]>
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 ([email protected]). | |
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 | |
b91ac374 JW |
82 | /* Can active pages be deactivated as part of reclaim? */ |
83 | #define DEACTIVATE_ANON 1 | |
84 | #define DEACTIVATE_FILE 2 | |
85 | unsigned int may_deactivate:2; | |
86 | unsigned int force_deactivate:1; | |
87 | unsigned int skipped_deactivate:1; | |
88 | ||
1276ad68 | 89 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
ee814fe2 JW |
90 | unsigned int may_writepage:1; |
91 | ||
92 | /* Can mapped pages be reclaimed? */ | |
93 | unsigned int may_unmap:1; | |
94 | ||
95 | /* Can pages be swapped as part of reclaim? */ | |
96 | unsigned int may_swap:1; | |
97 | ||
d6622f63 YX |
98 | /* |
99 | * Cgroups are not reclaimed below their configured memory.low, | |
100 | * unless we threaten to OOM. If any cgroups are skipped due to | |
101 | * memory.low and nothing was reclaimed, go back for memory.low. | |
102 | */ | |
103 | unsigned int memcg_low_reclaim:1; | |
104 | unsigned int memcg_low_skipped:1; | |
241994ed | 105 | |
ee814fe2 JW |
106 | unsigned int hibernation_mode:1; |
107 | ||
108 | /* One of the zones is ready for compaction */ | |
109 | unsigned int compaction_ready:1; | |
110 | ||
b91ac374 JW |
111 | /* There is easily reclaimable cold cache in the current node */ |
112 | unsigned int cache_trim_mode:1; | |
113 | ||
53138cea JW |
114 | /* The file pages on the current node are dangerously low */ |
115 | unsigned int file_is_tiny:1; | |
116 | ||
bb451fdf GT |
117 | /* Allocation order */ |
118 | s8 order; | |
119 | ||
120 | /* Scan (total_size >> priority) pages at once */ | |
121 | s8 priority; | |
122 | ||
123 | /* The highest zone to isolate pages for reclaim from */ | |
124 | s8 reclaim_idx; | |
125 | ||
126 | /* This context's GFP mask */ | |
127 | gfp_t gfp_mask; | |
128 | ||
ee814fe2 JW |
129 | /* Incremented by the number of inactive pages that were scanned */ |
130 | unsigned long nr_scanned; | |
131 | ||
132 | /* Number of pages freed so far during a call to shrink_zones() */ | |
133 | unsigned long nr_reclaimed; | |
d108c772 AR |
134 | |
135 | struct { | |
136 | unsigned int dirty; | |
137 | unsigned int unqueued_dirty; | |
138 | unsigned int congested; | |
139 | unsigned int writeback; | |
140 | unsigned int immediate; | |
141 | unsigned int file_taken; | |
142 | unsigned int taken; | |
143 | } nr; | |
e5ca8071 YS |
144 | |
145 | /* for recording the reclaimed slab by now */ | |
146 | struct reclaim_state reclaim_state; | |
1da177e4 LT |
147 | }; |
148 | ||
1da177e4 LT |
149 | #ifdef ARCH_HAS_PREFETCH |
150 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
151 | do { \ | |
152 | if ((_page)->lru.prev != _base) { \ | |
153 | struct page *prev; \ | |
154 | \ | |
155 | prev = lru_to_page(&(_page->lru)); \ | |
156 | prefetch(&prev->_field); \ | |
157 | } \ | |
158 | } while (0) | |
159 | #else | |
160 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
161 | #endif | |
162 | ||
163 | #ifdef ARCH_HAS_PREFETCHW | |
164 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
165 | do { \ | |
166 | if ((_page)->lru.prev != _base) { \ | |
167 | struct page *prev; \ | |
168 | \ | |
169 | prev = lru_to_page(&(_page->lru)); \ | |
170 | prefetchw(&prev->_field); \ | |
171 | } \ | |
172 | } while (0) | |
173 | #else | |
174 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
175 | #endif | |
176 | ||
177 | /* | |
178 | * From 0 .. 100. Higher means more swappy. | |
179 | */ | |
180 | int vm_swappiness = 60; | |
d0480be4 WSH |
181 | /* |
182 | * The total number of pages which are beyond the high watermark within all | |
183 | * zones. | |
184 | */ | |
185 | unsigned long vm_total_pages; | |
1da177e4 | 186 | |
0a432dcb YS |
187 | static void set_task_reclaim_state(struct task_struct *task, |
188 | struct reclaim_state *rs) | |
189 | { | |
190 | /* Check for an overwrite */ | |
191 | WARN_ON_ONCE(rs && task->reclaim_state); | |
192 | ||
193 | /* Check for the nulling of an already-nulled member */ | |
194 | WARN_ON_ONCE(!rs && !task->reclaim_state); | |
195 | ||
196 | task->reclaim_state = rs; | |
197 | } | |
198 | ||
1da177e4 LT |
199 | static LIST_HEAD(shrinker_list); |
200 | static DECLARE_RWSEM(shrinker_rwsem); | |
201 | ||
0a432dcb | 202 | #ifdef CONFIG_MEMCG |
7e010df5 KT |
203 | /* |
204 | * We allow subsystems to populate their shrinker-related | |
205 | * LRU lists before register_shrinker_prepared() is called | |
206 | * for the shrinker, since we don't want to impose | |
207 | * restrictions on their internal registration order. | |
208 | * In this case shrink_slab_memcg() may find corresponding | |
209 | * bit is set in the shrinkers map. | |
210 | * | |
211 | * This value is used by the function to detect registering | |
212 | * shrinkers and to skip do_shrink_slab() calls for them. | |
213 | */ | |
214 | #define SHRINKER_REGISTERING ((struct shrinker *)~0UL) | |
215 | ||
b4c2b231 KT |
216 | static DEFINE_IDR(shrinker_idr); |
217 | static int shrinker_nr_max; | |
218 | ||
219 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) | |
220 | { | |
221 | int id, ret = -ENOMEM; | |
222 | ||
223 | down_write(&shrinker_rwsem); | |
224 | /* This may call shrinker, so it must use down_read_trylock() */ | |
7e010df5 | 225 | id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL); |
b4c2b231 KT |
226 | if (id < 0) |
227 | goto unlock; | |
228 | ||
0a4465d3 KT |
229 | if (id >= shrinker_nr_max) { |
230 | if (memcg_expand_shrinker_maps(id)) { | |
231 | idr_remove(&shrinker_idr, id); | |
232 | goto unlock; | |
233 | } | |
234 | ||
b4c2b231 | 235 | shrinker_nr_max = id + 1; |
0a4465d3 | 236 | } |
b4c2b231 KT |
237 | shrinker->id = id; |
238 | ret = 0; | |
239 | unlock: | |
240 | up_write(&shrinker_rwsem); | |
241 | return ret; | |
242 | } | |
243 | ||
244 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
245 | { | |
246 | int id = shrinker->id; | |
247 | ||
248 | BUG_ON(id < 0); | |
249 | ||
250 | down_write(&shrinker_rwsem); | |
251 | idr_remove(&shrinker_idr, id); | |
252 | up_write(&shrinker_rwsem); | |
253 | } | |
b4c2b231 | 254 | |
b5ead35e | 255 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 256 | { |
b5ead35e | 257 | return sc->target_mem_cgroup; |
89b5fae5 | 258 | } |
97c9341f TH |
259 | |
260 | /** | |
b5ead35e | 261 | * writeback_throttling_sane - is the usual dirty throttling mechanism available? |
97c9341f TH |
262 | * @sc: scan_control in question |
263 | * | |
264 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
265 | * completely broken with the legacy memcg and direct stalling in | |
266 | * shrink_page_list() is used for throttling instead, which lacks all the | |
267 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
268 | * allocation and configurability. | |
269 | * | |
270 | * This function tests whether the vmscan currently in progress can assume | |
271 | * that the normal dirty throttling mechanism is operational. | |
272 | */ | |
b5ead35e | 273 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f | 274 | { |
b5ead35e | 275 | if (!cgroup_reclaim(sc)) |
97c9341f TH |
276 | return true; |
277 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 278 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
279 | return true; |
280 | #endif | |
281 | return false; | |
282 | } | |
91a45470 | 283 | #else |
0a432dcb YS |
284 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) |
285 | { | |
286 | return 0; | |
287 | } | |
288 | ||
289 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
290 | { | |
291 | } | |
292 | ||
b5ead35e | 293 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 294 | { |
b5ead35e | 295 | return false; |
89b5fae5 | 296 | } |
97c9341f | 297 | |
b5ead35e | 298 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f TH |
299 | { |
300 | return true; | |
301 | } | |
91a45470 KH |
302 | #endif |
303 | ||
5a1c84b4 MG |
304 | /* |
305 | * This misses isolated pages which are not accounted for to save counters. | |
306 | * As the data only determines if reclaim or compaction continues, it is | |
307 | * not expected that isolated pages will be a dominating factor. | |
308 | */ | |
309 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
310 | { | |
311 | unsigned long nr; | |
312 | ||
313 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
314 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
315 | if (get_nr_swap_pages() > 0) | |
316 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + | |
317 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
318 | ||
319 | return nr; | |
320 | } | |
321 | ||
fd538803 MH |
322 | /** |
323 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
324 | * @lruvec: lru vector | |
325 | * @lru: lru to use | |
326 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) | |
327 | */ | |
328 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) | |
c9f299d9 | 329 | { |
de3b0150 | 330 | unsigned long size = 0; |
fd538803 MH |
331 | int zid; |
332 | ||
de3b0150 | 333 | for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) { |
fd538803 | 334 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; |
c9f299d9 | 335 | |
fd538803 MH |
336 | if (!managed_zone(zone)) |
337 | continue; | |
338 | ||
339 | if (!mem_cgroup_disabled()) | |
de3b0150 | 340 | size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid); |
fd538803 | 341 | else |
de3b0150 | 342 | size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru); |
fd538803 | 343 | } |
de3b0150 | 344 | return size; |
b4536f0c MH |
345 | } |
346 | ||
1da177e4 | 347 | /* |
1d3d4437 | 348 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 349 | */ |
8e04944f | 350 | int prealloc_shrinker(struct shrinker *shrinker) |
1da177e4 | 351 | { |
b9726c26 | 352 | unsigned int size = sizeof(*shrinker->nr_deferred); |
1d3d4437 | 353 | |
1d3d4437 GC |
354 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
355 | size *= nr_node_ids; | |
356 | ||
357 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
358 | if (!shrinker->nr_deferred) | |
359 | return -ENOMEM; | |
b4c2b231 KT |
360 | |
361 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) { | |
362 | if (prealloc_memcg_shrinker(shrinker)) | |
363 | goto free_deferred; | |
364 | } | |
365 | ||
8e04944f | 366 | return 0; |
b4c2b231 KT |
367 | |
368 | free_deferred: | |
369 | kfree(shrinker->nr_deferred); | |
370 | shrinker->nr_deferred = NULL; | |
371 | return -ENOMEM; | |
8e04944f TH |
372 | } |
373 | ||
374 | void free_prealloced_shrinker(struct shrinker *shrinker) | |
375 | { | |
b4c2b231 KT |
376 | if (!shrinker->nr_deferred) |
377 | return; | |
378 | ||
379 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) | |
380 | unregister_memcg_shrinker(shrinker); | |
381 | ||
8e04944f TH |
382 | kfree(shrinker->nr_deferred); |
383 | shrinker->nr_deferred = NULL; | |
384 | } | |
1d3d4437 | 385 | |
8e04944f TH |
386 | void register_shrinker_prepared(struct shrinker *shrinker) |
387 | { | |
8e1f936b RR |
388 | down_write(&shrinker_rwsem); |
389 | list_add_tail(&shrinker->list, &shrinker_list); | |
42a9a53b | 390 | #ifdef CONFIG_MEMCG |
8df4a44c KT |
391 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) |
392 | idr_replace(&shrinker_idr, shrinker, shrinker->id); | |
7e010df5 | 393 | #endif |
8e1f936b | 394 | up_write(&shrinker_rwsem); |
8e04944f TH |
395 | } |
396 | ||
397 | int register_shrinker(struct shrinker *shrinker) | |
398 | { | |
399 | int err = prealloc_shrinker(shrinker); | |
400 | ||
401 | if (err) | |
402 | return err; | |
403 | register_shrinker_prepared(shrinker); | |
1d3d4437 | 404 | return 0; |
1da177e4 | 405 | } |
8e1f936b | 406 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
407 | |
408 | /* | |
409 | * Remove one | |
410 | */ | |
8e1f936b | 411 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 | 412 | { |
bb422a73 TH |
413 | if (!shrinker->nr_deferred) |
414 | return; | |
b4c2b231 KT |
415 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) |
416 | unregister_memcg_shrinker(shrinker); | |
1da177e4 LT |
417 | down_write(&shrinker_rwsem); |
418 | list_del(&shrinker->list); | |
419 | up_write(&shrinker_rwsem); | |
ae393321 | 420 | kfree(shrinker->nr_deferred); |
bb422a73 | 421 | shrinker->nr_deferred = NULL; |
1da177e4 | 422 | } |
8e1f936b | 423 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
424 | |
425 | #define SHRINK_BATCH 128 | |
1d3d4437 | 426 | |
cb731d6c | 427 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
9092c71b | 428 | struct shrinker *shrinker, int priority) |
1d3d4437 GC |
429 | { |
430 | unsigned long freed = 0; | |
431 | unsigned long long delta; | |
432 | long total_scan; | |
d5bc5fd3 | 433 | long freeable; |
1d3d4437 GC |
434 | long nr; |
435 | long new_nr; | |
436 | int nid = shrinkctl->nid; | |
437 | long batch_size = shrinker->batch ? shrinker->batch | |
438 | : SHRINK_BATCH; | |
5f33a080 | 439 | long scanned = 0, next_deferred; |
1d3d4437 | 440 | |
ac7fb3ad KT |
441 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
442 | nid = 0; | |
443 | ||
d5bc5fd3 | 444 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
9b996468 KT |
445 | if (freeable == 0 || freeable == SHRINK_EMPTY) |
446 | return freeable; | |
1d3d4437 GC |
447 | |
448 | /* | |
449 | * copy the current shrinker scan count into a local variable | |
450 | * and zero it so that other concurrent shrinker invocations | |
451 | * don't also do this scanning work. | |
452 | */ | |
453 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
454 | ||
455 | total_scan = nr; | |
4b85afbd JW |
456 | if (shrinker->seeks) { |
457 | delta = freeable >> priority; | |
458 | delta *= 4; | |
459 | do_div(delta, shrinker->seeks); | |
460 | } else { | |
461 | /* | |
462 | * These objects don't require any IO to create. Trim | |
463 | * them aggressively under memory pressure to keep | |
464 | * them from causing refetches in the IO caches. | |
465 | */ | |
466 | delta = freeable / 2; | |
467 | } | |
172b06c3 | 468 | |
1d3d4437 GC |
469 | total_scan += delta; |
470 | if (total_scan < 0) { | |
d75f773c | 471 | pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n", |
a0b02131 | 472 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 473 | total_scan = freeable; |
5f33a080 SL |
474 | next_deferred = nr; |
475 | } else | |
476 | next_deferred = total_scan; | |
1d3d4437 GC |
477 | |
478 | /* | |
479 | * We need to avoid excessive windup on filesystem shrinkers | |
480 | * due to large numbers of GFP_NOFS allocations causing the | |
481 | * shrinkers to return -1 all the time. This results in a large | |
482 | * nr being built up so when a shrink that can do some work | |
483 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 484 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
485 | * memory. |
486 | * | |
487 | * Hence only allow the shrinker to scan the entire cache when | |
488 | * a large delta change is calculated directly. | |
489 | */ | |
d5bc5fd3 VD |
490 | if (delta < freeable / 4) |
491 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
492 | |
493 | /* | |
494 | * Avoid risking looping forever due to too large nr value: | |
495 | * never try to free more than twice the estimate number of | |
496 | * freeable entries. | |
497 | */ | |
d5bc5fd3 VD |
498 | if (total_scan > freeable * 2) |
499 | total_scan = freeable * 2; | |
1d3d4437 GC |
500 | |
501 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
9092c71b | 502 | freeable, delta, total_scan, priority); |
1d3d4437 | 503 | |
0b1fb40a VD |
504 | /* |
505 | * Normally, we should not scan less than batch_size objects in one | |
506 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
507 | * than batch_size objects in total and we are really tight on memory, | |
508 | * we will try to reclaim all available objects, otherwise we can end | |
509 | * up failing allocations although there are plenty of reclaimable | |
510 | * objects spread over several slabs with usage less than the | |
511 | * batch_size. | |
512 | * | |
513 | * We detect the "tight on memory" situations by looking at the total | |
514 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 515 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
516 | * scanning at high prio and therefore should try to reclaim as much as |
517 | * possible. | |
518 | */ | |
519 | while (total_scan >= batch_size || | |
d5bc5fd3 | 520 | total_scan >= freeable) { |
a0b02131 | 521 | unsigned long ret; |
0b1fb40a | 522 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 523 | |
0b1fb40a | 524 | shrinkctl->nr_to_scan = nr_to_scan; |
d460acb5 | 525 | shrinkctl->nr_scanned = nr_to_scan; |
a0b02131 DC |
526 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
527 | if (ret == SHRINK_STOP) | |
528 | break; | |
529 | freed += ret; | |
1d3d4437 | 530 | |
d460acb5 CW |
531 | count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); |
532 | total_scan -= shrinkctl->nr_scanned; | |
533 | scanned += shrinkctl->nr_scanned; | |
1d3d4437 GC |
534 | |
535 | cond_resched(); | |
536 | } | |
537 | ||
5f33a080 SL |
538 | if (next_deferred >= scanned) |
539 | next_deferred -= scanned; | |
540 | else | |
541 | next_deferred = 0; | |
1d3d4437 GC |
542 | /* |
543 | * move the unused scan count back into the shrinker in a | |
544 | * manner that handles concurrent updates. If we exhausted the | |
545 | * scan, there is no need to do an update. | |
546 | */ | |
5f33a080 SL |
547 | if (next_deferred > 0) |
548 | new_nr = atomic_long_add_return(next_deferred, | |
1d3d4437 GC |
549 | &shrinker->nr_deferred[nid]); |
550 | else | |
551 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
552 | ||
df9024a8 | 553 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 554 | return freed; |
1495f230 YH |
555 | } |
556 | ||
0a432dcb | 557 | #ifdef CONFIG_MEMCG |
b0dedc49 KT |
558 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
559 | struct mem_cgroup *memcg, int priority) | |
560 | { | |
561 | struct memcg_shrinker_map *map; | |
b8e57efa KT |
562 | unsigned long ret, freed = 0; |
563 | int i; | |
b0dedc49 | 564 | |
0a432dcb | 565 | if (!mem_cgroup_online(memcg)) |
b0dedc49 KT |
566 | return 0; |
567 | ||
568 | if (!down_read_trylock(&shrinker_rwsem)) | |
569 | return 0; | |
570 | ||
571 | map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map, | |
572 | true); | |
573 | if (unlikely(!map)) | |
574 | goto unlock; | |
575 | ||
576 | for_each_set_bit(i, map->map, shrinker_nr_max) { | |
577 | struct shrink_control sc = { | |
578 | .gfp_mask = gfp_mask, | |
579 | .nid = nid, | |
580 | .memcg = memcg, | |
581 | }; | |
582 | struct shrinker *shrinker; | |
583 | ||
584 | shrinker = idr_find(&shrinker_idr, i); | |
7e010df5 KT |
585 | if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) { |
586 | if (!shrinker) | |
587 | clear_bit(i, map->map); | |
b0dedc49 KT |
588 | continue; |
589 | } | |
590 | ||
0a432dcb YS |
591 | /* Call non-slab shrinkers even though kmem is disabled */ |
592 | if (!memcg_kmem_enabled() && | |
593 | !(shrinker->flags & SHRINKER_NONSLAB)) | |
594 | continue; | |
595 | ||
b0dedc49 | 596 | ret = do_shrink_slab(&sc, shrinker, priority); |
f90280d6 KT |
597 | if (ret == SHRINK_EMPTY) { |
598 | clear_bit(i, map->map); | |
599 | /* | |
600 | * After the shrinker reported that it had no objects to | |
601 | * free, but before we cleared the corresponding bit in | |
602 | * the memcg shrinker map, a new object might have been | |
603 | * added. To make sure, we have the bit set in this | |
604 | * case, we invoke the shrinker one more time and reset | |
605 | * the bit if it reports that it is not empty anymore. | |
606 | * The memory barrier here pairs with the barrier in | |
607 | * memcg_set_shrinker_bit(): | |
608 | * | |
609 | * list_lru_add() shrink_slab_memcg() | |
610 | * list_add_tail() clear_bit() | |
611 | * <MB> <MB> | |
612 | * set_bit() do_shrink_slab() | |
613 | */ | |
614 | smp_mb__after_atomic(); | |
615 | ret = do_shrink_slab(&sc, shrinker, priority); | |
616 | if (ret == SHRINK_EMPTY) | |
617 | ret = 0; | |
618 | else | |
619 | memcg_set_shrinker_bit(memcg, nid, i); | |
620 | } | |
b0dedc49 KT |
621 | freed += ret; |
622 | ||
623 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
624 | freed = freed ? : 1; | |
625 | break; | |
626 | } | |
627 | } | |
628 | unlock: | |
629 | up_read(&shrinker_rwsem); | |
630 | return freed; | |
631 | } | |
0a432dcb | 632 | #else /* CONFIG_MEMCG */ |
b0dedc49 KT |
633 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
634 | struct mem_cgroup *memcg, int priority) | |
635 | { | |
636 | return 0; | |
637 | } | |
0a432dcb | 638 | #endif /* CONFIG_MEMCG */ |
b0dedc49 | 639 | |
6b4f7799 | 640 | /** |
cb731d6c | 641 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
642 | * @gfp_mask: allocation context |
643 | * @nid: node whose slab caches to target | |
cb731d6c | 644 | * @memcg: memory cgroup whose slab caches to target |
9092c71b | 645 | * @priority: the reclaim priority |
1da177e4 | 646 | * |
6b4f7799 | 647 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 648 | * |
6b4f7799 JW |
649 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
650 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 651 | * |
aeed1d32 VD |
652 | * @memcg specifies the memory cgroup to target. Unaware shrinkers |
653 | * are called only if it is the root cgroup. | |
cb731d6c | 654 | * |
9092c71b JB |
655 | * @priority is sc->priority, we take the number of objects and >> by priority |
656 | * in order to get the scan target. | |
b15e0905 | 657 | * |
6b4f7799 | 658 | * Returns the number of reclaimed slab objects. |
1da177e4 | 659 | */ |
cb731d6c VD |
660 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
661 | struct mem_cgroup *memcg, | |
9092c71b | 662 | int priority) |
1da177e4 | 663 | { |
b8e57efa | 664 | unsigned long ret, freed = 0; |
1da177e4 LT |
665 | struct shrinker *shrinker; |
666 | ||
fa1e512f YS |
667 | /* |
668 | * The root memcg might be allocated even though memcg is disabled | |
669 | * via "cgroup_disable=memory" boot parameter. This could make | |
670 | * mem_cgroup_is_root() return false, then just run memcg slab | |
671 | * shrink, but skip global shrink. This may result in premature | |
672 | * oom. | |
673 | */ | |
674 | if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg)) | |
b0dedc49 | 675 | return shrink_slab_memcg(gfp_mask, nid, memcg, priority); |
cb731d6c | 676 | |
e830c63a | 677 | if (!down_read_trylock(&shrinker_rwsem)) |
f06590bd | 678 | goto out; |
1da177e4 LT |
679 | |
680 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
681 | struct shrink_control sc = { |
682 | .gfp_mask = gfp_mask, | |
683 | .nid = nid, | |
cb731d6c | 684 | .memcg = memcg, |
6b4f7799 | 685 | }; |
ec97097b | 686 | |
9b996468 KT |
687 | ret = do_shrink_slab(&sc, shrinker, priority); |
688 | if (ret == SHRINK_EMPTY) | |
689 | ret = 0; | |
690 | freed += ret; | |
e496612c MK |
691 | /* |
692 | * Bail out if someone want to register a new shrinker to | |
693 | * prevent the regsitration from being stalled for long periods | |
694 | * by parallel ongoing shrinking. | |
695 | */ | |
696 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
697 | freed = freed ? : 1; | |
698 | break; | |
699 | } | |
1da177e4 | 700 | } |
6b4f7799 | 701 | |
1da177e4 | 702 | up_read(&shrinker_rwsem); |
f06590bd MK |
703 | out: |
704 | cond_resched(); | |
24f7c6b9 | 705 | return freed; |
1da177e4 LT |
706 | } |
707 | ||
cb731d6c VD |
708 | void drop_slab_node(int nid) |
709 | { | |
710 | unsigned long freed; | |
711 | ||
712 | do { | |
713 | struct mem_cgroup *memcg = NULL; | |
714 | ||
715 | freed = 0; | |
aeed1d32 | 716 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
cb731d6c | 717 | do { |
9092c71b | 718 | freed += shrink_slab(GFP_KERNEL, nid, memcg, 0); |
cb731d6c VD |
719 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); |
720 | } while (freed > 10); | |
721 | } | |
722 | ||
723 | void drop_slab(void) | |
724 | { | |
725 | int nid; | |
726 | ||
727 | for_each_online_node(nid) | |
728 | drop_slab_node(nid); | |
729 | } | |
730 | ||
1da177e4 LT |
731 | static inline int is_page_cache_freeable(struct page *page) |
732 | { | |
ceddc3a5 JW |
733 | /* |
734 | * A freeable page cache page is referenced only by the caller | |
67891fff MW |
735 | * that isolated the page, the page cache and optional buffer |
736 | * heads at page->private. | |
ceddc3a5 | 737 | */ |
67891fff | 738 | int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ? |
bd4c82c2 | 739 | HPAGE_PMD_NR : 1; |
67891fff | 740 | return page_count(page) - page_has_private(page) == 1 + page_cache_pins; |
1da177e4 LT |
741 | } |
742 | ||
cb16556d | 743 | static int may_write_to_inode(struct inode *inode) |
1da177e4 | 744 | { |
930d9152 | 745 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 746 | return 1; |
703c2708 | 747 | if (!inode_write_congested(inode)) |
1da177e4 | 748 | return 1; |
703c2708 | 749 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
750 | return 1; |
751 | return 0; | |
752 | } | |
753 | ||
754 | /* | |
755 | * We detected a synchronous write error writing a page out. Probably | |
756 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
757 | * fsync(), msync() or close(). | |
758 | * | |
759 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
760 | * prevents it from being freed up. But we have a ref on the page and once | |
761 | * that page is locked, the mapping is pinned. | |
762 | * | |
763 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
764 | * __GFP_FS. | |
765 | */ | |
766 | static void handle_write_error(struct address_space *mapping, | |
767 | struct page *page, int error) | |
768 | { | |
7eaceacc | 769 | lock_page(page); |
3e9f45bd GC |
770 | if (page_mapping(page) == mapping) |
771 | mapping_set_error(mapping, error); | |
1da177e4 LT |
772 | unlock_page(page); |
773 | } | |
774 | ||
04e62a29 CL |
775 | /* possible outcome of pageout() */ |
776 | typedef enum { | |
777 | /* failed to write page out, page is locked */ | |
778 | PAGE_KEEP, | |
779 | /* move page to the active list, page is locked */ | |
780 | PAGE_ACTIVATE, | |
781 | /* page has been sent to the disk successfully, page is unlocked */ | |
782 | PAGE_SUCCESS, | |
783 | /* page is clean and locked */ | |
784 | PAGE_CLEAN, | |
785 | } pageout_t; | |
786 | ||
1da177e4 | 787 | /* |
1742f19f AM |
788 | * pageout is called by shrink_page_list() for each dirty page. |
789 | * Calls ->writepage(). | |
1da177e4 | 790 | */ |
cb16556d | 791 | static pageout_t pageout(struct page *page, struct address_space *mapping) |
1da177e4 LT |
792 | { |
793 | /* | |
794 | * If the page is dirty, only perform writeback if that write | |
795 | * will be non-blocking. To prevent this allocation from being | |
796 | * stalled by pagecache activity. But note that there may be | |
797 | * stalls if we need to run get_block(). We could test | |
798 | * PagePrivate for that. | |
799 | * | |
8174202b | 800 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
801 | * this page's queue, we can perform writeback even if that |
802 | * will block. | |
803 | * | |
804 | * If the page is swapcache, write it back even if that would | |
805 | * block, for some throttling. This happens by accident, because | |
806 | * swap_backing_dev_info is bust: it doesn't reflect the | |
807 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
808 | */ |
809 | if (!is_page_cache_freeable(page)) | |
810 | return PAGE_KEEP; | |
811 | if (!mapping) { | |
812 | /* | |
813 | * Some data journaling orphaned pages can have | |
814 | * page->mapping == NULL while being dirty with clean buffers. | |
815 | */ | |
266cf658 | 816 | if (page_has_private(page)) { |
1da177e4 LT |
817 | if (try_to_free_buffers(page)) { |
818 | ClearPageDirty(page); | |
b1de0d13 | 819 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
820 | return PAGE_CLEAN; |
821 | } | |
822 | } | |
823 | return PAGE_KEEP; | |
824 | } | |
825 | if (mapping->a_ops->writepage == NULL) | |
826 | return PAGE_ACTIVATE; | |
cb16556d | 827 | if (!may_write_to_inode(mapping->host)) |
1da177e4 LT |
828 | return PAGE_KEEP; |
829 | ||
830 | if (clear_page_dirty_for_io(page)) { | |
831 | int res; | |
832 | struct writeback_control wbc = { | |
833 | .sync_mode = WB_SYNC_NONE, | |
834 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
835 | .range_start = 0, |
836 | .range_end = LLONG_MAX, | |
1da177e4 LT |
837 | .for_reclaim = 1, |
838 | }; | |
839 | ||
840 | SetPageReclaim(page); | |
841 | res = mapping->a_ops->writepage(page, &wbc); | |
842 | if (res < 0) | |
843 | handle_write_error(mapping, page, res); | |
994fc28c | 844 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
845 | ClearPageReclaim(page); |
846 | return PAGE_ACTIVATE; | |
847 | } | |
c661b078 | 848 | |
1da177e4 LT |
849 | if (!PageWriteback(page)) { |
850 | /* synchronous write or broken a_ops? */ | |
851 | ClearPageReclaim(page); | |
852 | } | |
3aa23851 | 853 | trace_mm_vmscan_writepage(page); |
c4a25635 | 854 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
855 | return PAGE_SUCCESS; |
856 | } | |
857 | ||
858 | return PAGE_CLEAN; | |
859 | } | |
860 | ||
a649fd92 | 861 | /* |
e286781d NP |
862 | * Same as remove_mapping, but if the page is removed from the mapping, it |
863 | * gets returned with a refcount of 0. | |
a649fd92 | 864 | */ |
a528910e | 865 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
b910718a | 866 | bool reclaimed, struct mem_cgroup *target_memcg) |
49d2e9cc | 867 | { |
c4843a75 | 868 | unsigned long flags; |
bd4c82c2 | 869 | int refcount; |
c4843a75 | 870 | |
28e4d965 NP |
871 | BUG_ON(!PageLocked(page)); |
872 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 873 | |
b93b0163 | 874 | xa_lock_irqsave(&mapping->i_pages, flags); |
49d2e9cc | 875 | /* |
0fd0e6b0 NP |
876 | * The non racy check for a busy page. |
877 | * | |
878 | * Must be careful with the order of the tests. When someone has | |
879 | * a ref to the page, it may be possible that they dirty it then | |
880 | * drop the reference. So if PageDirty is tested before page_count | |
881 | * here, then the following race may occur: | |
882 | * | |
883 | * get_user_pages(&page); | |
884 | * [user mapping goes away] | |
885 | * write_to(page); | |
886 | * !PageDirty(page) [good] | |
887 | * SetPageDirty(page); | |
888 | * put_page(page); | |
889 | * !page_count(page) [good, discard it] | |
890 | * | |
891 | * [oops, our write_to data is lost] | |
892 | * | |
893 | * Reversing the order of the tests ensures such a situation cannot | |
894 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 895 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
896 | * |
897 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
b93b0163 | 898 | * and thus under the i_pages lock, then this ordering is not required. |
49d2e9cc | 899 | */ |
906d278d | 900 | refcount = 1 + compound_nr(page); |
bd4c82c2 | 901 | if (!page_ref_freeze(page, refcount)) |
49d2e9cc | 902 | goto cannot_free; |
1c4c3b99 | 903 | /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */ |
e286781d | 904 | if (unlikely(PageDirty(page))) { |
bd4c82c2 | 905 | page_ref_unfreeze(page, refcount); |
49d2e9cc | 906 | goto cannot_free; |
e286781d | 907 | } |
49d2e9cc CL |
908 | |
909 | if (PageSwapCache(page)) { | |
910 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 911 | mem_cgroup_swapout(page, swap); |
4e17ec25 | 912 | __delete_from_swap_cache(page, swap); |
b93b0163 | 913 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
75f6d6d2 | 914 | put_swap_page(page, swap); |
e286781d | 915 | } else { |
6072d13c | 916 | void (*freepage)(struct page *); |
a528910e | 917 | void *shadow = NULL; |
6072d13c LT |
918 | |
919 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
920 | /* |
921 | * Remember a shadow entry for reclaimed file cache in | |
922 | * order to detect refaults, thus thrashing, later on. | |
923 | * | |
924 | * But don't store shadows in an address space that is | |
925 | * already exiting. This is not just an optizimation, | |
926 | * inode reclaim needs to empty out the radix tree or | |
927 | * the nodes are lost. Don't plant shadows behind its | |
928 | * back. | |
f9fe48be RZ |
929 | * |
930 | * We also don't store shadows for DAX mappings because the | |
931 | * only page cache pages found in these are zero pages | |
932 | * covering holes, and because we don't want to mix DAX | |
933 | * exceptional entries and shadow exceptional entries in the | |
b93b0163 | 934 | * same address_space. |
a528910e JW |
935 | */ |
936 | if (reclaimed && page_is_file_cache(page) && | |
f9fe48be | 937 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
b910718a | 938 | shadow = workingset_eviction(page, target_memcg); |
62cccb8c | 939 | __delete_from_page_cache(page, shadow); |
b93b0163 | 940 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
6072d13c LT |
941 | |
942 | if (freepage != NULL) | |
943 | freepage(page); | |
49d2e9cc CL |
944 | } |
945 | ||
49d2e9cc CL |
946 | return 1; |
947 | ||
948 | cannot_free: | |
b93b0163 | 949 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
49d2e9cc CL |
950 | return 0; |
951 | } | |
952 | ||
e286781d NP |
953 | /* |
954 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
955 | * someone else has a ref on the page, abort and return 0. If it was | |
956 | * successfully detached, return 1. Assumes the caller has a single ref on | |
957 | * this page. | |
958 | */ | |
959 | int remove_mapping(struct address_space *mapping, struct page *page) | |
960 | { | |
b910718a | 961 | if (__remove_mapping(mapping, page, false, NULL)) { |
e286781d NP |
962 | /* |
963 | * Unfreezing the refcount with 1 rather than 2 effectively | |
964 | * drops the pagecache ref for us without requiring another | |
965 | * atomic operation. | |
966 | */ | |
fe896d18 | 967 | page_ref_unfreeze(page, 1); |
e286781d NP |
968 | return 1; |
969 | } | |
970 | return 0; | |
971 | } | |
972 | ||
894bc310 LS |
973 | /** |
974 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
975 | * @page: page to be put back to appropriate lru list | |
976 | * | |
977 | * Add previously isolated @page to appropriate LRU list. | |
978 | * Page may still be unevictable for other reasons. | |
979 | * | |
980 | * lru_lock must not be held, interrupts must be enabled. | |
981 | */ | |
894bc310 LS |
982 | void putback_lru_page(struct page *page) |
983 | { | |
9c4e6b1a | 984 | lru_cache_add(page); |
894bc310 LS |
985 | put_page(page); /* drop ref from isolate */ |
986 | } | |
987 | ||
dfc8d636 JW |
988 | enum page_references { |
989 | PAGEREF_RECLAIM, | |
990 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 991 | PAGEREF_KEEP, |
dfc8d636 JW |
992 | PAGEREF_ACTIVATE, |
993 | }; | |
994 | ||
995 | static enum page_references page_check_references(struct page *page, | |
996 | struct scan_control *sc) | |
997 | { | |
64574746 | 998 | int referenced_ptes, referenced_page; |
dfc8d636 | 999 | unsigned long vm_flags; |
dfc8d636 | 1000 | |
c3ac9a8a JW |
1001 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
1002 | &vm_flags); | |
64574746 | 1003 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 1004 | |
dfc8d636 JW |
1005 | /* |
1006 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
1007 | * move the page to the unevictable list. | |
1008 | */ | |
1009 | if (vm_flags & VM_LOCKED) | |
1010 | return PAGEREF_RECLAIM; | |
1011 | ||
64574746 | 1012 | if (referenced_ptes) { |
e4898273 | 1013 | if (PageSwapBacked(page)) |
64574746 JW |
1014 | return PAGEREF_ACTIVATE; |
1015 | /* | |
1016 | * All mapped pages start out with page table | |
1017 | * references from the instantiating fault, so we need | |
1018 | * to look twice if a mapped file page is used more | |
1019 | * than once. | |
1020 | * | |
1021 | * Mark it and spare it for another trip around the | |
1022 | * inactive list. Another page table reference will | |
1023 | * lead to its activation. | |
1024 | * | |
1025 | * Note: the mark is set for activated pages as well | |
1026 | * so that recently deactivated but used pages are | |
1027 | * quickly recovered. | |
1028 | */ | |
1029 | SetPageReferenced(page); | |
1030 | ||
34dbc67a | 1031 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
1032 | return PAGEREF_ACTIVATE; |
1033 | ||
c909e993 KK |
1034 | /* |
1035 | * Activate file-backed executable pages after first usage. | |
1036 | */ | |
1037 | if (vm_flags & VM_EXEC) | |
1038 | return PAGEREF_ACTIVATE; | |
1039 | ||
64574746 JW |
1040 | return PAGEREF_KEEP; |
1041 | } | |
dfc8d636 JW |
1042 | |
1043 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 1044 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
1045 | return PAGEREF_RECLAIM_CLEAN; |
1046 | ||
1047 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
1048 | } |
1049 | ||
e2be15f6 MG |
1050 | /* Check if a page is dirty or under writeback */ |
1051 | static void page_check_dirty_writeback(struct page *page, | |
1052 | bool *dirty, bool *writeback) | |
1053 | { | |
b4597226 MG |
1054 | struct address_space *mapping; |
1055 | ||
e2be15f6 MG |
1056 | /* |
1057 | * Anonymous pages are not handled by flushers and must be written | |
1058 | * from reclaim context. Do not stall reclaim based on them | |
1059 | */ | |
802a3a92 SL |
1060 | if (!page_is_file_cache(page) || |
1061 | (PageAnon(page) && !PageSwapBacked(page))) { | |
e2be15f6 MG |
1062 | *dirty = false; |
1063 | *writeback = false; | |
1064 | return; | |
1065 | } | |
1066 | ||
1067 | /* By default assume that the page flags are accurate */ | |
1068 | *dirty = PageDirty(page); | |
1069 | *writeback = PageWriteback(page); | |
b4597226 MG |
1070 | |
1071 | /* Verify dirty/writeback state if the filesystem supports it */ | |
1072 | if (!page_has_private(page)) | |
1073 | return; | |
1074 | ||
1075 | mapping = page_mapping(page); | |
1076 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
1077 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
1078 | } |
1079 | ||
1da177e4 | 1080 | /* |
1742f19f | 1081 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 1082 | */ |
1742f19f | 1083 | static unsigned long shrink_page_list(struct list_head *page_list, |
599d0c95 | 1084 | struct pglist_data *pgdat, |
f84f6e2b | 1085 | struct scan_control *sc, |
02c6de8d | 1086 | enum ttu_flags ttu_flags, |
3c710c1a | 1087 | struct reclaim_stat *stat, |
8940b34a | 1088 | bool ignore_references) |
1da177e4 LT |
1089 | { |
1090 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 1091 | LIST_HEAD(free_pages); |
3c710c1a | 1092 | unsigned nr_reclaimed = 0; |
886cf190 | 1093 | unsigned pgactivate = 0; |
1da177e4 | 1094 | |
060f005f | 1095 | memset(stat, 0, sizeof(*stat)); |
1da177e4 LT |
1096 | cond_resched(); |
1097 | ||
1da177e4 LT |
1098 | while (!list_empty(page_list)) { |
1099 | struct address_space *mapping; | |
1100 | struct page *page; | |
1101 | int may_enter_fs; | |
8940b34a | 1102 | enum page_references references = PAGEREF_RECLAIM; |
e2be15f6 | 1103 | bool dirty, writeback; |
98879b3b | 1104 | unsigned int nr_pages; |
1da177e4 LT |
1105 | |
1106 | cond_resched(); | |
1107 | ||
1108 | page = lru_to_page(page_list); | |
1109 | list_del(&page->lru); | |
1110 | ||
529ae9aa | 1111 | if (!trylock_page(page)) |
1da177e4 LT |
1112 | goto keep; |
1113 | ||
309381fe | 1114 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 | 1115 | |
d8c6546b | 1116 | nr_pages = compound_nr(page); |
98879b3b YS |
1117 | |
1118 | /* Account the number of base pages even though THP */ | |
1119 | sc->nr_scanned += nr_pages; | |
80e43426 | 1120 | |
39b5f29a | 1121 | if (unlikely(!page_evictable(page))) |
ad6b6704 | 1122 | goto activate_locked; |
894bc310 | 1123 | |
a6dc60f8 | 1124 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
1125 | goto keep_locked; |
1126 | ||
c661b078 AW |
1127 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
1128 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
1129 | ||
e2be15f6 | 1130 | /* |
894befec | 1131 | * The number of dirty pages determines if a node is marked |
e2be15f6 MG |
1132 | * reclaim_congested which affects wait_iff_congested. kswapd |
1133 | * will stall and start writing pages if the tail of the LRU | |
1134 | * is all dirty unqueued pages. | |
1135 | */ | |
1136 | page_check_dirty_writeback(page, &dirty, &writeback); | |
1137 | if (dirty || writeback) | |
060f005f | 1138 | stat->nr_dirty++; |
e2be15f6 MG |
1139 | |
1140 | if (dirty && !writeback) | |
060f005f | 1141 | stat->nr_unqueued_dirty++; |
e2be15f6 | 1142 | |
d04e8acd MG |
1143 | /* |
1144 | * Treat this page as congested if the underlying BDI is or if | |
1145 | * pages are cycling through the LRU so quickly that the | |
1146 | * pages marked for immediate reclaim are making it to the | |
1147 | * end of the LRU a second time. | |
1148 | */ | |
e2be15f6 | 1149 | mapping = page_mapping(page); |
1da58ee2 | 1150 | if (((dirty || writeback) && mapping && |
703c2708 | 1151 | inode_write_congested(mapping->host)) || |
d04e8acd | 1152 | (writeback && PageReclaim(page))) |
060f005f | 1153 | stat->nr_congested++; |
e2be15f6 | 1154 | |
283aba9f MG |
1155 | /* |
1156 | * If a page at the tail of the LRU is under writeback, there | |
1157 | * are three cases to consider. | |
1158 | * | |
1159 | * 1) If reclaim is encountering an excessive number of pages | |
1160 | * under writeback and this page is both under writeback and | |
1161 | * PageReclaim then it indicates that pages are being queued | |
1162 | * for IO but are being recycled through the LRU before the | |
1163 | * IO can complete. Waiting on the page itself risks an | |
1164 | * indefinite stall if it is impossible to writeback the | |
1165 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
1166 | * note that the LRU is being scanned too quickly and the |
1167 | * caller can stall after page list has been processed. | |
283aba9f | 1168 | * |
97c9341f | 1169 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
1170 | * not marked for immediate reclaim, or the caller does not |
1171 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
1172 | * not to fs). In this case mark the page for immediate | |
97c9341f | 1173 | * reclaim and continue scanning. |
283aba9f | 1174 | * |
ecf5fc6e MH |
1175 | * Require may_enter_fs because we would wait on fs, which |
1176 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
1177 | * enter reclaim, and deadlock if it waits on a page for |
1178 | * which it is needed to do the write (loop masks off | |
1179 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1180 | * would probably show more reasons. | |
1181 | * | |
7fadc820 | 1182 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1183 | * PageReclaim. memcg does not have any dirty pages |
1184 | * throttling so we could easily OOM just because too many | |
1185 | * pages are in writeback and there is nothing else to | |
1186 | * reclaim. Wait for the writeback to complete. | |
c55e8d03 JW |
1187 | * |
1188 | * In cases 1) and 2) we activate the pages to get them out of | |
1189 | * the way while we continue scanning for clean pages on the | |
1190 | * inactive list and refilling from the active list. The | |
1191 | * observation here is that waiting for disk writes is more | |
1192 | * expensive than potentially causing reloads down the line. | |
1193 | * Since they're marked for immediate reclaim, they won't put | |
1194 | * memory pressure on the cache working set any longer than it | |
1195 | * takes to write them to disk. | |
283aba9f | 1196 | */ |
c661b078 | 1197 | if (PageWriteback(page)) { |
283aba9f MG |
1198 | /* Case 1 above */ |
1199 | if (current_is_kswapd() && | |
1200 | PageReclaim(page) && | |
599d0c95 | 1201 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
060f005f | 1202 | stat->nr_immediate++; |
c55e8d03 | 1203 | goto activate_locked; |
283aba9f MG |
1204 | |
1205 | /* Case 2 above */ | |
b5ead35e | 1206 | } else if (writeback_throttling_sane(sc) || |
ecf5fc6e | 1207 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1208 | /* |
1209 | * This is slightly racy - end_page_writeback() | |
1210 | * might have just cleared PageReclaim, then | |
1211 | * setting PageReclaim here end up interpreted | |
1212 | * as PageReadahead - but that does not matter | |
1213 | * enough to care. What we do want is for this | |
1214 | * page to have PageReclaim set next time memcg | |
1215 | * reclaim reaches the tests above, so it will | |
1216 | * then wait_on_page_writeback() to avoid OOM; | |
1217 | * and it's also appropriate in global reclaim. | |
1218 | */ | |
1219 | SetPageReclaim(page); | |
060f005f | 1220 | stat->nr_writeback++; |
c55e8d03 | 1221 | goto activate_locked; |
283aba9f MG |
1222 | |
1223 | /* Case 3 above */ | |
1224 | } else { | |
7fadc820 | 1225 | unlock_page(page); |
283aba9f | 1226 | wait_on_page_writeback(page); |
7fadc820 HD |
1227 | /* then go back and try same page again */ |
1228 | list_add_tail(&page->lru, page_list); | |
1229 | continue; | |
e62e384e | 1230 | } |
c661b078 | 1231 | } |
1da177e4 | 1232 | |
8940b34a | 1233 | if (!ignore_references) |
02c6de8d MK |
1234 | references = page_check_references(page, sc); |
1235 | ||
dfc8d636 JW |
1236 | switch (references) { |
1237 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1238 | goto activate_locked; |
64574746 | 1239 | case PAGEREF_KEEP: |
98879b3b | 1240 | stat->nr_ref_keep += nr_pages; |
64574746 | 1241 | goto keep_locked; |
dfc8d636 JW |
1242 | case PAGEREF_RECLAIM: |
1243 | case PAGEREF_RECLAIM_CLEAN: | |
1244 | ; /* try to reclaim the page below */ | |
1245 | } | |
1da177e4 | 1246 | |
1da177e4 LT |
1247 | /* |
1248 | * Anonymous process memory has backing store? | |
1249 | * Try to allocate it some swap space here. | |
802a3a92 | 1250 | * Lazyfree page could be freed directly |
1da177e4 | 1251 | */ |
bd4c82c2 YH |
1252 | if (PageAnon(page) && PageSwapBacked(page)) { |
1253 | if (!PageSwapCache(page)) { | |
1254 | if (!(sc->gfp_mask & __GFP_IO)) | |
1255 | goto keep_locked; | |
1256 | if (PageTransHuge(page)) { | |
1257 | /* cannot split THP, skip it */ | |
1258 | if (!can_split_huge_page(page, NULL)) | |
1259 | goto activate_locked; | |
1260 | /* | |
1261 | * Split pages without a PMD map right | |
1262 | * away. Chances are some or all of the | |
1263 | * tail pages can be freed without IO. | |
1264 | */ | |
1265 | if (!compound_mapcount(page) && | |
1266 | split_huge_page_to_list(page, | |
1267 | page_list)) | |
1268 | goto activate_locked; | |
1269 | } | |
1270 | if (!add_to_swap(page)) { | |
1271 | if (!PageTransHuge(page)) | |
98879b3b | 1272 | goto activate_locked_split; |
bd4c82c2 YH |
1273 | /* Fallback to swap normal pages */ |
1274 | if (split_huge_page_to_list(page, | |
1275 | page_list)) | |
1276 | goto activate_locked; | |
fe490cc0 YH |
1277 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1278 | count_vm_event(THP_SWPOUT_FALLBACK); | |
1279 | #endif | |
bd4c82c2 | 1280 | if (!add_to_swap(page)) |
98879b3b | 1281 | goto activate_locked_split; |
bd4c82c2 | 1282 | } |
0f074658 | 1283 | |
bd4c82c2 | 1284 | may_enter_fs = 1; |
1da177e4 | 1285 | |
bd4c82c2 YH |
1286 | /* Adding to swap updated mapping */ |
1287 | mapping = page_mapping(page); | |
1288 | } | |
7751b2da KS |
1289 | } else if (unlikely(PageTransHuge(page))) { |
1290 | /* Split file THP */ | |
1291 | if (split_huge_page_to_list(page, page_list)) | |
1292 | goto keep_locked; | |
e2be15f6 | 1293 | } |
1da177e4 | 1294 | |
98879b3b YS |
1295 | /* |
1296 | * THP may get split above, need minus tail pages and update | |
1297 | * nr_pages to avoid accounting tail pages twice. | |
1298 | * | |
1299 | * The tail pages that are added into swap cache successfully | |
1300 | * reach here. | |
1301 | */ | |
1302 | if ((nr_pages > 1) && !PageTransHuge(page)) { | |
1303 | sc->nr_scanned -= (nr_pages - 1); | |
1304 | nr_pages = 1; | |
1305 | } | |
1306 | ||
1da177e4 LT |
1307 | /* |
1308 | * The page is mapped into the page tables of one or more | |
1309 | * processes. Try to unmap it here. | |
1310 | */ | |
802a3a92 | 1311 | if (page_mapped(page)) { |
bd4c82c2 YH |
1312 | enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH; |
1313 | ||
1314 | if (unlikely(PageTransHuge(page))) | |
1315 | flags |= TTU_SPLIT_HUGE_PMD; | |
1316 | if (!try_to_unmap(page, flags)) { | |
98879b3b | 1317 | stat->nr_unmap_fail += nr_pages; |
1da177e4 | 1318 | goto activate_locked; |
1da177e4 LT |
1319 | } |
1320 | } | |
1321 | ||
1322 | if (PageDirty(page)) { | |
ee72886d | 1323 | /* |
4eda4823 JW |
1324 | * Only kswapd can writeback filesystem pages |
1325 | * to avoid risk of stack overflow. But avoid | |
1326 | * injecting inefficient single-page IO into | |
1327 | * flusher writeback as much as possible: only | |
1328 | * write pages when we've encountered many | |
1329 | * dirty pages, and when we've already scanned | |
1330 | * the rest of the LRU for clean pages and see | |
1331 | * the same dirty pages again (PageReclaim). | |
ee72886d | 1332 | */ |
f84f6e2b | 1333 | if (page_is_file_cache(page) && |
4eda4823 JW |
1334 | (!current_is_kswapd() || !PageReclaim(page) || |
1335 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { | |
49ea7eb6 MG |
1336 | /* |
1337 | * Immediately reclaim when written back. | |
1338 | * Similar in principal to deactivate_page() | |
1339 | * except we already have the page isolated | |
1340 | * and know it's dirty | |
1341 | */ | |
c4a25635 | 1342 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
49ea7eb6 MG |
1343 | SetPageReclaim(page); |
1344 | ||
c55e8d03 | 1345 | goto activate_locked; |
ee72886d MG |
1346 | } |
1347 | ||
dfc8d636 | 1348 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1349 | goto keep_locked; |
4dd4b920 | 1350 | if (!may_enter_fs) |
1da177e4 | 1351 | goto keep_locked; |
52a8363e | 1352 | if (!sc->may_writepage) |
1da177e4 LT |
1353 | goto keep_locked; |
1354 | ||
d950c947 MG |
1355 | /* |
1356 | * Page is dirty. Flush the TLB if a writable entry | |
1357 | * potentially exists to avoid CPU writes after IO | |
1358 | * starts and then write it out here. | |
1359 | */ | |
1360 | try_to_unmap_flush_dirty(); | |
cb16556d | 1361 | switch (pageout(page, mapping)) { |
1da177e4 LT |
1362 | case PAGE_KEEP: |
1363 | goto keep_locked; | |
1364 | case PAGE_ACTIVATE: | |
1365 | goto activate_locked; | |
1366 | case PAGE_SUCCESS: | |
7d3579e8 | 1367 | if (PageWriteback(page)) |
41ac1999 | 1368 | goto keep; |
7d3579e8 | 1369 | if (PageDirty(page)) |
1da177e4 | 1370 | goto keep; |
7d3579e8 | 1371 | |
1da177e4 LT |
1372 | /* |
1373 | * A synchronous write - probably a ramdisk. Go | |
1374 | * ahead and try to reclaim the page. | |
1375 | */ | |
529ae9aa | 1376 | if (!trylock_page(page)) |
1da177e4 LT |
1377 | goto keep; |
1378 | if (PageDirty(page) || PageWriteback(page)) | |
1379 | goto keep_locked; | |
1380 | mapping = page_mapping(page); | |
1381 | case PAGE_CLEAN: | |
1382 | ; /* try to free the page below */ | |
1383 | } | |
1384 | } | |
1385 | ||
1386 | /* | |
1387 | * If the page has buffers, try to free the buffer mappings | |
1388 | * associated with this page. If we succeed we try to free | |
1389 | * the page as well. | |
1390 | * | |
1391 | * We do this even if the page is PageDirty(). | |
1392 | * try_to_release_page() does not perform I/O, but it is | |
1393 | * possible for a page to have PageDirty set, but it is actually | |
1394 | * clean (all its buffers are clean). This happens if the | |
1395 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1396 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1397 | * try_to_release_page() will discover that cleanness and will |
1398 | * drop the buffers and mark the page clean - it can be freed. | |
1399 | * | |
1400 | * Rarely, pages can have buffers and no ->mapping. These are | |
1401 | * the pages which were not successfully invalidated in | |
1402 | * truncate_complete_page(). We try to drop those buffers here | |
1403 | * and if that worked, and the page is no longer mapped into | |
1404 | * process address space (page_count == 1) it can be freed. | |
1405 | * Otherwise, leave the page on the LRU so it is swappable. | |
1406 | */ | |
266cf658 | 1407 | if (page_has_private(page)) { |
1da177e4 LT |
1408 | if (!try_to_release_page(page, sc->gfp_mask)) |
1409 | goto activate_locked; | |
e286781d NP |
1410 | if (!mapping && page_count(page) == 1) { |
1411 | unlock_page(page); | |
1412 | if (put_page_testzero(page)) | |
1413 | goto free_it; | |
1414 | else { | |
1415 | /* | |
1416 | * rare race with speculative reference. | |
1417 | * the speculative reference will free | |
1418 | * this page shortly, so we may | |
1419 | * increment nr_reclaimed here (and | |
1420 | * leave it off the LRU). | |
1421 | */ | |
1422 | nr_reclaimed++; | |
1423 | continue; | |
1424 | } | |
1425 | } | |
1da177e4 LT |
1426 | } |
1427 | ||
802a3a92 SL |
1428 | if (PageAnon(page) && !PageSwapBacked(page)) { |
1429 | /* follow __remove_mapping for reference */ | |
1430 | if (!page_ref_freeze(page, 1)) | |
1431 | goto keep_locked; | |
1432 | if (PageDirty(page)) { | |
1433 | page_ref_unfreeze(page, 1); | |
1434 | goto keep_locked; | |
1435 | } | |
1da177e4 | 1436 | |
802a3a92 | 1437 | count_vm_event(PGLAZYFREED); |
2262185c | 1438 | count_memcg_page_event(page, PGLAZYFREED); |
b910718a JW |
1439 | } else if (!mapping || !__remove_mapping(mapping, page, true, |
1440 | sc->target_mem_cgroup)) | |
802a3a92 | 1441 | goto keep_locked; |
9a1ea439 HD |
1442 | |
1443 | unlock_page(page); | |
e286781d | 1444 | free_it: |
98879b3b YS |
1445 | /* |
1446 | * THP may get swapped out in a whole, need account | |
1447 | * all base pages. | |
1448 | */ | |
1449 | nr_reclaimed += nr_pages; | |
abe4c3b5 MG |
1450 | |
1451 | /* | |
1452 | * Is there need to periodically free_page_list? It would | |
1453 | * appear not as the counts should be low | |
1454 | */ | |
7ae88534 | 1455 | if (unlikely(PageTransHuge(page))) |
bd4c82c2 | 1456 | (*get_compound_page_dtor(page))(page); |
7ae88534 | 1457 | else |
bd4c82c2 | 1458 | list_add(&page->lru, &free_pages); |
1da177e4 LT |
1459 | continue; |
1460 | ||
98879b3b YS |
1461 | activate_locked_split: |
1462 | /* | |
1463 | * The tail pages that are failed to add into swap cache | |
1464 | * reach here. Fixup nr_scanned and nr_pages. | |
1465 | */ | |
1466 | if (nr_pages > 1) { | |
1467 | sc->nr_scanned -= (nr_pages - 1); | |
1468 | nr_pages = 1; | |
1469 | } | |
1da177e4 | 1470 | activate_locked: |
68a22394 | 1471 | /* Not a candidate for swapping, so reclaim swap space. */ |
ad6b6704 MK |
1472 | if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || |
1473 | PageMlocked(page))) | |
a2c43eed | 1474 | try_to_free_swap(page); |
309381fe | 1475 | VM_BUG_ON_PAGE(PageActive(page), page); |
ad6b6704 | 1476 | if (!PageMlocked(page)) { |
886cf190 | 1477 | int type = page_is_file_cache(page); |
ad6b6704 | 1478 | SetPageActive(page); |
98879b3b | 1479 | stat->nr_activate[type] += nr_pages; |
2262185c | 1480 | count_memcg_page_event(page, PGACTIVATE); |
ad6b6704 | 1481 | } |
1da177e4 LT |
1482 | keep_locked: |
1483 | unlock_page(page); | |
1484 | keep: | |
1485 | list_add(&page->lru, &ret_pages); | |
309381fe | 1486 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1487 | } |
abe4c3b5 | 1488 | |
98879b3b YS |
1489 | pgactivate = stat->nr_activate[0] + stat->nr_activate[1]; |
1490 | ||
747db954 | 1491 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1492 | try_to_unmap_flush(); |
2d4894b5 | 1493 | free_unref_page_list(&free_pages); |
abe4c3b5 | 1494 | |
1da177e4 | 1495 | list_splice(&ret_pages, page_list); |
886cf190 | 1496 | count_vm_events(PGACTIVATE, pgactivate); |
060f005f | 1497 | |
05ff5137 | 1498 | return nr_reclaimed; |
1da177e4 LT |
1499 | } |
1500 | ||
02c6de8d MK |
1501 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1502 | struct list_head *page_list) | |
1503 | { | |
1504 | struct scan_control sc = { | |
1505 | .gfp_mask = GFP_KERNEL, | |
1506 | .priority = DEF_PRIORITY, | |
1507 | .may_unmap = 1, | |
1508 | }; | |
060f005f | 1509 | struct reclaim_stat dummy_stat; |
3c710c1a | 1510 | unsigned long ret; |
02c6de8d MK |
1511 | struct page *page, *next; |
1512 | LIST_HEAD(clean_pages); | |
1513 | ||
1514 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e | 1515 | if (page_is_file_cache(page) && !PageDirty(page) && |
a58f2cef | 1516 | !__PageMovable(page) && !PageUnevictable(page)) { |
02c6de8d MK |
1517 | ClearPageActive(page); |
1518 | list_move(&page->lru, &clean_pages); | |
1519 | } | |
1520 | } | |
1521 | ||
599d0c95 | 1522 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
060f005f | 1523 | TTU_IGNORE_ACCESS, &dummy_stat, true); |
02c6de8d | 1524 | list_splice(&clean_pages, page_list); |
599d0c95 | 1525 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1526 | return ret; |
1527 | } | |
1528 | ||
5ad333eb AW |
1529 | /* |
1530 | * Attempt to remove the specified page from its LRU. Only take this page | |
1531 | * if it is of the appropriate PageActive status. Pages which are being | |
1532 | * freed elsewhere are also ignored. | |
1533 | * | |
1534 | * page: page to consider | |
1535 | * mode: one of the LRU isolation modes defined above | |
1536 | * | |
1537 | * returns 0 on success, -ve errno on failure. | |
1538 | */ | |
f3fd4a61 | 1539 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1540 | { |
1541 | int ret = -EINVAL; | |
1542 | ||
1543 | /* Only take pages on the LRU. */ | |
1544 | if (!PageLRU(page)) | |
1545 | return ret; | |
1546 | ||
e46a2879 MK |
1547 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1548 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1549 | return ret; |
1550 | ||
5ad333eb | 1551 | ret = -EBUSY; |
08e552c6 | 1552 | |
c8244935 MG |
1553 | /* |
1554 | * To minimise LRU disruption, the caller can indicate that it only | |
1555 | * wants to isolate pages it will be able to operate on without | |
1556 | * blocking - clean pages for the most part. | |
1557 | * | |
c8244935 MG |
1558 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
1559 | * that it is possible to migrate without blocking | |
1560 | */ | |
1276ad68 | 1561 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
c8244935 MG |
1562 | /* All the caller can do on PageWriteback is block */ |
1563 | if (PageWriteback(page)) | |
1564 | return ret; | |
1565 | ||
1566 | if (PageDirty(page)) { | |
1567 | struct address_space *mapping; | |
69d763fc | 1568 | bool migrate_dirty; |
c8244935 | 1569 | |
c8244935 MG |
1570 | /* |
1571 | * Only pages without mappings or that have a | |
1572 | * ->migratepage callback are possible to migrate | |
69d763fc MG |
1573 | * without blocking. However, we can be racing with |
1574 | * truncation so it's necessary to lock the page | |
1575 | * to stabilise the mapping as truncation holds | |
1576 | * the page lock until after the page is removed | |
1577 | * from the page cache. | |
c8244935 | 1578 | */ |
69d763fc MG |
1579 | if (!trylock_page(page)) |
1580 | return ret; | |
1581 | ||
c8244935 | 1582 | mapping = page_mapping(page); |
145e1a71 | 1583 | migrate_dirty = !mapping || mapping->a_ops->migratepage; |
69d763fc MG |
1584 | unlock_page(page); |
1585 | if (!migrate_dirty) | |
c8244935 MG |
1586 | return ret; |
1587 | } | |
1588 | } | |
39deaf85 | 1589 | |
f80c0673 MK |
1590 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1591 | return ret; | |
1592 | ||
5ad333eb AW |
1593 | if (likely(get_page_unless_zero(page))) { |
1594 | /* | |
1595 | * Be careful not to clear PageLRU until after we're | |
1596 | * sure the page is not being freed elsewhere -- the | |
1597 | * page release code relies on it. | |
1598 | */ | |
1599 | ClearPageLRU(page); | |
1600 | ret = 0; | |
1601 | } | |
1602 | ||
1603 | return ret; | |
1604 | } | |
1605 | ||
7ee36a14 MG |
1606 | |
1607 | /* | |
1608 | * Update LRU sizes after isolating pages. The LRU size updates must | |
1609 | * be complete before mem_cgroup_update_lru_size due to a santity check. | |
1610 | */ | |
1611 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 1612 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 1613 | { |
7ee36a14 MG |
1614 | int zid; |
1615 | ||
7ee36a14 MG |
1616 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1617 | if (!nr_zone_taken[zid]) | |
1618 | continue; | |
1619 | ||
1620 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); | |
7ee36a14 | 1621 | #ifdef CONFIG_MEMCG |
b4536f0c | 1622 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
7ee36a14 | 1623 | #endif |
b4536f0c MH |
1624 | } |
1625 | ||
7ee36a14 MG |
1626 | } |
1627 | ||
f4b7e272 AR |
1628 | /** |
1629 | * pgdat->lru_lock is heavily contended. Some of the functions that | |
1da177e4 LT |
1630 | * shrink the lists perform better by taking out a batch of pages |
1631 | * and working on them outside the LRU lock. | |
1632 | * | |
1633 | * For pagecache intensive workloads, this function is the hottest | |
1634 | * spot in the kernel (apart from copy_*_user functions). | |
1635 | * | |
1636 | * Appropriate locks must be held before calling this function. | |
1637 | * | |
791b48b6 | 1638 | * @nr_to_scan: The number of eligible pages to look through on the list. |
5dc35979 | 1639 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1640 | * @dst: The temp list to put pages on to. |
f626012d | 1641 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1642 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1643 | * @mode: One of the LRU isolation modes |
3cb99451 | 1644 | * @lru: LRU list id for isolating |
1da177e4 LT |
1645 | * |
1646 | * returns how many pages were moved onto *@dst. | |
1647 | */ | |
69e05944 | 1648 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1649 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1650 | unsigned long *nr_scanned, struct scan_control *sc, |
a9e7c39f | 1651 | enum lru_list lru) |
1da177e4 | 1652 | { |
75b00af7 | 1653 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1654 | unsigned long nr_taken = 0; |
599d0c95 | 1655 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 1656 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
3db65812 | 1657 | unsigned long skipped = 0; |
791b48b6 | 1658 | unsigned long scan, total_scan, nr_pages; |
b2e18757 | 1659 | LIST_HEAD(pages_skipped); |
a9e7c39f | 1660 | isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED); |
1da177e4 | 1661 | |
98879b3b | 1662 | total_scan = 0; |
791b48b6 | 1663 | scan = 0; |
98879b3b | 1664 | while (scan < nr_to_scan && !list_empty(src)) { |
5ad333eb | 1665 | struct page *page; |
5ad333eb | 1666 | |
1da177e4 LT |
1667 | page = lru_to_page(src); |
1668 | prefetchw_prev_lru_page(page, src, flags); | |
1669 | ||
309381fe | 1670 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1671 | |
d8c6546b | 1672 | nr_pages = compound_nr(page); |
98879b3b YS |
1673 | total_scan += nr_pages; |
1674 | ||
b2e18757 MG |
1675 | if (page_zonenum(page) > sc->reclaim_idx) { |
1676 | list_move(&page->lru, &pages_skipped); | |
98879b3b | 1677 | nr_skipped[page_zonenum(page)] += nr_pages; |
b2e18757 MG |
1678 | continue; |
1679 | } | |
1680 | ||
791b48b6 MK |
1681 | /* |
1682 | * Do not count skipped pages because that makes the function | |
1683 | * return with no isolated pages if the LRU mostly contains | |
1684 | * ineligible pages. This causes the VM to not reclaim any | |
1685 | * pages, triggering a premature OOM. | |
98879b3b YS |
1686 | * |
1687 | * Account all tail pages of THP. This would not cause | |
1688 | * premature OOM since __isolate_lru_page() returns -EBUSY | |
1689 | * only when the page is being freed somewhere else. | |
791b48b6 | 1690 | */ |
98879b3b | 1691 | scan += nr_pages; |
f3fd4a61 | 1692 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1693 | case 0: |
599d0c95 MG |
1694 | nr_taken += nr_pages; |
1695 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
5ad333eb | 1696 | list_move(&page->lru, dst); |
5ad333eb AW |
1697 | break; |
1698 | ||
1699 | case -EBUSY: | |
1700 | /* else it is being freed elsewhere */ | |
1701 | list_move(&page->lru, src); | |
1702 | continue; | |
46453a6e | 1703 | |
5ad333eb AW |
1704 | default: |
1705 | BUG(); | |
1706 | } | |
1da177e4 LT |
1707 | } |
1708 | ||
b2e18757 MG |
1709 | /* |
1710 | * Splice any skipped pages to the start of the LRU list. Note that | |
1711 | * this disrupts the LRU order when reclaiming for lower zones but | |
1712 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
1713 | * scanning would soon rescan the same pages to skip and put the | |
1714 | * system at risk of premature OOM. | |
1715 | */ | |
7cc30fcf MG |
1716 | if (!list_empty(&pages_skipped)) { |
1717 | int zid; | |
1718 | ||
3db65812 | 1719 | list_splice(&pages_skipped, src); |
7cc30fcf MG |
1720 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1721 | if (!nr_skipped[zid]) | |
1722 | continue; | |
1723 | ||
1724 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 1725 | skipped += nr_skipped[zid]; |
7cc30fcf MG |
1726 | } |
1727 | } | |
791b48b6 | 1728 | *nr_scanned = total_scan; |
1265e3a6 | 1729 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
791b48b6 | 1730 | total_scan, skipped, nr_taken, mode, lru); |
b4536f0c | 1731 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
1732 | return nr_taken; |
1733 | } | |
1734 | ||
62695a84 NP |
1735 | /** |
1736 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1737 | * @page: page to isolate from its LRU list | |
1738 | * | |
1739 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1740 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1741 | * | |
1742 | * Returns 0 if the page was removed from an LRU list. | |
1743 | * Returns -EBUSY if the page was not on an LRU list. | |
1744 | * | |
1745 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1746 | * the active list, it will have PageActive set. If it was found on |
1747 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1748 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1749 | * |
1750 | * The vmstat statistic corresponding to the list on which the page was | |
1751 | * found will be decremented. | |
1752 | * | |
1753 | * Restrictions: | |
a5d09bed | 1754 | * |
62695a84 NP |
1755 | * (1) Must be called with an elevated refcount on the page. This is a |
1756 | * fundamentnal difference from isolate_lru_pages (which is called | |
1757 | * without a stable reference). | |
1758 | * (2) the lru_lock must not be held. | |
1759 | * (3) interrupts must be enabled. | |
1760 | */ | |
1761 | int isolate_lru_page(struct page *page) | |
1762 | { | |
1763 | int ret = -EBUSY; | |
1764 | ||
309381fe | 1765 | VM_BUG_ON_PAGE(!page_count(page), page); |
cf2a82ee | 1766 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
0c917313 | 1767 | |
62695a84 | 1768 | if (PageLRU(page)) { |
f4b7e272 | 1769 | pg_data_t *pgdat = page_pgdat(page); |
fa9add64 | 1770 | struct lruvec *lruvec; |
62695a84 | 1771 | |
f4b7e272 AR |
1772 | spin_lock_irq(&pgdat->lru_lock); |
1773 | lruvec = mem_cgroup_page_lruvec(page, pgdat); | |
0c917313 | 1774 | if (PageLRU(page)) { |
894bc310 | 1775 | int lru = page_lru(page); |
0c917313 | 1776 | get_page(page); |
62695a84 | 1777 | ClearPageLRU(page); |
fa9add64 HD |
1778 | del_page_from_lru_list(page, lruvec, lru); |
1779 | ret = 0; | |
62695a84 | 1780 | } |
f4b7e272 | 1781 | spin_unlock_irq(&pgdat->lru_lock); |
62695a84 NP |
1782 | } |
1783 | return ret; | |
1784 | } | |
1785 | ||
35cd7815 | 1786 | /* |
d37dd5dc | 1787 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
178821b8 | 1788 | * then get rescheduled. When there are massive number of tasks doing page |
d37dd5dc FW |
1789 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, |
1790 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1791 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 1792 | */ |
599d0c95 | 1793 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
1794 | struct scan_control *sc) |
1795 | { | |
1796 | unsigned long inactive, isolated; | |
1797 | ||
1798 | if (current_is_kswapd()) | |
1799 | return 0; | |
1800 | ||
b5ead35e | 1801 | if (!writeback_throttling_sane(sc)) |
35cd7815 RR |
1802 | return 0; |
1803 | ||
1804 | if (file) { | |
599d0c95 MG |
1805 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1806 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 1807 | } else { |
599d0c95 MG |
1808 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1809 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
1810 | } |
1811 | ||
3cf23841 FW |
1812 | /* |
1813 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1814 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1815 | * deadlock. | |
1816 | */ | |
d0164adc | 1817 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1818 | inactive >>= 3; |
1819 | ||
35cd7815 RR |
1820 | return isolated > inactive; |
1821 | } | |
1822 | ||
a222f341 KT |
1823 | /* |
1824 | * This moves pages from @list to corresponding LRU list. | |
1825 | * | |
1826 | * We move them the other way if the page is referenced by one or more | |
1827 | * processes, from rmap. | |
1828 | * | |
1829 | * If the pages are mostly unmapped, the processing is fast and it is | |
1830 | * appropriate to hold zone_lru_lock across the whole operation. But if | |
1831 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1832 | * should drop zone_lru_lock around each page. It's impossible to balance | |
1833 | * this, so instead we remove the pages from the LRU while processing them. | |
1834 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1835 | * nobody will play with that bit on a non-LRU page. | |
1836 | * | |
1837 | * The downside is that we have to touch page->_refcount against each page. | |
1838 | * But we had to alter page->flags anyway. | |
1839 | * | |
1840 | * Returns the number of pages moved to the given lruvec. | |
1841 | */ | |
1842 | ||
1843 | static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec, | |
1844 | struct list_head *list) | |
66635629 | 1845 | { |
599d0c95 | 1846 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
a222f341 | 1847 | int nr_pages, nr_moved = 0; |
3f79768f | 1848 | LIST_HEAD(pages_to_free); |
a222f341 KT |
1849 | struct page *page; |
1850 | enum lru_list lru; | |
66635629 | 1851 | |
a222f341 KT |
1852 | while (!list_empty(list)) { |
1853 | page = lru_to_page(list); | |
309381fe | 1854 | VM_BUG_ON_PAGE(PageLRU(page), page); |
39b5f29a | 1855 | if (unlikely(!page_evictable(page))) { |
a222f341 | 1856 | list_del(&page->lru); |
599d0c95 | 1857 | spin_unlock_irq(&pgdat->lru_lock); |
66635629 | 1858 | putback_lru_page(page); |
599d0c95 | 1859 | spin_lock_irq(&pgdat->lru_lock); |
66635629 MG |
1860 | continue; |
1861 | } | |
599d0c95 | 1862 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 1863 | |
7a608572 | 1864 | SetPageLRU(page); |
66635629 | 1865 | lru = page_lru(page); |
a222f341 KT |
1866 | |
1867 | nr_pages = hpage_nr_pages(page); | |
1868 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); | |
1869 | list_move(&page->lru, &lruvec->lists[lru]); | |
fa9add64 | 1870 | |
2bcf8879 HD |
1871 | if (put_page_testzero(page)) { |
1872 | __ClearPageLRU(page); | |
1873 | __ClearPageActive(page); | |
fa9add64 | 1874 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1875 | |
1876 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1877 | spin_unlock_irq(&pgdat->lru_lock); |
2bcf8879 | 1878 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1879 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1880 | } else |
1881 | list_add(&page->lru, &pages_to_free); | |
a222f341 KT |
1882 | } else { |
1883 | nr_moved += nr_pages; | |
66635629 MG |
1884 | } |
1885 | } | |
66635629 | 1886 | |
3f79768f HD |
1887 | /* |
1888 | * To save our caller's stack, now use input list for pages to free. | |
1889 | */ | |
a222f341 KT |
1890 | list_splice(&pages_to_free, list); |
1891 | ||
1892 | return nr_moved; | |
66635629 MG |
1893 | } |
1894 | ||
399ba0b9 N |
1895 | /* |
1896 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1897 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1898 | * In that case we should only throttle if the backing device it is | |
1899 | * writing to is congested. In other cases it is safe to throttle. | |
1900 | */ | |
1901 | static int current_may_throttle(void) | |
1902 | { | |
1903 | return !(current->flags & PF_LESS_THROTTLE) || | |
1904 | current->backing_dev_info == NULL || | |
1905 | bdi_write_congested(current->backing_dev_info); | |
1906 | } | |
1907 | ||
1da177e4 | 1908 | /* |
b2e18757 | 1909 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 1910 | * of reclaimed pages |
1da177e4 | 1911 | */ |
66635629 | 1912 | static noinline_for_stack unsigned long |
1a93be0e | 1913 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1914 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1915 | { |
1916 | LIST_HEAD(page_list); | |
e247dbce | 1917 | unsigned long nr_scanned; |
05ff5137 | 1918 | unsigned long nr_reclaimed = 0; |
e247dbce | 1919 | unsigned long nr_taken; |
060f005f | 1920 | struct reclaim_stat stat; |
3cb99451 | 1921 | int file = is_file_lru(lru); |
f46b7912 | 1922 | enum vm_event_item item; |
599d0c95 | 1923 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1a93be0e | 1924 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
db73ee0d | 1925 | bool stalled = false; |
78dc583d | 1926 | |
599d0c95 | 1927 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
db73ee0d MH |
1928 | if (stalled) |
1929 | return 0; | |
1930 | ||
1931 | /* wait a bit for the reclaimer. */ | |
1932 | msleep(100); | |
1933 | stalled = true; | |
35cd7815 RR |
1934 | |
1935 | /* We are about to die and free our memory. Return now. */ | |
1936 | if (fatal_signal_pending(current)) | |
1937 | return SWAP_CLUSTER_MAX; | |
1938 | } | |
1939 | ||
1da177e4 | 1940 | lru_add_drain(); |
f80c0673 | 1941 | |
599d0c95 | 1942 | spin_lock_irq(&pgdat->lru_lock); |
b35ea17b | 1943 | |
5dc35979 | 1944 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
a9e7c39f | 1945 | &nr_scanned, sc, lru); |
95d918fc | 1946 | |
599d0c95 | 1947 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
9d5e6a9f | 1948 | reclaim_stat->recent_scanned[file] += nr_taken; |
95d918fc | 1949 | |
f46b7912 | 1950 | item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT; |
b5ead35e | 1951 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
1952 | __count_vm_events(item, nr_scanned); |
1953 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned); | |
599d0c95 | 1954 | spin_unlock_irq(&pgdat->lru_lock); |
b35ea17b | 1955 | |
d563c050 | 1956 | if (nr_taken == 0) |
66635629 | 1957 | return 0; |
5ad333eb | 1958 | |
a128ca71 | 1959 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0, |
3c710c1a | 1960 | &stat, false); |
c661b078 | 1961 | |
599d0c95 | 1962 | spin_lock_irq(&pgdat->lru_lock); |
3f79768f | 1963 | |
f46b7912 | 1964 | item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT; |
b5ead35e | 1965 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
1966 | __count_vm_events(item, nr_reclaimed); |
1967 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed); | |
b17f18af KT |
1968 | reclaim_stat->recent_rotated[0] += stat.nr_activate[0]; |
1969 | reclaim_stat->recent_rotated[1] += stat.nr_activate[1]; | |
a74609fa | 1970 | |
a222f341 | 1971 | move_pages_to_lru(lruvec, &page_list); |
3f79768f | 1972 | |
599d0c95 | 1973 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f | 1974 | |
599d0c95 | 1975 | spin_unlock_irq(&pgdat->lru_lock); |
3f79768f | 1976 | |
747db954 | 1977 | mem_cgroup_uncharge_list(&page_list); |
2d4894b5 | 1978 | free_unref_page_list(&page_list); |
e11da5b4 | 1979 | |
1c610d5f AR |
1980 | /* |
1981 | * If dirty pages are scanned that are not queued for IO, it | |
1982 | * implies that flushers are not doing their job. This can | |
1983 | * happen when memory pressure pushes dirty pages to the end of | |
1984 | * the LRU before the dirty limits are breached and the dirty | |
1985 | * data has expired. It can also happen when the proportion of | |
1986 | * dirty pages grows not through writes but through memory | |
1987 | * pressure reclaiming all the clean cache. And in some cases, | |
1988 | * the flushers simply cannot keep up with the allocation | |
1989 | * rate. Nudge the flusher threads in case they are asleep. | |
1990 | */ | |
1991 | if (stat.nr_unqueued_dirty == nr_taken) | |
1992 | wakeup_flusher_threads(WB_REASON_VMSCAN); | |
1993 | ||
d108c772 AR |
1994 | sc->nr.dirty += stat.nr_dirty; |
1995 | sc->nr.congested += stat.nr_congested; | |
1996 | sc->nr.unqueued_dirty += stat.nr_unqueued_dirty; | |
1997 | sc->nr.writeback += stat.nr_writeback; | |
1998 | sc->nr.immediate += stat.nr_immediate; | |
1999 | sc->nr.taken += nr_taken; | |
2000 | if (file) | |
2001 | sc->nr.file_taken += nr_taken; | |
8e950282 | 2002 | |
599d0c95 | 2003 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
d51d1e64 | 2004 | nr_scanned, nr_reclaimed, &stat, sc->priority, file); |
05ff5137 | 2005 | return nr_reclaimed; |
1da177e4 LT |
2006 | } |
2007 | ||
f626012d | 2008 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 2009 | struct lruvec *lruvec, |
f16015fb | 2010 | struct scan_control *sc, |
9e3b2f8c | 2011 | enum lru_list lru) |
1da177e4 | 2012 | { |
44c241f1 | 2013 | unsigned long nr_taken; |
f626012d | 2014 | unsigned long nr_scanned; |
6fe6b7e3 | 2015 | unsigned long vm_flags; |
1da177e4 | 2016 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 2017 | LIST_HEAD(l_active); |
b69408e8 | 2018 | LIST_HEAD(l_inactive); |
1da177e4 | 2019 | struct page *page; |
1a93be0e | 2020 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
9d998b4f MH |
2021 | unsigned nr_deactivate, nr_activate; |
2022 | unsigned nr_rotated = 0; | |
3cb99451 | 2023 | int file = is_file_lru(lru); |
599d0c95 | 2024 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
2025 | |
2026 | lru_add_drain(); | |
f80c0673 | 2027 | |
599d0c95 | 2028 | spin_lock_irq(&pgdat->lru_lock); |
925b7673 | 2029 | |
5dc35979 | 2030 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
a9e7c39f | 2031 | &nr_scanned, sc, lru); |
89b5fae5 | 2032 | |
599d0c95 | 2033 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
b7c46d15 | 2034 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 2035 | |
599d0c95 | 2036 | __count_vm_events(PGREFILL, nr_scanned); |
2fa2690c | 2037 | __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); |
9d5e6a9f | 2038 | |
599d0c95 | 2039 | spin_unlock_irq(&pgdat->lru_lock); |
1da177e4 | 2040 | |
1da177e4 LT |
2041 | while (!list_empty(&l_hold)) { |
2042 | cond_resched(); | |
2043 | page = lru_to_page(&l_hold); | |
2044 | list_del(&page->lru); | |
7e9cd484 | 2045 | |
39b5f29a | 2046 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
2047 | putback_lru_page(page); |
2048 | continue; | |
2049 | } | |
2050 | ||
cc715d99 MG |
2051 | if (unlikely(buffer_heads_over_limit)) { |
2052 | if (page_has_private(page) && trylock_page(page)) { | |
2053 | if (page_has_private(page)) | |
2054 | try_to_release_page(page, 0); | |
2055 | unlock_page(page); | |
2056 | } | |
2057 | } | |
2058 | ||
c3ac9a8a JW |
2059 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
2060 | &vm_flags)) { | |
9992af10 | 2061 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
2062 | /* |
2063 | * Identify referenced, file-backed active pages and | |
2064 | * give them one more trip around the active list. So | |
2065 | * that executable code get better chances to stay in | |
2066 | * memory under moderate memory pressure. Anon pages | |
2067 | * are not likely to be evicted by use-once streaming | |
2068 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
2069 | * so we ignore them here. | |
2070 | */ | |
41e20983 | 2071 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
2072 | list_add(&page->lru, &l_active); |
2073 | continue; | |
2074 | } | |
2075 | } | |
7e9cd484 | 2076 | |
5205e56e | 2077 | ClearPageActive(page); /* we are de-activating */ |
1899ad18 | 2078 | SetPageWorkingset(page); |
1da177e4 LT |
2079 | list_add(&page->lru, &l_inactive); |
2080 | } | |
2081 | ||
b555749a | 2082 | /* |
8cab4754 | 2083 | * Move pages back to the lru list. |
b555749a | 2084 | */ |
599d0c95 | 2085 | spin_lock_irq(&pgdat->lru_lock); |
556adecb | 2086 | /* |
8cab4754 WF |
2087 | * Count referenced pages from currently used mappings as rotated, |
2088 | * even though only some of them are actually re-activated. This | |
2089 | * helps balance scan pressure between file and anonymous pages in | |
7c0db9e9 | 2090 | * get_scan_count. |
7e9cd484 | 2091 | */ |
b7c46d15 | 2092 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 2093 | |
a222f341 KT |
2094 | nr_activate = move_pages_to_lru(lruvec, &l_active); |
2095 | nr_deactivate = move_pages_to_lru(lruvec, &l_inactive); | |
f372d89e KT |
2096 | /* Keep all free pages in l_active list */ |
2097 | list_splice(&l_inactive, &l_active); | |
9851ac13 KT |
2098 | |
2099 | __count_vm_events(PGDEACTIVATE, nr_deactivate); | |
2100 | __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate); | |
2101 | ||
599d0c95 MG |
2102 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
2103 | spin_unlock_irq(&pgdat->lru_lock); | |
2bcf8879 | 2104 | |
f372d89e KT |
2105 | mem_cgroup_uncharge_list(&l_active); |
2106 | free_unref_page_list(&l_active); | |
9d998b4f MH |
2107 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
2108 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
2109 | } |
2110 | ||
1a4e58cc MK |
2111 | unsigned long reclaim_pages(struct list_head *page_list) |
2112 | { | |
2113 | int nid = -1; | |
2114 | unsigned long nr_reclaimed = 0; | |
2115 | LIST_HEAD(node_page_list); | |
2116 | struct reclaim_stat dummy_stat; | |
2117 | struct page *page; | |
2118 | struct scan_control sc = { | |
2119 | .gfp_mask = GFP_KERNEL, | |
2120 | .priority = DEF_PRIORITY, | |
2121 | .may_writepage = 1, | |
2122 | .may_unmap = 1, | |
2123 | .may_swap = 1, | |
2124 | }; | |
2125 | ||
2126 | while (!list_empty(page_list)) { | |
2127 | page = lru_to_page(page_list); | |
2128 | if (nid == -1) { | |
2129 | nid = page_to_nid(page); | |
2130 | INIT_LIST_HEAD(&node_page_list); | |
2131 | } | |
2132 | ||
2133 | if (nid == page_to_nid(page)) { | |
2134 | ClearPageActive(page); | |
2135 | list_move(&page->lru, &node_page_list); | |
2136 | continue; | |
2137 | } | |
2138 | ||
2139 | nr_reclaimed += shrink_page_list(&node_page_list, | |
2140 | NODE_DATA(nid), | |
2141 | &sc, 0, | |
2142 | &dummy_stat, false); | |
2143 | while (!list_empty(&node_page_list)) { | |
2144 | page = lru_to_page(&node_page_list); | |
2145 | list_del(&page->lru); | |
2146 | putback_lru_page(page); | |
2147 | } | |
2148 | ||
2149 | nid = -1; | |
2150 | } | |
2151 | ||
2152 | if (!list_empty(&node_page_list)) { | |
2153 | nr_reclaimed += shrink_page_list(&node_page_list, | |
2154 | NODE_DATA(nid), | |
2155 | &sc, 0, | |
2156 | &dummy_stat, false); | |
2157 | while (!list_empty(&node_page_list)) { | |
2158 | page = lru_to_page(&node_page_list); | |
2159 | list_del(&page->lru); | |
2160 | putback_lru_page(page); | |
2161 | } | |
2162 | } | |
2163 | ||
2164 | return nr_reclaimed; | |
2165 | } | |
2166 | ||
b91ac374 JW |
2167 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
2168 | struct lruvec *lruvec, struct scan_control *sc) | |
2169 | { | |
2170 | if (is_active_lru(lru)) { | |
2171 | if (sc->may_deactivate & (1 << is_file_lru(lru))) | |
2172 | shrink_active_list(nr_to_scan, lruvec, sc, lru); | |
2173 | else | |
2174 | sc->skipped_deactivate = 1; | |
2175 | return 0; | |
2176 | } | |
2177 | ||
2178 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); | |
2179 | } | |
2180 | ||
59dc76b0 RR |
2181 | /* |
2182 | * The inactive anon list should be small enough that the VM never has | |
2183 | * to do too much work. | |
14797e23 | 2184 | * |
59dc76b0 RR |
2185 | * The inactive file list should be small enough to leave most memory |
2186 | * to the established workingset on the scan-resistant active list, | |
2187 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2188 | * |
59dc76b0 RR |
2189 | * Both inactive lists should also be large enough that each inactive |
2190 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 2191 | * |
2a2e4885 JW |
2192 | * If that fails and refaulting is observed, the inactive list grows. |
2193 | * | |
59dc76b0 | 2194 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
3a50d14d | 2195 | * on this LRU, maintained by the pageout code. An inactive_ratio |
59dc76b0 | 2196 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. |
56e49d21 | 2197 | * |
59dc76b0 RR |
2198 | * total target max |
2199 | * memory ratio inactive | |
2200 | * ------------------------------------- | |
2201 | * 10MB 1 5MB | |
2202 | * 100MB 1 50MB | |
2203 | * 1GB 3 250MB | |
2204 | * 10GB 10 0.9GB | |
2205 | * 100GB 31 3GB | |
2206 | * 1TB 101 10GB | |
2207 | * 10TB 320 32GB | |
56e49d21 | 2208 | */ |
b91ac374 | 2209 | static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru) |
56e49d21 | 2210 | { |
b91ac374 | 2211 | enum lru_list active_lru = inactive_lru + LRU_ACTIVE; |
2a2e4885 JW |
2212 | unsigned long inactive, active; |
2213 | unsigned long inactive_ratio; | |
59dc76b0 | 2214 | unsigned long gb; |
e3790144 | 2215 | |
b91ac374 JW |
2216 | inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru); |
2217 | active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru); | |
f8d1a311 | 2218 | |
b91ac374 JW |
2219 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
2220 | if (gb) | |
2221 | inactive_ratio = int_sqrt(10 * gb); | |
2222 | else | |
2223 | inactive_ratio = 1; | |
fd538803 | 2224 | |
59dc76b0 | 2225 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2226 | } |
2227 | ||
9a265114 JW |
2228 | enum scan_balance { |
2229 | SCAN_EQUAL, | |
2230 | SCAN_FRACT, | |
2231 | SCAN_ANON, | |
2232 | SCAN_FILE, | |
2233 | }; | |
2234 | ||
4f98a2fe RR |
2235 | /* |
2236 | * Determine how aggressively the anon and file LRU lists should be | |
2237 | * scanned. The relative value of each set of LRU lists is determined | |
2238 | * by looking at the fraction of the pages scanned we did rotate back | |
2239 | * onto the active list instead of evict. | |
2240 | * | |
be7bd59d WL |
2241 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2242 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 2243 | */ |
afaf07a6 JW |
2244 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, |
2245 | unsigned long *nr) | |
4f98a2fe | 2246 | { |
afaf07a6 | 2247 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); |
33377678 | 2248 | int swappiness = mem_cgroup_swappiness(memcg); |
9a265114 JW |
2249 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
2250 | u64 fraction[2]; | |
2251 | u64 denominator = 0; /* gcc */ | |
599d0c95 | 2252 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
4f98a2fe | 2253 | unsigned long anon_prio, file_prio; |
9a265114 | 2254 | enum scan_balance scan_balance; |
0bf1457f | 2255 | unsigned long anon, file; |
4f98a2fe | 2256 | unsigned long ap, fp; |
4111304d | 2257 | enum lru_list lru; |
76a33fc3 SL |
2258 | |
2259 | /* If we have no swap space, do not bother scanning anon pages. */ | |
d8b38438 | 2260 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
9a265114 | 2261 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2262 | goto out; |
2263 | } | |
4f98a2fe | 2264 | |
10316b31 JW |
2265 | /* |
2266 | * Global reclaim will swap to prevent OOM even with no | |
2267 | * swappiness, but memcg users want to use this knob to | |
2268 | * disable swapping for individual groups completely when | |
2269 | * using the memory controller's swap limit feature would be | |
2270 | * too expensive. | |
2271 | */ | |
b5ead35e | 2272 | if (cgroup_reclaim(sc) && !swappiness) { |
9a265114 | 2273 | scan_balance = SCAN_FILE; |
10316b31 JW |
2274 | goto out; |
2275 | } | |
2276 | ||
2277 | /* | |
2278 | * Do not apply any pressure balancing cleverness when the | |
2279 | * system is close to OOM, scan both anon and file equally | |
2280 | * (unless the swappiness setting disagrees with swapping). | |
2281 | */ | |
02695175 | 2282 | if (!sc->priority && swappiness) { |
9a265114 | 2283 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2284 | goto out; |
2285 | } | |
2286 | ||
62376251 | 2287 | /* |
53138cea | 2288 | * If the system is almost out of file pages, force-scan anon. |
62376251 | 2289 | */ |
b91ac374 | 2290 | if (sc->file_is_tiny) { |
53138cea JW |
2291 | scan_balance = SCAN_ANON; |
2292 | goto out; | |
62376251 JW |
2293 | } |
2294 | ||
7c5bd705 | 2295 | /* |
b91ac374 JW |
2296 | * If there is enough inactive page cache, we do not reclaim |
2297 | * anything from the anonymous working right now. | |
7c5bd705 | 2298 | */ |
b91ac374 | 2299 | if (sc->cache_trim_mode) { |
9a265114 | 2300 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2301 | goto out; |
2302 | } | |
2303 | ||
9a265114 JW |
2304 | scan_balance = SCAN_FRACT; |
2305 | ||
58c37f6e KM |
2306 | /* |
2307 | * With swappiness at 100, anonymous and file have the same priority. | |
2308 | * This scanning priority is essentially the inverse of IO cost. | |
2309 | */ | |
02695175 | 2310 | anon_prio = swappiness; |
75b00af7 | 2311 | file_prio = 200 - anon_prio; |
58c37f6e | 2312 | |
4f98a2fe RR |
2313 | /* |
2314 | * OK, so we have swap space and a fair amount of page cache | |
2315 | * pages. We use the recently rotated / recently scanned | |
2316 | * ratios to determine how valuable each cache is. | |
2317 | * | |
2318 | * Because workloads change over time (and to avoid overflow) | |
2319 | * we keep these statistics as a floating average, which ends | |
2320 | * up weighing recent references more than old ones. | |
2321 | * | |
2322 | * anon in [0], file in [1] | |
2323 | */ | |
2ab051e1 | 2324 | |
fd538803 MH |
2325 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + |
2326 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); | |
2327 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + | |
2328 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); | |
2ab051e1 | 2329 | |
599d0c95 | 2330 | spin_lock_irq(&pgdat->lru_lock); |
6e901571 | 2331 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
2332 | reclaim_stat->recent_scanned[0] /= 2; |
2333 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
2334 | } |
2335 | ||
6e901571 | 2336 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
2337 | reclaim_stat->recent_scanned[1] /= 2; |
2338 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
2339 | } |
2340 | ||
4f98a2fe | 2341 | /* |
00d8089c RR |
2342 | * The amount of pressure on anon vs file pages is inversely |
2343 | * proportional to the fraction of recently scanned pages on | |
2344 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2345 | */ |
fe35004f | 2346 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 2347 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2348 | |
fe35004f | 2349 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2350 | fp /= reclaim_stat->recent_rotated[1] + 1; |
599d0c95 | 2351 | spin_unlock_irq(&pgdat->lru_lock); |
4f98a2fe | 2352 | |
76a33fc3 SL |
2353 | fraction[0] = ap; |
2354 | fraction[1] = fp; | |
2355 | denominator = ap + fp + 1; | |
2356 | out: | |
688035f7 JW |
2357 | for_each_evictable_lru(lru) { |
2358 | int file = is_file_lru(lru); | |
9783aa99 | 2359 | unsigned long lruvec_size; |
688035f7 | 2360 | unsigned long scan; |
1bc63fb1 | 2361 | unsigned long protection; |
9783aa99 CD |
2362 | |
2363 | lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); | |
1bc63fb1 CD |
2364 | protection = mem_cgroup_protection(memcg, |
2365 | sc->memcg_low_reclaim); | |
9783aa99 | 2366 | |
1bc63fb1 | 2367 | if (protection) { |
9783aa99 CD |
2368 | /* |
2369 | * Scale a cgroup's reclaim pressure by proportioning | |
2370 | * its current usage to its memory.low or memory.min | |
2371 | * setting. | |
2372 | * | |
2373 | * This is important, as otherwise scanning aggression | |
2374 | * becomes extremely binary -- from nothing as we | |
2375 | * approach the memory protection threshold, to totally | |
2376 | * nominal as we exceed it. This results in requiring | |
2377 | * setting extremely liberal protection thresholds. It | |
2378 | * also means we simply get no protection at all if we | |
2379 | * set it too low, which is not ideal. | |
1bc63fb1 CD |
2380 | * |
2381 | * If there is any protection in place, we reduce scan | |
2382 | * pressure by how much of the total memory used is | |
2383 | * within protection thresholds. | |
9783aa99 | 2384 | * |
9de7ca46 CD |
2385 | * There is one special case: in the first reclaim pass, |
2386 | * we skip over all groups that are within their low | |
2387 | * protection. If that fails to reclaim enough pages to | |
2388 | * satisfy the reclaim goal, we come back and override | |
2389 | * the best-effort low protection. However, we still | |
2390 | * ideally want to honor how well-behaved groups are in | |
2391 | * that case instead of simply punishing them all | |
2392 | * equally. As such, we reclaim them based on how much | |
1bc63fb1 CD |
2393 | * memory they are using, reducing the scan pressure |
2394 | * again by how much of the total memory used is under | |
2395 | * hard protection. | |
9783aa99 | 2396 | */ |
1bc63fb1 CD |
2397 | unsigned long cgroup_size = mem_cgroup_size(memcg); |
2398 | ||
2399 | /* Avoid TOCTOU with earlier protection check */ | |
2400 | cgroup_size = max(cgroup_size, protection); | |
2401 | ||
2402 | scan = lruvec_size - lruvec_size * protection / | |
2403 | cgroup_size; | |
9783aa99 CD |
2404 | |
2405 | /* | |
1bc63fb1 | 2406 | * Minimally target SWAP_CLUSTER_MAX pages to keep |
9de7ca46 CD |
2407 | * reclaim moving forwards, avoiding decremeting |
2408 | * sc->priority further than desirable. | |
9783aa99 | 2409 | */ |
1bc63fb1 | 2410 | scan = max(scan, SWAP_CLUSTER_MAX); |
9783aa99 CD |
2411 | } else { |
2412 | scan = lruvec_size; | |
2413 | } | |
2414 | ||
2415 | scan >>= sc->priority; | |
6b4f7799 | 2416 | |
688035f7 JW |
2417 | /* |
2418 | * If the cgroup's already been deleted, make sure to | |
2419 | * scrape out the remaining cache. | |
2420 | */ | |
2421 | if (!scan && !mem_cgroup_online(memcg)) | |
9783aa99 | 2422 | scan = min(lruvec_size, SWAP_CLUSTER_MAX); |
6b4f7799 | 2423 | |
688035f7 JW |
2424 | switch (scan_balance) { |
2425 | case SCAN_EQUAL: | |
2426 | /* Scan lists relative to size */ | |
2427 | break; | |
2428 | case SCAN_FRACT: | |
9a265114 | 2429 | /* |
688035f7 JW |
2430 | * Scan types proportional to swappiness and |
2431 | * their relative recent reclaim efficiency. | |
68600f62 RG |
2432 | * Make sure we don't miss the last page |
2433 | * because of a round-off error. | |
9a265114 | 2434 | */ |
68600f62 RG |
2435 | scan = DIV64_U64_ROUND_UP(scan * fraction[file], |
2436 | denominator); | |
688035f7 JW |
2437 | break; |
2438 | case SCAN_FILE: | |
2439 | case SCAN_ANON: | |
2440 | /* Scan one type exclusively */ | |
2441 | if ((scan_balance == SCAN_FILE) != file) { | |
9783aa99 | 2442 | lruvec_size = 0; |
688035f7 JW |
2443 | scan = 0; |
2444 | } | |
2445 | break; | |
2446 | default: | |
2447 | /* Look ma, no brain */ | |
2448 | BUG(); | |
9a265114 | 2449 | } |
688035f7 | 2450 | |
688035f7 | 2451 | nr[lru] = scan; |
76a33fc3 | 2452 | } |
6e08a369 | 2453 | } |
4f98a2fe | 2454 | |
afaf07a6 | 2455 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) |
9b4f98cd JW |
2456 | { |
2457 | unsigned long nr[NR_LRU_LISTS]; | |
e82e0561 | 2458 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2459 | unsigned long nr_to_scan; |
2460 | enum lru_list lru; | |
2461 | unsigned long nr_reclaimed = 0; | |
2462 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2463 | struct blk_plug plug; | |
1a501907 | 2464 | bool scan_adjusted; |
9b4f98cd | 2465 | |
afaf07a6 | 2466 | get_scan_count(lruvec, sc, nr); |
9b4f98cd | 2467 | |
e82e0561 MG |
2468 | /* Record the original scan target for proportional adjustments later */ |
2469 | memcpy(targets, nr, sizeof(nr)); | |
2470 | ||
1a501907 MG |
2471 | /* |
2472 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2473 | * event that can occur when there is little memory pressure e.g. | |
2474 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2475 | * when the requested number of pages are reclaimed when scanning at | |
2476 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2477 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2478 | * do a batch of work at once. For memcg reclaim one check is made to | |
2479 | * abort proportional reclaim if either the file or anon lru has already | |
2480 | * dropped to zero at the first pass. | |
2481 | */ | |
b5ead35e | 2482 | scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() && |
1a501907 MG |
2483 | sc->priority == DEF_PRIORITY); |
2484 | ||
9b4f98cd JW |
2485 | blk_start_plug(&plug); |
2486 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2487 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2488 | unsigned long nr_anon, nr_file, percentage; |
2489 | unsigned long nr_scanned; | |
2490 | ||
9b4f98cd JW |
2491 | for_each_evictable_lru(lru) { |
2492 | if (nr[lru]) { | |
2493 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2494 | nr[lru] -= nr_to_scan; | |
2495 | ||
2496 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
3b991208 | 2497 | lruvec, sc); |
9b4f98cd JW |
2498 | } |
2499 | } | |
e82e0561 | 2500 | |
bd041733 MH |
2501 | cond_resched(); |
2502 | ||
e82e0561 MG |
2503 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
2504 | continue; | |
2505 | ||
e82e0561 MG |
2506 | /* |
2507 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2508 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2509 | * proportionally what was requested by get_scan_count(). We |
2510 | * stop reclaiming one LRU and reduce the amount scanning | |
2511 | * proportional to the original scan target. | |
2512 | */ | |
2513 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2514 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2515 | ||
1a501907 MG |
2516 | /* |
2517 | * It's just vindictive to attack the larger once the smaller | |
2518 | * has gone to zero. And given the way we stop scanning the | |
2519 | * smaller below, this makes sure that we only make one nudge | |
2520 | * towards proportionality once we've got nr_to_reclaim. | |
2521 | */ | |
2522 | if (!nr_file || !nr_anon) | |
2523 | break; | |
2524 | ||
e82e0561 MG |
2525 | if (nr_file > nr_anon) { |
2526 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2527 | targets[LRU_ACTIVE_ANON] + 1; | |
2528 | lru = LRU_BASE; | |
2529 | percentage = nr_anon * 100 / scan_target; | |
2530 | } else { | |
2531 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2532 | targets[LRU_ACTIVE_FILE] + 1; | |
2533 | lru = LRU_FILE; | |
2534 | percentage = nr_file * 100 / scan_target; | |
2535 | } | |
2536 | ||
2537 | /* Stop scanning the smaller of the LRU */ | |
2538 | nr[lru] = 0; | |
2539 | nr[lru + LRU_ACTIVE] = 0; | |
2540 | ||
2541 | /* | |
2542 | * Recalculate the other LRU scan count based on its original | |
2543 | * scan target and the percentage scanning already complete | |
2544 | */ | |
2545 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2546 | nr_scanned = targets[lru] - nr[lru]; | |
2547 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2548 | nr[lru] -= min(nr[lru], nr_scanned); | |
2549 | ||
2550 | lru += LRU_ACTIVE; | |
2551 | nr_scanned = targets[lru] - nr[lru]; | |
2552 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2553 | nr[lru] -= min(nr[lru], nr_scanned); | |
2554 | ||
2555 | scan_adjusted = true; | |
9b4f98cd JW |
2556 | } |
2557 | blk_finish_plug(&plug); | |
2558 | sc->nr_reclaimed += nr_reclaimed; | |
2559 | ||
2560 | /* | |
2561 | * Even if we did not try to evict anon pages at all, we want to | |
2562 | * rebalance the anon lru active/inactive ratio. | |
2563 | */ | |
b91ac374 | 2564 | if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON)) |
9b4f98cd JW |
2565 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2566 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
2567 | } |
2568 | ||
23b9da55 | 2569 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2570 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2571 | { |
d84da3f9 | 2572 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2573 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2574 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2575 | return true; |
2576 | ||
2577 | return false; | |
2578 | } | |
2579 | ||
3e7d3449 | 2580 | /* |
23b9da55 MG |
2581 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2582 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2583 | * true if more pages should be reclaimed such that when the page allocator | |
2584 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2585 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2586 | */ |
a9dd0a83 | 2587 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 | 2588 | unsigned long nr_reclaimed, |
3e7d3449 MG |
2589 | struct scan_control *sc) |
2590 | { | |
2591 | unsigned long pages_for_compaction; | |
2592 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 2593 | int z; |
3e7d3449 MG |
2594 | |
2595 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2596 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2597 | return false; |
2598 | ||
5ee04716 VB |
2599 | /* |
2600 | * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX | |
2601 | * number of pages that were scanned. This will return to the caller | |
2602 | * with the risk reclaim/compaction and the resulting allocation attempt | |
2603 | * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL | |
2604 | * allocations through requiring that the full LRU list has been scanned | |
2605 | * first, by assuming that zero delta of sc->nr_scanned means full LRU | |
2606 | * scan, but that approximation was wrong, and there were corner cases | |
2607 | * where always a non-zero amount of pages were scanned. | |
2608 | */ | |
2609 | if (!nr_reclaimed) | |
2610 | return false; | |
3e7d3449 | 2611 | |
3e7d3449 | 2612 | /* If compaction would go ahead or the allocation would succeed, stop */ |
a9dd0a83 MG |
2613 | for (z = 0; z <= sc->reclaim_idx; z++) { |
2614 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2615 | if (!managed_zone(zone)) |
a9dd0a83 MG |
2616 | continue; |
2617 | ||
2618 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 2619 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
2620 | case COMPACT_CONTINUE: |
2621 | return false; | |
2622 | default: | |
2623 | /* check next zone */ | |
2624 | ; | |
2625 | } | |
3e7d3449 | 2626 | } |
1c6c1597 HD |
2627 | |
2628 | /* | |
2629 | * If we have not reclaimed enough pages for compaction and the | |
2630 | * inactive lists are large enough, continue reclaiming | |
2631 | */ | |
2632 | pages_for_compaction = compact_gap(sc->order); | |
2633 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); | |
2634 | if (get_nr_swap_pages() > 0) | |
2635 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); | |
2636 | ||
5ee04716 | 2637 | return inactive_lru_pages > pages_for_compaction; |
3e7d3449 MG |
2638 | } |
2639 | ||
0f6a5cff | 2640 | static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 2641 | { |
0f6a5cff | 2642 | struct mem_cgroup *target_memcg = sc->target_mem_cgroup; |
d2af3397 | 2643 | struct mem_cgroup *memcg; |
1da177e4 | 2644 | |
0f6a5cff | 2645 | memcg = mem_cgroup_iter(target_memcg, NULL, NULL); |
d2af3397 | 2646 | do { |
afaf07a6 | 2647 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
d2af3397 JW |
2648 | unsigned long reclaimed; |
2649 | unsigned long scanned; | |
5660048c | 2650 | |
0f6a5cff | 2651 | switch (mem_cgroup_protected(target_memcg, memcg)) { |
d2af3397 JW |
2652 | case MEMCG_PROT_MIN: |
2653 | /* | |
2654 | * Hard protection. | |
2655 | * If there is no reclaimable memory, OOM. | |
2656 | */ | |
2657 | continue; | |
2658 | case MEMCG_PROT_LOW: | |
2659 | /* | |
2660 | * Soft protection. | |
2661 | * Respect the protection only as long as | |
2662 | * there is an unprotected supply | |
2663 | * of reclaimable memory from other cgroups. | |
2664 | */ | |
2665 | if (!sc->memcg_low_reclaim) { | |
2666 | sc->memcg_low_skipped = 1; | |
bf8d5d52 | 2667 | continue; |
241994ed | 2668 | } |
d2af3397 JW |
2669 | memcg_memory_event(memcg, MEMCG_LOW); |
2670 | break; | |
2671 | case MEMCG_PROT_NONE: | |
2672 | /* | |
2673 | * All protection thresholds breached. We may | |
2674 | * still choose to vary the scan pressure | |
2675 | * applied based on by how much the cgroup in | |
2676 | * question has exceeded its protection | |
2677 | * thresholds (see get_scan_count). | |
2678 | */ | |
2679 | break; | |
2680 | } | |
241994ed | 2681 | |
d2af3397 JW |
2682 | reclaimed = sc->nr_reclaimed; |
2683 | scanned = sc->nr_scanned; | |
afaf07a6 JW |
2684 | |
2685 | shrink_lruvec(lruvec, sc); | |
70ddf637 | 2686 | |
d2af3397 JW |
2687 | shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, |
2688 | sc->priority); | |
6b4f7799 | 2689 | |
d2af3397 JW |
2690 | /* Record the group's reclaim efficiency */ |
2691 | vmpressure(sc->gfp_mask, memcg, false, | |
2692 | sc->nr_scanned - scanned, | |
2693 | sc->nr_reclaimed - reclaimed); | |
70ddf637 | 2694 | |
0f6a5cff JW |
2695 | } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL))); |
2696 | } | |
2697 | ||
2698 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) | |
2699 | { | |
2700 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
0f6a5cff | 2701 | unsigned long nr_reclaimed, nr_scanned; |
1b05117d | 2702 | struct lruvec *target_lruvec; |
0f6a5cff | 2703 | bool reclaimable = false; |
b91ac374 | 2704 | unsigned long file; |
0f6a5cff | 2705 | |
1b05117d JW |
2706 | target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat); |
2707 | ||
0f6a5cff JW |
2708 | again: |
2709 | memset(&sc->nr, 0, sizeof(sc->nr)); | |
2710 | ||
2711 | nr_reclaimed = sc->nr_reclaimed; | |
2712 | nr_scanned = sc->nr_scanned; | |
2713 | ||
b91ac374 JW |
2714 | /* |
2715 | * Target desirable inactive:active list ratios for the anon | |
2716 | * and file LRU lists. | |
2717 | */ | |
2718 | if (!sc->force_deactivate) { | |
2719 | unsigned long refaults; | |
2720 | ||
2721 | if (inactive_is_low(target_lruvec, LRU_INACTIVE_ANON)) | |
2722 | sc->may_deactivate |= DEACTIVATE_ANON; | |
2723 | else | |
2724 | sc->may_deactivate &= ~DEACTIVATE_ANON; | |
2725 | ||
2726 | /* | |
2727 | * When refaults are being observed, it means a new | |
2728 | * workingset is being established. Deactivate to get | |
2729 | * rid of any stale active pages quickly. | |
2730 | */ | |
2731 | refaults = lruvec_page_state(target_lruvec, | |
2732 | WORKINGSET_ACTIVATE); | |
2733 | if (refaults != target_lruvec->refaults || | |
2734 | inactive_is_low(target_lruvec, LRU_INACTIVE_FILE)) | |
2735 | sc->may_deactivate |= DEACTIVATE_FILE; | |
2736 | else | |
2737 | sc->may_deactivate &= ~DEACTIVATE_FILE; | |
2738 | } else | |
2739 | sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE; | |
2740 | ||
2741 | /* | |
2742 | * If we have plenty of inactive file pages that aren't | |
2743 | * thrashing, try to reclaim those first before touching | |
2744 | * anonymous pages. | |
2745 | */ | |
2746 | file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE); | |
2747 | if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE)) | |
2748 | sc->cache_trim_mode = 1; | |
2749 | else | |
2750 | sc->cache_trim_mode = 0; | |
2751 | ||
53138cea JW |
2752 | /* |
2753 | * Prevent the reclaimer from falling into the cache trap: as | |
2754 | * cache pages start out inactive, every cache fault will tip | |
2755 | * the scan balance towards the file LRU. And as the file LRU | |
2756 | * shrinks, so does the window for rotation from references. | |
2757 | * This means we have a runaway feedback loop where a tiny | |
2758 | * thrashing file LRU becomes infinitely more attractive than | |
2759 | * anon pages. Try to detect this based on file LRU size. | |
2760 | */ | |
2761 | if (!cgroup_reclaim(sc)) { | |
53138cea | 2762 | unsigned long total_high_wmark = 0; |
b91ac374 JW |
2763 | unsigned long free, anon; |
2764 | int z; | |
53138cea JW |
2765 | |
2766 | free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); | |
2767 | file = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2768 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2769 | ||
2770 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2771 | struct zone *zone = &pgdat->node_zones[z]; | |
2772 | if (!managed_zone(zone)) | |
2773 | continue; | |
2774 | ||
2775 | total_high_wmark += high_wmark_pages(zone); | |
2776 | } | |
2777 | ||
b91ac374 JW |
2778 | /* |
2779 | * Consider anon: if that's low too, this isn't a | |
2780 | * runaway file reclaim problem, but rather just | |
2781 | * extreme pressure. Reclaim as per usual then. | |
2782 | */ | |
2783 | anon = node_page_state(pgdat, NR_INACTIVE_ANON); | |
2784 | ||
2785 | sc->file_is_tiny = | |
2786 | file + free <= total_high_wmark && | |
2787 | !(sc->may_deactivate & DEACTIVATE_ANON) && | |
2788 | anon >> sc->priority; | |
53138cea JW |
2789 | } |
2790 | ||
0f6a5cff | 2791 | shrink_node_memcgs(pgdat, sc); |
2344d7e4 | 2792 | |
d2af3397 JW |
2793 | if (reclaim_state) { |
2794 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2795 | reclaim_state->reclaimed_slab = 0; | |
2796 | } | |
d108c772 | 2797 | |
d2af3397 | 2798 | /* Record the subtree's reclaim efficiency */ |
1b05117d | 2799 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, |
d2af3397 JW |
2800 | sc->nr_scanned - nr_scanned, |
2801 | sc->nr_reclaimed - nr_reclaimed); | |
d108c772 | 2802 | |
d2af3397 JW |
2803 | if (sc->nr_reclaimed - nr_reclaimed) |
2804 | reclaimable = true; | |
d108c772 | 2805 | |
d2af3397 JW |
2806 | if (current_is_kswapd()) { |
2807 | /* | |
2808 | * If reclaim is isolating dirty pages under writeback, | |
2809 | * it implies that the long-lived page allocation rate | |
2810 | * is exceeding the page laundering rate. Either the | |
2811 | * global limits are not being effective at throttling | |
2812 | * processes due to the page distribution throughout | |
2813 | * zones or there is heavy usage of a slow backing | |
2814 | * device. The only option is to throttle from reclaim | |
2815 | * context which is not ideal as there is no guarantee | |
2816 | * the dirtying process is throttled in the same way | |
2817 | * balance_dirty_pages() manages. | |
2818 | * | |
2819 | * Once a node is flagged PGDAT_WRITEBACK, kswapd will | |
2820 | * count the number of pages under pages flagged for | |
2821 | * immediate reclaim and stall if any are encountered | |
2822 | * in the nr_immediate check below. | |
2823 | */ | |
2824 | if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken) | |
2825 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
d108c772 | 2826 | |
d2af3397 JW |
2827 | /* Allow kswapd to start writing pages during reclaim.*/ |
2828 | if (sc->nr.unqueued_dirty == sc->nr.file_taken) | |
2829 | set_bit(PGDAT_DIRTY, &pgdat->flags); | |
e3c1ac58 | 2830 | |
d108c772 | 2831 | /* |
d2af3397 JW |
2832 | * If kswapd scans pages marked marked for immediate |
2833 | * reclaim and under writeback (nr_immediate), it | |
2834 | * implies that pages are cycling through the LRU | |
2835 | * faster than they are written so also forcibly stall. | |
d108c772 | 2836 | */ |
d2af3397 JW |
2837 | if (sc->nr.immediate) |
2838 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
2839 | } | |
2840 | ||
2841 | /* | |
1b05117d JW |
2842 | * Tag a node/memcg as congested if all the dirty pages |
2843 | * scanned were backed by a congested BDI and | |
2844 | * wait_iff_congested will stall. | |
2845 | * | |
d2af3397 JW |
2846 | * Legacy memcg will stall in page writeback so avoid forcibly |
2847 | * stalling in wait_iff_congested(). | |
2848 | */ | |
1b05117d JW |
2849 | if ((current_is_kswapd() || |
2850 | (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) && | |
d2af3397 | 2851 | sc->nr.dirty && sc->nr.dirty == sc->nr.congested) |
1b05117d | 2852 | set_bit(LRUVEC_CONGESTED, &target_lruvec->flags); |
d2af3397 JW |
2853 | |
2854 | /* | |
2855 | * Stall direct reclaim for IO completions if underlying BDIs | |
2856 | * and node is congested. Allow kswapd to continue until it | |
2857 | * starts encountering unqueued dirty pages or cycling through | |
2858 | * the LRU too quickly. | |
2859 | */ | |
1b05117d JW |
2860 | if (!current_is_kswapd() && current_may_throttle() && |
2861 | !sc->hibernation_mode && | |
2862 | test_bit(LRUVEC_CONGESTED, &target_lruvec->flags)) | |
d2af3397 | 2863 | wait_iff_congested(BLK_RW_ASYNC, HZ/10); |
d108c772 | 2864 | |
d2af3397 JW |
2865 | if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
2866 | sc)) | |
2867 | goto again; | |
2344d7e4 | 2868 | |
c73322d0 JW |
2869 | /* |
2870 | * Kswapd gives up on balancing particular nodes after too | |
2871 | * many failures to reclaim anything from them and goes to | |
2872 | * sleep. On reclaim progress, reset the failure counter. A | |
2873 | * successful direct reclaim run will revive a dormant kswapd. | |
2874 | */ | |
2875 | if (reclaimable) | |
2876 | pgdat->kswapd_failures = 0; | |
2877 | ||
2344d7e4 | 2878 | return reclaimable; |
f16015fb JW |
2879 | } |
2880 | ||
53853e2d | 2881 | /* |
fdd4c614 VB |
2882 | * Returns true if compaction should go ahead for a costly-order request, or |
2883 | * the allocation would already succeed without compaction. Return false if we | |
2884 | * should reclaim first. | |
53853e2d | 2885 | */ |
4f588331 | 2886 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 2887 | { |
31483b6a | 2888 | unsigned long watermark; |
fdd4c614 | 2889 | enum compact_result suitable; |
fe4b1b24 | 2890 | |
fdd4c614 VB |
2891 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
2892 | if (suitable == COMPACT_SUCCESS) | |
2893 | /* Allocation should succeed already. Don't reclaim. */ | |
2894 | return true; | |
2895 | if (suitable == COMPACT_SKIPPED) | |
2896 | /* Compaction cannot yet proceed. Do reclaim. */ | |
2897 | return false; | |
fe4b1b24 | 2898 | |
53853e2d | 2899 | /* |
fdd4c614 VB |
2900 | * Compaction is already possible, but it takes time to run and there |
2901 | * are potentially other callers using the pages just freed. So proceed | |
2902 | * with reclaim to make a buffer of free pages available to give | |
2903 | * compaction a reasonable chance of completing and allocating the page. | |
2904 | * Note that we won't actually reclaim the whole buffer in one attempt | |
2905 | * as the target watermark in should_continue_reclaim() is lower. But if | |
2906 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 2907 | */ |
fdd4c614 | 2908 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 2909 | |
fdd4c614 | 2910 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
2911 | } |
2912 | ||
1da177e4 LT |
2913 | /* |
2914 | * This is the direct reclaim path, for page-allocating processes. We only | |
2915 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2916 | * request. | |
2917 | * | |
1da177e4 LT |
2918 | * If a zone is deemed to be full of pinned pages then just give it a light |
2919 | * scan then give up on it. | |
2920 | */ | |
0a0337e0 | 2921 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2922 | { |
dd1a239f | 2923 | struct zoneref *z; |
54a6eb5c | 2924 | struct zone *zone; |
0608f43d AM |
2925 | unsigned long nr_soft_reclaimed; |
2926 | unsigned long nr_soft_scanned; | |
619d0d76 | 2927 | gfp_t orig_mask; |
79dafcdc | 2928 | pg_data_t *last_pgdat = NULL; |
1cfb419b | 2929 | |
cc715d99 MG |
2930 | /* |
2931 | * If the number of buffer_heads in the machine exceeds the maximum | |
2932 | * allowed level, force direct reclaim to scan the highmem zone as | |
2933 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2934 | */ | |
619d0d76 | 2935 | orig_mask = sc->gfp_mask; |
b2e18757 | 2936 | if (buffer_heads_over_limit) { |
cc715d99 | 2937 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 2938 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 2939 | } |
cc715d99 | 2940 | |
d4debc66 | 2941 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 2942 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
2943 | /* |
2944 | * Take care memory controller reclaiming has small influence | |
2945 | * to global LRU. | |
2946 | */ | |
b5ead35e | 2947 | if (!cgroup_reclaim(sc)) { |
344736f2 VD |
2948 | if (!cpuset_zone_allowed(zone, |
2949 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2950 | continue; |
65ec02cb | 2951 | |
0b06496a JW |
2952 | /* |
2953 | * If we already have plenty of memory free for | |
2954 | * compaction in this zone, don't free any more. | |
2955 | * Even though compaction is invoked for any | |
2956 | * non-zero order, only frequent costly order | |
2957 | * reclamation is disruptive enough to become a | |
2958 | * noticeable problem, like transparent huge | |
2959 | * page allocations. | |
2960 | */ | |
2961 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2962 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 2963 | compaction_ready(zone, sc)) { |
0b06496a JW |
2964 | sc->compaction_ready = true; |
2965 | continue; | |
e0887c19 | 2966 | } |
0b06496a | 2967 | |
79dafcdc MG |
2968 | /* |
2969 | * Shrink each node in the zonelist once. If the | |
2970 | * zonelist is ordered by zone (not the default) then a | |
2971 | * node may be shrunk multiple times but in that case | |
2972 | * the user prefers lower zones being preserved. | |
2973 | */ | |
2974 | if (zone->zone_pgdat == last_pgdat) | |
2975 | continue; | |
2976 | ||
0608f43d AM |
2977 | /* |
2978 | * This steals pages from memory cgroups over softlimit | |
2979 | * and returns the number of reclaimed pages and | |
2980 | * scanned pages. This works for global memory pressure | |
2981 | * and balancing, not for a memcg's limit. | |
2982 | */ | |
2983 | nr_soft_scanned = 0; | |
ef8f2327 | 2984 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
2985 | sc->order, sc->gfp_mask, |
2986 | &nr_soft_scanned); | |
2987 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2988 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2989 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2990 | } |
408d8544 | 2991 | |
79dafcdc MG |
2992 | /* See comment about same check for global reclaim above */ |
2993 | if (zone->zone_pgdat == last_pgdat) | |
2994 | continue; | |
2995 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 2996 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 2997 | } |
e0c23279 | 2998 | |
619d0d76 WY |
2999 | /* |
3000 | * Restore to original mask to avoid the impact on the caller if we | |
3001 | * promoted it to __GFP_HIGHMEM. | |
3002 | */ | |
3003 | sc->gfp_mask = orig_mask; | |
1da177e4 | 3004 | } |
4f98a2fe | 3005 | |
b910718a | 3006 | static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat) |
2a2e4885 | 3007 | { |
b910718a JW |
3008 | struct lruvec *target_lruvec; |
3009 | unsigned long refaults; | |
2a2e4885 | 3010 | |
b910718a JW |
3011 | target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat); |
3012 | refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE); | |
3013 | target_lruvec->refaults = refaults; | |
2a2e4885 JW |
3014 | } |
3015 | ||
1da177e4 LT |
3016 | /* |
3017 | * This is the main entry point to direct page reclaim. | |
3018 | * | |
3019 | * If a full scan of the inactive list fails to free enough memory then we | |
3020 | * are "out of memory" and something needs to be killed. | |
3021 | * | |
3022 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
3023 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
3024 | * caller can't do much about. We kick the writeback threads and take explicit |
3025 | * naps in the hope that some of these pages can be written. But if the | |
3026 | * allocating task holds filesystem locks which prevent writeout this might not | |
3027 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
3028 | * |
3029 | * returns: 0, if no pages reclaimed | |
3030 | * else, the number of pages reclaimed | |
1da177e4 | 3031 | */ |
dac1d27b | 3032 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 3033 | struct scan_control *sc) |
1da177e4 | 3034 | { |
241994ed | 3035 | int initial_priority = sc->priority; |
2a2e4885 JW |
3036 | pg_data_t *last_pgdat; |
3037 | struct zoneref *z; | |
3038 | struct zone *zone; | |
241994ed | 3039 | retry: |
873b4771 KK |
3040 | delayacct_freepages_start(); |
3041 | ||
b5ead35e | 3042 | if (!cgroup_reclaim(sc)) |
7cc30fcf | 3043 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 3044 | |
9e3b2f8c | 3045 | do { |
70ddf637 AV |
3046 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
3047 | sc->priority); | |
66e1707b | 3048 | sc->nr_scanned = 0; |
0a0337e0 | 3049 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 3050 | |
bb21c7ce | 3051 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
3052 | break; |
3053 | ||
3054 | if (sc->compaction_ready) | |
3055 | break; | |
1da177e4 | 3056 | |
0e50ce3b MK |
3057 | /* |
3058 | * If we're getting trouble reclaiming, start doing | |
3059 | * writepage even in laptop mode. | |
3060 | */ | |
3061 | if (sc->priority < DEF_PRIORITY - 2) | |
3062 | sc->may_writepage = 1; | |
0b06496a | 3063 | } while (--sc->priority >= 0); |
bb21c7ce | 3064 | |
2a2e4885 JW |
3065 | last_pgdat = NULL; |
3066 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
3067 | sc->nodemask) { | |
3068 | if (zone->zone_pgdat == last_pgdat) | |
3069 | continue; | |
3070 | last_pgdat = zone->zone_pgdat; | |
1b05117d | 3071 | |
2a2e4885 | 3072 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); |
1b05117d JW |
3073 | |
3074 | if (cgroup_reclaim(sc)) { | |
3075 | struct lruvec *lruvec; | |
3076 | ||
3077 | lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, | |
3078 | zone->zone_pgdat); | |
3079 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
3080 | } | |
2a2e4885 JW |
3081 | } |
3082 | ||
873b4771 KK |
3083 | delayacct_freepages_end(); |
3084 | ||
bb21c7ce KM |
3085 | if (sc->nr_reclaimed) |
3086 | return sc->nr_reclaimed; | |
3087 | ||
0cee34fd | 3088 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 3089 | if (sc->compaction_ready) |
7335084d MG |
3090 | return 1; |
3091 | ||
b91ac374 JW |
3092 | /* |
3093 | * We make inactive:active ratio decisions based on the node's | |
3094 | * composition of memory, but a restrictive reclaim_idx or a | |
3095 | * memory.low cgroup setting can exempt large amounts of | |
3096 | * memory from reclaim. Neither of which are very common, so | |
3097 | * instead of doing costly eligibility calculations of the | |
3098 | * entire cgroup subtree up front, we assume the estimates are | |
3099 | * good, and retry with forcible deactivation if that fails. | |
3100 | */ | |
3101 | if (sc->skipped_deactivate) { | |
3102 | sc->priority = initial_priority; | |
3103 | sc->force_deactivate = 1; | |
3104 | sc->skipped_deactivate = 0; | |
3105 | goto retry; | |
3106 | } | |
3107 | ||
241994ed | 3108 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
d6622f63 | 3109 | if (sc->memcg_low_skipped) { |
241994ed | 3110 | sc->priority = initial_priority; |
b91ac374 JW |
3111 | sc->force_deactivate = 0; |
3112 | sc->skipped_deactivate = 0; | |
d6622f63 YX |
3113 | sc->memcg_low_reclaim = 1; |
3114 | sc->memcg_low_skipped = 0; | |
241994ed JW |
3115 | goto retry; |
3116 | } | |
3117 | ||
bb21c7ce | 3118 | return 0; |
1da177e4 LT |
3119 | } |
3120 | ||
c73322d0 | 3121 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
5515061d MG |
3122 | { |
3123 | struct zone *zone; | |
3124 | unsigned long pfmemalloc_reserve = 0; | |
3125 | unsigned long free_pages = 0; | |
3126 | int i; | |
3127 | bool wmark_ok; | |
3128 | ||
c73322d0 JW |
3129 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
3130 | return true; | |
3131 | ||
5515061d MG |
3132 | for (i = 0; i <= ZONE_NORMAL; i++) { |
3133 | zone = &pgdat->node_zones[i]; | |
d450abd8 JW |
3134 | if (!managed_zone(zone)) |
3135 | continue; | |
3136 | ||
3137 | if (!zone_reclaimable_pages(zone)) | |
675becce MG |
3138 | continue; |
3139 | ||
5515061d MG |
3140 | pfmemalloc_reserve += min_wmark_pages(zone); |
3141 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
3142 | } | |
3143 | ||
675becce MG |
3144 | /* If there are no reserves (unexpected config) then do not throttle */ |
3145 | if (!pfmemalloc_reserve) | |
3146 | return true; | |
3147 | ||
5515061d MG |
3148 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
3149 | ||
3150 | /* kswapd must be awake if processes are being throttled */ | |
3151 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
38087d9b | 3152 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
5515061d MG |
3153 | (enum zone_type)ZONE_NORMAL); |
3154 | wake_up_interruptible(&pgdat->kswapd_wait); | |
3155 | } | |
3156 | ||
3157 | return wmark_ok; | |
3158 | } | |
3159 | ||
3160 | /* | |
3161 | * Throttle direct reclaimers if backing storage is backed by the network | |
3162 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
3163 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
3164 | * when the low watermark is reached. |
3165 | * | |
3166 | * Returns true if a fatal signal was delivered during throttling. If this | |
3167 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 3168 | */ |
50694c28 | 3169 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
3170 | nodemask_t *nodemask) |
3171 | { | |
675becce | 3172 | struct zoneref *z; |
5515061d | 3173 | struct zone *zone; |
675becce | 3174 | pg_data_t *pgdat = NULL; |
5515061d MG |
3175 | |
3176 | /* | |
3177 | * Kernel threads should not be throttled as they may be indirectly | |
3178 | * responsible for cleaning pages necessary for reclaim to make forward | |
3179 | * progress. kjournald for example may enter direct reclaim while | |
3180 | * committing a transaction where throttling it could forcing other | |
3181 | * processes to block on log_wait_commit(). | |
3182 | */ | |
3183 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
3184 | goto out; |
3185 | ||
3186 | /* | |
3187 | * If a fatal signal is pending, this process should not throttle. | |
3188 | * It should return quickly so it can exit and free its memory | |
3189 | */ | |
3190 | if (fatal_signal_pending(current)) | |
3191 | goto out; | |
5515061d | 3192 | |
675becce MG |
3193 | /* |
3194 | * Check if the pfmemalloc reserves are ok by finding the first node | |
3195 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
3196 | * GFP_KERNEL will be required for allocating network buffers when | |
3197 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
3198 | * | |
3199 | * Throttling is based on the first usable node and throttled processes | |
3200 | * wait on a queue until kswapd makes progress and wakes them. There | |
3201 | * is an affinity then between processes waking up and where reclaim | |
3202 | * progress has been made assuming the process wakes on the same node. | |
3203 | * More importantly, processes running on remote nodes will not compete | |
3204 | * for remote pfmemalloc reserves and processes on different nodes | |
3205 | * should make reasonable progress. | |
3206 | */ | |
3207 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 3208 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
3209 | if (zone_idx(zone) > ZONE_NORMAL) |
3210 | continue; | |
3211 | ||
3212 | /* Throttle based on the first usable node */ | |
3213 | pgdat = zone->zone_pgdat; | |
c73322d0 | 3214 | if (allow_direct_reclaim(pgdat)) |
675becce MG |
3215 | goto out; |
3216 | break; | |
3217 | } | |
3218 | ||
3219 | /* If no zone was usable by the allocation flags then do not throttle */ | |
3220 | if (!pgdat) | |
50694c28 | 3221 | goto out; |
5515061d | 3222 | |
68243e76 MG |
3223 | /* Account for the throttling */ |
3224 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
3225 | ||
5515061d MG |
3226 | /* |
3227 | * If the caller cannot enter the filesystem, it's possible that it | |
3228 | * is due to the caller holding an FS lock or performing a journal | |
3229 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
3230 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
3231 | * blocked waiting on the same lock. Instead, throttle for up to a | |
3232 | * second before continuing. | |
3233 | */ | |
3234 | if (!(gfp_mask & __GFP_FS)) { | |
3235 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
c73322d0 | 3236 | allow_direct_reclaim(pgdat), HZ); |
50694c28 MG |
3237 | |
3238 | goto check_pending; | |
5515061d MG |
3239 | } |
3240 | ||
3241 | /* Throttle until kswapd wakes the process */ | |
3242 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
c73322d0 | 3243 | allow_direct_reclaim(pgdat)); |
50694c28 MG |
3244 | |
3245 | check_pending: | |
3246 | if (fatal_signal_pending(current)) | |
3247 | return true; | |
3248 | ||
3249 | out: | |
3250 | return false; | |
5515061d MG |
3251 | } |
3252 | ||
dac1d27b | 3253 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 3254 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 3255 | { |
33906bc5 | 3256 | unsigned long nr_reclaimed; |
66e1707b | 3257 | struct scan_control sc = { |
ee814fe2 | 3258 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
f2f43e56 | 3259 | .gfp_mask = current_gfp_context(gfp_mask), |
b2e18757 | 3260 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
3261 | .order = order, |
3262 | .nodemask = nodemask, | |
3263 | .priority = DEF_PRIORITY, | |
66e1707b | 3264 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3265 | .may_unmap = 1, |
2e2e4259 | 3266 | .may_swap = 1, |
66e1707b BS |
3267 | }; |
3268 | ||
bb451fdf GT |
3269 | /* |
3270 | * scan_control uses s8 fields for order, priority, and reclaim_idx. | |
3271 | * Confirm they are large enough for max values. | |
3272 | */ | |
3273 | BUILD_BUG_ON(MAX_ORDER > S8_MAX); | |
3274 | BUILD_BUG_ON(DEF_PRIORITY > S8_MAX); | |
3275 | BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX); | |
3276 | ||
5515061d | 3277 | /* |
50694c28 MG |
3278 | * Do not enter reclaim if fatal signal was delivered while throttled. |
3279 | * 1 is returned so that the page allocator does not OOM kill at this | |
3280 | * point. | |
5515061d | 3281 | */ |
f2f43e56 | 3282 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
5515061d MG |
3283 | return 1; |
3284 | ||
1732d2b0 | 3285 | set_task_reclaim_state(current, &sc.reclaim_state); |
3481c37f | 3286 | trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask); |
33906bc5 | 3287 | |
3115cd91 | 3288 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
3289 | |
3290 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
1732d2b0 | 3291 | set_task_reclaim_state(current, NULL); |
33906bc5 MG |
3292 | |
3293 | return nr_reclaimed; | |
66e1707b BS |
3294 | } |
3295 | ||
c255a458 | 3296 | #ifdef CONFIG_MEMCG |
66e1707b | 3297 | |
d2e5fb92 | 3298 | /* Only used by soft limit reclaim. Do not reuse for anything else. */ |
a9dd0a83 | 3299 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 3300 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 3301 | pg_data_t *pgdat, |
0ae5e89c | 3302 | unsigned long *nr_scanned) |
4e416953 | 3303 | { |
afaf07a6 | 3304 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
4e416953 | 3305 | struct scan_control sc = { |
b8f5c566 | 3306 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 3307 | .target_mem_cgroup = memcg, |
4e416953 BS |
3308 | .may_writepage = !laptop_mode, |
3309 | .may_unmap = 1, | |
b2e18757 | 3310 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 3311 | .may_swap = !noswap, |
4e416953 | 3312 | }; |
0ae5e89c | 3313 | |
d2e5fb92 MH |
3314 | WARN_ON_ONCE(!current->reclaim_state); |
3315 | ||
4e416953 BS |
3316 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3317 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 3318 | |
9e3b2f8c | 3319 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
3481c37f | 3320 | sc.gfp_mask); |
bdce6d9e | 3321 | |
4e416953 BS |
3322 | /* |
3323 | * NOTE: Although we can get the priority field, using it | |
3324 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 3325 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
3326 | * will pick up pages from other mem cgroup's as well. We hack |
3327 | * the priority and make it zero. | |
3328 | */ | |
afaf07a6 | 3329 | shrink_lruvec(lruvec, &sc); |
bdce6d9e KM |
3330 | |
3331 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
3332 | ||
0ae5e89c | 3333 | *nr_scanned = sc.nr_scanned; |
0308f7cf | 3334 | |
4e416953 BS |
3335 | return sc.nr_reclaimed; |
3336 | } | |
3337 | ||
72835c86 | 3338 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 3339 | unsigned long nr_pages, |
a7885eb8 | 3340 | gfp_t gfp_mask, |
b70a2a21 | 3341 | bool may_swap) |
66e1707b | 3342 | { |
bdce6d9e | 3343 | unsigned long nr_reclaimed; |
eb414681 | 3344 | unsigned long pflags; |
499118e9 | 3345 | unsigned int noreclaim_flag; |
66e1707b | 3346 | struct scan_control sc = { |
b70a2a21 | 3347 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 3348 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
a09ed5e0 | 3349 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
b2e18757 | 3350 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
3351 | .target_mem_cgroup = memcg, |
3352 | .priority = DEF_PRIORITY, | |
3353 | .may_writepage = !laptop_mode, | |
3354 | .may_unmap = 1, | |
b70a2a21 | 3355 | .may_swap = may_swap, |
a09ed5e0 | 3356 | }; |
889976db | 3357 | /* |
fa40d1ee SB |
3358 | * Traverse the ZONELIST_FALLBACK zonelist of the current node to put |
3359 | * equal pressure on all the nodes. This is based on the assumption that | |
3360 | * the reclaim does not bail out early. | |
889976db | 3361 | */ |
fa40d1ee | 3362 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
889976db | 3363 | |
fa40d1ee | 3364 | set_task_reclaim_state(current, &sc.reclaim_state); |
bdce6d9e | 3365 | |
3481c37f | 3366 | trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask); |
bdce6d9e | 3367 | |
eb414681 | 3368 | psi_memstall_enter(&pflags); |
499118e9 | 3369 | noreclaim_flag = memalloc_noreclaim_save(); |
eb414681 | 3370 | |
3115cd91 | 3371 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
eb414681 | 3372 | |
499118e9 | 3373 | memalloc_noreclaim_restore(noreclaim_flag); |
eb414681 | 3374 | psi_memstall_leave(&pflags); |
bdce6d9e KM |
3375 | |
3376 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
1732d2b0 | 3377 | set_task_reclaim_state(current, NULL); |
bdce6d9e KM |
3378 | |
3379 | return nr_reclaimed; | |
66e1707b BS |
3380 | } |
3381 | #endif | |
3382 | ||
1d82de61 | 3383 | static void age_active_anon(struct pglist_data *pgdat, |
ef8f2327 | 3384 | struct scan_control *sc) |
f16015fb | 3385 | { |
b95a2f2d | 3386 | struct mem_cgroup *memcg; |
b91ac374 | 3387 | struct lruvec *lruvec; |
f16015fb | 3388 | |
b95a2f2d JW |
3389 | if (!total_swap_pages) |
3390 | return; | |
3391 | ||
b91ac374 JW |
3392 | lruvec = mem_cgroup_lruvec(NULL, pgdat); |
3393 | if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON)) | |
3394 | return; | |
3395 | ||
b95a2f2d JW |
3396 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
3397 | do { | |
b91ac374 JW |
3398 | lruvec = mem_cgroup_lruvec(memcg, pgdat); |
3399 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
3400 | sc, LRU_ACTIVE_ANON); | |
b95a2f2d JW |
3401 | memcg = mem_cgroup_iter(NULL, memcg, NULL); |
3402 | } while (memcg); | |
f16015fb JW |
3403 | } |
3404 | ||
1c30844d MG |
3405 | static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx) |
3406 | { | |
3407 | int i; | |
3408 | struct zone *zone; | |
3409 | ||
3410 | /* | |
3411 | * Check for watermark boosts top-down as the higher zones | |
3412 | * are more likely to be boosted. Both watermarks and boosts | |
3413 | * should not be checked at the time time as reclaim would | |
3414 | * start prematurely when there is no boosting and a lower | |
3415 | * zone is balanced. | |
3416 | */ | |
3417 | for (i = classzone_idx; i >= 0; i--) { | |
3418 | zone = pgdat->node_zones + i; | |
3419 | if (!managed_zone(zone)) | |
3420 | continue; | |
3421 | ||
3422 | if (zone->watermark_boost) | |
3423 | return true; | |
3424 | } | |
3425 | ||
3426 | return false; | |
3427 | } | |
3428 | ||
e716f2eb MG |
3429 | /* |
3430 | * Returns true if there is an eligible zone balanced for the request order | |
3431 | * and classzone_idx | |
3432 | */ | |
3433 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) | |
60cefed4 | 3434 | { |
e716f2eb MG |
3435 | int i; |
3436 | unsigned long mark = -1; | |
3437 | struct zone *zone; | |
60cefed4 | 3438 | |
1c30844d MG |
3439 | /* |
3440 | * Check watermarks bottom-up as lower zones are more likely to | |
3441 | * meet watermarks. | |
3442 | */ | |
e716f2eb MG |
3443 | for (i = 0; i <= classzone_idx; i++) { |
3444 | zone = pgdat->node_zones + i; | |
6256c6b4 | 3445 | |
e716f2eb MG |
3446 | if (!managed_zone(zone)) |
3447 | continue; | |
3448 | ||
3449 | mark = high_wmark_pages(zone); | |
3450 | if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) | |
3451 | return true; | |
3452 | } | |
3453 | ||
3454 | /* | |
3455 | * If a node has no populated zone within classzone_idx, it does not | |
3456 | * need balancing by definition. This can happen if a zone-restricted | |
3457 | * allocation tries to wake a remote kswapd. | |
3458 | */ | |
3459 | if (mark == -1) | |
3460 | return true; | |
3461 | ||
3462 | return false; | |
60cefed4 JW |
3463 | } |
3464 | ||
631b6e08 MG |
3465 | /* Clear pgdat state for congested, dirty or under writeback. */ |
3466 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
3467 | { | |
1b05117d JW |
3468 | struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat); |
3469 | ||
3470 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
631b6e08 MG |
3471 | clear_bit(PGDAT_DIRTY, &pgdat->flags); |
3472 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
3473 | } | |
3474 | ||
5515061d MG |
3475 | /* |
3476 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3477 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3478 | * | |
3479 | * Returns true if kswapd is ready to sleep | |
3480 | */ | |
d9f21d42 | 3481 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f50de2d3 | 3482 | { |
5515061d | 3483 | /* |
9e5e3661 | 3484 | * The throttled processes are normally woken up in balance_pgdat() as |
c73322d0 | 3485 | * soon as allow_direct_reclaim() is true. But there is a potential |
9e5e3661 VB |
3486 | * race between when kswapd checks the watermarks and a process gets |
3487 | * throttled. There is also a potential race if processes get | |
3488 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3489 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3490 | * the wake up checks. If kswapd is going to sleep, no process should | |
3491 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3492 | * the wake up is premature, processes will wake kswapd and get | |
3493 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3494 | * that here we are under prepare_to_wait(). | |
5515061d | 3495 | */ |
9e5e3661 VB |
3496 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3497 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3498 | |
c73322d0 JW |
3499 | /* Hopeless node, leave it to direct reclaim */ |
3500 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3501 | return true; | |
3502 | ||
e716f2eb MG |
3503 | if (pgdat_balanced(pgdat, order, classzone_idx)) { |
3504 | clear_pgdat_congested(pgdat); | |
3505 | return true; | |
1d82de61 MG |
3506 | } |
3507 | ||
333b0a45 | 3508 | return false; |
f50de2d3 MG |
3509 | } |
3510 | ||
75485363 | 3511 | /* |
1d82de61 MG |
3512 | * kswapd shrinks a node of pages that are at or below the highest usable |
3513 | * zone that is currently unbalanced. | |
b8e83b94 MG |
3514 | * |
3515 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3516 | * reclaim or if the lack of progress was due to pages under writeback. |
3517 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3518 | */ |
1d82de61 | 3519 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 3520 | struct scan_control *sc) |
75485363 | 3521 | { |
1d82de61 MG |
3522 | struct zone *zone; |
3523 | int z; | |
75485363 | 3524 | |
1d82de61 MG |
3525 | /* Reclaim a number of pages proportional to the number of zones */ |
3526 | sc->nr_to_reclaim = 0; | |
970a39a3 | 3527 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 3528 | zone = pgdat->node_zones + z; |
6aa303de | 3529 | if (!managed_zone(zone)) |
1d82de61 | 3530 | continue; |
7c954f6d | 3531 | |
1d82de61 MG |
3532 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
3533 | } | |
7c954f6d MG |
3534 | |
3535 | /* | |
1d82de61 MG |
3536 | * Historically care was taken to put equal pressure on all zones but |
3537 | * now pressure is applied based on node LRU order. | |
7c954f6d | 3538 | */ |
970a39a3 | 3539 | shrink_node(pgdat, sc); |
283aba9f | 3540 | |
7c954f6d | 3541 | /* |
1d82de61 MG |
3542 | * Fragmentation may mean that the system cannot be rebalanced for |
3543 | * high-order allocations. If twice the allocation size has been | |
3544 | * reclaimed then recheck watermarks only at order-0 to prevent | |
3545 | * excessive reclaim. Assume that a process requested a high-order | |
3546 | * can direct reclaim/compact. | |
7c954f6d | 3547 | */ |
9861a62c | 3548 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 3549 | sc->order = 0; |
7c954f6d | 3550 | |
b8e83b94 | 3551 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3552 | } |
3553 | ||
1da177e4 | 3554 | /* |
1d82de61 MG |
3555 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
3556 | * that are eligible for use by the caller until at least one zone is | |
3557 | * balanced. | |
1da177e4 | 3558 | * |
1d82de61 | 3559 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
3560 | * |
3561 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 3562 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
8bb4e7a2 | 3563 | * found to have free_pages <= high_wmark_pages(zone), any page in that zone |
1d82de61 MG |
3564 | * or lower is eligible for reclaim until at least one usable zone is |
3565 | * balanced. | |
1da177e4 | 3566 | */ |
accf6242 | 3567 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
1da177e4 | 3568 | { |
1da177e4 | 3569 | int i; |
0608f43d AM |
3570 | unsigned long nr_soft_reclaimed; |
3571 | unsigned long nr_soft_scanned; | |
eb414681 | 3572 | unsigned long pflags; |
1c30844d MG |
3573 | unsigned long nr_boost_reclaim; |
3574 | unsigned long zone_boosts[MAX_NR_ZONES] = { 0, }; | |
3575 | bool boosted; | |
1d82de61 | 3576 | struct zone *zone; |
179e9639 AM |
3577 | struct scan_control sc = { |
3578 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3579 | .order = order, |
a6dc60f8 | 3580 | .may_unmap = 1, |
179e9639 | 3581 | }; |
93781325 | 3582 | |
1732d2b0 | 3583 | set_task_reclaim_state(current, &sc.reclaim_state); |
eb414681 | 3584 | psi_memstall_enter(&pflags); |
93781325 OS |
3585 | __fs_reclaim_acquire(); |
3586 | ||
f8891e5e | 3587 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3588 | |
1c30844d MG |
3589 | /* |
3590 | * Account for the reclaim boost. Note that the zone boost is left in | |
3591 | * place so that parallel allocations that are near the watermark will | |
3592 | * stall or direct reclaim until kswapd is finished. | |
3593 | */ | |
3594 | nr_boost_reclaim = 0; | |
3595 | for (i = 0; i <= classzone_idx; i++) { | |
3596 | zone = pgdat->node_zones + i; | |
3597 | if (!managed_zone(zone)) | |
3598 | continue; | |
3599 | ||
3600 | nr_boost_reclaim += zone->watermark_boost; | |
3601 | zone_boosts[i] = zone->watermark_boost; | |
3602 | } | |
3603 | boosted = nr_boost_reclaim; | |
3604 | ||
3605 | restart: | |
3606 | sc.priority = DEF_PRIORITY; | |
9e3b2f8c | 3607 | do { |
c73322d0 | 3608 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
b8e83b94 | 3609 | bool raise_priority = true; |
1c30844d | 3610 | bool balanced; |
93781325 | 3611 | bool ret; |
b8e83b94 | 3612 | |
84c7a777 | 3613 | sc.reclaim_idx = classzone_idx; |
1da177e4 | 3614 | |
86c79f6b | 3615 | /* |
84c7a777 MG |
3616 | * If the number of buffer_heads exceeds the maximum allowed |
3617 | * then consider reclaiming from all zones. This has a dual | |
3618 | * purpose -- on 64-bit systems it is expected that | |
3619 | * buffer_heads are stripped during active rotation. On 32-bit | |
3620 | * systems, highmem pages can pin lowmem memory and shrinking | |
3621 | * buffers can relieve lowmem pressure. Reclaim may still not | |
3622 | * go ahead if all eligible zones for the original allocation | |
3623 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
3624 | */ |
3625 | if (buffer_heads_over_limit) { | |
3626 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
3627 | zone = pgdat->node_zones + i; | |
6aa303de | 3628 | if (!managed_zone(zone)) |
86c79f6b | 3629 | continue; |
cc715d99 | 3630 | |
970a39a3 | 3631 | sc.reclaim_idx = i; |
e1dbeda6 | 3632 | break; |
1da177e4 | 3633 | } |
1da177e4 | 3634 | } |
dafcb73e | 3635 | |
86c79f6b | 3636 | /* |
1c30844d MG |
3637 | * If the pgdat is imbalanced then ignore boosting and preserve |
3638 | * the watermarks for a later time and restart. Note that the | |
3639 | * zone watermarks will be still reset at the end of balancing | |
3640 | * on the grounds that the normal reclaim should be enough to | |
3641 | * re-evaluate if boosting is required when kswapd next wakes. | |
3642 | */ | |
3643 | balanced = pgdat_balanced(pgdat, sc.order, classzone_idx); | |
3644 | if (!balanced && nr_boost_reclaim) { | |
3645 | nr_boost_reclaim = 0; | |
3646 | goto restart; | |
3647 | } | |
3648 | ||
3649 | /* | |
3650 | * If boosting is not active then only reclaim if there are no | |
3651 | * eligible zones. Note that sc.reclaim_idx is not used as | |
3652 | * buffer_heads_over_limit may have adjusted it. | |
86c79f6b | 3653 | */ |
1c30844d | 3654 | if (!nr_boost_reclaim && balanced) |
e716f2eb | 3655 | goto out; |
e1dbeda6 | 3656 | |
1c30844d MG |
3657 | /* Limit the priority of boosting to avoid reclaim writeback */ |
3658 | if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2) | |
3659 | raise_priority = false; | |
3660 | ||
3661 | /* | |
3662 | * Do not writeback or swap pages for boosted reclaim. The | |
3663 | * intent is to relieve pressure not issue sub-optimal IO | |
3664 | * from reclaim context. If no pages are reclaimed, the | |
3665 | * reclaim will be aborted. | |
3666 | */ | |
3667 | sc.may_writepage = !laptop_mode && !nr_boost_reclaim; | |
3668 | sc.may_swap = !nr_boost_reclaim; | |
1c30844d | 3669 | |
1d82de61 MG |
3670 | /* |
3671 | * Do some background aging of the anon list, to give | |
3672 | * pages a chance to be referenced before reclaiming. All | |
3673 | * pages are rotated regardless of classzone as this is | |
3674 | * about consistent aging. | |
3675 | */ | |
ef8f2327 | 3676 | age_active_anon(pgdat, &sc); |
1d82de61 | 3677 | |
b7ea3c41 MG |
3678 | /* |
3679 | * If we're getting trouble reclaiming, start doing writepage | |
3680 | * even in laptop mode. | |
3681 | */ | |
047d72c3 | 3682 | if (sc.priority < DEF_PRIORITY - 2) |
b7ea3c41 MG |
3683 | sc.may_writepage = 1; |
3684 | ||
1d82de61 MG |
3685 | /* Call soft limit reclaim before calling shrink_node. */ |
3686 | sc.nr_scanned = 0; | |
3687 | nr_soft_scanned = 0; | |
ef8f2327 | 3688 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
3689 | sc.gfp_mask, &nr_soft_scanned); |
3690 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3691 | ||
1da177e4 | 3692 | /* |
1d82de61 MG |
3693 | * There should be no need to raise the scanning priority if |
3694 | * enough pages are already being scanned that that high | |
3695 | * watermark would be met at 100% efficiency. | |
1da177e4 | 3696 | */ |
970a39a3 | 3697 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 3698 | raise_priority = false; |
5515061d MG |
3699 | |
3700 | /* | |
3701 | * If the low watermark is met there is no need for processes | |
3702 | * to be throttled on pfmemalloc_wait as they should not be | |
3703 | * able to safely make forward progress. Wake them | |
3704 | */ | |
3705 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
c73322d0 | 3706 | allow_direct_reclaim(pgdat)) |
cfc51155 | 3707 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3708 | |
b8e83b94 | 3709 | /* Check if kswapd should be suspending */ |
93781325 OS |
3710 | __fs_reclaim_release(); |
3711 | ret = try_to_freeze(); | |
3712 | __fs_reclaim_acquire(); | |
3713 | if (ret || kthread_should_stop()) | |
b8e83b94 | 3714 | break; |
8357376d | 3715 | |
73ce02e9 | 3716 | /* |
b8e83b94 MG |
3717 | * Raise priority if scanning rate is too low or there was no |
3718 | * progress in reclaiming pages | |
73ce02e9 | 3719 | */ |
c73322d0 | 3720 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
1c30844d MG |
3721 | nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed); |
3722 | ||
3723 | /* | |
3724 | * If reclaim made no progress for a boost, stop reclaim as | |
3725 | * IO cannot be queued and it could be an infinite loop in | |
3726 | * extreme circumstances. | |
3727 | */ | |
3728 | if (nr_boost_reclaim && !nr_reclaimed) | |
3729 | break; | |
3730 | ||
c73322d0 | 3731 | if (raise_priority || !nr_reclaimed) |
b8e83b94 | 3732 | sc.priority--; |
1d82de61 | 3733 | } while (sc.priority >= 1); |
1da177e4 | 3734 | |
c73322d0 JW |
3735 | if (!sc.nr_reclaimed) |
3736 | pgdat->kswapd_failures++; | |
3737 | ||
b8e83b94 | 3738 | out: |
1c30844d MG |
3739 | /* If reclaim was boosted, account for the reclaim done in this pass */ |
3740 | if (boosted) { | |
3741 | unsigned long flags; | |
3742 | ||
3743 | for (i = 0; i <= classzone_idx; i++) { | |
3744 | if (!zone_boosts[i]) | |
3745 | continue; | |
3746 | ||
3747 | /* Increments are under the zone lock */ | |
3748 | zone = pgdat->node_zones + i; | |
3749 | spin_lock_irqsave(&zone->lock, flags); | |
3750 | zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]); | |
3751 | spin_unlock_irqrestore(&zone->lock, flags); | |
3752 | } | |
3753 | ||
3754 | /* | |
3755 | * As there is now likely space, wakeup kcompact to defragment | |
3756 | * pageblocks. | |
3757 | */ | |
3758 | wakeup_kcompactd(pgdat, pageblock_order, classzone_idx); | |
3759 | } | |
3760 | ||
2a2e4885 | 3761 | snapshot_refaults(NULL, pgdat); |
93781325 | 3762 | __fs_reclaim_release(); |
eb414681 | 3763 | psi_memstall_leave(&pflags); |
1732d2b0 | 3764 | set_task_reclaim_state(current, NULL); |
e5ca8071 | 3765 | |
0abdee2b | 3766 | /* |
1d82de61 MG |
3767 | * Return the order kswapd stopped reclaiming at as |
3768 | * prepare_kswapd_sleep() takes it into account. If another caller | |
3769 | * entered the allocator slow path while kswapd was awake, order will | |
3770 | * remain at the higher level. | |
0abdee2b | 3771 | */ |
1d82de61 | 3772 | return sc.order; |
1da177e4 LT |
3773 | } |
3774 | ||
e716f2eb | 3775 | /* |
dffcac2c SB |
3776 | * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be |
3777 | * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not | |
3778 | * a valid index then either kswapd runs for first time or kswapd couldn't sleep | |
3779 | * after previous reclaim attempt (node is still unbalanced). In that case | |
3780 | * return the zone index of the previous kswapd reclaim cycle. | |
e716f2eb MG |
3781 | */ |
3782 | static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, | |
dffcac2c | 3783 | enum zone_type prev_classzone_idx) |
e716f2eb MG |
3784 | { |
3785 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
dffcac2c SB |
3786 | return prev_classzone_idx; |
3787 | return pgdat->kswapd_classzone_idx; | |
e716f2eb MG |
3788 | } |
3789 | ||
38087d9b MG |
3790 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
3791 | unsigned int classzone_idx) | |
f0bc0a60 KM |
3792 | { |
3793 | long remaining = 0; | |
3794 | DEFINE_WAIT(wait); | |
3795 | ||
3796 | if (freezing(current) || kthread_should_stop()) | |
3797 | return; | |
3798 | ||
3799 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3800 | ||
333b0a45 SG |
3801 | /* |
3802 | * Try to sleep for a short interval. Note that kcompactd will only be | |
3803 | * woken if it is possible to sleep for a short interval. This is | |
3804 | * deliberate on the assumption that if reclaim cannot keep an | |
3805 | * eligible zone balanced that it's also unlikely that compaction will | |
3806 | * succeed. | |
3807 | */ | |
d9f21d42 | 3808 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
fd901c95 VB |
3809 | /* |
3810 | * Compaction records what page blocks it recently failed to | |
3811 | * isolate pages from and skips them in the future scanning. | |
3812 | * When kswapd is going to sleep, it is reasonable to assume | |
3813 | * that pages and compaction may succeed so reset the cache. | |
3814 | */ | |
3815 | reset_isolation_suitable(pgdat); | |
3816 | ||
3817 | /* | |
3818 | * We have freed the memory, now we should compact it to make | |
3819 | * allocation of the requested order possible. | |
3820 | */ | |
38087d9b | 3821 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); |
fd901c95 | 3822 | |
f0bc0a60 | 3823 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
3824 | |
3825 | /* | |
3826 | * If woken prematurely then reset kswapd_classzone_idx and | |
3827 | * order. The values will either be from a wakeup request or | |
3828 | * the previous request that slept prematurely. | |
3829 | */ | |
3830 | if (remaining) { | |
e716f2eb | 3831 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
38087d9b MG |
3832 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); |
3833 | } | |
3834 | ||
f0bc0a60 KM |
3835 | finish_wait(&pgdat->kswapd_wait, &wait); |
3836 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3837 | } | |
3838 | ||
3839 | /* | |
3840 | * After a short sleep, check if it was a premature sleep. If not, then | |
3841 | * go fully to sleep until explicitly woken up. | |
3842 | */ | |
d9f21d42 MG |
3843 | if (!remaining && |
3844 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { | |
f0bc0a60 KM |
3845 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3846 | ||
3847 | /* | |
3848 | * vmstat counters are not perfectly accurate and the estimated | |
3849 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3850 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3851 | * watermarks being breached while under pressure, we reduce the | |
3852 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3853 | * them before going back to sleep. | |
3854 | */ | |
3855 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
3856 | |
3857 | if (!kthread_should_stop()) | |
3858 | schedule(); | |
3859 | ||
f0bc0a60 KM |
3860 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3861 | } else { | |
3862 | if (remaining) | |
3863 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3864 | else | |
3865 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3866 | } | |
3867 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3868 | } | |
3869 | ||
1da177e4 LT |
3870 | /* |
3871 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3872 | * from the init process. |
1da177e4 LT |
3873 | * |
3874 | * This basically trickles out pages so that we have _some_ | |
3875 | * free memory available even if there is no other activity | |
3876 | * that frees anything up. This is needed for things like routing | |
3877 | * etc, where we otherwise might have all activity going on in | |
3878 | * asynchronous contexts that cannot page things out. | |
3879 | * | |
3880 | * If there are applications that are active memory-allocators | |
3881 | * (most normal use), this basically shouldn't matter. | |
3882 | */ | |
3883 | static int kswapd(void *p) | |
3884 | { | |
e716f2eb MG |
3885 | unsigned int alloc_order, reclaim_order; |
3886 | unsigned int classzone_idx = MAX_NR_ZONES - 1; | |
1da177e4 LT |
3887 | pg_data_t *pgdat = (pg_data_t*)p; |
3888 | struct task_struct *tsk = current; | |
a70f7302 | 3889 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3890 | |
174596a0 | 3891 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3892 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3893 | |
3894 | /* | |
3895 | * Tell the memory management that we're a "memory allocator", | |
3896 | * and that if we need more memory we should get access to it | |
3897 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3898 | * never get caught in the normal page freeing logic. | |
3899 | * | |
3900 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3901 | * you need a small amount of memory in order to be able to | |
3902 | * page out something else, and this flag essentially protects | |
3903 | * us from recursively trying to free more memory as we're | |
3904 | * trying to free the first piece of memory in the first place). | |
3905 | */ | |
930d9152 | 3906 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3907 | set_freezable(); |
1da177e4 | 3908 | |
e716f2eb MG |
3909 | pgdat->kswapd_order = 0; |
3910 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; | |
1da177e4 | 3911 | for ( ; ; ) { |
6f6313d4 | 3912 | bool ret; |
3e1d1d28 | 3913 | |
e716f2eb MG |
3914 | alloc_order = reclaim_order = pgdat->kswapd_order; |
3915 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); | |
3916 | ||
38087d9b MG |
3917 | kswapd_try_sleep: |
3918 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
3919 | classzone_idx); | |
215ddd66 | 3920 | |
38087d9b MG |
3921 | /* Read the new order and classzone_idx */ |
3922 | alloc_order = reclaim_order = pgdat->kswapd_order; | |
dffcac2c | 3923 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
38087d9b | 3924 | pgdat->kswapd_order = 0; |
e716f2eb | 3925 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; |
1da177e4 | 3926 | |
8fe23e05 DR |
3927 | ret = try_to_freeze(); |
3928 | if (kthread_should_stop()) | |
3929 | break; | |
3930 | ||
3931 | /* | |
3932 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3933 | * after returning from the refrigerator | |
3934 | */ | |
38087d9b MG |
3935 | if (ret) |
3936 | continue; | |
3937 | ||
3938 | /* | |
3939 | * Reclaim begins at the requested order but if a high-order | |
3940 | * reclaim fails then kswapd falls back to reclaiming for | |
3941 | * order-0. If that happens, kswapd will consider sleeping | |
3942 | * for the order it finished reclaiming at (reclaim_order) | |
3943 | * but kcompactd is woken to compact for the original | |
3944 | * request (alloc_order). | |
3945 | */ | |
e5146b12 MG |
3946 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, |
3947 | alloc_order); | |
38087d9b MG |
3948 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); |
3949 | if (reclaim_order < alloc_order) | |
3950 | goto kswapd_try_sleep; | |
1da177e4 | 3951 | } |
b0a8cc58 | 3952 | |
71abdc15 | 3953 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
71abdc15 | 3954 | |
1da177e4 LT |
3955 | return 0; |
3956 | } | |
3957 | ||
3958 | /* | |
5ecd9d40 DR |
3959 | * A zone is low on free memory or too fragmented for high-order memory. If |
3960 | * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's | |
3961 | * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim | |
3962 | * has failed or is not needed, still wake up kcompactd if only compaction is | |
3963 | * needed. | |
1da177e4 | 3964 | */ |
5ecd9d40 DR |
3965 | void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, |
3966 | enum zone_type classzone_idx) | |
1da177e4 LT |
3967 | { |
3968 | pg_data_t *pgdat; | |
3969 | ||
6aa303de | 3970 | if (!managed_zone(zone)) |
1da177e4 LT |
3971 | return; |
3972 | ||
5ecd9d40 | 3973 | if (!cpuset_zone_allowed(zone, gfp_flags)) |
1da177e4 | 3974 | return; |
88f5acf8 | 3975 | pgdat = zone->zone_pgdat; |
dffcac2c SB |
3976 | |
3977 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
3978 | pgdat->kswapd_classzone_idx = classzone_idx; | |
3979 | else | |
3980 | pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, | |
3981 | classzone_idx); | |
38087d9b | 3982 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); |
8d0986e2 | 3983 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3984 | return; |
e1a55637 | 3985 | |
5ecd9d40 DR |
3986 | /* Hopeless node, leave it to direct reclaim if possible */ |
3987 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES || | |
1c30844d MG |
3988 | (pgdat_balanced(pgdat, order, classzone_idx) && |
3989 | !pgdat_watermark_boosted(pgdat, classzone_idx))) { | |
5ecd9d40 DR |
3990 | /* |
3991 | * There may be plenty of free memory available, but it's too | |
3992 | * fragmented for high-order allocations. Wake up kcompactd | |
3993 | * and rely on compaction_suitable() to determine if it's | |
3994 | * needed. If it fails, it will defer subsequent attempts to | |
3995 | * ratelimit its work. | |
3996 | */ | |
3997 | if (!(gfp_flags & __GFP_DIRECT_RECLAIM)) | |
3998 | wakeup_kcompactd(pgdat, order, classzone_idx); | |
e716f2eb | 3999 | return; |
5ecd9d40 | 4000 | } |
88f5acf8 | 4001 | |
5ecd9d40 DR |
4002 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order, |
4003 | gfp_flags); | |
8d0986e2 | 4004 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
4005 | } |
4006 | ||
c6f37f12 | 4007 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 4008 | /* |
7b51755c | 4009 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
4010 | * freed pages. |
4011 | * | |
4012 | * Rather than trying to age LRUs the aim is to preserve the overall | |
4013 | * LRU order by reclaiming preferentially | |
4014 | * inactive > active > active referenced > active mapped | |
1da177e4 | 4015 | */ |
7b51755c | 4016 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 4017 | { |
d6277db4 | 4018 | struct scan_control sc = { |
ee814fe2 | 4019 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 4020 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 4021 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 4022 | .priority = DEF_PRIORITY, |
d6277db4 | 4023 | .may_writepage = 1, |
ee814fe2 JW |
4024 | .may_unmap = 1, |
4025 | .may_swap = 1, | |
7b51755c | 4026 | .hibernation_mode = 1, |
1da177e4 | 4027 | }; |
a09ed5e0 | 4028 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c | 4029 | unsigned long nr_reclaimed; |
499118e9 | 4030 | unsigned int noreclaim_flag; |
1da177e4 | 4031 | |
d92a8cfc | 4032 | fs_reclaim_acquire(sc.gfp_mask); |
93781325 | 4033 | noreclaim_flag = memalloc_noreclaim_save(); |
1732d2b0 | 4034 | set_task_reclaim_state(current, &sc.reclaim_state); |
d6277db4 | 4035 | |
3115cd91 | 4036 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 4037 | |
1732d2b0 | 4038 | set_task_reclaim_state(current, NULL); |
499118e9 | 4039 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 4040 | fs_reclaim_release(sc.gfp_mask); |
d6277db4 | 4041 | |
7b51755c | 4042 | return nr_reclaimed; |
1da177e4 | 4043 | } |
c6f37f12 | 4044 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 4045 | |
1da177e4 LT |
4046 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
4047 | not required for correctness. So if the last cpu in a node goes | |
4048 | away, we get changed to run anywhere: as the first one comes back, | |
4049 | restore their cpu bindings. */ | |
517bbed9 | 4050 | static int kswapd_cpu_online(unsigned int cpu) |
1da177e4 | 4051 | { |
58c0a4a7 | 4052 | int nid; |
1da177e4 | 4053 | |
517bbed9 SAS |
4054 | for_each_node_state(nid, N_MEMORY) { |
4055 | pg_data_t *pgdat = NODE_DATA(nid); | |
4056 | const struct cpumask *mask; | |
a70f7302 | 4057 | |
517bbed9 | 4058 | mask = cpumask_of_node(pgdat->node_id); |
c5f59f08 | 4059 | |
517bbed9 SAS |
4060 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
4061 | /* One of our CPUs online: restore mask */ | |
4062 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
1da177e4 | 4063 | } |
517bbed9 | 4064 | return 0; |
1da177e4 | 4065 | } |
1da177e4 | 4066 | |
3218ae14 YG |
4067 | /* |
4068 | * This kswapd start function will be called by init and node-hot-add. | |
4069 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
4070 | */ | |
4071 | int kswapd_run(int nid) | |
4072 | { | |
4073 | pg_data_t *pgdat = NODE_DATA(nid); | |
4074 | int ret = 0; | |
4075 | ||
4076 | if (pgdat->kswapd) | |
4077 | return 0; | |
4078 | ||
4079 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
4080 | if (IS_ERR(pgdat->kswapd)) { | |
4081 | /* failure at boot is fatal */ | |
c6202adf | 4082 | BUG_ON(system_state < SYSTEM_RUNNING); |
d5dc0ad9 GS |
4083 | pr_err("Failed to start kswapd on node %d\n", nid); |
4084 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 4085 | pgdat->kswapd = NULL; |
3218ae14 YG |
4086 | } |
4087 | return ret; | |
4088 | } | |
4089 | ||
8fe23e05 | 4090 | /* |
d8adde17 | 4091 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 4092 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
4093 | */ |
4094 | void kswapd_stop(int nid) | |
4095 | { | |
4096 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
4097 | ||
d8adde17 | 4098 | if (kswapd) { |
8fe23e05 | 4099 | kthread_stop(kswapd); |
d8adde17 JL |
4100 | NODE_DATA(nid)->kswapd = NULL; |
4101 | } | |
8fe23e05 DR |
4102 | } |
4103 | ||
1da177e4 LT |
4104 | static int __init kswapd_init(void) |
4105 | { | |
517bbed9 | 4106 | int nid, ret; |
69e05944 | 4107 | |
1da177e4 | 4108 | swap_setup(); |
48fb2e24 | 4109 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 4110 | kswapd_run(nid); |
517bbed9 SAS |
4111 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
4112 | "mm/vmscan:online", kswapd_cpu_online, | |
4113 | NULL); | |
4114 | WARN_ON(ret < 0); | |
1da177e4 LT |
4115 | return 0; |
4116 | } | |
4117 | ||
4118 | module_init(kswapd_init) | |
9eeff239 CL |
4119 | |
4120 | #ifdef CONFIG_NUMA | |
4121 | /* | |
a5f5f91d | 4122 | * Node reclaim mode |
9eeff239 | 4123 | * |
a5f5f91d | 4124 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 4125 | * the watermarks. |
9eeff239 | 4126 | */ |
a5f5f91d | 4127 | int node_reclaim_mode __read_mostly; |
9eeff239 | 4128 | |
1b2ffb78 | 4129 | #define RECLAIM_OFF 0 |
7d03431c | 4130 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 | 4131 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
95bbc0c7 | 4132 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
1b2ffb78 | 4133 | |
a92f7126 | 4134 | /* |
a5f5f91d | 4135 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
4136 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
4137 | * a zone. | |
4138 | */ | |
a5f5f91d | 4139 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 4140 | |
9614634f | 4141 | /* |
a5f5f91d | 4142 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
4143 | * occur. |
4144 | */ | |
4145 | int sysctl_min_unmapped_ratio = 1; | |
4146 | ||
0ff38490 CL |
4147 | /* |
4148 | * If the number of slab pages in a zone grows beyond this percentage then | |
4149 | * slab reclaim needs to occur. | |
4150 | */ | |
4151 | int sysctl_min_slab_ratio = 5; | |
4152 | ||
11fb9989 | 4153 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 4154 | { |
11fb9989 MG |
4155 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
4156 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
4157 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
4158 | |
4159 | /* | |
4160 | * It's possible for there to be more file mapped pages than | |
4161 | * accounted for by the pages on the file LRU lists because | |
4162 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
4163 | */ | |
4164 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
4165 | } | |
4166 | ||
4167 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 4168 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 4169 | { |
d031a157 AM |
4170 | unsigned long nr_pagecache_reclaimable; |
4171 | unsigned long delta = 0; | |
90afa5de MG |
4172 | |
4173 | /* | |
95bbc0c7 | 4174 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 4175 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 4176 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
4177 | * a better estimate |
4178 | */ | |
a5f5f91d MG |
4179 | if (node_reclaim_mode & RECLAIM_UNMAP) |
4180 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 4181 | else |
a5f5f91d | 4182 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
4183 | |
4184 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
4185 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
4186 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
4187 | |
4188 | /* Watch for any possible underflows due to delta */ | |
4189 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
4190 | delta = nr_pagecache_reclaimable; | |
4191 | ||
4192 | return nr_pagecache_reclaimable - delta; | |
4193 | } | |
4194 | ||
9eeff239 | 4195 | /* |
a5f5f91d | 4196 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 4197 | */ |
a5f5f91d | 4198 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 4199 | { |
7fb2d46d | 4200 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 4201 | const unsigned long nr_pages = 1 << order; |
9eeff239 | 4202 | struct task_struct *p = current; |
499118e9 | 4203 | unsigned int noreclaim_flag; |
179e9639 | 4204 | struct scan_control sc = { |
62b726c1 | 4205 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
f2f43e56 | 4206 | .gfp_mask = current_gfp_context(gfp_mask), |
bd2f6199 | 4207 | .order = order, |
a5f5f91d MG |
4208 | .priority = NODE_RECLAIM_PRIORITY, |
4209 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
4210 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 4211 | .may_swap = 1, |
f2f43e56 | 4212 | .reclaim_idx = gfp_zone(gfp_mask), |
179e9639 | 4213 | }; |
9eeff239 | 4214 | |
132bb8cf YS |
4215 | trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order, |
4216 | sc.gfp_mask); | |
4217 | ||
9eeff239 | 4218 | cond_resched(); |
93781325 | 4219 | fs_reclaim_acquire(sc.gfp_mask); |
d4f7796e | 4220 | /* |
95bbc0c7 | 4221 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 4222 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 4223 | * and RECLAIM_UNMAP. |
d4f7796e | 4224 | */ |
499118e9 VB |
4225 | noreclaim_flag = memalloc_noreclaim_save(); |
4226 | p->flags |= PF_SWAPWRITE; | |
1732d2b0 | 4227 | set_task_reclaim_state(p, &sc.reclaim_state); |
c84db23c | 4228 | |
a5f5f91d | 4229 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
0ff38490 | 4230 | /* |
894befec | 4231 | * Free memory by calling shrink node with increasing |
0ff38490 CL |
4232 | * priorities until we have enough memory freed. |
4233 | */ | |
0ff38490 | 4234 | do { |
970a39a3 | 4235 | shrink_node(pgdat, &sc); |
9e3b2f8c | 4236 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 4237 | } |
c84db23c | 4238 | |
1732d2b0 | 4239 | set_task_reclaim_state(p, NULL); |
499118e9 VB |
4240 | current->flags &= ~PF_SWAPWRITE; |
4241 | memalloc_noreclaim_restore(noreclaim_flag); | |
93781325 | 4242 | fs_reclaim_release(sc.gfp_mask); |
132bb8cf YS |
4243 | |
4244 | trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed); | |
4245 | ||
a79311c1 | 4246 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 4247 | } |
179e9639 | 4248 | |
a5f5f91d | 4249 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 4250 | { |
d773ed6b | 4251 | int ret; |
179e9639 AM |
4252 | |
4253 | /* | |
a5f5f91d | 4254 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 4255 | * slab pages if we are over the defined limits. |
34aa1330 | 4256 | * |
9614634f CL |
4257 | * A small portion of unmapped file backed pages is needed for |
4258 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
4259 | * thrown out if the node is overallocated. So we do not reclaim |
4260 | * if less than a specified percentage of the node is used by | |
9614634f | 4261 | * unmapped file backed pages. |
179e9639 | 4262 | */ |
a5f5f91d | 4263 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
385386cf | 4264 | node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) |
a5f5f91d | 4265 | return NODE_RECLAIM_FULL; |
179e9639 AM |
4266 | |
4267 | /* | |
d773ed6b | 4268 | * Do not scan if the allocation should not be delayed. |
179e9639 | 4269 | */ |
d0164adc | 4270 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 4271 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
4272 | |
4273 | /* | |
a5f5f91d | 4274 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
4275 | * have associated processors. This will favor the local processor |
4276 | * over remote processors and spread off node memory allocations | |
4277 | * as wide as possible. | |
4278 | */ | |
a5f5f91d MG |
4279 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
4280 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 4281 | |
a5f5f91d MG |
4282 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
4283 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 4284 | |
a5f5f91d MG |
4285 | ret = __node_reclaim(pgdat, gfp_mask, order); |
4286 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 4287 | |
24cf7251 MG |
4288 | if (!ret) |
4289 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
4290 | ||
d773ed6b | 4291 | return ret; |
179e9639 | 4292 | } |
9eeff239 | 4293 | #endif |
894bc310 | 4294 | |
894bc310 LS |
4295 | /* |
4296 | * page_evictable - test whether a page is evictable | |
4297 | * @page: the page to test | |
894bc310 LS |
4298 | * |
4299 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 4300 | * lists vs unevictable list. |
894bc310 LS |
4301 | * |
4302 | * Reasons page might not be evictable: | |
ba9ddf49 | 4303 | * (1) page's mapping marked unevictable |
b291f000 | 4304 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 4305 | * |
894bc310 | 4306 | */ |
39b5f29a | 4307 | int page_evictable(struct page *page) |
894bc310 | 4308 | { |
e92bb4dd YH |
4309 | int ret; |
4310 | ||
4311 | /* Prevent address_space of inode and swap cache from being freed */ | |
4312 | rcu_read_lock(); | |
4313 | ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); | |
4314 | rcu_read_unlock(); | |
4315 | return ret; | |
894bc310 | 4316 | } |
89e004ea LS |
4317 | |
4318 | /** | |
64e3d12f KHY |
4319 | * check_move_unevictable_pages - check pages for evictability and move to |
4320 | * appropriate zone lru list | |
4321 | * @pvec: pagevec with lru pages to check | |
89e004ea | 4322 | * |
64e3d12f KHY |
4323 | * Checks pages for evictability, if an evictable page is in the unevictable |
4324 | * lru list, moves it to the appropriate evictable lru list. This function | |
4325 | * should be only used for lru pages. | |
89e004ea | 4326 | */ |
64e3d12f | 4327 | void check_move_unevictable_pages(struct pagevec *pvec) |
89e004ea | 4328 | { |
925b7673 | 4329 | struct lruvec *lruvec; |
785b99fe | 4330 | struct pglist_data *pgdat = NULL; |
24513264 HD |
4331 | int pgscanned = 0; |
4332 | int pgrescued = 0; | |
4333 | int i; | |
89e004ea | 4334 | |
64e3d12f KHY |
4335 | for (i = 0; i < pvec->nr; i++) { |
4336 | struct page *page = pvec->pages[i]; | |
785b99fe | 4337 | struct pglist_data *pagepgdat = page_pgdat(page); |
89e004ea | 4338 | |
24513264 | 4339 | pgscanned++; |
785b99fe MG |
4340 | if (pagepgdat != pgdat) { |
4341 | if (pgdat) | |
4342 | spin_unlock_irq(&pgdat->lru_lock); | |
4343 | pgdat = pagepgdat; | |
4344 | spin_lock_irq(&pgdat->lru_lock); | |
24513264 | 4345 | } |
785b99fe | 4346 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
89e004ea | 4347 | |
24513264 HD |
4348 | if (!PageLRU(page) || !PageUnevictable(page)) |
4349 | continue; | |
89e004ea | 4350 | |
39b5f29a | 4351 | if (page_evictable(page)) { |
24513264 HD |
4352 | enum lru_list lru = page_lru_base_type(page); |
4353 | ||
309381fe | 4354 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 4355 | ClearPageUnevictable(page); |
fa9add64 HD |
4356 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
4357 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 4358 | pgrescued++; |
89e004ea | 4359 | } |
24513264 | 4360 | } |
89e004ea | 4361 | |
785b99fe | 4362 | if (pgdat) { |
24513264 HD |
4363 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
4364 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
785b99fe | 4365 | spin_unlock_irq(&pgdat->lru_lock); |
89e004ea | 4366 | } |
89e004ea | 4367 | } |
64e3d12f | 4368 | EXPORT_SYMBOL_GPL(check_move_unevictable_pages); |