]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
1da177e4 LT |
2 | /* |
3 | * linux/mm/vmscan.c | |
4 | * | |
5 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
6 | * | |
7 | * Swap reorganised 29.12.95, Stephen Tweedie. | |
8 | * kswapd added: 7.1.96 sct | |
9 | * Removed kswapd_ctl limits, and swap out as many pages as needed | |
10 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. | |
11 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). | |
12 | * Multiqueue VM started 5.8.00, Rik van Riel. | |
13 | */ | |
14 | ||
b1de0d13 MH |
15 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
16 | ||
1da177e4 | 17 | #include <linux/mm.h> |
5b3cc15a | 18 | #include <linux/sched/mm.h> |
1da177e4 | 19 | #include <linux/module.h> |
5a0e3ad6 | 20 | #include <linux/gfp.h> |
1da177e4 LT |
21 | #include <linux/kernel_stat.h> |
22 | #include <linux/swap.h> | |
23 | #include <linux/pagemap.h> | |
24 | #include <linux/init.h> | |
25 | #include <linux/highmem.h> | |
70ddf637 | 26 | #include <linux/vmpressure.h> |
e129b5c2 | 27 | #include <linux/vmstat.h> |
1da177e4 LT |
28 | #include <linux/file.h> |
29 | #include <linux/writeback.h> | |
30 | #include <linux/blkdev.h> | |
31 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
32 | buffer_heads_over_limit */ | |
33 | #include <linux/mm_inline.h> | |
1da177e4 LT |
34 | #include <linux/backing-dev.h> |
35 | #include <linux/rmap.h> | |
36 | #include <linux/topology.h> | |
37 | #include <linux/cpu.h> | |
38 | #include <linux/cpuset.h> | |
3e7d3449 | 39 | #include <linux/compaction.h> |
1da177e4 LT |
40 | #include <linux/notifier.h> |
41 | #include <linux/rwsem.h> | |
248a0301 | 42 | #include <linux/delay.h> |
3218ae14 | 43 | #include <linux/kthread.h> |
7dfb7103 | 44 | #include <linux/freezer.h> |
66e1707b | 45 | #include <linux/memcontrol.h> |
873b4771 | 46 | #include <linux/delayacct.h> |
af936a16 | 47 | #include <linux/sysctl.h> |
929bea7c | 48 | #include <linux/oom.h> |
268bb0ce | 49 | #include <linux/prefetch.h> |
b1de0d13 | 50 | #include <linux/printk.h> |
f9fe48be | 51 | #include <linux/dax.h> |
1da177e4 LT |
52 | |
53 | #include <asm/tlbflush.h> | |
54 | #include <asm/div64.h> | |
55 | ||
56 | #include <linux/swapops.h> | |
117aad1e | 57 | #include <linux/balloon_compaction.h> |
1da177e4 | 58 | |
0f8053a5 NP |
59 | #include "internal.h" |
60 | ||
33906bc5 MG |
61 | #define CREATE_TRACE_POINTS |
62 | #include <trace/events/vmscan.h> | |
63 | ||
1da177e4 | 64 | struct scan_control { |
22fba335 KM |
65 | /* How many pages shrink_list() should reclaim */ |
66 | unsigned long nr_to_reclaim; | |
67 | ||
ee814fe2 JW |
68 | /* |
69 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
70 | * are scanned. | |
71 | */ | |
72 | nodemask_t *nodemask; | |
9e3b2f8c | 73 | |
f16015fb JW |
74 | /* |
75 | * The memory cgroup that hit its limit and as a result is the | |
76 | * primary target of this reclaim invocation. | |
77 | */ | |
78 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 79 | |
1276ad68 | 80 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
ee814fe2 JW |
81 | unsigned int may_writepage:1; |
82 | ||
83 | /* Can mapped pages be reclaimed? */ | |
84 | unsigned int may_unmap:1; | |
85 | ||
86 | /* Can pages be swapped as part of reclaim? */ | |
87 | unsigned int may_swap:1; | |
88 | ||
d6622f63 YX |
89 | /* |
90 | * Cgroups are not reclaimed below their configured memory.low, | |
91 | * unless we threaten to OOM. If any cgroups are skipped due to | |
92 | * memory.low and nothing was reclaimed, go back for memory.low. | |
93 | */ | |
94 | unsigned int memcg_low_reclaim:1; | |
95 | unsigned int memcg_low_skipped:1; | |
241994ed | 96 | |
ee814fe2 JW |
97 | unsigned int hibernation_mode:1; |
98 | ||
99 | /* One of the zones is ready for compaction */ | |
100 | unsigned int compaction_ready:1; | |
101 | ||
bb451fdf GT |
102 | /* Allocation order */ |
103 | s8 order; | |
104 | ||
105 | /* Scan (total_size >> priority) pages at once */ | |
106 | s8 priority; | |
107 | ||
108 | /* The highest zone to isolate pages for reclaim from */ | |
109 | s8 reclaim_idx; | |
110 | ||
111 | /* This context's GFP mask */ | |
112 | gfp_t gfp_mask; | |
113 | ||
ee814fe2 JW |
114 | /* Incremented by the number of inactive pages that were scanned */ |
115 | unsigned long nr_scanned; | |
116 | ||
117 | /* Number of pages freed so far during a call to shrink_zones() */ | |
118 | unsigned long nr_reclaimed; | |
d108c772 AR |
119 | |
120 | struct { | |
121 | unsigned int dirty; | |
122 | unsigned int unqueued_dirty; | |
123 | unsigned int congested; | |
124 | unsigned int writeback; | |
125 | unsigned int immediate; | |
126 | unsigned int file_taken; | |
127 | unsigned int taken; | |
128 | } nr; | |
1da177e4 LT |
129 | }; |
130 | ||
1da177e4 LT |
131 | #ifdef ARCH_HAS_PREFETCH |
132 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
133 | do { \ | |
134 | if ((_page)->lru.prev != _base) { \ | |
135 | struct page *prev; \ | |
136 | \ | |
137 | prev = lru_to_page(&(_page->lru)); \ | |
138 | prefetch(&prev->_field); \ | |
139 | } \ | |
140 | } while (0) | |
141 | #else | |
142 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
143 | #endif | |
144 | ||
145 | #ifdef ARCH_HAS_PREFETCHW | |
146 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
147 | do { \ | |
148 | if ((_page)->lru.prev != _base) { \ | |
149 | struct page *prev; \ | |
150 | \ | |
151 | prev = lru_to_page(&(_page->lru)); \ | |
152 | prefetchw(&prev->_field); \ | |
153 | } \ | |
154 | } while (0) | |
155 | #else | |
156 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
157 | #endif | |
158 | ||
159 | /* | |
160 | * From 0 .. 100. Higher means more swappy. | |
161 | */ | |
162 | int vm_swappiness = 60; | |
d0480be4 WSH |
163 | /* |
164 | * The total number of pages which are beyond the high watermark within all | |
165 | * zones. | |
166 | */ | |
167 | unsigned long vm_total_pages; | |
1da177e4 LT |
168 | |
169 | static LIST_HEAD(shrinker_list); | |
170 | static DECLARE_RWSEM(shrinker_rwsem); | |
171 | ||
b4c2b231 KT |
172 | #ifdef CONFIG_MEMCG_KMEM |
173 | static DEFINE_IDR(shrinker_idr); | |
174 | static int shrinker_nr_max; | |
175 | ||
176 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) | |
177 | { | |
178 | int id, ret = -ENOMEM; | |
179 | ||
180 | down_write(&shrinker_rwsem); | |
181 | /* This may call shrinker, so it must use down_read_trylock() */ | |
182 | id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL); | |
183 | if (id < 0) | |
184 | goto unlock; | |
185 | ||
0a4465d3 KT |
186 | if (id >= shrinker_nr_max) { |
187 | if (memcg_expand_shrinker_maps(id)) { | |
188 | idr_remove(&shrinker_idr, id); | |
189 | goto unlock; | |
190 | } | |
191 | ||
b4c2b231 | 192 | shrinker_nr_max = id + 1; |
0a4465d3 | 193 | } |
b4c2b231 KT |
194 | shrinker->id = id; |
195 | ret = 0; | |
196 | unlock: | |
197 | up_write(&shrinker_rwsem); | |
198 | return ret; | |
199 | } | |
200 | ||
201 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
202 | { | |
203 | int id = shrinker->id; | |
204 | ||
205 | BUG_ON(id < 0); | |
206 | ||
207 | down_write(&shrinker_rwsem); | |
208 | idr_remove(&shrinker_idr, id); | |
209 | up_write(&shrinker_rwsem); | |
210 | } | |
211 | #else /* CONFIG_MEMCG_KMEM */ | |
212 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) | |
213 | { | |
214 | return 0; | |
215 | } | |
216 | ||
217 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
218 | { | |
219 | } | |
220 | #endif /* CONFIG_MEMCG_KMEM */ | |
221 | ||
c255a458 | 222 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
223 | static bool global_reclaim(struct scan_control *sc) |
224 | { | |
f16015fb | 225 | return !sc->target_mem_cgroup; |
89b5fae5 | 226 | } |
97c9341f TH |
227 | |
228 | /** | |
229 | * sane_reclaim - is the usual dirty throttling mechanism operational? | |
230 | * @sc: scan_control in question | |
231 | * | |
232 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
233 | * completely broken with the legacy memcg and direct stalling in | |
234 | * shrink_page_list() is used for throttling instead, which lacks all the | |
235 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
236 | * allocation and configurability. | |
237 | * | |
238 | * This function tests whether the vmscan currently in progress can assume | |
239 | * that the normal dirty throttling mechanism is operational. | |
240 | */ | |
241 | static bool sane_reclaim(struct scan_control *sc) | |
242 | { | |
243 | struct mem_cgroup *memcg = sc->target_mem_cgroup; | |
244 | ||
245 | if (!memcg) | |
246 | return true; | |
247 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 248 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
249 | return true; |
250 | #endif | |
251 | return false; | |
252 | } | |
e3c1ac58 AR |
253 | |
254 | static void set_memcg_congestion(pg_data_t *pgdat, | |
255 | struct mem_cgroup *memcg, | |
256 | bool congested) | |
257 | { | |
258 | struct mem_cgroup_per_node *mn; | |
259 | ||
260 | if (!memcg) | |
261 | return; | |
262 | ||
263 | mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id); | |
264 | WRITE_ONCE(mn->congested, congested); | |
265 | } | |
266 | ||
267 | static bool memcg_congested(pg_data_t *pgdat, | |
268 | struct mem_cgroup *memcg) | |
269 | { | |
270 | struct mem_cgroup_per_node *mn; | |
271 | ||
272 | mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id); | |
273 | return READ_ONCE(mn->congested); | |
274 | ||
275 | } | |
91a45470 | 276 | #else |
89b5fae5 JW |
277 | static bool global_reclaim(struct scan_control *sc) |
278 | { | |
279 | return true; | |
280 | } | |
97c9341f TH |
281 | |
282 | static bool sane_reclaim(struct scan_control *sc) | |
283 | { | |
284 | return true; | |
285 | } | |
e3c1ac58 AR |
286 | |
287 | static inline void set_memcg_congestion(struct pglist_data *pgdat, | |
288 | struct mem_cgroup *memcg, bool congested) | |
289 | { | |
290 | } | |
291 | ||
292 | static inline bool memcg_congested(struct pglist_data *pgdat, | |
293 | struct mem_cgroup *memcg) | |
294 | { | |
295 | return false; | |
296 | ||
297 | } | |
91a45470 KH |
298 | #endif |
299 | ||
5a1c84b4 MG |
300 | /* |
301 | * This misses isolated pages which are not accounted for to save counters. | |
302 | * As the data only determines if reclaim or compaction continues, it is | |
303 | * not expected that isolated pages will be a dominating factor. | |
304 | */ | |
305 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
306 | { | |
307 | unsigned long nr; | |
308 | ||
309 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
310 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
311 | if (get_nr_swap_pages() > 0) | |
312 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + | |
313 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
314 | ||
315 | return nr; | |
316 | } | |
317 | ||
fd538803 MH |
318 | /** |
319 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
320 | * @lruvec: lru vector | |
321 | * @lru: lru to use | |
322 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) | |
323 | */ | |
324 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) | |
c9f299d9 | 325 | { |
fd538803 MH |
326 | unsigned long lru_size; |
327 | int zid; | |
328 | ||
c3c787e8 | 329 | if (!mem_cgroup_disabled()) |
fd538803 MH |
330 | lru_size = mem_cgroup_get_lru_size(lruvec, lru); |
331 | else | |
332 | lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); | |
a3d8e054 | 333 | |
fd538803 MH |
334 | for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) { |
335 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; | |
336 | unsigned long size; | |
c9f299d9 | 337 | |
fd538803 MH |
338 | if (!managed_zone(zone)) |
339 | continue; | |
340 | ||
341 | if (!mem_cgroup_disabled()) | |
342 | size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid); | |
343 | else | |
344 | size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid], | |
345 | NR_ZONE_LRU_BASE + lru); | |
346 | lru_size -= min(size, lru_size); | |
347 | } | |
348 | ||
349 | return lru_size; | |
b4536f0c | 350 | |
b4536f0c MH |
351 | } |
352 | ||
1da177e4 | 353 | /* |
1d3d4437 | 354 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 355 | */ |
8e04944f | 356 | int prealloc_shrinker(struct shrinker *shrinker) |
1da177e4 | 357 | { |
1d3d4437 GC |
358 | size_t size = sizeof(*shrinker->nr_deferred); |
359 | ||
1d3d4437 GC |
360 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
361 | size *= nr_node_ids; | |
362 | ||
363 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
364 | if (!shrinker->nr_deferred) | |
365 | return -ENOMEM; | |
b4c2b231 | 366 | |
b0dedc49 KT |
367 | /* |
368 | * There is a window between prealloc_shrinker() | |
369 | * and register_shrinker_prepared(). We don't want | |
370 | * to clear bit of a shrinker in such the state | |
371 | * in shrink_slab_memcg(), since this will impose | |
372 | * restrictions on a code registering a shrinker | |
373 | * (they would have to guarantee, their LRU lists | |
374 | * are empty till shrinker is completely registered). | |
375 | * So, we differ the situation, when 1)a shrinker | |
376 | * is semi-registered (id is assigned, but it has | |
377 | * not yet linked to shrinker_list) and 2)shrinker | |
378 | * is not registered (id is not assigned). | |
379 | */ | |
380 | INIT_LIST_HEAD(&shrinker->list); | |
381 | ||
b4c2b231 KT |
382 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) { |
383 | if (prealloc_memcg_shrinker(shrinker)) | |
384 | goto free_deferred; | |
385 | } | |
386 | ||
8e04944f | 387 | return 0; |
b4c2b231 KT |
388 | |
389 | free_deferred: | |
390 | kfree(shrinker->nr_deferred); | |
391 | shrinker->nr_deferred = NULL; | |
392 | return -ENOMEM; | |
8e04944f TH |
393 | } |
394 | ||
395 | void free_prealloced_shrinker(struct shrinker *shrinker) | |
396 | { | |
b4c2b231 KT |
397 | if (!shrinker->nr_deferred) |
398 | return; | |
399 | ||
400 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) | |
401 | unregister_memcg_shrinker(shrinker); | |
402 | ||
8e04944f TH |
403 | kfree(shrinker->nr_deferred); |
404 | shrinker->nr_deferred = NULL; | |
405 | } | |
1d3d4437 | 406 | |
8e04944f TH |
407 | void register_shrinker_prepared(struct shrinker *shrinker) |
408 | { | |
8e1f936b RR |
409 | down_write(&shrinker_rwsem); |
410 | list_add_tail(&shrinker->list, &shrinker_list); | |
411 | up_write(&shrinker_rwsem); | |
8e04944f TH |
412 | } |
413 | ||
414 | int register_shrinker(struct shrinker *shrinker) | |
415 | { | |
416 | int err = prealloc_shrinker(shrinker); | |
417 | ||
418 | if (err) | |
419 | return err; | |
420 | register_shrinker_prepared(shrinker); | |
1d3d4437 | 421 | return 0; |
1da177e4 | 422 | } |
8e1f936b | 423 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
424 | |
425 | /* | |
426 | * Remove one | |
427 | */ | |
8e1f936b | 428 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 | 429 | { |
bb422a73 TH |
430 | if (!shrinker->nr_deferred) |
431 | return; | |
b4c2b231 KT |
432 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) |
433 | unregister_memcg_shrinker(shrinker); | |
1da177e4 LT |
434 | down_write(&shrinker_rwsem); |
435 | list_del(&shrinker->list); | |
436 | up_write(&shrinker_rwsem); | |
ae393321 | 437 | kfree(shrinker->nr_deferred); |
bb422a73 | 438 | shrinker->nr_deferred = NULL; |
1da177e4 | 439 | } |
8e1f936b | 440 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
441 | |
442 | #define SHRINK_BATCH 128 | |
1d3d4437 | 443 | |
cb731d6c | 444 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
9092c71b | 445 | struct shrinker *shrinker, int priority) |
1d3d4437 GC |
446 | { |
447 | unsigned long freed = 0; | |
448 | unsigned long long delta; | |
449 | long total_scan; | |
d5bc5fd3 | 450 | long freeable; |
1d3d4437 GC |
451 | long nr; |
452 | long new_nr; | |
453 | int nid = shrinkctl->nid; | |
454 | long batch_size = shrinker->batch ? shrinker->batch | |
455 | : SHRINK_BATCH; | |
5f33a080 | 456 | long scanned = 0, next_deferred; |
1d3d4437 | 457 | |
d5bc5fd3 VD |
458 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
459 | if (freeable == 0) | |
1d3d4437 GC |
460 | return 0; |
461 | ||
462 | /* | |
463 | * copy the current shrinker scan count into a local variable | |
464 | * and zero it so that other concurrent shrinker invocations | |
465 | * don't also do this scanning work. | |
466 | */ | |
467 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
468 | ||
469 | total_scan = nr; | |
9092c71b JB |
470 | delta = freeable >> priority; |
471 | delta *= 4; | |
472 | do_div(delta, shrinker->seeks); | |
1d3d4437 GC |
473 | total_scan += delta; |
474 | if (total_scan < 0) { | |
8612c663 | 475 | pr_err("shrink_slab: %pF negative objects to delete nr=%ld\n", |
a0b02131 | 476 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 477 | total_scan = freeable; |
5f33a080 SL |
478 | next_deferred = nr; |
479 | } else | |
480 | next_deferred = total_scan; | |
1d3d4437 GC |
481 | |
482 | /* | |
483 | * We need to avoid excessive windup on filesystem shrinkers | |
484 | * due to large numbers of GFP_NOFS allocations causing the | |
485 | * shrinkers to return -1 all the time. This results in a large | |
486 | * nr being built up so when a shrink that can do some work | |
487 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 488 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
489 | * memory. |
490 | * | |
491 | * Hence only allow the shrinker to scan the entire cache when | |
492 | * a large delta change is calculated directly. | |
493 | */ | |
d5bc5fd3 VD |
494 | if (delta < freeable / 4) |
495 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
496 | |
497 | /* | |
498 | * Avoid risking looping forever due to too large nr value: | |
499 | * never try to free more than twice the estimate number of | |
500 | * freeable entries. | |
501 | */ | |
d5bc5fd3 VD |
502 | if (total_scan > freeable * 2) |
503 | total_scan = freeable * 2; | |
1d3d4437 GC |
504 | |
505 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
9092c71b | 506 | freeable, delta, total_scan, priority); |
1d3d4437 | 507 | |
0b1fb40a VD |
508 | /* |
509 | * Normally, we should not scan less than batch_size objects in one | |
510 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
511 | * than batch_size objects in total and we are really tight on memory, | |
512 | * we will try to reclaim all available objects, otherwise we can end | |
513 | * up failing allocations although there are plenty of reclaimable | |
514 | * objects spread over several slabs with usage less than the | |
515 | * batch_size. | |
516 | * | |
517 | * We detect the "tight on memory" situations by looking at the total | |
518 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 519 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
520 | * scanning at high prio and therefore should try to reclaim as much as |
521 | * possible. | |
522 | */ | |
523 | while (total_scan >= batch_size || | |
d5bc5fd3 | 524 | total_scan >= freeable) { |
a0b02131 | 525 | unsigned long ret; |
0b1fb40a | 526 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 527 | |
0b1fb40a | 528 | shrinkctl->nr_to_scan = nr_to_scan; |
d460acb5 | 529 | shrinkctl->nr_scanned = nr_to_scan; |
a0b02131 DC |
530 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
531 | if (ret == SHRINK_STOP) | |
532 | break; | |
533 | freed += ret; | |
1d3d4437 | 534 | |
d460acb5 CW |
535 | count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); |
536 | total_scan -= shrinkctl->nr_scanned; | |
537 | scanned += shrinkctl->nr_scanned; | |
1d3d4437 GC |
538 | |
539 | cond_resched(); | |
540 | } | |
541 | ||
5f33a080 SL |
542 | if (next_deferred >= scanned) |
543 | next_deferred -= scanned; | |
544 | else | |
545 | next_deferred = 0; | |
1d3d4437 GC |
546 | /* |
547 | * move the unused scan count back into the shrinker in a | |
548 | * manner that handles concurrent updates. If we exhausted the | |
549 | * scan, there is no need to do an update. | |
550 | */ | |
5f33a080 SL |
551 | if (next_deferred > 0) |
552 | new_nr = atomic_long_add_return(next_deferred, | |
1d3d4437 GC |
553 | &shrinker->nr_deferred[nid]); |
554 | else | |
555 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
556 | ||
df9024a8 | 557 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 558 | return freed; |
1495f230 YH |
559 | } |
560 | ||
b0dedc49 KT |
561 | #ifdef CONFIG_MEMCG_KMEM |
562 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, | |
563 | struct mem_cgroup *memcg, int priority) | |
564 | { | |
565 | struct memcg_shrinker_map *map; | |
566 | unsigned long freed = 0; | |
567 | int ret, i; | |
568 | ||
569 | if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)) | |
570 | return 0; | |
571 | ||
572 | if (!down_read_trylock(&shrinker_rwsem)) | |
573 | return 0; | |
574 | ||
575 | map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map, | |
576 | true); | |
577 | if (unlikely(!map)) | |
578 | goto unlock; | |
579 | ||
580 | for_each_set_bit(i, map->map, shrinker_nr_max) { | |
581 | struct shrink_control sc = { | |
582 | .gfp_mask = gfp_mask, | |
583 | .nid = nid, | |
584 | .memcg = memcg, | |
585 | }; | |
586 | struct shrinker *shrinker; | |
587 | ||
588 | shrinker = idr_find(&shrinker_idr, i); | |
589 | if (unlikely(!shrinker)) { | |
590 | clear_bit(i, map->map); | |
591 | continue; | |
592 | } | |
593 | ||
594 | /* See comment in prealloc_shrinker() */ | |
595 | if (unlikely(list_empty(&shrinker->list))) | |
596 | continue; | |
597 | ||
598 | ret = do_shrink_slab(&sc, shrinker, priority); | |
599 | freed += ret; | |
600 | ||
601 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
602 | freed = freed ? : 1; | |
603 | break; | |
604 | } | |
605 | } | |
606 | unlock: | |
607 | up_read(&shrinker_rwsem); | |
608 | return freed; | |
609 | } | |
610 | #else /* CONFIG_MEMCG_KMEM */ | |
611 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, | |
612 | struct mem_cgroup *memcg, int priority) | |
613 | { | |
614 | return 0; | |
615 | } | |
616 | #endif /* CONFIG_MEMCG_KMEM */ | |
617 | ||
6b4f7799 | 618 | /** |
cb731d6c | 619 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
620 | * @gfp_mask: allocation context |
621 | * @nid: node whose slab caches to target | |
cb731d6c | 622 | * @memcg: memory cgroup whose slab caches to target |
9092c71b | 623 | * @priority: the reclaim priority |
1da177e4 | 624 | * |
6b4f7799 | 625 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 626 | * |
6b4f7799 JW |
627 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
628 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 629 | * |
aeed1d32 VD |
630 | * @memcg specifies the memory cgroup to target. Unaware shrinkers |
631 | * are called only if it is the root cgroup. | |
cb731d6c | 632 | * |
9092c71b JB |
633 | * @priority is sc->priority, we take the number of objects and >> by priority |
634 | * in order to get the scan target. | |
b15e0905 | 635 | * |
6b4f7799 | 636 | * Returns the number of reclaimed slab objects. |
1da177e4 | 637 | */ |
cb731d6c VD |
638 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
639 | struct mem_cgroup *memcg, | |
9092c71b | 640 | int priority) |
1da177e4 LT |
641 | { |
642 | struct shrinker *shrinker; | |
24f7c6b9 | 643 | unsigned long freed = 0; |
1da177e4 | 644 | |
aeed1d32 | 645 | if (!mem_cgroup_is_root(memcg)) |
b0dedc49 | 646 | return shrink_slab_memcg(gfp_mask, nid, memcg, priority); |
cb731d6c | 647 | |
e830c63a | 648 | if (!down_read_trylock(&shrinker_rwsem)) |
f06590bd | 649 | goto out; |
1da177e4 LT |
650 | |
651 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
652 | struct shrink_control sc = { |
653 | .gfp_mask = gfp_mask, | |
654 | .nid = nid, | |
cb731d6c | 655 | .memcg = memcg, |
6b4f7799 | 656 | }; |
ec97097b | 657 | |
6b4f7799 JW |
658 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
659 | sc.nid = 0; | |
1da177e4 | 660 | |
9092c71b | 661 | freed += do_shrink_slab(&sc, shrinker, priority); |
e496612c MK |
662 | /* |
663 | * Bail out if someone want to register a new shrinker to | |
664 | * prevent the regsitration from being stalled for long periods | |
665 | * by parallel ongoing shrinking. | |
666 | */ | |
667 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
668 | freed = freed ? : 1; | |
669 | break; | |
670 | } | |
1da177e4 | 671 | } |
6b4f7799 | 672 | |
1da177e4 | 673 | up_read(&shrinker_rwsem); |
f06590bd MK |
674 | out: |
675 | cond_resched(); | |
24f7c6b9 | 676 | return freed; |
1da177e4 LT |
677 | } |
678 | ||
cb731d6c VD |
679 | void drop_slab_node(int nid) |
680 | { | |
681 | unsigned long freed; | |
682 | ||
683 | do { | |
684 | struct mem_cgroup *memcg = NULL; | |
685 | ||
686 | freed = 0; | |
aeed1d32 | 687 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
cb731d6c | 688 | do { |
9092c71b | 689 | freed += shrink_slab(GFP_KERNEL, nid, memcg, 0); |
cb731d6c VD |
690 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); |
691 | } while (freed > 10); | |
692 | } | |
693 | ||
694 | void drop_slab(void) | |
695 | { | |
696 | int nid; | |
697 | ||
698 | for_each_online_node(nid) | |
699 | drop_slab_node(nid); | |
700 | } | |
701 | ||
1da177e4 LT |
702 | static inline int is_page_cache_freeable(struct page *page) |
703 | { | |
ceddc3a5 JW |
704 | /* |
705 | * A freeable page cache page is referenced only by the caller | |
706 | * that isolated the page, the page cache radix tree and | |
707 | * optional buffer heads at page->private. | |
708 | */ | |
bd4c82c2 HY |
709 | int radix_pins = PageTransHuge(page) && PageSwapCache(page) ? |
710 | HPAGE_PMD_NR : 1; | |
711 | return page_count(page) - page_has_private(page) == 1 + radix_pins; | |
1da177e4 LT |
712 | } |
713 | ||
703c2708 | 714 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) |
1da177e4 | 715 | { |
930d9152 | 716 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 717 | return 1; |
703c2708 | 718 | if (!inode_write_congested(inode)) |
1da177e4 | 719 | return 1; |
703c2708 | 720 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
721 | return 1; |
722 | return 0; | |
723 | } | |
724 | ||
725 | /* | |
726 | * We detected a synchronous write error writing a page out. Probably | |
727 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
728 | * fsync(), msync() or close(). | |
729 | * | |
730 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
731 | * prevents it from being freed up. But we have a ref on the page and once | |
732 | * that page is locked, the mapping is pinned. | |
733 | * | |
734 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
735 | * __GFP_FS. | |
736 | */ | |
737 | static void handle_write_error(struct address_space *mapping, | |
738 | struct page *page, int error) | |
739 | { | |
7eaceacc | 740 | lock_page(page); |
3e9f45bd GC |
741 | if (page_mapping(page) == mapping) |
742 | mapping_set_error(mapping, error); | |
1da177e4 LT |
743 | unlock_page(page); |
744 | } | |
745 | ||
04e62a29 CL |
746 | /* possible outcome of pageout() */ |
747 | typedef enum { | |
748 | /* failed to write page out, page is locked */ | |
749 | PAGE_KEEP, | |
750 | /* move page to the active list, page is locked */ | |
751 | PAGE_ACTIVATE, | |
752 | /* page has been sent to the disk successfully, page is unlocked */ | |
753 | PAGE_SUCCESS, | |
754 | /* page is clean and locked */ | |
755 | PAGE_CLEAN, | |
756 | } pageout_t; | |
757 | ||
1da177e4 | 758 | /* |
1742f19f AM |
759 | * pageout is called by shrink_page_list() for each dirty page. |
760 | * Calls ->writepage(). | |
1da177e4 | 761 | */ |
c661b078 | 762 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 763 | struct scan_control *sc) |
1da177e4 LT |
764 | { |
765 | /* | |
766 | * If the page is dirty, only perform writeback if that write | |
767 | * will be non-blocking. To prevent this allocation from being | |
768 | * stalled by pagecache activity. But note that there may be | |
769 | * stalls if we need to run get_block(). We could test | |
770 | * PagePrivate for that. | |
771 | * | |
8174202b | 772 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
773 | * this page's queue, we can perform writeback even if that |
774 | * will block. | |
775 | * | |
776 | * If the page is swapcache, write it back even if that would | |
777 | * block, for some throttling. This happens by accident, because | |
778 | * swap_backing_dev_info is bust: it doesn't reflect the | |
779 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
780 | */ |
781 | if (!is_page_cache_freeable(page)) | |
782 | return PAGE_KEEP; | |
783 | if (!mapping) { | |
784 | /* | |
785 | * Some data journaling orphaned pages can have | |
786 | * page->mapping == NULL while being dirty with clean buffers. | |
787 | */ | |
266cf658 | 788 | if (page_has_private(page)) { |
1da177e4 LT |
789 | if (try_to_free_buffers(page)) { |
790 | ClearPageDirty(page); | |
b1de0d13 | 791 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
792 | return PAGE_CLEAN; |
793 | } | |
794 | } | |
795 | return PAGE_KEEP; | |
796 | } | |
797 | if (mapping->a_ops->writepage == NULL) | |
798 | return PAGE_ACTIVATE; | |
703c2708 | 799 | if (!may_write_to_inode(mapping->host, sc)) |
1da177e4 LT |
800 | return PAGE_KEEP; |
801 | ||
802 | if (clear_page_dirty_for_io(page)) { | |
803 | int res; | |
804 | struct writeback_control wbc = { | |
805 | .sync_mode = WB_SYNC_NONE, | |
806 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
807 | .range_start = 0, |
808 | .range_end = LLONG_MAX, | |
1da177e4 LT |
809 | .for_reclaim = 1, |
810 | }; | |
811 | ||
812 | SetPageReclaim(page); | |
813 | res = mapping->a_ops->writepage(page, &wbc); | |
814 | if (res < 0) | |
815 | handle_write_error(mapping, page, res); | |
994fc28c | 816 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
817 | ClearPageReclaim(page); |
818 | return PAGE_ACTIVATE; | |
819 | } | |
c661b078 | 820 | |
1da177e4 LT |
821 | if (!PageWriteback(page)) { |
822 | /* synchronous write or broken a_ops? */ | |
823 | ClearPageReclaim(page); | |
824 | } | |
3aa23851 | 825 | trace_mm_vmscan_writepage(page); |
c4a25635 | 826 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
827 | return PAGE_SUCCESS; |
828 | } | |
829 | ||
830 | return PAGE_CLEAN; | |
831 | } | |
832 | ||
a649fd92 | 833 | /* |
e286781d NP |
834 | * Same as remove_mapping, but if the page is removed from the mapping, it |
835 | * gets returned with a refcount of 0. | |
a649fd92 | 836 | */ |
a528910e JW |
837 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
838 | bool reclaimed) | |
49d2e9cc | 839 | { |
c4843a75 | 840 | unsigned long flags; |
bd4c82c2 | 841 | int refcount; |
c4843a75 | 842 | |
28e4d965 NP |
843 | BUG_ON(!PageLocked(page)); |
844 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 845 | |
b93b0163 | 846 | xa_lock_irqsave(&mapping->i_pages, flags); |
49d2e9cc | 847 | /* |
0fd0e6b0 NP |
848 | * The non racy check for a busy page. |
849 | * | |
850 | * Must be careful with the order of the tests. When someone has | |
851 | * a ref to the page, it may be possible that they dirty it then | |
852 | * drop the reference. So if PageDirty is tested before page_count | |
853 | * here, then the following race may occur: | |
854 | * | |
855 | * get_user_pages(&page); | |
856 | * [user mapping goes away] | |
857 | * write_to(page); | |
858 | * !PageDirty(page) [good] | |
859 | * SetPageDirty(page); | |
860 | * put_page(page); | |
861 | * !page_count(page) [good, discard it] | |
862 | * | |
863 | * [oops, our write_to data is lost] | |
864 | * | |
865 | * Reversing the order of the tests ensures such a situation cannot | |
866 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 867 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
868 | * |
869 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
b93b0163 | 870 | * and thus under the i_pages lock, then this ordering is not required. |
49d2e9cc | 871 | */ |
bd4c82c2 HY |
872 | if (unlikely(PageTransHuge(page)) && PageSwapCache(page)) |
873 | refcount = 1 + HPAGE_PMD_NR; | |
874 | else | |
875 | refcount = 2; | |
876 | if (!page_ref_freeze(page, refcount)) | |
49d2e9cc | 877 | goto cannot_free; |
e286781d NP |
878 | /* note: atomic_cmpxchg in page_freeze_refs provides the smp_rmb */ |
879 | if (unlikely(PageDirty(page))) { | |
bd4c82c2 | 880 | page_ref_unfreeze(page, refcount); |
49d2e9cc | 881 | goto cannot_free; |
e286781d | 882 | } |
49d2e9cc CL |
883 | |
884 | if (PageSwapCache(page)) { | |
885 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 886 | mem_cgroup_swapout(page, swap); |
49d2e9cc | 887 | __delete_from_swap_cache(page); |
b93b0163 | 888 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
75f6d6d2 | 889 | put_swap_page(page, swap); |
e286781d | 890 | } else { |
6072d13c | 891 | void (*freepage)(struct page *); |
a528910e | 892 | void *shadow = NULL; |
6072d13c LT |
893 | |
894 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
895 | /* |
896 | * Remember a shadow entry for reclaimed file cache in | |
897 | * order to detect refaults, thus thrashing, later on. | |
898 | * | |
899 | * But don't store shadows in an address space that is | |
900 | * already exiting. This is not just an optizimation, | |
901 | * inode reclaim needs to empty out the radix tree or | |
902 | * the nodes are lost. Don't plant shadows behind its | |
903 | * back. | |
f9fe48be RZ |
904 | * |
905 | * We also don't store shadows for DAX mappings because the | |
906 | * only page cache pages found in these are zero pages | |
907 | * covering holes, and because we don't want to mix DAX | |
908 | * exceptional entries and shadow exceptional entries in the | |
b93b0163 | 909 | * same address_space. |
a528910e JW |
910 | */ |
911 | if (reclaimed && page_is_file_cache(page) && | |
f9fe48be | 912 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
a528910e | 913 | shadow = workingset_eviction(mapping, page); |
62cccb8c | 914 | __delete_from_page_cache(page, shadow); |
b93b0163 | 915 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
6072d13c LT |
916 | |
917 | if (freepage != NULL) | |
918 | freepage(page); | |
49d2e9cc CL |
919 | } |
920 | ||
49d2e9cc CL |
921 | return 1; |
922 | ||
923 | cannot_free: | |
b93b0163 | 924 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
49d2e9cc CL |
925 | return 0; |
926 | } | |
927 | ||
e286781d NP |
928 | /* |
929 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
930 | * someone else has a ref on the page, abort and return 0. If it was | |
931 | * successfully detached, return 1. Assumes the caller has a single ref on | |
932 | * this page. | |
933 | */ | |
934 | int remove_mapping(struct address_space *mapping, struct page *page) | |
935 | { | |
a528910e | 936 | if (__remove_mapping(mapping, page, false)) { |
e286781d NP |
937 | /* |
938 | * Unfreezing the refcount with 1 rather than 2 effectively | |
939 | * drops the pagecache ref for us without requiring another | |
940 | * atomic operation. | |
941 | */ | |
fe896d18 | 942 | page_ref_unfreeze(page, 1); |
e286781d NP |
943 | return 1; |
944 | } | |
945 | return 0; | |
946 | } | |
947 | ||
894bc310 LS |
948 | /** |
949 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
950 | * @page: page to be put back to appropriate lru list | |
951 | * | |
952 | * Add previously isolated @page to appropriate LRU list. | |
953 | * Page may still be unevictable for other reasons. | |
954 | * | |
955 | * lru_lock must not be held, interrupts must be enabled. | |
956 | */ | |
894bc310 LS |
957 | void putback_lru_page(struct page *page) |
958 | { | |
9c4e6b1a | 959 | lru_cache_add(page); |
894bc310 LS |
960 | put_page(page); /* drop ref from isolate */ |
961 | } | |
962 | ||
dfc8d636 JW |
963 | enum page_references { |
964 | PAGEREF_RECLAIM, | |
965 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 966 | PAGEREF_KEEP, |
dfc8d636 JW |
967 | PAGEREF_ACTIVATE, |
968 | }; | |
969 | ||
970 | static enum page_references page_check_references(struct page *page, | |
971 | struct scan_control *sc) | |
972 | { | |
64574746 | 973 | int referenced_ptes, referenced_page; |
dfc8d636 | 974 | unsigned long vm_flags; |
dfc8d636 | 975 | |
c3ac9a8a JW |
976 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
977 | &vm_flags); | |
64574746 | 978 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 979 | |
dfc8d636 JW |
980 | /* |
981 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
982 | * move the page to the unevictable list. | |
983 | */ | |
984 | if (vm_flags & VM_LOCKED) | |
985 | return PAGEREF_RECLAIM; | |
986 | ||
64574746 | 987 | if (referenced_ptes) { |
e4898273 | 988 | if (PageSwapBacked(page)) |
64574746 JW |
989 | return PAGEREF_ACTIVATE; |
990 | /* | |
991 | * All mapped pages start out with page table | |
992 | * references from the instantiating fault, so we need | |
993 | * to look twice if a mapped file page is used more | |
994 | * than once. | |
995 | * | |
996 | * Mark it and spare it for another trip around the | |
997 | * inactive list. Another page table reference will | |
998 | * lead to its activation. | |
999 | * | |
1000 | * Note: the mark is set for activated pages as well | |
1001 | * so that recently deactivated but used pages are | |
1002 | * quickly recovered. | |
1003 | */ | |
1004 | SetPageReferenced(page); | |
1005 | ||
34dbc67a | 1006 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
1007 | return PAGEREF_ACTIVATE; |
1008 | ||
c909e993 KK |
1009 | /* |
1010 | * Activate file-backed executable pages after first usage. | |
1011 | */ | |
1012 | if (vm_flags & VM_EXEC) | |
1013 | return PAGEREF_ACTIVATE; | |
1014 | ||
64574746 JW |
1015 | return PAGEREF_KEEP; |
1016 | } | |
dfc8d636 JW |
1017 | |
1018 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 1019 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
1020 | return PAGEREF_RECLAIM_CLEAN; |
1021 | ||
1022 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
1023 | } |
1024 | ||
e2be15f6 MG |
1025 | /* Check if a page is dirty or under writeback */ |
1026 | static void page_check_dirty_writeback(struct page *page, | |
1027 | bool *dirty, bool *writeback) | |
1028 | { | |
b4597226 MG |
1029 | struct address_space *mapping; |
1030 | ||
e2be15f6 MG |
1031 | /* |
1032 | * Anonymous pages are not handled by flushers and must be written | |
1033 | * from reclaim context. Do not stall reclaim based on them | |
1034 | */ | |
802a3a92 SL |
1035 | if (!page_is_file_cache(page) || |
1036 | (PageAnon(page) && !PageSwapBacked(page))) { | |
e2be15f6 MG |
1037 | *dirty = false; |
1038 | *writeback = false; | |
1039 | return; | |
1040 | } | |
1041 | ||
1042 | /* By default assume that the page flags are accurate */ | |
1043 | *dirty = PageDirty(page); | |
1044 | *writeback = PageWriteback(page); | |
b4597226 MG |
1045 | |
1046 | /* Verify dirty/writeback state if the filesystem supports it */ | |
1047 | if (!page_has_private(page)) | |
1048 | return; | |
1049 | ||
1050 | mapping = page_mapping(page); | |
1051 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
1052 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
1053 | } |
1054 | ||
1da177e4 | 1055 | /* |
1742f19f | 1056 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 1057 | */ |
1742f19f | 1058 | static unsigned long shrink_page_list(struct list_head *page_list, |
599d0c95 | 1059 | struct pglist_data *pgdat, |
f84f6e2b | 1060 | struct scan_control *sc, |
02c6de8d | 1061 | enum ttu_flags ttu_flags, |
3c710c1a | 1062 | struct reclaim_stat *stat, |
02c6de8d | 1063 | bool force_reclaim) |
1da177e4 LT |
1064 | { |
1065 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 1066 | LIST_HEAD(free_pages); |
1da177e4 | 1067 | int pgactivate = 0; |
3c710c1a MH |
1068 | unsigned nr_unqueued_dirty = 0; |
1069 | unsigned nr_dirty = 0; | |
1070 | unsigned nr_congested = 0; | |
1071 | unsigned nr_reclaimed = 0; | |
1072 | unsigned nr_writeback = 0; | |
1073 | unsigned nr_immediate = 0; | |
5bccd166 MH |
1074 | unsigned nr_ref_keep = 0; |
1075 | unsigned nr_unmap_fail = 0; | |
1da177e4 LT |
1076 | |
1077 | cond_resched(); | |
1078 | ||
1da177e4 LT |
1079 | while (!list_empty(page_list)) { |
1080 | struct address_space *mapping; | |
1081 | struct page *page; | |
1082 | int may_enter_fs; | |
02c6de8d | 1083 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
e2be15f6 | 1084 | bool dirty, writeback; |
1da177e4 LT |
1085 | |
1086 | cond_resched(); | |
1087 | ||
1088 | page = lru_to_page(page_list); | |
1089 | list_del(&page->lru); | |
1090 | ||
529ae9aa | 1091 | if (!trylock_page(page)) |
1da177e4 LT |
1092 | goto keep; |
1093 | ||
309381fe | 1094 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 LT |
1095 | |
1096 | sc->nr_scanned++; | |
80e43426 | 1097 | |
39b5f29a | 1098 | if (unlikely(!page_evictable(page))) |
ad6b6704 | 1099 | goto activate_locked; |
894bc310 | 1100 | |
a6dc60f8 | 1101 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
1102 | goto keep_locked; |
1103 | ||
1da177e4 | 1104 | /* Double the slab pressure for mapped and swapcache pages */ |
802a3a92 SL |
1105 | if ((page_mapped(page) || PageSwapCache(page)) && |
1106 | !(PageAnon(page) && !PageSwapBacked(page))) | |
1da177e4 LT |
1107 | sc->nr_scanned++; |
1108 | ||
c661b078 AW |
1109 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
1110 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
1111 | ||
e2be15f6 | 1112 | /* |
894befec | 1113 | * The number of dirty pages determines if a node is marked |
e2be15f6 MG |
1114 | * reclaim_congested which affects wait_iff_congested. kswapd |
1115 | * will stall and start writing pages if the tail of the LRU | |
1116 | * is all dirty unqueued pages. | |
1117 | */ | |
1118 | page_check_dirty_writeback(page, &dirty, &writeback); | |
1119 | if (dirty || writeback) | |
1120 | nr_dirty++; | |
1121 | ||
1122 | if (dirty && !writeback) | |
1123 | nr_unqueued_dirty++; | |
1124 | ||
d04e8acd MG |
1125 | /* |
1126 | * Treat this page as congested if the underlying BDI is or if | |
1127 | * pages are cycling through the LRU so quickly that the | |
1128 | * pages marked for immediate reclaim are making it to the | |
1129 | * end of the LRU a second time. | |
1130 | */ | |
e2be15f6 | 1131 | mapping = page_mapping(page); |
1da58ee2 | 1132 | if (((dirty || writeback) && mapping && |
703c2708 | 1133 | inode_write_congested(mapping->host)) || |
d04e8acd | 1134 | (writeback && PageReclaim(page))) |
e2be15f6 MG |
1135 | nr_congested++; |
1136 | ||
283aba9f MG |
1137 | /* |
1138 | * If a page at the tail of the LRU is under writeback, there | |
1139 | * are three cases to consider. | |
1140 | * | |
1141 | * 1) If reclaim is encountering an excessive number of pages | |
1142 | * under writeback and this page is both under writeback and | |
1143 | * PageReclaim then it indicates that pages are being queued | |
1144 | * for IO but are being recycled through the LRU before the | |
1145 | * IO can complete. Waiting on the page itself risks an | |
1146 | * indefinite stall if it is impossible to writeback the | |
1147 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
1148 | * note that the LRU is being scanned too quickly and the |
1149 | * caller can stall after page list has been processed. | |
283aba9f | 1150 | * |
97c9341f | 1151 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
1152 | * not marked for immediate reclaim, or the caller does not |
1153 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
1154 | * not to fs). In this case mark the page for immediate | |
97c9341f | 1155 | * reclaim and continue scanning. |
283aba9f | 1156 | * |
ecf5fc6e MH |
1157 | * Require may_enter_fs because we would wait on fs, which |
1158 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
1159 | * enter reclaim, and deadlock if it waits on a page for |
1160 | * which it is needed to do the write (loop masks off | |
1161 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1162 | * would probably show more reasons. | |
1163 | * | |
7fadc820 | 1164 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1165 | * PageReclaim. memcg does not have any dirty pages |
1166 | * throttling so we could easily OOM just because too many | |
1167 | * pages are in writeback and there is nothing else to | |
1168 | * reclaim. Wait for the writeback to complete. | |
c55e8d03 JW |
1169 | * |
1170 | * In cases 1) and 2) we activate the pages to get them out of | |
1171 | * the way while we continue scanning for clean pages on the | |
1172 | * inactive list and refilling from the active list. The | |
1173 | * observation here is that waiting for disk writes is more | |
1174 | * expensive than potentially causing reloads down the line. | |
1175 | * Since they're marked for immediate reclaim, they won't put | |
1176 | * memory pressure on the cache working set any longer than it | |
1177 | * takes to write them to disk. | |
283aba9f | 1178 | */ |
c661b078 | 1179 | if (PageWriteback(page)) { |
283aba9f MG |
1180 | /* Case 1 above */ |
1181 | if (current_is_kswapd() && | |
1182 | PageReclaim(page) && | |
599d0c95 | 1183 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
b1a6f21e | 1184 | nr_immediate++; |
c55e8d03 | 1185 | goto activate_locked; |
283aba9f MG |
1186 | |
1187 | /* Case 2 above */ | |
97c9341f | 1188 | } else if (sane_reclaim(sc) || |
ecf5fc6e | 1189 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1190 | /* |
1191 | * This is slightly racy - end_page_writeback() | |
1192 | * might have just cleared PageReclaim, then | |
1193 | * setting PageReclaim here end up interpreted | |
1194 | * as PageReadahead - but that does not matter | |
1195 | * enough to care. What we do want is for this | |
1196 | * page to have PageReclaim set next time memcg | |
1197 | * reclaim reaches the tests above, so it will | |
1198 | * then wait_on_page_writeback() to avoid OOM; | |
1199 | * and it's also appropriate in global reclaim. | |
1200 | */ | |
1201 | SetPageReclaim(page); | |
e62e384e | 1202 | nr_writeback++; |
c55e8d03 | 1203 | goto activate_locked; |
283aba9f MG |
1204 | |
1205 | /* Case 3 above */ | |
1206 | } else { | |
7fadc820 | 1207 | unlock_page(page); |
283aba9f | 1208 | wait_on_page_writeback(page); |
7fadc820 HD |
1209 | /* then go back and try same page again */ |
1210 | list_add_tail(&page->lru, page_list); | |
1211 | continue; | |
e62e384e | 1212 | } |
c661b078 | 1213 | } |
1da177e4 | 1214 | |
02c6de8d MK |
1215 | if (!force_reclaim) |
1216 | references = page_check_references(page, sc); | |
1217 | ||
dfc8d636 JW |
1218 | switch (references) { |
1219 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1220 | goto activate_locked; |
64574746 | 1221 | case PAGEREF_KEEP: |
5bccd166 | 1222 | nr_ref_keep++; |
64574746 | 1223 | goto keep_locked; |
dfc8d636 JW |
1224 | case PAGEREF_RECLAIM: |
1225 | case PAGEREF_RECLAIM_CLEAN: | |
1226 | ; /* try to reclaim the page below */ | |
1227 | } | |
1da177e4 | 1228 | |
1da177e4 LT |
1229 | /* |
1230 | * Anonymous process memory has backing store? | |
1231 | * Try to allocate it some swap space here. | |
802a3a92 | 1232 | * Lazyfree page could be freed directly |
1da177e4 | 1233 | */ |
bd4c82c2 HY |
1234 | if (PageAnon(page) && PageSwapBacked(page)) { |
1235 | if (!PageSwapCache(page)) { | |
1236 | if (!(sc->gfp_mask & __GFP_IO)) | |
1237 | goto keep_locked; | |
1238 | if (PageTransHuge(page)) { | |
1239 | /* cannot split THP, skip it */ | |
1240 | if (!can_split_huge_page(page, NULL)) | |
1241 | goto activate_locked; | |
1242 | /* | |
1243 | * Split pages without a PMD map right | |
1244 | * away. Chances are some or all of the | |
1245 | * tail pages can be freed without IO. | |
1246 | */ | |
1247 | if (!compound_mapcount(page) && | |
1248 | split_huge_page_to_list(page, | |
1249 | page_list)) | |
1250 | goto activate_locked; | |
1251 | } | |
1252 | if (!add_to_swap(page)) { | |
1253 | if (!PageTransHuge(page)) | |
1254 | goto activate_locked; | |
1255 | /* Fallback to swap normal pages */ | |
1256 | if (split_huge_page_to_list(page, | |
1257 | page_list)) | |
1258 | goto activate_locked; | |
fe490cc0 HY |
1259 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1260 | count_vm_event(THP_SWPOUT_FALLBACK); | |
1261 | #endif | |
bd4c82c2 HY |
1262 | if (!add_to_swap(page)) |
1263 | goto activate_locked; | |
1264 | } | |
0f074658 | 1265 | |
bd4c82c2 | 1266 | may_enter_fs = 1; |
1da177e4 | 1267 | |
bd4c82c2 HY |
1268 | /* Adding to swap updated mapping */ |
1269 | mapping = page_mapping(page); | |
1270 | } | |
7751b2da KS |
1271 | } else if (unlikely(PageTransHuge(page))) { |
1272 | /* Split file THP */ | |
1273 | if (split_huge_page_to_list(page, page_list)) | |
1274 | goto keep_locked; | |
e2be15f6 | 1275 | } |
1da177e4 LT |
1276 | |
1277 | /* | |
1278 | * The page is mapped into the page tables of one or more | |
1279 | * processes. Try to unmap it here. | |
1280 | */ | |
802a3a92 | 1281 | if (page_mapped(page)) { |
bd4c82c2 HY |
1282 | enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH; |
1283 | ||
1284 | if (unlikely(PageTransHuge(page))) | |
1285 | flags |= TTU_SPLIT_HUGE_PMD; | |
1286 | if (!try_to_unmap(page, flags)) { | |
5bccd166 | 1287 | nr_unmap_fail++; |
1da177e4 | 1288 | goto activate_locked; |
1da177e4 LT |
1289 | } |
1290 | } | |
1291 | ||
1292 | if (PageDirty(page)) { | |
ee72886d | 1293 | /* |
4eda4823 JW |
1294 | * Only kswapd can writeback filesystem pages |
1295 | * to avoid risk of stack overflow. But avoid | |
1296 | * injecting inefficient single-page IO into | |
1297 | * flusher writeback as much as possible: only | |
1298 | * write pages when we've encountered many | |
1299 | * dirty pages, and when we've already scanned | |
1300 | * the rest of the LRU for clean pages and see | |
1301 | * the same dirty pages again (PageReclaim). | |
ee72886d | 1302 | */ |
f84f6e2b | 1303 | if (page_is_file_cache(page) && |
4eda4823 JW |
1304 | (!current_is_kswapd() || !PageReclaim(page) || |
1305 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { | |
49ea7eb6 MG |
1306 | /* |
1307 | * Immediately reclaim when written back. | |
1308 | * Similar in principal to deactivate_page() | |
1309 | * except we already have the page isolated | |
1310 | * and know it's dirty | |
1311 | */ | |
c4a25635 | 1312 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
49ea7eb6 MG |
1313 | SetPageReclaim(page); |
1314 | ||
c55e8d03 | 1315 | goto activate_locked; |
ee72886d MG |
1316 | } |
1317 | ||
dfc8d636 | 1318 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1319 | goto keep_locked; |
4dd4b920 | 1320 | if (!may_enter_fs) |
1da177e4 | 1321 | goto keep_locked; |
52a8363e | 1322 | if (!sc->may_writepage) |
1da177e4 LT |
1323 | goto keep_locked; |
1324 | ||
d950c947 MG |
1325 | /* |
1326 | * Page is dirty. Flush the TLB if a writable entry | |
1327 | * potentially exists to avoid CPU writes after IO | |
1328 | * starts and then write it out here. | |
1329 | */ | |
1330 | try_to_unmap_flush_dirty(); | |
7d3579e8 | 1331 | switch (pageout(page, mapping, sc)) { |
1da177e4 LT |
1332 | case PAGE_KEEP: |
1333 | goto keep_locked; | |
1334 | case PAGE_ACTIVATE: | |
1335 | goto activate_locked; | |
1336 | case PAGE_SUCCESS: | |
7d3579e8 | 1337 | if (PageWriteback(page)) |
41ac1999 | 1338 | goto keep; |
7d3579e8 | 1339 | if (PageDirty(page)) |
1da177e4 | 1340 | goto keep; |
7d3579e8 | 1341 | |
1da177e4 LT |
1342 | /* |
1343 | * A synchronous write - probably a ramdisk. Go | |
1344 | * ahead and try to reclaim the page. | |
1345 | */ | |
529ae9aa | 1346 | if (!trylock_page(page)) |
1da177e4 LT |
1347 | goto keep; |
1348 | if (PageDirty(page) || PageWriteback(page)) | |
1349 | goto keep_locked; | |
1350 | mapping = page_mapping(page); | |
1351 | case PAGE_CLEAN: | |
1352 | ; /* try to free the page below */ | |
1353 | } | |
1354 | } | |
1355 | ||
1356 | /* | |
1357 | * If the page has buffers, try to free the buffer mappings | |
1358 | * associated with this page. If we succeed we try to free | |
1359 | * the page as well. | |
1360 | * | |
1361 | * We do this even if the page is PageDirty(). | |
1362 | * try_to_release_page() does not perform I/O, but it is | |
1363 | * possible for a page to have PageDirty set, but it is actually | |
1364 | * clean (all its buffers are clean). This happens if the | |
1365 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1366 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1367 | * try_to_release_page() will discover that cleanness and will |
1368 | * drop the buffers and mark the page clean - it can be freed. | |
1369 | * | |
1370 | * Rarely, pages can have buffers and no ->mapping. These are | |
1371 | * the pages which were not successfully invalidated in | |
1372 | * truncate_complete_page(). We try to drop those buffers here | |
1373 | * and if that worked, and the page is no longer mapped into | |
1374 | * process address space (page_count == 1) it can be freed. | |
1375 | * Otherwise, leave the page on the LRU so it is swappable. | |
1376 | */ | |
266cf658 | 1377 | if (page_has_private(page)) { |
1da177e4 LT |
1378 | if (!try_to_release_page(page, sc->gfp_mask)) |
1379 | goto activate_locked; | |
e286781d NP |
1380 | if (!mapping && page_count(page) == 1) { |
1381 | unlock_page(page); | |
1382 | if (put_page_testzero(page)) | |
1383 | goto free_it; | |
1384 | else { | |
1385 | /* | |
1386 | * rare race with speculative reference. | |
1387 | * the speculative reference will free | |
1388 | * this page shortly, so we may | |
1389 | * increment nr_reclaimed here (and | |
1390 | * leave it off the LRU). | |
1391 | */ | |
1392 | nr_reclaimed++; | |
1393 | continue; | |
1394 | } | |
1395 | } | |
1da177e4 LT |
1396 | } |
1397 | ||
802a3a92 SL |
1398 | if (PageAnon(page) && !PageSwapBacked(page)) { |
1399 | /* follow __remove_mapping for reference */ | |
1400 | if (!page_ref_freeze(page, 1)) | |
1401 | goto keep_locked; | |
1402 | if (PageDirty(page)) { | |
1403 | page_ref_unfreeze(page, 1); | |
1404 | goto keep_locked; | |
1405 | } | |
1da177e4 | 1406 | |
802a3a92 | 1407 | count_vm_event(PGLAZYFREED); |
2262185c | 1408 | count_memcg_page_event(page, PGLAZYFREED); |
802a3a92 SL |
1409 | } else if (!mapping || !__remove_mapping(mapping, page, true)) |
1410 | goto keep_locked; | |
a978d6f5 NP |
1411 | /* |
1412 | * At this point, we have no other references and there is | |
1413 | * no way to pick any more up (removed from LRU, removed | |
1414 | * from pagecache). Can use non-atomic bitops now (and | |
1415 | * we obviously don't have to worry about waking up a process | |
1416 | * waiting on the page lock, because there are no references. | |
1417 | */ | |
48c935ad | 1418 | __ClearPageLocked(page); |
e286781d | 1419 | free_it: |
05ff5137 | 1420 | nr_reclaimed++; |
abe4c3b5 MG |
1421 | |
1422 | /* | |
1423 | * Is there need to periodically free_page_list? It would | |
1424 | * appear not as the counts should be low | |
1425 | */ | |
bd4c82c2 HY |
1426 | if (unlikely(PageTransHuge(page))) { |
1427 | mem_cgroup_uncharge(page); | |
1428 | (*get_compound_page_dtor(page))(page); | |
1429 | } else | |
1430 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
1431 | continue; |
1432 | ||
1433 | activate_locked: | |
68a22394 | 1434 | /* Not a candidate for swapping, so reclaim swap space. */ |
ad6b6704 MK |
1435 | if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || |
1436 | PageMlocked(page))) | |
a2c43eed | 1437 | try_to_free_swap(page); |
309381fe | 1438 | VM_BUG_ON_PAGE(PageActive(page), page); |
ad6b6704 MK |
1439 | if (!PageMlocked(page)) { |
1440 | SetPageActive(page); | |
1441 | pgactivate++; | |
2262185c | 1442 | count_memcg_page_event(page, PGACTIVATE); |
ad6b6704 | 1443 | } |
1da177e4 LT |
1444 | keep_locked: |
1445 | unlock_page(page); | |
1446 | keep: | |
1447 | list_add(&page->lru, &ret_pages); | |
309381fe | 1448 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1449 | } |
abe4c3b5 | 1450 | |
747db954 | 1451 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1452 | try_to_unmap_flush(); |
2d4894b5 | 1453 | free_unref_page_list(&free_pages); |
abe4c3b5 | 1454 | |
1da177e4 | 1455 | list_splice(&ret_pages, page_list); |
f8891e5e | 1456 | count_vm_events(PGACTIVATE, pgactivate); |
0a31bc97 | 1457 | |
3c710c1a MH |
1458 | if (stat) { |
1459 | stat->nr_dirty = nr_dirty; | |
1460 | stat->nr_congested = nr_congested; | |
1461 | stat->nr_unqueued_dirty = nr_unqueued_dirty; | |
1462 | stat->nr_writeback = nr_writeback; | |
1463 | stat->nr_immediate = nr_immediate; | |
5bccd166 MH |
1464 | stat->nr_activate = pgactivate; |
1465 | stat->nr_ref_keep = nr_ref_keep; | |
1466 | stat->nr_unmap_fail = nr_unmap_fail; | |
3c710c1a | 1467 | } |
05ff5137 | 1468 | return nr_reclaimed; |
1da177e4 LT |
1469 | } |
1470 | ||
02c6de8d MK |
1471 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1472 | struct list_head *page_list) | |
1473 | { | |
1474 | struct scan_control sc = { | |
1475 | .gfp_mask = GFP_KERNEL, | |
1476 | .priority = DEF_PRIORITY, | |
1477 | .may_unmap = 1, | |
1478 | }; | |
3c710c1a | 1479 | unsigned long ret; |
02c6de8d MK |
1480 | struct page *page, *next; |
1481 | LIST_HEAD(clean_pages); | |
1482 | ||
1483 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e | 1484 | if (page_is_file_cache(page) && !PageDirty(page) && |
b1123ea6 | 1485 | !__PageMovable(page)) { |
02c6de8d MK |
1486 | ClearPageActive(page); |
1487 | list_move(&page->lru, &clean_pages); | |
1488 | } | |
1489 | } | |
1490 | ||
599d0c95 | 1491 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
a128ca71 | 1492 | TTU_IGNORE_ACCESS, NULL, true); |
02c6de8d | 1493 | list_splice(&clean_pages, page_list); |
599d0c95 | 1494 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1495 | return ret; |
1496 | } | |
1497 | ||
5ad333eb AW |
1498 | /* |
1499 | * Attempt to remove the specified page from its LRU. Only take this page | |
1500 | * if it is of the appropriate PageActive status. Pages which are being | |
1501 | * freed elsewhere are also ignored. | |
1502 | * | |
1503 | * page: page to consider | |
1504 | * mode: one of the LRU isolation modes defined above | |
1505 | * | |
1506 | * returns 0 on success, -ve errno on failure. | |
1507 | */ | |
f3fd4a61 | 1508 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1509 | { |
1510 | int ret = -EINVAL; | |
1511 | ||
1512 | /* Only take pages on the LRU. */ | |
1513 | if (!PageLRU(page)) | |
1514 | return ret; | |
1515 | ||
e46a2879 MK |
1516 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1517 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1518 | return ret; |
1519 | ||
5ad333eb | 1520 | ret = -EBUSY; |
08e552c6 | 1521 | |
c8244935 MG |
1522 | /* |
1523 | * To minimise LRU disruption, the caller can indicate that it only | |
1524 | * wants to isolate pages it will be able to operate on without | |
1525 | * blocking - clean pages for the most part. | |
1526 | * | |
c8244935 MG |
1527 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
1528 | * that it is possible to migrate without blocking | |
1529 | */ | |
1276ad68 | 1530 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
c8244935 MG |
1531 | /* All the caller can do on PageWriteback is block */ |
1532 | if (PageWriteback(page)) | |
1533 | return ret; | |
1534 | ||
1535 | if (PageDirty(page)) { | |
1536 | struct address_space *mapping; | |
69d763fc | 1537 | bool migrate_dirty; |
c8244935 | 1538 | |
c8244935 MG |
1539 | /* |
1540 | * Only pages without mappings or that have a | |
1541 | * ->migratepage callback are possible to migrate | |
69d763fc MG |
1542 | * without blocking. However, we can be racing with |
1543 | * truncation so it's necessary to lock the page | |
1544 | * to stabilise the mapping as truncation holds | |
1545 | * the page lock until after the page is removed | |
1546 | * from the page cache. | |
c8244935 | 1547 | */ |
69d763fc MG |
1548 | if (!trylock_page(page)) |
1549 | return ret; | |
1550 | ||
c8244935 | 1551 | mapping = page_mapping(page); |
145e1a71 | 1552 | migrate_dirty = !mapping || mapping->a_ops->migratepage; |
69d763fc MG |
1553 | unlock_page(page); |
1554 | if (!migrate_dirty) | |
c8244935 MG |
1555 | return ret; |
1556 | } | |
1557 | } | |
39deaf85 | 1558 | |
f80c0673 MK |
1559 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1560 | return ret; | |
1561 | ||
5ad333eb AW |
1562 | if (likely(get_page_unless_zero(page))) { |
1563 | /* | |
1564 | * Be careful not to clear PageLRU until after we're | |
1565 | * sure the page is not being freed elsewhere -- the | |
1566 | * page release code relies on it. | |
1567 | */ | |
1568 | ClearPageLRU(page); | |
1569 | ret = 0; | |
1570 | } | |
1571 | ||
1572 | return ret; | |
1573 | } | |
1574 | ||
7ee36a14 MG |
1575 | |
1576 | /* | |
1577 | * Update LRU sizes after isolating pages. The LRU size updates must | |
1578 | * be complete before mem_cgroup_update_lru_size due to a santity check. | |
1579 | */ | |
1580 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 1581 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 1582 | { |
7ee36a14 MG |
1583 | int zid; |
1584 | ||
7ee36a14 MG |
1585 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1586 | if (!nr_zone_taken[zid]) | |
1587 | continue; | |
1588 | ||
1589 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); | |
7ee36a14 | 1590 | #ifdef CONFIG_MEMCG |
b4536f0c | 1591 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
7ee36a14 | 1592 | #endif |
b4536f0c MH |
1593 | } |
1594 | ||
7ee36a14 MG |
1595 | } |
1596 | ||
1da177e4 | 1597 | /* |
a52633d8 | 1598 | * zone_lru_lock is heavily contended. Some of the functions that |
1da177e4 LT |
1599 | * shrink the lists perform better by taking out a batch of pages |
1600 | * and working on them outside the LRU lock. | |
1601 | * | |
1602 | * For pagecache intensive workloads, this function is the hottest | |
1603 | * spot in the kernel (apart from copy_*_user functions). | |
1604 | * | |
1605 | * Appropriate locks must be held before calling this function. | |
1606 | * | |
791b48b6 | 1607 | * @nr_to_scan: The number of eligible pages to look through on the list. |
5dc35979 | 1608 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1609 | * @dst: The temp list to put pages on to. |
f626012d | 1610 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1611 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1612 | * @mode: One of the LRU isolation modes |
3cb99451 | 1613 | * @lru: LRU list id for isolating |
1da177e4 LT |
1614 | * |
1615 | * returns how many pages were moved onto *@dst. | |
1616 | */ | |
69e05944 | 1617 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1618 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1619 | unsigned long *nr_scanned, struct scan_control *sc, |
3cb99451 | 1620 | isolate_mode_t mode, enum lru_list lru) |
1da177e4 | 1621 | { |
75b00af7 | 1622 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1623 | unsigned long nr_taken = 0; |
599d0c95 | 1624 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 1625 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
3db65812 | 1626 | unsigned long skipped = 0; |
791b48b6 | 1627 | unsigned long scan, total_scan, nr_pages; |
b2e18757 | 1628 | LIST_HEAD(pages_skipped); |
1da177e4 | 1629 | |
791b48b6 MK |
1630 | scan = 0; |
1631 | for (total_scan = 0; | |
1632 | scan < nr_to_scan && nr_taken < nr_to_scan && !list_empty(src); | |
1633 | total_scan++) { | |
5ad333eb | 1634 | struct page *page; |
5ad333eb | 1635 | |
1da177e4 LT |
1636 | page = lru_to_page(src); |
1637 | prefetchw_prev_lru_page(page, src, flags); | |
1638 | ||
309381fe | 1639 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1640 | |
b2e18757 MG |
1641 | if (page_zonenum(page) > sc->reclaim_idx) { |
1642 | list_move(&page->lru, &pages_skipped); | |
7cc30fcf | 1643 | nr_skipped[page_zonenum(page)]++; |
b2e18757 MG |
1644 | continue; |
1645 | } | |
1646 | ||
791b48b6 MK |
1647 | /* |
1648 | * Do not count skipped pages because that makes the function | |
1649 | * return with no isolated pages if the LRU mostly contains | |
1650 | * ineligible pages. This causes the VM to not reclaim any | |
1651 | * pages, triggering a premature OOM. | |
1652 | */ | |
1653 | scan++; | |
f3fd4a61 | 1654 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1655 | case 0: |
599d0c95 MG |
1656 | nr_pages = hpage_nr_pages(page); |
1657 | nr_taken += nr_pages; | |
1658 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
5ad333eb | 1659 | list_move(&page->lru, dst); |
5ad333eb AW |
1660 | break; |
1661 | ||
1662 | case -EBUSY: | |
1663 | /* else it is being freed elsewhere */ | |
1664 | list_move(&page->lru, src); | |
1665 | continue; | |
46453a6e | 1666 | |
5ad333eb AW |
1667 | default: |
1668 | BUG(); | |
1669 | } | |
1da177e4 LT |
1670 | } |
1671 | ||
b2e18757 MG |
1672 | /* |
1673 | * Splice any skipped pages to the start of the LRU list. Note that | |
1674 | * this disrupts the LRU order when reclaiming for lower zones but | |
1675 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
1676 | * scanning would soon rescan the same pages to skip and put the | |
1677 | * system at risk of premature OOM. | |
1678 | */ | |
7cc30fcf MG |
1679 | if (!list_empty(&pages_skipped)) { |
1680 | int zid; | |
1681 | ||
3db65812 | 1682 | list_splice(&pages_skipped, src); |
7cc30fcf MG |
1683 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1684 | if (!nr_skipped[zid]) | |
1685 | continue; | |
1686 | ||
1687 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 1688 | skipped += nr_skipped[zid]; |
7cc30fcf MG |
1689 | } |
1690 | } | |
791b48b6 | 1691 | *nr_scanned = total_scan; |
1265e3a6 | 1692 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
791b48b6 | 1693 | total_scan, skipped, nr_taken, mode, lru); |
b4536f0c | 1694 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
1695 | return nr_taken; |
1696 | } | |
1697 | ||
62695a84 NP |
1698 | /** |
1699 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1700 | * @page: page to isolate from its LRU list | |
1701 | * | |
1702 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1703 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1704 | * | |
1705 | * Returns 0 if the page was removed from an LRU list. | |
1706 | * Returns -EBUSY if the page was not on an LRU list. | |
1707 | * | |
1708 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1709 | * the active list, it will have PageActive set. If it was found on |
1710 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1711 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1712 | * |
1713 | * The vmstat statistic corresponding to the list on which the page was | |
1714 | * found will be decremented. | |
1715 | * | |
1716 | * Restrictions: | |
a5d09bed | 1717 | * |
62695a84 NP |
1718 | * (1) Must be called with an elevated refcount on the page. This is a |
1719 | * fundamentnal difference from isolate_lru_pages (which is called | |
1720 | * without a stable reference). | |
1721 | * (2) the lru_lock must not be held. | |
1722 | * (3) interrupts must be enabled. | |
1723 | */ | |
1724 | int isolate_lru_page(struct page *page) | |
1725 | { | |
1726 | int ret = -EBUSY; | |
1727 | ||
309381fe | 1728 | VM_BUG_ON_PAGE(!page_count(page), page); |
cf2a82ee | 1729 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
0c917313 | 1730 | |
62695a84 NP |
1731 | if (PageLRU(page)) { |
1732 | struct zone *zone = page_zone(page); | |
fa9add64 | 1733 | struct lruvec *lruvec; |
62695a84 | 1734 | |
a52633d8 | 1735 | spin_lock_irq(zone_lru_lock(zone)); |
599d0c95 | 1736 | lruvec = mem_cgroup_page_lruvec(page, zone->zone_pgdat); |
0c917313 | 1737 | if (PageLRU(page)) { |
894bc310 | 1738 | int lru = page_lru(page); |
0c917313 | 1739 | get_page(page); |
62695a84 | 1740 | ClearPageLRU(page); |
fa9add64 HD |
1741 | del_page_from_lru_list(page, lruvec, lru); |
1742 | ret = 0; | |
62695a84 | 1743 | } |
a52633d8 | 1744 | spin_unlock_irq(zone_lru_lock(zone)); |
62695a84 NP |
1745 | } |
1746 | return ret; | |
1747 | } | |
1748 | ||
35cd7815 | 1749 | /* |
d37dd5dc FW |
1750 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1751 | * then get resheduled. When there are massive number of tasks doing page | |
1752 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1753 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1754 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 1755 | */ |
599d0c95 | 1756 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
1757 | struct scan_control *sc) |
1758 | { | |
1759 | unsigned long inactive, isolated; | |
1760 | ||
1761 | if (current_is_kswapd()) | |
1762 | return 0; | |
1763 | ||
97c9341f | 1764 | if (!sane_reclaim(sc)) |
35cd7815 RR |
1765 | return 0; |
1766 | ||
1767 | if (file) { | |
599d0c95 MG |
1768 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1769 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 1770 | } else { |
599d0c95 MG |
1771 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1772 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
1773 | } |
1774 | ||
3cf23841 FW |
1775 | /* |
1776 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1777 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1778 | * deadlock. | |
1779 | */ | |
d0164adc | 1780 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1781 | inactive >>= 3; |
1782 | ||
35cd7815 RR |
1783 | return isolated > inactive; |
1784 | } | |
1785 | ||
66635629 | 1786 | static noinline_for_stack void |
75b00af7 | 1787 | putback_inactive_pages(struct lruvec *lruvec, struct list_head *page_list) |
66635629 | 1788 | { |
27ac81d8 | 1789 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
599d0c95 | 1790 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3f79768f | 1791 | LIST_HEAD(pages_to_free); |
66635629 | 1792 | |
66635629 MG |
1793 | /* |
1794 | * Put back any unfreeable pages. | |
1795 | */ | |
66635629 | 1796 | while (!list_empty(page_list)) { |
3f79768f | 1797 | struct page *page = lru_to_page(page_list); |
66635629 | 1798 | int lru; |
3f79768f | 1799 | |
309381fe | 1800 | VM_BUG_ON_PAGE(PageLRU(page), page); |
66635629 | 1801 | list_del(&page->lru); |
39b5f29a | 1802 | if (unlikely(!page_evictable(page))) { |
599d0c95 | 1803 | spin_unlock_irq(&pgdat->lru_lock); |
66635629 | 1804 | putback_lru_page(page); |
599d0c95 | 1805 | spin_lock_irq(&pgdat->lru_lock); |
66635629 MG |
1806 | continue; |
1807 | } | |
fa9add64 | 1808 | |
599d0c95 | 1809 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 1810 | |
7a608572 | 1811 | SetPageLRU(page); |
66635629 | 1812 | lru = page_lru(page); |
fa9add64 HD |
1813 | add_page_to_lru_list(page, lruvec, lru); |
1814 | ||
66635629 MG |
1815 | if (is_active_lru(lru)) { |
1816 | int file = is_file_lru(lru); | |
9992af10 RR |
1817 | int numpages = hpage_nr_pages(page); |
1818 | reclaim_stat->recent_rotated[file] += numpages; | |
66635629 | 1819 | } |
2bcf8879 HD |
1820 | if (put_page_testzero(page)) { |
1821 | __ClearPageLRU(page); | |
1822 | __ClearPageActive(page); | |
fa9add64 | 1823 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1824 | |
1825 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1826 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1827 | mem_cgroup_uncharge(page); |
2bcf8879 | 1828 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1829 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1830 | } else |
1831 | list_add(&page->lru, &pages_to_free); | |
66635629 MG |
1832 | } |
1833 | } | |
66635629 | 1834 | |
3f79768f HD |
1835 | /* |
1836 | * To save our caller's stack, now use input list for pages to free. | |
1837 | */ | |
1838 | list_splice(&pages_to_free, page_list); | |
66635629 MG |
1839 | } |
1840 | ||
399ba0b9 N |
1841 | /* |
1842 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1843 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1844 | * In that case we should only throttle if the backing device it is | |
1845 | * writing to is congested. In other cases it is safe to throttle. | |
1846 | */ | |
1847 | static int current_may_throttle(void) | |
1848 | { | |
1849 | return !(current->flags & PF_LESS_THROTTLE) || | |
1850 | current->backing_dev_info == NULL || | |
1851 | bdi_write_congested(current->backing_dev_info); | |
1852 | } | |
1853 | ||
1da177e4 | 1854 | /* |
b2e18757 | 1855 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 1856 | * of reclaimed pages |
1da177e4 | 1857 | */ |
66635629 | 1858 | static noinline_for_stack unsigned long |
1a93be0e | 1859 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1860 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1861 | { |
1862 | LIST_HEAD(page_list); | |
e247dbce | 1863 | unsigned long nr_scanned; |
05ff5137 | 1864 | unsigned long nr_reclaimed = 0; |
e247dbce | 1865 | unsigned long nr_taken; |
3c710c1a | 1866 | struct reclaim_stat stat = {}; |
f3fd4a61 | 1867 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 1868 | int file = is_file_lru(lru); |
599d0c95 | 1869 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1a93be0e | 1870 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
db73ee0d | 1871 | bool stalled = false; |
78dc583d | 1872 | |
599d0c95 | 1873 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
db73ee0d MH |
1874 | if (stalled) |
1875 | return 0; | |
1876 | ||
1877 | /* wait a bit for the reclaimer. */ | |
1878 | msleep(100); | |
1879 | stalled = true; | |
35cd7815 RR |
1880 | |
1881 | /* We are about to die and free our memory. Return now. */ | |
1882 | if (fatal_signal_pending(current)) | |
1883 | return SWAP_CLUSTER_MAX; | |
1884 | } | |
1885 | ||
1da177e4 | 1886 | lru_add_drain(); |
f80c0673 MK |
1887 | |
1888 | if (!sc->may_unmap) | |
61317289 | 1889 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 1890 | |
599d0c95 | 1891 | spin_lock_irq(&pgdat->lru_lock); |
b35ea17b | 1892 | |
5dc35979 KK |
1893 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
1894 | &nr_scanned, sc, isolate_mode, lru); | |
95d918fc | 1895 | |
599d0c95 | 1896 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
9d5e6a9f | 1897 | reclaim_stat->recent_scanned[file] += nr_taken; |
95d918fc | 1898 | |
2262185c RG |
1899 | if (current_is_kswapd()) { |
1900 | if (global_reclaim(sc)) | |
599d0c95 | 1901 | __count_vm_events(PGSCAN_KSWAPD, nr_scanned); |
2262185c RG |
1902 | count_memcg_events(lruvec_memcg(lruvec), PGSCAN_KSWAPD, |
1903 | nr_scanned); | |
1904 | } else { | |
1905 | if (global_reclaim(sc)) | |
599d0c95 | 1906 | __count_vm_events(PGSCAN_DIRECT, nr_scanned); |
2262185c RG |
1907 | count_memcg_events(lruvec_memcg(lruvec), PGSCAN_DIRECT, |
1908 | nr_scanned); | |
e247dbce | 1909 | } |
599d0c95 | 1910 | spin_unlock_irq(&pgdat->lru_lock); |
b35ea17b | 1911 | |
d563c050 | 1912 | if (nr_taken == 0) |
66635629 | 1913 | return 0; |
5ad333eb | 1914 | |
a128ca71 | 1915 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0, |
3c710c1a | 1916 | &stat, false); |
c661b078 | 1917 | |
599d0c95 | 1918 | spin_lock_irq(&pgdat->lru_lock); |
3f79768f | 1919 | |
2262185c RG |
1920 | if (current_is_kswapd()) { |
1921 | if (global_reclaim(sc)) | |
599d0c95 | 1922 | __count_vm_events(PGSTEAL_KSWAPD, nr_reclaimed); |
2262185c RG |
1923 | count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_KSWAPD, |
1924 | nr_reclaimed); | |
1925 | } else { | |
1926 | if (global_reclaim(sc)) | |
599d0c95 | 1927 | __count_vm_events(PGSTEAL_DIRECT, nr_reclaimed); |
2262185c RG |
1928 | count_memcg_events(lruvec_memcg(lruvec), PGSTEAL_DIRECT, |
1929 | nr_reclaimed); | |
904249aa | 1930 | } |
a74609fa | 1931 | |
27ac81d8 | 1932 | putback_inactive_pages(lruvec, &page_list); |
3f79768f | 1933 | |
599d0c95 | 1934 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f | 1935 | |
599d0c95 | 1936 | spin_unlock_irq(&pgdat->lru_lock); |
3f79768f | 1937 | |
747db954 | 1938 | mem_cgroup_uncharge_list(&page_list); |
2d4894b5 | 1939 | free_unref_page_list(&page_list); |
e11da5b4 | 1940 | |
1c610d5f AR |
1941 | /* |
1942 | * If dirty pages are scanned that are not queued for IO, it | |
1943 | * implies that flushers are not doing their job. This can | |
1944 | * happen when memory pressure pushes dirty pages to the end of | |
1945 | * the LRU before the dirty limits are breached and the dirty | |
1946 | * data has expired. It can also happen when the proportion of | |
1947 | * dirty pages grows not through writes but through memory | |
1948 | * pressure reclaiming all the clean cache. And in some cases, | |
1949 | * the flushers simply cannot keep up with the allocation | |
1950 | * rate. Nudge the flusher threads in case they are asleep. | |
1951 | */ | |
1952 | if (stat.nr_unqueued_dirty == nr_taken) | |
1953 | wakeup_flusher_threads(WB_REASON_VMSCAN); | |
1954 | ||
d108c772 AR |
1955 | sc->nr.dirty += stat.nr_dirty; |
1956 | sc->nr.congested += stat.nr_congested; | |
1957 | sc->nr.unqueued_dirty += stat.nr_unqueued_dirty; | |
1958 | sc->nr.writeback += stat.nr_writeback; | |
1959 | sc->nr.immediate += stat.nr_immediate; | |
1960 | sc->nr.taken += nr_taken; | |
1961 | if (file) | |
1962 | sc->nr.file_taken += nr_taken; | |
8e950282 | 1963 | |
599d0c95 | 1964 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
d51d1e64 | 1965 | nr_scanned, nr_reclaimed, &stat, sc->priority, file); |
05ff5137 | 1966 | return nr_reclaimed; |
1da177e4 LT |
1967 | } |
1968 | ||
1969 | /* | |
1970 | * This moves pages from the active list to the inactive list. | |
1971 | * | |
1972 | * We move them the other way if the page is referenced by one or more | |
1973 | * processes, from rmap. | |
1974 | * | |
1975 | * If the pages are mostly unmapped, the processing is fast and it is | |
a52633d8 | 1976 | * appropriate to hold zone_lru_lock across the whole operation. But if |
1da177e4 | 1977 | * the pages are mapped, the processing is slow (page_referenced()) so we |
a52633d8 | 1978 | * should drop zone_lru_lock around each page. It's impossible to balance |
1da177e4 LT |
1979 | * this, so instead we remove the pages from the LRU while processing them. |
1980 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1981 | * nobody will play with that bit on a non-LRU page. | |
1982 | * | |
0139aa7b | 1983 | * The downside is that we have to touch page->_refcount against each page. |
1da177e4 | 1984 | * But we had to alter page->flags anyway. |
9d998b4f MH |
1985 | * |
1986 | * Returns the number of pages moved to the given lru. | |
1da177e4 | 1987 | */ |
1cfb419b | 1988 | |
9d998b4f | 1989 | static unsigned move_active_pages_to_lru(struct lruvec *lruvec, |
3eb4140f | 1990 | struct list_head *list, |
2bcf8879 | 1991 | struct list_head *pages_to_free, |
3eb4140f WF |
1992 | enum lru_list lru) |
1993 | { | |
599d0c95 | 1994 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
3eb4140f | 1995 | struct page *page; |
fa9add64 | 1996 | int nr_pages; |
9d998b4f | 1997 | int nr_moved = 0; |
3eb4140f | 1998 | |
3eb4140f WF |
1999 | while (!list_empty(list)) { |
2000 | page = lru_to_page(list); | |
599d0c95 | 2001 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
3eb4140f | 2002 | |
309381fe | 2003 | VM_BUG_ON_PAGE(PageLRU(page), page); |
3eb4140f WF |
2004 | SetPageLRU(page); |
2005 | ||
fa9add64 | 2006 | nr_pages = hpage_nr_pages(page); |
599d0c95 | 2007 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); |
925b7673 | 2008 | list_move(&page->lru, &lruvec->lists[lru]); |
3eb4140f | 2009 | |
2bcf8879 HD |
2010 | if (put_page_testzero(page)) { |
2011 | __ClearPageLRU(page); | |
2012 | __ClearPageActive(page); | |
fa9add64 | 2013 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
2014 | |
2015 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 2016 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 2017 | mem_cgroup_uncharge(page); |
2bcf8879 | 2018 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 2019 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
2020 | } else |
2021 | list_add(&page->lru, pages_to_free); | |
9d998b4f MH |
2022 | } else { |
2023 | nr_moved += nr_pages; | |
3eb4140f WF |
2024 | } |
2025 | } | |
9d5e6a9f | 2026 | |
2262185c | 2027 | if (!is_active_lru(lru)) { |
f0958906 | 2028 | __count_vm_events(PGDEACTIVATE, nr_moved); |
2262185c RG |
2029 | count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, |
2030 | nr_moved); | |
2031 | } | |
9d998b4f MH |
2032 | |
2033 | return nr_moved; | |
3eb4140f | 2034 | } |
1cfb419b | 2035 | |
f626012d | 2036 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 2037 | struct lruvec *lruvec, |
f16015fb | 2038 | struct scan_control *sc, |
9e3b2f8c | 2039 | enum lru_list lru) |
1da177e4 | 2040 | { |
44c241f1 | 2041 | unsigned long nr_taken; |
f626012d | 2042 | unsigned long nr_scanned; |
6fe6b7e3 | 2043 | unsigned long vm_flags; |
1da177e4 | 2044 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 2045 | LIST_HEAD(l_active); |
b69408e8 | 2046 | LIST_HEAD(l_inactive); |
1da177e4 | 2047 | struct page *page; |
1a93be0e | 2048 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
9d998b4f MH |
2049 | unsigned nr_deactivate, nr_activate; |
2050 | unsigned nr_rotated = 0; | |
f3fd4a61 | 2051 | isolate_mode_t isolate_mode = 0; |
3cb99451 | 2052 | int file = is_file_lru(lru); |
599d0c95 | 2053 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
2054 | |
2055 | lru_add_drain(); | |
f80c0673 MK |
2056 | |
2057 | if (!sc->may_unmap) | |
61317289 | 2058 | isolate_mode |= ISOLATE_UNMAPPED; |
f80c0673 | 2059 | |
599d0c95 | 2060 | spin_lock_irq(&pgdat->lru_lock); |
925b7673 | 2061 | |
5dc35979 KK |
2062 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
2063 | &nr_scanned, sc, isolate_mode, lru); | |
89b5fae5 | 2064 | |
599d0c95 | 2065 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
b7c46d15 | 2066 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 2067 | |
599d0c95 | 2068 | __count_vm_events(PGREFILL, nr_scanned); |
2262185c | 2069 | count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); |
9d5e6a9f | 2070 | |
599d0c95 | 2071 | spin_unlock_irq(&pgdat->lru_lock); |
1da177e4 | 2072 | |
1da177e4 LT |
2073 | while (!list_empty(&l_hold)) { |
2074 | cond_resched(); | |
2075 | page = lru_to_page(&l_hold); | |
2076 | list_del(&page->lru); | |
7e9cd484 | 2077 | |
39b5f29a | 2078 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
2079 | putback_lru_page(page); |
2080 | continue; | |
2081 | } | |
2082 | ||
cc715d99 MG |
2083 | if (unlikely(buffer_heads_over_limit)) { |
2084 | if (page_has_private(page) && trylock_page(page)) { | |
2085 | if (page_has_private(page)) | |
2086 | try_to_release_page(page, 0); | |
2087 | unlock_page(page); | |
2088 | } | |
2089 | } | |
2090 | ||
c3ac9a8a JW |
2091 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
2092 | &vm_flags)) { | |
9992af10 | 2093 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
2094 | /* |
2095 | * Identify referenced, file-backed active pages and | |
2096 | * give them one more trip around the active list. So | |
2097 | * that executable code get better chances to stay in | |
2098 | * memory under moderate memory pressure. Anon pages | |
2099 | * are not likely to be evicted by use-once streaming | |
2100 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
2101 | * so we ignore them here. | |
2102 | */ | |
41e20983 | 2103 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
2104 | list_add(&page->lru, &l_active); |
2105 | continue; | |
2106 | } | |
2107 | } | |
7e9cd484 | 2108 | |
5205e56e | 2109 | ClearPageActive(page); /* we are de-activating */ |
1da177e4 LT |
2110 | list_add(&page->lru, &l_inactive); |
2111 | } | |
2112 | ||
b555749a | 2113 | /* |
8cab4754 | 2114 | * Move pages back to the lru list. |
b555749a | 2115 | */ |
599d0c95 | 2116 | spin_lock_irq(&pgdat->lru_lock); |
556adecb | 2117 | /* |
8cab4754 WF |
2118 | * Count referenced pages from currently used mappings as rotated, |
2119 | * even though only some of them are actually re-activated. This | |
2120 | * helps balance scan pressure between file and anonymous pages in | |
7c0db9e9 | 2121 | * get_scan_count. |
7e9cd484 | 2122 | */ |
b7c46d15 | 2123 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 2124 | |
9d998b4f MH |
2125 | nr_activate = move_active_pages_to_lru(lruvec, &l_active, &l_hold, lru); |
2126 | nr_deactivate = move_active_pages_to_lru(lruvec, &l_inactive, &l_hold, lru - LRU_ACTIVE); | |
599d0c95 MG |
2127 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
2128 | spin_unlock_irq(&pgdat->lru_lock); | |
2bcf8879 | 2129 | |
747db954 | 2130 | mem_cgroup_uncharge_list(&l_hold); |
2d4894b5 | 2131 | free_unref_page_list(&l_hold); |
9d998b4f MH |
2132 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
2133 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
2134 | } |
2135 | ||
59dc76b0 RR |
2136 | /* |
2137 | * The inactive anon list should be small enough that the VM never has | |
2138 | * to do too much work. | |
14797e23 | 2139 | * |
59dc76b0 RR |
2140 | * The inactive file list should be small enough to leave most memory |
2141 | * to the established workingset on the scan-resistant active list, | |
2142 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2143 | * |
59dc76b0 RR |
2144 | * Both inactive lists should also be large enough that each inactive |
2145 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 2146 | * |
2a2e4885 JW |
2147 | * If that fails and refaulting is observed, the inactive list grows. |
2148 | * | |
59dc76b0 | 2149 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
3a50d14d | 2150 | * on this LRU, maintained by the pageout code. An inactive_ratio |
59dc76b0 | 2151 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. |
56e49d21 | 2152 | * |
59dc76b0 RR |
2153 | * total target max |
2154 | * memory ratio inactive | |
2155 | * ------------------------------------- | |
2156 | * 10MB 1 5MB | |
2157 | * 100MB 1 50MB | |
2158 | * 1GB 3 250MB | |
2159 | * 10GB 10 0.9GB | |
2160 | * 100GB 31 3GB | |
2161 | * 1TB 101 10GB | |
2162 | * 10TB 320 32GB | |
56e49d21 | 2163 | */ |
f8d1a311 | 2164 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file, |
2a2e4885 JW |
2165 | struct mem_cgroup *memcg, |
2166 | struct scan_control *sc, bool actual_reclaim) | |
56e49d21 | 2167 | { |
fd538803 | 2168 | enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; |
2a2e4885 JW |
2169 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
2170 | enum lru_list inactive_lru = file * LRU_FILE; | |
2171 | unsigned long inactive, active; | |
2172 | unsigned long inactive_ratio; | |
2173 | unsigned long refaults; | |
59dc76b0 | 2174 | unsigned long gb; |
e3790144 | 2175 | |
59dc76b0 RR |
2176 | /* |
2177 | * If we don't have swap space, anonymous page deactivation | |
2178 | * is pointless. | |
2179 | */ | |
2180 | if (!file && !total_swap_pages) | |
2181 | return false; | |
56e49d21 | 2182 | |
fd538803 MH |
2183 | inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); |
2184 | active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); | |
f8d1a311 | 2185 | |
2a2e4885 | 2186 | if (memcg) |
ccda7f43 | 2187 | refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); |
b39415b2 | 2188 | else |
2a2e4885 JW |
2189 | refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); |
2190 | ||
2191 | /* | |
2192 | * When refaults are being observed, it means a new workingset | |
2193 | * is being established. Disable active list protection to get | |
2194 | * rid of the stale workingset quickly. | |
2195 | */ | |
2196 | if (file && actual_reclaim && lruvec->refaults != refaults) { | |
2197 | inactive_ratio = 0; | |
2198 | } else { | |
2199 | gb = (inactive + active) >> (30 - PAGE_SHIFT); | |
2200 | if (gb) | |
2201 | inactive_ratio = int_sqrt(10 * gb); | |
2202 | else | |
2203 | inactive_ratio = 1; | |
2204 | } | |
59dc76b0 | 2205 | |
2a2e4885 JW |
2206 | if (actual_reclaim) |
2207 | trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx, | |
2208 | lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive, | |
2209 | lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active, | |
2210 | inactive_ratio, file); | |
fd538803 | 2211 | |
59dc76b0 | 2212 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2213 | } |
2214 | ||
4f98a2fe | 2215 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
2a2e4885 JW |
2216 | struct lruvec *lruvec, struct mem_cgroup *memcg, |
2217 | struct scan_control *sc) | |
b69408e8 | 2218 | { |
b39415b2 | 2219 | if (is_active_lru(lru)) { |
2a2e4885 JW |
2220 | if (inactive_list_is_low(lruvec, is_file_lru(lru), |
2221 | memcg, sc, true)) | |
1a93be0e | 2222 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
2223 | return 0; |
2224 | } | |
2225 | ||
1a93be0e | 2226 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
2227 | } |
2228 | ||
9a265114 JW |
2229 | enum scan_balance { |
2230 | SCAN_EQUAL, | |
2231 | SCAN_FRACT, | |
2232 | SCAN_ANON, | |
2233 | SCAN_FILE, | |
2234 | }; | |
2235 | ||
4f98a2fe RR |
2236 | /* |
2237 | * Determine how aggressively the anon and file LRU lists should be | |
2238 | * scanned. The relative value of each set of LRU lists is determined | |
2239 | * by looking at the fraction of the pages scanned we did rotate back | |
2240 | * onto the active list instead of evict. | |
2241 | * | |
be7bd59d WL |
2242 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2243 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 2244 | */ |
33377678 | 2245 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, |
6b4f7799 JW |
2246 | struct scan_control *sc, unsigned long *nr, |
2247 | unsigned long *lru_pages) | |
4f98a2fe | 2248 | { |
33377678 | 2249 | int swappiness = mem_cgroup_swappiness(memcg); |
9a265114 JW |
2250 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
2251 | u64 fraction[2]; | |
2252 | u64 denominator = 0; /* gcc */ | |
599d0c95 | 2253 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
4f98a2fe | 2254 | unsigned long anon_prio, file_prio; |
9a265114 | 2255 | enum scan_balance scan_balance; |
0bf1457f | 2256 | unsigned long anon, file; |
4f98a2fe | 2257 | unsigned long ap, fp; |
4111304d | 2258 | enum lru_list lru; |
76a33fc3 SL |
2259 | |
2260 | /* If we have no swap space, do not bother scanning anon pages. */ | |
d8b38438 | 2261 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
9a265114 | 2262 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2263 | goto out; |
2264 | } | |
4f98a2fe | 2265 | |
10316b31 JW |
2266 | /* |
2267 | * Global reclaim will swap to prevent OOM even with no | |
2268 | * swappiness, but memcg users want to use this knob to | |
2269 | * disable swapping for individual groups completely when | |
2270 | * using the memory controller's swap limit feature would be | |
2271 | * too expensive. | |
2272 | */ | |
02695175 | 2273 | if (!global_reclaim(sc) && !swappiness) { |
9a265114 | 2274 | scan_balance = SCAN_FILE; |
10316b31 JW |
2275 | goto out; |
2276 | } | |
2277 | ||
2278 | /* | |
2279 | * Do not apply any pressure balancing cleverness when the | |
2280 | * system is close to OOM, scan both anon and file equally | |
2281 | * (unless the swappiness setting disagrees with swapping). | |
2282 | */ | |
02695175 | 2283 | if (!sc->priority && swappiness) { |
9a265114 | 2284 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2285 | goto out; |
2286 | } | |
2287 | ||
62376251 JW |
2288 | /* |
2289 | * Prevent the reclaimer from falling into the cache trap: as | |
2290 | * cache pages start out inactive, every cache fault will tip | |
2291 | * the scan balance towards the file LRU. And as the file LRU | |
2292 | * shrinks, so does the window for rotation from references. | |
2293 | * This means we have a runaway feedback loop where a tiny | |
2294 | * thrashing file LRU becomes infinitely more attractive than | |
2295 | * anon pages. Try to detect this based on file LRU size. | |
2296 | */ | |
2297 | if (global_reclaim(sc)) { | |
599d0c95 MG |
2298 | unsigned long pgdatfile; |
2299 | unsigned long pgdatfree; | |
2300 | int z; | |
2301 | unsigned long total_high_wmark = 0; | |
2ab051e1 | 2302 | |
599d0c95 MG |
2303 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
2304 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2305 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2306 | ||
2307 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2308 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2309 | if (!managed_zone(zone)) |
599d0c95 MG |
2310 | continue; |
2311 | ||
2312 | total_high_wmark += high_wmark_pages(zone); | |
2313 | } | |
62376251 | 2314 | |
599d0c95 | 2315 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
06226226 DR |
2316 | /* |
2317 | * Force SCAN_ANON if there are enough inactive | |
2318 | * anonymous pages on the LRU in eligible zones. | |
2319 | * Otherwise, the small LRU gets thrashed. | |
2320 | */ | |
2321 | if (!inactive_list_is_low(lruvec, false, memcg, sc, false) && | |
2322 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx) | |
2323 | >> sc->priority) { | |
2324 | scan_balance = SCAN_ANON; | |
2325 | goto out; | |
2326 | } | |
62376251 JW |
2327 | } |
2328 | } | |
2329 | ||
7c5bd705 | 2330 | /* |
316bda0e VD |
2331 | * If there is enough inactive page cache, i.e. if the size of the |
2332 | * inactive list is greater than that of the active list *and* the | |
2333 | * inactive list actually has some pages to scan on this priority, we | |
2334 | * do not reclaim anything from the anonymous working set right now. | |
2335 | * Without the second condition we could end up never scanning an | |
2336 | * lruvec even if it has plenty of old anonymous pages unless the | |
2337 | * system is under heavy pressure. | |
7c5bd705 | 2338 | */ |
2a2e4885 | 2339 | if (!inactive_list_is_low(lruvec, true, memcg, sc, false) && |
71ab6cfe | 2340 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { |
9a265114 | 2341 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2342 | goto out; |
2343 | } | |
2344 | ||
9a265114 JW |
2345 | scan_balance = SCAN_FRACT; |
2346 | ||
58c37f6e KM |
2347 | /* |
2348 | * With swappiness at 100, anonymous and file have the same priority. | |
2349 | * This scanning priority is essentially the inverse of IO cost. | |
2350 | */ | |
02695175 | 2351 | anon_prio = swappiness; |
75b00af7 | 2352 | file_prio = 200 - anon_prio; |
58c37f6e | 2353 | |
4f98a2fe RR |
2354 | /* |
2355 | * OK, so we have swap space and a fair amount of page cache | |
2356 | * pages. We use the recently rotated / recently scanned | |
2357 | * ratios to determine how valuable each cache is. | |
2358 | * | |
2359 | * Because workloads change over time (and to avoid overflow) | |
2360 | * we keep these statistics as a floating average, which ends | |
2361 | * up weighing recent references more than old ones. | |
2362 | * | |
2363 | * anon in [0], file in [1] | |
2364 | */ | |
2ab051e1 | 2365 | |
fd538803 MH |
2366 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + |
2367 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); | |
2368 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + | |
2369 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); | |
2ab051e1 | 2370 | |
599d0c95 | 2371 | spin_lock_irq(&pgdat->lru_lock); |
6e901571 | 2372 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
2373 | reclaim_stat->recent_scanned[0] /= 2; |
2374 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
2375 | } |
2376 | ||
6e901571 | 2377 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
2378 | reclaim_stat->recent_scanned[1] /= 2; |
2379 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
2380 | } |
2381 | ||
4f98a2fe | 2382 | /* |
00d8089c RR |
2383 | * The amount of pressure on anon vs file pages is inversely |
2384 | * proportional to the fraction of recently scanned pages on | |
2385 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2386 | */ |
fe35004f | 2387 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 2388 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2389 | |
fe35004f | 2390 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2391 | fp /= reclaim_stat->recent_rotated[1] + 1; |
599d0c95 | 2392 | spin_unlock_irq(&pgdat->lru_lock); |
4f98a2fe | 2393 | |
76a33fc3 SL |
2394 | fraction[0] = ap; |
2395 | fraction[1] = fp; | |
2396 | denominator = ap + fp + 1; | |
2397 | out: | |
688035f7 JW |
2398 | *lru_pages = 0; |
2399 | for_each_evictable_lru(lru) { | |
2400 | int file = is_file_lru(lru); | |
2401 | unsigned long size; | |
2402 | unsigned long scan; | |
6b4f7799 | 2403 | |
688035f7 JW |
2404 | size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); |
2405 | scan = size >> sc->priority; | |
2406 | /* | |
2407 | * If the cgroup's already been deleted, make sure to | |
2408 | * scrape out the remaining cache. | |
2409 | */ | |
2410 | if (!scan && !mem_cgroup_online(memcg)) | |
2411 | scan = min(size, SWAP_CLUSTER_MAX); | |
6b4f7799 | 2412 | |
688035f7 JW |
2413 | switch (scan_balance) { |
2414 | case SCAN_EQUAL: | |
2415 | /* Scan lists relative to size */ | |
2416 | break; | |
2417 | case SCAN_FRACT: | |
9a265114 | 2418 | /* |
688035f7 JW |
2419 | * Scan types proportional to swappiness and |
2420 | * their relative recent reclaim efficiency. | |
9a265114 | 2421 | */ |
688035f7 JW |
2422 | scan = div64_u64(scan * fraction[file], |
2423 | denominator); | |
2424 | break; | |
2425 | case SCAN_FILE: | |
2426 | case SCAN_ANON: | |
2427 | /* Scan one type exclusively */ | |
2428 | if ((scan_balance == SCAN_FILE) != file) { | |
2429 | size = 0; | |
2430 | scan = 0; | |
2431 | } | |
2432 | break; | |
2433 | default: | |
2434 | /* Look ma, no brain */ | |
2435 | BUG(); | |
9a265114 | 2436 | } |
688035f7 JW |
2437 | |
2438 | *lru_pages += size; | |
2439 | nr[lru] = scan; | |
76a33fc3 | 2440 | } |
6e08a369 | 2441 | } |
4f98a2fe | 2442 | |
9b4f98cd | 2443 | /* |
a9dd0a83 | 2444 | * This is a basic per-node page freer. Used by both kswapd and direct reclaim. |
9b4f98cd | 2445 | */ |
a9dd0a83 | 2446 | static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, |
33377678 | 2447 | struct scan_control *sc, unsigned long *lru_pages) |
9b4f98cd | 2448 | { |
ef8f2327 | 2449 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
9b4f98cd | 2450 | unsigned long nr[NR_LRU_LISTS]; |
e82e0561 | 2451 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2452 | unsigned long nr_to_scan; |
2453 | enum lru_list lru; | |
2454 | unsigned long nr_reclaimed = 0; | |
2455 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2456 | struct blk_plug plug; | |
1a501907 | 2457 | bool scan_adjusted; |
9b4f98cd | 2458 | |
33377678 | 2459 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); |
9b4f98cd | 2460 | |
e82e0561 MG |
2461 | /* Record the original scan target for proportional adjustments later */ |
2462 | memcpy(targets, nr, sizeof(nr)); | |
2463 | ||
1a501907 MG |
2464 | /* |
2465 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2466 | * event that can occur when there is little memory pressure e.g. | |
2467 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2468 | * when the requested number of pages are reclaimed when scanning at | |
2469 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2470 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2471 | * do a batch of work at once. For memcg reclaim one check is made to | |
2472 | * abort proportional reclaim if either the file or anon lru has already | |
2473 | * dropped to zero at the first pass. | |
2474 | */ | |
2475 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2476 | sc->priority == DEF_PRIORITY); | |
2477 | ||
9b4f98cd JW |
2478 | blk_start_plug(&plug); |
2479 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2480 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2481 | unsigned long nr_anon, nr_file, percentage; |
2482 | unsigned long nr_scanned; | |
2483 | ||
9b4f98cd JW |
2484 | for_each_evictable_lru(lru) { |
2485 | if (nr[lru]) { | |
2486 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2487 | nr[lru] -= nr_to_scan; | |
2488 | ||
2489 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
2a2e4885 | 2490 | lruvec, memcg, sc); |
9b4f98cd JW |
2491 | } |
2492 | } | |
e82e0561 | 2493 | |
bd041733 MH |
2494 | cond_resched(); |
2495 | ||
e82e0561 MG |
2496 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
2497 | continue; | |
2498 | ||
e82e0561 MG |
2499 | /* |
2500 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2501 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2502 | * proportionally what was requested by get_scan_count(). We |
2503 | * stop reclaiming one LRU and reduce the amount scanning | |
2504 | * proportional to the original scan target. | |
2505 | */ | |
2506 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2507 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2508 | ||
1a501907 MG |
2509 | /* |
2510 | * It's just vindictive to attack the larger once the smaller | |
2511 | * has gone to zero. And given the way we stop scanning the | |
2512 | * smaller below, this makes sure that we only make one nudge | |
2513 | * towards proportionality once we've got nr_to_reclaim. | |
2514 | */ | |
2515 | if (!nr_file || !nr_anon) | |
2516 | break; | |
2517 | ||
e82e0561 MG |
2518 | if (nr_file > nr_anon) { |
2519 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2520 | targets[LRU_ACTIVE_ANON] + 1; | |
2521 | lru = LRU_BASE; | |
2522 | percentage = nr_anon * 100 / scan_target; | |
2523 | } else { | |
2524 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2525 | targets[LRU_ACTIVE_FILE] + 1; | |
2526 | lru = LRU_FILE; | |
2527 | percentage = nr_file * 100 / scan_target; | |
2528 | } | |
2529 | ||
2530 | /* Stop scanning the smaller of the LRU */ | |
2531 | nr[lru] = 0; | |
2532 | nr[lru + LRU_ACTIVE] = 0; | |
2533 | ||
2534 | /* | |
2535 | * Recalculate the other LRU scan count based on its original | |
2536 | * scan target and the percentage scanning already complete | |
2537 | */ | |
2538 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2539 | nr_scanned = targets[lru] - nr[lru]; | |
2540 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2541 | nr[lru] -= min(nr[lru], nr_scanned); | |
2542 | ||
2543 | lru += LRU_ACTIVE; | |
2544 | nr_scanned = targets[lru] - nr[lru]; | |
2545 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2546 | nr[lru] -= min(nr[lru], nr_scanned); | |
2547 | ||
2548 | scan_adjusted = true; | |
9b4f98cd JW |
2549 | } |
2550 | blk_finish_plug(&plug); | |
2551 | sc->nr_reclaimed += nr_reclaimed; | |
2552 | ||
2553 | /* | |
2554 | * Even if we did not try to evict anon pages at all, we want to | |
2555 | * rebalance the anon lru active/inactive ratio. | |
2556 | */ | |
2a2e4885 | 2557 | if (inactive_list_is_low(lruvec, false, memcg, sc, true)) |
9b4f98cd JW |
2558 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2559 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
2560 | } |
2561 | ||
23b9da55 | 2562 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2563 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2564 | { |
d84da3f9 | 2565 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2566 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2567 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2568 | return true; |
2569 | ||
2570 | return false; | |
2571 | } | |
2572 | ||
3e7d3449 | 2573 | /* |
23b9da55 MG |
2574 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2575 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2576 | * true if more pages should be reclaimed such that when the page allocator | |
2577 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2578 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2579 | */ |
a9dd0a83 | 2580 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 MG |
2581 | unsigned long nr_reclaimed, |
2582 | unsigned long nr_scanned, | |
2583 | struct scan_control *sc) | |
2584 | { | |
2585 | unsigned long pages_for_compaction; | |
2586 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 2587 | int z; |
3e7d3449 MG |
2588 | |
2589 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2590 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2591 | return false; |
2592 | ||
2876592f | 2593 | /* Consider stopping depending on scan and reclaim activity */ |
dcda9b04 | 2594 | if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) { |
2876592f | 2595 | /* |
dcda9b04 | 2596 | * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the |
2876592f MG |
2597 | * full LRU list has been scanned and we are still failing |
2598 | * to reclaim pages. This full LRU scan is potentially | |
dcda9b04 | 2599 | * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed |
2876592f MG |
2600 | */ |
2601 | if (!nr_reclaimed && !nr_scanned) | |
2602 | return false; | |
2603 | } else { | |
2604 | /* | |
dcda9b04 | 2605 | * For non-__GFP_RETRY_MAYFAIL allocations which can presumably |
2876592f MG |
2606 | * fail without consequence, stop if we failed to reclaim |
2607 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2608 | * pages that were scanned. This will return to the | |
2609 | * caller faster at the risk reclaim/compaction and | |
2610 | * the resulting allocation attempt fails | |
2611 | */ | |
2612 | if (!nr_reclaimed) | |
2613 | return false; | |
2614 | } | |
3e7d3449 MG |
2615 | |
2616 | /* | |
2617 | * If we have not reclaimed enough pages for compaction and the | |
2618 | * inactive lists are large enough, continue reclaiming | |
2619 | */ | |
9861a62c | 2620 | pages_for_compaction = compact_gap(sc->order); |
a9dd0a83 | 2621 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); |
ec8acf20 | 2622 | if (get_nr_swap_pages() > 0) |
a9dd0a83 | 2623 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
3e7d3449 MG |
2624 | if (sc->nr_reclaimed < pages_for_compaction && |
2625 | inactive_lru_pages > pages_for_compaction) | |
2626 | return true; | |
2627 | ||
2628 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
a9dd0a83 MG |
2629 | for (z = 0; z <= sc->reclaim_idx; z++) { |
2630 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2631 | if (!managed_zone(zone)) |
a9dd0a83 MG |
2632 | continue; |
2633 | ||
2634 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 2635 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
2636 | case COMPACT_CONTINUE: |
2637 | return false; | |
2638 | default: | |
2639 | /* check next zone */ | |
2640 | ; | |
2641 | } | |
3e7d3449 | 2642 | } |
a9dd0a83 | 2643 | return true; |
3e7d3449 MG |
2644 | } |
2645 | ||
e3c1ac58 AR |
2646 | static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg) |
2647 | { | |
2648 | return test_bit(PGDAT_CONGESTED, &pgdat->flags) || | |
2649 | (memcg && memcg_congested(pgdat, memcg)); | |
2650 | } | |
2651 | ||
970a39a3 | 2652 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 2653 | { |
cb731d6c | 2654 | struct reclaim_state *reclaim_state = current->reclaim_state; |
f0fdc5e8 | 2655 | unsigned long nr_reclaimed, nr_scanned; |
2344d7e4 | 2656 | bool reclaimable = false; |
1da177e4 | 2657 | |
9b4f98cd JW |
2658 | do { |
2659 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2660 | struct mem_cgroup_reclaim_cookie reclaim = { | |
ef8f2327 | 2661 | .pgdat = pgdat, |
9b4f98cd JW |
2662 | .priority = sc->priority, |
2663 | }; | |
a9dd0a83 | 2664 | unsigned long node_lru_pages = 0; |
694fbc0f | 2665 | struct mem_cgroup *memcg; |
3e7d3449 | 2666 | |
d108c772 AR |
2667 | memset(&sc->nr, 0, sizeof(sc->nr)); |
2668 | ||
9b4f98cd JW |
2669 | nr_reclaimed = sc->nr_reclaimed; |
2670 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2671 | |
694fbc0f AM |
2672 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2673 | do { | |
6b4f7799 | 2674 | unsigned long lru_pages; |
8e8ae645 | 2675 | unsigned long reclaimed; |
cb731d6c | 2676 | unsigned long scanned; |
5660048c | 2677 | |
bf8d5d52 RG |
2678 | switch (mem_cgroup_protected(root, memcg)) { |
2679 | case MEMCG_PROT_MIN: | |
2680 | /* | |
2681 | * Hard protection. | |
2682 | * If there is no reclaimable memory, OOM. | |
2683 | */ | |
2684 | continue; | |
2685 | case MEMCG_PROT_LOW: | |
2686 | /* | |
2687 | * Soft protection. | |
2688 | * Respect the protection only as long as | |
2689 | * there is an unprotected supply | |
2690 | * of reclaimable memory from other cgroups. | |
2691 | */ | |
d6622f63 YX |
2692 | if (!sc->memcg_low_reclaim) { |
2693 | sc->memcg_low_skipped = 1; | |
241994ed | 2694 | continue; |
d6622f63 | 2695 | } |
e27be240 | 2696 | memcg_memory_event(memcg, MEMCG_LOW); |
bf8d5d52 RG |
2697 | break; |
2698 | case MEMCG_PROT_NONE: | |
2699 | break; | |
241994ed JW |
2700 | } |
2701 | ||
8e8ae645 | 2702 | reclaimed = sc->nr_reclaimed; |
cb731d6c | 2703 | scanned = sc->nr_scanned; |
a9dd0a83 MG |
2704 | shrink_node_memcg(pgdat, memcg, sc, &lru_pages); |
2705 | node_lru_pages += lru_pages; | |
f16015fb | 2706 | |
aeed1d32 VD |
2707 | shrink_slab(sc->gfp_mask, pgdat->node_id, |
2708 | memcg, sc->priority); | |
cb731d6c | 2709 | |
8e8ae645 JW |
2710 | /* Record the group's reclaim efficiency */ |
2711 | vmpressure(sc->gfp_mask, memcg, false, | |
2712 | sc->nr_scanned - scanned, | |
2713 | sc->nr_reclaimed - reclaimed); | |
2714 | ||
9b4f98cd | 2715 | /* |
a394cb8e MH |
2716 | * Direct reclaim and kswapd have to scan all memory |
2717 | * cgroups to fulfill the overall scan target for the | |
a9dd0a83 | 2718 | * node. |
a394cb8e MH |
2719 | * |
2720 | * Limit reclaim, on the other hand, only cares about | |
2721 | * nr_to_reclaim pages to be reclaimed and it will | |
2722 | * retry with decreasing priority if one round over the | |
2723 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2724 | */ |
a394cb8e MH |
2725 | if (!global_reclaim(sc) && |
2726 | sc->nr_reclaimed >= sc->nr_to_reclaim) { | |
9b4f98cd JW |
2727 | mem_cgroup_iter_break(root, memcg); |
2728 | break; | |
2729 | } | |
241994ed | 2730 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
70ddf637 | 2731 | |
cb731d6c VD |
2732 | if (reclaim_state) { |
2733 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2734 | reclaim_state->reclaimed_slab = 0; | |
6b4f7799 JW |
2735 | } |
2736 | ||
8e8ae645 JW |
2737 | /* Record the subtree's reclaim efficiency */ |
2738 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
70ddf637 AV |
2739 | sc->nr_scanned - nr_scanned, |
2740 | sc->nr_reclaimed - nr_reclaimed); | |
2741 | ||
2344d7e4 JW |
2742 | if (sc->nr_reclaimed - nr_reclaimed) |
2743 | reclaimable = true; | |
2744 | ||
e3c1ac58 AR |
2745 | if (current_is_kswapd()) { |
2746 | /* | |
2747 | * If reclaim is isolating dirty pages under writeback, | |
2748 | * it implies that the long-lived page allocation rate | |
2749 | * is exceeding the page laundering rate. Either the | |
2750 | * global limits are not being effective at throttling | |
2751 | * processes due to the page distribution throughout | |
2752 | * zones or there is heavy usage of a slow backing | |
2753 | * device. The only option is to throttle from reclaim | |
2754 | * context which is not ideal as there is no guarantee | |
2755 | * the dirtying process is throttled in the same way | |
2756 | * balance_dirty_pages() manages. | |
2757 | * | |
2758 | * Once a node is flagged PGDAT_WRITEBACK, kswapd will | |
2759 | * count the number of pages under pages flagged for | |
2760 | * immediate reclaim and stall if any are encountered | |
2761 | * in the nr_immediate check below. | |
2762 | */ | |
2763 | if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken) | |
2764 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
d108c772 | 2765 | |
d108c772 AR |
2766 | /* |
2767 | * Tag a node as congested if all the dirty pages | |
2768 | * scanned were backed by a congested BDI and | |
2769 | * wait_iff_congested will stall. | |
2770 | */ | |
2771 | if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) | |
2772 | set_bit(PGDAT_CONGESTED, &pgdat->flags); | |
2773 | ||
2774 | /* Allow kswapd to start writing pages during reclaim.*/ | |
2775 | if (sc->nr.unqueued_dirty == sc->nr.file_taken) | |
2776 | set_bit(PGDAT_DIRTY, &pgdat->flags); | |
2777 | ||
2778 | /* | |
2779 | * If kswapd scans pages marked marked for immediate | |
2780 | * reclaim and under writeback (nr_immediate), it | |
2781 | * implies that pages are cycling through the LRU | |
2782 | * faster than they are written so also forcibly stall. | |
2783 | */ | |
2784 | if (sc->nr.immediate) | |
2785 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
2786 | } | |
2787 | ||
e3c1ac58 AR |
2788 | /* |
2789 | * Legacy memcg will stall in page writeback so avoid forcibly | |
2790 | * stalling in wait_iff_congested(). | |
2791 | */ | |
2792 | if (!global_reclaim(sc) && sane_reclaim(sc) && | |
2793 | sc->nr.dirty && sc->nr.dirty == sc->nr.congested) | |
2794 | set_memcg_congestion(pgdat, root, true); | |
2795 | ||
d108c772 AR |
2796 | /* |
2797 | * Stall direct reclaim for IO completions if underlying BDIs | |
2798 | * and node is congested. Allow kswapd to continue until it | |
2799 | * starts encountering unqueued dirty pages or cycling through | |
2800 | * the LRU too quickly. | |
2801 | */ | |
2802 | if (!sc->hibernation_mode && !current_is_kswapd() && | |
e3c1ac58 AR |
2803 | current_may_throttle() && pgdat_memcg_congested(pgdat, root)) |
2804 | wait_iff_congested(BLK_RW_ASYNC, HZ/10); | |
d108c772 | 2805 | |
a9dd0a83 | 2806 | } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
9b4f98cd | 2807 | sc->nr_scanned - nr_scanned, sc)); |
2344d7e4 | 2808 | |
c73322d0 JW |
2809 | /* |
2810 | * Kswapd gives up on balancing particular nodes after too | |
2811 | * many failures to reclaim anything from them and goes to | |
2812 | * sleep. On reclaim progress, reset the failure counter. A | |
2813 | * successful direct reclaim run will revive a dormant kswapd. | |
2814 | */ | |
2815 | if (reclaimable) | |
2816 | pgdat->kswapd_failures = 0; | |
2817 | ||
2344d7e4 | 2818 | return reclaimable; |
f16015fb JW |
2819 | } |
2820 | ||
53853e2d | 2821 | /* |
fdd4c614 VB |
2822 | * Returns true if compaction should go ahead for a costly-order request, or |
2823 | * the allocation would already succeed without compaction. Return false if we | |
2824 | * should reclaim first. | |
53853e2d | 2825 | */ |
4f588331 | 2826 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 2827 | { |
31483b6a | 2828 | unsigned long watermark; |
fdd4c614 | 2829 | enum compact_result suitable; |
fe4b1b24 | 2830 | |
fdd4c614 VB |
2831 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
2832 | if (suitable == COMPACT_SUCCESS) | |
2833 | /* Allocation should succeed already. Don't reclaim. */ | |
2834 | return true; | |
2835 | if (suitable == COMPACT_SKIPPED) | |
2836 | /* Compaction cannot yet proceed. Do reclaim. */ | |
2837 | return false; | |
fe4b1b24 | 2838 | |
53853e2d | 2839 | /* |
fdd4c614 VB |
2840 | * Compaction is already possible, but it takes time to run and there |
2841 | * are potentially other callers using the pages just freed. So proceed | |
2842 | * with reclaim to make a buffer of free pages available to give | |
2843 | * compaction a reasonable chance of completing and allocating the page. | |
2844 | * Note that we won't actually reclaim the whole buffer in one attempt | |
2845 | * as the target watermark in should_continue_reclaim() is lower. But if | |
2846 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 2847 | */ |
fdd4c614 | 2848 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 2849 | |
fdd4c614 | 2850 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
2851 | } |
2852 | ||
1da177e4 LT |
2853 | /* |
2854 | * This is the direct reclaim path, for page-allocating processes. We only | |
2855 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2856 | * request. | |
2857 | * | |
1da177e4 LT |
2858 | * If a zone is deemed to be full of pinned pages then just give it a light |
2859 | * scan then give up on it. | |
2860 | */ | |
0a0337e0 | 2861 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2862 | { |
dd1a239f | 2863 | struct zoneref *z; |
54a6eb5c | 2864 | struct zone *zone; |
0608f43d AM |
2865 | unsigned long nr_soft_reclaimed; |
2866 | unsigned long nr_soft_scanned; | |
619d0d76 | 2867 | gfp_t orig_mask; |
79dafcdc | 2868 | pg_data_t *last_pgdat = NULL; |
1cfb419b | 2869 | |
cc715d99 MG |
2870 | /* |
2871 | * If the number of buffer_heads in the machine exceeds the maximum | |
2872 | * allowed level, force direct reclaim to scan the highmem zone as | |
2873 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2874 | */ | |
619d0d76 | 2875 | orig_mask = sc->gfp_mask; |
b2e18757 | 2876 | if (buffer_heads_over_limit) { |
cc715d99 | 2877 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 2878 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 2879 | } |
cc715d99 | 2880 | |
d4debc66 | 2881 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 2882 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
2883 | /* |
2884 | * Take care memory controller reclaiming has small influence | |
2885 | * to global LRU. | |
2886 | */ | |
89b5fae5 | 2887 | if (global_reclaim(sc)) { |
344736f2 VD |
2888 | if (!cpuset_zone_allowed(zone, |
2889 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2890 | continue; |
65ec02cb | 2891 | |
0b06496a JW |
2892 | /* |
2893 | * If we already have plenty of memory free for | |
2894 | * compaction in this zone, don't free any more. | |
2895 | * Even though compaction is invoked for any | |
2896 | * non-zero order, only frequent costly order | |
2897 | * reclamation is disruptive enough to become a | |
2898 | * noticeable problem, like transparent huge | |
2899 | * page allocations. | |
2900 | */ | |
2901 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2902 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 2903 | compaction_ready(zone, sc)) { |
0b06496a JW |
2904 | sc->compaction_ready = true; |
2905 | continue; | |
e0887c19 | 2906 | } |
0b06496a | 2907 | |
79dafcdc MG |
2908 | /* |
2909 | * Shrink each node in the zonelist once. If the | |
2910 | * zonelist is ordered by zone (not the default) then a | |
2911 | * node may be shrunk multiple times but in that case | |
2912 | * the user prefers lower zones being preserved. | |
2913 | */ | |
2914 | if (zone->zone_pgdat == last_pgdat) | |
2915 | continue; | |
2916 | ||
0608f43d AM |
2917 | /* |
2918 | * This steals pages from memory cgroups over softlimit | |
2919 | * and returns the number of reclaimed pages and | |
2920 | * scanned pages. This works for global memory pressure | |
2921 | * and balancing, not for a memcg's limit. | |
2922 | */ | |
2923 | nr_soft_scanned = 0; | |
ef8f2327 | 2924 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
2925 | sc->order, sc->gfp_mask, |
2926 | &nr_soft_scanned); | |
2927 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2928 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2929 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2930 | } |
408d8544 | 2931 | |
79dafcdc MG |
2932 | /* See comment about same check for global reclaim above */ |
2933 | if (zone->zone_pgdat == last_pgdat) | |
2934 | continue; | |
2935 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 2936 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 2937 | } |
e0c23279 | 2938 | |
619d0d76 WY |
2939 | /* |
2940 | * Restore to original mask to avoid the impact on the caller if we | |
2941 | * promoted it to __GFP_HIGHMEM. | |
2942 | */ | |
2943 | sc->gfp_mask = orig_mask; | |
1da177e4 | 2944 | } |
4f98a2fe | 2945 | |
2a2e4885 JW |
2946 | static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat) |
2947 | { | |
2948 | struct mem_cgroup *memcg; | |
2949 | ||
2950 | memcg = mem_cgroup_iter(root_memcg, NULL, NULL); | |
2951 | do { | |
2952 | unsigned long refaults; | |
2953 | struct lruvec *lruvec; | |
2954 | ||
2955 | if (memcg) | |
ccda7f43 | 2956 | refaults = memcg_page_state(memcg, WORKINGSET_ACTIVATE); |
2a2e4885 JW |
2957 | else |
2958 | refaults = node_page_state(pgdat, WORKINGSET_ACTIVATE); | |
2959 | ||
2960 | lruvec = mem_cgroup_lruvec(pgdat, memcg); | |
2961 | lruvec->refaults = refaults; | |
2962 | } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL))); | |
2963 | } | |
2964 | ||
1da177e4 LT |
2965 | /* |
2966 | * This is the main entry point to direct page reclaim. | |
2967 | * | |
2968 | * If a full scan of the inactive list fails to free enough memory then we | |
2969 | * are "out of memory" and something needs to be killed. | |
2970 | * | |
2971 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2972 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2973 | * caller can't do much about. We kick the writeback threads and take explicit |
2974 | * naps in the hope that some of these pages can be written. But if the | |
2975 | * allocating task holds filesystem locks which prevent writeout this might not | |
2976 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2977 | * |
2978 | * returns: 0, if no pages reclaimed | |
2979 | * else, the number of pages reclaimed | |
1da177e4 | 2980 | */ |
dac1d27b | 2981 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 2982 | struct scan_control *sc) |
1da177e4 | 2983 | { |
241994ed | 2984 | int initial_priority = sc->priority; |
2a2e4885 JW |
2985 | pg_data_t *last_pgdat; |
2986 | struct zoneref *z; | |
2987 | struct zone *zone; | |
241994ed | 2988 | retry: |
873b4771 KK |
2989 | delayacct_freepages_start(); |
2990 | ||
89b5fae5 | 2991 | if (global_reclaim(sc)) |
7cc30fcf | 2992 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 2993 | |
9e3b2f8c | 2994 | do { |
70ddf637 AV |
2995 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
2996 | sc->priority); | |
66e1707b | 2997 | sc->nr_scanned = 0; |
0a0337e0 | 2998 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 2999 | |
bb21c7ce | 3000 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
3001 | break; |
3002 | ||
3003 | if (sc->compaction_ready) | |
3004 | break; | |
1da177e4 | 3005 | |
0e50ce3b MK |
3006 | /* |
3007 | * If we're getting trouble reclaiming, start doing | |
3008 | * writepage even in laptop mode. | |
3009 | */ | |
3010 | if (sc->priority < DEF_PRIORITY - 2) | |
3011 | sc->may_writepage = 1; | |
0b06496a | 3012 | } while (--sc->priority >= 0); |
bb21c7ce | 3013 | |
2a2e4885 JW |
3014 | last_pgdat = NULL; |
3015 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
3016 | sc->nodemask) { | |
3017 | if (zone->zone_pgdat == last_pgdat) | |
3018 | continue; | |
3019 | last_pgdat = zone->zone_pgdat; | |
3020 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); | |
e3c1ac58 | 3021 | set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false); |
2a2e4885 JW |
3022 | } |
3023 | ||
873b4771 KK |
3024 | delayacct_freepages_end(); |
3025 | ||
bb21c7ce KM |
3026 | if (sc->nr_reclaimed) |
3027 | return sc->nr_reclaimed; | |
3028 | ||
0cee34fd | 3029 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 3030 | if (sc->compaction_ready) |
7335084d MG |
3031 | return 1; |
3032 | ||
241994ed | 3033 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
d6622f63 | 3034 | if (sc->memcg_low_skipped) { |
241994ed | 3035 | sc->priority = initial_priority; |
d6622f63 YX |
3036 | sc->memcg_low_reclaim = 1; |
3037 | sc->memcg_low_skipped = 0; | |
241994ed JW |
3038 | goto retry; |
3039 | } | |
3040 | ||
bb21c7ce | 3041 | return 0; |
1da177e4 LT |
3042 | } |
3043 | ||
c73322d0 | 3044 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
5515061d MG |
3045 | { |
3046 | struct zone *zone; | |
3047 | unsigned long pfmemalloc_reserve = 0; | |
3048 | unsigned long free_pages = 0; | |
3049 | int i; | |
3050 | bool wmark_ok; | |
3051 | ||
c73322d0 JW |
3052 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
3053 | return true; | |
3054 | ||
5515061d MG |
3055 | for (i = 0; i <= ZONE_NORMAL; i++) { |
3056 | zone = &pgdat->node_zones[i]; | |
d450abd8 JW |
3057 | if (!managed_zone(zone)) |
3058 | continue; | |
3059 | ||
3060 | if (!zone_reclaimable_pages(zone)) | |
675becce MG |
3061 | continue; |
3062 | ||
5515061d MG |
3063 | pfmemalloc_reserve += min_wmark_pages(zone); |
3064 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
3065 | } | |
3066 | ||
675becce MG |
3067 | /* If there are no reserves (unexpected config) then do not throttle */ |
3068 | if (!pfmemalloc_reserve) | |
3069 | return true; | |
3070 | ||
5515061d MG |
3071 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
3072 | ||
3073 | /* kswapd must be awake if processes are being throttled */ | |
3074 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
38087d9b | 3075 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
5515061d MG |
3076 | (enum zone_type)ZONE_NORMAL); |
3077 | wake_up_interruptible(&pgdat->kswapd_wait); | |
3078 | } | |
3079 | ||
3080 | return wmark_ok; | |
3081 | } | |
3082 | ||
3083 | /* | |
3084 | * Throttle direct reclaimers if backing storage is backed by the network | |
3085 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
3086 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
3087 | * when the low watermark is reached. |
3088 | * | |
3089 | * Returns true if a fatal signal was delivered during throttling. If this | |
3090 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 3091 | */ |
50694c28 | 3092 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
3093 | nodemask_t *nodemask) |
3094 | { | |
675becce | 3095 | struct zoneref *z; |
5515061d | 3096 | struct zone *zone; |
675becce | 3097 | pg_data_t *pgdat = NULL; |
5515061d MG |
3098 | |
3099 | /* | |
3100 | * Kernel threads should not be throttled as they may be indirectly | |
3101 | * responsible for cleaning pages necessary for reclaim to make forward | |
3102 | * progress. kjournald for example may enter direct reclaim while | |
3103 | * committing a transaction where throttling it could forcing other | |
3104 | * processes to block on log_wait_commit(). | |
3105 | */ | |
3106 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
3107 | goto out; |
3108 | ||
3109 | /* | |
3110 | * If a fatal signal is pending, this process should not throttle. | |
3111 | * It should return quickly so it can exit and free its memory | |
3112 | */ | |
3113 | if (fatal_signal_pending(current)) | |
3114 | goto out; | |
5515061d | 3115 | |
675becce MG |
3116 | /* |
3117 | * Check if the pfmemalloc reserves are ok by finding the first node | |
3118 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
3119 | * GFP_KERNEL will be required for allocating network buffers when | |
3120 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
3121 | * | |
3122 | * Throttling is based on the first usable node and throttled processes | |
3123 | * wait on a queue until kswapd makes progress and wakes them. There | |
3124 | * is an affinity then between processes waking up and where reclaim | |
3125 | * progress has been made assuming the process wakes on the same node. | |
3126 | * More importantly, processes running on remote nodes will not compete | |
3127 | * for remote pfmemalloc reserves and processes on different nodes | |
3128 | * should make reasonable progress. | |
3129 | */ | |
3130 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 3131 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
3132 | if (zone_idx(zone) > ZONE_NORMAL) |
3133 | continue; | |
3134 | ||
3135 | /* Throttle based on the first usable node */ | |
3136 | pgdat = zone->zone_pgdat; | |
c73322d0 | 3137 | if (allow_direct_reclaim(pgdat)) |
675becce MG |
3138 | goto out; |
3139 | break; | |
3140 | } | |
3141 | ||
3142 | /* If no zone was usable by the allocation flags then do not throttle */ | |
3143 | if (!pgdat) | |
50694c28 | 3144 | goto out; |
5515061d | 3145 | |
68243e76 MG |
3146 | /* Account for the throttling */ |
3147 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
3148 | ||
5515061d MG |
3149 | /* |
3150 | * If the caller cannot enter the filesystem, it's possible that it | |
3151 | * is due to the caller holding an FS lock or performing a journal | |
3152 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
3153 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
3154 | * blocked waiting on the same lock. Instead, throttle for up to a | |
3155 | * second before continuing. | |
3156 | */ | |
3157 | if (!(gfp_mask & __GFP_FS)) { | |
3158 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
c73322d0 | 3159 | allow_direct_reclaim(pgdat), HZ); |
50694c28 MG |
3160 | |
3161 | goto check_pending; | |
5515061d MG |
3162 | } |
3163 | ||
3164 | /* Throttle until kswapd wakes the process */ | |
3165 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
c73322d0 | 3166 | allow_direct_reclaim(pgdat)); |
50694c28 MG |
3167 | |
3168 | check_pending: | |
3169 | if (fatal_signal_pending(current)) | |
3170 | return true; | |
3171 | ||
3172 | out: | |
3173 | return false; | |
5515061d MG |
3174 | } |
3175 | ||
dac1d27b | 3176 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 3177 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 3178 | { |
33906bc5 | 3179 | unsigned long nr_reclaimed; |
66e1707b | 3180 | struct scan_control sc = { |
ee814fe2 | 3181 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
f2f43e56 | 3182 | .gfp_mask = current_gfp_context(gfp_mask), |
b2e18757 | 3183 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
3184 | .order = order, |
3185 | .nodemask = nodemask, | |
3186 | .priority = DEF_PRIORITY, | |
66e1707b | 3187 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3188 | .may_unmap = 1, |
2e2e4259 | 3189 | .may_swap = 1, |
66e1707b BS |
3190 | }; |
3191 | ||
bb451fdf GT |
3192 | /* |
3193 | * scan_control uses s8 fields for order, priority, and reclaim_idx. | |
3194 | * Confirm they are large enough for max values. | |
3195 | */ | |
3196 | BUILD_BUG_ON(MAX_ORDER > S8_MAX); | |
3197 | BUILD_BUG_ON(DEF_PRIORITY > S8_MAX); | |
3198 | BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX); | |
3199 | ||
5515061d | 3200 | /* |
50694c28 MG |
3201 | * Do not enter reclaim if fatal signal was delivered while throttled. |
3202 | * 1 is returned so that the page allocator does not OOM kill at this | |
3203 | * point. | |
5515061d | 3204 | */ |
f2f43e56 | 3205 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
5515061d MG |
3206 | return 1; |
3207 | ||
33906bc5 MG |
3208 | trace_mm_vmscan_direct_reclaim_begin(order, |
3209 | sc.may_writepage, | |
f2f43e56 | 3210 | sc.gfp_mask, |
e5146b12 | 3211 | sc.reclaim_idx); |
33906bc5 | 3212 | |
3115cd91 | 3213 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
3214 | |
3215 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
3216 | ||
3217 | return nr_reclaimed; | |
66e1707b BS |
3218 | } |
3219 | ||
c255a458 | 3220 | #ifdef CONFIG_MEMCG |
66e1707b | 3221 | |
a9dd0a83 | 3222 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 3223 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 3224 | pg_data_t *pgdat, |
0ae5e89c | 3225 | unsigned long *nr_scanned) |
4e416953 BS |
3226 | { |
3227 | struct scan_control sc = { | |
b8f5c566 | 3228 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 3229 | .target_mem_cgroup = memcg, |
4e416953 BS |
3230 | .may_writepage = !laptop_mode, |
3231 | .may_unmap = 1, | |
b2e18757 | 3232 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 3233 | .may_swap = !noswap, |
4e416953 | 3234 | }; |
6b4f7799 | 3235 | unsigned long lru_pages; |
0ae5e89c | 3236 | |
4e416953 BS |
3237 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3238 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 3239 | |
9e3b2f8c | 3240 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
bdce6d9e | 3241 | sc.may_writepage, |
e5146b12 MG |
3242 | sc.gfp_mask, |
3243 | sc.reclaim_idx); | |
bdce6d9e | 3244 | |
4e416953 BS |
3245 | /* |
3246 | * NOTE: Although we can get the priority field, using it | |
3247 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 3248 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
3249 | * will pick up pages from other mem cgroup's as well. We hack |
3250 | * the priority and make it zero. | |
3251 | */ | |
ef8f2327 | 3252 | shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); |
bdce6d9e KM |
3253 | |
3254 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
3255 | ||
0ae5e89c | 3256 | *nr_scanned = sc.nr_scanned; |
4e416953 BS |
3257 | return sc.nr_reclaimed; |
3258 | } | |
3259 | ||
72835c86 | 3260 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 3261 | unsigned long nr_pages, |
a7885eb8 | 3262 | gfp_t gfp_mask, |
b70a2a21 | 3263 | bool may_swap) |
66e1707b | 3264 | { |
4e416953 | 3265 | struct zonelist *zonelist; |
bdce6d9e | 3266 | unsigned long nr_reclaimed; |
889976db | 3267 | int nid; |
499118e9 | 3268 | unsigned int noreclaim_flag; |
66e1707b | 3269 | struct scan_control sc = { |
b70a2a21 | 3270 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 3271 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
a09ed5e0 | 3272 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
b2e18757 | 3273 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
3274 | .target_mem_cgroup = memcg, |
3275 | .priority = DEF_PRIORITY, | |
3276 | .may_writepage = !laptop_mode, | |
3277 | .may_unmap = 1, | |
b70a2a21 | 3278 | .may_swap = may_swap, |
a09ed5e0 | 3279 | }; |
66e1707b | 3280 | |
889976db YH |
3281 | /* |
3282 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
3283 | * take care of from where we get pages. So the node where we start the | |
3284 | * scan does not need to be the current node. | |
3285 | */ | |
72835c86 | 3286 | nid = mem_cgroup_select_victim_node(memcg); |
889976db | 3287 | |
c9634cf0 | 3288 | zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; |
bdce6d9e KM |
3289 | |
3290 | trace_mm_vmscan_memcg_reclaim_begin(0, | |
3291 | sc.may_writepage, | |
e5146b12 MG |
3292 | sc.gfp_mask, |
3293 | sc.reclaim_idx); | |
bdce6d9e | 3294 | |
499118e9 | 3295 | noreclaim_flag = memalloc_noreclaim_save(); |
3115cd91 | 3296 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
499118e9 | 3297 | memalloc_noreclaim_restore(noreclaim_flag); |
bdce6d9e KM |
3298 | |
3299 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
3300 | ||
3301 | return nr_reclaimed; | |
66e1707b BS |
3302 | } |
3303 | #endif | |
3304 | ||
1d82de61 | 3305 | static void age_active_anon(struct pglist_data *pgdat, |
ef8f2327 | 3306 | struct scan_control *sc) |
f16015fb | 3307 | { |
b95a2f2d | 3308 | struct mem_cgroup *memcg; |
f16015fb | 3309 | |
b95a2f2d JW |
3310 | if (!total_swap_pages) |
3311 | return; | |
3312 | ||
3313 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
3314 | do { | |
ef8f2327 | 3315 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
b95a2f2d | 3316 | |
2a2e4885 | 3317 | if (inactive_list_is_low(lruvec, false, memcg, sc, true)) |
1a93be0e | 3318 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 3319 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
3320 | |
3321 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
3322 | } while (memcg); | |
f16015fb JW |
3323 | } |
3324 | ||
e716f2eb MG |
3325 | /* |
3326 | * Returns true if there is an eligible zone balanced for the request order | |
3327 | * and classzone_idx | |
3328 | */ | |
3329 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) | |
60cefed4 | 3330 | { |
e716f2eb MG |
3331 | int i; |
3332 | unsigned long mark = -1; | |
3333 | struct zone *zone; | |
60cefed4 | 3334 | |
e716f2eb MG |
3335 | for (i = 0; i <= classzone_idx; i++) { |
3336 | zone = pgdat->node_zones + i; | |
6256c6b4 | 3337 | |
e716f2eb MG |
3338 | if (!managed_zone(zone)) |
3339 | continue; | |
3340 | ||
3341 | mark = high_wmark_pages(zone); | |
3342 | if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) | |
3343 | return true; | |
3344 | } | |
3345 | ||
3346 | /* | |
3347 | * If a node has no populated zone within classzone_idx, it does not | |
3348 | * need balancing by definition. This can happen if a zone-restricted | |
3349 | * allocation tries to wake a remote kswapd. | |
3350 | */ | |
3351 | if (mark == -1) | |
3352 | return true; | |
3353 | ||
3354 | return false; | |
60cefed4 JW |
3355 | } |
3356 | ||
631b6e08 MG |
3357 | /* Clear pgdat state for congested, dirty or under writeback. */ |
3358 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
3359 | { | |
3360 | clear_bit(PGDAT_CONGESTED, &pgdat->flags); | |
3361 | clear_bit(PGDAT_DIRTY, &pgdat->flags); | |
3362 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
3363 | } | |
3364 | ||
5515061d MG |
3365 | /* |
3366 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3367 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3368 | * | |
3369 | * Returns true if kswapd is ready to sleep | |
3370 | */ | |
d9f21d42 | 3371 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f50de2d3 | 3372 | { |
5515061d | 3373 | /* |
9e5e3661 | 3374 | * The throttled processes are normally woken up in balance_pgdat() as |
c73322d0 | 3375 | * soon as allow_direct_reclaim() is true. But there is a potential |
9e5e3661 VB |
3376 | * race between when kswapd checks the watermarks and a process gets |
3377 | * throttled. There is also a potential race if processes get | |
3378 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3379 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3380 | * the wake up checks. If kswapd is going to sleep, no process should | |
3381 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3382 | * the wake up is premature, processes will wake kswapd and get | |
3383 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3384 | * that here we are under prepare_to_wait(). | |
5515061d | 3385 | */ |
9e5e3661 VB |
3386 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3387 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3388 | |
c73322d0 JW |
3389 | /* Hopeless node, leave it to direct reclaim */ |
3390 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3391 | return true; | |
3392 | ||
e716f2eb MG |
3393 | if (pgdat_balanced(pgdat, order, classzone_idx)) { |
3394 | clear_pgdat_congested(pgdat); | |
3395 | return true; | |
1d82de61 MG |
3396 | } |
3397 | ||
333b0a45 | 3398 | return false; |
f50de2d3 MG |
3399 | } |
3400 | ||
75485363 | 3401 | /* |
1d82de61 MG |
3402 | * kswapd shrinks a node of pages that are at or below the highest usable |
3403 | * zone that is currently unbalanced. | |
b8e83b94 MG |
3404 | * |
3405 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3406 | * reclaim or if the lack of progress was due to pages under writeback. |
3407 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3408 | */ |
1d82de61 | 3409 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 3410 | struct scan_control *sc) |
75485363 | 3411 | { |
1d82de61 MG |
3412 | struct zone *zone; |
3413 | int z; | |
75485363 | 3414 | |
1d82de61 MG |
3415 | /* Reclaim a number of pages proportional to the number of zones */ |
3416 | sc->nr_to_reclaim = 0; | |
970a39a3 | 3417 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 3418 | zone = pgdat->node_zones + z; |
6aa303de | 3419 | if (!managed_zone(zone)) |
1d82de61 | 3420 | continue; |
7c954f6d | 3421 | |
1d82de61 MG |
3422 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
3423 | } | |
7c954f6d MG |
3424 | |
3425 | /* | |
1d82de61 MG |
3426 | * Historically care was taken to put equal pressure on all zones but |
3427 | * now pressure is applied based on node LRU order. | |
7c954f6d | 3428 | */ |
970a39a3 | 3429 | shrink_node(pgdat, sc); |
283aba9f | 3430 | |
7c954f6d | 3431 | /* |
1d82de61 MG |
3432 | * Fragmentation may mean that the system cannot be rebalanced for |
3433 | * high-order allocations. If twice the allocation size has been | |
3434 | * reclaimed then recheck watermarks only at order-0 to prevent | |
3435 | * excessive reclaim. Assume that a process requested a high-order | |
3436 | * can direct reclaim/compact. | |
7c954f6d | 3437 | */ |
9861a62c | 3438 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 3439 | sc->order = 0; |
7c954f6d | 3440 | |
b8e83b94 | 3441 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3442 | } |
3443 | ||
1da177e4 | 3444 | /* |
1d82de61 MG |
3445 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
3446 | * that are eligible for use by the caller until at least one zone is | |
3447 | * balanced. | |
1da177e4 | 3448 | * |
1d82de61 | 3449 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
3450 | * |
3451 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 3452 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
1d82de61 MG |
3453 | * found to have free_pages <= high_wmark_pages(zone), any page is that zone |
3454 | * or lower is eligible for reclaim until at least one usable zone is | |
3455 | * balanced. | |
1da177e4 | 3456 | */ |
accf6242 | 3457 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
1da177e4 | 3458 | { |
1da177e4 | 3459 | int i; |
0608f43d AM |
3460 | unsigned long nr_soft_reclaimed; |
3461 | unsigned long nr_soft_scanned; | |
1d82de61 | 3462 | struct zone *zone; |
179e9639 AM |
3463 | struct scan_control sc = { |
3464 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3465 | .order = order, |
b8e83b94 | 3466 | .priority = DEF_PRIORITY, |
ee814fe2 | 3467 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3468 | .may_unmap = 1, |
2e2e4259 | 3469 | .may_swap = 1, |
179e9639 | 3470 | }; |
93781325 OS |
3471 | |
3472 | __fs_reclaim_acquire(); | |
3473 | ||
f8891e5e | 3474 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3475 | |
9e3b2f8c | 3476 | do { |
c73322d0 | 3477 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
b8e83b94 | 3478 | bool raise_priority = true; |
93781325 | 3479 | bool ret; |
b8e83b94 | 3480 | |
84c7a777 | 3481 | sc.reclaim_idx = classzone_idx; |
1da177e4 | 3482 | |
86c79f6b | 3483 | /* |
84c7a777 MG |
3484 | * If the number of buffer_heads exceeds the maximum allowed |
3485 | * then consider reclaiming from all zones. This has a dual | |
3486 | * purpose -- on 64-bit systems it is expected that | |
3487 | * buffer_heads are stripped during active rotation. On 32-bit | |
3488 | * systems, highmem pages can pin lowmem memory and shrinking | |
3489 | * buffers can relieve lowmem pressure. Reclaim may still not | |
3490 | * go ahead if all eligible zones for the original allocation | |
3491 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
3492 | */ |
3493 | if (buffer_heads_over_limit) { | |
3494 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
3495 | zone = pgdat->node_zones + i; | |
6aa303de | 3496 | if (!managed_zone(zone)) |
86c79f6b | 3497 | continue; |
cc715d99 | 3498 | |
970a39a3 | 3499 | sc.reclaim_idx = i; |
e1dbeda6 | 3500 | break; |
1da177e4 | 3501 | } |
1da177e4 | 3502 | } |
dafcb73e | 3503 | |
86c79f6b | 3504 | /* |
e716f2eb MG |
3505 | * Only reclaim if there are no eligible zones. Note that |
3506 | * sc.reclaim_idx is not used as buffer_heads_over_limit may | |
3507 | * have adjusted it. | |
86c79f6b | 3508 | */ |
e716f2eb MG |
3509 | if (pgdat_balanced(pgdat, sc.order, classzone_idx)) |
3510 | goto out; | |
e1dbeda6 | 3511 | |
1d82de61 MG |
3512 | /* |
3513 | * Do some background aging of the anon list, to give | |
3514 | * pages a chance to be referenced before reclaiming. All | |
3515 | * pages are rotated regardless of classzone as this is | |
3516 | * about consistent aging. | |
3517 | */ | |
ef8f2327 | 3518 | age_active_anon(pgdat, &sc); |
1d82de61 | 3519 | |
b7ea3c41 MG |
3520 | /* |
3521 | * If we're getting trouble reclaiming, start doing writepage | |
3522 | * even in laptop mode. | |
3523 | */ | |
047d72c3 | 3524 | if (sc.priority < DEF_PRIORITY - 2) |
b7ea3c41 MG |
3525 | sc.may_writepage = 1; |
3526 | ||
1d82de61 MG |
3527 | /* Call soft limit reclaim before calling shrink_node. */ |
3528 | sc.nr_scanned = 0; | |
3529 | nr_soft_scanned = 0; | |
ef8f2327 | 3530 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
3531 | sc.gfp_mask, &nr_soft_scanned); |
3532 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3533 | ||
1da177e4 | 3534 | /* |
1d82de61 MG |
3535 | * There should be no need to raise the scanning priority if |
3536 | * enough pages are already being scanned that that high | |
3537 | * watermark would be met at 100% efficiency. | |
1da177e4 | 3538 | */ |
970a39a3 | 3539 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 3540 | raise_priority = false; |
5515061d MG |
3541 | |
3542 | /* | |
3543 | * If the low watermark is met there is no need for processes | |
3544 | * to be throttled on pfmemalloc_wait as they should not be | |
3545 | * able to safely make forward progress. Wake them | |
3546 | */ | |
3547 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
c73322d0 | 3548 | allow_direct_reclaim(pgdat)) |
cfc51155 | 3549 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3550 | |
b8e83b94 | 3551 | /* Check if kswapd should be suspending */ |
93781325 OS |
3552 | __fs_reclaim_release(); |
3553 | ret = try_to_freeze(); | |
3554 | __fs_reclaim_acquire(); | |
3555 | if (ret || kthread_should_stop()) | |
b8e83b94 | 3556 | break; |
8357376d | 3557 | |
73ce02e9 | 3558 | /* |
b8e83b94 MG |
3559 | * Raise priority if scanning rate is too low or there was no |
3560 | * progress in reclaiming pages | |
73ce02e9 | 3561 | */ |
c73322d0 JW |
3562 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
3563 | if (raise_priority || !nr_reclaimed) | |
b8e83b94 | 3564 | sc.priority--; |
1d82de61 | 3565 | } while (sc.priority >= 1); |
1da177e4 | 3566 | |
c73322d0 JW |
3567 | if (!sc.nr_reclaimed) |
3568 | pgdat->kswapd_failures++; | |
3569 | ||
b8e83b94 | 3570 | out: |
2a2e4885 | 3571 | snapshot_refaults(NULL, pgdat); |
93781325 | 3572 | __fs_reclaim_release(); |
0abdee2b | 3573 | /* |
1d82de61 MG |
3574 | * Return the order kswapd stopped reclaiming at as |
3575 | * prepare_kswapd_sleep() takes it into account. If another caller | |
3576 | * entered the allocator slow path while kswapd was awake, order will | |
3577 | * remain at the higher level. | |
0abdee2b | 3578 | */ |
1d82de61 | 3579 | return sc.order; |
1da177e4 LT |
3580 | } |
3581 | ||
e716f2eb MG |
3582 | /* |
3583 | * pgdat->kswapd_classzone_idx is the highest zone index that a recent | |
3584 | * allocation request woke kswapd for. When kswapd has not woken recently, | |
3585 | * the value is MAX_NR_ZONES which is not a valid index. This compares a | |
3586 | * given classzone and returns it or the highest classzone index kswapd | |
3587 | * was recently woke for. | |
3588 | */ | |
3589 | static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, | |
3590 | enum zone_type classzone_idx) | |
3591 | { | |
3592 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
3593 | return classzone_idx; | |
3594 | ||
3595 | return max(pgdat->kswapd_classzone_idx, classzone_idx); | |
3596 | } | |
3597 | ||
38087d9b MG |
3598 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
3599 | unsigned int classzone_idx) | |
f0bc0a60 KM |
3600 | { |
3601 | long remaining = 0; | |
3602 | DEFINE_WAIT(wait); | |
3603 | ||
3604 | if (freezing(current) || kthread_should_stop()) | |
3605 | return; | |
3606 | ||
3607 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3608 | ||
333b0a45 SG |
3609 | /* |
3610 | * Try to sleep for a short interval. Note that kcompactd will only be | |
3611 | * woken if it is possible to sleep for a short interval. This is | |
3612 | * deliberate on the assumption that if reclaim cannot keep an | |
3613 | * eligible zone balanced that it's also unlikely that compaction will | |
3614 | * succeed. | |
3615 | */ | |
d9f21d42 | 3616 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
fd901c95 VB |
3617 | /* |
3618 | * Compaction records what page blocks it recently failed to | |
3619 | * isolate pages from and skips them in the future scanning. | |
3620 | * When kswapd is going to sleep, it is reasonable to assume | |
3621 | * that pages and compaction may succeed so reset the cache. | |
3622 | */ | |
3623 | reset_isolation_suitable(pgdat); | |
3624 | ||
3625 | /* | |
3626 | * We have freed the memory, now we should compact it to make | |
3627 | * allocation of the requested order possible. | |
3628 | */ | |
38087d9b | 3629 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); |
fd901c95 | 3630 | |
f0bc0a60 | 3631 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
3632 | |
3633 | /* | |
3634 | * If woken prematurely then reset kswapd_classzone_idx and | |
3635 | * order. The values will either be from a wakeup request or | |
3636 | * the previous request that slept prematurely. | |
3637 | */ | |
3638 | if (remaining) { | |
e716f2eb | 3639 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
38087d9b MG |
3640 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); |
3641 | } | |
3642 | ||
f0bc0a60 KM |
3643 | finish_wait(&pgdat->kswapd_wait, &wait); |
3644 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3645 | } | |
3646 | ||
3647 | /* | |
3648 | * After a short sleep, check if it was a premature sleep. If not, then | |
3649 | * go fully to sleep until explicitly woken up. | |
3650 | */ | |
d9f21d42 MG |
3651 | if (!remaining && |
3652 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { | |
f0bc0a60 KM |
3653 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3654 | ||
3655 | /* | |
3656 | * vmstat counters are not perfectly accurate and the estimated | |
3657 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3658 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3659 | * watermarks being breached while under pressure, we reduce the | |
3660 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3661 | * them before going back to sleep. | |
3662 | */ | |
3663 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
3664 | |
3665 | if (!kthread_should_stop()) | |
3666 | schedule(); | |
3667 | ||
f0bc0a60 KM |
3668 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3669 | } else { | |
3670 | if (remaining) | |
3671 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3672 | else | |
3673 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3674 | } | |
3675 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3676 | } | |
3677 | ||
1da177e4 LT |
3678 | /* |
3679 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3680 | * from the init process. |
1da177e4 LT |
3681 | * |
3682 | * This basically trickles out pages so that we have _some_ | |
3683 | * free memory available even if there is no other activity | |
3684 | * that frees anything up. This is needed for things like routing | |
3685 | * etc, where we otherwise might have all activity going on in | |
3686 | * asynchronous contexts that cannot page things out. | |
3687 | * | |
3688 | * If there are applications that are active memory-allocators | |
3689 | * (most normal use), this basically shouldn't matter. | |
3690 | */ | |
3691 | static int kswapd(void *p) | |
3692 | { | |
e716f2eb MG |
3693 | unsigned int alloc_order, reclaim_order; |
3694 | unsigned int classzone_idx = MAX_NR_ZONES - 1; | |
1da177e4 LT |
3695 | pg_data_t *pgdat = (pg_data_t*)p; |
3696 | struct task_struct *tsk = current; | |
f0bc0a60 | 3697 | |
1da177e4 LT |
3698 | struct reclaim_state reclaim_state = { |
3699 | .reclaimed_slab = 0, | |
3700 | }; | |
a70f7302 | 3701 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3702 | |
174596a0 | 3703 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3704 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3705 | current->reclaim_state = &reclaim_state; |
3706 | ||
3707 | /* | |
3708 | * Tell the memory management that we're a "memory allocator", | |
3709 | * and that if we need more memory we should get access to it | |
3710 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3711 | * never get caught in the normal page freeing logic. | |
3712 | * | |
3713 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3714 | * you need a small amount of memory in order to be able to | |
3715 | * page out something else, and this flag essentially protects | |
3716 | * us from recursively trying to free more memory as we're | |
3717 | * trying to free the first piece of memory in the first place). | |
3718 | */ | |
930d9152 | 3719 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3720 | set_freezable(); |
1da177e4 | 3721 | |
e716f2eb MG |
3722 | pgdat->kswapd_order = 0; |
3723 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; | |
1da177e4 | 3724 | for ( ; ; ) { |
6f6313d4 | 3725 | bool ret; |
3e1d1d28 | 3726 | |
e716f2eb MG |
3727 | alloc_order = reclaim_order = pgdat->kswapd_order; |
3728 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); | |
3729 | ||
38087d9b MG |
3730 | kswapd_try_sleep: |
3731 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
3732 | classzone_idx); | |
215ddd66 | 3733 | |
38087d9b MG |
3734 | /* Read the new order and classzone_idx */ |
3735 | alloc_order = reclaim_order = pgdat->kswapd_order; | |
e716f2eb | 3736 | classzone_idx = kswapd_classzone_idx(pgdat, 0); |
38087d9b | 3737 | pgdat->kswapd_order = 0; |
e716f2eb | 3738 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; |
1da177e4 | 3739 | |
8fe23e05 DR |
3740 | ret = try_to_freeze(); |
3741 | if (kthread_should_stop()) | |
3742 | break; | |
3743 | ||
3744 | /* | |
3745 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3746 | * after returning from the refrigerator | |
3747 | */ | |
38087d9b MG |
3748 | if (ret) |
3749 | continue; | |
3750 | ||
3751 | /* | |
3752 | * Reclaim begins at the requested order but if a high-order | |
3753 | * reclaim fails then kswapd falls back to reclaiming for | |
3754 | * order-0. If that happens, kswapd will consider sleeping | |
3755 | * for the order it finished reclaiming at (reclaim_order) | |
3756 | * but kcompactd is woken to compact for the original | |
3757 | * request (alloc_order). | |
3758 | */ | |
e5146b12 MG |
3759 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, |
3760 | alloc_order); | |
38087d9b MG |
3761 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); |
3762 | if (reclaim_order < alloc_order) | |
3763 | goto kswapd_try_sleep; | |
1da177e4 | 3764 | } |
b0a8cc58 | 3765 | |
71abdc15 | 3766 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
b0a8cc58 | 3767 | current->reclaim_state = NULL; |
71abdc15 | 3768 | |
1da177e4 LT |
3769 | return 0; |
3770 | } | |
3771 | ||
3772 | /* | |
5ecd9d40 DR |
3773 | * A zone is low on free memory or too fragmented for high-order memory. If |
3774 | * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's | |
3775 | * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim | |
3776 | * has failed or is not needed, still wake up kcompactd if only compaction is | |
3777 | * needed. | |
1da177e4 | 3778 | */ |
5ecd9d40 DR |
3779 | void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, |
3780 | enum zone_type classzone_idx) | |
1da177e4 LT |
3781 | { |
3782 | pg_data_t *pgdat; | |
3783 | ||
6aa303de | 3784 | if (!managed_zone(zone)) |
1da177e4 LT |
3785 | return; |
3786 | ||
5ecd9d40 | 3787 | if (!cpuset_zone_allowed(zone, gfp_flags)) |
1da177e4 | 3788 | return; |
88f5acf8 | 3789 | pgdat = zone->zone_pgdat; |
e716f2eb MG |
3790 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, |
3791 | classzone_idx); | |
38087d9b | 3792 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); |
8d0986e2 | 3793 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3794 | return; |
e1a55637 | 3795 | |
5ecd9d40 DR |
3796 | /* Hopeless node, leave it to direct reclaim if possible */ |
3797 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES || | |
3798 | pgdat_balanced(pgdat, order, classzone_idx)) { | |
3799 | /* | |
3800 | * There may be plenty of free memory available, but it's too | |
3801 | * fragmented for high-order allocations. Wake up kcompactd | |
3802 | * and rely on compaction_suitable() to determine if it's | |
3803 | * needed. If it fails, it will defer subsequent attempts to | |
3804 | * ratelimit its work. | |
3805 | */ | |
3806 | if (!(gfp_flags & __GFP_DIRECT_RECLAIM)) | |
3807 | wakeup_kcompactd(pgdat, order, classzone_idx); | |
e716f2eb | 3808 | return; |
5ecd9d40 | 3809 | } |
88f5acf8 | 3810 | |
5ecd9d40 DR |
3811 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order, |
3812 | gfp_flags); | |
8d0986e2 | 3813 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3814 | } |
3815 | ||
c6f37f12 | 3816 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3817 | /* |
7b51755c | 3818 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3819 | * freed pages. |
3820 | * | |
3821 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3822 | * LRU order by reclaiming preferentially | |
3823 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3824 | */ |
7b51755c | 3825 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3826 | { |
d6277db4 | 3827 | struct reclaim_state reclaim_state; |
d6277db4 | 3828 | struct scan_control sc = { |
ee814fe2 | 3829 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 3830 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 3831 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 3832 | .priority = DEF_PRIORITY, |
d6277db4 | 3833 | .may_writepage = 1, |
ee814fe2 JW |
3834 | .may_unmap = 1, |
3835 | .may_swap = 1, | |
7b51755c | 3836 | .hibernation_mode = 1, |
1da177e4 | 3837 | }; |
a09ed5e0 | 3838 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c KM |
3839 | struct task_struct *p = current; |
3840 | unsigned long nr_reclaimed; | |
499118e9 | 3841 | unsigned int noreclaim_flag; |
1da177e4 | 3842 | |
d92a8cfc | 3843 | fs_reclaim_acquire(sc.gfp_mask); |
93781325 | 3844 | noreclaim_flag = memalloc_noreclaim_save(); |
7b51755c KM |
3845 | reclaim_state.reclaimed_slab = 0; |
3846 | p->reclaim_state = &reclaim_state; | |
d6277db4 | 3847 | |
3115cd91 | 3848 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 3849 | |
7b51755c | 3850 | p->reclaim_state = NULL; |
499118e9 | 3851 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 3852 | fs_reclaim_release(sc.gfp_mask); |
d6277db4 | 3853 | |
7b51755c | 3854 | return nr_reclaimed; |
1da177e4 | 3855 | } |
c6f37f12 | 3856 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3857 | |
1da177e4 LT |
3858 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3859 | not required for correctness. So if the last cpu in a node goes | |
3860 | away, we get changed to run anywhere: as the first one comes back, | |
3861 | restore their cpu bindings. */ | |
517bbed9 | 3862 | static int kswapd_cpu_online(unsigned int cpu) |
1da177e4 | 3863 | { |
58c0a4a7 | 3864 | int nid; |
1da177e4 | 3865 | |
517bbed9 SAS |
3866 | for_each_node_state(nid, N_MEMORY) { |
3867 | pg_data_t *pgdat = NODE_DATA(nid); | |
3868 | const struct cpumask *mask; | |
a70f7302 | 3869 | |
517bbed9 | 3870 | mask = cpumask_of_node(pgdat->node_id); |
c5f59f08 | 3871 | |
517bbed9 SAS |
3872 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
3873 | /* One of our CPUs online: restore mask */ | |
3874 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
1da177e4 | 3875 | } |
517bbed9 | 3876 | return 0; |
1da177e4 | 3877 | } |
1da177e4 | 3878 | |
3218ae14 YG |
3879 | /* |
3880 | * This kswapd start function will be called by init and node-hot-add. | |
3881 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
3882 | */ | |
3883 | int kswapd_run(int nid) | |
3884 | { | |
3885 | pg_data_t *pgdat = NODE_DATA(nid); | |
3886 | int ret = 0; | |
3887 | ||
3888 | if (pgdat->kswapd) | |
3889 | return 0; | |
3890 | ||
3891 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
3892 | if (IS_ERR(pgdat->kswapd)) { | |
3893 | /* failure at boot is fatal */ | |
c6202adf | 3894 | BUG_ON(system_state < SYSTEM_RUNNING); |
d5dc0ad9 GS |
3895 | pr_err("Failed to start kswapd on node %d\n", nid); |
3896 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 3897 | pgdat->kswapd = NULL; |
3218ae14 YG |
3898 | } |
3899 | return ret; | |
3900 | } | |
3901 | ||
8fe23e05 | 3902 | /* |
d8adde17 | 3903 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 3904 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
3905 | */ |
3906 | void kswapd_stop(int nid) | |
3907 | { | |
3908 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
3909 | ||
d8adde17 | 3910 | if (kswapd) { |
8fe23e05 | 3911 | kthread_stop(kswapd); |
d8adde17 JL |
3912 | NODE_DATA(nid)->kswapd = NULL; |
3913 | } | |
8fe23e05 DR |
3914 | } |
3915 | ||
1da177e4 LT |
3916 | static int __init kswapd_init(void) |
3917 | { | |
517bbed9 | 3918 | int nid, ret; |
69e05944 | 3919 | |
1da177e4 | 3920 | swap_setup(); |
48fb2e24 | 3921 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 3922 | kswapd_run(nid); |
517bbed9 SAS |
3923 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
3924 | "mm/vmscan:online", kswapd_cpu_online, | |
3925 | NULL); | |
3926 | WARN_ON(ret < 0); | |
1da177e4 LT |
3927 | return 0; |
3928 | } | |
3929 | ||
3930 | module_init(kswapd_init) | |
9eeff239 CL |
3931 | |
3932 | #ifdef CONFIG_NUMA | |
3933 | /* | |
a5f5f91d | 3934 | * Node reclaim mode |
9eeff239 | 3935 | * |
a5f5f91d | 3936 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 3937 | * the watermarks. |
9eeff239 | 3938 | */ |
a5f5f91d | 3939 | int node_reclaim_mode __read_mostly; |
9eeff239 | 3940 | |
1b2ffb78 | 3941 | #define RECLAIM_OFF 0 |
7d03431c | 3942 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 | 3943 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
95bbc0c7 | 3944 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
1b2ffb78 | 3945 | |
a92f7126 | 3946 | /* |
a5f5f91d | 3947 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
3948 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
3949 | * a zone. | |
3950 | */ | |
a5f5f91d | 3951 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 3952 | |
9614634f | 3953 | /* |
a5f5f91d | 3954 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
3955 | * occur. |
3956 | */ | |
3957 | int sysctl_min_unmapped_ratio = 1; | |
3958 | ||
0ff38490 CL |
3959 | /* |
3960 | * If the number of slab pages in a zone grows beyond this percentage then | |
3961 | * slab reclaim needs to occur. | |
3962 | */ | |
3963 | int sysctl_min_slab_ratio = 5; | |
3964 | ||
11fb9989 | 3965 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 3966 | { |
11fb9989 MG |
3967 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
3968 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
3969 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
3970 | |
3971 | /* | |
3972 | * It's possible for there to be more file mapped pages than | |
3973 | * accounted for by the pages on the file LRU lists because | |
3974 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
3975 | */ | |
3976 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
3977 | } | |
3978 | ||
3979 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 3980 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 3981 | { |
d031a157 AM |
3982 | unsigned long nr_pagecache_reclaimable; |
3983 | unsigned long delta = 0; | |
90afa5de MG |
3984 | |
3985 | /* | |
95bbc0c7 | 3986 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 3987 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 3988 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
3989 | * a better estimate |
3990 | */ | |
a5f5f91d MG |
3991 | if (node_reclaim_mode & RECLAIM_UNMAP) |
3992 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 3993 | else |
a5f5f91d | 3994 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
3995 | |
3996 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
3997 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
3998 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
3999 | |
4000 | /* Watch for any possible underflows due to delta */ | |
4001 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
4002 | delta = nr_pagecache_reclaimable; | |
4003 | ||
4004 | return nr_pagecache_reclaimable - delta; | |
4005 | } | |
4006 | ||
9eeff239 | 4007 | /* |
a5f5f91d | 4008 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 4009 | */ |
a5f5f91d | 4010 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 4011 | { |
7fb2d46d | 4012 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 4013 | const unsigned long nr_pages = 1 << order; |
9eeff239 CL |
4014 | struct task_struct *p = current; |
4015 | struct reclaim_state reclaim_state; | |
499118e9 | 4016 | unsigned int noreclaim_flag; |
179e9639 | 4017 | struct scan_control sc = { |
62b726c1 | 4018 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
f2f43e56 | 4019 | .gfp_mask = current_gfp_context(gfp_mask), |
bd2f6199 | 4020 | .order = order, |
a5f5f91d MG |
4021 | .priority = NODE_RECLAIM_PRIORITY, |
4022 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
4023 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 4024 | .may_swap = 1, |
f2f43e56 | 4025 | .reclaim_idx = gfp_zone(gfp_mask), |
179e9639 | 4026 | }; |
9eeff239 | 4027 | |
9eeff239 | 4028 | cond_resched(); |
93781325 | 4029 | fs_reclaim_acquire(sc.gfp_mask); |
d4f7796e | 4030 | /* |
95bbc0c7 | 4031 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 4032 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 4033 | * and RECLAIM_UNMAP. |
d4f7796e | 4034 | */ |
499118e9 VB |
4035 | noreclaim_flag = memalloc_noreclaim_save(); |
4036 | p->flags |= PF_SWAPWRITE; | |
9eeff239 CL |
4037 | reclaim_state.reclaimed_slab = 0; |
4038 | p->reclaim_state = &reclaim_state; | |
c84db23c | 4039 | |
a5f5f91d | 4040 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
0ff38490 | 4041 | /* |
894befec | 4042 | * Free memory by calling shrink node with increasing |
0ff38490 CL |
4043 | * priorities until we have enough memory freed. |
4044 | */ | |
0ff38490 | 4045 | do { |
970a39a3 | 4046 | shrink_node(pgdat, &sc); |
9e3b2f8c | 4047 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 4048 | } |
c84db23c | 4049 | |
9eeff239 | 4050 | p->reclaim_state = NULL; |
499118e9 VB |
4051 | current->flags &= ~PF_SWAPWRITE; |
4052 | memalloc_noreclaim_restore(noreclaim_flag); | |
93781325 | 4053 | fs_reclaim_release(sc.gfp_mask); |
a79311c1 | 4054 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 4055 | } |
179e9639 | 4056 | |
a5f5f91d | 4057 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 4058 | { |
d773ed6b | 4059 | int ret; |
179e9639 AM |
4060 | |
4061 | /* | |
a5f5f91d | 4062 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 4063 | * slab pages if we are over the defined limits. |
34aa1330 | 4064 | * |
9614634f CL |
4065 | * A small portion of unmapped file backed pages is needed for |
4066 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
4067 | * thrown out if the node is overallocated. So we do not reclaim |
4068 | * if less than a specified percentage of the node is used by | |
9614634f | 4069 | * unmapped file backed pages. |
179e9639 | 4070 | */ |
a5f5f91d | 4071 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
385386cf | 4072 | node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) |
a5f5f91d | 4073 | return NODE_RECLAIM_FULL; |
179e9639 AM |
4074 | |
4075 | /* | |
d773ed6b | 4076 | * Do not scan if the allocation should not be delayed. |
179e9639 | 4077 | */ |
d0164adc | 4078 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 4079 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
4080 | |
4081 | /* | |
a5f5f91d | 4082 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
4083 | * have associated processors. This will favor the local processor |
4084 | * over remote processors and spread off node memory allocations | |
4085 | * as wide as possible. | |
4086 | */ | |
a5f5f91d MG |
4087 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
4088 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 4089 | |
a5f5f91d MG |
4090 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
4091 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 4092 | |
a5f5f91d MG |
4093 | ret = __node_reclaim(pgdat, gfp_mask, order); |
4094 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 4095 | |
24cf7251 MG |
4096 | if (!ret) |
4097 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
4098 | ||
d773ed6b | 4099 | return ret; |
179e9639 | 4100 | } |
9eeff239 | 4101 | #endif |
894bc310 | 4102 | |
894bc310 LS |
4103 | /* |
4104 | * page_evictable - test whether a page is evictable | |
4105 | * @page: the page to test | |
894bc310 LS |
4106 | * |
4107 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 4108 | * lists vs unevictable list. |
894bc310 LS |
4109 | * |
4110 | * Reasons page might not be evictable: | |
ba9ddf49 | 4111 | * (1) page's mapping marked unevictable |
b291f000 | 4112 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 4113 | * |
894bc310 | 4114 | */ |
39b5f29a | 4115 | int page_evictable(struct page *page) |
894bc310 | 4116 | { |
e92bb4dd HY |
4117 | int ret; |
4118 | ||
4119 | /* Prevent address_space of inode and swap cache from being freed */ | |
4120 | rcu_read_lock(); | |
4121 | ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); | |
4122 | rcu_read_unlock(); | |
4123 | return ret; | |
894bc310 | 4124 | } |
89e004ea | 4125 | |
85046579 | 4126 | #ifdef CONFIG_SHMEM |
89e004ea | 4127 | /** |
24513264 HD |
4128 | * check_move_unevictable_pages - check pages for evictability and move to appropriate zone lru list |
4129 | * @pages: array of pages to check | |
4130 | * @nr_pages: number of pages to check | |
89e004ea | 4131 | * |
24513264 | 4132 | * Checks pages for evictability and moves them to the appropriate lru list. |
85046579 HD |
4133 | * |
4134 | * This function is only used for SysV IPC SHM_UNLOCK. | |
89e004ea | 4135 | */ |
24513264 | 4136 | void check_move_unevictable_pages(struct page **pages, int nr_pages) |
89e004ea | 4137 | { |
925b7673 | 4138 | struct lruvec *lruvec; |
785b99fe | 4139 | struct pglist_data *pgdat = NULL; |
24513264 HD |
4140 | int pgscanned = 0; |
4141 | int pgrescued = 0; | |
4142 | int i; | |
89e004ea | 4143 | |
24513264 HD |
4144 | for (i = 0; i < nr_pages; i++) { |
4145 | struct page *page = pages[i]; | |
785b99fe | 4146 | struct pglist_data *pagepgdat = page_pgdat(page); |
89e004ea | 4147 | |
24513264 | 4148 | pgscanned++; |
785b99fe MG |
4149 | if (pagepgdat != pgdat) { |
4150 | if (pgdat) | |
4151 | spin_unlock_irq(&pgdat->lru_lock); | |
4152 | pgdat = pagepgdat; | |
4153 | spin_lock_irq(&pgdat->lru_lock); | |
24513264 | 4154 | } |
785b99fe | 4155 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
89e004ea | 4156 | |
24513264 HD |
4157 | if (!PageLRU(page) || !PageUnevictable(page)) |
4158 | continue; | |
89e004ea | 4159 | |
39b5f29a | 4160 | if (page_evictable(page)) { |
24513264 HD |
4161 | enum lru_list lru = page_lru_base_type(page); |
4162 | ||
309381fe | 4163 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 4164 | ClearPageUnevictable(page); |
fa9add64 HD |
4165 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
4166 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 4167 | pgrescued++; |
89e004ea | 4168 | } |
24513264 | 4169 | } |
89e004ea | 4170 | |
785b99fe | 4171 | if (pgdat) { |
24513264 HD |
4172 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
4173 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
785b99fe | 4174 | spin_unlock_irq(&pgdat->lru_lock); |
89e004ea | 4175 | } |
89e004ea | 4176 | } |
85046579 | 4177 | #endif /* CONFIG_SHMEM */ |