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