycling through the LRU so quickly that the
953 * pages marked for immediate reclaim are making it to the
954 * end of the LRU a second time.
955 */
e2be15f6 956 mapping = page_mapping(page);
1da58ee2 957 if (((dirty || writeback) && mapping &&
703c2708 958 inode_write_congested(mapping->host)) ||
d04e8acd 959 (writeback && PageReclaim(page)))
e2be15f6
MG
960 nr_congested++;
961
283aba9f
MG
962 /*
963 * If a page at the tail of the LRU is under writeback, there
964 * are three cases to consider.
965 *
966 * 1) If reclaim is encountering an excessive number of pages
967 * under writeback and this page is both under writeback and
968 * PageReclaim then it indicates that pages are being queued
969 * for IO but are being recycled through the LRU before the
970 * IO can complete. Waiting on the page itself risks an
971 * indefinite stall if it is impossible to writeback the
972 * page due to IO error or disconnected storage so instead
b1a6f21e
MG
973 * note that the LRU is being scanned too quickly and the
974 * caller can stall after page list has been processed.
283aba9f 975 *
97c9341f 976 * 2) Global or new memcg reclaim encounters a page that is
ecf5fc6e
MH
977 * not marked for immediate reclaim, or the caller does not
978 * have __GFP_FS (or __GFP_IO if it's simply going to swap,
979 * not to fs). In this case mark the page for immediate
97c9341f 980 * reclaim and continue scanning.
283aba9f 981 *
ecf5fc6e
MH
982 * Require may_enter_fs because we would wait on fs, which
983 * may not have submitted IO yet. And the loop driver might
283aba9f
MG
984 * enter reclaim, and deadlock if it waits on a page for
985 * which it is needed to do the write (loop masks off
986 * __GFP_IO|__GFP_FS for this reason); but more thought
987 * would probably show more reasons.
988 *
7fadc820 989 * 3) Legacy memcg encounters a page that is already marked
283aba9f
MG
990 * PageReclaim. memcg does not have any dirty pages
991 * throttling so we could easily OOM just because too many
992 * pages are in writeback and there is nothing else to
993 * reclaim. Wait for the writeback to complete.
994 */
c661b078 995 if (PageWriteback(page)) {
283aba9f
MG
996 /* Case 1 above */
997 if (current_is_kswapd() &&
998 PageReclaim(page) &&
57054651 999 test_bit(ZONE_WRITEBACK, &zone->flags)) {
b1a6f21e
MG
1000 nr_immediate++;
1001 goto keep_locked;
283aba9f
MG
1002
1003 /* Case 2 above */
97c9341f 1004 } else if (sane_reclaim(sc) ||
ecf5fc6e 1005 !PageReclaim(page) || !may_enter_fs) {
c3b94f44
HD
1006 /*
1007 * This is slightly racy - end_page_writeback()
1008 * might have just cleared PageReclaim, then
1009 * setting PageReclaim here end up interpreted
1010 * as PageReadahead - but that does not matter
1011 * enough to care. What we do want is for this
1012 * page to have PageReclaim set next time memcg
1013 * reclaim reaches the tests above, so it will
1014 * then wait_on_page_writeback() to avoid OOM;
1015 * and it's also appropriate in global reclaim.
1016 */
1017 SetPageReclaim(page);
e62e384e 1018 nr_writeback++;
c3b94f44 1019 goto keep_locked;
283aba9f
MG
1020
1021 /* Case 3 above */
1022 } else {
7fadc820 1023 unlock_page(page);
283aba9f 1024 wait_on_page_writeback(page);
7fadc820
HD
1025 /* then go back and try same page again */
1026 list_add_tail(&page->lru, page_list);
1027 continue;
e62e384e 1028 }
c661b078 1029 }
1da177e4 1030
02c6de8d
MK
1031 if (!force_reclaim)
1032 references = page_check_references(page, sc);
1033
dfc8d636
JW
1034 switch (references) {
1035 case PAGEREF_ACTIVATE:
1da177e4 1036 goto activate_locked;
64574746
JW
1037 case PAGEREF_KEEP:
1038 goto keep_locked;
dfc8d636
JW
1039 case PAGEREF_RECLAIM:
1040 case PAGEREF_RECLAIM_CLEAN:
1041 ; /* try to reclaim the page below */
1042 }
1da177e4 1043
1da177e4
LT
1044 /*
1045 * Anonymous process memory has backing store?
1046 * Try to allocate it some swap space here.
1047 */
b291f000 1048 if (PageAnon(page) && !PageSwapCache(page)) {
63eb6b93
HD
1049 if (!(sc->gfp_mask & __GFP_IO))
1050 goto keep_locked;
5bc7b8ac 1051 if (!add_to_swap(page, page_list))
1da177e4 1052 goto activate_locked;
854e9ed0 1053 lazyfree = true;
63eb6b93 1054 may_enter_fs = 1;
1da177e4 1055
e2be15f6
MG
1056 /* Adding to swap updated mapping */
1057 mapping = page_mapping(page);
1058 }
1da177e4
LT
1059
1060 /*
1061 * The page is mapped into the page tables of one or more
1062 * processes. Try to unmap it here.
1063 */
1064 if (page_mapped(page) && mapping) {
854e9ed0
MK
1065 switch (ret = try_to_unmap(page, lazyfree ?
1066 (ttu_flags | TTU_BATCH_FLUSH | TTU_LZFREE) :
1067 (ttu_flags | TTU_BATCH_FLUSH))) {
1da177e4
LT
1068 case SWAP_FAIL:
1069 goto activate_locked;
1070 case SWAP_AGAIN:
1071 goto keep_locked;
b291f000
NP
1072 case SWAP_MLOCK:
1073 goto cull_mlocked;
854e9ed0
MK
1074 case SWAP_LZFREE:
1075 goto lazyfree;
1da177e4
LT
1076 case SWAP_SUCCESS:
1077 ; /* try to free the page below */
1078 }
1079 }
1080
1081 if (PageDirty(page)) {
ee72886d
MG
1082 /*
1083 * Only kswapd can writeback filesystem pages to
d43006d5
MG
1084 * avoid risk of stack overflow but only writeback
1085 * if many dirty pages have been encountered.
ee72886d 1086 */
f84f6e2b 1087 if (page_is_file_cache(page) &&
9e3b2f8c 1088 (!current_is_kswapd() ||
57054651 1089 !test_bit(ZONE_DIRTY, &zone->flags))) {
49ea7eb6
MG
1090 /*
1091 * Immediately reclaim when written back.
1092 * Similar in principal to deactivate_page()
1093 * except we already have the page isolated
1094 * and know it's dirty
1095 */
1096 inc_zone_page_state(page, NR_VMSCAN_IMMEDIATE);
1097 SetPageReclaim(page);
1098
ee72886d
MG
1099 goto keep_locked;
1100 }
1101
dfc8d636 1102 if (references == PAGEREF_RECLAIM_CLEAN)
1da177e4 1103 goto keep_locked;
4dd4b920 1104 if (!may_enter_fs)
1da177e4 1105 goto keep_locked;
52a8363e 1106 if (!sc->may_writepage)
1da177e4
LT
1107 goto keep_locked;
1108
d950c947
MG
1109 /*
1110 * Page is dirty. Flush the TLB if a writable entry
1111 * potentially exists to avoid CPU writes after IO
1112 * starts and then write it out here.
1113 */
1114 try_to_unmap_flush_dirty();
7d3579e8 1115 switch (pageout(page, mapping, sc)) {
1da177e4
LT
1116 case PAGE_KEEP:
1117 goto keep_locked;
1118 case PAGE_ACTIVATE:
1119 goto activate_locked;
1120 case PAGE_SUCCESS:
7d3579e8 1121 if (PageWriteback(page))
41ac1999 1122 goto keep;
7d3579e8 1123 if (PageDirty(page))
1da177e4 1124 goto keep;
7d3579e8 1125
1da177e4
LT
1126 /*
1127 * A synchronous write - probably a ramdisk. Go
1128 * ahead and try to reclaim the page.
1129 */
529ae9aa 1130 if (!trylock_page(page))
1da177e4
LT
1131 goto keep;
1132 if (PageDirty(page) || PageWriteback(page))
1133 goto keep_locked;
1134 mapping = page_mapping(page);
1135 case PAGE_CLEAN:
1136 ; /* try to free the page below */
1137 }
1138 }
1139
1140 /*
1141 * If the page has buffers, try to free the buffer mappings
1142 * associated with this page. If we succeed we try to free
1143 * the page as well.
1144 *
1145 * We do this even if the page is PageDirty().
1146 * try_to_release_page() does not perform I/O, but it is
1147 * possible for a page to have PageDirty set, but it is actually
1148 * clean (all its buffers are clean). This happens if the
1149 * buffers were written out directly, with submit_bh(). ext3
894bc310 1150 * will do this, as well as the blockdev mapping.
1da177e4
LT
1151 * try_to_release_page() will discover that cleanness and will
1152 * drop the buffers and mark the page clean - it can be freed.
1153 *
1154 * Rarely, pages can have buffers and no ->mapping. These are
1155 * the pages which were not successfully invalidated in
1156 * truncate_complete_page(). We try to drop those buffers here
1157 * and if that worked, and the page is no longer mapped into
1158 * process address space (page_count == 1) it can be freed.
1159 * Otherwise, leave the page on the LRU so it is swappable.
1160 */
266cf658 1161 if (page_has_private(page)) {
1da177e4
LT
1162 if (!try_to_release_page(page, sc->gfp_mask))
1163 goto activate_locked;
e286781d
NP
1164 if (!mapping && page_count(page) == 1) {
1165 unlock_page(page);
1166 if (put_page_testzero(page))
1167 goto free_it;
1168 else {
1169 /*
1170 * rare race with speculative reference.
1171 * the speculative reference will free
1172 * this page shortly, so we may
1173 * increment nr_reclaimed here (and
1174 * leave it off the LRU).
1175 */
1176 nr_reclaimed++;
1177 continue;
1178 }
1179 }
1da177e4
LT
1180 }
1181
854e9ed0 1182lazyfree:
a528910e 1183 if (!mapping || !__remove_mapping(mapping, page, true))
49d2e9cc 1184 goto keep_locked;
1da177e4 1185
a978d6f5
NP
1186 /*
1187 * At this point, we have no other references and there is
1188 * no way to pick any more up (removed from LRU, removed
1189 * from pagecache). Can use non-atomic bitops now (and
1190 * we obviously don't have to worry about waking up a process
1191 * waiting on the page lock, because there are no references.
1192 */
48c935ad 1193 __ClearPageLocked(page);
e286781d 1194free_it:
854e9ed0
MK
1195 if (ret == SWAP_LZFREE)
1196 count_vm_event(PGLAZYFREED);
1197
05ff5137 1198 nr_reclaimed++;
abe4c3b5
MG
1199
1200 /*
1201 * Is there need to periodically free_page_list? It would
1202 * appear not as the counts should be low
1203 */
1204 list_add(&page->lru, &free_pages);
1da177e4
LT
1205 continue;
1206
b291f000 1207cull_mlocked:
63d6c5ad
HD
1208 if (PageSwapCache(page))
1209 try_to_free_swap(page);
b291f000 1210 unlock_page(page);
c54839a7 1211 list_add(&page->lru, &ret_pages);
b291f000
NP
1212 continue;
1213
1da177e4 1214activate_locked:
68a22394 1215 /* Not a candidate for swapping, so reclaim swap space. */
5ccc5aba 1216 if (PageSwapCache(page) && mem_cgroup_swap_full(page))
a2c43eed 1217 try_to_free_swap(page);
309381fe 1218 VM_BUG_ON_PAGE(PageActive(page), page);
1da177e4
LT
1219 SetPageActive(page);
1220 pgactivate++;
1221keep_locked:
1222 unlock_page(page);
1223keep:
1224 list_add(&page->lru, &ret_pages);
309381fe 1225 VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page);
1da177e4 1226 }
abe4c3b5 1227
747db954 1228 mem_cgroup_uncharge_list(&free_pages);
72b252ae 1229 try_to_unmap_flush();
b745bc85 1230 free_hot_cold_page_list(&free_pages, true);
abe4c3b5 1231
1da177e4 1232 list_splice(&ret_pages, page_list);
f8891e5e 1233 count_vm_events(PGACTIVATE, pgactivate);
0a31bc97 1234
8e950282
MG
1235 *ret_nr_dirty += nr_dirty;
1236 *ret_nr_congested += nr_congested;
d43006d5 1237 *ret_nr_unqueued_dirty += nr_unqueued_dirty;
92df3a72 1238 *ret_nr_writeback += nr_writeback;
b1a6f21e 1239 *ret_nr_immediate += nr_immediate;
05ff5137 1240 return nr_reclaimed;
1da177e4
LT
1241}
1242
02c6de8d
MK
1243unsigned long reclaim_clean_pages_from_list(struct zone *zone,
1244 struct list_head *page_list)
1245{
1246 struct scan_control sc = {
1247 .gfp_mask = GFP_KERNEL,
1248 .priority = DEF_PRIORITY,
1249 .may_unmap = 1,
1250 };
8e950282 1251 unsigned long ret, dummy1, dummy2, dummy3, dummy4, dummy5;
02c6de8d
MK
1252 struct page *page, *next;
1253 LIST_HEAD(clean_pages);
1254
1255 list_for_each_entry_safe(page, next, page_list, lru) {
117aad1e
RA
1256 if (page_is_file_cache(page) && !PageDirty(page) &&
1257 !isolated_balloon_page(page)) {
02c6de8d
MK
1258 ClearPageActive(page);
1259 list_move(&page->lru, &clean_pages);
1260 }
1261 }
1262
1263 ret = shrink_page_list(&clean_pages, zone, &sc,
8e950282
MG
1264 TTU_UNMAP|TTU_IGNORE_ACCESS,
1265 &dummy1, &dummy2, &dummy3, &dummy4, &dummy5, true);
02c6de8d 1266 list_splice(&clean_pages, page_list);
83da7510 1267 mod_zone_page_state(zone, NR_ISOLATED_FILE, -ret);
02c6de8d
MK
1268 return ret;
1269}
1270
5ad333eb
AW
1271/*
1272 * Attempt to remove the specified page from its LRU. Only take this page
1273 * if it is of the appropriate PageActive status. Pages which are being
1274 * freed elsewhere are also ignored.
1275 *
1276 * page: page to consider
1277 * mode: one of the LRU isolation modes defined above
1278 *
1279 * returns 0 on success, -ve errno on failure.
1280 */
f3fd4a61 1281int __isolate_lru_page(struct page *page, isolate_mode_t mode)
5ad333eb
AW
1282{
1283 int ret = -EINVAL;
1284
1285 /* Only take pages on the LRU. */
1286 if (!PageLRU(page))
1287 return ret;
1288
e46a2879
MK
1289 /* Compaction should not handle unevictable pages but CMA can do so */
1290 if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE))
894bc310
LS
1291 return ret;
1292
5ad333eb 1293 ret = -EBUSY;
08e552c6 1294
c8244935
MG
1295 /*
1296 * To minimise LRU disruption, the caller can indicate that it only
1297 * wants to isolate pages it will be able to operate on without
1298 * blocking - clean pages for the most part.
1299 *
1300 * ISOLATE_CLEAN means that only clean pages should be isolated. This
1301 * is used by reclaim when it is cannot write to backing storage
1302 *
1303 * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages
1304 * that it is possible to migrate without blocking
1305 */
1306 if (mode & (ISOLATE_CLEAN|ISOLATE_ASYNC_MIGRATE)) {
1307 /* All the caller can do on PageWriteback is block */
1308 if (PageWriteback(page))
1309 return ret;
1310
1311 if (PageDirty(page)) {
1312 struct address_space *mapping;
1313
1314 /* ISOLATE_CLEAN means only clean pages */
1315 if (mode & ISOLATE_CLEAN)
1316 return ret;
1317
1318 /*
1319 * Only pages without mappings or that have a
1320 * ->migratepage callback are possible to migrate
1321 * without blocking
1322 */
1323 mapping = page_mapping(page);
1324 if (mapping && !mapping->a_ops->migratepage)
1325 return ret;
1326 }
1327 }
39deaf85 1328
f80c0673
MK
1329 if ((mode & ISOLATE_UNMAPPED) && page_mapped(page))
1330 return ret;
1331
5ad333eb
AW
1332 if (likely(get_page_unless_zero(page))) {
1333 /*
1334 * Be careful not to clear PageLRU until after we're
1335 * sure the page is not being freed elsewhere -- the
1336 * page release code relies on it.
1337 */
1338 ClearPageLRU(page);
1339 ret = 0;
1340 }
1341
1342 return ret;
1343}
1344
1da177e4
LT
1345/*
1346 * zone->lru_lock is heavily contended. Some of the functions that
1347 * shrink the lists perform better by taking out a batch of pages
1348 * and working on them outside the LRU lock.
1349 *
1350 * For pagecache intensive workloads, this function is the hottest
1351 * spot in the kernel (apart from copy_*_user functions).
1352 *
1353 * Appropriate locks must be held before calling this function.
1354 *
1355 * @nr_to_scan: The number of pages to look through on the list.
5dc35979 1356 * @lruvec: The LRU vector to pull pages from.
1da177e4 1357 * @dst: The temp list to put pages on to.
f626012d 1358 * @nr_scanned: The number of pages that were scanned.
fe2c2a10 1359 * @sc: The scan_control struct for this reclaim session
5ad333eb 1360 * @mode: One of the LRU isolation modes
3cb99451 1361 * @lru: LRU list id for isolating
1da177e4
LT
1362 *
1363 * returns how many pages were moved onto *@dst.
1364 */
69e05944 1365static unsigned long isolate_lru_pages(unsigned long nr_to_scan,
5dc35979 1366 struct lruvec *lruvec, struct list_head *dst,
fe2c2a10 1367 unsigned long *nr_scanned, struct scan_control *sc,
3cb99451 1368 isolate_mode_t mode, enum lru_list lru)
1da177e4 1369{
75b00af7 1370 struct list_head *src = &lruvec->lists[lru];
69e05944 1371 unsigned long nr_taken = 0;
c9b02d97 1372 unsigned long scan;
1da177e4 1373
0b802f10
VD
1374 for (scan = 0; scan < nr_to_scan && nr_taken < nr_to_scan &&
1375 !list_empty(src); scan++) {
5ad333eb 1376 struct page *page;
5ad333eb 1377
1da177e4
LT
1378 page = lru_to_page(src);
1379 prefetchw_prev_lru_page(page, src, flags);
1380
309381fe 1381 VM_BUG_ON_PAGE(!PageLRU(page), page);
8d438f96 1382
f3fd4a61 1383 switch (__isolate_lru_page(page, mode)) {
5ad333eb 1384 case 0:
9d5e6a9f 1385 nr_taken += hpage_nr_pages(page);
5ad333eb 1386 list_move(&page->lru, dst);
5ad333eb
AW
1387 break;
1388
1389 case -EBUSY:
1390 /* else it is being freed elsewhere */
1391 list_move(&page->lru, src);
1392 continue;
46453a6e 1393
5ad333eb
AW
1394 default:
1395 BUG();
1396 }
1da177e4
LT
1397 }
1398
f626012d 1399 *nr_scanned = scan;
75b00af7
HD
1400 trace_mm_vmscan_lru_isolate(sc->order, nr_to_scan, scan,
1401 nr_taken, mode, is_file_lru(lru));
1da177e4
LT
1402 return nr_taken;
1403}
1404
62695a84
NP
1405/**
1406 * isolate_lru_page - tries to isolate a page from its LRU list
1407 * @page: page to isolate from its LRU list
1408 *
1409 * Isolates a @page from an LRU list, clears PageLRU and adjusts the
1410 * vmstat statistic corresponding to whatever LRU list the page was on.
1411 *
1412 * Returns 0 if the page was removed from an LRU list.
1413 * Returns -EBUSY if the page was not on an LRU list.
1414 *
1415 * The returned page will have PageLRU() cleared. If it was found on
894bc310
LS
1416 * the active list, it will have PageActive set. If it was found on
1417 * the unevictable list, it will have the PageUnevictable bit set. That flag
1418 * may need to be cleared by the caller before letting the page go.
62695a84
NP
1419 *
1420 * The vmstat statistic corresponding to the list on which the page was
1421 * found will be decremented.
1422 *
1423 * Restrictions:
1424 * (1) Must be called with an elevated refcount on the page. This is a
1425 * fundamentnal difference from isolate_lru_pages (which is called
1426 * without a stable reference).
1427 * (2) the lru_lock must not be held.
1428 * (3) interrupts must be enabled.
1429 */
1430int isolate_lru_page(struct page *page)
1431{
1432 int ret = -EBUSY;
1433
309381fe 1434 VM_BUG_ON_PAGE(!page_count(page), page);
cf2a82ee 1435 WARN_RATELIMIT(PageTail(page), "trying to isolate tail page");
0c917313 1436
62695a84
NP
1437 if (PageLRU(page)) {
1438 struct zone *zone = page_zone(page);
fa9add64 1439 struct lruvec *lruvec;
62695a84
NP
1440
1441 spin_lock_irq(&zone->lru_lock);
fa9add64 1442 lruvec = mem_cgroup_page_lruvec(page, zone);
0c917313 1443 if (PageLRU(page)) {
894bc310 1444 int lru = page_lru(page);
0c917313 1445 get_page(page);
62695a84 1446 ClearPageLRU(page);
fa9add64
HD
1447 del_page_from_lru_list(page, lruvec, lru);
1448 ret = 0;
62695a84
NP
1449 }
1450 spin_unlock_irq(&zone->lru_lock);
1451 }
1452 return ret;
1453}
1454
35cd7815 1455/*
d37dd5dc
FW
1456 * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and
1457 * then get resheduled. When there are massive number of tasks doing page
1458 * allocation, such sleeping direct reclaimers may keep piling up on each CPU,
1459 * the LRU list will go small and be scanned faster than necessary, leading to
1460 * unnecessary swapping, thrashing and OOM.
35cd7815
RR
1461 */
1462static int too_many_isolated(struct zone *zone, int file,
1463 struct scan_control *sc)
1464{
1465 unsigned long inactive, isolated;
1466
1467 if (current_is_kswapd())
1468 return 0;
1469
97c9341f 1470 if (!sane_reclaim(sc))
35cd7815
RR
1471 return 0;
1472
1473 if (file) {
1474 inactive = zone_page_state(zone, NR_INACTIVE_FILE);
1475 isolated = zone_page_state(zone, NR_ISOLATED_FILE);
1476 } else {
1477 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1478 isolated = zone_page_state(zone, NR_ISOLATED_ANON);
1479 }
1480
3cf23841
FW
1481 /*
1482 * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they
1483 * won't get blocked by normal direct-reclaimers, forming a circular
1484 * deadlock.
1485 */
d0164adc 1486 if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS))
3cf23841
FW
1487 inactive >>= 3;
1488
35cd7815
RR
1489 return isolated > inactive;
1490}
1491
66635629 1492static noinline_for_stack void
75b00af7 1493putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list)
66635629 1494{
27ac81d8
KK
1495 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1496 struct zone *zone = lruvec_zone(lruvec);
3f79768f 1497 LIST_HEAD(pages_to_free);
66635629 1498
66635629
MG
1499 /*
1500 * Put back any unfreeable pages.
1501 */
66635629 1502 while (!list_empty(page_list)) {
3f79768f 1503 struct page *page = lru_to_page(page_list);
66635629 1504 int lru;
3f79768f 1505
309381fe 1506 VM_BUG_ON_PAGE(PageLRU(page), page);
66635629 1507 list_del(&page->lru);
39b5f29a 1508 if (unlikely(!page_evictable(page))) {
66635629
MG
1509 spin_unlock_irq(&zone->lru_lock);
1510 putback_lru_page(page);
1511 spin_lock_irq(&zone->lru_lock);
1512 continue;
1513 }
fa9add64
HD
1514
1515 lruvec = mem_cgroup_page_lruvec(page, zone);
1516
7a608572 1517 SetPageLRU(page);
66635629 1518 lru = page_lru(page);
fa9add64
HD
1519 add_page_to_lru_list(page, lruvec, lru);
1520
66635629
MG
1521 if (is_active_lru(lru)) {
1522 int file = is_file_lru(lru);
9992af10
RR
1523 int numpages = hpage_nr_pages(page);
1524 reclaim_stat->recent_rotated[file] += numpages;
66635629 1525 }
2bcf8879
HD
1526 if (put_page_testzero(page)) {
1527 __ClearPageLRU(page);
1528 __ClearPageActive(page);
fa9add64 1529 del_page_from_lru_list(page, lruvec, lru);
2bcf8879
HD
1530
1531 if (unlikely(PageCompound(page))) {
1532 spin_unlock_irq(&zone->lru_lock);
747db954 1533 mem_cgroup_uncharge(page);
2bcf8879
HD
1534 (*get_compound_page_dtor(page))(page);
1535 spin_lock_irq(&zone->lru_lock);
1536 } else
1537 list_add(&page->lru, &pages_to_free);
66635629
MG
1538 }
1539 }
66635629 1540
3f79768f
HD
1541 /*
1542 * To save our caller's stack, now use input list for pages to free.
1543 */
1544 list_splice(&pages_to_free, page_list);
66635629
MG
1545}
1546
399ba0b9
N
1547/*
1548 * If a kernel thread (such as nfsd for loop-back mounts) services
1549 * a backing device by writing to the page cache it sets PF_LESS_THROTTLE.
1550 * In that case we should only throttle if the backing device it is
1551 * writing to is congested. In other cases it is safe to throttle.
1552 */
1553static int current_may_throttle(void)
1554{
1555 return !(current->flags & PF_LESS_THROTTLE) ||
1556 current->backing_dev_info == NULL ||
1557 bdi_write_congested(current->backing_dev_info);
1558}
1559
1da177e4 1560/*
1742f19f
AM
1561 * shrink_inactive_list() is a helper for shrink_zone(). It returns the number
1562 * of reclaimed pages
1da177e4 1563 */
66635629 1564static noinline_for_stack unsigned long
1a93be0e 1565shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec,
9e3b2f8c 1566 struct scan_control *sc, enum lru_list lru)
1da177e4
LT
1567{
1568 LIST_HEAD(page_list);
e247dbce 1569 unsigned long nr_scanned;
05ff5137 1570 unsigned long nr_reclaimed = 0;
e247dbce 1571 unsigned long nr_taken;
8e950282
MG
1572 unsigned long nr_dirty = 0;
1573 unsigned long nr_congested = 0;
e2be15f6 1574 unsigned long nr_unqueued_dirty = 0;
92df3a72 1575 unsigned long nr_writeback = 0;
b1a6f21e 1576 unsigned long nr_immediate = 0;
f3fd4a61 1577 isolate_mode_t isolate_mode = 0;
3cb99451 1578 int file = is_file_lru(lru);
1a93be0e
KK
1579 struct zone *zone = lruvec_zone(lruvec);
1580 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
78dc583d 1581
35cd7815 1582 while (unlikely(too_many_isolated(zone, file, sc))) {
58355c78 1583 congestion_wait(BLK_RW_ASYNC, HZ/10);
35cd7815
RR
1584
1585 /* We are about to die and free our memory. Return now. */
1586 if (fatal_signal_pending(current))
1587 return SWAP_CLUSTER_MAX;
1588 }
1589
1da177e4 1590 lru_add_drain();
f80c0673
MK
1591
1592 if (!sc->may_unmap)
61317289 1593 isolate_mode |= ISOLATE_UNMAPPED;
f80c0673 1594 if (!sc->may_writepage)
61317289 1595 isolate_mode |= ISOLATE_CLEAN;
f80c0673 1596
1da177e4 1597 spin_lock_irq(&zone->lru_lock);
b35ea17b 1598
5dc35979
KK
1599 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list,
1600 &nr_scanned, sc, isolate_mode, lru);
95d918fc 1601
9d5e6a9f 1602 update_lru_size(lruvec, lru, -nr_taken);
95d918fc 1603 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
9d5e6a9f 1604 reclaim_stat->recent_scanned[file] += nr_taken;
95d918fc 1605
89b5fae5 1606 if (global_reclaim(sc)) {
0d5d823a 1607 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
e247dbce 1608 if (current_is_kswapd())
75b00af7 1609 __count_zone_vm_events(PGSCAN_KSWAPD, zone, nr_scanned);
e247dbce 1610 else
75b00af7 1611 __count_zone_vm_events(PGSCAN_DIRECT, zone, nr_scanned);
e247dbce 1612 }
d563c050 1613 spin_unlock_irq(&zone->lru_lock);
b35ea17b 1614
d563c050 1615 if (nr_taken == 0)
66635629 1616 return 0;
5ad333eb 1617
02c6de8d 1618 nr_reclaimed = shrink_page_list(&page_list, zone, sc, TTU_UNMAP,
8e950282
MG
1619 &nr_dirty, &nr_unqueued_dirty, &nr_congested,
1620 &nr_writeback, &nr_immediate,
1621 false);
c661b078 1622
3f79768f
HD
1623 spin_lock_irq(&zone->lru_lock);
1624
904249aa
YH
1625 if (global_reclaim(sc)) {
1626 if (current_is_kswapd())
1627 __count_zone_vm_events(PGSTEAL_KSWAPD, zone,
1628 nr_reclaimed);
1629 else
1630 __count_zone_vm_events(PGSTEAL_DIRECT, zone,
1631 nr_reclaimed);
1632 }
a74609fa 1633
27ac81d8 1634 putback_inactive_pages(lruvec, &page_list);
3f79768f 1635
95d918fc 1636 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
3f79768f
HD
1637
1638 spin_unlock_irq(&zone->lru_lock);
1639
747db954 1640 mem_cgroup_uncharge_list(&page_list);
b745bc85 1641 free_hot_cold_page_list(&page_list, true);
e11da5b4 1642
92df3a72
MG
1643 /*
1644 * If reclaim is isolating dirty pages under writeback, it implies
1645 * that the long-lived page allocation rate is exceeding the page
1646 * laundering rate. Either the global limits are not being effective
1647 * at throttling processes due to the page distribution throughout
1648 * zones or there is heavy usage of a slow backing device. The
1649 * only option is to throttle from reclaim context which is not ideal
1650 * as there is no guarantee the dirtying process is throttled in the
1651 * same way balance_dirty_pages() manages.
1652 *
8e950282
MG
1653 * Once a zone is flagged ZONE_WRITEBACK, kswapd will count the number
1654 * of pages under pages flagged for immediate reclaim and stall if any
1655 * are encountered in the nr_immediate check below.
92df3a72 1656 */
918fc718 1657 if (nr_writeback && nr_writeback == nr_taken)
57054651 1658 set_bit(ZONE_WRITEBACK, &zone->flags);
92df3a72 1659
d43006d5 1660 /*
97c9341f
TH
1661 * Legacy memcg will stall in page writeback so avoid forcibly
1662 * stalling here.
d43006d5 1663 */
97c9341f 1664 if (sane_reclaim(sc)) {
8e950282
MG
1665 /*
1666 * Tag a zone as congested if all the dirty pages scanned were
1667 * backed by a congested BDI and wait_iff_congested will stall.
1668 */
1669 if (nr_dirty && nr_dirty == nr_congested)
57054651 1670 set_bit(ZONE_CONGESTED, &zone->flags);
8e950282 1671
b1a6f21e
MG
1672 /*
1673 * If dirty pages are scanned that are not queued for IO, it
1674 * implies that flushers are not keeping up. In this case, flag
57054651
JW
1675 * the zone ZONE_DIRTY and kswapd will start writing pages from
1676 * reclaim context.
b1a6f21e
MG
1677 */
1678 if (nr_unqueued_dirty == nr_taken)
57054651 1679 set_bit(ZONE_DIRTY, &zone->flags);
b1a6f21e
MG
1680
1681 /*
b738d764
LT
1682 * If kswapd scans pages marked marked for immediate
1683 * reclaim and under writeback (nr_immediate), it implies
1684 * that pages are cycling through the LRU faster than
b1a6f21e
MG
1685 * they are written so also forcibly stall.
1686 */
b738d764 1687 if (nr_immediate && current_may_throttle())
b1a6f21e 1688 congestion_wait(BLK_RW_ASYNC, HZ/10);
e2be15f6 1689 }
d43006d5 1690
8e950282
MG
1691 /*
1692 * Stall direct reclaim for IO completions if underlying BDIs or zone
1693 * is congested. Allow kswapd to continue until it starts encountering
1694 * unqueued dirty pages or cycling through the LRU too quickly.
1695 */
399ba0b9
N
1696 if (!sc->hibernation_mode && !current_is_kswapd() &&
1697 current_may_throttle())
8e950282
MG
1698 wait_iff_congested(zone, BLK_RW_ASYNC, HZ/10);
1699
ba5e9579 1700 trace_mm_vmscan_lru_shrink_inactive(zone, nr_scanned, nr_reclaimed,
1701 sc->priority, file);
05ff5137 1702 return nr_reclaimed;
1da177e4
LT
1703}
1704
1705/*
1706 * This moves pages from the active list to the inactive list.
1707 *
1708 * We move them the other way if the page is referenced by one or more
1709 * processes, from rmap.
1710 *
1711 * If the pages are mostly unmapped, the processing is fast and it is
1712 * appropriate to hold zone->lru_lock across the whole operation. But if
1713 * the pages are mapped, the processing is slow (page_referenced()) so we
1714 * should drop zone->lru_lock around each page. It's impossible to balance
1715 * this, so instead we remove the pages from the LRU while processing them.
1716 * It is safe to rely on PG_active against the non-LRU pages in here because
1717 * nobody will play with that bit on a non-LRU page.
1718 *
0139aa7b 1719 * The downside is that we have to touch page->_refcount against each page.
1da177e4
LT
1720 * But we had to alter page->flags anyway.
1721 */
1cfb419b 1722
fa9add64 1723static void move_active_pages_to_lru(struct lruvec *lruvec,
3eb4140f 1724 struct list_head *list,
2bcf8879 1725 struct list_head *pages_to_free,
3eb4140f
WF
1726 enum lru_list lru)
1727{
fa9add64 1728 struct zone *zone = lruvec_zone(lruvec);
3eb4140f 1729 unsigned long pgmoved = 0;
3eb4140f 1730 struct page *page;
fa9add64 1731 int nr_pages;
3eb4140f 1732
3eb4140f
WF
1733 while (!list_empty(list)) {
1734 page = lru_to_page(list);
fa9add64 1735 lruvec = mem_cgroup_page_lruvec(page, zone);
3eb4140f 1736
309381fe 1737 VM_BUG_ON_PAGE(PageLRU(page), page);
3eb4140f
WF
1738 SetPageLRU(page);
1739
fa9add64 1740 nr_pages = hpage_nr_pages(page);
9d5e6a9f 1741 update_lru_size(lruvec, lru, nr_pages);
925b7673 1742 list_move(&page->lru, &lruvec->lists[lru]);
fa9add64 1743 pgmoved += nr_pages;
3eb4140f 1744
2bcf8879
HD
1745 if (put_page_testzero(page)) {
1746 __ClearPageLRU(page);
1747 __ClearPageActive(page);
fa9add64 1748 del_page_from_lru_list(page, lruvec, lru);
2bcf8879
HD
1749
1750 if (unlikely(PageCompound(page))) {
1751 spin_unlock_irq(&zone->lru_lock);
747db954 1752 mem_cgroup_uncharge(page);
2bcf8879
HD
1753 (*get_compound_page_dtor(page))(page);
1754 spin_lock_irq(&zone->lru_lock);
1755 } else
1756 list_add(&page->lru, pages_to_free);
3eb4140f
WF
1757 }
1758 }
9d5e6a9f 1759
3eb4140f
WF
1760 if (!is_active_lru(lru))
1761 __count_vm_events(PGDEACTIVATE, pgmoved);
1762}
1cfb419b 1763
f626012d 1764static void shrink_active_list(unsigned long nr_to_scan,
1a93be0e 1765 struct lruvec *lruvec,
f16015fb 1766 struct scan_control *sc,
9e3b2f8c 1767 enum lru_list lru)
1da177e4 1768{
44c241f1 1769 unsigned long nr_taken;
f626012d 1770 unsigned long nr_scanned;
6fe6b7e3 1771 unsigned long vm_flags;
1da177e4 1772 LIST_HEAD(l_hold); /* The pages which were snipped off */
8cab4754 1773 LIST_HEAD(l_active);
b69408e8 1774 LIST_HEAD(l_inactive);
1da177e4 1775 struct page *page;
1a93be0e 1776 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
44c241f1 1777 unsigned long nr_rotated = 0;
f3fd4a61 1778 isolate_mode_t isolate_mode = 0;
3cb99451 1779 int file = is_file_lru(lru);
1a93be0e 1780 struct zone *zone = lruvec_zone(lruvec);
1da177e4
LT
1781
1782 lru_add_drain();
f80c0673
MK
1783
1784 if (!sc->may_unmap)
61317289 1785 isolate_mode |= ISOLATE_UNMAPPED;
f80c0673 1786 if (!sc->may_writepage)
61317289 1787 isolate_mode |= ISOLATE_CLEAN;
f80c0673 1788
1da177e4 1789 spin_lock_irq(&zone->lru_lock);
925b7673 1790
5dc35979
KK
1791 nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold,
1792 &nr_scanned, sc, isolate_mode, lru);
89b5fae5 1793
9d5e6a9f
HD
1794 update_lru_size(lruvec, lru, -nr_taken);
1795 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, nr_taken);
b7c46d15 1796 reclaim_stat->recent_scanned[file] += nr_taken;
1cfb419b 1797
9d5e6a9f
HD
1798 if (global_reclaim(sc))
1799 __mod_zone_page_state(zone, NR_PAGES_SCANNED, nr_scanned);
f626012d 1800 __count_zone_vm_events(PGREFILL, zone, nr_scanned);
9d5e6a9f 1801
1da177e4
LT
1802 spin_unlock_irq(&zone->lru_lock);
1803
1da177e4
LT
1804 while (!list_empty(&l_hold)) {
1805 cond_resched();
1806 page = lru_to_page(&l_hold);
1807 list_del(&page->lru);
7e9cd484 1808
39b5f29a 1809 if (unlikely(!page_evictable(page))) {
894bc310
LS
1810 putback_lru_page(page);
1811 continue;
1812 }
1813
cc715d99
MG
1814 if (unlikely(buffer_heads_over_limit)) {
1815 if (page_has_private(page) && trylock_page(page)) {
1816 if (page_has_private(page))
1817 try_to_release_page(page, 0);
1818 unlock_page(page);
1819 }
1820 }
1821
c3ac9a8a
JW
1822 if (page_referenced(page, 0, sc->target_mem_cgroup,
1823 &vm_flags)) {
9992af10 1824 nr_rotated += hpage_nr_pages(page);
8cab4754
WF
1825 /*
1826 * Identify referenced, file-backed active pages and
1827 * give them one more trip around the active list. So
1828 * that executable code get better chances to stay in
1829 * memory under moderate memory pressure. Anon pages
1830 * are not likely to be evicted by use-once streaming
1831 * IO, plus JVM can create lots of anon VM_EXEC pages,
1832 * so we ignore them here.
1833 */
41e20983 1834 if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) {
8cab4754
WF
1835 list_add(&page->lru, &l_active);
1836 continue;
1837 }
1838 }
7e9cd484 1839
5205e56e 1840 ClearPageActive(page); /* we are de-activating */
1da177e4
LT
1841 list_add(&page->lru, &l_inactive);
1842 }
1843
b555749a 1844 /*
8cab4754 1845 * Move pages back to the lru list.
b555749a 1846 */
2a1dc509 1847 spin_lock_irq(&zone->lru_lock);
556adecb 1848 /*
8cab4754
WF
1849 * Count referenced pages from currently used mappings as rotated,
1850 * even though only some of them are actually re-activated. This
1851 * helps balance scan pressure between file and anonymous pages in
7c0db9e9 1852 * get_scan_count.
7e9cd484 1853 */
b7c46d15 1854 reclaim_stat->recent_rotated[file] += nr_rotated;
556adecb 1855
fa9add64
HD
1856 move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru);
1857 move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE);
a731286d 1858 __mod_zone_page_state(zone, NR_ISOLATED_ANON + file, -nr_taken);
f8891e5e 1859 spin_unlock_irq(&zone->lru_lock);
2bcf8879 1860
747db954 1861 mem_cgroup_uncharge_list(&l_hold);
b745bc85 1862 free_hot_cold_page_list(&l_hold, true);
1da177e4
LT
1863}
1864
74e3f3c3 1865#ifdef CONFIG_SWAP
42e2e457 1866static bool inactive_anon_is_low_global(struct zone *zone)
f89eb90e
KM
1867{
1868 unsigned long active, inactive;
1869
1870 active = zone_page_state(zone, NR_ACTIVE_ANON);
1871 inactive = zone_page_state(zone, NR_INACTIVE_ANON);
1872
29d06bbb 1873 return inactive * zone->inactive_ratio < active;
f89eb90e
KM
1874}
1875
14797e23
KM
1876/**
1877 * inactive_anon_is_low - check if anonymous pages need to be deactivated
c56d5c7d 1878 * @lruvec: LRU vector to check
14797e23
KM
1879 *
1880 * Returns true if the zone does not have enough inactive anon pages,
1881 * meaning some active anon pages need to be deactivated.
1882 */
42e2e457 1883static bool inactive_anon_is_low(struct lruvec *lruvec)
14797e23 1884{
74e3f3c3
MK
1885 /*
1886 * If we don't have swap space, anonymous page deactivation
1887 * is pointless.
1888 */
1889 if (!total_swap_pages)
42e2e457 1890 return false;
74e3f3c3 1891
c3c787e8 1892 if (!mem_cgroup_disabled())
c56d5c7d 1893 return mem_cgroup_inactive_anon_is_low(lruvec);
f16015fb 1894
c56d5c7d 1895 return inactive_anon_is_low_global(lruvec_zone(lruvec));
14797e23 1896}
74e3f3c3 1897#else
42e2e457 1898static inline bool inactive_anon_is_low(struct lruvec *lruvec)
74e3f3c3 1899{
42e2e457 1900 return false;
74e3f3c3
MK
1901}
1902#endif
14797e23 1903
56e49d21
RR
1904/**
1905 * inactive_file_is_low - check if file pages need to be deactivated
c56d5c7d 1906 * @lruvec: LRU vector to check
56e49d21
RR
1907 *
1908 * When the system is doing streaming IO, memory pressure here
1909 * ensures that active file pages get deactivated, until more
1910 * than half of the file pages are on the inactive list.
1911 *
1912 * Once we get to that situation, protect the system's working
1913 * set from being evicted by disabling active file page aging.
1914 *
1915 * This uses a different ratio than the anonymous pages, because
1916 * the page cache uses a use-once replacement algorithm.
1917 */
42e2e457 1918static bool inactive_file_is_low(struct lruvec *lruvec)
56e49d21 1919{
e3790144
JW
1920 unsigned long inactive;
1921 unsigned long active;
1922
23047a96
JW
1923 inactive = lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
1924 active = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE);
56e49d21 1925
e3790144 1926 return active > inactive;
56e49d21
RR
1927}
1928
42e2e457 1929static bool inactive_list_is_low(struct lruvec *lruvec, enum lru_list lru)
b39415b2 1930{
75b00af7 1931 if (is_file_lru(lru))
c56d5c7d 1932 return inactive_file_is_low(lruvec);
b39415b2 1933 else
c56d5c7d 1934 return inactive_anon_is_low(lruvec);
b39415b2
RR
1935}
1936
4f98a2fe 1937static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan,
1a93be0e 1938 struct lruvec *lruvec, struct scan_control *sc)
b69408e8 1939{
b39415b2 1940 if (is_active_lru(lru)) {
75b00af7 1941 if (inactive_list_is_low(lruvec, lru))
1a93be0e 1942 shrink_active_list(nr_to_scan, lruvec, sc, lru);
556adecb
RR
1943 return 0;
1944 }
1945
1a93be0e 1946 return shrink_inactive_list(nr_to_scan, lruvec, sc, lru);
4f98a2fe
RR
1947}
1948
9a265114
JW
1949enum scan_balance {
1950 SCAN_EQUAL,
1951 SCAN_FRACT,
1952 SCAN_ANON,
1953 SCAN_FILE,
1954};
1955
4f98a2fe
RR
1956/*
1957 * Determine how aggressively the anon and file LRU lists should be
1958 * scanned. The relative value of each set of LRU lists is determined
1959 * by looking at the fraction of the pages scanned we did rotate back
1960 * onto the active list instead of evict.
1961 *
be7bd59d
WL
1962 * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan
1963 * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan
4f98a2fe 1964 */
33377678 1965static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg,
6b4f7799
JW
1966 struct scan_control *sc, unsigned long *nr,
1967 unsigned long *lru_pages)
4f98a2fe 1968{
33377678 1969 int swappiness = mem_cgroup_swappiness(memcg);
9a265114
JW
1970 struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat;
1971 u64 fraction[2];
1972 u64 denominator = 0; /* gcc */
1973 struct zone *zone = lruvec_zone(lruvec);
4f98a2fe 1974 unsigned long anon_prio, file_prio;
9a265114 1975 enum scan_balance scan_balance;
0bf1457f 1976 unsigned long anon, file;
9a265114 1977 bool force_scan = false;
4f98a2fe 1978 unsigned long ap, fp;
4111304d 1979 enum lru_list lru;
6f04f48d
SS
1980 bool some_scanned;
1981 int pass;
246e87a9 1982
f11c0ca5
JW
1983 /*
1984 * If the zone or memcg is small, nr[l] can be 0. This
1985 * results in no scanning on this priority and a potential
1986 * priority drop. Global direct reclaim can go to the next
1987 * zone and tends to have no problems. Global kswapd is for
1988 * zone balancing and it needs to scan a minimum amount. When
1989 * reclaiming for a memcg, a priority drop can cause high
1990 * latencies, so it's better to scan a minimum amount there as
1991 * well.
1992 */
90cbc250
VD
1993 if (current_is_kswapd()) {
1994 if (!zone_reclaimable(zone))
1995 force_scan = true;
eb01aaab 1996 if (!mem_cgroup_online(memcg))
90cbc250
VD
1997 force_scan = true;
1998 }
89b5fae5 1999 if (!global_reclaim(sc))
a4d3e9e7 2000 force_scan = true;
76a33fc3
SL
2001
2002 /* If we have no swap space, do not bother scanning anon pages. */
d8b38438 2003 if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) {
9a265114 2004 scan_balance = SCAN_FILE;
76a33fc3
SL
2005 goto out;
2006 }
4f98a2fe 2007
10316b31
JW
2008 /*
2009 * Global reclaim will swap to prevent OOM even with no
2010 * swappiness, but memcg users want to use this knob to
2011 * disable swapping for individual groups completely when
2012 * using the memory controller's swap limit feature would be
2013 * too expensive.
2014 */
02695175 2015 if (!global_reclaim(sc) && !swappiness) {
9a265114 2016 scan_balance = SCAN_FILE;
10316b31
JW
2017 goto out;
2018 }
2019
2020 /*
2021 * Do not apply any pressure balancing cleverness when the
2022 * system is close to OOM, scan both anon and file equally
2023 * (unless the swappiness setting disagrees with swapping).
2024 */
02695175 2025 if (!sc->priority && swappiness) {
9a265114 2026 scan_balance = SCAN_EQUAL;
10316b31
JW
2027 goto out;
2028 }
2029
62376251
JW
2030 /*
2031 * Prevent the reclaimer from falling into the cache trap: as
2032 * cache pages start out inactive, every cache fault will tip
2033 * the scan balance towards the file LRU. And as the file LRU
2034 * shrinks, so does the window for rotation from references.
2035 * This means we have a runaway feedback loop where a tiny
2036 * thrashing file LRU becomes infinitely more attractive than
2037 * anon pages. Try to detect this based on file LRU size.
2038 */
2039 if (global_reclaim(sc)) {
2ab051e1
JM
2040 unsigned long zonefile;
2041 unsigned long zonefree;
2042
2043 zonefree = zone_page_state(zone, NR_FREE_PAGES);
2044 zonefile = zone_page_state(zone, NR_ACTIVE_FILE) +
2045 zone_page_state(zone, NR_INACTIVE_FILE);
62376251 2046
2ab051e1 2047 if (unlikely(zonefile + zonefree <= high_wmark_pages(zone))) {
62376251
JW
2048 scan_balance = SCAN_ANON;
2049 goto out;
2050 }
2051 }
2052
7c5bd705 2053 /*
316bda0e
VD
2054 * If there is enough inactive page cache, i.e. if the size of the
2055 * inactive list is greater than that of the active list *and* the
2056 * inactive list actually has some pages to scan on this priority, we
2057 * do not reclaim anything from the anonymous working set right now.
2058 * Without the second condition we could end up never scanning an
2059 * lruvec even if it has plenty of old anonymous pages unless the
2060 * system is under heavy pressure.
7c5bd705 2061 */
316bda0e 2062 if (!inactive_file_is_low(lruvec) &&
23047a96 2063 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE) >> sc->priority) {
9a265114 2064 scan_balance = SCAN_FILE;
7c5bd705
JW
2065 goto out;
2066 }
2067
9a265114
JW
2068 scan_balance = SCAN_FRACT;
2069
58c37f6e
KM
2070 /*
2071 * With swappiness at 100, anonymous and file have the same priority.
2072 * This scanning priority is essentially the inverse of IO cost.
2073 */
02695175 2074 anon_prio = swappiness;
75b00af7 2075 file_prio = 200 - anon_prio;
58c37f6e 2076
4f98a2fe
RR
2077 /*
2078 * OK, so we have swap space and a fair amount of page cache
2079 * pages. We use the recently rotated / recently scanned
2080 * ratios to determine how valuable each cache is.
2081 *
2082 * Because workloads change over time (and to avoid overflow)
2083 * we keep these statistics as a floating average, which ends
2084 * up weighing recent references more than old ones.
2085 *
2086 * anon in [0], file in [1]
2087 */
2ab051e1 2088
23047a96
JW
2089 anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON) +
2090 lruvec_lru_size(lruvec, LRU_INACTIVE_ANON);
2091 file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE) +
2092 lruvec_lru_size(lruvec, LRU_INACTIVE_FILE);
2ab051e1 2093
90126375 2094 spin_lock_irq(&zone->lru_lock);
6e901571 2095 if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) {
6e901571
KM
2096 reclaim_stat->recent_scanned[0] /= 2;
2097 reclaim_stat->recent_rotated[0] /= 2;
4f98a2fe
RR
2098 }
2099
6e901571 2100 if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) {
6e901571
KM
2101 reclaim_stat->recent_scanned[1] /= 2;
2102 reclaim_stat->recent_rotated[1] /= 2;
4f98a2fe
RR
2103 }
2104
4f98a2fe 2105 /*
00d8089c
RR
2106 * The amount of pressure on anon vs file pages is inversely
2107 * proportional to the fraction of recently scanned pages on
2108 * each list that were recently referenced and in active use.
4f98a2fe 2109 */
fe35004f 2110 ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1);
6e901571 2111 ap /= reclaim_stat->recent_rotated[0] + 1;
4f98a2fe 2112
fe35004f 2113 fp = file_prio * (reclaim_stat->recent_scanned[1] + 1);
6e901571 2114 fp /= reclaim_stat->recent_rotated[1] + 1;
90126375 2115 spin_unlock_irq(&zone->lru_lock);
4f98a2fe 2116
76a33fc3
SL
2117 fraction[0] = ap;
2118 fraction[1] = fp;
2119 denominator = ap + fp + 1;
2120out:
6f04f48d
SS
2121 some_scanned = false;
2122 /* Only use force_scan on second pass. */
2123 for (pass = 0; !some_scanned && pass < 2; pass++) {
6b4f7799 2124 *lru_pages = 0;
6f04f48d
SS
2125 for_each_evictable_lru(lru) {
2126 int file = is_file_lru(lru);
2127 unsigned long size;
2128 unsigned long scan;
6e08a369 2129
23047a96 2130 size = lruvec_lru_size(lruvec, lru);
6f04f48d 2131 scan = size >> sc->priority;
9a265114 2132
6f04f48d
SS
2133 if (!scan && pass && force_scan)
2134 scan = min(size, SWAP_CLUSTER_MAX);
9a265114 2135
6f04f48d
SS
2136 switch (scan_balance) {
2137 case SCAN_EQUAL:
2138 /* Scan lists relative to size */
2139 break;
2140 case SCAN_FRACT:
2141 /*
2142 * Scan types proportional to swappiness and
2143 * their relative recent reclaim efficiency.
2144 */
2145 scan = div64_u64(scan * fraction[file],
2146 denominator);
2147 break;
2148 case SCAN_FILE:
2149 case SCAN_ANON:
2150 /* Scan one type exclusively */
6b4f7799
JW
2151 if ((scan_balance == SCAN_FILE) != file) {
2152 size = 0;
6f04f48d 2153 scan = 0;
6b4f7799 2154 }
6f04f48d
SS
2155 break;
2156 default:
2157 /* Look ma, no brain */
2158 BUG();
2159 }
6b4f7799
JW
2160
2161 *lru_pages += size;
6f04f48d 2162 nr[lru] = scan;
6b4f7799 2163
9a265114 2164 /*
6f04f48d
SS
2165 * Skip the second pass and don't force_scan,
2166 * if we found something to scan.
9a265114 2167 */
6f04f48d 2168 some_scanned |= !!scan;
9a265114 2169 }
76a33fc3 2170 }
6e08a369 2171}
4f98a2fe 2172
72b252ae
MG
2173#ifdef CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH
2174static void init_tlb_ubc(void)
2175{
2176 /*
2177 * This deliberately does not clear the cpumask as it's expensive
2178 * and unnecessary. If there happens to be data in there then the
2179 * first SWAP_CLUSTER_MAX pages will send an unnecessary IPI and
2180 * then will be cleared.
2181 */
2182 current->tlb_ubc.flush_required = false;
2183}
2184#else
2185static inline void init_tlb_ubc(void)
2186{
2187}
2188#endif /* CONFIG_ARCH_WANT_BATCHED_UNMAP_TLB_FLUSH */
2189
9b4f98cd
JW
2190/*
2191 * This is a basic per-zone page freer. Used by both kswapd and direct reclaim.
2192 */
33377678
VD
2193static void shrink_zone_memcg(struct zone *zone, struct mem_cgroup *memcg,
2194 struct scan_control *sc, unsigned long *lru_pages)
9b4f98cd 2195{
33377678 2196 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
9b4f98cd 2197 unsigned long nr[NR_LRU_LISTS];
e82e0561 2198 unsigned long targets[NR_LRU_LISTS];
9b4f98cd
JW
2199 unsigned long nr_to_scan;
2200 enum lru_list lru;
2201 unsigned long nr_reclaimed = 0;
2202 unsigned long nr_to_reclaim = sc->nr_to_reclaim;
2203 struct blk_plug plug;
1a501907 2204 bool scan_adjusted;
9b4f98cd 2205
33377678 2206 get_scan_count(lruvec, memcg, sc, nr, lru_pages);
9b4f98cd 2207
e82e0561
MG
2208 /* Record the original scan target for proportional adjustments later */
2209 memcpy(targets, nr, sizeof(nr));
2210
1a501907
MG
2211 /*
2212 * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal
2213 * event that can occur when there is little memory pressure e.g.
2214 * multiple streaming readers/writers. Hence, we do not abort scanning
2215 * when the requested number of pages are reclaimed when scanning at
2216 * DEF_PRIORITY on the assumption that the fact we are direct
2217 * reclaiming implies that kswapd is not keeping up and it is best to
2218 * do a batch of work at once. For memcg reclaim one check is made to
2219 * abort proportional reclaim if either the file or anon lru has already
2220 * dropped to zero at the first pass.
2221 */
2222 scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() &&
2223 sc->priority == DEF_PRIORITY);
2224
72b252ae
MG
2225 init_tlb_ubc();
2226
9b4f98cd
JW
2227 blk_start_plug(&plug);
2228 while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] ||
2229 nr[LRU_INACTIVE_FILE]) {
e82e0561
MG
2230 unsigned long nr_anon, nr_file, percentage;
2231 unsigned long nr_scanned;
2232
9b4f98cd
JW
2233 for_each_evictable_lru(lru) {
2234 if (nr[lru]) {
2235 nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX);
2236 nr[lru] -= nr_to_scan;
2237
2238 nr_reclaimed += shrink_list(lru, nr_to_scan,
2239 lruvec, sc);
2240 }
2241 }
e82e0561
MG
2242
2243 if (nr_reclaimed < nr_to_reclaim || scan_adjusted)
2244 continue;
2245
e82e0561
MG
2246 /*
2247 * For kswapd and memcg, reclaim at least the number of pages
1a501907 2248 * requested. Ensure that the anon and file LRUs are scanned
e82e0561
MG
2249 * proportionally what was requested by get_scan_count(). We
2250 * stop reclaiming one LRU and reduce the amount scanning
2251 * proportional to the original scan target.
2252 */
2253 nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE];
2254 nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON];
2255
1a501907
MG
2256 /*
2257 * It's just vindictive to attack the larger once the smaller
2258 * has gone to zero. And given the way we stop scanning the
2259 * smaller below, this makes sure that we only make one nudge
2260 * towards proportionality once we've got nr_to_reclaim.
2261 */
2262 if (!nr_file || !nr_anon)
2263 break;
2264
e82e0561
MG
2265 if (nr_file > nr_anon) {
2266 unsigned long scan_target = targets[LRU_INACTIVE_ANON] +
2267 targets[LRU_ACTIVE_ANON] + 1;
2268 lru = LRU_BASE;
2269 percentage = nr_anon * 100 / scan_target;
2270 } else {
2271 unsigned long scan_target = targets[LRU_INACTIVE_FILE] +
2272 targets[LRU_ACTIVE_FILE] + 1;
2273 lru = LRU_FILE;
2274 percentage = nr_file * 100 / scan_target;
2275 }
2276
2277 /* Stop scanning the smaller of the LRU */
2278 nr[lru] = 0;
2279 nr[lru + LRU_ACTIVE] = 0;
2280
2281 /*
2282 * Recalculate the other LRU scan count based on its original
2283 * scan target and the percentage scanning already complete
2284 */
2285 lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE;
2286 nr_scanned = targets[lru] - nr[lru];
2287 nr[lru] = targets[lru] * (100 - percentage) / 100;
2288 nr[lru] -= min(nr[lru], nr_scanned);
2289
2290 lru += LRU_ACTIVE;
2291 nr_scanned = targets[lru] - nr[lru];
2292 nr[lru] = targets[lru] * (100 - percentage) / 100;
2293 nr[lru] -= min(nr[lru], nr_scanned);
2294
2295 scan_adjusted = true;
9b4f98cd
JW
2296 }
2297 blk_finish_plug(&plug);
2298 sc->nr_reclaimed += nr_reclaimed;
2299
2300 /*
2301 * Even if we did not try to evict anon pages at all, we want to
2302 * rebalance the anon lru active/inactive ratio.
2303 */
2304 if (inactive_anon_is_low(lruvec))
2305 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
2306 sc, LRU_ACTIVE_ANON);
2307
2308 throttle_vm_writeout(sc->gfp_mask);
2309}
2310
23b9da55 2311/* Use reclaim/compaction for costly allocs or under memory pressure */
9e3b2f8c 2312static bool in_reclaim_compaction(struct scan_control *sc)
23b9da55 2313{
d84da3f9 2314 if (IS_ENABLED(CONFIG_COMPACTION) && sc->order &&
23b9da55 2315 (sc->order > PAGE_ALLOC_COSTLY_ORDER ||
9e3b2f8c 2316 sc->priority < DEF_PRIORITY - 2))
23b9da55
MG
2317 return true;
2318
2319 return false;
2320}
2321
3e7d3449 2322/*
23b9da55
MG
2323 * Reclaim/compaction is used for high-order allocation requests. It reclaims
2324 * order-0 pages before compacting the zone. should_continue_reclaim() returns
2325 * true if more pages should be reclaimed such that when the page allocator
2326 * calls try_to_compact_zone() that it will have enough free pages to succeed.
2327 * It will give up earlier than that if there is difficulty reclaiming pages.
3e7d3449 2328 */
9b4f98cd 2329static inline bool should_continue_reclaim(struct zone *zone,
3e7d3449
MG
2330 unsigned long nr_reclaimed,
2331 unsigned long nr_scanned,
2332 struct scan_control *sc)
2333{
2334 unsigned long pages_for_compaction;
2335 unsigned long inactive_lru_pages;
2336
2337 /* If not in reclaim/compaction mode, stop */
9e3b2f8c 2338 if (!in_reclaim_compaction(sc))
3e7d3449
MG
2339 return false;
2340
2876592f
MG
2341 /* Consider stopping depending on scan and reclaim activity */
2342 if (sc->gfp_mask & __GFP_REPEAT) {
2343 /*
2344 * For __GFP_REPEAT allocations, stop reclaiming if the
2345 * full LRU list has been scanned and we are still failing
2346 * to reclaim pages. This full LRU scan is potentially
2347 * expensive but a __GFP_REPEAT caller really wants to succeed
2348 */
2349 if (!nr_reclaimed && !nr_scanned)
2350 return false;
2351 } else {
2352 /*
2353 * For non-__GFP_REPEAT allocations which can presumably
2354 * fail without consequence, stop if we failed to reclaim
2355 * any pages from the last SWAP_CLUSTER_MAX number of
2356 * pages that were scanned. This will return to the
2357 * caller faster at the risk reclaim/compaction and
2358 * the resulting allocation attempt fails
2359 */
2360 if (!nr_reclaimed)
2361 return false;
2362 }
3e7d3449
MG
2363
2364 /*
2365 * If we have not reclaimed enough pages for compaction and the
2366 * inactive lists are large enough, continue reclaiming
2367 */
2368 pages_for_compaction = (2UL << sc->order);
9b4f98cd 2369 inactive_lru_pages = zone_page_state(zone, NR_INACTIVE_FILE);
ec8acf20 2370 if (get_nr_swap_pages() > 0)
9b4f98cd 2371 inactive_lru_pages += zone_page_state(zone, NR_INACTIVE_ANON);
3e7d3449
MG
2372 if (sc->nr_reclaimed < pages_for_compaction &&
2373 inactive_lru_pages > pages_for_compaction)
2374 return true;
2375
2376 /* If compaction would go ahead or the allocation would succeed, stop */
ebff3980 2377 switch (compaction_suitable(zone, sc->order, 0, 0)) {
3e7d3449
MG
2378 case COMPACT_PARTIAL:
2379 case COMPACT_CONTINUE:
2380 return false;
2381 default:
2382 return true;
2383 }
2384}
2385
6b4f7799
JW
2386static bool shrink_zone(struct zone *zone, struct scan_control *sc,
2387 bool is_classzone)
1da177e4 2388{
cb731d6c 2389 struct reclaim_state *reclaim_state = current->reclaim_state;
f0fdc5e8 2390 unsigned long nr_reclaimed, nr_scanned;
2344d7e4 2391 bool reclaimable = false;
1da177e4 2392
9b4f98cd
JW
2393 do {
2394 struct mem_cgroup *root = sc->target_mem_cgroup;
2395 struct mem_cgroup_reclaim_cookie reclaim = {
2396 .zone = zone,
2397 .priority = sc->priority,
2398 };
6b4f7799 2399 unsigned long zone_lru_pages = 0;
694fbc0f 2400 struct mem_cgroup *memcg;
3e7d3449 2401
9b4f98cd
JW
2402 nr_reclaimed = sc->nr_reclaimed;
2403 nr_scanned = sc->nr_scanned;
1da177e4 2404
694fbc0f
AM
2405 memcg = mem_cgroup_iter(root, NULL, &reclaim);
2406 do {
6b4f7799 2407 unsigned long lru_pages;
8e8ae645 2408 unsigned long reclaimed;
cb731d6c 2409 unsigned long scanned;
5660048c 2410
241994ed
JW
2411 if (mem_cgroup_low(root, memcg)) {
2412 if (!sc->may_thrash)
2413 continue;
2414 mem_cgroup_events(memcg, MEMCG_LOW, 1);
2415 }
2416
8e8ae645 2417 reclaimed = sc->nr_reclaimed;
cb731d6c 2418 scanned = sc->nr_scanned;
f9be23d6 2419
33377678 2420 shrink_zone_memcg(zone, memcg, sc, &lru_pages);
6b4f7799 2421 zone_lru_pages += lru_pages;
f16015fb 2422
cb731d6c
VD
2423 if (memcg && is_classzone)
2424 shrink_slab(sc->gfp_mask, zone_to_nid(zone),
2425 memcg, sc->nr_scanned - scanned,
2426 lru_pages);
2427
8e8ae645
JW
2428 /* Record the group's reclaim efficiency */
2429 vmpressure(sc->gfp_mask, memcg, false,
2430 sc->nr_scanned - scanned,
2431 sc->nr_reclaimed - reclaimed);
2432
9b4f98cd 2433 /*
a394cb8e
MH
2434 * Direct reclaim and kswapd have to scan all memory
2435 * cgroups to fulfill the overall scan target for the
9b4f98cd 2436 * zone.
a394cb8e
MH
2437 *
2438 * Limit reclaim, on the other hand, only cares about
2439 * nr_to_reclaim pages to be reclaimed and it will
2440 * retry with decreasing priority if one round over the
2441 * whole hierarchy is not sufficient.
9b4f98cd 2442 */
a394cb8e
MH
2443 if (!global_reclaim(sc) &&
2444 sc->nr_reclaimed >= sc->nr_to_reclaim) {
9b4f98cd
JW
2445 mem_cgroup_iter_break(root, memcg);
2446 break;
2447 }
241994ed 2448 } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim)));
70ddf637 2449
6b4f7799
JW
2450 /*
2451 * Shrink the slab caches in the same proportion that
2452 * the eligible LRU pages were scanned.
2453 */
cb731d6c
VD
2454 if (global_reclaim(sc) && is_classzone)
2455 shrink_slab(sc->gfp_mask, zone_to_nid(zone), NULL,
2456 sc->nr_scanned - nr_scanned,
2457 zone_lru_pages);
2458
2459 if (reclaim_state) {
2460 sc->nr_reclaimed += reclaim_state->reclaimed_slab;
2461 reclaim_state->reclaimed_slab = 0;
6b4f7799
JW
2462 }
2463
8e8ae645
JW
2464 /* Record the subtree's reclaim efficiency */
2465 vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true,
70ddf637
AV
2466 sc->nr_scanned - nr_scanned,
2467 sc->nr_reclaimed - nr_reclaimed);
2468
2344d7e4
JW
2469 if (sc->nr_reclaimed - nr_reclaimed)
2470 reclaimable = true;
2471
9b4f98cd
JW
2472 } while (should_continue_reclaim(zone, sc->nr_reclaimed - nr_reclaimed,
2473 sc->nr_scanned - nr_scanned, sc));
2344d7e4
JW
2474
2475 return reclaimable;
f16015fb
JW
2476}
2477
53853e2d
VB
2478/*
2479 * Returns true if compaction should go ahead for a high-order request, or
2480 * the high-order allocation would succeed without compaction.
2481 */
0b06496a 2482static inline bool compaction_ready(struct zone *zone, int order)
fe4b1b24
MG
2483{
2484 unsigned long balance_gap, watermark;
2485 bool watermark_ok;
2486
fe4b1b24
MG
2487 /*
2488 * Compaction takes time to run and there are potentially other
2489 * callers using the pages just freed. Continue reclaiming until
2490 * there is a buffer of free pages available to give compaction
2491 * a reasonable chance of completing and allocating the page
2492 */
4be89a34
JZ
2493 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
2494 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
0b06496a 2495 watermark = high_wmark_pages(zone) + balance_gap + (2UL << order);
e2b19197 2496 watermark_ok = zone_watermark_ok_safe(zone, 0, watermark, 0);
fe4b1b24
MG
2497
2498 /*
2499 * If compaction is deferred, reclaim up to a point where
2500 * compaction will have a chance of success when re-enabled
2501 */
0b06496a 2502 if (compaction_deferred(zone, order))
fe4b1b24
MG
2503 return watermark_ok;
2504
53853e2d
VB
2505 /*
2506 * If compaction is not ready to start and allocation is not likely
2507 * to succeed without it, then keep reclaiming.
2508 */
ebff3980 2509 if (compaction_suitable(zone, order, 0, 0) == COMPACT_SKIPPED)
fe4b1b24
MG
2510 return false;
2511
2512 return watermark_ok;
2513}
2514
1da177e4
LT
2515/*
2516 * This is the direct reclaim path, for page-allocating processes. We only
2517 * try to reclaim pages from zones which will satisfy the caller's allocation
2518 * request.
2519 *
41858966
MG
2520 * We reclaim from a zone even if that zone is over high_wmark_pages(zone).
2521 * Because:
1da177e4
LT
2522 * a) The caller may be trying to free *extra* pages to satisfy a higher-order
2523 * allocation or
41858966
MG
2524 * b) The target zone may be at high_wmark_pages(zone) but the lower zones
2525 * must go *over* high_wmark_pages(zone) to satisfy the `incremental min'
2526 * zone defense algorithm.
1da177e4 2527 *
1da177e4
LT
2528 * If a zone is deemed to be full of pinned pages then just give it a light
2529 * scan then give up on it.
2344d7e4
JW
2530 *
2531 * Returns true if a zone was reclaimable.
1da177e4 2532 */
2344d7e4 2533static bool shrink_zones(struct zonelist *zonelist, struct scan_control *sc)
1da177e4 2534{
dd1a239f 2535 struct zoneref *z;
54a6eb5c 2536 struct zone *zone;
0608f43d
AM
2537 unsigned long nr_soft_reclaimed;
2538 unsigned long nr_soft_scanned;
619d0d76 2539 gfp_t orig_mask;
9bbc04ee 2540 enum zone_type requested_highidx = gfp_zone(sc->gfp_mask);
2344d7e4 2541 bool reclaimable = false;
1cfb419b 2542
cc715d99
MG
2543 /*
2544 * If the number of buffer_heads in the machine exceeds the maximum
2545 * allowed level, force direct reclaim to scan the highmem zone as
2546 * highmem pages could be pinning lowmem pages storing buffer_heads
2547 */
619d0d76 2548 orig_mask = sc->gfp_mask;
cc715d99
MG
2549 if (buffer_heads_over_limit)
2550 sc->gfp_mask |= __GFP_HIGHMEM;
2551
d4debc66 2552 for_each_zone_zonelist_nodemask(zone, z, zonelist,
7bf52fb8 2553 gfp_zone(sc->gfp_mask), sc->nodemask) {
6b4f7799
JW
2554 enum zone_type classzone_idx;
2555
f3fe6512 2556 if (!populated_zone(zone))
1da177e4 2557 continue;
6b4f7799
JW
2558
2559 classzone_idx = requested_highidx;
2560 while (!populated_zone(zone->zone_pgdat->node_zones +
2561 classzone_idx))
2562 classzone_idx--;
2563
1cfb419b
KH
2564 /*
2565 * Take care memory controller reclaiming has small influence
2566 * to global LRU.
2567 */
89b5fae5 2568 if (global_reclaim(sc)) {
344736f2
VD
2569 if (!cpuset_zone_allowed(zone,
2570 GFP_KERNEL | __GFP_HARDWALL))
1cfb419b 2571 continue;
65ec02cb 2572
6e543d57
LD
2573 if (sc->priority != DEF_PRIORITY &&
2574 !zone_reclaimable(zone))
1cfb419b 2575 continue; /* Let kswapd poll it */
0b06496a
JW
2576
2577 /*
2578 * If we already have plenty of memory free for
2579 * compaction in this zone, don't free any more.
2580 * Even though compaction is invoked for any
2581 * non-zero order, only frequent costly order
2582 * reclamation is disruptive enough to become a
2583 * noticeable problem, like transparent huge
2584 * page allocations.
2585 */
2586 if (IS_ENABLED(CONFIG_COMPACTION) &&
2587 sc->order > PAGE_ALLOC_COSTLY_ORDER &&
2588 zonelist_zone_idx(z) <= requested_highidx &&
2589 compaction_ready(zone, sc->order)) {
2590 sc->compaction_ready = true;
2591 continue;
e0887c19 2592 }
0b06496a 2593
0608f43d
AM
2594 /*
2595 * This steals pages from memory cgroups over softlimit
2596 * and returns the number of reclaimed pages and
2597 * scanned pages. This works for global memory pressure
2598 * and balancing, not for a memcg's limit.
2599 */
2600 nr_soft_scanned = 0;
2601 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
2602 sc->order, sc->gfp_mask,
2603 &nr_soft_scanned);
2604 sc->nr_reclaimed += nr_soft_reclaimed;
2605 sc->nr_scanned += nr_soft_scanned;
2344d7e4
JW
2606 if (nr_soft_reclaimed)
2607 reclaimable = true;
ac34a1a3 2608 /* need some check for avoid more shrink_zone() */
1cfb419b 2609 }
408d8544 2610
6b4f7799 2611 if (shrink_zone(zone, sc, zone_idx(zone) == classzone_idx))
2344d7e4
JW
2612 reclaimable = true;
2613
2614 if (global_reclaim(sc) &&
2615 !reclaimable && zone_reclaimable(zone))
2616 reclaimable = true;
1da177e4 2617 }
e0c23279 2618
619d0d76
WY
2619 /*
2620 * Restore to original mask to avoid the impact on the caller if we
2621 * promoted it to __GFP_HIGHMEM.
2622 */
2623 sc->gfp_mask = orig_mask;
d1908362 2624
2344d7e4 2625 return reclaimable;
1da177e4 2626}
4f98a2fe 2627
1da177e4
LT
2628/*
2629 * This is the main entry point to direct page reclaim.
2630 *
2631 * If a full scan of the inactive list fails to free enough memory then we
2632 * are "out of memory" and something needs to be killed.
2633 *
2634 * If the caller is !__GFP_FS then the probability of a failure is reasonably
2635 * high - the zone may be full of dirty or under-writeback pages, which this
5b0830cb
JA
2636 * caller can't do much about. We kick the writeback threads and take explicit
2637 * naps in the hope that some of these pages can be written. But if the
2638 * allocating task holds filesystem locks which prevent writeout this might not
2639 * work, and the allocation attempt will fail.
a41f24ea
NA
2640 *
2641 * returns: 0, if no pages reclaimed
2642 * else, the number of pages reclaimed
1da177e4 2643 */
dac1d27b 2644static unsigned long do_try_to_free_pages(struct zonelist *zonelist,
3115cd91 2645 struct scan_control *sc)
1da177e4 2646{
241994ed 2647 int initial_priority = sc->priority;
69e05944 2648 unsigned long total_scanned = 0;
22fba335 2649 unsigned long writeback_threshold;
2344d7e4 2650 bool zones_reclaimable;
241994ed 2651retry:
873b4771
KK
2652 delayacct_freepages_start();
2653
89b5fae5 2654 if (global_reclaim(sc))
1cfb419b 2655 count_vm_event(ALLOCSTALL);
1da177e4 2656
9e3b2f8c 2657 do {
70ddf637
AV
2658 vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup,
2659 sc->priority);
66e1707b 2660 sc->nr_scanned = 0;
2344d7e4 2661 zones_reclaimable = shrink_zones(zonelist, sc);
c6a8a8c5 2662
66e1707b 2663 total_scanned += sc->nr_scanned;
bb21c7ce 2664 if (sc->nr_reclaimed >= sc->nr_to_reclaim)
0b06496a
JW
2665 break;
2666
2667 if (sc->compaction_ready)
2668 break;
1da177e4 2669
0e50ce3b
MK
2670 /*
2671 * If we're getting trouble reclaiming, start doing
2672 * writepage even in laptop mode.
2673 */
2674 if (sc->priority < DEF_PRIORITY - 2)
2675 sc->may_writepage = 1;
2676
1da177e4
LT
2677 /*
2678 * Try to write back as many pages as we just scanned. This
2679 * tends to cause slow streaming writers to write data to the
2680 * disk smoothly, at the dirtying rate, which is nice. But
2681 * that's undesirable in laptop mode, where we *want* lumpy
2682 * writeout. So in laptop mode, write out the whole world.
2683 */
22fba335
KM
2684 writeback_threshold = sc->nr_to_reclaim + sc->nr_to_reclaim / 2;
2685 if (total_scanned > writeback_threshold) {
0e175a18
CW
2686 wakeup_flusher_threads(laptop_mode ? 0 : total_scanned,
2687 WB_REASON_TRY_TO_FREE_PAGES);
66e1707b 2688 sc->may_writepage = 1;
1da177e4 2689 }
0b06496a 2690 } while (--sc->priority >= 0);
bb21c7ce 2691
873b4771
KK
2692 delayacct_freepages_end();
2693
bb21c7ce
KM
2694 if (sc->nr_reclaimed)
2695 return sc->nr_reclaimed;
2696
0cee34fd 2697 /* Aborted reclaim to try compaction? don't OOM, then */
0b06496a 2698 if (sc->compaction_ready)
7335084d
MG
2699 return 1;
2700
241994ed
JW
2701 /* Untapped cgroup reserves? Don't OOM, retry. */
2702 if (!sc->may_thrash) {
2703 sc->priority = initial_priority;
2704 sc->may_thrash = 1;
2705 goto retry;
2706 }
2707
2344d7e4
JW
2708 /* Any of the zones still reclaimable? Don't OOM. */
2709 if (zones_reclaimable)
bb21c7ce
KM
2710 return 1;
2711
2712 return 0;
1da177e4
LT
2713}
2714
5515061d
MG
2715static bool pfmemalloc_watermark_ok(pg_data_t *pgdat)
2716{
2717 struct zone *zone;
2718 unsigned long pfmemalloc_reserve = 0;
2719 unsigned long free_pages = 0;
2720 int i;
2721 bool wmark_ok;
2722
2723 for (i = 0; i <= ZONE_NORMAL; i++) {
2724 zone = &pgdat->node_zones[i];
f012a84a
NA
2725 if (!populated_zone(zone) ||
2726 zone_reclaimable_pages(zone) == 0)
675becce
MG
2727 continue;
2728
5515061d
MG
2729 pfmemalloc_reserve += min_wmark_pages(zone);
2730 free_pages += zone_page_state(zone, NR_FREE_PAGES);
2731 }
2732
675becce
MG
2733 /* If there are no reserves (unexpected config) then do not throttle */
2734 if (!pfmemalloc_reserve)
2735 return true;
2736
5515061d
MG
2737 wmark_ok = free_pages > pfmemalloc_reserve / 2;
2738
2739 /* kswapd must be awake if processes are being throttled */
2740 if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) {
2741 pgdat->classzone_idx = min(pgdat->classzone_idx,
2742 (enum zone_type)ZONE_NORMAL);
2743 wake_up_interruptible(&pgdat->kswapd_wait);
2744 }
2745
2746 return wmark_ok;
2747}
2748
2749/*
2750 * Throttle direct reclaimers if backing storage is backed by the network
2751 * and the PFMEMALLOC reserve for the preferred node is getting dangerously
2752 * depleted. kswapd will continue to make progress and wake the processes
50694c28
MG
2753 * when the low watermark is reached.
2754 *
2755 * Returns true if a fatal signal was delivered during throttling. If this
2756 * happens, the page allocator should not consider triggering the OOM killer.
5515061d 2757 */
50694c28 2758static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist,
5515061d
MG
2759 nodemask_t *nodemask)
2760{
675becce 2761 struct zoneref *z;
5515061d 2762 struct zone *zone;
675becce 2763 pg_data_t *pgdat = NULL;
5515061d
MG
2764
2765 /*
2766 * Kernel threads should not be throttled as they may be indirectly
2767 * responsible for cleaning pages necessary for reclaim to make forward
2768 * progress. kjournald for example may enter direct reclaim while
2769 * committing a transaction where throttling it could forcing other
2770 * processes to block on log_wait_commit().
2771 */
2772 if (current->flags & PF_KTHREAD)
50694c28
MG
2773 goto out;
2774
2775 /*
2776 * If a fatal signal is pending, this process should not throttle.
2777 * It should return quickly so it can exit and free its memory
2778 */
2779 if (fatal_signal_pending(current))
2780 goto out;
5515061d 2781
675becce
MG
2782 /*
2783 * Check if the pfmemalloc reserves are ok by finding the first node
2784 * with a usable ZONE_NORMAL or lower zone. The expectation is that
2785 * GFP_KERNEL will be required for allocating network buffers when
2786 * swapping over the network so ZONE_HIGHMEM is unusable.
2787 *
2788 * Throttling is based on the first usable node and throttled processes
2789 * wait on a queue until kswapd makes progress and wakes them. There
2790 * is an affinity then between processes waking up and where reclaim
2791 * progress has been made assuming the process wakes on the same node.
2792 * More importantly, processes running on remote nodes will not compete
2793 * for remote pfmemalloc reserves and processes on different nodes
2794 * should make reasonable progress.
2795 */
2796 for_each_zone_zonelist_nodemask(zone, z, zonelist,
17636faa 2797 gfp_zone(gfp_mask), nodemask) {
675becce
MG
2798 if (zone_idx(zone) > ZONE_NORMAL)
2799 continue;
2800
2801 /* Throttle based on the first usable node */
2802 pgdat = zone->zone_pgdat;
2803 if (pfmemalloc_watermark_ok(pgdat))
2804 goto out;
2805 break;
2806 }
2807
2808 /* If no zone was usable by the allocation flags then do not throttle */
2809 if (!pgdat)
50694c28 2810 goto out;
5515061d 2811
68243e76
MG
2812 /* Account for the throttling */
2813 count_vm_event(PGSCAN_DIRECT_THROTTLE);
2814
5515061d
MG
2815 /*
2816 * If the caller cannot enter the filesystem, it's possible that it
2817 * is due to the caller holding an FS lock or performing a journal
2818 * transaction in the case of a filesystem like ext[3|4]. In this case,
2819 * it is not safe to block on pfmemalloc_wait as kswapd could be
2820 * blocked waiting on the same lock. Instead, throttle for up to a
2821 * second before continuing.
2822 */
2823 if (!(gfp_mask & __GFP_FS)) {
2824 wait_event_interruptible_timeout(pgdat->pfmemalloc_wait,
2825 pfmemalloc_watermark_ok(pgdat), HZ);
50694c28
MG
2826
2827 goto check_pending;
5515061d
MG
2828 }
2829
2830 /* Throttle until kswapd wakes the process */
2831 wait_event_killable(zone->zone_pgdat->pfmemalloc_wait,
2832 pfmemalloc_watermark_ok(pgdat));
50694c28
MG
2833
2834check_pending:
2835 if (fatal_signal_pending(current))
2836 return true;
2837
2838out:
2839 return false;
5515061d
MG
2840}
2841
dac1d27b 2842unsigned long try_to_free_pages(struct zonelist *zonelist, int order,
327c0e96 2843 gfp_t gfp_mask, nodemask_t *nodemask)
66e1707b 2844{
33906bc5 2845 unsigned long nr_reclaimed;
66e1707b 2846 struct scan_control sc = {
ee814fe2 2847 .nr_to_reclaim = SWAP_CLUSTER_MAX,
21caf2fc 2848 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
ee814fe2
JW
2849 .order = order,
2850 .nodemask = nodemask,
2851 .priority = DEF_PRIORITY,
66e1707b 2852 .may_writepage = !laptop_mode,
a6dc60f8 2853 .may_unmap = 1,
2e2e4259 2854 .may_swap = 1,
66e1707b
BS
2855 };
2856
5515061d 2857 /*
50694c28
MG
2858 * Do not enter reclaim if fatal signal was delivered while throttled.
2859 * 1 is returned so that the page allocator does not OOM kill at this
2860 * point.
5515061d 2861 */
50694c28 2862 if (throttle_direct_reclaim(gfp_mask, zonelist, nodemask))
5515061d
MG
2863 return 1;
2864
33906bc5
MG
2865 trace_mm_vmscan_direct_reclaim_begin(order,
2866 sc.may_writepage,
2867 gfp_mask);
2868
3115cd91 2869 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
33906bc5
MG
2870
2871 trace_mm_vmscan_direct_reclaim_end(nr_reclaimed);
2872
2873 return nr_reclaimed;
66e1707b
BS
2874}
2875
c255a458 2876#ifdef CONFIG_MEMCG
66e1707b 2877
72835c86 2878unsigned long mem_cgroup_shrink_node_zone(struct mem_cgroup *memcg,
4e416953 2879 gfp_t gfp_mask, bool noswap,
0ae5e89c
YH
2880 struct zone *zone,
2881 unsigned long *nr_scanned)
4e416953
BS
2882{
2883 struct scan_control sc = {
b8f5c566 2884 .nr_to_reclaim = SWAP_CLUSTER_MAX,
ee814fe2 2885 .target_mem_cgroup = memcg,
4e416953
BS
2886 .may_writepage = !laptop_mode,
2887 .may_unmap = 1,
2888 .may_swap = !noswap,
4e416953 2889 };
6b4f7799 2890 unsigned long lru_pages;
0ae5e89c 2891
4e416953
BS
2892 sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2893 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK);
bdce6d9e 2894
9e3b2f8c 2895 trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order,
bdce6d9e
KM
2896 sc.may_writepage,
2897 sc.gfp_mask);
2898
4e416953
BS
2899 /*
2900 * NOTE: Although we can get the priority field, using it
2901 * here is not a good idea, since it limits the pages we can scan.
2902 * if we don't reclaim here, the shrink_zone from balance_pgdat
2903 * will pick up pages from other mem cgroup's as well. We hack
2904 * the priority and make it zero.
2905 */
33377678 2906 shrink_zone_memcg(zone, memcg, &sc, &lru_pages);
bdce6d9e
KM
2907
2908 trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed);
2909
0ae5e89c 2910 *nr_scanned = sc.nr_scanned;
4e416953
BS
2911 return sc.nr_reclaimed;
2912}
2913
72835c86 2914unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg,
b70a2a21 2915 unsigned long nr_pages,
a7885eb8 2916 gfp_t gfp_mask,
b70a2a21 2917 bool may_swap)
66e1707b 2918{
4e416953 2919 struct zonelist *zonelist;
bdce6d9e 2920 unsigned long nr_reclaimed;
889976db 2921 int nid;
66e1707b 2922 struct scan_control sc = {
b70a2a21 2923 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
a09ed5e0
YH
2924 .gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) |
2925 (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK),
ee814fe2
JW
2926 .target_mem_cgroup = memcg,
2927 .priority = DEF_PRIORITY,
2928 .may_writepage = !laptop_mode,
2929 .may_unmap = 1,
b70a2a21 2930 .may_swap = may_swap,
a09ed5e0 2931 };
66e1707b 2932
889976db
YH
2933 /*
2934 * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't
2935 * take care of from where we get pages. So the node where we start the
2936 * scan does not need to be the current node.
2937 */
72835c86 2938 nid = mem_cgroup_select_victim_node(memcg);
889976db
YH
2939
2940 zonelist = NODE_DATA(nid)->node_zonelists;
bdce6d9e
KM
2941
2942 trace_mm_vmscan_memcg_reclaim_begin(0,
2943 sc.may_writepage,
2944 sc.gfp_mask);
2945
3115cd91 2946 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
bdce6d9e
KM
2947
2948 trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed);
2949
2950 return nr_reclaimed;
66e1707b
BS
2951}
2952#endif
2953
9e3b2f8c 2954static void age_active_anon(struct zone *zone, struct scan_control *sc)
f16015fb 2955{
b95a2f2d 2956 struct mem_cgroup *memcg;
f16015fb 2957
b95a2f2d
JW
2958 if (!total_swap_pages)
2959 return;
2960
2961 memcg = mem_cgroup_iter(NULL, NULL, NULL);
2962 do {
c56d5c7d 2963 struct lruvec *lruvec = mem_cgroup_zone_lruvec(zone, memcg);
b95a2f2d 2964
c56d5c7d 2965 if (inactive_anon_is_low(lruvec))
1a93be0e 2966 shrink_active_list(SWAP_CLUSTER_MAX, lruvec,
9e3b2f8c 2967 sc, LRU_ACTIVE_ANON);
b95a2f2d
JW
2968
2969 memcg = mem_cgroup_iter(NULL, memcg, NULL);
2970 } while (memcg);
f16015fb
JW
2971}
2972
accf6242
VB
2973static bool zone_balanced(struct zone *zone, int order, bool highorder,
2974 unsigned long balance_gap, int classzone_idx)
60cefed4 2975{
accf6242 2976 unsigned long mark = high_wmark_pages(zone) + balance_gap;
60cefed4 2977
accf6242
VB
2978 /*
2979 * When checking from pgdat_balanced(), kswapd should stop and sleep
2980 * when it reaches the high order-0 watermark and let kcompactd take
2981 * over. Other callers such as wakeup_kswapd() want to determine the
2982 * true high-order watermark.
2983 */
2984 if (IS_ENABLED(CONFIG_COMPACTION) && !highorder) {
2985 mark += (1UL << order);
2986 order = 0;
2987 }
60cefed4 2988
accf6242 2989 return zone_watermark_ok_safe(zone, order, mark, classzone_idx);
60cefed4
JW
2990}
2991
1741c877 2992/*
4ae0a48b
ZC
2993 * pgdat_balanced() is used when checking if a node is balanced.
2994 *
2995 * For order-0, all zones must be balanced!
2996 *
2997 * For high-order allocations only zones that meet watermarks and are in a
2998 * zone allowed by the callers classzone_idx are added to balanced_pages. The
2999 * total of balanced pages must be at least 25% of the zones allowed by
3000 * classzone_idx for the node to be considered balanced. Forcing all zones to
3001 * be balanced for high orders can cause excessive reclaim when there are
3002 * imbalanced zones.
1741c877
MG
3003 * The choice of 25% is due to
3004 * o a 16M DMA zone that is balanced will not balance a zone on any
3005 * reasonable sized machine
3006 * o On all other machines, the top zone must be at least a reasonable
25985edc 3007 * percentage of the middle zones. For example, on 32-bit x86, highmem
1741c877
MG
3008 * would need to be at least 256M for it to be balance a whole node.
3009 * Similarly, on x86-64 the Normal zone would need to be at least 1G
3010 * to balance a node on its own. These seemed like reasonable ratios.
3011 */
4ae0a48b 3012static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx)
1741c877 3013{
b40da049 3014 unsigned long managed_pages = 0;
4ae0a48b 3015 unsigned long balanced_pages = 0;
1741c877
MG
3016 int i;
3017
4ae0a48b
ZC
3018 /* Check the watermark levels */
3019 for (i = 0; i <= classzone_idx; i++) {
3020 struct zone *zone = pgdat->node_zones + i;
1741c877 3021
4ae0a48b
ZC
3022 if (!populated_zone(zone))
3023 continue;
3024
b40da049 3025 managed_pages += zone->managed_pages;
4ae0a48b
ZC
3026
3027 /*
3028 * A special case here:
3029 *
3030 * balance_pgdat() skips over all_unreclaimable after
3031 * DEF_PRIORITY. Effectively, it considers them balanced so
3032 * they must be considered balanced here as well!
3033 */
6e543d57 3034 if (!zone_reclaimable(zone)) {
b40da049 3035 balanced_pages += zone->managed_pages;
4ae0a48b
ZC
3036 continue;
3037 }
3038
accf6242 3039 if (zone_balanced(zone, order, false, 0, i))
b40da049 3040 balanced_pages += zone->managed_pages;
4ae0a48b
ZC
3041 else if (!order)
3042 return false;
3043 }
3044
3045 if (order)
b40da049 3046 return balanced_pages >= (managed_pages >> 2);
4ae0a48b
ZC
3047 else
3048 return true;
1741c877
MG
3049}
3050
5515061d
MG
3051/*
3052 * Prepare kswapd for sleeping. This verifies that there are no processes
3053 * waiting in throttle_direct_reclaim() and that watermarks have been met.
3054 *
3055 * Returns true if kswapd is ready to sleep
3056 */
3057static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, long remaining,
dc83edd9 3058 int classzone_idx)
f50de2d3 3059{
f50de2d3
MG
3060 /* If a direct reclaimer woke kswapd within HZ/10, it's premature */
3061 if (remaining)
5515061d
MG
3062 return false;
3063
3064 /*
9e5e3661
VB
3065 * The throttled processes are normally woken up in balance_pgdat() as
3066 * soon as pfmemalloc_watermark_ok() is true. But there is a potential
3067 * race between when kswapd checks the watermarks and a process gets
3068 * throttled. There is also a potential race if processes get
3069 * throttled, kswapd wakes, a large process exits thereby balancing the
3070 * zones, which causes kswapd to exit balance_pgdat() before reaching
3071 * the wake up checks. If kswapd is going to sleep, no process should
3072 * be sleeping on pfmemalloc_wait, so wake them now if necessary. If
3073 * the wake up is premature, processes will wake kswapd and get
3074 * throttled again. The difference from wake ups in balance_pgdat() is
3075 * that here we are under prepare_to_wait().
5515061d 3076 */
9e5e3661
VB
3077 if (waitqueue_active(&pgdat->pfmemalloc_wait))
3078 wake_up_all(&pgdat->pfmemalloc_wait);
f50de2d3 3079
4ae0a48b 3080 return pgdat_balanced(pgdat, order, classzone_idx);
f50de2d3
MG
3081}
3082
75485363
MG
3083/*
3084 * kswapd shrinks the zone by the number of pages required to reach
3085 * the high watermark.
b8e83b94
MG
3086 *
3087 * Returns true if kswapd scanned at least the requested number of pages to
283aba9f
MG
3088 * reclaim or if the lack of progress was due to pages under writeback.
3089 * This is used to determine if the scanning priority needs to be raised.
75485363 3090 */
b8e83b94 3091static bool kswapd_shrink_zone(struct zone *zone,
7c954f6d 3092 int classzone_idx,
accf6242 3093 struct scan_control *sc)
75485363 3094{
7c954f6d 3095 unsigned long balance_gap;
7c954f6d 3096 bool lowmem_pressure;
75485363
MG
3097
3098 /* Reclaim above the high watermark. */
3099 sc->nr_to_reclaim = max(SWAP_CLUSTER_MAX, high_wmark_pages(zone));
7c954f6d 3100
7c954f6d
MG
3101 /*
3102 * We put equal pressure on every zone, unless one zone has way too
3103 * many pages free already. The "too many pages" is defined as the
3104 * high wmark plus a "gap" where the gap is either the low
3105 * watermark or 1% of the zone, whichever is smaller.
3106 */
4be89a34
JZ
3107 balance_gap = min(low_wmark_pages(zone), DIV_ROUND_UP(
3108 zone->managed_pages, KSWAPD_ZONE_BALANCE_GAP_RATIO));
7c954f6d
MG
3109
3110 /*
3111 * If there is no low memory pressure or the zone is balanced then no
3112 * reclaim is necessary
3113 */
3114 lowmem_pressure = (buffer_heads_over_limit && is_highmem(zone));
accf6242 3115 if (!lowmem_pressure && zone_balanced(zone, sc->order, false,
7c954f6d
MG
3116 balance_gap, classzone_idx))
3117 return true;
3118
6b4f7799 3119 shrink_zone(zone, sc, zone_idx(zone) == classzone_idx);
75485363 3120
57054651 3121 clear_bit(ZONE_WRITEBACK, &zone->flags);
283aba9f 3122
7c954f6d
MG
3123 /*
3124 * If a zone reaches its high watermark, consider it to be no longer
3125 * congested. It's possible there are dirty pages backed by congested
3126 * BDIs but as pressure is relieved, speculatively avoid congestion
3127 * waits.
3128 */
6e543d57 3129 if (zone_reclaimable(zone) &&
accf6242 3130 zone_balanced(zone, sc->order, false, 0, classzone_idx)) {
57054651
JW
3131 clear_bit(ZONE_CONGESTED, &zone->flags);
3132 clear_bit(ZONE_DIRTY, &zone->flags);
7c954f6d
MG
3133 }
3134
b8e83b94 3135 return sc->nr_scanned >= sc->nr_to_reclaim;
75485363
MG
3136}
3137
1da177e4
LT
3138/*
3139 * For kswapd, balance_pgdat() will work across all this node's zones until
41858966 3140 * they are all at high_wmark_pages(zone).
1da177e4 3141 *
accf6242 3142 * Returns the highest zone idx kswapd was reclaiming at
1da177e4
LT
3143 *
3144 * There is special handling here for zones which are full of pinned pages.
3145 * This can happen if the pages are all mlocked, or if they are all used by
3146 * device drivers (say, ZONE_DMA). Or if they are all in use by hugetlb.
3147 * What we do is to detect the case where all pages in the zone have been
3148 * scanned twice and there has been zero successful reclaim. Mark the zone as
3149 * dead and from now on, only perform a short scan. Basically we're polling
3150 * the zone for when the problem goes away.
3151 *
3152 * kswapd scans the zones in the highmem->normal->dma direction. It skips
41858966
MG
3153 * zones which have free_pages > high_wmark_pages(zone), but once a zone is
3154 * found to have free_pages <= high_wmark_pages(zone), we scan that zone and the
3155 * lower zones regardless of the number of free pages in the lower zones. This
3156 * interoperates with the page allocator fallback scheme to ensure that aging
3157 * of pages is balanced across the zones.
1da177e4 3158 */
accf6242 3159static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
1da177e4 3160{
1da177e4 3161 int i;
99504748 3162 int end_zone = 0; /* Inclusive. 0 = ZONE_DMA */
0608f43d
AM
3163 unsigned long nr_soft_reclaimed;
3164 unsigned long nr_soft_scanned;
179e9639
AM
3165 struct scan_control sc = {
3166 .gfp_mask = GFP_KERNEL,
ee814fe2 3167 .order = order,
b8e83b94 3168 .priority = DEF_PRIORITY,
ee814fe2 3169 .may_writepage = !laptop_mode,
a6dc60f8 3170 .may_unmap = 1,
2e2e4259 3171 .may_swap = 1,
179e9639 3172 };
f8891e5e 3173 count_vm_event(PAGEOUTRUN);
1da177e4 3174
9e3b2f8c 3175 do {
b8e83b94
MG
3176 bool raise_priority = true;
3177
3178 sc.nr_reclaimed = 0;
1da177e4 3179
d6277db4
RW
3180 /*
3181 * Scan in the highmem->dma direction for the highest
3182 * zone which needs scanning
3183 */
3184 for (i = pgdat->nr_zones - 1; i >= 0; i--) {
3185 struct zone *zone = pgdat->node_zones + i;
1da177e4 3186
d6277db4
RW
3187 if (!populated_zone(zone))
3188 continue;
1da177e4 3189
6e543d57
LD
3190 if (sc.priority != DEF_PRIORITY &&
3191 !zone_reclaimable(zone))
d6277db4 3192 continue;
1da177e4 3193
556adecb
RR
3194 /*
3195 * Do some background aging of the anon list, to give
3196 * pages a chance to be referenced before reclaiming.
3197 */
9e3b2f8c 3198 age_active_anon(zone, &sc);
556adecb 3199
cc715d99
MG
3200 /*
3201 * If the number of buffer_heads in the machine
3202 * exceeds the maximum allowed level and this node
3203 * has a highmem zone, force kswapd to reclaim from
3204 * it to relieve lowmem pressure.
3205 */
3206 if (buffer_heads_over_limit && is_highmem_idx(i)) {
3207 end_zone = i;
3208 break;
3209 }
3210
accf6242 3211 if (!zone_balanced(zone, order, false, 0, 0)) {
d6277db4 3212 end_zone = i;
e1dbeda6 3213 break;
439423f6 3214 } else {
d43006d5
MG
3215 /*
3216 * If balanced, clear the dirty and congested
3217 * flags
3218 */
57054651
JW
3219 clear_bit(ZONE_CONGESTED, &zone->flags);
3220 clear_bit(ZONE_DIRTY, &zone->flags);
1da177e4 3221 }
1da177e4 3222 }
dafcb73e 3223
b8e83b94 3224 if (i < 0)
e1dbeda6
AM
3225 goto out;
3226
b7ea3c41
MG
3227 /*
3228 * If we're getting trouble reclaiming, start doing writepage
3229 * even in laptop mode.
3230 */
3231 if (sc.priority < DEF_PRIORITY - 2)
3232 sc.may_writepage = 1;
3233
1da177e4
LT
3234 /*
3235 * Now scan the zone in the dma->highmem direction, stopping
3236 * at the last zone which needs scanning.
3237 *
3238 * We do this because the page allocator works in the opposite
3239 * direction. This prevents the page allocator from allocating
3240 * pages behind kswapd's direction of progress, which would
3241 * cause too much scanning of the lower zones.
3242 */
3243 for (i = 0; i <= end_zone; i++) {
3244 struct zone *zone = pgdat->node_zones + i;
3245
f3fe6512 3246 if (!populated_zone(zone))
1da177e4
LT
3247 continue;
3248
6e543d57
LD
3249 if (sc.priority != DEF_PRIORITY &&
3250 !zone_reclaimable(zone))
1da177e4
LT
3251 continue;
3252
1da177e4 3253 sc.nr_scanned = 0;
4e416953 3254
0608f43d
AM
3255 nr_soft_scanned = 0;
3256 /*
3257 * Call soft limit reclaim before calling shrink_zone.
3258 */
3259 nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone,
3260 order, sc.gfp_mask,
3261 &nr_soft_scanned);
3262 sc.nr_reclaimed += nr_soft_reclaimed;
3263
32a4330d 3264 /*
7c954f6d
MG
3265 * There should be no need to raise the scanning
3266 * priority if enough pages are already being scanned
3267 * that that high watermark would be met at 100%
3268 * efficiency.
fe2c2a10 3269 */
accf6242 3270 if (kswapd_shrink_zone(zone, end_zone, &sc))
7c954f6d 3271 raise_priority = false;
1da177e4 3272 }
5515061d
MG
3273
3274 /*
3275 * If the low watermark is met there is no need for processes
3276 * to be throttled on pfmemalloc_wait as they should not be
3277 * able to safely make forward progress. Wake them
3278 */
3279 if (waitqueue_active(&pgdat->pfmemalloc_wait) &&
3280 pfmemalloc_watermark_ok(pgdat))
cfc51155 3281 wake_up_all(&pgdat->pfmemalloc_wait);
5515061d 3282
b8e83b94
MG
3283 /* Check if kswapd should be suspending */
3284 if (try_to_freeze() || kthread_should_stop())
3285 break;
8357376d 3286
73ce02e9 3287 /*
b8e83b94
MG
3288 * Raise priority if scanning rate is too low or there was no
3289 * progress in reclaiming pages
73ce02e9 3290 */
b8e83b94
MG
3291 if (raise_priority || !sc.nr_reclaimed)
3292 sc.priority--;
9aa41348 3293 } while (sc.priority >= 1 &&
accf6242 3294 !pgdat_balanced(pgdat, order, classzone_idx));
1da177e4 3295
b8e83b94 3296out:
0abdee2b 3297 /*
accf6242
VB
3298 * Return the highest zone idx we were reclaiming at so
3299 * prepare_kswapd_sleep() makes the same decisions as here.
0abdee2b 3300 */
accf6242 3301 return end_zone;
1da177e4
LT
3302}
3303
accf6242
VB
3304static void kswapd_try_to_sleep(pg_data_t *pgdat, int order,
3305 int classzone_idx, int balanced_classzone_idx)
f0bc0a60
KM
3306{
3307 long remaining = 0;
3308 DEFINE_WAIT(wait);
3309
3310 if (freezing(current) || kthread_should_stop())
3311 return;
3312
3313 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3314
3315 /* Try to sleep for a short interval */
accf6242
VB
3316 if (prepare_kswapd_sleep(pgdat, order, remaining,
3317 balanced_classzone_idx)) {
fd901c95
VB
3318 /*
3319 * Compaction records what page blocks it recently failed to
3320 * isolate pages from and skips them in the future scanning.
3321 * When kswapd is going to sleep, it is reasonable to assume
3322 * that pages and compaction may succeed so reset the cache.
3323 */
3324 reset_isolation_suitable(pgdat);
3325
3326 /*
3327 * We have freed the memory, now we should compact it to make
3328 * allocation of the requested order possible.
3329 */
3330 wakeup_kcompactd(pgdat, order, classzone_idx);
3331
f0bc0a60
KM
3332 remaining = schedule_timeout(HZ/10);
3333 finish_wait(&pgdat->kswapd_wait, &wait);
3334 prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE);
3335 }
3336
3337 /*
3338 * After a short sleep, check if it was a premature sleep. If not, then
3339 * go fully to sleep until explicitly woken up.
3340 */
accf6242
VB
3341 if (prepare_kswapd_sleep(pgdat, order, remaining,
3342 balanced_classzone_idx)) {
f0bc0a60
KM
3343 trace_mm_vmscan_kswapd_sleep(pgdat->node_id);
3344
3345 /*
3346 * vmstat counters are not perfectly accurate and the estimated
3347 * value for counters such as NR_FREE_PAGES can deviate from the
3348 * true value by nr_online_cpus * threshold. To avoid the zone
3349 * watermarks being breached while under pressure, we reduce the
3350 * per-cpu vmstat threshold while kswapd is awake and restore
3351 * them before going back to sleep.
3352 */
3353 set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold);
1c7e7f6c
AK
3354
3355 if (!kthread_should_stop())
3356 schedule();
3357
f0bc0a60
KM
3358 set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold);
3359 } else {
3360 if (remaining)
3361 count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY);
3362 else
3363 count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY);
3364 }
3365 finish_wait(&pgdat->kswapd_wait, &wait);
3366}
3367
1da177e4
LT
3368/*
3369 * The background pageout daemon, started as a kernel thread
4f98a2fe 3370 * from the init process.
1da177e4
LT
3371 *
3372 * This basically trickles out pages so that we have _some_
3373 * free memory available even if there is no other activity
3374 * that frees anything up. This is needed for things like routing
3375 * etc, where we otherwise might have all activity going on in
3376 * asynchronous contexts that cannot page things out.
3377 *
3378 * If there are applications that are active memory-allocators
3379 * (most normal use), this basically shouldn't matter.
3380 */
3381static int kswapd(void *p)
3382{
215ddd66
MG
3383 unsigned long order, new_order;
3384 int classzone_idx, new_classzone_idx;
d2ebd0f6 3385 int balanced_classzone_idx;
1da177e4
LT
3386 pg_data_t *pgdat = (pg_data_t*)p;
3387 struct task_struct *tsk = current;
f0bc0a60 3388
1da177e4
LT
3389 struct reclaim_state reclaim_state = {
3390 .reclaimed_slab = 0,
3391 };
a70f7302 3392 const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id);
1da177e4 3393
cf40bd16
NP
3394 lockdep_set_current_reclaim_state(GFP_KERNEL);
3395
174596a0 3396 if (!cpumask_empty(cpumask))
c5f59f08 3397 set_cpus_allowed_ptr(tsk, cpumask);
1da177e4
LT
3398 current->reclaim_state = &reclaim_state;
3399
3400 /*
3401 * Tell the memory management that we're a "memory allocator",
3402 * and that if we need more memory we should get access to it
3403 * regardless (see "__alloc_pages()"). "kswapd" should
3404 * never get caught in the normal page freeing logic.
3405 *
3406 * (Kswapd normally doesn't need memory anyway, but sometimes
3407 * you need a small amount of memory in order to be able to
3408 * page out something else, and this flag essentially protects
3409 * us from recursively trying to free more memory as we're
3410 * trying to free the first piece of memory in the first place).
3411 */
930d9152 3412 tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD;
83144186 3413 set_freezable();
1da177e4 3414
215ddd66
MG
3415 order = new_order = 0;
3416 classzone_idx = new_classzone_idx = pgdat->nr_zones - 1;
d2ebd0f6 3417 balanced_classzone_idx = classzone_idx;
1da177e4 3418 for ( ; ; ) {
6f6313d4 3419 bool ret;
3e1d1d28 3420
215ddd66 3421 /*
accf6242
VB
3422 * While we were reclaiming, there might have been another
3423 * wakeup, so check the values.
215ddd66 3424 */
accf6242
VB
3425 new_order = pgdat->kswapd_max_order;
3426 new_classzone_idx = pgdat->classzone_idx;
3427 pgdat->kswapd_max_order = 0;
3428 pgdat->classzone_idx = pgdat->nr_zones - 1;
215ddd66 3429
99504748 3430 if (order < new_order || classzone_idx > new_classzone_idx) {
1da177e4
LT
3431 /*
3432 * Don't sleep if someone wants a larger 'order'
99504748 3433 * allocation or has tigher zone constraints
1da177e4
LT
3434 */
3435 order = new_order;
99504748 3436 classzone_idx = new_classzone_idx;
1da177e4 3437 } else {
accf6242 3438 kswapd_try_to_sleep(pgdat, order, classzone_idx,
d2ebd0f6 3439 balanced_classzone_idx);
1da177e4 3440 order = pgdat->kswapd_max_order;
99504748 3441 classzone_idx = pgdat->classzone_idx;
f0dfcde0
AS
3442 new_order = order;
3443 new_classzone_idx = classzone_idx;
4d40502e 3444 pgdat->kswapd_max_order = 0;
215ddd66 3445 pgdat->classzone_idx = pgdat->nr_zones - 1;
1da177e4 3446 }
1da177e4 3447
8fe23e05
DR
3448 ret = try_to_freeze();
3449 if (kthread_should_stop())
3450 break;
3451
3452 /*
3453 * We can speed up thawing tasks if we don't call balance_pgdat
3454 * after returning from the refrigerator
3455 */
33906bc5
MG
3456 if (!ret) {
3457 trace_mm_vmscan_kswapd_wake(pgdat->node_id, order);
accf6242
VB
3458 balanced_classzone_idx = balance_pgdat(pgdat, order,
3459 classzone_idx);
33906bc5 3460 }
1da177e4 3461 }
b0a8cc58 3462
71abdc15 3463 tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD);
b0a8cc58 3464 current->reclaim_state = NULL;
71abdc15
JW
3465 lockdep_clear_current_reclaim_state();
3466
1da177e4
LT
3467 return 0;
3468}
3469
3470/*
3471 * A zone is low on free memory, so wake its kswapd task to service it.
3472 */
99504748 3473void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
1da177e4
LT
3474{
3475 pg_data_t *pgdat;
3476
f3fe6512 3477 if (!populated_zone(zone))
1da177e4
LT
3478 return;
3479
344736f2 3480 if (!cpuset_zone_allowed(zone, GFP_KERNEL | __GFP_HARDWALL))
1da177e4 3481 return;
88f5acf8 3482 pgdat = zone->zone_pgdat;
99504748 3483 if (pgdat->kswapd_max_order < order) {
1da177e4 3484 pgdat->kswapd_max_order = order;
99504748
MG
3485 pgdat->classzone_idx = min(pgdat->classzone_idx, classzone_idx);
3486 }
8d0986e2 3487 if (!waitqueue_active(&pgdat->kswapd_wait))
1da177e4 3488 return;
accf6242 3489 if (zone_balanced(zone, order, true, 0, 0))
88f5acf8
MG
3490 return;
3491
3492 trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
8d0986e2 3493 wake_up_interruptible(&pgdat->kswapd_wait);
1da177e4
LT
3494}
3495
c6f37f12 3496#ifdef CONFIG_HIBERNATION
1da177e4 3497/*
7b51755c 3498 * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of
d6277db4
RW
3499 * freed pages.
3500 *
3501 * Rather than trying to age LRUs the aim is to preserve the overall
3502 * LRU order by reclaiming preferentially
3503 * inactive > active > active referenced > active mapped
1da177e4 3504 */
7b51755c 3505unsigned long shrink_all_memory(unsigned long nr_to_reclaim)
1da177e4 3506{
d6277db4 3507 struct reclaim_state reclaim_state;
d6277db4 3508 struct scan_control sc = {
ee814fe2 3509 .nr_to_reclaim = nr_to_reclaim,
7b51755c 3510 .gfp_mask = GFP_HIGHUSER_MOVABLE,
ee814fe2 3511 .priority = DEF_PRIORITY,
d6277db4 3512 .may_writepage = 1,
ee814fe2
JW
3513 .may_unmap = 1,
3514 .may_swap = 1,
7b51755c 3515 .hibernation_mode = 1,
1da177e4 3516 };
a09ed5e0 3517 struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask);
7b51755c
KM
3518 struct task_struct *p = current;
3519 unsigned long nr_reclaimed;
1da177e4 3520
7b51755c
KM
3521 p->flags |= PF_MEMALLOC;
3522 lockdep_set_current_reclaim_state(sc.gfp_mask);
3523 reclaim_state.reclaimed_slab = 0;
3524 p->reclaim_state = &reclaim_state;
d6277db4 3525
3115cd91 3526 nr_reclaimed = do_try_to_free_pages(zonelist, &sc);
d979677c 3527
7b51755c
KM
3528 p->reclaim_state = NULL;
3529 lockdep_clear_current_reclaim_state();
3530 p->flags &= ~PF_MEMALLOC;
d6277db4 3531
7b51755c 3532 return nr_reclaimed;
1da177e4 3533}
c6f37f12 3534#endif /* CONFIG_HIBERNATION */
1da177e4 3535
1da177e4
LT
3536/* It's optimal to keep kswapds on the same CPUs as their memory, but
3537 not required for correctness. So if the last cpu in a node goes
3538 away, we get changed to run anywhere: as the first one comes back,
3539 restore their cpu bindings. */
fcb35a9b
GKH
3540static int cpu_callback(struct notifier_block *nfb, unsigned long action,
3541 void *hcpu)
1da177e4 3542{
58c0a4a7 3543 int nid;
1da177e4 3544
8bb78442 3545 if (action == CPU_ONLINE || action == CPU_ONLINE_FROZEN) {
48fb2e24 3546 for_each_node_state(nid, N_MEMORY) {
c5f59f08 3547 pg_data_t *pgdat = NODE_DATA(nid);
a70f7302
RR
3548 const struct cpumask *mask;
3549
3550 mask = cpumask_of_node(pgdat->node_id);
c5f59f08 3551
3e597945 3552 if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids)
1da177e4 3553 /* One of our CPUs online: restore mask */
c5f59f08 3554 set_cpus_allowed_ptr(pgdat->kswapd, mask);
1da177e4
LT
3555 }
3556 }
3557 return NOTIFY_OK;
3558}
1da177e4 3559
3218ae14
YG
3560/*
3561 * This kswapd start function will be called by init and node-hot-add.
3562 * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added.
3563 */
3564int kswapd_run(int nid)
3565{
3566 pg_data_t *pgdat = NODE_DATA(nid);
3567 int ret = 0;
3568
3569 if (pgdat->kswapd)
3570 return 0;
3571
3572 pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid);
3573 if (IS_ERR(pgdat->kswapd)) {
3574 /* failure at boot is fatal */
3575 BUG_ON(system_state == SYSTEM_BOOTING);
d5dc0ad9
GS
3576 pr_err("Failed to start kswapd on node %d\n", nid);
3577 ret = PTR_ERR(pgdat->kswapd);
d72515b8 3578 pgdat->kswapd = NULL;
3218ae14
YG
3579 }
3580 return ret;
3581}
3582
8fe23e05 3583/*
d8adde17 3584 * Called by memory hotplug when all memory in a node is offlined. Caller must
bfc8c901 3585 * hold mem_hotplug_begin/end().
8fe23e05
DR
3586 */
3587void kswapd_stop(int nid)
3588{
3589 struct task_struct *kswapd = NODE_DATA(nid)->kswapd;
3590
d8adde17 3591 if (kswapd) {
8fe23e05 3592 kthread_stop(kswapd);
d8adde17
JL
3593 NODE_DATA(nid)->kswapd = NULL;
3594 }
8fe23e05
DR
3595}
3596
1da177e4
LT
3597static int __init kswapd_init(void)
3598{
3218ae14 3599 int nid;
69e05944 3600
1da177e4 3601 swap_setup();
48fb2e24 3602 for_each_node_state(nid, N_MEMORY)
3218ae14 3603 kswapd_run(nid);
1da177e4
LT
3604 hotcpu_notifier(cpu_callback, 0);
3605 return 0;
3606}
3607
3608module_init(kswapd_init)
9eeff239
CL
3609
3610#ifdef CONFIG_NUMA
3611/*
3612 * Zone reclaim mode
3613 *
3614 * If non-zero call zone_reclaim when the number of free pages falls below
3615 * the watermarks.
9eeff239
CL
3616 */
3617int zone_reclaim_mode __read_mostly;
3618
1b2ffb78 3619#define RECLAIM_OFF 0
7d03431c 3620#define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */
1b2ffb78 3621#define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */
95bbc0c7 3622#define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */
1b2ffb78 3623
a92f7126
CL
3624/*
3625 * Priority for ZONE_RECLAIM. This determines the fraction of pages
3626 * of a node considered for each zone_reclaim. 4 scans 1/16th of
3627 * a zone.
3628 */
3629#define ZONE_RECLAIM_PRIORITY 4
3630
9614634f
CL
3631/*
3632 * Percentage of pages in a zone that must be unmapped for zone_reclaim to
3633 * occur.
3634 */
3635int sysctl_min_unmapped_ratio = 1;
3636
0ff38490
CL
3637/*
3638 * If the number of slab pages in a zone grows beyond this percentage then
3639 * slab reclaim needs to occur.
3640 */
3641int sysctl_min_slab_ratio = 5;
3642
90afa5de
MG
3643static inline unsigned long zone_unmapped_file_pages(struct zone *zone)
3644{
3645 unsigned long file_mapped = zone_page_state(zone, NR_FILE_MAPPED);
3646 unsigned long file_lru = zone_page_state(zone, NR_INACTIVE_FILE) +
3647 zone_page_state(zone, NR_ACTIVE_FILE);
3648
3649 /*
3650 * It's possible for there to be more file mapped pages than
3651 * accounted for by the pages on the file LRU lists because
3652 * tmpfs pages accounted for as ANON can also be FILE_MAPPED
3653 */
3654 return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0;
3655}
3656
3657/* Work out how many page cache pages we can reclaim in this reclaim_mode */
d031a157 3658static unsigned long zone_pagecache_reclaimable(struct zone *zone)
90afa5de 3659{
d031a157
AM
3660 unsigned long nr_pagecache_reclaimable;
3661 unsigned long delta = 0;
90afa5de
MG
3662
3663 /*
95bbc0c7 3664 * If RECLAIM_UNMAP is set, then all file pages are considered
90afa5de
MG
3665 * potentially reclaimable. Otherwise, we have to worry about
3666 * pages like swapcache and zone_unmapped_file_pages() provides
3667 * a better estimate
3668 */
95bbc0c7 3669 if (zone_reclaim_mode & RECLAIM_UNMAP)
90afa5de
MG
3670 nr_pagecache_reclaimable = zone_page_state(zone, NR_FILE_PAGES);
3671 else
3672 nr_pagecache_reclaimable = zone_unmapped_file_pages(zone);
3673
3674 /* If we can't clean pages, remove dirty pages from consideration */
3675 if (!(zone_reclaim_mode & RECLAIM_WRITE))
3676 delta += zone_page_state(zone, NR_FILE_DIRTY);
3677
3678 /* Watch for any possible underflows due to delta */
3679 if (unlikely(delta > nr_pagecache_reclaimable))
3680 delta = nr_pagecache_reclaimable;
3681
3682 return nr_pagecache_reclaimable - delta;
3683}
3684
9eeff239
CL
3685/*
3686 * Try to free up some pages from this zone through reclaim.
3687 */
179e9639 3688static int __zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
9eeff239 3689{
7fb2d46d 3690 /* Minimum pages needed in order to stay on node */
69e05944 3691 const unsigned long nr_pages = 1 << order;
9eeff239
CL
3692 struct task_struct *p = current;
3693 struct reclaim_state reclaim_state;
179e9639 3694 struct scan_control sc = {
62b726c1 3695 .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX),
21caf2fc 3696 .gfp_mask = (gfp_mask = memalloc_noio_flags(gfp_mask)),
bd2f6199 3697 .order = order,
9e3b2f8c 3698 .priority = ZONE_RECLAIM_PRIORITY,
ee814fe2 3699 .may_writepage = !!(zone_reclaim_mode & RECLAIM_WRITE),
95bbc0c7 3700 .may_unmap = !!(zone_reclaim_mode & RECLAIM_UNMAP),
ee814fe2 3701 .may_swap = 1,
179e9639 3702 };
9eeff239 3703
9eeff239 3704 cond_resched();
d4f7796e 3705 /*
95bbc0c7 3706 * We need to be able to allocate from the reserves for RECLAIM_UNMAP
d4f7796e 3707 * and we also need to be able to write out pages for RECLAIM_WRITE
95bbc0c7 3708 * and RECLAIM_UNMAP.
d4f7796e
CL
3709 */
3710 p->flags |= PF_MEMALLOC | PF_SWAPWRITE;
76ca542d 3711 lockdep_set_current_reclaim_state(gfp_mask);
9eeff239
CL
3712 reclaim_state.reclaimed_slab = 0;
3713 p->reclaim_state = &reclaim_state;
c84db23c 3714
90afa5de 3715 if (zone_pagecache_reclaimable(zone) > zone->min_unmapped_pages) {
0ff38490
CL
3716 /*
3717 * Free memory by calling shrink zone with increasing
3718 * priorities until we have enough memory freed.
3719 */
0ff38490 3720 do {
6b4f7799 3721 shrink_zone(zone, &sc, true);
9e3b2f8c 3722 } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0);
0ff38490 3723 }
c84db23c 3724
9eeff239 3725 p->reclaim_state = NULL;
d4f7796e 3726 current->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE);
76ca542d 3727 lockdep_clear_current_reclaim_state();
a79311c1 3728 return sc.nr_reclaimed >= nr_pages;
9eeff239 3729}
179e9639
AM
3730
3731int zone_reclaim(struct zone *zone, gfp_t gfp_mask, unsigned int order)
3732{
179e9639 3733 int node_id;
d773ed6b 3734 int ret;
179e9639
AM
3735
3736 /*
0ff38490
CL
3737 * Zone reclaim reclaims unmapped file backed pages and
3738 * slab pages if we are over the defined limits.
34aa1330 3739 *
9614634f
CL
3740 * A small portion of unmapped file backed pages is needed for
3741 * file I/O otherwise pages read by file I/O will be immediately
3742 * thrown out if the zone is overallocated. So we do not reclaim
3743 * if less than a specified percentage of the zone is used by
3744 * unmapped file backed pages.
179e9639 3745 */
90afa5de
MG
3746 if (zone_pagecache_reclaimable(zone) <= zone->min_unmapped_pages &&
3747 zone_page_state(zone, NR_SLAB_RECLAIMABLE) <= zone->min_slab_pages)
fa5e084e 3748 return ZONE_RECLAIM_FULL;
179e9639 3749
6e543d57 3750 if (!zone_reclaimable(zone))
fa5e084e 3751 return ZONE_RECLAIM_FULL;
d773ed6b 3752
179e9639 3753 /*
d773ed6b 3754 * Do not scan if the allocation should not be delayed.
179e9639 3755 */
d0164adc 3756 if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC))
fa5e084e 3757 return ZONE_RECLAIM_NOSCAN;
179e9639
AM
3758
3759 /*
3760 * Only run zone reclaim on the local zone or on zones that do not
3761 * have associated processors. This will favor the local processor
3762 * over remote processors and spread off node memory allocations
3763 * as wide as possible.
3764 */
89fa3024 3765 node_id = zone_to_nid(zone);
37c0708d 3766 if (node_state(node_id, N_CPU) && node_id != numa_node_id())
fa5e084e 3767 return ZONE_RECLAIM_NOSCAN;
d773ed6b 3768
57054651 3769 if (test_and_set_bit(ZONE_RECLAIM_LOCKED, &zone->flags))
fa5e084e
MG
3770 return ZONE_RECLAIM_NOSCAN;
3771
d773ed6b 3772 ret = __zone_reclaim(zone, gfp_mask, order);
57054651 3773 clear_bit(ZONE_RECLAIM_LOCKED, &zone->flags);
d773ed6b 3774
24cf7251
MG
3775 if (!ret)
3776 count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED);
3777
d773ed6b 3778 return ret;
179e9639 3779}
9eeff239 3780#endif
894bc310 3781
894bc310
LS
3782/*
3783 * page_evictable - test whether a page is evictable
3784 * @page: the page to test
894bc310
LS
3785 *
3786 * Test whether page is evictable--i.e., should be placed on active/inactive
39b5f29a 3787 * lists vs unevictable list.
894bc310
LS
3788 *
3789 * Reasons page might not be evictable:
ba9ddf49 3790 * (1) page's mapping marked unevictable
b291f000 3791 * (2) page is part of an mlocked VMA
ba9ddf49 3792 *
894bc310 3793 */
39b5f29a 3794int page_evictable(struct page *page)
894bc310 3795{
39b5f29a 3796 return !mapping_unevictable(page_mapping(page)) && !PageMlocked(page);
894bc310 3797}
89e004ea 3798
85046579 3799#ifdef CONFIG_SHMEM
89e004ea 3800/**
24513264
HD
3801 * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list
3802 * @pages: array of pages to check
3803 * @nr_pages: number of pages to check
89e004ea 3804 *
24513264 3805 * Checks pages for evictability and moves them to the appropriate lru list.
85046579
HD
3806 *
3807 * This function is only used for SysV IPC SHM_UNLOCK.
89e004ea 3808 */
24513264 3809void check_move_unevictable_pages(struct page **pages, int nr_pages)
89e004ea 3810{
925b7673 3811 struct lruvec *lruvec;
24513264
HD
3812 struct zone *zone = NULL;
3813 int pgscanned = 0;
3814 int pgrescued = 0;
3815 int i;
89e004ea 3816
24513264
HD
3817 for (i = 0; i < nr_pages; i++) {
3818 struct page *page = pages[i];
3819 struct zone *pagezone;
89e004ea 3820
24513264
HD
3821 pgscanned++;
3822 pagezone = page_zone(page);
3823 if (pagezone != zone) {
3824 if (zone)
3825 spin_unlock_irq(&zone->lru_lock);
3826 zone = pagezone;
3827 spin_lock_irq(&zone->lru_lock);
3828 }
fa9add64 3829 lruvec = mem_cgroup_page_lruvec(page, zone);
89e004ea 3830
24513264
HD
3831 if (!PageLRU(page) || !PageUnevictable(page))
3832 continue;
89e004ea 3833
39b5f29a 3834 if (page_evictable(page)) {
24513264
HD
3835 enum lru_list lru = page_lru_base_type(page);
3836
309381fe 3837 VM_BUG_ON_PAGE(PageActive(page), page);
24513264 3838 ClearPageUnevictable(page);
fa9add64
HD
3839 del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE);
3840 add_page_to_lru_list(page, lruvec, lru);
24513264 3841 pgrescued++;
89e004ea 3842 }
24513264 3843 }
89e004ea 3844
24513264
HD
3845 if (zone) {
3846 __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued);
3847 __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned);
3848 spin_unlock_irq(&zone->lru_lock);
89e004ea 3849 }
89e004ea 3850}
85046579 3851#endif /* CONFIG_SHMEM */