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> |
64e3d12f | 49 | #include <linux/pagevec.h> |
268bb0ce | 50 | #include <linux/prefetch.h> |
b1de0d13 | 51 | #include <linux/printk.h> |
f9fe48be | 52 | #include <linux/dax.h> |
eb414681 | 53 | #include <linux/psi.h> |
1da177e4 LT |
54 | |
55 | #include <asm/tlbflush.h> | |
56 | #include <asm/div64.h> | |
57 | ||
58 | #include <linux/swapops.h> | |
117aad1e | 59 | #include <linux/balloon_compaction.h> |
1da177e4 | 60 | |
0f8053a5 NP |
61 | #include "internal.h" |
62 | ||
33906bc5 MG |
63 | #define CREATE_TRACE_POINTS |
64 | #include <trace/events/vmscan.h> | |
65 | ||
1da177e4 | 66 | struct scan_control { |
22fba335 KM |
67 | /* How many pages shrink_list() should reclaim */ |
68 | unsigned long nr_to_reclaim; | |
69 | ||
ee814fe2 JW |
70 | /* |
71 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
72 | * are scanned. | |
73 | */ | |
74 | nodemask_t *nodemask; | |
9e3b2f8c | 75 | |
f16015fb JW |
76 | /* |
77 | * The memory cgroup that hit its limit and as a result is the | |
78 | * primary target of this reclaim invocation. | |
79 | */ | |
80 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 81 | |
1276ad68 | 82 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
ee814fe2 JW |
83 | unsigned int may_writepage:1; |
84 | ||
85 | /* Can mapped pages be reclaimed? */ | |
86 | unsigned int may_unmap:1; | |
87 | ||
88 | /* Can pages be swapped as part of reclaim? */ | |
89 | unsigned int may_swap:1; | |
90 | ||
1c30844d MG |
91 | /* e.g. boosted watermark reclaim leaves slabs alone */ |
92 | unsigned int may_shrinkslab:1; | |
93 | ||
d6622f63 YX |
94 | /* |
95 | * Cgroups are not reclaimed below their configured memory.low, | |
96 | * unless we threaten to OOM. If any cgroups are skipped due to | |
97 | * memory.low and nothing was reclaimed, go back for memory.low. | |
98 | */ | |
99 | unsigned int memcg_low_reclaim:1; | |
100 | unsigned int memcg_low_skipped:1; | |
241994ed | 101 | |
ee814fe2 JW |
102 | unsigned int hibernation_mode:1; |
103 | ||
104 | /* One of the zones is ready for compaction */ | |
105 | unsigned int compaction_ready:1; | |
106 | ||
bb451fdf GT |
107 | /* Allocation order */ |
108 | s8 order; | |
109 | ||
110 | /* Scan (total_size >> priority) pages at once */ | |
111 | s8 priority; | |
112 | ||
113 | /* The highest zone to isolate pages for reclaim from */ | |
114 | s8 reclaim_idx; | |
115 | ||
116 | /* This context's GFP mask */ | |
117 | gfp_t gfp_mask; | |
118 | ||
ee814fe2 JW |
119 | /* Incremented by the number of inactive pages that were scanned */ |
120 | unsigned long nr_scanned; | |
121 | ||
122 | /* Number of pages freed so far during a call to shrink_zones() */ | |
123 | unsigned long nr_reclaimed; | |
d108c772 AR |
124 | |
125 | struct { | |
126 | unsigned int dirty; | |
127 | unsigned int unqueued_dirty; | |
128 | unsigned int congested; | |
129 | unsigned int writeback; | |
130 | unsigned int immediate; | |
131 | unsigned int file_taken; | |
132 | unsigned int taken; | |
133 | } nr; | |
e5ca8071 YS |
134 | |
135 | /* for recording the reclaimed slab by now */ | |
136 | struct reclaim_state reclaim_state; | |
1da177e4 LT |
137 | }; |
138 | ||
1da177e4 LT |
139 | #ifdef ARCH_HAS_PREFETCH |
140 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
141 | do { \ | |
142 | if ((_page)->lru.prev != _base) { \ | |
143 | struct page *prev; \ | |
144 | \ | |
145 | prev = lru_to_page(&(_page->lru)); \ | |
146 | prefetch(&prev->_field); \ | |
147 | } \ | |
148 | } while (0) | |
149 | #else | |
150 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
151 | #endif | |
152 | ||
153 | #ifdef ARCH_HAS_PREFETCHW | |
154 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
155 | do { \ | |
156 | if ((_page)->lru.prev != _base) { \ | |
157 | struct page *prev; \ | |
158 | \ | |
159 | prev = lru_to_page(&(_page->lru)); \ | |
160 | prefetchw(&prev->_field); \ | |
161 | } \ | |
162 | } while (0) | |
163 | #else | |
164 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
165 | #endif | |
166 | ||
167 | /* | |
168 | * From 0 .. 100. Higher means more swappy. | |
169 | */ | |
170 | int vm_swappiness = 60; | |
d0480be4 WSH |
171 | /* |
172 | * The total number of pages which are beyond the high watermark within all | |
173 | * zones. | |
174 | */ | |
175 | unsigned long vm_total_pages; | |
1da177e4 LT |
176 | |
177 | static LIST_HEAD(shrinker_list); | |
178 | static DECLARE_RWSEM(shrinker_rwsem); | |
179 | ||
b4c2b231 | 180 | #ifdef CONFIG_MEMCG_KMEM |
7e010df5 KT |
181 | |
182 | /* | |
183 | * We allow subsystems to populate their shrinker-related | |
184 | * LRU lists before register_shrinker_prepared() is called | |
185 | * for the shrinker, since we don't want to impose | |
186 | * restrictions on their internal registration order. | |
187 | * In this case shrink_slab_memcg() may find corresponding | |
188 | * bit is set in the shrinkers map. | |
189 | * | |
190 | * This value is used by the function to detect registering | |
191 | * shrinkers and to skip do_shrink_slab() calls for them. | |
192 | */ | |
193 | #define SHRINKER_REGISTERING ((struct shrinker *)~0UL) | |
194 | ||
b4c2b231 KT |
195 | static DEFINE_IDR(shrinker_idr); |
196 | static int shrinker_nr_max; | |
197 | ||
198 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) | |
199 | { | |
200 | int id, ret = -ENOMEM; | |
201 | ||
202 | down_write(&shrinker_rwsem); | |
203 | /* This may call shrinker, so it must use down_read_trylock() */ | |
7e010df5 | 204 | id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL); |
b4c2b231 KT |
205 | if (id < 0) |
206 | goto unlock; | |
207 | ||
0a4465d3 KT |
208 | if (id >= shrinker_nr_max) { |
209 | if (memcg_expand_shrinker_maps(id)) { | |
210 | idr_remove(&shrinker_idr, id); | |
211 | goto unlock; | |
212 | } | |
213 | ||
b4c2b231 | 214 | shrinker_nr_max = id + 1; |
0a4465d3 | 215 | } |
b4c2b231 KT |
216 | shrinker->id = id; |
217 | ret = 0; | |
218 | unlock: | |
219 | up_write(&shrinker_rwsem); | |
220 | return ret; | |
221 | } | |
222 | ||
223 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
224 | { | |
225 | int id = shrinker->id; | |
226 | ||
227 | BUG_ON(id < 0); | |
228 | ||
229 | down_write(&shrinker_rwsem); | |
230 | idr_remove(&shrinker_idr, id); | |
231 | up_write(&shrinker_rwsem); | |
232 | } | |
233 | #else /* CONFIG_MEMCG_KMEM */ | |
234 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) | |
235 | { | |
236 | return 0; | |
237 | } | |
238 | ||
239 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
240 | { | |
241 | } | |
242 | #endif /* CONFIG_MEMCG_KMEM */ | |
243 | ||
1732d2b0 AM |
244 | static void set_task_reclaim_state(struct task_struct *task, |
245 | struct reclaim_state *rs) | |
246 | { | |
247 | /* Check for an overwrite */ | |
248 | WARN_ON_ONCE(rs && task->reclaim_state); | |
249 | ||
250 | /* Check for the nulling of an already-nulled member */ | |
251 | WARN_ON_ONCE(!rs && !task->reclaim_state); | |
252 | ||
253 | task->reclaim_state = rs; | |
254 | } | |
255 | ||
c255a458 | 256 | #ifdef CONFIG_MEMCG |
89b5fae5 JW |
257 | static bool global_reclaim(struct scan_control *sc) |
258 | { | |
f16015fb | 259 | return !sc->target_mem_cgroup; |
89b5fae5 | 260 | } |
97c9341f TH |
261 | |
262 | /** | |
263 | * sane_reclaim - is the usual dirty throttling mechanism operational? | |
264 | * @sc: scan_control in question | |
265 | * | |
266 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
267 | * completely broken with the legacy memcg and direct stalling in | |
268 | * shrink_page_list() is used for throttling instead, which lacks all the | |
269 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
270 | * allocation and configurability. | |
271 | * | |
272 | * This function tests whether the vmscan currently in progress can assume | |
273 | * that the normal dirty throttling mechanism is operational. | |
274 | */ | |
275 | static bool sane_reclaim(struct scan_control *sc) | |
276 | { | |
277 | struct mem_cgroup *memcg = sc->target_mem_cgroup; | |
278 | ||
279 | if (!memcg) | |
280 | return true; | |
281 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 282 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
283 | return true; |
284 | #endif | |
285 | return false; | |
286 | } | |
e3c1ac58 AR |
287 | |
288 | static void set_memcg_congestion(pg_data_t *pgdat, | |
289 | struct mem_cgroup *memcg, | |
290 | bool congested) | |
291 | { | |
292 | struct mem_cgroup_per_node *mn; | |
293 | ||
294 | if (!memcg) | |
295 | return; | |
296 | ||
297 | mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id); | |
298 | WRITE_ONCE(mn->congested, congested); | |
299 | } | |
300 | ||
301 | static bool memcg_congested(pg_data_t *pgdat, | |
302 | struct mem_cgroup *memcg) | |
303 | { | |
304 | struct mem_cgroup_per_node *mn; | |
305 | ||
306 | mn = mem_cgroup_nodeinfo(memcg, pgdat->node_id); | |
307 | return READ_ONCE(mn->congested); | |
308 | ||
309 | } | |
91a45470 | 310 | #else |
89b5fae5 JW |
311 | static bool global_reclaim(struct scan_control *sc) |
312 | { | |
313 | return true; | |
314 | } | |
97c9341f TH |
315 | |
316 | static bool sane_reclaim(struct scan_control *sc) | |
317 | { | |
318 | return true; | |
319 | } | |
e3c1ac58 AR |
320 | |
321 | static inline void set_memcg_congestion(struct pglist_data *pgdat, | |
322 | struct mem_cgroup *memcg, bool congested) | |
323 | { | |
324 | } | |
325 | ||
326 | static inline bool memcg_congested(struct pglist_data *pgdat, | |
327 | struct mem_cgroup *memcg) | |
328 | { | |
329 | return false; | |
330 | ||
331 | } | |
91a45470 KH |
332 | #endif |
333 | ||
5a1c84b4 MG |
334 | /* |
335 | * This misses isolated pages which are not accounted for to save counters. | |
336 | * As the data only determines if reclaim or compaction continues, it is | |
337 | * not expected that isolated pages will be a dominating factor. | |
338 | */ | |
339 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
340 | { | |
341 | unsigned long nr; | |
342 | ||
343 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
344 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
345 | if (get_nr_swap_pages() > 0) | |
346 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + | |
347 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
348 | ||
349 | return nr; | |
350 | } | |
351 | ||
fd538803 MH |
352 | /** |
353 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
354 | * @lruvec: lru vector | |
355 | * @lru: lru to use | |
356 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) | |
357 | */ | |
358 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) | |
c9f299d9 | 359 | { |
fd538803 MH |
360 | unsigned long lru_size; |
361 | int zid; | |
362 | ||
c3c787e8 | 363 | if (!mem_cgroup_disabled()) |
205b20cc | 364 | lru_size = lruvec_page_state_local(lruvec, NR_LRU_BASE + lru); |
fd538803 MH |
365 | else |
366 | lru_size = node_page_state(lruvec_pgdat(lruvec), NR_LRU_BASE + lru); | |
a3d8e054 | 367 | |
fd538803 MH |
368 | for (zid = zone_idx + 1; zid < MAX_NR_ZONES; zid++) { |
369 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; | |
370 | unsigned long size; | |
c9f299d9 | 371 | |
fd538803 MH |
372 | if (!managed_zone(zone)) |
373 | continue; | |
374 | ||
375 | if (!mem_cgroup_disabled()) | |
376 | size = mem_cgroup_get_zone_lru_size(lruvec, lru, zid); | |
377 | else | |
378 | size = zone_page_state(&lruvec_pgdat(lruvec)->node_zones[zid], | |
379 | NR_ZONE_LRU_BASE + lru); | |
380 | lru_size -= min(size, lru_size); | |
381 | } | |
382 | ||
383 | return lru_size; | |
b4536f0c | 384 | |
b4536f0c MH |
385 | } |
386 | ||
1da177e4 | 387 | /* |
1d3d4437 | 388 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 389 | */ |
8e04944f | 390 | int prealloc_shrinker(struct shrinker *shrinker) |
1da177e4 | 391 | { |
b9726c26 | 392 | unsigned int size = sizeof(*shrinker->nr_deferred); |
1d3d4437 | 393 | |
1d3d4437 GC |
394 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
395 | size *= nr_node_ids; | |
396 | ||
397 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
398 | if (!shrinker->nr_deferred) | |
399 | return -ENOMEM; | |
b4c2b231 KT |
400 | |
401 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) { | |
402 | if (prealloc_memcg_shrinker(shrinker)) | |
403 | goto free_deferred; | |
404 | } | |
405 | ||
8e04944f | 406 | return 0; |
b4c2b231 KT |
407 | |
408 | free_deferred: | |
409 | kfree(shrinker->nr_deferred); | |
410 | shrinker->nr_deferred = NULL; | |
411 | return -ENOMEM; | |
8e04944f TH |
412 | } |
413 | ||
414 | void free_prealloced_shrinker(struct shrinker *shrinker) | |
415 | { | |
b4c2b231 KT |
416 | if (!shrinker->nr_deferred) |
417 | return; | |
418 | ||
419 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) | |
420 | unregister_memcg_shrinker(shrinker); | |
421 | ||
8e04944f TH |
422 | kfree(shrinker->nr_deferred); |
423 | shrinker->nr_deferred = NULL; | |
424 | } | |
1d3d4437 | 425 | |
8e04944f TH |
426 | void register_shrinker_prepared(struct shrinker *shrinker) |
427 | { | |
8e1f936b RR |
428 | down_write(&shrinker_rwsem); |
429 | list_add_tail(&shrinker->list, &shrinker_list); | |
7e010df5 | 430 | #ifdef CONFIG_MEMCG_KMEM |
8df4a44c KT |
431 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) |
432 | idr_replace(&shrinker_idr, shrinker, shrinker->id); | |
7e010df5 | 433 | #endif |
8e1f936b | 434 | up_write(&shrinker_rwsem); |
8e04944f TH |
435 | } |
436 | ||
437 | int register_shrinker(struct shrinker *shrinker) | |
438 | { | |
439 | int err = prealloc_shrinker(shrinker); | |
440 | ||
441 | if (err) | |
442 | return err; | |
443 | register_shrinker_prepared(shrinker); | |
1d3d4437 | 444 | return 0; |
1da177e4 | 445 | } |
8e1f936b | 446 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
447 | |
448 | /* | |
449 | * Remove one | |
450 | */ | |
8e1f936b | 451 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 | 452 | { |
bb422a73 TH |
453 | if (!shrinker->nr_deferred) |
454 | return; | |
b4c2b231 KT |
455 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) |
456 | unregister_memcg_shrinker(shrinker); | |
1da177e4 LT |
457 | down_write(&shrinker_rwsem); |
458 | list_del(&shrinker->list); | |
459 | up_write(&shrinker_rwsem); | |
ae393321 | 460 | kfree(shrinker->nr_deferred); |
bb422a73 | 461 | shrinker->nr_deferred = NULL; |
1da177e4 | 462 | } |
8e1f936b | 463 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
464 | |
465 | #define SHRINK_BATCH 128 | |
1d3d4437 | 466 | |
cb731d6c | 467 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
9092c71b | 468 | struct shrinker *shrinker, int priority) |
1d3d4437 GC |
469 | { |
470 | unsigned long freed = 0; | |
471 | unsigned long long delta; | |
472 | long total_scan; | |
d5bc5fd3 | 473 | long freeable; |
1d3d4437 GC |
474 | long nr; |
475 | long new_nr; | |
476 | int nid = shrinkctl->nid; | |
477 | long batch_size = shrinker->batch ? shrinker->batch | |
478 | : SHRINK_BATCH; | |
5f33a080 | 479 | long scanned = 0, next_deferred; |
1d3d4437 | 480 | |
ac7fb3ad KT |
481 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
482 | nid = 0; | |
483 | ||
d5bc5fd3 | 484 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
9b996468 KT |
485 | if (freeable == 0 || freeable == SHRINK_EMPTY) |
486 | return freeable; | |
1d3d4437 GC |
487 | |
488 | /* | |
489 | * copy the current shrinker scan count into a local variable | |
490 | * and zero it so that other concurrent shrinker invocations | |
491 | * don't also do this scanning work. | |
492 | */ | |
493 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
494 | ||
495 | total_scan = nr; | |
4b85afbd JW |
496 | if (shrinker->seeks) { |
497 | delta = freeable >> priority; | |
498 | delta *= 4; | |
499 | do_div(delta, shrinker->seeks); | |
500 | } else { | |
501 | /* | |
502 | * These objects don't require any IO to create. Trim | |
503 | * them aggressively under memory pressure to keep | |
504 | * them from causing refetches in the IO caches. | |
505 | */ | |
506 | delta = freeable / 2; | |
507 | } | |
172b06c3 | 508 | |
1d3d4437 GC |
509 | total_scan += delta; |
510 | if (total_scan < 0) { | |
d75f773c | 511 | pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n", |
a0b02131 | 512 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 513 | total_scan = freeable; |
5f33a080 SL |
514 | next_deferred = nr; |
515 | } else | |
516 | next_deferred = total_scan; | |
1d3d4437 GC |
517 | |
518 | /* | |
519 | * We need to avoid excessive windup on filesystem shrinkers | |
520 | * due to large numbers of GFP_NOFS allocations causing the | |
521 | * shrinkers to return -1 all the time. This results in a large | |
522 | * nr being built up so when a shrink that can do some work | |
523 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 524 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
525 | * memory. |
526 | * | |
527 | * Hence only allow the shrinker to scan the entire cache when | |
528 | * a large delta change is calculated directly. | |
529 | */ | |
d5bc5fd3 VD |
530 | if (delta < freeable / 4) |
531 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
532 | |
533 | /* | |
534 | * Avoid risking looping forever due to too large nr value: | |
535 | * never try to free more than twice the estimate number of | |
536 | * freeable entries. | |
537 | */ | |
d5bc5fd3 VD |
538 | if (total_scan > freeable * 2) |
539 | total_scan = freeable * 2; | |
1d3d4437 GC |
540 | |
541 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
9092c71b | 542 | freeable, delta, total_scan, priority); |
1d3d4437 | 543 | |
0b1fb40a VD |
544 | /* |
545 | * Normally, we should not scan less than batch_size objects in one | |
546 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
547 | * than batch_size objects in total and we are really tight on memory, | |
548 | * we will try to reclaim all available objects, otherwise we can end | |
549 | * up failing allocations although there are plenty of reclaimable | |
550 | * objects spread over several slabs with usage less than the | |
551 | * batch_size. | |
552 | * | |
553 | * We detect the "tight on memory" situations by looking at the total | |
554 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 555 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
556 | * scanning at high prio and therefore should try to reclaim as much as |
557 | * possible. | |
558 | */ | |
559 | while (total_scan >= batch_size || | |
d5bc5fd3 | 560 | total_scan >= freeable) { |
a0b02131 | 561 | unsigned long ret; |
0b1fb40a | 562 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 563 | |
0b1fb40a | 564 | shrinkctl->nr_to_scan = nr_to_scan; |
d460acb5 | 565 | shrinkctl->nr_scanned = nr_to_scan; |
a0b02131 DC |
566 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
567 | if (ret == SHRINK_STOP) | |
568 | break; | |
569 | freed += ret; | |
1d3d4437 | 570 | |
d460acb5 CW |
571 | count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); |
572 | total_scan -= shrinkctl->nr_scanned; | |
573 | scanned += shrinkctl->nr_scanned; | |
1d3d4437 GC |
574 | |
575 | cond_resched(); | |
576 | } | |
577 | ||
5f33a080 SL |
578 | if (next_deferred >= scanned) |
579 | next_deferred -= scanned; | |
580 | else | |
581 | next_deferred = 0; | |
1d3d4437 GC |
582 | /* |
583 | * move the unused scan count back into the shrinker in a | |
584 | * manner that handles concurrent updates. If we exhausted the | |
585 | * scan, there is no need to do an update. | |
586 | */ | |
5f33a080 SL |
587 | if (next_deferred > 0) |
588 | new_nr = atomic_long_add_return(next_deferred, | |
1d3d4437 GC |
589 | &shrinker->nr_deferred[nid]); |
590 | else | |
591 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
592 | ||
df9024a8 | 593 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 594 | return freed; |
1495f230 YH |
595 | } |
596 | ||
b0dedc49 KT |
597 | #ifdef CONFIG_MEMCG_KMEM |
598 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, | |
599 | struct mem_cgroup *memcg, int priority) | |
600 | { | |
601 | struct memcg_shrinker_map *map; | |
b8e57efa KT |
602 | unsigned long ret, freed = 0; |
603 | int i; | |
b0dedc49 KT |
604 | |
605 | if (!memcg_kmem_enabled() || !mem_cgroup_online(memcg)) | |
606 | return 0; | |
607 | ||
608 | if (!down_read_trylock(&shrinker_rwsem)) | |
609 | return 0; | |
610 | ||
611 | map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map, | |
612 | true); | |
613 | if (unlikely(!map)) | |
614 | goto unlock; | |
615 | ||
616 | for_each_set_bit(i, map->map, shrinker_nr_max) { | |
617 | struct shrink_control sc = { | |
618 | .gfp_mask = gfp_mask, | |
619 | .nid = nid, | |
620 | .memcg = memcg, | |
621 | }; | |
622 | struct shrinker *shrinker; | |
623 | ||
624 | shrinker = idr_find(&shrinker_idr, i); | |
7e010df5 KT |
625 | if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) { |
626 | if (!shrinker) | |
627 | clear_bit(i, map->map); | |
b0dedc49 KT |
628 | continue; |
629 | } | |
630 | ||
b0dedc49 | 631 | ret = do_shrink_slab(&sc, shrinker, priority); |
f90280d6 KT |
632 | if (ret == SHRINK_EMPTY) { |
633 | clear_bit(i, map->map); | |
634 | /* | |
635 | * After the shrinker reported that it had no objects to | |
636 | * free, but before we cleared the corresponding bit in | |
637 | * the memcg shrinker map, a new object might have been | |
638 | * added. To make sure, we have the bit set in this | |
639 | * case, we invoke the shrinker one more time and reset | |
640 | * the bit if it reports that it is not empty anymore. | |
641 | * The memory barrier here pairs with the barrier in | |
642 | * memcg_set_shrinker_bit(): | |
643 | * | |
644 | * list_lru_add() shrink_slab_memcg() | |
645 | * list_add_tail() clear_bit() | |
646 | * <MB> <MB> | |
647 | * set_bit() do_shrink_slab() | |
648 | */ | |
649 | smp_mb__after_atomic(); | |
650 | ret = do_shrink_slab(&sc, shrinker, priority); | |
651 | if (ret == SHRINK_EMPTY) | |
652 | ret = 0; | |
653 | else | |
654 | memcg_set_shrinker_bit(memcg, nid, i); | |
655 | } | |
b0dedc49 KT |
656 | freed += ret; |
657 | ||
658 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
659 | freed = freed ? : 1; | |
660 | break; | |
661 | } | |
662 | } | |
663 | unlock: | |
664 | up_read(&shrinker_rwsem); | |
665 | return freed; | |
666 | } | |
667 | #else /* CONFIG_MEMCG_KMEM */ | |
668 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, | |
669 | struct mem_cgroup *memcg, int priority) | |
670 | { | |
671 | return 0; | |
672 | } | |
673 | #endif /* CONFIG_MEMCG_KMEM */ | |
674 | ||
6b4f7799 | 675 | /** |
cb731d6c | 676 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
677 | * @gfp_mask: allocation context |
678 | * @nid: node whose slab caches to target | |
cb731d6c | 679 | * @memcg: memory cgroup whose slab caches to target |
9092c71b | 680 | * @priority: the reclaim priority |
1da177e4 | 681 | * |
6b4f7799 | 682 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 683 | * |
6b4f7799 JW |
684 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
685 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 686 | * |
aeed1d32 VD |
687 | * @memcg specifies the memory cgroup to target. Unaware shrinkers |
688 | * are called only if it is the root cgroup. | |
cb731d6c | 689 | * |
9092c71b JB |
690 | * @priority is sc->priority, we take the number of objects and >> by priority |
691 | * in order to get the scan target. | |
b15e0905 | 692 | * |
6b4f7799 | 693 | * Returns the number of reclaimed slab objects. |
1da177e4 | 694 | */ |
cb731d6c VD |
695 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
696 | struct mem_cgroup *memcg, | |
9092c71b | 697 | int priority) |
1da177e4 | 698 | { |
b8e57efa | 699 | unsigned long ret, freed = 0; |
1da177e4 LT |
700 | struct shrinker *shrinker; |
701 | ||
fa1e512f YS |
702 | /* |
703 | * The root memcg might be allocated even though memcg is disabled | |
704 | * via "cgroup_disable=memory" boot parameter. This could make | |
705 | * mem_cgroup_is_root() return false, then just run memcg slab | |
706 | * shrink, but skip global shrink. This may result in premature | |
707 | * oom. | |
708 | */ | |
709 | if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg)) | |
b0dedc49 | 710 | return shrink_slab_memcg(gfp_mask, nid, memcg, priority); |
cb731d6c | 711 | |
e830c63a | 712 | if (!down_read_trylock(&shrinker_rwsem)) |
f06590bd | 713 | goto out; |
1da177e4 LT |
714 | |
715 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
716 | struct shrink_control sc = { |
717 | .gfp_mask = gfp_mask, | |
718 | .nid = nid, | |
cb731d6c | 719 | .memcg = memcg, |
6b4f7799 | 720 | }; |
ec97097b | 721 | |
9b996468 KT |
722 | ret = do_shrink_slab(&sc, shrinker, priority); |
723 | if (ret == SHRINK_EMPTY) | |
724 | ret = 0; | |
725 | freed += ret; | |
e496612c MK |
726 | /* |
727 | * Bail out if someone want to register a new shrinker to | |
728 | * prevent the regsitration from being stalled for long periods | |
729 | * by parallel ongoing shrinking. | |
730 | */ | |
731 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
732 | freed = freed ? : 1; | |
733 | break; | |
734 | } | |
1da177e4 | 735 | } |
6b4f7799 | 736 | |
1da177e4 | 737 | up_read(&shrinker_rwsem); |
f06590bd MK |
738 | out: |
739 | cond_resched(); | |
24f7c6b9 | 740 | return freed; |
1da177e4 LT |
741 | } |
742 | ||
cb731d6c VD |
743 | void drop_slab_node(int nid) |
744 | { | |
745 | unsigned long freed; | |
746 | ||
747 | do { | |
748 | struct mem_cgroup *memcg = NULL; | |
749 | ||
750 | freed = 0; | |
aeed1d32 | 751 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
cb731d6c | 752 | do { |
9092c71b | 753 | freed += shrink_slab(GFP_KERNEL, nid, memcg, 0); |
cb731d6c VD |
754 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); |
755 | } while (freed > 10); | |
756 | } | |
757 | ||
758 | void drop_slab(void) | |
759 | { | |
760 | int nid; | |
761 | ||
762 | for_each_online_node(nid) | |
763 | drop_slab_node(nid); | |
764 | } | |
765 | ||
1da177e4 LT |
766 | static inline int is_page_cache_freeable(struct page *page) |
767 | { | |
ceddc3a5 JW |
768 | /* |
769 | * A freeable page cache page is referenced only by the caller | |
67891fff MW |
770 | * that isolated the page, the page cache and optional buffer |
771 | * heads at page->private. | |
ceddc3a5 | 772 | */ |
67891fff | 773 | int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ? |
bd4c82c2 | 774 | HPAGE_PMD_NR : 1; |
67891fff | 775 | return page_count(page) - page_has_private(page) == 1 + page_cache_pins; |
1da177e4 LT |
776 | } |
777 | ||
703c2708 | 778 | static int may_write_to_inode(struct inode *inode, struct scan_control *sc) |
1da177e4 | 779 | { |
930d9152 | 780 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 781 | return 1; |
703c2708 | 782 | if (!inode_write_congested(inode)) |
1da177e4 | 783 | return 1; |
703c2708 | 784 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
785 | return 1; |
786 | return 0; | |
787 | } | |
788 | ||
789 | /* | |
790 | * We detected a synchronous write error writing a page out. Probably | |
791 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
792 | * fsync(), msync() or close(). | |
793 | * | |
794 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
795 | * prevents it from being freed up. But we have a ref on the page and once | |
796 | * that page is locked, the mapping is pinned. | |
797 | * | |
798 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
799 | * __GFP_FS. | |
800 | */ | |
801 | static void handle_write_error(struct address_space *mapping, | |
802 | struct page *page, int error) | |
803 | { | |
7eaceacc | 804 | lock_page(page); |
3e9f45bd GC |
805 | if (page_mapping(page) == mapping) |
806 | mapping_set_error(mapping, error); | |
1da177e4 LT |
807 | unlock_page(page); |
808 | } | |
809 | ||
04e62a29 CL |
810 | /* possible outcome of pageout() */ |
811 | typedef enum { | |
812 | /* failed to write page out, page is locked */ | |
813 | PAGE_KEEP, | |
814 | /* move page to the active list, page is locked */ | |
815 | PAGE_ACTIVATE, | |
816 | /* page has been sent to the disk successfully, page is unlocked */ | |
817 | PAGE_SUCCESS, | |
818 | /* page is clean and locked */ | |
819 | PAGE_CLEAN, | |
820 | } pageout_t; | |
821 | ||
1da177e4 | 822 | /* |
1742f19f AM |
823 | * pageout is called by shrink_page_list() for each dirty page. |
824 | * Calls ->writepage(). | |
1da177e4 | 825 | */ |
c661b078 | 826 | static pageout_t pageout(struct page *page, struct address_space *mapping, |
7d3579e8 | 827 | struct scan_control *sc) |
1da177e4 LT |
828 | { |
829 | /* | |
830 | * If the page is dirty, only perform writeback if that write | |
831 | * will be non-blocking. To prevent this allocation from being | |
832 | * stalled by pagecache activity. But note that there may be | |
833 | * stalls if we need to run get_block(). We could test | |
834 | * PagePrivate for that. | |
835 | * | |
8174202b | 836 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
837 | * this page's queue, we can perform writeback even if that |
838 | * will block. | |
839 | * | |
840 | * If the page is swapcache, write it back even if that would | |
841 | * block, for some throttling. This happens by accident, because | |
842 | * swap_backing_dev_info is bust: it doesn't reflect the | |
843 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
844 | */ |
845 | if (!is_page_cache_freeable(page)) | |
846 | return PAGE_KEEP; | |
847 | if (!mapping) { | |
848 | /* | |
849 | * Some data journaling orphaned pages can have | |
850 | * page->mapping == NULL while being dirty with clean buffers. | |
851 | */ | |
266cf658 | 852 | if (page_has_private(page)) { |
1da177e4 LT |
853 | if (try_to_free_buffers(page)) { |
854 | ClearPageDirty(page); | |
b1de0d13 | 855 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
856 | return PAGE_CLEAN; |
857 | } | |
858 | } | |
859 | return PAGE_KEEP; | |
860 | } | |
861 | if (mapping->a_ops->writepage == NULL) | |
862 | return PAGE_ACTIVATE; | |
703c2708 | 863 | if (!may_write_to_inode(mapping->host, sc)) |
1da177e4 LT |
864 | return PAGE_KEEP; |
865 | ||
866 | if (clear_page_dirty_for_io(page)) { | |
867 | int res; | |
868 | struct writeback_control wbc = { | |
869 | .sync_mode = WB_SYNC_NONE, | |
870 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
871 | .range_start = 0, |
872 | .range_end = LLONG_MAX, | |
1da177e4 LT |
873 | .for_reclaim = 1, |
874 | }; | |
875 | ||
876 | SetPageReclaim(page); | |
877 | res = mapping->a_ops->writepage(page, &wbc); | |
878 | if (res < 0) | |
879 | handle_write_error(mapping, page, res); | |
994fc28c | 880 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
881 | ClearPageReclaim(page); |
882 | return PAGE_ACTIVATE; | |
883 | } | |
c661b078 | 884 | |
1da177e4 LT |
885 | if (!PageWriteback(page)) { |
886 | /* synchronous write or broken a_ops? */ | |
887 | ClearPageReclaim(page); | |
888 | } | |
3aa23851 | 889 | trace_mm_vmscan_writepage(page); |
c4a25635 | 890 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
891 | return PAGE_SUCCESS; |
892 | } | |
893 | ||
894 | return PAGE_CLEAN; | |
895 | } | |
896 | ||
a649fd92 | 897 | /* |
e286781d NP |
898 | * Same as remove_mapping, but if the page is removed from the mapping, it |
899 | * gets returned with a refcount of 0. | |
a649fd92 | 900 | */ |
a528910e JW |
901 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
902 | bool reclaimed) | |
49d2e9cc | 903 | { |
c4843a75 | 904 | unsigned long flags; |
bd4c82c2 | 905 | int refcount; |
c4843a75 | 906 | |
28e4d965 NP |
907 | BUG_ON(!PageLocked(page)); |
908 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 909 | |
b93b0163 | 910 | xa_lock_irqsave(&mapping->i_pages, flags); |
49d2e9cc | 911 | /* |
0fd0e6b0 NP |
912 | * The non racy check for a busy page. |
913 | * | |
914 | * Must be careful with the order of the tests. When someone has | |
915 | * a ref to the page, it may be possible that they dirty it then | |
916 | * drop the reference. So if PageDirty is tested before page_count | |
917 | * here, then the following race may occur: | |
918 | * | |
919 | * get_user_pages(&page); | |
920 | * [user mapping goes away] | |
921 | * write_to(page); | |
922 | * !PageDirty(page) [good] | |
923 | * SetPageDirty(page); | |
924 | * put_page(page); | |
925 | * !page_count(page) [good, discard it] | |
926 | * | |
927 | * [oops, our write_to data is lost] | |
928 | * | |
929 | * Reversing the order of the tests ensures such a situation cannot | |
930 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 931 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
932 | * |
933 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
b93b0163 | 934 | * and thus under the i_pages lock, then this ordering is not required. |
49d2e9cc | 935 | */ |
bd4c82c2 HY |
936 | if (unlikely(PageTransHuge(page)) && PageSwapCache(page)) |
937 | refcount = 1 + HPAGE_PMD_NR; | |
938 | else | |
939 | refcount = 2; | |
940 | if (!page_ref_freeze(page, refcount)) | |
49d2e9cc | 941 | goto cannot_free; |
1c4c3b99 | 942 | /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */ |
e286781d | 943 | if (unlikely(PageDirty(page))) { |
bd4c82c2 | 944 | page_ref_unfreeze(page, refcount); |
49d2e9cc | 945 | goto cannot_free; |
e286781d | 946 | } |
49d2e9cc CL |
947 | |
948 | if (PageSwapCache(page)) { | |
949 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 950 | mem_cgroup_swapout(page, swap); |
4e17ec25 | 951 | __delete_from_swap_cache(page, swap); |
b93b0163 | 952 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
75f6d6d2 | 953 | put_swap_page(page, swap); |
e286781d | 954 | } else { |
6072d13c | 955 | void (*freepage)(struct page *); |
a528910e | 956 | void *shadow = NULL; |
6072d13c LT |
957 | |
958 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
959 | /* |
960 | * Remember a shadow entry for reclaimed file cache in | |
961 | * order to detect refaults, thus thrashing, later on. | |
962 | * | |
963 | * But don't store shadows in an address space that is | |
964 | * already exiting. This is not just an optizimation, | |
965 | * inode reclaim needs to empty out the radix tree or | |
966 | * the nodes are lost. Don't plant shadows behind its | |
967 | * back. | |
f9fe48be RZ |
968 | * |
969 | * We also don't store shadows for DAX mappings because the | |
970 | * only page cache pages found in these are zero pages | |
971 | * covering holes, and because we don't want to mix DAX | |
972 | * exceptional entries and shadow exceptional entries in the | |
b93b0163 | 973 | * same address_space. |
a528910e JW |
974 | */ |
975 | if (reclaimed && page_is_file_cache(page) && | |
f9fe48be | 976 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
a7ca12f9 | 977 | shadow = workingset_eviction(page); |
62cccb8c | 978 | __delete_from_page_cache(page, shadow); |
b93b0163 | 979 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
6072d13c LT |
980 | |
981 | if (freepage != NULL) | |
982 | freepage(page); | |
49d2e9cc CL |
983 | } |
984 | ||
49d2e9cc CL |
985 | return 1; |
986 | ||
987 | cannot_free: | |
b93b0163 | 988 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
49d2e9cc CL |
989 | return 0; |
990 | } | |
991 | ||
e286781d NP |
992 | /* |
993 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
994 | * someone else has a ref on the page, abort and return 0. If it was | |
995 | * successfully detached, return 1. Assumes the caller has a single ref on | |
996 | * this page. | |
997 | */ | |
998 | int remove_mapping(struct address_space *mapping, struct page *page) | |
999 | { | |
a528910e | 1000 | if (__remove_mapping(mapping, page, false)) { |
e286781d NP |
1001 | /* |
1002 | * Unfreezing the refcount with 1 rather than 2 effectively | |
1003 | * drops the pagecache ref for us without requiring another | |
1004 | * atomic operation. | |
1005 | */ | |
fe896d18 | 1006 | page_ref_unfreeze(page, 1); |
e286781d NP |
1007 | return 1; |
1008 | } | |
1009 | return 0; | |
1010 | } | |
1011 | ||
894bc310 LS |
1012 | /** |
1013 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
1014 | * @page: page to be put back to appropriate lru list | |
1015 | * | |
1016 | * Add previously isolated @page to appropriate LRU list. | |
1017 | * Page may still be unevictable for other reasons. | |
1018 | * | |
1019 | * lru_lock must not be held, interrupts must be enabled. | |
1020 | */ | |
894bc310 LS |
1021 | void putback_lru_page(struct page *page) |
1022 | { | |
9c4e6b1a | 1023 | lru_cache_add(page); |
894bc310 LS |
1024 | put_page(page); /* drop ref from isolate */ |
1025 | } | |
1026 | ||
dfc8d636 JW |
1027 | enum page_references { |
1028 | PAGEREF_RECLAIM, | |
1029 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 1030 | PAGEREF_KEEP, |
dfc8d636 JW |
1031 | PAGEREF_ACTIVATE, |
1032 | }; | |
1033 | ||
1034 | static enum page_references page_check_references(struct page *page, | |
1035 | struct scan_control *sc) | |
1036 | { | |
64574746 | 1037 | int referenced_ptes, referenced_page; |
dfc8d636 | 1038 | unsigned long vm_flags; |
dfc8d636 | 1039 | |
c3ac9a8a JW |
1040 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
1041 | &vm_flags); | |
64574746 | 1042 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 1043 | |
dfc8d636 JW |
1044 | /* |
1045 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
1046 | * move the page to the unevictable list. | |
1047 | */ | |
1048 | if (vm_flags & VM_LOCKED) | |
1049 | return PAGEREF_RECLAIM; | |
1050 | ||
64574746 | 1051 | if (referenced_ptes) { |
e4898273 | 1052 | if (PageSwapBacked(page)) |
64574746 JW |
1053 | return PAGEREF_ACTIVATE; |
1054 | /* | |
1055 | * All mapped pages start out with page table | |
1056 | * references from the instantiating fault, so we need | |
1057 | * to look twice if a mapped file page is used more | |
1058 | * than once. | |
1059 | * | |
1060 | * Mark it and spare it for another trip around the | |
1061 | * inactive list. Another page table reference will | |
1062 | * lead to its activation. | |
1063 | * | |
1064 | * Note: the mark is set for activated pages as well | |
1065 | * so that recently deactivated but used pages are | |
1066 | * quickly recovered. | |
1067 | */ | |
1068 | SetPageReferenced(page); | |
1069 | ||
34dbc67a | 1070 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
1071 | return PAGEREF_ACTIVATE; |
1072 | ||
c909e993 KK |
1073 | /* |
1074 | * Activate file-backed executable pages after first usage. | |
1075 | */ | |
1076 | if (vm_flags & VM_EXEC) | |
1077 | return PAGEREF_ACTIVATE; | |
1078 | ||
64574746 JW |
1079 | return PAGEREF_KEEP; |
1080 | } | |
dfc8d636 JW |
1081 | |
1082 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 1083 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
1084 | return PAGEREF_RECLAIM_CLEAN; |
1085 | ||
1086 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
1087 | } |
1088 | ||
e2be15f6 MG |
1089 | /* Check if a page is dirty or under writeback */ |
1090 | static void page_check_dirty_writeback(struct page *page, | |
1091 | bool *dirty, bool *writeback) | |
1092 | { | |
b4597226 MG |
1093 | struct address_space *mapping; |
1094 | ||
e2be15f6 MG |
1095 | /* |
1096 | * Anonymous pages are not handled by flushers and must be written | |
1097 | * from reclaim context. Do not stall reclaim based on them | |
1098 | */ | |
802a3a92 SL |
1099 | if (!page_is_file_cache(page) || |
1100 | (PageAnon(page) && !PageSwapBacked(page))) { | |
e2be15f6 MG |
1101 | *dirty = false; |
1102 | *writeback = false; | |
1103 | return; | |
1104 | } | |
1105 | ||
1106 | /* By default assume that the page flags are accurate */ | |
1107 | *dirty = PageDirty(page); | |
1108 | *writeback = PageWriteback(page); | |
b4597226 MG |
1109 | |
1110 | /* Verify dirty/writeback state if the filesystem supports it */ | |
1111 | if (!page_has_private(page)) | |
1112 | return; | |
1113 | ||
1114 | mapping = page_mapping(page); | |
1115 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
1116 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
1117 | } |
1118 | ||
1da177e4 | 1119 | /* |
1742f19f | 1120 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 1121 | */ |
1742f19f | 1122 | static unsigned long shrink_page_list(struct list_head *page_list, |
599d0c95 | 1123 | struct pglist_data *pgdat, |
f84f6e2b | 1124 | struct scan_control *sc, |
02c6de8d | 1125 | enum ttu_flags ttu_flags, |
3c710c1a | 1126 | struct reclaim_stat *stat, |
02c6de8d | 1127 | bool force_reclaim) |
1da177e4 LT |
1128 | { |
1129 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 1130 | LIST_HEAD(free_pages); |
3c710c1a | 1131 | unsigned nr_reclaimed = 0; |
886cf190 | 1132 | unsigned pgactivate = 0; |
1da177e4 | 1133 | |
060f005f | 1134 | memset(stat, 0, sizeof(*stat)); |
1da177e4 LT |
1135 | cond_resched(); |
1136 | ||
1da177e4 LT |
1137 | while (!list_empty(page_list)) { |
1138 | struct address_space *mapping; | |
1139 | struct page *page; | |
1140 | int may_enter_fs; | |
02c6de8d | 1141 | enum page_references references = PAGEREF_RECLAIM_CLEAN; |
e2be15f6 | 1142 | bool dirty, writeback; |
98879b3b | 1143 | unsigned int nr_pages; |
1da177e4 LT |
1144 | |
1145 | cond_resched(); | |
1146 | ||
1147 | page = lru_to_page(page_list); | |
1148 | list_del(&page->lru); | |
1149 | ||
529ae9aa | 1150 | if (!trylock_page(page)) |
1da177e4 LT |
1151 | goto keep; |
1152 | ||
309381fe | 1153 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 | 1154 | |
98879b3b YS |
1155 | nr_pages = 1 << compound_order(page); |
1156 | ||
1157 | /* Account the number of base pages even though THP */ | |
1158 | sc->nr_scanned += nr_pages; | |
80e43426 | 1159 | |
39b5f29a | 1160 | if (unlikely(!page_evictable(page))) |
ad6b6704 | 1161 | goto activate_locked; |
894bc310 | 1162 | |
a6dc60f8 | 1163 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
1164 | goto keep_locked; |
1165 | ||
c661b078 AW |
1166 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
1167 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
1168 | ||
e2be15f6 | 1169 | /* |
894befec | 1170 | * The number of dirty pages determines if a node is marked |
e2be15f6 MG |
1171 | * reclaim_congested which affects wait_iff_congested. kswapd |
1172 | * will stall and start writing pages if the tail of the LRU | |
1173 | * is all dirty unqueued pages. | |
1174 | */ | |
1175 | page_check_dirty_writeback(page, &dirty, &writeback); | |
1176 | if (dirty || writeback) | |
060f005f | 1177 | stat->nr_dirty++; |
e2be15f6 MG |
1178 | |
1179 | if (dirty && !writeback) | |
060f005f | 1180 | stat->nr_unqueued_dirty++; |
e2be15f6 | 1181 | |
d04e8acd MG |
1182 | /* |
1183 | * Treat this page as congested if the underlying BDI is or if | |
1184 | * pages are cycling through the LRU so quickly that the | |
1185 | * pages marked for immediate reclaim are making it to the | |
1186 | * end of the LRU a second time. | |
1187 | */ | |
e2be15f6 | 1188 | mapping = page_mapping(page); |
1da58ee2 | 1189 | if (((dirty || writeback) && mapping && |
703c2708 | 1190 | inode_write_congested(mapping->host)) || |
d04e8acd | 1191 | (writeback && PageReclaim(page))) |
060f005f | 1192 | stat->nr_congested++; |
e2be15f6 | 1193 | |
283aba9f MG |
1194 | /* |
1195 | * If a page at the tail of the LRU is under writeback, there | |
1196 | * are three cases to consider. | |
1197 | * | |
1198 | * 1) If reclaim is encountering an excessive number of pages | |
1199 | * under writeback and this page is both under writeback and | |
1200 | * PageReclaim then it indicates that pages are being queued | |
1201 | * for IO but are being recycled through the LRU before the | |
1202 | * IO can complete. Waiting on the page itself risks an | |
1203 | * indefinite stall if it is impossible to writeback the | |
1204 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
1205 | * note that the LRU is being scanned too quickly and the |
1206 | * caller can stall after page list has been processed. | |
283aba9f | 1207 | * |
97c9341f | 1208 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
1209 | * not marked for immediate reclaim, or the caller does not |
1210 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
1211 | * not to fs). In this case mark the page for immediate | |
97c9341f | 1212 | * reclaim and continue scanning. |
283aba9f | 1213 | * |
ecf5fc6e MH |
1214 | * Require may_enter_fs because we would wait on fs, which |
1215 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
1216 | * enter reclaim, and deadlock if it waits on a page for |
1217 | * which it is needed to do the write (loop masks off | |
1218 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1219 | * would probably show more reasons. | |
1220 | * | |
7fadc820 | 1221 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1222 | * PageReclaim. memcg does not have any dirty pages |
1223 | * throttling so we could easily OOM just because too many | |
1224 | * pages are in writeback and there is nothing else to | |
1225 | * reclaim. Wait for the writeback to complete. | |
c55e8d03 JW |
1226 | * |
1227 | * In cases 1) and 2) we activate the pages to get them out of | |
1228 | * the way while we continue scanning for clean pages on the | |
1229 | * inactive list and refilling from the active list. The | |
1230 | * observation here is that waiting for disk writes is more | |
1231 | * expensive than potentially causing reloads down the line. | |
1232 | * Since they're marked for immediate reclaim, they won't put | |
1233 | * memory pressure on the cache working set any longer than it | |
1234 | * takes to write them to disk. | |
283aba9f | 1235 | */ |
c661b078 | 1236 | if (PageWriteback(page)) { |
283aba9f MG |
1237 | /* Case 1 above */ |
1238 | if (current_is_kswapd() && | |
1239 | PageReclaim(page) && | |
599d0c95 | 1240 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
060f005f | 1241 | stat->nr_immediate++; |
c55e8d03 | 1242 | goto activate_locked; |
283aba9f MG |
1243 | |
1244 | /* Case 2 above */ | |
97c9341f | 1245 | } else if (sane_reclaim(sc) || |
ecf5fc6e | 1246 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1247 | /* |
1248 | * This is slightly racy - end_page_writeback() | |
1249 | * might have just cleared PageReclaim, then | |
1250 | * setting PageReclaim here end up interpreted | |
1251 | * as PageReadahead - but that does not matter | |
1252 | * enough to care. What we do want is for this | |
1253 | * page to have PageReclaim set next time memcg | |
1254 | * reclaim reaches the tests above, so it will | |
1255 | * then wait_on_page_writeback() to avoid OOM; | |
1256 | * and it's also appropriate in global reclaim. | |
1257 | */ | |
1258 | SetPageReclaim(page); | |
060f005f | 1259 | stat->nr_writeback++; |
c55e8d03 | 1260 | goto activate_locked; |
283aba9f MG |
1261 | |
1262 | /* Case 3 above */ | |
1263 | } else { | |
7fadc820 | 1264 | unlock_page(page); |
283aba9f | 1265 | wait_on_page_writeback(page); |
7fadc820 HD |
1266 | /* then go back and try same page again */ |
1267 | list_add_tail(&page->lru, page_list); | |
1268 | continue; | |
e62e384e | 1269 | } |
c661b078 | 1270 | } |
1da177e4 | 1271 | |
02c6de8d MK |
1272 | if (!force_reclaim) |
1273 | references = page_check_references(page, sc); | |
1274 | ||
dfc8d636 JW |
1275 | switch (references) { |
1276 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1277 | goto activate_locked; |
64574746 | 1278 | case PAGEREF_KEEP: |
98879b3b | 1279 | stat->nr_ref_keep += nr_pages; |
64574746 | 1280 | goto keep_locked; |
dfc8d636 JW |
1281 | case PAGEREF_RECLAIM: |
1282 | case PAGEREF_RECLAIM_CLEAN: | |
1283 | ; /* try to reclaim the page below */ | |
1284 | } | |
1da177e4 | 1285 | |
1da177e4 LT |
1286 | /* |
1287 | * Anonymous process memory has backing store? | |
1288 | * Try to allocate it some swap space here. | |
802a3a92 | 1289 | * Lazyfree page could be freed directly |
1da177e4 | 1290 | */ |
bd4c82c2 HY |
1291 | if (PageAnon(page) && PageSwapBacked(page)) { |
1292 | if (!PageSwapCache(page)) { | |
1293 | if (!(sc->gfp_mask & __GFP_IO)) | |
1294 | goto keep_locked; | |
1295 | if (PageTransHuge(page)) { | |
1296 | /* cannot split THP, skip it */ | |
1297 | if (!can_split_huge_page(page, NULL)) | |
1298 | goto activate_locked; | |
1299 | /* | |
1300 | * Split pages without a PMD map right | |
1301 | * away. Chances are some or all of the | |
1302 | * tail pages can be freed without IO. | |
1303 | */ | |
1304 | if (!compound_mapcount(page) && | |
1305 | split_huge_page_to_list(page, | |
1306 | page_list)) | |
1307 | goto activate_locked; | |
1308 | } | |
1309 | if (!add_to_swap(page)) { | |
1310 | if (!PageTransHuge(page)) | |
98879b3b | 1311 | goto activate_locked_split; |
bd4c82c2 HY |
1312 | /* Fallback to swap normal pages */ |
1313 | if (split_huge_page_to_list(page, | |
1314 | page_list)) | |
1315 | goto activate_locked; | |
fe490cc0 HY |
1316 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1317 | count_vm_event(THP_SWPOUT_FALLBACK); | |
1318 | #endif | |
bd4c82c2 | 1319 | if (!add_to_swap(page)) |
98879b3b | 1320 | goto activate_locked_split; |
bd4c82c2 | 1321 | } |
0f074658 | 1322 | |
bd4c82c2 | 1323 | may_enter_fs = 1; |
1da177e4 | 1324 | |
bd4c82c2 HY |
1325 | /* Adding to swap updated mapping */ |
1326 | mapping = page_mapping(page); | |
1327 | } | |
7751b2da KS |
1328 | } else if (unlikely(PageTransHuge(page))) { |
1329 | /* Split file THP */ | |
1330 | if (split_huge_page_to_list(page, page_list)) | |
1331 | goto keep_locked; | |
e2be15f6 | 1332 | } |
1da177e4 | 1333 | |
98879b3b YS |
1334 | /* |
1335 | * THP may get split above, need minus tail pages and update | |
1336 | * nr_pages to avoid accounting tail pages twice. | |
1337 | * | |
1338 | * The tail pages that are added into swap cache successfully | |
1339 | * reach here. | |
1340 | */ | |
1341 | if ((nr_pages > 1) && !PageTransHuge(page)) { | |
1342 | sc->nr_scanned -= (nr_pages - 1); | |
1343 | nr_pages = 1; | |
1344 | } | |
1345 | ||
1da177e4 LT |
1346 | /* |
1347 | * The page is mapped into the page tables of one or more | |
1348 | * processes. Try to unmap it here. | |
1349 | */ | |
802a3a92 | 1350 | if (page_mapped(page)) { |
bd4c82c2 HY |
1351 | enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH; |
1352 | ||
1353 | if (unlikely(PageTransHuge(page))) | |
1354 | flags |= TTU_SPLIT_HUGE_PMD; | |
1355 | if (!try_to_unmap(page, flags)) { | |
98879b3b | 1356 | stat->nr_unmap_fail += nr_pages; |
1da177e4 | 1357 | goto activate_locked; |
1da177e4 LT |
1358 | } |
1359 | } | |
1360 | ||
1361 | if (PageDirty(page)) { | |
ee72886d | 1362 | /* |
4eda4823 JW |
1363 | * Only kswapd can writeback filesystem pages |
1364 | * to avoid risk of stack overflow. But avoid | |
1365 | * injecting inefficient single-page IO into | |
1366 | * flusher writeback as much as possible: only | |
1367 | * write pages when we've encountered many | |
1368 | * dirty pages, and when we've already scanned | |
1369 | * the rest of the LRU for clean pages and see | |
1370 | * the same dirty pages again (PageReclaim). | |
ee72886d | 1371 | */ |
f84f6e2b | 1372 | if (page_is_file_cache(page) && |
4eda4823 JW |
1373 | (!current_is_kswapd() || !PageReclaim(page) || |
1374 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { | |
49ea7eb6 MG |
1375 | /* |
1376 | * Immediately reclaim when written back. | |
1377 | * Similar in principal to deactivate_page() | |
1378 | * except we already have the page isolated | |
1379 | * and know it's dirty | |
1380 | */ | |
c4a25635 | 1381 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
49ea7eb6 MG |
1382 | SetPageReclaim(page); |
1383 | ||
c55e8d03 | 1384 | goto activate_locked; |
ee72886d MG |
1385 | } |
1386 | ||
dfc8d636 | 1387 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1388 | goto keep_locked; |
4dd4b920 | 1389 | if (!may_enter_fs) |
1da177e4 | 1390 | goto keep_locked; |
52a8363e | 1391 | if (!sc->may_writepage) |
1da177e4 LT |
1392 | goto keep_locked; |
1393 | ||
d950c947 MG |
1394 | /* |
1395 | * Page is dirty. Flush the TLB if a writable entry | |
1396 | * potentially exists to avoid CPU writes after IO | |
1397 | * starts and then write it out here. | |
1398 | */ | |
1399 | try_to_unmap_flush_dirty(); | |
7d3579e8 | 1400 | switch (pageout(page, mapping, sc)) { |
1da177e4 LT |
1401 | case PAGE_KEEP: |
1402 | goto keep_locked; | |
1403 | case PAGE_ACTIVATE: | |
1404 | goto activate_locked; | |
1405 | case PAGE_SUCCESS: | |
7d3579e8 | 1406 | if (PageWriteback(page)) |
41ac1999 | 1407 | goto keep; |
7d3579e8 | 1408 | if (PageDirty(page)) |
1da177e4 | 1409 | goto keep; |
7d3579e8 | 1410 | |
1da177e4 LT |
1411 | /* |
1412 | * A synchronous write - probably a ramdisk. Go | |
1413 | * ahead and try to reclaim the page. | |
1414 | */ | |
529ae9aa | 1415 | if (!trylock_page(page)) |
1da177e4 LT |
1416 | goto keep; |
1417 | if (PageDirty(page) || PageWriteback(page)) | |
1418 | goto keep_locked; | |
1419 | mapping = page_mapping(page); | |
1420 | case PAGE_CLEAN: | |
1421 | ; /* try to free the page below */ | |
1422 | } | |
1423 | } | |
1424 | ||
1425 | /* | |
1426 | * If the page has buffers, try to free the buffer mappings | |
1427 | * associated with this page. If we succeed we try to free | |
1428 | * the page as well. | |
1429 | * | |
1430 | * We do this even if the page is PageDirty(). | |
1431 | * try_to_release_page() does not perform I/O, but it is | |
1432 | * possible for a page to have PageDirty set, but it is actually | |
1433 | * clean (all its buffers are clean). This happens if the | |
1434 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1435 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1436 | * try_to_release_page() will discover that cleanness and will |
1437 | * drop the buffers and mark the page clean - it can be freed. | |
1438 | * | |
1439 | * Rarely, pages can have buffers and no ->mapping. These are | |
1440 | * the pages which were not successfully invalidated in | |
1441 | * truncate_complete_page(). We try to drop those buffers here | |
1442 | * and if that worked, and the page is no longer mapped into | |
1443 | * process address space (page_count == 1) it can be freed. | |
1444 | * Otherwise, leave the page on the LRU so it is swappable. | |
1445 | */ | |
266cf658 | 1446 | if (page_has_private(page)) { |
1da177e4 LT |
1447 | if (!try_to_release_page(page, sc->gfp_mask)) |
1448 | goto activate_locked; | |
e286781d NP |
1449 | if (!mapping && page_count(page) == 1) { |
1450 | unlock_page(page); | |
1451 | if (put_page_testzero(page)) | |
1452 | goto free_it; | |
1453 | else { | |
1454 | /* | |
1455 | * rare race with speculative reference. | |
1456 | * the speculative reference will free | |
1457 | * this page shortly, so we may | |
1458 | * increment nr_reclaimed here (and | |
1459 | * leave it off the LRU). | |
1460 | */ | |
1461 | nr_reclaimed++; | |
1462 | continue; | |
1463 | } | |
1464 | } | |
1da177e4 LT |
1465 | } |
1466 | ||
802a3a92 SL |
1467 | if (PageAnon(page) && !PageSwapBacked(page)) { |
1468 | /* follow __remove_mapping for reference */ | |
1469 | if (!page_ref_freeze(page, 1)) | |
1470 | goto keep_locked; | |
1471 | if (PageDirty(page)) { | |
1472 | page_ref_unfreeze(page, 1); | |
1473 | goto keep_locked; | |
1474 | } | |
1da177e4 | 1475 | |
802a3a92 | 1476 | count_vm_event(PGLAZYFREED); |
2262185c | 1477 | count_memcg_page_event(page, PGLAZYFREED); |
802a3a92 SL |
1478 | } else if (!mapping || !__remove_mapping(mapping, page, true)) |
1479 | goto keep_locked; | |
9a1ea439 HD |
1480 | |
1481 | unlock_page(page); | |
e286781d | 1482 | free_it: |
98879b3b YS |
1483 | /* |
1484 | * THP may get swapped out in a whole, need account | |
1485 | * all base pages. | |
1486 | */ | |
1487 | nr_reclaimed += nr_pages; | |
abe4c3b5 MG |
1488 | |
1489 | /* | |
1490 | * Is there need to periodically free_page_list? It would | |
1491 | * appear not as the counts should be low | |
1492 | */ | |
bd4c82c2 HY |
1493 | if (unlikely(PageTransHuge(page))) { |
1494 | mem_cgroup_uncharge(page); | |
1495 | (*get_compound_page_dtor(page))(page); | |
1496 | } else | |
1497 | list_add(&page->lru, &free_pages); | |
1da177e4 LT |
1498 | continue; |
1499 | ||
98879b3b YS |
1500 | activate_locked_split: |
1501 | /* | |
1502 | * The tail pages that are failed to add into swap cache | |
1503 | * reach here. Fixup nr_scanned and nr_pages. | |
1504 | */ | |
1505 | if (nr_pages > 1) { | |
1506 | sc->nr_scanned -= (nr_pages - 1); | |
1507 | nr_pages = 1; | |
1508 | } | |
1da177e4 | 1509 | activate_locked: |
68a22394 | 1510 | /* Not a candidate for swapping, so reclaim swap space. */ |
ad6b6704 MK |
1511 | if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || |
1512 | PageMlocked(page))) | |
a2c43eed | 1513 | try_to_free_swap(page); |
309381fe | 1514 | VM_BUG_ON_PAGE(PageActive(page), page); |
ad6b6704 | 1515 | if (!PageMlocked(page)) { |
886cf190 | 1516 | int type = page_is_file_cache(page); |
ad6b6704 | 1517 | SetPageActive(page); |
98879b3b | 1518 | stat->nr_activate[type] += nr_pages; |
2262185c | 1519 | count_memcg_page_event(page, PGACTIVATE); |
ad6b6704 | 1520 | } |
1da177e4 LT |
1521 | keep_locked: |
1522 | unlock_page(page); | |
1523 | keep: | |
1524 | list_add(&page->lru, &ret_pages); | |
309381fe | 1525 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1526 | } |
abe4c3b5 | 1527 | |
98879b3b YS |
1528 | pgactivate = stat->nr_activate[0] + stat->nr_activate[1]; |
1529 | ||
747db954 | 1530 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1531 | try_to_unmap_flush(); |
2d4894b5 | 1532 | free_unref_page_list(&free_pages); |
abe4c3b5 | 1533 | |
1da177e4 | 1534 | list_splice(&ret_pages, page_list); |
886cf190 | 1535 | count_vm_events(PGACTIVATE, pgactivate); |
060f005f | 1536 | |
05ff5137 | 1537 | return nr_reclaimed; |
1da177e4 LT |
1538 | } |
1539 | ||
02c6de8d MK |
1540 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1541 | struct list_head *page_list) | |
1542 | { | |
1543 | struct scan_control sc = { | |
1544 | .gfp_mask = GFP_KERNEL, | |
1545 | .priority = DEF_PRIORITY, | |
1546 | .may_unmap = 1, | |
1547 | }; | |
060f005f | 1548 | struct reclaim_stat dummy_stat; |
3c710c1a | 1549 | unsigned long ret; |
02c6de8d MK |
1550 | struct page *page, *next; |
1551 | LIST_HEAD(clean_pages); | |
1552 | ||
1553 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e | 1554 | if (page_is_file_cache(page) && !PageDirty(page) && |
a58f2cef | 1555 | !__PageMovable(page) && !PageUnevictable(page)) { |
02c6de8d MK |
1556 | ClearPageActive(page); |
1557 | list_move(&page->lru, &clean_pages); | |
1558 | } | |
1559 | } | |
1560 | ||
599d0c95 | 1561 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
060f005f | 1562 | TTU_IGNORE_ACCESS, &dummy_stat, true); |
02c6de8d | 1563 | list_splice(&clean_pages, page_list); |
599d0c95 | 1564 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1565 | return ret; |
1566 | } | |
1567 | ||
5ad333eb AW |
1568 | /* |
1569 | * Attempt to remove the specified page from its LRU. Only take this page | |
1570 | * if it is of the appropriate PageActive status. Pages which are being | |
1571 | * freed elsewhere are also ignored. | |
1572 | * | |
1573 | * page: page to consider | |
1574 | * mode: one of the LRU isolation modes defined above | |
1575 | * | |
1576 | * returns 0 on success, -ve errno on failure. | |
1577 | */ | |
f3fd4a61 | 1578 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1579 | { |
1580 | int ret = -EINVAL; | |
1581 | ||
1582 | /* Only take pages on the LRU. */ | |
1583 | if (!PageLRU(page)) | |
1584 | return ret; | |
1585 | ||
e46a2879 MK |
1586 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1587 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1588 | return ret; |
1589 | ||
5ad333eb | 1590 | ret = -EBUSY; |
08e552c6 | 1591 | |
c8244935 MG |
1592 | /* |
1593 | * To minimise LRU disruption, the caller can indicate that it only | |
1594 | * wants to isolate pages it will be able to operate on without | |
1595 | * blocking - clean pages for the most part. | |
1596 | * | |
c8244935 MG |
1597 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
1598 | * that it is possible to migrate without blocking | |
1599 | */ | |
1276ad68 | 1600 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
c8244935 MG |
1601 | /* All the caller can do on PageWriteback is block */ |
1602 | if (PageWriteback(page)) | |
1603 | return ret; | |
1604 | ||
1605 | if (PageDirty(page)) { | |
1606 | struct address_space *mapping; | |
69d763fc | 1607 | bool migrate_dirty; |
c8244935 | 1608 | |
c8244935 MG |
1609 | /* |
1610 | * Only pages without mappings or that have a | |
1611 | * ->migratepage callback are possible to migrate | |
69d763fc MG |
1612 | * without blocking. However, we can be racing with |
1613 | * truncation so it's necessary to lock the page | |
1614 | * to stabilise the mapping as truncation holds | |
1615 | * the page lock until after the page is removed | |
1616 | * from the page cache. | |
c8244935 | 1617 | */ |
69d763fc MG |
1618 | if (!trylock_page(page)) |
1619 | return ret; | |
1620 | ||
c8244935 | 1621 | mapping = page_mapping(page); |
145e1a71 | 1622 | migrate_dirty = !mapping || mapping->a_ops->migratepage; |
69d763fc MG |
1623 | unlock_page(page); |
1624 | if (!migrate_dirty) | |
c8244935 MG |
1625 | return ret; |
1626 | } | |
1627 | } | |
39deaf85 | 1628 | |
f80c0673 MK |
1629 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1630 | return ret; | |
1631 | ||
5ad333eb AW |
1632 | if (likely(get_page_unless_zero(page))) { |
1633 | /* | |
1634 | * Be careful not to clear PageLRU until after we're | |
1635 | * sure the page is not being freed elsewhere -- the | |
1636 | * page release code relies on it. | |
1637 | */ | |
1638 | ClearPageLRU(page); | |
1639 | ret = 0; | |
1640 | } | |
1641 | ||
1642 | return ret; | |
1643 | } | |
1644 | ||
7ee36a14 MG |
1645 | |
1646 | /* | |
1647 | * Update LRU sizes after isolating pages. The LRU size updates must | |
1648 | * be complete before mem_cgroup_update_lru_size due to a santity check. | |
1649 | */ | |
1650 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 1651 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 1652 | { |
7ee36a14 MG |
1653 | int zid; |
1654 | ||
7ee36a14 MG |
1655 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1656 | if (!nr_zone_taken[zid]) | |
1657 | continue; | |
1658 | ||
1659 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); | |
7ee36a14 | 1660 | #ifdef CONFIG_MEMCG |
b4536f0c | 1661 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
7ee36a14 | 1662 | #endif |
b4536f0c MH |
1663 | } |
1664 | ||
7ee36a14 MG |
1665 | } |
1666 | ||
f4b7e272 AR |
1667 | /** |
1668 | * pgdat->lru_lock is heavily contended. Some of the functions that | |
1da177e4 LT |
1669 | * shrink the lists perform better by taking out a batch of pages |
1670 | * and working on them outside the LRU lock. | |
1671 | * | |
1672 | * For pagecache intensive workloads, this function is the hottest | |
1673 | * spot in the kernel (apart from copy_*_user functions). | |
1674 | * | |
1675 | * Appropriate locks must be held before calling this function. | |
1676 | * | |
791b48b6 | 1677 | * @nr_to_scan: The number of eligible pages to look through on the list. |
5dc35979 | 1678 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1679 | * @dst: The temp list to put pages on to. |
f626012d | 1680 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1681 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1682 | * @mode: One of the LRU isolation modes |
3cb99451 | 1683 | * @lru: LRU list id for isolating |
1da177e4 LT |
1684 | * |
1685 | * returns how many pages were moved onto *@dst. | |
1686 | */ | |
69e05944 | 1687 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1688 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1689 | unsigned long *nr_scanned, struct scan_control *sc, |
a9e7c39f | 1690 | enum lru_list lru) |
1da177e4 | 1691 | { |
75b00af7 | 1692 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1693 | unsigned long nr_taken = 0; |
599d0c95 | 1694 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 1695 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
3db65812 | 1696 | unsigned long skipped = 0; |
791b48b6 | 1697 | unsigned long scan, total_scan, nr_pages; |
b2e18757 | 1698 | LIST_HEAD(pages_skipped); |
a9e7c39f | 1699 | isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED); |
1da177e4 | 1700 | |
98879b3b | 1701 | total_scan = 0; |
791b48b6 | 1702 | scan = 0; |
98879b3b | 1703 | while (scan < nr_to_scan && !list_empty(src)) { |
5ad333eb | 1704 | struct page *page; |
5ad333eb | 1705 | |
1da177e4 LT |
1706 | page = lru_to_page(src); |
1707 | prefetchw_prev_lru_page(page, src, flags); | |
1708 | ||
309381fe | 1709 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1710 | |
98879b3b YS |
1711 | nr_pages = 1 << compound_order(page); |
1712 | total_scan += nr_pages; | |
1713 | ||
b2e18757 MG |
1714 | if (page_zonenum(page) > sc->reclaim_idx) { |
1715 | list_move(&page->lru, &pages_skipped); | |
98879b3b | 1716 | nr_skipped[page_zonenum(page)] += nr_pages; |
b2e18757 MG |
1717 | continue; |
1718 | } | |
1719 | ||
791b48b6 MK |
1720 | /* |
1721 | * Do not count skipped pages because that makes the function | |
1722 | * return with no isolated pages if the LRU mostly contains | |
1723 | * ineligible pages. This causes the VM to not reclaim any | |
1724 | * pages, triggering a premature OOM. | |
98879b3b YS |
1725 | * |
1726 | * Account all tail pages of THP. This would not cause | |
1727 | * premature OOM since __isolate_lru_page() returns -EBUSY | |
1728 | * only when the page is being freed somewhere else. | |
791b48b6 | 1729 | */ |
98879b3b | 1730 | scan += nr_pages; |
f3fd4a61 | 1731 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1732 | case 0: |
599d0c95 MG |
1733 | nr_taken += nr_pages; |
1734 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
5ad333eb | 1735 | list_move(&page->lru, dst); |
5ad333eb AW |
1736 | break; |
1737 | ||
1738 | case -EBUSY: | |
1739 | /* else it is being freed elsewhere */ | |
1740 | list_move(&page->lru, src); | |
1741 | continue; | |
46453a6e | 1742 | |
5ad333eb AW |
1743 | default: |
1744 | BUG(); | |
1745 | } | |
1da177e4 LT |
1746 | } |
1747 | ||
b2e18757 MG |
1748 | /* |
1749 | * Splice any skipped pages to the start of the LRU list. Note that | |
1750 | * this disrupts the LRU order when reclaiming for lower zones but | |
1751 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
1752 | * scanning would soon rescan the same pages to skip and put the | |
1753 | * system at risk of premature OOM. | |
1754 | */ | |
7cc30fcf MG |
1755 | if (!list_empty(&pages_skipped)) { |
1756 | int zid; | |
1757 | ||
3db65812 | 1758 | list_splice(&pages_skipped, src); |
7cc30fcf MG |
1759 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1760 | if (!nr_skipped[zid]) | |
1761 | continue; | |
1762 | ||
1763 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 1764 | skipped += nr_skipped[zid]; |
7cc30fcf MG |
1765 | } |
1766 | } | |
791b48b6 | 1767 | *nr_scanned = total_scan; |
1265e3a6 | 1768 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
791b48b6 | 1769 | total_scan, skipped, nr_taken, mode, lru); |
b4536f0c | 1770 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
1771 | return nr_taken; |
1772 | } | |
1773 | ||
62695a84 NP |
1774 | /** |
1775 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1776 | * @page: page to isolate from its LRU list | |
1777 | * | |
1778 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1779 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1780 | * | |
1781 | * Returns 0 if the page was removed from an LRU list. | |
1782 | * Returns -EBUSY if the page was not on an LRU list. | |
1783 | * | |
1784 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1785 | * the active list, it will have PageActive set. If it was found on |
1786 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1787 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1788 | * |
1789 | * The vmstat statistic corresponding to the list on which the page was | |
1790 | * found will be decremented. | |
1791 | * | |
1792 | * Restrictions: | |
a5d09bed | 1793 | * |
62695a84 NP |
1794 | * (1) Must be called with an elevated refcount on the page. This is a |
1795 | * fundamentnal difference from isolate_lru_pages (which is called | |
1796 | * without a stable reference). | |
1797 | * (2) the lru_lock must not be held. | |
1798 | * (3) interrupts must be enabled. | |
1799 | */ | |
1800 | int isolate_lru_page(struct page *page) | |
1801 | { | |
1802 | int ret = -EBUSY; | |
1803 | ||
309381fe | 1804 | VM_BUG_ON_PAGE(!page_count(page), page); |
cf2a82ee | 1805 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
0c917313 | 1806 | |
62695a84 | 1807 | if (PageLRU(page)) { |
f4b7e272 | 1808 | pg_data_t *pgdat = page_pgdat(page); |
fa9add64 | 1809 | struct lruvec *lruvec; |
62695a84 | 1810 | |
f4b7e272 AR |
1811 | spin_lock_irq(&pgdat->lru_lock); |
1812 | lruvec = mem_cgroup_page_lruvec(page, pgdat); | |
0c917313 | 1813 | if (PageLRU(page)) { |
894bc310 | 1814 | int lru = page_lru(page); |
0c917313 | 1815 | get_page(page); |
62695a84 | 1816 | ClearPageLRU(page); |
fa9add64 HD |
1817 | del_page_from_lru_list(page, lruvec, lru); |
1818 | ret = 0; | |
62695a84 | 1819 | } |
f4b7e272 | 1820 | spin_unlock_irq(&pgdat->lru_lock); |
62695a84 NP |
1821 | } |
1822 | return ret; | |
1823 | } | |
1824 | ||
35cd7815 | 1825 | /* |
d37dd5dc FW |
1826 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1827 | * then get resheduled. When there are massive number of tasks doing page | |
1828 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1829 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1830 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 1831 | */ |
599d0c95 | 1832 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
1833 | struct scan_control *sc) |
1834 | { | |
1835 | unsigned long inactive, isolated; | |
1836 | ||
1837 | if (current_is_kswapd()) | |
1838 | return 0; | |
1839 | ||
97c9341f | 1840 | if (!sane_reclaim(sc)) |
35cd7815 RR |
1841 | return 0; |
1842 | ||
1843 | if (file) { | |
599d0c95 MG |
1844 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1845 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 1846 | } else { |
599d0c95 MG |
1847 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1848 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
1849 | } |
1850 | ||
3cf23841 FW |
1851 | /* |
1852 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1853 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1854 | * deadlock. | |
1855 | */ | |
d0164adc | 1856 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1857 | inactive >>= 3; |
1858 | ||
35cd7815 RR |
1859 | return isolated > inactive; |
1860 | } | |
1861 | ||
a222f341 KT |
1862 | /* |
1863 | * This moves pages from @list to corresponding LRU list. | |
1864 | * | |
1865 | * We move them the other way if the page is referenced by one or more | |
1866 | * processes, from rmap. | |
1867 | * | |
1868 | * If the pages are mostly unmapped, the processing is fast and it is | |
1869 | * appropriate to hold zone_lru_lock across the whole operation. But if | |
1870 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1871 | * should drop zone_lru_lock around each page. It's impossible to balance | |
1872 | * this, so instead we remove the pages from the LRU while processing them. | |
1873 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1874 | * nobody will play with that bit on a non-LRU page. | |
1875 | * | |
1876 | * The downside is that we have to touch page->_refcount against each page. | |
1877 | * But we had to alter page->flags anyway. | |
1878 | * | |
1879 | * Returns the number of pages moved to the given lruvec. | |
1880 | */ | |
1881 | ||
1882 | static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec, | |
1883 | struct list_head *list) | |
66635629 | 1884 | { |
599d0c95 | 1885 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
a222f341 | 1886 | int nr_pages, nr_moved = 0; |
3f79768f | 1887 | LIST_HEAD(pages_to_free); |
a222f341 KT |
1888 | struct page *page; |
1889 | enum lru_list lru; | |
66635629 | 1890 | |
a222f341 KT |
1891 | while (!list_empty(list)) { |
1892 | page = lru_to_page(list); | |
309381fe | 1893 | VM_BUG_ON_PAGE(PageLRU(page), page); |
39b5f29a | 1894 | if (unlikely(!page_evictable(page))) { |
a222f341 | 1895 | list_del(&page->lru); |
599d0c95 | 1896 | spin_unlock_irq(&pgdat->lru_lock); |
66635629 | 1897 | putback_lru_page(page); |
599d0c95 | 1898 | spin_lock_irq(&pgdat->lru_lock); |
66635629 MG |
1899 | continue; |
1900 | } | |
599d0c95 | 1901 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 1902 | |
7a608572 | 1903 | SetPageLRU(page); |
66635629 | 1904 | lru = page_lru(page); |
a222f341 KT |
1905 | |
1906 | nr_pages = hpage_nr_pages(page); | |
1907 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); | |
1908 | list_move(&page->lru, &lruvec->lists[lru]); | |
fa9add64 | 1909 | |
2bcf8879 HD |
1910 | if (put_page_testzero(page)) { |
1911 | __ClearPageLRU(page); | |
1912 | __ClearPageActive(page); | |
fa9add64 | 1913 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1914 | |
1915 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1916 | spin_unlock_irq(&pgdat->lru_lock); |
747db954 | 1917 | mem_cgroup_uncharge(page); |
2bcf8879 | 1918 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1919 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1920 | } else |
1921 | list_add(&page->lru, &pages_to_free); | |
a222f341 KT |
1922 | } else { |
1923 | nr_moved += nr_pages; | |
66635629 MG |
1924 | } |
1925 | } | |
66635629 | 1926 | |
3f79768f HD |
1927 | /* |
1928 | * To save our caller's stack, now use input list for pages to free. | |
1929 | */ | |
a222f341 KT |
1930 | list_splice(&pages_to_free, list); |
1931 | ||
1932 | return nr_moved; | |
66635629 MG |
1933 | } |
1934 | ||
399ba0b9 N |
1935 | /* |
1936 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1937 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1938 | * In that case we should only throttle if the backing device it is | |
1939 | * writing to is congested. In other cases it is safe to throttle. | |
1940 | */ | |
1941 | static int current_may_throttle(void) | |
1942 | { | |
1943 | return !(current->flags & PF_LESS_THROTTLE) || | |
1944 | current->backing_dev_info == NULL || | |
1945 | bdi_write_congested(current->backing_dev_info); | |
1946 | } | |
1947 | ||
1da177e4 | 1948 | /* |
b2e18757 | 1949 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 1950 | * of reclaimed pages |
1da177e4 | 1951 | */ |
66635629 | 1952 | static noinline_for_stack unsigned long |
1a93be0e | 1953 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1954 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1955 | { |
1956 | LIST_HEAD(page_list); | |
e247dbce | 1957 | unsigned long nr_scanned; |
05ff5137 | 1958 | unsigned long nr_reclaimed = 0; |
e247dbce | 1959 | unsigned long nr_taken; |
060f005f | 1960 | struct reclaim_stat stat; |
3cb99451 | 1961 | int file = is_file_lru(lru); |
f46b7912 | 1962 | enum vm_event_item item; |
599d0c95 | 1963 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1a93be0e | 1964 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
db73ee0d | 1965 | bool stalled = false; |
78dc583d | 1966 | |
599d0c95 | 1967 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
db73ee0d MH |
1968 | if (stalled) |
1969 | return 0; | |
1970 | ||
1971 | /* wait a bit for the reclaimer. */ | |
1972 | msleep(100); | |
1973 | stalled = true; | |
35cd7815 RR |
1974 | |
1975 | /* We are about to die and free our memory. Return now. */ | |
1976 | if (fatal_signal_pending(current)) | |
1977 | return SWAP_CLUSTER_MAX; | |
1978 | } | |
1979 | ||
1da177e4 | 1980 | lru_add_drain(); |
f80c0673 | 1981 | |
599d0c95 | 1982 | spin_lock_irq(&pgdat->lru_lock); |
b35ea17b | 1983 | |
5dc35979 | 1984 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
a9e7c39f | 1985 | &nr_scanned, sc, lru); |
95d918fc | 1986 | |
599d0c95 | 1987 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
9d5e6a9f | 1988 | reclaim_stat->recent_scanned[file] += nr_taken; |
95d918fc | 1989 | |
f46b7912 KT |
1990 | item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT; |
1991 | if (global_reclaim(sc)) | |
1992 | __count_vm_events(item, nr_scanned); | |
1993 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned); | |
599d0c95 | 1994 | spin_unlock_irq(&pgdat->lru_lock); |
b35ea17b | 1995 | |
d563c050 | 1996 | if (nr_taken == 0) |
66635629 | 1997 | return 0; |
5ad333eb | 1998 | |
a128ca71 | 1999 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0, |
3c710c1a | 2000 | &stat, false); |
c661b078 | 2001 | |
599d0c95 | 2002 | spin_lock_irq(&pgdat->lru_lock); |
3f79768f | 2003 | |
f46b7912 KT |
2004 | item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT; |
2005 | if (global_reclaim(sc)) | |
2006 | __count_vm_events(item, nr_reclaimed); | |
2007 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed); | |
b17f18af KT |
2008 | reclaim_stat->recent_rotated[0] += stat.nr_activate[0]; |
2009 | reclaim_stat->recent_rotated[1] += stat.nr_activate[1]; | |
a74609fa | 2010 | |
a222f341 | 2011 | move_pages_to_lru(lruvec, &page_list); |
3f79768f | 2012 | |
599d0c95 | 2013 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f | 2014 | |
599d0c95 | 2015 | spin_unlock_irq(&pgdat->lru_lock); |
3f79768f | 2016 | |
747db954 | 2017 | mem_cgroup_uncharge_list(&page_list); |
2d4894b5 | 2018 | free_unref_page_list(&page_list); |
e11da5b4 | 2019 | |
1c610d5f AR |
2020 | /* |
2021 | * If dirty pages are scanned that are not queued for IO, it | |
2022 | * implies that flushers are not doing their job. This can | |
2023 | * happen when memory pressure pushes dirty pages to the end of | |
2024 | * the LRU before the dirty limits are breached and the dirty | |
2025 | * data has expired. It can also happen when the proportion of | |
2026 | * dirty pages grows not through writes but through memory | |
2027 | * pressure reclaiming all the clean cache. And in some cases, | |
2028 | * the flushers simply cannot keep up with the allocation | |
2029 | * rate. Nudge the flusher threads in case they are asleep. | |
2030 | */ | |
2031 | if (stat.nr_unqueued_dirty == nr_taken) | |
2032 | wakeup_flusher_threads(WB_REASON_VMSCAN); | |
2033 | ||
d108c772 AR |
2034 | sc->nr.dirty += stat.nr_dirty; |
2035 | sc->nr.congested += stat.nr_congested; | |
2036 | sc->nr.unqueued_dirty += stat.nr_unqueued_dirty; | |
2037 | sc->nr.writeback += stat.nr_writeback; | |
2038 | sc->nr.immediate += stat.nr_immediate; | |
2039 | sc->nr.taken += nr_taken; | |
2040 | if (file) | |
2041 | sc->nr.file_taken += nr_taken; | |
8e950282 | 2042 | |
599d0c95 | 2043 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
d51d1e64 | 2044 | nr_scanned, nr_reclaimed, &stat, sc->priority, file); |
05ff5137 | 2045 | return nr_reclaimed; |
1da177e4 LT |
2046 | } |
2047 | ||
f626012d | 2048 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 2049 | struct lruvec *lruvec, |
f16015fb | 2050 | struct scan_control *sc, |
9e3b2f8c | 2051 | enum lru_list lru) |
1da177e4 | 2052 | { |
44c241f1 | 2053 | unsigned long nr_taken; |
f626012d | 2054 | unsigned long nr_scanned; |
6fe6b7e3 | 2055 | unsigned long vm_flags; |
1da177e4 | 2056 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 2057 | LIST_HEAD(l_active); |
b69408e8 | 2058 | LIST_HEAD(l_inactive); |
1da177e4 | 2059 | struct page *page; |
1a93be0e | 2060 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
9d998b4f MH |
2061 | unsigned nr_deactivate, nr_activate; |
2062 | unsigned nr_rotated = 0; | |
3cb99451 | 2063 | int file = is_file_lru(lru); |
599d0c95 | 2064 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
2065 | |
2066 | lru_add_drain(); | |
f80c0673 | 2067 | |
599d0c95 | 2068 | spin_lock_irq(&pgdat->lru_lock); |
925b7673 | 2069 | |
5dc35979 | 2070 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
a9e7c39f | 2071 | &nr_scanned, sc, lru); |
89b5fae5 | 2072 | |
599d0c95 | 2073 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
b7c46d15 | 2074 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 2075 | |
599d0c95 | 2076 | __count_vm_events(PGREFILL, nr_scanned); |
2fa2690c | 2077 | __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); |
9d5e6a9f | 2078 | |
599d0c95 | 2079 | spin_unlock_irq(&pgdat->lru_lock); |
1da177e4 | 2080 | |
1da177e4 LT |
2081 | while (!list_empty(&l_hold)) { |
2082 | cond_resched(); | |
2083 | page = lru_to_page(&l_hold); | |
2084 | list_del(&page->lru); | |
7e9cd484 | 2085 | |
39b5f29a | 2086 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
2087 | putback_lru_page(page); |
2088 | continue; | |
2089 | } | |
2090 | ||
cc715d99 MG |
2091 | if (unlikely(buffer_heads_over_limit)) { |
2092 | if (page_has_private(page) && trylock_page(page)) { | |
2093 | if (page_has_private(page)) | |
2094 | try_to_release_page(page, 0); | |
2095 | unlock_page(page); | |
2096 | } | |
2097 | } | |
2098 | ||
c3ac9a8a JW |
2099 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
2100 | &vm_flags)) { | |
9992af10 | 2101 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
2102 | /* |
2103 | * Identify referenced, file-backed active pages and | |
2104 | * give them one more trip around the active list. So | |
2105 | * that executable code get better chances to stay in | |
2106 | * memory under moderate memory pressure. Anon pages | |
2107 | * are not likely to be evicted by use-once streaming | |
2108 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
2109 | * so we ignore them here. | |
2110 | */ | |
41e20983 | 2111 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
2112 | list_add(&page->lru, &l_active); |
2113 | continue; | |
2114 | } | |
2115 | } | |
7e9cd484 | 2116 | |
5205e56e | 2117 | ClearPageActive(page); /* we are de-activating */ |
1899ad18 | 2118 | SetPageWorkingset(page); |
1da177e4 LT |
2119 | list_add(&page->lru, &l_inactive); |
2120 | } | |
2121 | ||
b555749a | 2122 | /* |
8cab4754 | 2123 | * Move pages back to the lru list. |
b555749a | 2124 | */ |
599d0c95 | 2125 | spin_lock_irq(&pgdat->lru_lock); |
556adecb | 2126 | /* |
8cab4754 WF |
2127 | * Count referenced pages from currently used mappings as rotated, |
2128 | * even though only some of them are actually re-activated. This | |
2129 | * helps balance scan pressure between file and anonymous pages in | |
7c0db9e9 | 2130 | * get_scan_count. |
7e9cd484 | 2131 | */ |
b7c46d15 | 2132 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 2133 | |
a222f341 KT |
2134 | nr_activate = move_pages_to_lru(lruvec, &l_active); |
2135 | nr_deactivate = move_pages_to_lru(lruvec, &l_inactive); | |
f372d89e KT |
2136 | /* Keep all free pages in l_active list */ |
2137 | list_splice(&l_inactive, &l_active); | |
9851ac13 KT |
2138 | |
2139 | __count_vm_events(PGDEACTIVATE, nr_deactivate); | |
2140 | __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate); | |
2141 | ||
599d0c95 MG |
2142 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
2143 | spin_unlock_irq(&pgdat->lru_lock); | |
2bcf8879 | 2144 | |
f372d89e KT |
2145 | mem_cgroup_uncharge_list(&l_active); |
2146 | free_unref_page_list(&l_active); | |
9d998b4f MH |
2147 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
2148 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
2149 | } |
2150 | ||
59dc76b0 RR |
2151 | /* |
2152 | * The inactive anon list should be small enough that the VM never has | |
2153 | * to do too much work. | |
14797e23 | 2154 | * |
59dc76b0 RR |
2155 | * The inactive file list should be small enough to leave most memory |
2156 | * to the established workingset on the scan-resistant active list, | |
2157 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2158 | * |
59dc76b0 RR |
2159 | * Both inactive lists should also be large enough that each inactive |
2160 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 2161 | * |
2a2e4885 JW |
2162 | * If that fails and refaulting is observed, the inactive list grows. |
2163 | * | |
59dc76b0 | 2164 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
3a50d14d | 2165 | * on this LRU, maintained by the pageout code. An inactive_ratio |
59dc76b0 | 2166 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. |
56e49d21 | 2167 | * |
59dc76b0 RR |
2168 | * total target max |
2169 | * memory ratio inactive | |
2170 | * ------------------------------------- | |
2171 | * 10MB 1 5MB | |
2172 | * 100MB 1 50MB | |
2173 | * 1GB 3 250MB | |
2174 | * 10GB 10 0.9GB | |
2175 | * 100GB 31 3GB | |
2176 | * 1TB 101 10GB | |
2177 | * 10TB 320 32GB | |
56e49d21 | 2178 | */ |
f8d1a311 | 2179 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file, |
2c012a4a | 2180 | struct scan_control *sc, bool trace) |
56e49d21 | 2181 | { |
fd538803 | 2182 | enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; |
2a2e4885 JW |
2183 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
2184 | enum lru_list inactive_lru = file * LRU_FILE; | |
2185 | unsigned long inactive, active; | |
2186 | unsigned long inactive_ratio; | |
2187 | unsigned long refaults; | |
59dc76b0 | 2188 | unsigned long gb; |
e3790144 | 2189 | |
59dc76b0 RR |
2190 | /* |
2191 | * If we don't have swap space, anonymous page deactivation | |
2192 | * is pointless. | |
2193 | */ | |
2194 | if (!file && !total_swap_pages) | |
2195 | return false; | |
56e49d21 | 2196 | |
fd538803 MH |
2197 | inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); |
2198 | active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); | |
f8d1a311 | 2199 | |
2a2e4885 JW |
2200 | /* |
2201 | * When refaults are being observed, it means a new workingset | |
2202 | * is being established. Disable active list protection to get | |
2203 | * rid of the stale workingset quickly. | |
2204 | */ | |
205b20cc | 2205 | refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE); |
2c012a4a | 2206 | if (file && lruvec->refaults != refaults) { |
2a2e4885 JW |
2207 | inactive_ratio = 0; |
2208 | } else { | |
2209 | gb = (inactive + active) >> (30 - PAGE_SHIFT); | |
2210 | if (gb) | |
2211 | inactive_ratio = int_sqrt(10 * gb); | |
2212 | else | |
2213 | inactive_ratio = 1; | |
2214 | } | |
59dc76b0 | 2215 | |
2c012a4a | 2216 | if (trace) |
2a2e4885 JW |
2217 | trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx, |
2218 | lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive, | |
2219 | lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active, | |
2220 | inactive_ratio, file); | |
fd538803 | 2221 | |
59dc76b0 | 2222 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2223 | } |
2224 | ||
4f98a2fe | 2225 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
3b991208 | 2226 | struct lruvec *lruvec, struct scan_control *sc) |
b69408e8 | 2227 | { |
b39415b2 | 2228 | if (is_active_lru(lru)) { |
3b991208 | 2229 | if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true)) |
1a93be0e | 2230 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
2231 | return 0; |
2232 | } | |
2233 | ||
1a93be0e | 2234 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
2235 | } |
2236 | ||
9a265114 JW |
2237 | enum scan_balance { |
2238 | SCAN_EQUAL, | |
2239 | SCAN_FRACT, | |
2240 | SCAN_ANON, | |
2241 | SCAN_FILE, | |
2242 | }; | |
2243 | ||
4f98a2fe RR |
2244 | /* |
2245 | * Determine how aggressively the anon and file LRU lists should be | |
2246 | * scanned. The relative value of each set of LRU lists is determined | |
2247 | * by looking at the fraction of the pages scanned we did rotate back | |
2248 | * onto the active list instead of evict. | |
2249 | * | |
be7bd59d WL |
2250 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2251 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 2252 | */ |
33377678 | 2253 | static void get_scan_count(struct lruvec *lruvec, struct mem_cgroup *memcg, |
6b4f7799 JW |
2254 | struct scan_control *sc, unsigned long *nr, |
2255 | unsigned long *lru_pages) | |
4f98a2fe | 2256 | { |
33377678 | 2257 | int swappiness = mem_cgroup_swappiness(memcg); |
9a265114 JW |
2258 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
2259 | u64 fraction[2]; | |
2260 | u64 denominator = 0; /* gcc */ | |
599d0c95 | 2261 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
4f98a2fe | 2262 | unsigned long anon_prio, file_prio; |
9a265114 | 2263 | enum scan_balance scan_balance; |
0bf1457f | 2264 | unsigned long anon, file; |
4f98a2fe | 2265 | unsigned long ap, fp; |
4111304d | 2266 | enum lru_list lru; |
76a33fc3 SL |
2267 | |
2268 | /* If we have no swap space, do not bother scanning anon pages. */ | |
d8b38438 | 2269 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
9a265114 | 2270 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2271 | goto out; |
2272 | } | |
4f98a2fe | 2273 | |
10316b31 JW |
2274 | /* |
2275 | * Global reclaim will swap to prevent OOM even with no | |
2276 | * swappiness, but memcg users want to use this knob to | |
2277 | * disable swapping for individual groups completely when | |
2278 | * using the memory controller's swap limit feature would be | |
2279 | * too expensive. | |
2280 | */ | |
02695175 | 2281 | if (!global_reclaim(sc) && !swappiness) { |
9a265114 | 2282 | scan_balance = SCAN_FILE; |
10316b31 JW |
2283 | goto out; |
2284 | } | |
2285 | ||
2286 | /* | |
2287 | * Do not apply any pressure balancing cleverness when the | |
2288 | * system is close to OOM, scan both anon and file equally | |
2289 | * (unless the swappiness setting disagrees with swapping). | |
2290 | */ | |
02695175 | 2291 | if (!sc->priority && swappiness) { |
9a265114 | 2292 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2293 | goto out; |
2294 | } | |
2295 | ||
62376251 JW |
2296 | /* |
2297 | * Prevent the reclaimer from falling into the cache trap: as | |
2298 | * cache pages start out inactive, every cache fault will tip | |
2299 | * the scan balance towards the file LRU. And as the file LRU | |
2300 | * shrinks, so does the window for rotation from references. | |
2301 | * This means we have a runaway feedback loop where a tiny | |
2302 | * thrashing file LRU becomes infinitely more attractive than | |
2303 | * anon pages. Try to detect this based on file LRU size. | |
2304 | */ | |
2305 | if (global_reclaim(sc)) { | |
599d0c95 MG |
2306 | unsigned long pgdatfile; |
2307 | unsigned long pgdatfree; | |
2308 | int z; | |
2309 | unsigned long total_high_wmark = 0; | |
2ab051e1 | 2310 | |
599d0c95 MG |
2311 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
2312 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2313 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2314 | ||
2315 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2316 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2317 | if (!managed_zone(zone)) |
599d0c95 MG |
2318 | continue; |
2319 | ||
2320 | total_high_wmark += high_wmark_pages(zone); | |
2321 | } | |
62376251 | 2322 | |
599d0c95 | 2323 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
06226226 DR |
2324 | /* |
2325 | * Force SCAN_ANON if there are enough inactive | |
2326 | * anonymous pages on the LRU in eligible zones. | |
2327 | * Otherwise, the small LRU gets thrashed. | |
2328 | */ | |
3b991208 | 2329 | if (!inactive_list_is_low(lruvec, false, sc, false) && |
06226226 DR |
2330 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx) |
2331 | >> sc->priority) { | |
2332 | scan_balance = SCAN_ANON; | |
2333 | goto out; | |
2334 | } | |
62376251 JW |
2335 | } |
2336 | } | |
2337 | ||
7c5bd705 | 2338 | /* |
316bda0e VD |
2339 | * If there is enough inactive page cache, i.e. if the size of the |
2340 | * inactive list is greater than that of the active list *and* the | |
2341 | * inactive list actually has some pages to scan on this priority, we | |
2342 | * do not reclaim anything from the anonymous working set right now. | |
2343 | * Without the second condition we could end up never scanning an | |
2344 | * lruvec even if it has plenty of old anonymous pages unless the | |
2345 | * system is under heavy pressure. | |
7c5bd705 | 2346 | */ |
3b991208 | 2347 | if (!inactive_list_is_low(lruvec, true, sc, false) && |
71ab6cfe | 2348 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { |
9a265114 | 2349 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2350 | goto out; |
2351 | } | |
2352 | ||
9a265114 JW |
2353 | scan_balance = SCAN_FRACT; |
2354 | ||
58c37f6e KM |
2355 | /* |
2356 | * With swappiness at 100, anonymous and file have the same priority. | |
2357 | * This scanning priority is essentially the inverse of IO cost. | |
2358 | */ | |
02695175 | 2359 | anon_prio = swappiness; |
75b00af7 | 2360 | file_prio = 200 - anon_prio; |
58c37f6e | 2361 | |
4f98a2fe RR |
2362 | /* |
2363 | * OK, so we have swap space and a fair amount of page cache | |
2364 | * pages. We use the recently rotated / recently scanned | |
2365 | * ratios to determine how valuable each cache is. | |
2366 | * | |
2367 | * Because workloads change over time (and to avoid overflow) | |
2368 | * we keep these statistics as a floating average, which ends | |
2369 | * up weighing recent references more than old ones. | |
2370 | * | |
2371 | * anon in [0], file in [1] | |
2372 | */ | |
2ab051e1 | 2373 | |
fd538803 MH |
2374 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + |
2375 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); | |
2376 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + | |
2377 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); | |
2ab051e1 | 2378 | |
599d0c95 | 2379 | spin_lock_irq(&pgdat->lru_lock); |
6e901571 | 2380 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
2381 | reclaim_stat->recent_scanned[0] /= 2; |
2382 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
2383 | } |
2384 | ||
6e901571 | 2385 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
2386 | reclaim_stat->recent_scanned[1] /= 2; |
2387 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
2388 | } |
2389 | ||
4f98a2fe | 2390 | /* |
00d8089c RR |
2391 | * The amount of pressure on anon vs file pages is inversely |
2392 | * proportional to the fraction of recently scanned pages on | |
2393 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2394 | */ |
fe35004f | 2395 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 2396 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2397 | |
fe35004f | 2398 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2399 | fp /= reclaim_stat->recent_rotated[1] + 1; |
599d0c95 | 2400 | spin_unlock_irq(&pgdat->lru_lock); |
4f98a2fe | 2401 | |
76a33fc3 SL |
2402 | fraction[0] = ap; |
2403 | fraction[1] = fp; | |
2404 | denominator = ap + fp + 1; | |
2405 | out: | |
688035f7 JW |
2406 | *lru_pages = 0; |
2407 | for_each_evictable_lru(lru) { | |
2408 | int file = is_file_lru(lru); | |
2409 | unsigned long size; | |
2410 | unsigned long scan; | |
6b4f7799 | 2411 | |
688035f7 JW |
2412 | size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); |
2413 | scan = size >> sc->priority; | |
2414 | /* | |
2415 | * If the cgroup's already been deleted, make sure to | |
2416 | * scrape out the remaining cache. | |
2417 | */ | |
2418 | if (!scan && !mem_cgroup_online(memcg)) | |
2419 | scan = min(size, SWAP_CLUSTER_MAX); | |
6b4f7799 | 2420 | |
688035f7 JW |
2421 | switch (scan_balance) { |
2422 | case SCAN_EQUAL: | |
2423 | /* Scan lists relative to size */ | |
2424 | break; | |
2425 | case SCAN_FRACT: | |
9a265114 | 2426 | /* |
688035f7 JW |
2427 | * Scan types proportional to swappiness and |
2428 | * their relative recent reclaim efficiency. | |
68600f62 RG |
2429 | * Make sure we don't miss the last page |
2430 | * because of a round-off error. | |
9a265114 | 2431 | */ |
68600f62 RG |
2432 | scan = DIV64_U64_ROUND_UP(scan * fraction[file], |
2433 | denominator); | |
688035f7 JW |
2434 | break; |
2435 | case SCAN_FILE: | |
2436 | case SCAN_ANON: | |
2437 | /* Scan one type exclusively */ | |
2438 | if ((scan_balance == SCAN_FILE) != file) { | |
2439 | size = 0; | |
2440 | scan = 0; | |
2441 | } | |
2442 | break; | |
2443 | default: | |
2444 | /* Look ma, no brain */ | |
2445 | BUG(); | |
9a265114 | 2446 | } |
688035f7 JW |
2447 | |
2448 | *lru_pages += size; | |
2449 | nr[lru] = scan; | |
76a33fc3 | 2450 | } |
6e08a369 | 2451 | } |
4f98a2fe | 2452 | |
9b4f98cd | 2453 | /* |
a9dd0a83 | 2454 | * This is a basic per-node page freer. Used by both kswapd and direct reclaim. |
9b4f98cd | 2455 | */ |
a9dd0a83 | 2456 | static void shrink_node_memcg(struct pglist_data *pgdat, struct mem_cgroup *memcg, |
33377678 | 2457 | struct scan_control *sc, unsigned long *lru_pages) |
9b4f98cd | 2458 | { |
ef8f2327 | 2459 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
9b4f98cd | 2460 | unsigned long nr[NR_LRU_LISTS]; |
e82e0561 | 2461 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2462 | unsigned long nr_to_scan; |
2463 | enum lru_list lru; | |
2464 | unsigned long nr_reclaimed = 0; | |
2465 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2466 | struct blk_plug plug; | |
1a501907 | 2467 | bool scan_adjusted; |
9b4f98cd | 2468 | |
33377678 | 2469 | get_scan_count(lruvec, memcg, sc, nr, lru_pages); |
9b4f98cd | 2470 | |
e82e0561 MG |
2471 | /* Record the original scan target for proportional adjustments later */ |
2472 | memcpy(targets, nr, sizeof(nr)); | |
2473 | ||
1a501907 MG |
2474 | /* |
2475 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2476 | * event that can occur when there is little memory pressure e.g. | |
2477 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2478 | * when the requested number of pages are reclaimed when scanning at | |
2479 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2480 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2481 | * do a batch of work at once. For memcg reclaim one check is made to | |
2482 | * abort proportional reclaim if either the file or anon lru has already | |
2483 | * dropped to zero at the first pass. | |
2484 | */ | |
2485 | scan_adjusted = (global_reclaim(sc) && !current_is_kswapd() && | |
2486 | sc->priority == DEF_PRIORITY); | |
2487 | ||
9b4f98cd JW |
2488 | blk_start_plug(&plug); |
2489 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2490 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2491 | unsigned long nr_anon, nr_file, percentage; |
2492 | unsigned long nr_scanned; | |
2493 | ||
9b4f98cd JW |
2494 | for_each_evictable_lru(lru) { |
2495 | if (nr[lru]) { | |
2496 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2497 | nr[lru] -= nr_to_scan; | |
2498 | ||
2499 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
3b991208 | 2500 | lruvec, sc); |
9b4f98cd JW |
2501 | } |
2502 | } | |
e82e0561 | 2503 | |
bd041733 MH |
2504 | cond_resched(); |
2505 | ||
e82e0561 MG |
2506 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
2507 | continue; | |
2508 | ||
e82e0561 MG |
2509 | /* |
2510 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2511 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2512 | * proportionally what was requested by get_scan_count(). We |
2513 | * stop reclaiming one LRU and reduce the amount scanning | |
2514 | * proportional to the original scan target. | |
2515 | */ | |
2516 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2517 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2518 | ||
1a501907 MG |
2519 | /* |
2520 | * It's just vindictive to attack the larger once the smaller | |
2521 | * has gone to zero. And given the way we stop scanning the | |
2522 | * smaller below, this makes sure that we only make one nudge | |
2523 | * towards proportionality once we've got nr_to_reclaim. | |
2524 | */ | |
2525 | if (!nr_file || !nr_anon) | |
2526 | break; | |
2527 | ||
e82e0561 MG |
2528 | if (nr_file > nr_anon) { |
2529 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2530 | targets[LRU_ACTIVE_ANON] + 1; | |
2531 | lru = LRU_BASE; | |
2532 | percentage = nr_anon * 100 / scan_target; | |
2533 | } else { | |
2534 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2535 | targets[LRU_ACTIVE_FILE] + 1; | |
2536 | lru = LRU_FILE; | |
2537 | percentage = nr_file * 100 / scan_target; | |
2538 | } | |
2539 | ||
2540 | /* Stop scanning the smaller of the LRU */ | |
2541 | nr[lru] = 0; | |
2542 | nr[lru + LRU_ACTIVE] = 0; | |
2543 | ||
2544 | /* | |
2545 | * Recalculate the other LRU scan count based on its original | |
2546 | * scan target and the percentage scanning already complete | |
2547 | */ | |
2548 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2549 | nr_scanned = targets[lru] - nr[lru]; | |
2550 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2551 | nr[lru] -= min(nr[lru], nr_scanned); | |
2552 | ||
2553 | lru += LRU_ACTIVE; | |
2554 | nr_scanned = targets[lru] - nr[lru]; | |
2555 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2556 | nr[lru] -= min(nr[lru], nr_scanned); | |
2557 | ||
2558 | scan_adjusted = true; | |
9b4f98cd JW |
2559 | } |
2560 | blk_finish_plug(&plug); | |
2561 | sc->nr_reclaimed += nr_reclaimed; | |
2562 | ||
2563 | /* | |
2564 | * Even if we did not try to evict anon pages at all, we want to | |
2565 | * rebalance the anon lru active/inactive ratio. | |
2566 | */ | |
3b991208 | 2567 | if (inactive_list_is_low(lruvec, false, sc, true)) |
9b4f98cd JW |
2568 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2569 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
2570 | } |
2571 | ||
23b9da55 | 2572 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2573 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2574 | { |
d84da3f9 | 2575 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2576 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2577 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2578 | return true; |
2579 | ||
2580 | return false; | |
2581 | } | |
2582 | ||
3e7d3449 | 2583 | /* |
23b9da55 MG |
2584 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2585 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2586 | * true if more pages should be reclaimed such that when the page allocator | |
2587 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2588 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2589 | */ |
a9dd0a83 | 2590 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 MG |
2591 | unsigned long nr_reclaimed, |
2592 | unsigned long nr_scanned, | |
2593 | struct scan_control *sc) | |
2594 | { | |
2595 | unsigned long pages_for_compaction; | |
2596 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 2597 | int z; |
3e7d3449 MG |
2598 | |
2599 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2600 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2601 | return false; |
2602 | ||
2876592f | 2603 | /* Consider stopping depending on scan and reclaim activity */ |
dcda9b04 | 2604 | if (sc->gfp_mask & __GFP_RETRY_MAYFAIL) { |
2876592f | 2605 | /* |
dcda9b04 | 2606 | * For __GFP_RETRY_MAYFAIL allocations, stop reclaiming if the |
2876592f MG |
2607 | * full LRU list has been scanned and we are still failing |
2608 | * to reclaim pages. This full LRU scan is potentially | |
dcda9b04 | 2609 | * expensive but a __GFP_RETRY_MAYFAIL caller really wants to succeed |
2876592f MG |
2610 | */ |
2611 | if (!nr_reclaimed && !nr_scanned) | |
2612 | return false; | |
2613 | } else { | |
2614 | /* | |
dcda9b04 | 2615 | * For non-__GFP_RETRY_MAYFAIL allocations which can presumably |
2876592f MG |
2616 | * fail without consequence, stop if we failed to reclaim |
2617 | * any pages from the last SWAP_CLUSTER_MAX number of | |
2618 | * pages that were scanned. This will return to the | |
2619 | * caller faster at the risk reclaim/compaction and | |
2620 | * the resulting allocation attempt fails | |
2621 | */ | |
2622 | if (!nr_reclaimed) | |
2623 | return false; | |
2624 | } | |
3e7d3449 MG |
2625 | |
2626 | /* | |
2627 | * If we have not reclaimed enough pages for compaction and the | |
2628 | * inactive lists are large enough, continue reclaiming | |
2629 | */ | |
9861a62c | 2630 | pages_for_compaction = compact_gap(sc->order); |
a9dd0a83 | 2631 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); |
ec8acf20 | 2632 | if (get_nr_swap_pages() > 0) |
a9dd0a83 | 2633 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
3e7d3449 MG |
2634 | if (sc->nr_reclaimed < pages_for_compaction && |
2635 | inactive_lru_pages > pages_for_compaction) | |
2636 | return true; | |
2637 | ||
2638 | /* If compaction would go ahead or the allocation would succeed, stop */ | |
a9dd0a83 MG |
2639 | for (z = 0; z <= sc->reclaim_idx; z++) { |
2640 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2641 | if (!managed_zone(zone)) |
a9dd0a83 MG |
2642 | continue; |
2643 | ||
2644 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 2645 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
2646 | case COMPACT_CONTINUE: |
2647 | return false; | |
2648 | default: | |
2649 | /* check next zone */ | |
2650 | ; | |
2651 | } | |
3e7d3449 | 2652 | } |
a9dd0a83 | 2653 | return true; |
3e7d3449 MG |
2654 | } |
2655 | ||
e3c1ac58 AR |
2656 | static bool pgdat_memcg_congested(pg_data_t *pgdat, struct mem_cgroup *memcg) |
2657 | { | |
2658 | return test_bit(PGDAT_CONGESTED, &pgdat->flags) || | |
2659 | (memcg && memcg_congested(pgdat, memcg)); | |
2660 | } | |
2661 | ||
970a39a3 | 2662 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 2663 | { |
cb731d6c | 2664 | struct reclaim_state *reclaim_state = current->reclaim_state; |
f0fdc5e8 | 2665 | unsigned long nr_reclaimed, nr_scanned; |
2344d7e4 | 2666 | bool reclaimable = false; |
1da177e4 | 2667 | |
9b4f98cd JW |
2668 | do { |
2669 | struct mem_cgroup *root = sc->target_mem_cgroup; | |
2670 | struct mem_cgroup_reclaim_cookie reclaim = { | |
ef8f2327 | 2671 | .pgdat = pgdat, |
9b4f98cd JW |
2672 | .priority = sc->priority, |
2673 | }; | |
a9dd0a83 | 2674 | unsigned long node_lru_pages = 0; |
694fbc0f | 2675 | struct mem_cgroup *memcg; |
3e7d3449 | 2676 | |
d108c772 AR |
2677 | memset(&sc->nr, 0, sizeof(sc->nr)); |
2678 | ||
9b4f98cd JW |
2679 | nr_reclaimed = sc->nr_reclaimed; |
2680 | nr_scanned = sc->nr_scanned; | |
1da177e4 | 2681 | |
694fbc0f AM |
2682 | memcg = mem_cgroup_iter(root, NULL, &reclaim); |
2683 | do { | |
6b4f7799 | 2684 | unsigned long lru_pages; |
8e8ae645 | 2685 | unsigned long reclaimed; |
cb731d6c | 2686 | unsigned long scanned; |
5660048c | 2687 | |
bf8d5d52 RG |
2688 | switch (mem_cgroup_protected(root, memcg)) { |
2689 | case MEMCG_PROT_MIN: | |
2690 | /* | |
2691 | * Hard protection. | |
2692 | * If there is no reclaimable memory, OOM. | |
2693 | */ | |
2694 | continue; | |
2695 | case MEMCG_PROT_LOW: | |
2696 | /* | |
2697 | * Soft protection. | |
2698 | * Respect the protection only as long as | |
2699 | * there is an unprotected supply | |
2700 | * of reclaimable memory from other cgroups. | |
2701 | */ | |
d6622f63 YX |
2702 | if (!sc->memcg_low_reclaim) { |
2703 | sc->memcg_low_skipped = 1; | |
241994ed | 2704 | continue; |
d6622f63 | 2705 | } |
e27be240 | 2706 | memcg_memory_event(memcg, MEMCG_LOW); |
bf8d5d52 RG |
2707 | break; |
2708 | case MEMCG_PROT_NONE: | |
2709 | break; | |
241994ed JW |
2710 | } |
2711 | ||
8e8ae645 | 2712 | reclaimed = sc->nr_reclaimed; |
cb731d6c | 2713 | scanned = sc->nr_scanned; |
a9dd0a83 MG |
2714 | shrink_node_memcg(pgdat, memcg, sc, &lru_pages); |
2715 | node_lru_pages += lru_pages; | |
f16015fb | 2716 | |
1c30844d MG |
2717 | if (sc->may_shrinkslab) { |
2718 | shrink_slab(sc->gfp_mask, pgdat->node_id, | |
aeed1d32 | 2719 | memcg, sc->priority); |
1c30844d | 2720 | } |
cb731d6c | 2721 | |
8e8ae645 JW |
2722 | /* Record the group's reclaim efficiency */ |
2723 | vmpressure(sc->gfp_mask, memcg, false, | |
2724 | sc->nr_scanned - scanned, | |
2725 | sc->nr_reclaimed - reclaimed); | |
2726 | ||
9b4f98cd | 2727 | /* |
2bb0f34f YS |
2728 | * Kswapd have to scan all memory cgroups to fulfill |
2729 | * the overall scan target for the node. | |
a394cb8e MH |
2730 | * |
2731 | * Limit reclaim, on the other hand, only cares about | |
2732 | * nr_to_reclaim pages to be reclaimed and it will | |
2733 | * retry with decreasing priority if one round over the | |
2734 | * whole hierarchy is not sufficient. | |
9b4f98cd | 2735 | */ |
2bb0f34f | 2736 | if (!current_is_kswapd() && |
a394cb8e | 2737 | sc->nr_reclaimed >= sc->nr_to_reclaim) { |
9b4f98cd JW |
2738 | mem_cgroup_iter_break(root, memcg); |
2739 | break; | |
2740 | } | |
241994ed | 2741 | } while ((memcg = mem_cgroup_iter(root, memcg, &reclaim))); |
70ddf637 | 2742 | |
cb731d6c VD |
2743 | if (reclaim_state) { |
2744 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2745 | reclaim_state->reclaimed_slab = 0; | |
6b4f7799 JW |
2746 | } |
2747 | ||
8e8ae645 JW |
2748 | /* Record the subtree's reclaim efficiency */ |
2749 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
70ddf637 AV |
2750 | sc->nr_scanned - nr_scanned, |
2751 | sc->nr_reclaimed - nr_reclaimed); | |
2752 | ||
2344d7e4 JW |
2753 | if (sc->nr_reclaimed - nr_reclaimed) |
2754 | reclaimable = true; | |
2755 | ||
e3c1ac58 AR |
2756 | if (current_is_kswapd()) { |
2757 | /* | |
2758 | * If reclaim is isolating dirty pages under writeback, | |
2759 | * it implies that the long-lived page allocation rate | |
2760 | * is exceeding the page laundering rate. Either the | |
2761 | * global limits are not being effective at throttling | |
2762 | * processes due to the page distribution throughout | |
2763 | * zones or there is heavy usage of a slow backing | |
2764 | * device. The only option is to throttle from reclaim | |
2765 | * context which is not ideal as there is no guarantee | |
2766 | * the dirtying process is throttled in the same way | |
2767 | * balance_dirty_pages() manages. | |
2768 | * | |
2769 | * Once a node is flagged PGDAT_WRITEBACK, kswapd will | |
2770 | * count the number of pages under pages flagged for | |
2771 | * immediate reclaim and stall if any are encountered | |
2772 | * in the nr_immediate check below. | |
2773 | */ | |
2774 | if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken) | |
2775 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
d108c772 | 2776 | |
d108c772 AR |
2777 | /* |
2778 | * Tag a node as congested if all the dirty pages | |
2779 | * scanned were backed by a congested BDI and | |
2780 | * wait_iff_congested will stall. | |
2781 | */ | |
2782 | if (sc->nr.dirty && sc->nr.dirty == sc->nr.congested) | |
2783 | set_bit(PGDAT_CONGESTED, &pgdat->flags); | |
2784 | ||
2785 | /* Allow kswapd to start writing pages during reclaim.*/ | |
2786 | if (sc->nr.unqueued_dirty == sc->nr.file_taken) | |
2787 | set_bit(PGDAT_DIRTY, &pgdat->flags); | |
2788 | ||
2789 | /* | |
2790 | * If kswapd scans pages marked marked for immediate | |
2791 | * reclaim and under writeback (nr_immediate), it | |
2792 | * implies that pages are cycling through the LRU | |
2793 | * faster than they are written so also forcibly stall. | |
2794 | */ | |
2795 | if (sc->nr.immediate) | |
2796 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
2797 | } | |
2798 | ||
e3c1ac58 AR |
2799 | /* |
2800 | * Legacy memcg will stall in page writeback so avoid forcibly | |
2801 | * stalling in wait_iff_congested(). | |
2802 | */ | |
2803 | if (!global_reclaim(sc) && sane_reclaim(sc) && | |
2804 | sc->nr.dirty && sc->nr.dirty == sc->nr.congested) | |
2805 | set_memcg_congestion(pgdat, root, true); | |
2806 | ||
d108c772 AR |
2807 | /* |
2808 | * Stall direct reclaim for IO completions if underlying BDIs | |
2809 | * and node is congested. Allow kswapd to continue until it | |
2810 | * starts encountering unqueued dirty pages or cycling through | |
2811 | * the LRU too quickly. | |
2812 | */ | |
2813 | if (!sc->hibernation_mode && !current_is_kswapd() && | |
e3c1ac58 AR |
2814 | current_may_throttle() && pgdat_memcg_congested(pgdat, root)) |
2815 | wait_iff_congested(BLK_RW_ASYNC, HZ/10); | |
d108c772 | 2816 | |
a9dd0a83 | 2817 | } while (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
9b4f98cd | 2818 | sc->nr_scanned - nr_scanned, sc)); |
2344d7e4 | 2819 | |
c73322d0 JW |
2820 | /* |
2821 | * Kswapd gives up on balancing particular nodes after too | |
2822 | * many failures to reclaim anything from them and goes to | |
2823 | * sleep. On reclaim progress, reset the failure counter. A | |
2824 | * successful direct reclaim run will revive a dormant kswapd. | |
2825 | */ | |
2826 | if (reclaimable) | |
2827 | pgdat->kswapd_failures = 0; | |
2828 | ||
2344d7e4 | 2829 | return reclaimable; |
f16015fb JW |
2830 | } |
2831 | ||
53853e2d | 2832 | /* |
fdd4c614 VB |
2833 | * Returns true if compaction should go ahead for a costly-order request, or |
2834 | * the allocation would already succeed without compaction. Return false if we | |
2835 | * should reclaim first. | |
53853e2d | 2836 | */ |
4f588331 | 2837 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 2838 | { |
31483b6a | 2839 | unsigned long watermark; |
fdd4c614 | 2840 | enum compact_result suitable; |
fe4b1b24 | 2841 | |
fdd4c614 VB |
2842 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
2843 | if (suitable == COMPACT_SUCCESS) | |
2844 | /* Allocation should succeed already. Don't reclaim. */ | |
2845 | return true; | |
2846 | if (suitable == COMPACT_SKIPPED) | |
2847 | /* Compaction cannot yet proceed. Do reclaim. */ | |
2848 | return false; | |
fe4b1b24 | 2849 | |
53853e2d | 2850 | /* |
fdd4c614 VB |
2851 | * Compaction is already possible, but it takes time to run and there |
2852 | * are potentially other callers using the pages just freed. So proceed | |
2853 | * with reclaim to make a buffer of free pages available to give | |
2854 | * compaction a reasonable chance of completing and allocating the page. | |
2855 | * Note that we won't actually reclaim the whole buffer in one attempt | |
2856 | * as the target watermark in should_continue_reclaim() is lower. But if | |
2857 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 2858 | */ |
fdd4c614 | 2859 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 2860 | |
fdd4c614 | 2861 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
2862 | } |
2863 | ||
1da177e4 LT |
2864 | /* |
2865 | * This is the direct reclaim path, for page-allocating processes. We only | |
2866 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2867 | * request. | |
2868 | * | |
1da177e4 LT |
2869 | * If a zone is deemed to be full of pinned pages then just give it a light |
2870 | * scan then give up on it. | |
2871 | */ | |
0a0337e0 | 2872 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2873 | { |
dd1a239f | 2874 | struct zoneref *z; |
54a6eb5c | 2875 | struct zone *zone; |
0608f43d AM |
2876 | unsigned long nr_soft_reclaimed; |
2877 | unsigned long nr_soft_scanned; | |
619d0d76 | 2878 | gfp_t orig_mask; |
79dafcdc | 2879 | pg_data_t *last_pgdat = NULL; |
1cfb419b | 2880 | |
cc715d99 MG |
2881 | /* |
2882 | * If the number of buffer_heads in the machine exceeds the maximum | |
2883 | * allowed level, force direct reclaim to scan the highmem zone as | |
2884 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2885 | */ | |
619d0d76 | 2886 | orig_mask = sc->gfp_mask; |
b2e18757 | 2887 | if (buffer_heads_over_limit) { |
cc715d99 | 2888 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 2889 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 2890 | } |
cc715d99 | 2891 | |
d4debc66 | 2892 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 2893 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
2894 | /* |
2895 | * Take care memory controller reclaiming has small influence | |
2896 | * to global LRU. | |
2897 | */ | |
89b5fae5 | 2898 | if (global_reclaim(sc)) { |
344736f2 VD |
2899 | if (!cpuset_zone_allowed(zone, |
2900 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2901 | continue; |
65ec02cb | 2902 | |
0b06496a JW |
2903 | /* |
2904 | * If we already have plenty of memory free for | |
2905 | * compaction in this zone, don't free any more. | |
2906 | * Even though compaction is invoked for any | |
2907 | * non-zero order, only frequent costly order | |
2908 | * reclamation is disruptive enough to become a | |
2909 | * noticeable problem, like transparent huge | |
2910 | * page allocations. | |
2911 | */ | |
2912 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2913 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 2914 | compaction_ready(zone, sc)) { |
0b06496a JW |
2915 | sc->compaction_ready = true; |
2916 | continue; | |
e0887c19 | 2917 | } |
0b06496a | 2918 | |
79dafcdc MG |
2919 | /* |
2920 | * Shrink each node in the zonelist once. If the | |
2921 | * zonelist is ordered by zone (not the default) then a | |
2922 | * node may be shrunk multiple times but in that case | |
2923 | * the user prefers lower zones being preserved. | |
2924 | */ | |
2925 | if (zone->zone_pgdat == last_pgdat) | |
2926 | continue; | |
2927 | ||
0608f43d AM |
2928 | /* |
2929 | * This steals pages from memory cgroups over softlimit | |
2930 | * and returns the number of reclaimed pages and | |
2931 | * scanned pages. This works for global memory pressure | |
2932 | * and balancing, not for a memcg's limit. | |
2933 | */ | |
2934 | nr_soft_scanned = 0; | |
ef8f2327 | 2935 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
2936 | sc->order, sc->gfp_mask, |
2937 | &nr_soft_scanned); | |
2938 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2939 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2940 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2941 | } |
408d8544 | 2942 | |
79dafcdc MG |
2943 | /* See comment about same check for global reclaim above */ |
2944 | if (zone->zone_pgdat == last_pgdat) | |
2945 | continue; | |
2946 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 2947 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 2948 | } |
e0c23279 | 2949 | |
619d0d76 WY |
2950 | /* |
2951 | * Restore to original mask to avoid the impact on the caller if we | |
2952 | * promoted it to __GFP_HIGHMEM. | |
2953 | */ | |
2954 | sc->gfp_mask = orig_mask; | |
1da177e4 | 2955 | } |
4f98a2fe | 2956 | |
2a2e4885 JW |
2957 | static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat) |
2958 | { | |
2959 | struct mem_cgroup *memcg; | |
2960 | ||
2961 | memcg = mem_cgroup_iter(root_memcg, NULL, NULL); | |
2962 | do { | |
2963 | unsigned long refaults; | |
2964 | struct lruvec *lruvec; | |
2965 | ||
2a2e4885 | 2966 | lruvec = mem_cgroup_lruvec(pgdat, memcg); |
205b20cc | 2967 | refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE); |
2a2e4885 JW |
2968 | lruvec->refaults = refaults; |
2969 | } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL))); | |
2970 | } | |
2971 | ||
1da177e4 LT |
2972 | /* |
2973 | * This is the main entry point to direct page reclaim. | |
2974 | * | |
2975 | * If a full scan of the inactive list fails to free enough memory then we | |
2976 | * are "out of memory" and something needs to be killed. | |
2977 | * | |
2978 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2979 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2980 | * caller can't do much about. We kick the writeback threads and take explicit |
2981 | * naps in the hope that some of these pages can be written. But if the | |
2982 | * allocating task holds filesystem locks which prevent writeout this might not | |
2983 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2984 | * |
2985 | * returns: 0, if no pages reclaimed | |
2986 | * else, the number of pages reclaimed | |
1da177e4 | 2987 | */ |
dac1d27b | 2988 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 2989 | struct scan_control *sc) |
1da177e4 | 2990 | { |
241994ed | 2991 | int initial_priority = sc->priority; |
2a2e4885 JW |
2992 | pg_data_t *last_pgdat; |
2993 | struct zoneref *z; | |
2994 | struct zone *zone; | |
241994ed | 2995 | retry: |
873b4771 KK |
2996 | delayacct_freepages_start(); |
2997 | ||
89b5fae5 | 2998 | if (global_reclaim(sc)) |
7cc30fcf | 2999 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 3000 | |
9e3b2f8c | 3001 | do { |
70ddf637 AV |
3002 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
3003 | sc->priority); | |
66e1707b | 3004 | sc->nr_scanned = 0; |
0a0337e0 | 3005 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 3006 | |
bb21c7ce | 3007 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
3008 | break; |
3009 | ||
3010 | if (sc->compaction_ready) | |
3011 | break; | |
1da177e4 | 3012 | |
0e50ce3b MK |
3013 | /* |
3014 | * If we're getting trouble reclaiming, start doing | |
3015 | * writepage even in laptop mode. | |
3016 | */ | |
3017 | if (sc->priority < DEF_PRIORITY - 2) | |
3018 | sc->may_writepage = 1; | |
0b06496a | 3019 | } while (--sc->priority >= 0); |
bb21c7ce | 3020 | |
2a2e4885 JW |
3021 | last_pgdat = NULL; |
3022 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
3023 | sc->nodemask) { | |
3024 | if (zone->zone_pgdat == last_pgdat) | |
3025 | continue; | |
3026 | last_pgdat = zone->zone_pgdat; | |
3027 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); | |
e3c1ac58 | 3028 | set_memcg_congestion(last_pgdat, sc->target_mem_cgroup, false); |
2a2e4885 JW |
3029 | } |
3030 | ||
873b4771 KK |
3031 | delayacct_freepages_end(); |
3032 | ||
bb21c7ce KM |
3033 | if (sc->nr_reclaimed) |
3034 | return sc->nr_reclaimed; | |
3035 | ||
0cee34fd | 3036 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 3037 | if (sc->compaction_ready) |
7335084d MG |
3038 | return 1; |
3039 | ||
241994ed | 3040 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
d6622f63 | 3041 | if (sc->memcg_low_skipped) { |
241994ed | 3042 | sc->priority = initial_priority; |
d6622f63 YX |
3043 | sc->memcg_low_reclaim = 1; |
3044 | sc->memcg_low_skipped = 0; | |
241994ed JW |
3045 | goto retry; |
3046 | } | |
3047 | ||
bb21c7ce | 3048 | return 0; |
1da177e4 LT |
3049 | } |
3050 | ||
c73322d0 | 3051 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
5515061d MG |
3052 | { |
3053 | struct zone *zone; | |
3054 | unsigned long pfmemalloc_reserve = 0; | |
3055 | unsigned long free_pages = 0; | |
3056 | int i; | |
3057 | bool wmark_ok; | |
3058 | ||
c73322d0 JW |
3059 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
3060 | return true; | |
3061 | ||
5515061d MG |
3062 | for (i = 0; i <= ZONE_NORMAL; i++) { |
3063 | zone = &pgdat->node_zones[i]; | |
d450abd8 JW |
3064 | if (!managed_zone(zone)) |
3065 | continue; | |
3066 | ||
3067 | if (!zone_reclaimable_pages(zone)) | |
675becce MG |
3068 | continue; |
3069 | ||
5515061d MG |
3070 | pfmemalloc_reserve += min_wmark_pages(zone); |
3071 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
3072 | } | |
3073 | ||
675becce MG |
3074 | /* If there are no reserves (unexpected config) then do not throttle */ |
3075 | if (!pfmemalloc_reserve) | |
3076 | return true; | |
3077 | ||
5515061d MG |
3078 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
3079 | ||
3080 | /* kswapd must be awake if processes are being throttled */ | |
3081 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
38087d9b | 3082 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
5515061d MG |
3083 | (enum zone_type)ZONE_NORMAL); |
3084 | wake_up_interruptible(&pgdat->kswapd_wait); | |
3085 | } | |
3086 | ||
3087 | return wmark_ok; | |
3088 | } | |
3089 | ||
3090 | /* | |
3091 | * Throttle direct reclaimers if backing storage is backed by the network | |
3092 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
3093 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
3094 | * when the low watermark is reached. |
3095 | * | |
3096 | * Returns true if a fatal signal was delivered during throttling. If this | |
3097 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 3098 | */ |
50694c28 | 3099 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
3100 | nodemask_t *nodemask) |
3101 | { | |
675becce | 3102 | struct zoneref *z; |
5515061d | 3103 | struct zone *zone; |
675becce | 3104 | pg_data_t *pgdat = NULL; |
5515061d MG |
3105 | |
3106 | /* | |
3107 | * Kernel threads should not be throttled as they may be indirectly | |
3108 | * responsible for cleaning pages necessary for reclaim to make forward | |
3109 | * progress. kjournald for example may enter direct reclaim while | |
3110 | * committing a transaction where throttling it could forcing other | |
3111 | * processes to block on log_wait_commit(). | |
3112 | */ | |
3113 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
3114 | goto out; |
3115 | ||
3116 | /* | |
3117 | * If a fatal signal is pending, this process should not throttle. | |
3118 | * It should return quickly so it can exit and free its memory | |
3119 | */ | |
3120 | if (fatal_signal_pending(current)) | |
3121 | goto out; | |
5515061d | 3122 | |
675becce MG |
3123 | /* |
3124 | * Check if the pfmemalloc reserves are ok by finding the first node | |
3125 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
3126 | * GFP_KERNEL will be required for allocating network buffers when | |
3127 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
3128 | * | |
3129 | * Throttling is based on the first usable node and throttled processes | |
3130 | * wait on a queue until kswapd makes progress and wakes them. There | |
3131 | * is an affinity then between processes waking up and where reclaim | |
3132 | * progress has been made assuming the process wakes on the same node. | |
3133 | * More importantly, processes running on remote nodes will not compete | |
3134 | * for remote pfmemalloc reserves and processes on different nodes | |
3135 | * should make reasonable progress. | |
3136 | */ | |
3137 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 3138 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
3139 | if (zone_idx(zone) > ZONE_NORMAL) |
3140 | continue; | |
3141 | ||
3142 | /* Throttle based on the first usable node */ | |
3143 | pgdat = zone->zone_pgdat; | |
c73322d0 | 3144 | if (allow_direct_reclaim(pgdat)) |
675becce MG |
3145 | goto out; |
3146 | break; | |
3147 | } | |
3148 | ||
3149 | /* If no zone was usable by the allocation flags then do not throttle */ | |
3150 | if (!pgdat) | |
50694c28 | 3151 | goto out; |
5515061d | 3152 | |
68243e76 MG |
3153 | /* Account for the throttling */ |
3154 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
3155 | ||
5515061d MG |
3156 | /* |
3157 | * If the caller cannot enter the filesystem, it's possible that it | |
3158 | * is due to the caller holding an FS lock or performing a journal | |
3159 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
3160 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
3161 | * blocked waiting on the same lock. Instead, throttle for up to a | |
3162 | * second before continuing. | |
3163 | */ | |
3164 | if (!(gfp_mask & __GFP_FS)) { | |
3165 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
c73322d0 | 3166 | allow_direct_reclaim(pgdat), HZ); |
50694c28 MG |
3167 | |
3168 | goto check_pending; | |
5515061d MG |
3169 | } |
3170 | ||
3171 | /* Throttle until kswapd wakes the process */ | |
3172 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
c73322d0 | 3173 | allow_direct_reclaim(pgdat)); |
50694c28 MG |
3174 | |
3175 | check_pending: | |
3176 | if (fatal_signal_pending(current)) | |
3177 | return true; | |
3178 | ||
3179 | out: | |
3180 | return false; | |
5515061d MG |
3181 | } |
3182 | ||
dac1d27b | 3183 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 3184 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 3185 | { |
33906bc5 | 3186 | unsigned long nr_reclaimed; |
66e1707b | 3187 | struct scan_control sc = { |
ee814fe2 | 3188 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
f2f43e56 | 3189 | .gfp_mask = current_gfp_context(gfp_mask), |
b2e18757 | 3190 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
3191 | .order = order, |
3192 | .nodemask = nodemask, | |
3193 | .priority = DEF_PRIORITY, | |
66e1707b | 3194 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3195 | .may_unmap = 1, |
2e2e4259 | 3196 | .may_swap = 1, |
1c30844d | 3197 | .may_shrinkslab = 1, |
66e1707b BS |
3198 | }; |
3199 | ||
bb451fdf GT |
3200 | /* |
3201 | * scan_control uses s8 fields for order, priority, and reclaim_idx. | |
3202 | * Confirm they are large enough for max values. | |
3203 | */ | |
3204 | BUILD_BUG_ON(MAX_ORDER > S8_MAX); | |
3205 | BUILD_BUG_ON(DEF_PRIORITY > S8_MAX); | |
3206 | BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX); | |
3207 | ||
5515061d | 3208 | /* |
50694c28 MG |
3209 | * Do not enter reclaim if fatal signal was delivered while throttled. |
3210 | * 1 is returned so that the page allocator does not OOM kill at this | |
3211 | * point. | |
5515061d | 3212 | */ |
f2f43e56 | 3213 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
5515061d MG |
3214 | return 1; |
3215 | ||
1732d2b0 | 3216 | set_task_reclaim_state(current, &sc.reclaim_state); |
3481c37f | 3217 | trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask); |
33906bc5 | 3218 | |
3115cd91 | 3219 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
3220 | |
3221 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
1732d2b0 | 3222 | set_task_reclaim_state(current, NULL); |
33906bc5 MG |
3223 | |
3224 | return nr_reclaimed; | |
66e1707b BS |
3225 | } |
3226 | ||
c255a458 | 3227 | #ifdef CONFIG_MEMCG |
66e1707b | 3228 | |
a9dd0a83 | 3229 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 3230 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 3231 | pg_data_t *pgdat, |
0ae5e89c | 3232 | unsigned long *nr_scanned) |
4e416953 BS |
3233 | { |
3234 | struct scan_control sc = { | |
b8f5c566 | 3235 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 3236 | .target_mem_cgroup = memcg, |
4e416953 BS |
3237 | .may_writepage = !laptop_mode, |
3238 | .may_unmap = 1, | |
b2e18757 | 3239 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 3240 | .may_swap = !noswap, |
1c30844d | 3241 | .may_shrinkslab = 1, |
4e416953 | 3242 | }; |
6b4f7799 | 3243 | unsigned long lru_pages; |
0ae5e89c | 3244 | |
1732d2b0 | 3245 | set_task_reclaim_state(current, &sc.reclaim_state); |
4e416953 BS |
3246 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3247 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 3248 | |
9e3b2f8c | 3249 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
3481c37f | 3250 | sc.gfp_mask); |
bdce6d9e | 3251 | |
4e416953 BS |
3252 | /* |
3253 | * NOTE: Although we can get the priority field, using it | |
3254 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 3255 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
3256 | * will pick up pages from other mem cgroup's as well. We hack |
3257 | * the priority and make it zero. | |
3258 | */ | |
ef8f2327 | 3259 | shrink_node_memcg(pgdat, memcg, &sc, &lru_pages); |
bdce6d9e KM |
3260 | |
3261 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
3262 | ||
1732d2b0 | 3263 | set_task_reclaim_state(current, NULL); |
0ae5e89c | 3264 | *nr_scanned = sc.nr_scanned; |
0308f7cf | 3265 | |
4e416953 BS |
3266 | return sc.nr_reclaimed; |
3267 | } | |
3268 | ||
72835c86 | 3269 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 3270 | unsigned long nr_pages, |
a7885eb8 | 3271 | gfp_t gfp_mask, |
b70a2a21 | 3272 | bool may_swap) |
66e1707b | 3273 | { |
4e416953 | 3274 | struct zonelist *zonelist; |
bdce6d9e | 3275 | unsigned long nr_reclaimed; |
eb414681 | 3276 | unsigned long pflags; |
889976db | 3277 | int nid; |
499118e9 | 3278 | unsigned int noreclaim_flag; |
66e1707b | 3279 | struct scan_control sc = { |
b70a2a21 | 3280 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 3281 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
a09ed5e0 | 3282 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
b2e18757 | 3283 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
3284 | .target_mem_cgroup = memcg, |
3285 | .priority = DEF_PRIORITY, | |
3286 | .may_writepage = !laptop_mode, | |
3287 | .may_unmap = 1, | |
b70a2a21 | 3288 | .may_swap = may_swap, |
1c30844d | 3289 | .may_shrinkslab = 1, |
a09ed5e0 | 3290 | }; |
66e1707b | 3291 | |
1732d2b0 | 3292 | set_task_reclaim_state(current, &sc.reclaim_state); |
889976db YH |
3293 | /* |
3294 | * Unlike direct reclaim via alloc_pages(), memcg's reclaim doesn't | |
3295 | * take care of from where we get pages. So the node where we start the | |
3296 | * scan does not need to be the current node. | |
3297 | */ | |
72835c86 | 3298 | nid = mem_cgroup_select_victim_node(memcg); |
889976db | 3299 | |
c9634cf0 | 3300 | zonelist = &NODE_DATA(nid)->node_zonelists[ZONELIST_FALLBACK]; |
bdce6d9e | 3301 | |
3481c37f | 3302 | trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask); |
bdce6d9e | 3303 | |
eb414681 | 3304 | psi_memstall_enter(&pflags); |
499118e9 | 3305 | noreclaim_flag = memalloc_noreclaim_save(); |
eb414681 | 3306 | |
3115cd91 | 3307 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
eb414681 | 3308 | |
499118e9 | 3309 | memalloc_noreclaim_restore(noreclaim_flag); |
eb414681 | 3310 | psi_memstall_leave(&pflags); |
bdce6d9e KM |
3311 | |
3312 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
1732d2b0 | 3313 | set_task_reclaim_state(current, NULL); |
bdce6d9e KM |
3314 | |
3315 | return nr_reclaimed; | |
66e1707b BS |
3316 | } |
3317 | #endif | |
3318 | ||
1d82de61 | 3319 | static void age_active_anon(struct pglist_data *pgdat, |
ef8f2327 | 3320 | struct scan_control *sc) |
f16015fb | 3321 | { |
b95a2f2d | 3322 | struct mem_cgroup *memcg; |
f16015fb | 3323 | |
b95a2f2d JW |
3324 | if (!total_swap_pages) |
3325 | return; | |
3326 | ||
3327 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
3328 | do { | |
ef8f2327 | 3329 | struct lruvec *lruvec = mem_cgroup_lruvec(pgdat, memcg); |
b95a2f2d | 3330 | |
3b991208 | 3331 | if (inactive_list_is_low(lruvec, false, sc, true)) |
1a93be0e | 3332 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 3333 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
3334 | |
3335 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
3336 | } while (memcg); | |
f16015fb JW |
3337 | } |
3338 | ||
1c30844d MG |
3339 | static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx) |
3340 | { | |
3341 | int i; | |
3342 | struct zone *zone; | |
3343 | ||
3344 | /* | |
3345 | * Check for watermark boosts top-down as the higher zones | |
3346 | * are more likely to be boosted. Both watermarks and boosts | |
3347 | * should not be checked at the time time as reclaim would | |
3348 | * start prematurely when there is no boosting and a lower | |
3349 | * zone is balanced. | |
3350 | */ | |
3351 | for (i = classzone_idx; i >= 0; i--) { | |
3352 | zone = pgdat->node_zones + i; | |
3353 | if (!managed_zone(zone)) | |
3354 | continue; | |
3355 | ||
3356 | if (zone->watermark_boost) | |
3357 | return true; | |
3358 | } | |
3359 | ||
3360 | return false; | |
3361 | } | |
3362 | ||
e716f2eb MG |
3363 | /* |
3364 | * Returns true if there is an eligible zone balanced for the request order | |
3365 | * and classzone_idx | |
3366 | */ | |
3367 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) | |
60cefed4 | 3368 | { |
e716f2eb MG |
3369 | int i; |
3370 | unsigned long mark = -1; | |
3371 | struct zone *zone; | |
60cefed4 | 3372 | |
1c30844d MG |
3373 | /* |
3374 | * Check watermarks bottom-up as lower zones are more likely to | |
3375 | * meet watermarks. | |
3376 | */ | |
e716f2eb MG |
3377 | for (i = 0; i <= classzone_idx; i++) { |
3378 | zone = pgdat->node_zones + i; | |
6256c6b4 | 3379 | |
e716f2eb MG |
3380 | if (!managed_zone(zone)) |
3381 | continue; | |
3382 | ||
3383 | mark = high_wmark_pages(zone); | |
3384 | if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) | |
3385 | return true; | |
3386 | } | |
3387 | ||
3388 | /* | |
3389 | * If a node has no populated zone within classzone_idx, it does not | |
3390 | * need balancing by definition. This can happen if a zone-restricted | |
3391 | * allocation tries to wake a remote kswapd. | |
3392 | */ | |
3393 | if (mark == -1) | |
3394 | return true; | |
3395 | ||
3396 | return false; | |
60cefed4 JW |
3397 | } |
3398 | ||
631b6e08 MG |
3399 | /* Clear pgdat state for congested, dirty or under writeback. */ |
3400 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
3401 | { | |
3402 | clear_bit(PGDAT_CONGESTED, &pgdat->flags); | |
3403 | clear_bit(PGDAT_DIRTY, &pgdat->flags); | |
3404 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
3405 | } | |
3406 | ||
5515061d MG |
3407 | /* |
3408 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3409 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3410 | * | |
3411 | * Returns true if kswapd is ready to sleep | |
3412 | */ | |
d9f21d42 | 3413 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f50de2d3 | 3414 | { |
5515061d | 3415 | /* |
9e5e3661 | 3416 | * The throttled processes are normally woken up in balance_pgdat() as |
c73322d0 | 3417 | * soon as allow_direct_reclaim() is true. But there is a potential |
9e5e3661 VB |
3418 | * race between when kswapd checks the watermarks and a process gets |
3419 | * throttled. There is also a potential race if processes get | |
3420 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3421 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3422 | * the wake up checks. If kswapd is going to sleep, no process should | |
3423 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3424 | * the wake up is premature, processes will wake kswapd and get | |
3425 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3426 | * that here we are under prepare_to_wait(). | |
5515061d | 3427 | */ |
9e5e3661 VB |
3428 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3429 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3430 | |
c73322d0 JW |
3431 | /* Hopeless node, leave it to direct reclaim */ |
3432 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3433 | return true; | |
3434 | ||
e716f2eb MG |
3435 | if (pgdat_balanced(pgdat, order, classzone_idx)) { |
3436 | clear_pgdat_congested(pgdat); | |
3437 | return true; | |
1d82de61 MG |
3438 | } |
3439 | ||
333b0a45 | 3440 | return false; |
f50de2d3 MG |
3441 | } |
3442 | ||
75485363 | 3443 | /* |
1d82de61 MG |
3444 | * kswapd shrinks a node of pages that are at or below the highest usable |
3445 | * zone that is currently unbalanced. | |
b8e83b94 MG |
3446 | * |
3447 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3448 | * reclaim or if the lack of progress was due to pages under writeback. |
3449 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3450 | */ |
1d82de61 | 3451 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 3452 | struct scan_control *sc) |
75485363 | 3453 | { |
1d82de61 MG |
3454 | struct zone *zone; |
3455 | int z; | |
75485363 | 3456 | |
1d82de61 MG |
3457 | /* Reclaim a number of pages proportional to the number of zones */ |
3458 | sc->nr_to_reclaim = 0; | |
970a39a3 | 3459 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 3460 | zone = pgdat->node_zones + z; |
6aa303de | 3461 | if (!managed_zone(zone)) |
1d82de61 | 3462 | continue; |
7c954f6d | 3463 | |
1d82de61 MG |
3464 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
3465 | } | |
7c954f6d MG |
3466 | |
3467 | /* | |
1d82de61 MG |
3468 | * Historically care was taken to put equal pressure on all zones but |
3469 | * now pressure is applied based on node LRU order. | |
7c954f6d | 3470 | */ |
970a39a3 | 3471 | shrink_node(pgdat, sc); |
283aba9f | 3472 | |
7c954f6d | 3473 | /* |
1d82de61 MG |
3474 | * Fragmentation may mean that the system cannot be rebalanced for |
3475 | * high-order allocations. If twice the allocation size has been | |
3476 | * reclaimed then recheck watermarks only at order-0 to prevent | |
3477 | * excessive reclaim. Assume that a process requested a high-order | |
3478 | * can direct reclaim/compact. | |
7c954f6d | 3479 | */ |
9861a62c | 3480 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 3481 | sc->order = 0; |
7c954f6d | 3482 | |
b8e83b94 | 3483 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3484 | } |
3485 | ||
1da177e4 | 3486 | /* |
1d82de61 MG |
3487 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
3488 | * that are eligible for use by the caller until at least one zone is | |
3489 | * balanced. | |
1da177e4 | 3490 | * |
1d82de61 | 3491 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
3492 | * |
3493 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 3494 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
8bb4e7a2 | 3495 | * found to have free_pages <= high_wmark_pages(zone), any page in that zone |
1d82de61 MG |
3496 | * or lower is eligible for reclaim until at least one usable zone is |
3497 | * balanced. | |
1da177e4 | 3498 | */ |
accf6242 | 3499 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
1da177e4 | 3500 | { |
1da177e4 | 3501 | int i; |
0608f43d AM |
3502 | unsigned long nr_soft_reclaimed; |
3503 | unsigned long nr_soft_scanned; | |
eb414681 | 3504 | unsigned long pflags; |
1c30844d MG |
3505 | unsigned long nr_boost_reclaim; |
3506 | unsigned long zone_boosts[MAX_NR_ZONES] = { 0, }; | |
3507 | bool boosted; | |
1d82de61 | 3508 | struct zone *zone; |
179e9639 AM |
3509 | struct scan_control sc = { |
3510 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3511 | .order = order, |
a6dc60f8 | 3512 | .may_unmap = 1, |
179e9639 | 3513 | }; |
93781325 | 3514 | |
1732d2b0 | 3515 | set_task_reclaim_state(current, &sc.reclaim_state); |
eb414681 | 3516 | psi_memstall_enter(&pflags); |
93781325 OS |
3517 | __fs_reclaim_acquire(); |
3518 | ||
f8891e5e | 3519 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3520 | |
1c30844d MG |
3521 | /* |
3522 | * Account for the reclaim boost. Note that the zone boost is left in | |
3523 | * place so that parallel allocations that are near the watermark will | |
3524 | * stall or direct reclaim until kswapd is finished. | |
3525 | */ | |
3526 | nr_boost_reclaim = 0; | |
3527 | for (i = 0; i <= classzone_idx; i++) { | |
3528 | zone = pgdat->node_zones + i; | |
3529 | if (!managed_zone(zone)) | |
3530 | continue; | |
3531 | ||
3532 | nr_boost_reclaim += zone->watermark_boost; | |
3533 | zone_boosts[i] = zone->watermark_boost; | |
3534 | } | |
3535 | boosted = nr_boost_reclaim; | |
3536 | ||
3537 | restart: | |
3538 | sc.priority = DEF_PRIORITY; | |
9e3b2f8c | 3539 | do { |
c73322d0 | 3540 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
b8e83b94 | 3541 | bool raise_priority = true; |
1c30844d | 3542 | bool balanced; |
93781325 | 3543 | bool ret; |
b8e83b94 | 3544 | |
84c7a777 | 3545 | sc.reclaim_idx = classzone_idx; |
1da177e4 | 3546 | |
86c79f6b | 3547 | /* |
84c7a777 MG |
3548 | * If the number of buffer_heads exceeds the maximum allowed |
3549 | * then consider reclaiming from all zones. This has a dual | |
3550 | * purpose -- on 64-bit systems it is expected that | |
3551 | * buffer_heads are stripped during active rotation. On 32-bit | |
3552 | * systems, highmem pages can pin lowmem memory and shrinking | |
3553 | * buffers can relieve lowmem pressure. Reclaim may still not | |
3554 | * go ahead if all eligible zones for the original allocation | |
3555 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
3556 | */ |
3557 | if (buffer_heads_over_limit) { | |
3558 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
3559 | zone = pgdat->node_zones + i; | |
6aa303de | 3560 | if (!managed_zone(zone)) |
86c79f6b | 3561 | continue; |
cc715d99 | 3562 | |
970a39a3 | 3563 | sc.reclaim_idx = i; |
e1dbeda6 | 3564 | break; |
1da177e4 | 3565 | } |
1da177e4 | 3566 | } |
dafcb73e | 3567 | |
86c79f6b | 3568 | /* |
1c30844d MG |
3569 | * If the pgdat is imbalanced then ignore boosting and preserve |
3570 | * the watermarks for a later time and restart. Note that the | |
3571 | * zone watermarks will be still reset at the end of balancing | |
3572 | * on the grounds that the normal reclaim should be enough to | |
3573 | * re-evaluate if boosting is required when kswapd next wakes. | |
3574 | */ | |
3575 | balanced = pgdat_balanced(pgdat, sc.order, classzone_idx); | |
3576 | if (!balanced && nr_boost_reclaim) { | |
3577 | nr_boost_reclaim = 0; | |
3578 | goto restart; | |
3579 | } | |
3580 | ||
3581 | /* | |
3582 | * If boosting is not active then only reclaim if there are no | |
3583 | * eligible zones. Note that sc.reclaim_idx is not used as | |
3584 | * buffer_heads_over_limit may have adjusted it. | |
86c79f6b | 3585 | */ |
1c30844d | 3586 | if (!nr_boost_reclaim && balanced) |
e716f2eb | 3587 | goto out; |
e1dbeda6 | 3588 | |
1c30844d MG |
3589 | /* Limit the priority of boosting to avoid reclaim writeback */ |
3590 | if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2) | |
3591 | raise_priority = false; | |
3592 | ||
3593 | /* | |
3594 | * Do not writeback or swap pages for boosted reclaim. The | |
3595 | * intent is to relieve pressure not issue sub-optimal IO | |
3596 | * from reclaim context. If no pages are reclaimed, the | |
3597 | * reclaim will be aborted. | |
3598 | */ | |
3599 | sc.may_writepage = !laptop_mode && !nr_boost_reclaim; | |
3600 | sc.may_swap = !nr_boost_reclaim; | |
3601 | sc.may_shrinkslab = !nr_boost_reclaim; | |
3602 | ||
1d82de61 MG |
3603 | /* |
3604 | * Do some background aging of the anon list, to give | |
3605 | * pages a chance to be referenced before reclaiming. All | |
3606 | * pages are rotated regardless of classzone as this is | |
3607 | * about consistent aging. | |
3608 | */ | |
ef8f2327 | 3609 | age_active_anon(pgdat, &sc); |
1d82de61 | 3610 | |
b7ea3c41 MG |
3611 | /* |
3612 | * If we're getting trouble reclaiming, start doing writepage | |
3613 | * even in laptop mode. | |
3614 | */ | |
047d72c3 | 3615 | if (sc.priority < DEF_PRIORITY - 2) |
b7ea3c41 MG |
3616 | sc.may_writepage = 1; |
3617 | ||
1d82de61 MG |
3618 | /* Call soft limit reclaim before calling shrink_node. */ |
3619 | sc.nr_scanned = 0; | |
3620 | nr_soft_scanned = 0; | |
ef8f2327 | 3621 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
3622 | sc.gfp_mask, &nr_soft_scanned); |
3623 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3624 | ||
1da177e4 | 3625 | /* |
1d82de61 MG |
3626 | * There should be no need to raise the scanning priority if |
3627 | * enough pages are already being scanned that that high | |
3628 | * watermark would be met at 100% efficiency. | |
1da177e4 | 3629 | */ |
970a39a3 | 3630 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 3631 | raise_priority = false; |
5515061d MG |
3632 | |
3633 | /* | |
3634 | * If the low watermark is met there is no need for processes | |
3635 | * to be throttled on pfmemalloc_wait as they should not be | |
3636 | * able to safely make forward progress. Wake them | |
3637 | */ | |
3638 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
c73322d0 | 3639 | allow_direct_reclaim(pgdat)) |
cfc51155 | 3640 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3641 | |
b8e83b94 | 3642 | /* Check if kswapd should be suspending */ |
93781325 OS |
3643 | __fs_reclaim_release(); |
3644 | ret = try_to_freeze(); | |
3645 | __fs_reclaim_acquire(); | |
3646 | if (ret || kthread_should_stop()) | |
b8e83b94 | 3647 | break; |
8357376d | 3648 | |
73ce02e9 | 3649 | /* |
b8e83b94 MG |
3650 | * Raise priority if scanning rate is too low or there was no |
3651 | * progress in reclaiming pages | |
73ce02e9 | 3652 | */ |
c73322d0 | 3653 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
1c30844d MG |
3654 | nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed); |
3655 | ||
3656 | /* | |
3657 | * If reclaim made no progress for a boost, stop reclaim as | |
3658 | * IO cannot be queued and it could be an infinite loop in | |
3659 | * extreme circumstances. | |
3660 | */ | |
3661 | if (nr_boost_reclaim && !nr_reclaimed) | |
3662 | break; | |
3663 | ||
c73322d0 | 3664 | if (raise_priority || !nr_reclaimed) |
b8e83b94 | 3665 | sc.priority--; |
1d82de61 | 3666 | } while (sc.priority >= 1); |
1da177e4 | 3667 | |
c73322d0 JW |
3668 | if (!sc.nr_reclaimed) |
3669 | pgdat->kswapd_failures++; | |
3670 | ||
b8e83b94 | 3671 | out: |
1c30844d MG |
3672 | /* If reclaim was boosted, account for the reclaim done in this pass */ |
3673 | if (boosted) { | |
3674 | unsigned long flags; | |
3675 | ||
3676 | for (i = 0; i <= classzone_idx; i++) { | |
3677 | if (!zone_boosts[i]) | |
3678 | continue; | |
3679 | ||
3680 | /* Increments are under the zone lock */ | |
3681 | zone = pgdat->node_zones + i; | |
3682 | spin_lock_irqsave(&zone->lock, flags); | |
3683 | zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]); | |
3684 | spin_unlock_irqrestore(&zone->lock, flags); | |
3685 | } | |
3686 | ||
3687 | /* | |
3688 | * As there is now likely space, wakeup kcompact to defragment | |
3689 | * pageblocks. | |
3690 | */ | |
3691 | wakeup_kcompactd(pgdat, pageblock_order, classzone_idx); | |
3692 | } | |
3693 | ||
2a2e4885 | 3694 | snapshot_refaults(NULL, pgdat); |
93781325 | 3695 | __fs_reclaim_release(); |
eb414681 | 3696 | psi_memstall_leave(&pflags); |
1732d2b0 | 3697 | set_task_reclaim_state(current, NULL); |
e5ca8071 | 3698 | |
0abdee2b | 3699 | /* |
1d82de61 MG |
3700 | * Return the order kswapd stopped reclaiming at as |
3701 | * prepare_kswapd_sleep() takes it into account. If another caller | |
3702 | * entered the allocator slow path while kswapd was awake, order will | |
3703 | * remain at the higher level. | |
0abdee2b | 3704 | */ |
1d82de61 | 3705 | return sc.order; |
1da177e4 LT |
3706 | } |
3707 | ||
e716f2eb | 3708 | /* |
dffcac2c SB |
3709 | * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be |
3710 | * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not | |
3711 | * a valid index then either kswapd runs for first time or kswapd couldn't sleep | |
3712 | * after previous reclaim attempt (node is still unbalanced). In that case | |
3713 | * return the zone index of the previous kswapd reclaim cycle. | |
e716f2eb MG |
3714 | */ |
3715 | static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, | |
dffcac2c | 3716 | enum zone_type prev_classzone_idx) |
e716f2eb MG |
3717 | { |
3718 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
dffcac2c SB |
3719 | return prev_classzone_idx; |
3720 | return pgdat->kswapd_classzone_idx; | |
e716f2eb MG |
3721 | } |
3722 | ||
38087d9b MG |
3723 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
3724 | unsigned int classzone_idx) | |
f0bc0a60 KM |
3725 | { |
3726 | long remaining = 0; | |
3727 | DEFINE_WAIT(wait); | |
3728 | ||
3729 | if (freezing(current) || kthread_should_stop()) | |
3730 | return; | |
3731 | ||
3732 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3733 | ||
333b0a45 SG |
3734 | /* |
3735 | * Try to sleep for a short interval. Note that kcompactd will only be | |
3736 | * woken if it is possible to sleep for a short interval. This is | |
3737 | * deliberate on the assumption that if reclaim cannot keep an | |
3738 | * eligible zone balanced that it's also unlikely that compaction will | |
3739 | * succeed. | |
3740 | */ | |
d9f21d42 | 3741 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
fd901c95 VB |
3742 | /* |
3743 | * Compaction records what page blocks it recently failed to | |
3744 | * isolate pages from and skips them in the future scanning. | |
3745 | * When kswapd is going to sleep, it is reasonable to assume | |
3746 | * that pages and compaction may succeed so reset the cache. | |
3747 | */ | |
3748 | reset_isolation_suitable(pgdat); | |
3749 | ||
3750 | /* | |
3751 | * We have freed the memory, now we should compact it to make | |
3752 | * allocation of the requested order possible. | |
3753 | */ | |
38087d9b | 3754 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); |
fd901c95 | 3755 | |
f0bc0a60 | 3756 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
3757 | |
3758 | /* | |
3759 | * If woken prematurely then reset kswapd_classzone_idx and | |
3760 | * order. The values will either be from a wakeup request or | |
3761 | * the previous request that slept prematurely. | |
3762 | */ | |
3763 | if (remaining) { | |
e716f2eb | 3764 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
38087d9b MG |
3765 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); |
3766 | } | |
3767 | ||
f0bc0a60 KM |
3768 | finish_wait(&pgdat->kswapd_wait, &wait); |
3769 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3770 | } | |
3771 | ||
3772 | /* | |
3773 | * After a short sleep, check if it was a premature sleep. If not, then | |
3774 | * go fully to sleep until explicitly woken up. | |
3775 | */ | |
d9f21d42 MG |
3776 | if (!remaining && |
3777 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { | |
f0bc0a60 KM |
3778 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3779 | ||
3780 | /* | |
3781 | * vmstat counters are not perfectly accurate and the estimated | |
3782 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3783 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3784 | * watermarks being breached while under pressure, we reduce the | |
3785 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3786 | * them before going back to sleep. | |
3787 | */ | |
3788 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
3789 | |
3790 | if (!kthread_should_stop()) | |
3791 | schedule(); | |
3792 | ||
f0bc0a60 KM |
3793 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3794 | } else { | |
3795 | if (remaining) | |
3796 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3797 | else | |
3798 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3799 | } | |
3800 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3801 | } | |
3802 | ||
1da177e4 LT |
3803 | /* |
3804 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3805 | * from the init process. |
1da177e4 LT |
3806 | * |
3807 | * This basically trickles out pages so that we have _some_ | |
3808 | * free memory available even if there is no other activity | |
3809 | * that frees anything up. This is needed for things like routing | |
3810 | * etc, where we otherwise might have all activity going on in | |
3811 | * asynchronous contexts that cannot page things out. | |
3812 | * | |
3813 | * If there are applications that are active memory-allocators | |
3814 | * (most normal use), this basically shouldn't matter. | |
3815 | */ | |
3816 | static int kswapd(void *p) | |
3817 | { | |
e716f2eb MG |
3818 | unsigned int alloc_order, reclaim_order; |
3819 | unsigned int classzone_idx = MAX_NR_ZONES - 1; | |
1da177e4 LT |
3820 | pg_data_t *pgdat = (pg_data_t*)p; |
3821 | struct task_struct *tsk = current; | |
a70f7302 | 3822 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3823 | |
174596a0 | 3824 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3825 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3826 | |
3827 | /* | |
3828 | * Tell the memory management that we're a "memory allocator", | |
3829 | * and that if we need more memory we should get access to it | |
3830 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3831 | * never get caught in the normal page freeing logic. | |
3832 | * | |
3833 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3834 | * you need a small amount of memory in order to be able to | |
3835 | * page out something else, and this flag essentially protects | |
3836 | * us from recursively trying to free more memory as we're | |
3837 | * trying to free the first piece of memory in the first place). | |
3838 | */ | |
930d9152 | 3839 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3840 | set_freezable(); |
1da177e4 | 3841 | |
e716f2eb MG |
3842 | pgdat->kswapd_order = 0; |
3843 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; | |
1da177e4 | 3844 | for ( ; ; ) { |
6f6313d4 | 3845 | bool ret; |
3e1d1d28 | 3846 | |
e716f2eb MG |
3847 | alloc_order = reclaim_order = pgdat->kswapd_order; |
3848 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); | |
3849 | ||
38087d9b MG |
3850 | kswapd_try_sleep: |
3851 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
3852 | classzone_idx); | |
215ddd66 | 3853 | |
38087d9b MG |
3854 | /* Read the new order and classzone_idx */ |
3855 | alloc_order = reclaim_order = pgdat->kswapd_order; | |
dffcac2c | 3856 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
38087d9b | 3857 | pgdat->kswapd_order = 0; |
e716f2eb | 3858 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; |
1da177e4 | 3859 | |
8fe23e05 DR |
3860 | ret = try_to_freeze(); |
3861 | if (kthread_should_stop()) | |
3862 | break; | |
3863 | ||
3864 | /* | |
3865 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3866 | * after returning from the refrigerator | |
3867 | */ | |
38087d9b MG |
3868 | if (ret) |
3869 | continue; | |
3870 | ||
3871 | /* | |
3872 | * Reclaim begins at the requested order but if a high-order | |
3873 | * reclaim fails then kswapd falls back to reclaiming for | |
3874 | * order-0. If that happens, kswapd will consider sleeping | |
3875 | * for the order it finished reclaiming at (reclaim_order) | |
3876 | * but kcompactd is woken to compact for the original | |
3877 | * request (alloc_order). | |
3878 | */ | |
e5146b12 MG |
3879 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, |
3880 | alloc_order); | |
38087d9b MG |
3881 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); |
3882 | if (reclaim_order < alloc_order) | |
3883 | goto kswapd_try_sleep; | |
1da177e4 | 3884 | } |
b0a8cc58 | 3885 | |
71abdc15 | 3886 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
71abdc15 | 3887 | |
1da177e4 LT |
3888 | return 0; |
3889 | } | |
3890 | ||
3891 | /* | |
5ecd9d40 DR |
3892 | * A zone is low on free memory or too fragmented for high-order memory. If |
3893 | * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's | |
3894 | * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim | |
3895 | * has failed or is not needed, still wake up kcompactd if only compaction is | |
3896 | * needed. | |
1da177e4 | 3897 | */ |
5ecd9d40 DR |
3898 | void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, |
3899 | enum zone_type classzone_idx) | |
1da177e4 LT |
3900 | { |
3901 | pg_data_t *pgdat; | |
3902 | ||
6aa303de | 3903 | if (!managed_zone(zone)) |
1da177e4 LT |
3904 | return; |
3905 | ||
5ecd9d40 | 3906 | if (!cpuset_zone_allowed(zone, gfp_flags)) |
1da177e4 | 3907 | return; |
88f5acf8 | 3908 | pgdat = zone->zone_pgdat; |
dffcac2c SB |
3909 | |
3910 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
3911 | pgdat->kswapd_classzone_idx = classzone_idx; | |
3912 | else | |
3913 | pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, | |
3914 | classzone_idx); | |
38087d9b | 3915 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); |
8d0986e2 | 3916 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3917 | return; |
e1a55637 | 3918 | |
5ecd9d40 DR |
3919 | /* Hopeless node, leave it to direct reclaim if possible */ |
3920 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES || | |
1c30844d MG |
3921 | (pgdat_balanced(pgdat, order, classzone_idx) && |
3922 | !pgdat_watermark_boosted(pgdat, classzone_idx))) { | |
5ecd9d40 DR |
3923 | /* |
3924 | * There may be plenty of free memory available, but it's too | |
3925 | * fragmented for high-order allocations. Wake up kcompactd | |
3926 | * and rely on compaction_suitable() to determine if it's | |
3927 | * needed. If it fails, it will defer subsequent attempts to | |
3928 | * ratelimit its work. | |
3929 | */ | |
3930 | if (!(gfp_flags & __GFP_DIRECT_RECLAIM)) | |
3931 | wakeup_kcompactd(pgdat, order, classzone_idx); | |
e716f2eb | 3932 | return; |
5ecd9d40 | 3933 | } |
88f5acf8 | 3934 | |
5ecd9d40 DR |
3935 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order, |
3936 | gfp_flags); | |
8d0986e2 | 3937 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3938 | } |
3939 | ||
c6f37f12 | 3940 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3941 | /* |
7b51755c | 3942 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3943 | * freed pages. |
3944 | * | |
3945 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3946 | * LRU order by reclaiming preferentially | |
3947 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3948 | */ |
7b51755c | 3949 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3950 | { |
d6277db4 | 3951 | struct scan_control sc = { |
ee814fe2 | 3952 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 3953 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 3954 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 3955 | .priority = DEF_PRIORITY, |
d6277db4 | 3956 | .may_writepage = 1, |
ee814fe2 JW |
3957 | .may_unmap = 1, |
3958 | .may_swap = 1, | |
7b51755c | 3959 | .hibernation_mode = 1, |
1da177e4 | 3960 | }; |
a09ed5e0 | 3961 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c | 3962 | unsigned long nr_reclaimed; |
499118e9 | 3963 | unsigned int noreclaim_flag; |
1da177e4 | 3964 | |
d92a8cfc | 3965 | fs_reclaim_acquire(sc.gfp_mask); |
93781325 | 3966 | noreclaim_flag = memalloc_noreclaim_save(); |
1732d2b0 | 3967 | set_task_reclaim_state(current, &sc.reclaim_state); |
d6277db4 | 3968 | |
3115cd91 | 3969 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 3970 | |
1732d2b0 | 3971 | set_task_reclaim_state(current, NULL); |
499118e9 | 3972 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 3973 | fs_reclaim_release(sc.gfp_mask); |
d6277db4 | 3974 | |
7b51755c | 3975 | return nr_reclaimed; |
1da177e4 | 3976 | } |
c6f37f12 | 3977 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3978 | |
1da177e4 LT |
3979 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3980 | not required for correctness. So if the last cpu in a node goes | |
3981 | away, we get changed to run anywhere: as the first one comes back, | |
3982 | restore their cpu bindings. */ | |
517bbed9 | 3983 | static int kswapd_cpu_online(unsigned int cpu) |
1da177e4 | 3984 | { |
58c0a4a7 | 3985 | int nid; |
1da177e4 | 3986 | |
517bbed9 SAS |
3987 | for_each_node_state(nid, N_MEMORY) { |
3988 | pg_data_t *pgdat = NODE_DATA(nid); | |
3989 | const struct cpumask *mask; | |
a70f7302 | 3990 | |
517bbed9 | 3991 | mask = cpumask_of_node(pgdat->node_id); |
c5f59f08 | 3992 | |
517bbed9 SAS |
3993 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
3994 | /* One of our CPUs online: restore mask */ | |
3995 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
1da177e4 | 3996 | } |
517bbed9 | 3997 | return 0; |
1da177e4 | 3998 | } |
1da177e4 | 3999 | |
3218ae14 YG |
4000 | /* |
4001 | * This kswapd start function will be called by init and node-hot-add. | |
4002 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
4003 | */ | |
4004 | int kswapd_run(int nid) | |
4005 | { | |
4006 | pg_data_t *pgdat = NODE_DATA(nid); | |
4007 | int ret = 0; | |
4008 | ||
4009 | if (pgdat->kswapd) | |
4010 | return 0; | |
4011 | ||
4012 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
4013 | if (IS_ERR(pgdat->kswapd)) { | |
4014 | /* failure at boot is fatal */ | |
c6202adf | 4015 | BUG_ON(system_state < SYSTEM_RUNNING); |
d5dc0ad9 GS |
4016 | pr_err("Failed to start kswapd on node %d\n", nid); |
4017 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 4018 | pgdat->kswapd = NULL; |
3218ae14 YG |
4019 | } |
4020 | return ret; | |
4021 | } | |
4022 | ||
8fe23e05 | 4023 | /* |
d8adde17 | 4024 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 4025 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
4026 | */ |
4027 | void kswapd_stop(int nid) | |
4028 | { | |
4029 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
4030 | ||
d8adde17 | 4031 | if (kswapd) { |
8fe23e05 | 4032 | kthread_stop(kswapd); |
d8adde17 JL |
4033 | NODE_DATA(nid)->kswapd = NULL; |
4034 | } | |
8fe23e05 DR |
4035 | } |
4036 | ||
1da177e4 LT |
4037 | static int __init kswapd_init(void) |
4038 | { | |
517bbed9 | 4039 | int nid, ret; |
69e05944 | 4040 | |
1da177e4 | 4041 | swap_setup(); |
48fb2e24 | 4042 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 4043 | kswapd_run(nid); |
517bbed9 SAS |
4044 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
4045 | "mm/vmscan:online", kswapd_cpu_online, | |
4046 | NULL); | |
4047 | WARN_ON(ret < 0); | |
1da177e4 LT |
4048 | return 0; |
4049 | } | |
4050 | ||
4051 | module_init(kswapd_init) | |
9eeff239 CL |
4052 | |
4053 | #ifdef CONFIG_NUMA | |
4054 | /* | |
a5f5f91d | 4055 | * Node reclaim mode |
9eeff239 | 4056 | * |
a5f5f91d | 4057 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 4058 | * the watermarks. |
9eeff239 | 4059 | */ |
a5f5f91d | 4060 | int node_reclaim_mode __read_mostly; |
9eeff239 | 4061 | |
1b2ffb78 | 4062 | #define RECLAIM_OFF 0 |
7d03431c | 4063 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 | 4064 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
95bbc0c7 | 4065 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
1b2ffb78 | 4066 | |
a92f7126 | 4067 | /* |
a5f5f91d | 4068 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
4069 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
4070 | * a zone. | |
4071 | */ | |
a5f5f91d | 4072 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 4073 | |
9614634f | 4074 | /* |
a5f5f91d | 4075 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
4076 | * occur. |
4077 | */ | |
4078 | int sysctl_min_unmapped_ratio = 1; | |
4079 | ||
0ff38490 CL |
4080 | /* |
4081 | * If the number of slab pages in a zone grows beyond this percentage then | |
4082 | * slab reclaim needs to occur. | |
4083 | */ | |
4084 | int sysctl_min_slab_ratio = 5; | |
4085 | ||
11fb9989 | 4086 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 4087 | { |
11fb9989 MG |
4088 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
4089 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
4090 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
4091 | |
4092 | /* | |
4093 | * It's possible for there to be more file mapped pages than | |
4094 | * accounted for by the pages on the file LRU lists because | |
4095 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
4096 | */ | |
4097 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
4098 | } | |
4099 | ||
4100 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 4101 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 4102 | { |
d031a157 AM |
4103 | unsigned long nr_pagecache_reclaimable; |
4104 | unsigned long delta = 0; | |
90afa5de MG |
4105 | |
4106 | /* | |
95bbc0c7 | 4107 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 4108 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 4109 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
4110 | * a better estimate |
4111 | */ | |
a5f5f91d MG |
4112 | if (node_reclaim_mode & RECLAIM_UNMAP) |
4113 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 4114 | else |
a5f5f91d | 4115 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
4116 | |
4117 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
4118 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
4119 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
4120 | |
4121 | /* Watch for any possible underflows due to delta */ | |
4122 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
4123 | delta = nr_pagecache_reclaimable; | |
4124 | ||
4125 | return nr_pagecache_reclaimable - delta; | |
4126 | } | |
4127 | ||
9eeff239 | 4128 | /* |
a5f5f91d | 4129 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 4130 | */ |
a5f5f91d | 4131 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 4132 | { |
7fb2d46d | 4133 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 4134 | const unsigned long nr_pages = 1 << order; |
9eeff239 | 4135 | struct task_struct *p = current; |
499118e9 | 4136 | unsigned int noreclaim_flag; |
179e9639 | 4137 | struct scan_control sc = { |
62b726c1 | 4138 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
f2f43e56 | 4139 | .gfp_mask = current_gfp_context(gfp_mask), |
bd2f6199 | 4140 | .order = order, |
a5f5f91d MG |
4141 | .priority = NODE_RECLAIM_PRIORITY, |
4142 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
4143 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 4144 | .may_swap = 1, |
f2f43e56 | 4145 | .reclaim_idx = gfp_zone(gfp_mask), |
179e9639 | 4146 | }; |
9eeff239 | 4147 | |
132bb8cf YS |
4148 | trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order, |
4149 | sc.gfp_mask); | |
4150 | ||
9eeff239 | 4151 | cond_resched(); |
93781325 | 4152 | fs_reclaim_acquire(sc.gfp_mask); |
d4f7796e | 4153 | /* |
95bbc0c7 | 4154 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 4155 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 4156 | * and RECLAIM_UNMAP. |
d4f7796e | 4157 | */ |
499118e9 VB |
4158 | noreclaim_flag = memalloc_noreclaim_save(); |
4159 | p->flags |= PF_SWAPWRITE; | |
1732d2b0 | 4160 | set_task_reclaim_state(p, &sc.reclaim_state); |
c84db23c | 4161 | |
a5f5f91d | 4162 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
0ff38490 | 4163 | /* |
894befec | 4164 | * Free memory by calling shrink node with increasing |
0ff38490 CL |
4165 | * priorities until we have enough memory freed. |
4166 | */ | |
0ff38490 | 4167 | do { |
970a39a3 | 4168 | shrink_node(pgdat, &sc); |
9e3b2f8c | 4169 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 4170 | } |
c84db23c | 4171 | |
1732d2b0 | 4172 | set_task_reclaim_state(p, NULL); |
499118e9 VB |
4173 | current->flags &= ~PF_SWAPWRITE; |
4174 | memalloc_noreclaim_restore(noreclaim_flag); | |
93781325 | 4175 | fs_reclaim_release(sc.gfp_mask); |
132bb8cf YS |
4176 | |
4177 | trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed); | |
4178 | ||
a79311c1 | 4179 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 4180 | } |
179e9639 | 4181 | |
a5f5f91d | 4182 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 4183 | { |
d773ed6b | 4184 | int ret; |
179e9639 AM |
4185 | |
4186 | /* | |
a5f5f91d | 4187 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 4188 | * slab pages if we are over the defined limits. |
34aa1330 | 4189 | * |
9614634f CL |
4190 | * A small portion of unmapped file backed pages is needed for |
4191 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
4192 | * thrown out if the node is overallocated. So we do not reclaim |
4193 | * if less than a specified percentage of the node is used by | |
9614634f | 4194 | * unmapped file backed pages. |
179e9639 | 4195 | */ |
a5f5f91d | 4196 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
385386cf | 4197 | node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) |
a5f5f91d | 4198 | return NODE_RECLAIM_FULL; |
179e9639 AM |
4199 | |
4200 | /* | |
d773ed6b | 4201 | * Do not scan if the allocation should not be delayed. |
179e9639 | 4202 | */ |
d0164adc | 4203 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 4204 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
4205 | |
4206 | /* | |
a5f5f91d | 4207 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
4208 | * have associated processors. This will favor the local processor |
4209 | * over remote processors and spread off node memory allocations | |
4210 | * as wide as possible. | |
4211 | */ | |
a5f5f91d MG |
4212 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
4213 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 4214 | |
a5f5f91d MG |
4215 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
4216 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 4217 | |
a5f5f91d MG |
4218 | ret = __node_reclaim(pgdat, gfp_mask, order); |
4219 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 4220 | |
24cf7251 MG |
4221 | if (!ret) |
4222 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
4223 | ||
d773ed6b | 4224 | return ret; |
179e9639 | 4225 | } |
9eeff239 | 4226 | #endif |
894bc310 | 4227 | |
894bc310 LS |
4228 | /* |
4229 | * page_evictable - test whether a page is evictable | |
4230 | * @page: the page to test | |
894bc310 LS |
4231 | * |
4232 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 4233 | * lists vs unevictable list. |
894bc310 LS |
4234 | * |
4235 | * Reasons page might not be evictable: | |
ba9ddf49 | 4236 | * (1) page's mapping marked unevictable |
b291f000 | 4237 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 4238 | * |
894bc310 | 4239 | */ |
39b5f29a | 4240 | int page_evictable(struct page *page) |
894bc310 | 4241 | { |
e92bb4dd HY |
4242 | int ret; |
4243 | ||
4244 | /* Prevent address_space of inode and swap cache from being freed */ | |
4245 | rcu_read_lock(); | |
4246 | ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); | |
4247 | rcu_read_unlock(); | |
4248 | return ret; | |
894bc310 | 4249 | } |
89e004ea LS |
4250 | |
4251 | /** | |
64e3d12f KHY |
4252 | * check_move_unevictable_pages - check pages for evictability and move to |
4253 | * appropriate zone lru list | |
4254 | * @pvec: pagevec with lru pages to check | |
89e004ea | 4255 | * |
64e3d12f KHY |
4256 | * Checks pages for evictability, if an evictable page is in the unevictable |
4257 | * lru list, moves it to the appropriate evictable lru list. This function | |
4258 | * should be only used for lru pages. | |
89e004ea | 4259 | */ |
64e3d12f | 4260 | void check_move_unevictable_pages(struct pagevec *pvec) |
89e004ea | 4261 | { |
925b7673 | 4262 | struct lruvec *lruvec; |
785b99fe | 4263 | struct pglist_data *pgdat = NULL; |
24513264 HD |
4264 | int pgscanned = 0; |
4265 | int pgrescued = 0; | |
4266 | int i; | |
89e004ea | 4267 | |
64e3d12f KHY |
4268 | for (i = 0; i < pvec->nr; i++) { |
4269 | struct page *page = pvec->pages[i]; | |
785b99fe | 4270 | struct pglist_data *pagepgdat = page_pgdat(page); |
89e004ea | 4271 | |
24513264 | 4272 | pgscanned++; |
785b99fe MG |
4273 | if (pagepgdat != pgdat) { |
4274 | if (pgdat) | |
4275 | spin_unlock_irq(&pgdat->lru_lock); | |
4276 | pgdat = pagepgdat; | |
4277 | spin_lock_irq(&pgdat->lru_lock); | |
24513264 | 4278 | } |
785b99fe | 4279 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
89e004ea | 4280 | |
24513264 HD |
4281 | if (!PageLRU(page) || !PageUnevictable(page)) |
4282 | continue; | |
89e004ea | 4283 | |
39b5f29a | 4284 | if (page_evictable(page)) { |
24513264 HD |
4285 | enum lru_list lru = page_lru_base_type(page); |
4286 | ||
309381fe | 4287 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 4288 | ClearPageUnevictable(page); |
fa9add64 HD |
4289 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
4290 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 4291 | pgrescued++; |
89e004ea | 4292 | } |
24513264 | 4293 | } |
89e004ea | 4294 | |
785b99fe | 4295 | if (pgdat) { |
24513264 HD |
4296 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
4297 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
785b99fe | 4298 | spin_unlock_irq(&pgdat->lru_lock); |
89e004ea | 4299 | } |
89e004ea | 4300 | } |
64e3d12f | 4301 | EXPORT_SYMBOL_GPL(check_move_unevictable_pages); |