]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
1da177e4 | 2 | /* |
1da177e4 LT |
3 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds |
4 | * | |
5 | * Swap reorganised 29.12.95, Stephen Tweedie. | |
6 | * kswapd added: 7.1.96 sct | |
7 | * Removed kswapd_ctl limits, and swap out as many pages as needed | |
8 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. | |
9 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). | |
10 | * Multiqueue VM started 5.8.00, Rik van Riel. | |
11 | */ | |
12 | ||
b1de0d13 MH |
13 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
14 | ||
1da177e4 | 15 | #include <linux/mm.h> |
5b3cc15a | 16 | #include <linux/sched/mm.h> |
1da177e4 | 17 | #include <linux/module.h> |
5a0e3ad6 | 18 | #include <linux/gfp.h> |
1da177e4 LT |
19 | #include <linux/kernel_stat.h> |
20 | #include <linux/swap.h> | |
21 | #include <linux/pagemap.h> | |
22 | #include <linux/init.h> | |
23 | #include <linux/highmem.h> | |
70ddf637 | 24 | #include <linux/vmpressure.h> |
e129b5c2 | 25 | #include <linux/vmstat.h> |
1da177e4 LT |
26 | #include <linux/file.h> |
27 | #include <linux/writeback.h> | |
28 | #include <linux/blkdev.h> | |
29 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
30 | buffer_heads_over_limit */ | |
31 | #include <linux/mm_inline.h> | |
1da177e4 LT |
32 | #include <linux/backing-dev.h> |
33 | #include <linux/rmap.h> | |
34 | #include <linux/topology.h> | |
35 | #include <linux/cpu.h> | |
36 | #include <linux/cpuset.h> | |
3e7d3449 | 37 | #include <linux/compaction.h> |
1da177e4 LT |
38 | #include <linux/notifier.h> |
39 | #include <linux/rwsem.h> | |
248a0301 | 40 | #include <linux/delay.h> |
3218ae14 | 41 | #include <linux/kthread.h> |
7dfb7103 | 42 | #include <linux/freezer.h> |
66e1707b | 43 | #include <linux/memcontrol.h> |
873b4771 | 44 | #include <linux/delayacct.h> |
af936a16 | 45 | #include <linux/sysctl.h> |
929bea7c | 46 | #include <linux/oom.h> |
64e3d12f | 47 | #include <linux/pagevec.h> |
268bb0ce | 48 | #include <linux/prefetch.h> |
b1de0d13 | 49 | #include <linux/printk.h> |
f9fe48be | 50 | #include <linux/dax.h> |
eb414681 | 51 | #include <linux/psi.h> |
1da177e4 LT |
52 | |
53 | #include <asm/tlbflush.h> | |
54 | #include <asm/div64.h> | |
55 | ||
56 | #include <linux/swapops.h> | |
117aad1e | 57 | #include <linux/balloon_compaction.h> |
1da177e4 | 58 | |
0f8053a5 NP |
59 | #include "internal.h" |
60 | ||
33906bc5 MG |
61 | #define CREATE_TRACE_POINTS |
62 | #include <trace/events/vmscan.h> | |
63 | ||
1da177e4 | 64 | struct scan_control { |
22fba335 KM |
65 | /* How many pages shrink_list() should reclaim */ |
66 | unsigned long nr_to_reclaim; | |
67 | ||
ee814fe2 JW |
68 | /* |
69 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
70 | * are scanned. | |
71 | */ | |
72 | nodemask_t *nodemask; | |
9e3b2f8c | 73 | |
f16015fb JW |
74 | /* |
75 | * The memory cgroup that hit its limit and as a result is the | |
76 | * primary target of this reclaim invocation. | |
77 | */ | |
78 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 79 | |
7cf111bc JW |
80 | /* |
81 | * Scan pressure balancing between anon and file LRUs | |
82 | */ | |
83 | unsigned long anon_cost; | |
84 | unsigned long file_cost; | |
85 | ||
b91ac374 JW |
86 | /* Can active pages be deactivated as part of reclaim? */ |
87 | #define DEACTIVATE_ANON 1 | |
88 | #define DEACTIVATE_FILE 2 | |
89 | unsigned int may_deactivate:2; | |
90 | unsigned int force_deactivate:1; | |
91 | unsigned int skipped_deactivate:1; | |
92 | ||
1276ad68 | 93 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
ee814fe2 JW |
94 | unsigned int may_writepage:1; |
95 | ||
96 | /* Can mapped pages be reclaimed? */ | |
97 | unsigned int may_unmap:1; | |
98 | ||
99 | /* Can pages be swapped as part of reclaim? */ | |
100 | unsigned int may_swap:1; | |
101 | ||
d6622f63 YX |
102 | /* |
103 | * Cgroups are not reclaimed below their configured memory.low, | |
104 | * unless we threaten to OOM. If any cgroups are skipped due to | |
105 | * memory.low and nothing was reclaimed, go back for memory.low. | |
106 | */ | |
107 | unsigned int memcg_low_reclaim:1; | |
108 | unsigned int memcg_low_skipped:1; | |
241994ed | 109 | |
ee814fe2 JW |
110 | unsigned int hibernation_mode:1; |
111 | ||
112 | /* One of the zones is ready for compaction */ | |
113 | unsigned int compaction_ready:1; | |
114 | ||
b91ac374 JW |
115 | /* There is easily reclaimable cold cache in the current node */ |
116 | unsigned int cache_trim_mode:1; | |
117 | ||
53138cea JW |
118 | /* The file pages on the current node are dangerously low */ |
119 | unsigned int file_is_tiny:1; | |
120 | ||
bb451fdf GT |
121 | /* Allocation order */ |
122 | s8 order; | |
123 | ||
124 | /* Scan (total_size >> priority) pages at once */ | |
125 | s8 priority; | |
126 | ||
127 | /* The highest zone to isolate pages for reclaim from */ | |
128 | s8 reclaim_idx; | |
129 | ||
130 | /* This context's GFP mask */ | |
131 | gfp_t gfp_mask; | |
132 | ||
ee814fe2 JW |
133 | /* Incremented by the number of inactive pages that were scanned */ |
134 | unsigned long nr_scanned; | |
135 | ||
136 | /* Number of pages freed so far during a call to shrink_zones() */ | |
137 | unsigned long nr_reclaimed; | |
d108c772 AR |
138 | |
139 | struct { | |
140 | unsigned int dirty; | |
141 | unsigned int unqueued_dirty; | |
142 | unsigned int congested; | |
143 | unsigned int writeback; | |
144 | unsigned int immediate; | |
145 | unsigned int file_taken; | |
146 | unsigned int taken; | |
147 | } nr; | |
e5ca8071 YS |
148 | |
149 | /* for recording the reclaimed slab by now */ | |
150 | struct reclaim_state reclaim_state; | |
1da177e4 LT |
151 | }; |
152 | ||
1da177e4 LT |
153 | #ifdef ARCH_HAS_PREFETCHW |
154 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
155 | do { \ | |
156 | if ((_page)->lru.prev != _base) { \ | |
157 | struct page *prev; \ | |
158 | \ | |
159 | prev = lru_to_page(&(_page->lru)); \ | |
160 | prefetchw(&prev->_field); \ | |
161 | } \ | |
162 | } while (0) | |
163 | #else | |
164 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
165 | #endif | |
166 | ||
167 | /* | |
c843966c | 168 | * From 0 .. 200. Higher means more swappy. |
1da177e4 LT |
169 | */ |
170 | int vm_swappiness = 60; | |
1da177e4 | 171 | |
0a432dcb YS |
172 | static void set_task_reclaim_state(struct task_struct *task, |
173 | struct reclaim_state *rs) | |
174 | { | |
175 | /* Check for an overwrite */ | |
176 | WARN_ON_ONCE(rs && task->reclaim_state); | |
177 | ||
178 | /* Check for the nulling of an already-nulled member */ | |
179 | WARN_ON_ONCE(!rs && !task->reclaim_state); | |
180 | ||
181 | task->reclaim_state = rs; | |
182 | } | |
183 | ||
1da177e4 LT |
184 | static LIST_HEAD(shrinker_list); |
185 | static DECLARE_RWSEM(shrinker_rwsem); | |
186 | ||
0a432dcb | 187 | #ifdef CONFIG_MEMCG |
7e010df5 KT |
188 | /* |
189 | * We allow subsystems to populate their shrinker-related | |
190 | * LRU lists before register_shrinker_prepared() is called | |
191 | * for the shrinker, since we don't want to impose | |
192 | * restrictions on their internal registration order. | |
193 | * In this case shrink_slab_memcg() may find corresponding | |
194 | * bit is set in the shrinkers map. | |
195 | * | |
196 | * This value is used by the function to detect registering | |
197 | * shrinkers and to skip do_shrink_slab() calls for them. | |
198 | */ | |
199 | #define SHRINKER_REGISTERING ((struct shrinker *)~0UL) | |
200 | ||
b4c2b231 KT |
201 | static DEFINE_IDR(shrinker_idr); |
202 | static int shrinker_nr_max; | |
203 | ||
204 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) | |
205 | { | |
206 | int id, ret = -ENOMEM; | |
207 | ||
208 | down_write(&shrinker_rwsem); | |
209 | /* This may call shrinker, so it must use down_read_trylock() */ | |
7e010df5 | 210 | id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL); |
b4c2b231 KT |
211 | if (id < 0) |
212 | goto unlock; | |
213 | ||
0a4465d3 KT |
214 | if (id >= shrinker_nr_max) { |
215 | if (memcg_expand_shrinker_maps(id)) { | |
216 | idr_remove(&shrinker_idr, id); | |
217 | goto unlock; | |
218 | } | |
219 | ||
b4c2b231 | 220 | shrinker_nr_max = id + 1; |
0a4465d3 | 221 | } |
b4c2b231 KT |
222 | shrinker->id = id; |
223 | ret = 0; | |
224 | unlock: | |
225 | up_write(&shrinker_rwsem); | |
226 | return ret; | |
227 | } | |
228 | ||
229 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
230 | { | |
231 | int id = shrinker->id; | |
232 | ||
233 | BUG_ON(id < 0); | |
234 | ||
235 | down_write(&shrinker_rwsem); | |
236 | idr_remove(&shrinker_idr, id); | |
237 | up_write(&shrinker_rwsem); | |
238 | } | |
b4c2b231 | 239 | |
b5ead35e | 240 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 241 | { |
b5ead35e | 242 | return sc->target_mem_cgroup; |
89b5fae5 | 243 | } |
97c9341f TH |
244 | |
245 | /** | |
b5ead35e | 246 | * writeback_throttling_sane - is the usual dirty throttling mechanism available? |
97c9341f TH |
247 | * @sc: scan_control in question |
248 | * | |
249 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
250 | * completely broken with the legacy memcg and direct stalling in | |
251 | * shrink_page_list() is used for throttling instead, which lacks all the | |
252 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
253 | * allocation and configurability. | |
254 | * | |
255 | * This function tests whether the vmscan currently in progress can assume | |
256 | * that the normal dirty throttling mechanism is operational. | |
257 | */ | |
b5ead35e | 258 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f | 259 | { |
b5ead35e | 260 | if (!cgroup_reclaim(sc)) |
97c9341f TH |
261 | return true; |
262 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 263 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
264 | return true; |
265 | #endif | |
266 | return false; | |
267 | } | |
91a45470 | 268 | #else |
0a432dcb YS |
269 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) |
270 | { | |
271 | return 0; | |
272 | } | |
273 | ||
274 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
275 | { | |
276 | } | |
277 | ||
b5ead35e | 278 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 279 | { |
b5ead35e | 280 | return false; |
89b5fae5 | 281 | } |
97c9341f | 282 | |
b5ead35e | 283 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f TH |
284 | { |
285 | return true; | |
286 | } | |
91a45470 KH |
287 | #endif |
288 | ||
5a1c84b4 MG |
289 | /* |
290 | * This misses isolated pages which are not accounted for to save counters. | |
291 | * As the data only determines if reclaim or compaction continues, it is | |
292 | * not expected that isolated pages will be a dominating factor. | |
293 | */ | |
294 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
295 | { | |
296 | unsigned long nr; | |
297 | ||
298 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
299 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
300 | if (get_nr_swap_pages() > 0) | |
301 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + | |
302 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
303 | ||
304 | return nr; | |
305 | } | |
306 | ||
fd538803 MH |
307 | /** |
308 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
309 | * @lruvec: lru vector | |
310 | * @lru: lru to use | |
311 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) | |
312 | */ | |
2091339d YZ |
313 | static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, |
314 | int zone_idx) | |
c9f299d9 | 315 | { |
de3b0150 | 316 | unsigned long size = 0; |
fd538803 MH |
317 | int zid; |
318 | ||
de3b0150 | 319 | for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) { |
fd538803 | 320 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; |
c9f299d9 | 321 | |
fd538803 MH |
322 | if (!managed_zone(zone)) |
323 | continue; | |
324 | ||
325 | if (!mem_cgroup_disabled()) | |
de3b0150 | 326 | size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid); |
fd538803 | 327 | else |
de3b0150 | 328 | size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru); |
fd538803 | 329 | } |
de3b0150 | 330 | return size; |
b4536f0c MH |
331 | } |
332 | ||
1da177e4 | 333 | /* |
1d3d4437 | 334 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 335 | */ |
8e04944f | 336 | int prealloc_shrinker(struct shrinker *shrinker) |
1da177e4 | 337 | { |
b9726c26 | 338 | unsigned int size = sizeof(*shrinker->nr_deferred); |
1d3d4437 | 339 | |
1d3d4437 GC |
340 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
341 | size *= nr_node_ids; | |
342 | ||
343 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
344 | if (!shrinker->nr_deferred) | |
345 | return -ENOMEM; | |
b4c2b231 KT |
346 | |
347 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) { | |
348 | if (prealloc_memcg_shrinker(shrinker)) | |
349 | goto free_deferred; | |
350 | } | |
351 | ||
8e04944f | 352 | return 0; |
b4c2b231 KT |
353 | |
354 | free_deferred: | |
355 | kfree(shrinker->nr_deferred); | |
356 | shrinker->nr_deferred = NULL; | |
357 | return -ENOMEM; | |
8e04944f TH |
358 | } |
359 | ||
360 | void free_prealloced_shrinker(struct shrinker *shrinker) | |
361 | { | |
b4c2b231 KT |
362 | if (!shrinker->nr_deferred) |
363 | return; | |
364 | ||
365 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) | |
366 | unregister_memcg_shrinker(shrinker); | |
367 | ||
8e04944f TH |
368 | kfree(shrinker->nr_deferred); |
369 | shrinker->nr_deferred = NULL; | |
370 | } | |
1d3d4437 | 371 | |
8e04944f TH |
372 | void register_shrinker_prepared(struct shrinker *shrinker) |
373 | { | |
8e1f936b RR |
374 | down_write(&shrinker_rwsem); |
375 | list_add_tail(&shrinker->list, &shrinker_list); | |
42a9a53b | 376 | #ifdef CONFIG_MEMCG |
8df4a44c KT |
377 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) |
378 | idr_replace(&shrinker_idr, shrinker, shrinker->id); | |
7e010df5 | 379 | #endif |
8e1f936b | 380 | up_write(&shrinker_rwsem); |
8e04944f TH |
381 | } |
382 | ||
383 | int register_shrinker(struct shrinker *shrinker) | |
384 | { | |
385 | int err = prealloc_shrinker(shrinker); | |
386 | ||
387 | if (err) | |
388 | return err; | |
389 | register_shrinker_prepared(shrinker); | |
1d3d4437 | 390 | return 0; |
1da177e4 | 391 | } |
8e1f936b | 392 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
393 | |
394 | /* | |
395 | * Remove one | |
396 | */ | |
8e1f936b | 397 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 | 398 | { |
bb422a73 TH |
399 | if (!shrinker->nr_deferred) |
400 | return; | |
b4c2b231 KT |
401 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) |
402 | unregister_memcg_shrinker(shrinker); | |
1da177e4 LT |
403 | down_write(&shrinker_rwsem); |
404 | list_del(&shrinker->list); | |
405 | up_write(&shrinker_rwsem); | |
ae393321 | 406 | kfree(shrinker->nr_deferred); |
bb422a73 | 407 | shrinker->nr_deferred = NULL; |
1da177e4 | 408 | } |
8e1f936b | 409 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
410 | |
411 | #define SHRINK_BATCH 128 | |
1d3d4437 | 412 | |
cb731d6c | 413 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
9092c71b | 414 | struct shrinker *shrinker, int priority) |
1d3d4437 GC |
415 | { |
416 | unsigned long freed = 0; | |
417 | unsigned long long delta; | |
418 | long total_scan; | |
d5bc5fd3 | 419 | long freeable; |
1d3d4437 GC |
420 | long nr; |
421 | long new_nr; | |
422 | int nid = shrinkctl->nid; | |
423 | long batch_size = shrinker->batch ? shrinker->batch | |
424 | : SHRINK_BATCH; | |
5f33a080 | 425 | long scanned = 0, next_deferred; |
1d3d4437 | 426 | |
ac7fb3ad KT |
427 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
428 | nid = 0; | |
429 | ||
d5bc5fd3 | 430 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
9b996468 KT |
431 | if (freeable == 0 || freeable == SHRINK_EMPTY) |
432 | return freeable; | |
1d3d4437 GC |
433 | |
434 | /* | |
435 | * copy the current shrinker scan count into a local variable | |
436 | * and zero it so that other concurrent shrinker invocations | |
437 | * don't also do this scanning work. | |
438 | */ | |
439 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
440 | ||
441 | total_scan = nr; | |
4b85afbd JW |
442 | if (shrinker->seeks) { |
443 | delta = freeable >> priority; | |
444 | delta *= 4; | |
445 | do_div(delta, shrinker->seeks); | |
446 | } else { | |
447 | /* | |
448 | * These objects don't require any IO to create. Trim | |
449 | * them aggressively under memory pressure to keep | |
450 | * them from causing refetches in the IO caches. | |
451 | */ | |
452 | delta = freeable / 2; | |
453 | } | |
172b06c3 | 454 | |
1d3d4437 GC |
455 | total_scan += delta; |
456 | if (total_scan < 0) { | |
d75f773c | 457 | pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n", |
a0b02131 | 458 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 459 | total_scan = freeable; |
5f33a080 SL |
460 | next_deferred = nr; |
461 | } else | |
462 | next_deferred = total_scan; | |
1d3d4437 GC |
463 | |
464 | /* | |
465 | * We need to avoid excessive windup on filesystem shrinkers | |
466 | * due to large numbers of GFP_NOFS allocations causing the | |
467 | * shrinkers to return -1 all the time. This results in a large | |
468 | * nr being built up so when a shrink that can do some work | |
469 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 470 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
471 | * memory. |
472 | * | |
473 | * Hence only allow the shrinker to scan the entire cache when | |
474 | * a large delta change is calculated directly. | |
475 | */ | |
d5bc5fd3 VD |
476 | if (delta < freeable / 4) |
477 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
478 | |
479 | /* | |
480 | * Avoid risking looping forever due to too large nr value: | |
481 | * never try to free more than twice the estimate number of | |
482 | * freeable entries. | |
483 | */ | |
d5bc5fd3 VD |
484 | if (total_scan > freeable * 2) |
485 | total_scan = freeable * 2; | |
1d3d4437 GC |
486 | |
487 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
9092c71b | 488 | freeable, delta, total_scan, priority); |
1d3d4437 | 489 | |
0b1fb40a VD |
490 | /* |
491 | * Normally, we should not scan less than batch_size objects in one | |
492 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
493 | * than batch_size objects in total and we are really tight on memory, | |
494 | * we will try to reclaim all available objects, otherwise we can end | |
495 | * up failing allocations although there are plenty of reclaimable | |
496 | * objects spread over several slabs with usage less than the | |
497 | * batch_size. | |
498 | * | |
499 | * We detect the "tight on memory" situations by looking at the total | |
500 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 501 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
502 | * scanning at high prio and therefore should try to reclaim as much as |
503 | * possible. | |
504 | */ | |
505 | while (total_scan >= batch_size || | |
d5bc5fd3 | 506 | total_scan >= freeable) { |
a0b02131 | 507 | unsigned long ret; |
0b1fb40a | 508 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 509 | |
0b1fb40a | 510 | shrinkctl->nr_to_scan = nr_to_scan; |
d460acb5 | 511 | shrinkctl->nr_scanned = nr_to_scan; |
a0b02131 DC |
512 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
513 | if (ret == SHRINK_STOP) | |
514 | break; | |
515 | freed += ret; | |
1d3d4437 | 516 | |
d460acb5 CW |
517 | count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); |
518 | total_scan -= shrinkctl->nr_scanned; | |
519 | scanned += shrinkctl->nr_scanned; | |
1d3d4437 GC |
520 | |
521 | cond_resched(); | |
522 | } | |
523 | ||
5f33a080 SL |
524 | if (next_deferred >= scanned) |
525 | next_deferred -= scanned; | |
526 | else | |
527 | next_deferred = 0; | |
1d3d4437 GC |
528 | /* |
529 | * move the unused scan count back into the shrinker in a | |
530 | * manner that handles concurrent updates. If we exhausted the | |
531 | * scan, there is no need to do an update. | |
532 | */ | |
5f33a080 SL |
533 | if (next_deferred > 0) |
534 | new_nr = atomic_long_add_return(next_deferred, | |
1d3d4437 GC |
535 | &shrinker->nr_deferred[nid]); |
536 | else | |
537 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
538 | ||
8efb4b59 | 539 | trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 540 | return freed; |
1495f230 YH |
541 | } |
542 | ||
0a432dcb | 543 | #ifdef CONFIG_MEMCG |
b0dedc49 KT |
544 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
545 | struct mem_cgroup *memcg, int priority) | |
546 | { | |
547 | struct memcg_shrinker_map *map; | |
b8e57efa KT |
548 | unsigned long ret, freed = 0; |
549 | int i; | |
b0dedc49 | 550 | |
0a432dcb | 551 | if (!mem_cgroup_online(memcg)) |
b0dedc49 KT |
552 | return 0; |
553 | ||
554 | if (!down_read_trylock(&shrinker_rwsem)) | |
555 | return 0; | |
556 | ||
557 | map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map, | |
558 | true); | |
559 | if (unlikely(!map)) | |
560 | goto unlock; | |
561 | ||
562 | for_each_set_bit(i, map->map, shrinker_nr_max) { | |
563 | struct shrink_control sc = { | |
564 | .gfp_mask = gfp_mask, | |
565 | .nid = nid, | |
566 | .memcg = memcg, | |
567 | }; | |
568 | struct shrinker *shrinker; | |
569 | ||
570 | shrinker = idr_find(&shrinker_idr, i); | |
7e010df5 KT |
571 | if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) { |
572 | if (!shrinker) | |
573 | clear_bit(i, map->map); | |
b0dedc49 KT |
574 | continue; |
575 | } | |
576 | ||
0a432dcb YS |
577 | /* Call non-slab shrinkers even though kmem is disabled */ |
578 | if (!memcg_kmem_enabled() && | |
579 | !(shrinker->flags & SHRINKER_NONSLAB)) | |
580 | continue; | |
581 | ||
b0dedc49 | 582 | ret = do_shrink_slab(&sc, shrinker, priority); |
f90280d6 KT |
583 | if (ret == SHRINK_EMPTY) { |
584 | clear_bit(i, map->map); | |
585 | /* | |
586 | * After the shrinker reported that it had no objects to | |
587 | * free, but before we cleared the corresponding bit in | |
588 | * the memcg shrinker map, a new object might have been | |
589 | * added. To make sure, we have the bit set in this | |
590 | * case, we invoke the shrinker one more time and reset | |
591 | * the bit if it reports that it is not empty anymore. | |
592 | * The memory barrier here pairs with the barrier in | |
593 | * memcg_set_shrinker_bit(): | |
594 | * | |
595 | * list_lru_add() shrink_slab_memcg() | |
596 | * list_add_tail() clear_bit() | |
597 | * <MB> <MB> | |
598 | * set_bit() do_shrink_slab() | |
599 | */ | |
600 | smp_mb__after_atomic(); | |
601 | ret = do_shrink_slab(&sc, shrinker, priority); | |
602 | if (ret == SHRINK_EMPTY) | |
603 | ret = 0; | |
604 | else | |
605 | memcg_set_shrinker_bit(memcg, nid, i); | |
606 | } | |
b0dedc49 KT |
607 | freed += ret; |
608 | ||
609 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
610 | freed = freed ? : 1; | |
611 | break; | |
612 | } | |
613 | } | |
614 | unlock: | |
615 | up_read(&shrinker_rwsem); | |
616 | return freed; | |
617 | } | |
0a432dcb | 618 | #else /* CONFIG_MEMCG */ |
b0dedc49 KT |
619 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
620 | struct mem_cgroup *memcg, int priority) | |
621 | { | |
622 | return 0; | |
623 | } | |
0a432dcb | 624 | #endif /* CONFIG_MEMCG */ |
b0dedc49 | 625 | |
6b4f7799 | 626 | /** |
cb731d6c | 627 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
628 | * @gfp_mask: allocation context |
629 | * @nid: node whose slab caches to target | |
cb731d6c | 630 | * @memcg: memory cgroup whose slab caches to target |
9092c71b | 631 | * @priority: the reclaim priority |
1da177e4 | 632 | * |
6b4f7799 | 633 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 634 | * |
6b4f7799 JW |
635 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
636 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 637 | * |
aeed1d32 VD |
638 | * @memcg specifies the memory cgroup to target. Unaware shrinkers |
639 | * are called only if it is the root cgroup. | |
cb731d6c | 640 | * |
9092c71b JB |
641 | * @priority is sc->priority, we take the number of objects and >> by priority |
642 | * in order to get the scan target. | |
b15e0905 | 643 | * |
6b4f7799 | 644 | * Returns the number of reclaimed slab objects. |
1da177e4 | 645 | */ |
cb731d6c VD |
646 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
647 | struct mem_cgroup *memcg, | |
9092c71b | 648 | int priority) |
1da177e4 | 649 | { |
b8e57efa | 650 | unsigned long ret, freed = 0; |
1da177e4 LT |
651 | struct shrinker *shrinker; |
652 | ||
fa1e512f YS |
653 | /* |
654 | * The root memcg might be allocated even though memcg is disabled | |
655 | * via "cgroup_disable=memory" boot parameter. This could make | |
656 | * mem_cgroup_is_root() return false, then just run memcg slab | |
657 | * shrink, but skip global shrink. This may result in premature | |
658 | * oom. | |
659 | */ | |
660 | if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg)) | |
b0dedc49 | 661 | return shrink_slab_memcg(gfp_mask, nid, memcg, priority); |
cb731d6c | 662 | |
e830c63a | 663 | if (!down_read_trylock(&shrinker_rwsem)) |
f06590bd | 664 | goto out; |
1da177e4 LT |
665 | |
666 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
667 | struct shrink_control sc = { |
668 | .gfp_mask = gfp_mask, | |
669 | .nid = nid, | |
cb731d6c | 670 | .memcg = memcg, |
6b4f7799 | 671 | }; |
ec97097b | 672 | |
9b996468 KT |
673 | ret = do_shrink_slab(&sc, shrinker, priority); |
674 | if (ret == SHRINK_EMPTY) | |
675 | ret = 0; | |
676 | freed += ret; | |
e496612c MK |
677 | /* |
678 | * Bail out if someone want to register a new shrinker to | |
55b65a57 | 679 | * prevent the registration from being stalled for long periods |
e496612c MK |
680 | * by parallel ongoing shrinking. |
681 | */ | |
682 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
683 | freed = freed ? : 1; | |
684 | break; | |
685 | } | |
1da177e4 | 686 | } |
6b4f7799 | 687 | |
1da177e4 | 688 | up_read(&shrinker_rwsem); |
f06590bd MK |
689 | out: |
690 | cond_resched(); | |
24f7c6b9 | 691 | return freed; |
1da177e4 LT |
692 | } |
693 | ||
cb731d6c VD |
694 | void drop_slab_node(int nid) |
695 | { | |
696 | unsigned long freed; | |
697 | ||
698 | do { | |
699 | struct mem_cgroup *memcg = NULL; | |
700 | ||
069c411d CZ |
701 | if (fatal_signal_pending(current)) |
702 | return; | |
703 | ||
cb731d6c | 704 | freed = 0; |
aeed1d32 | 705 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
cb731d6c | 706 | do { |
9092c71b | 707 | freed += shrink_slab(GFP_KERNEL, nid, memcg, 0); |
cb731d6c VD |
708 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); |
709 | } while (freed > 10); | |
710 | } | |
711 | ||
712 | void drop_slab(void) | |
713 | { | |
714 | int nid; | |
715 | ||
716 | for_each_online_node(nid) | |
717 | drop_slab_node(nid); | |
718 | } | |
719 | ||
1da177e4 LT |
720 | static inline int is_page_cache_freeable(struct page *page) |
721 | { | |
ceddc3a5 JW |
722 | /* |
723 | * A freeable page cache page is referenced only by the caller | |
67891fff MW |
724 | * that isolated the page, the page cache and optional buffer |
725 | * heads at page->private. | |
ceddc3a5 | 726 | */ |
3efe62e4 | 727 | int page_cache_pins = thp_nr_pages(page); |
67891fff | 728 | return page_count(page) - page_has_private(page) == 1 + page_cache_pins; |
1da177e4 LT |
729 | } |
730 | ||
cb16556d | 731 | static int may_write_to_inode(struct inode *inode) |
1da177e4 | 732 | { |
930d9152 | 733 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 734 | return 1; |
703c2708 | 735 | if (!inode_write_congested(inode)) |
1da177e4 | 736 | return 1; |
703c2708 | 737 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
738 | return 1; |
739 | return 0; | |
740 | } | |
741 | ||
742 | /* | |
743 | * We detected a synchronous write error writing a page out. Probably | |
744 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
745 | * fsync(), msync() or close(). | |
746 | * | |
747 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
748 | * prevents it from being freed up. But we have a ref on the page and once | |
749 | * that page is locked, the mapping is pinned. | |
750 | * | |
751 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
752 | * __GFP_FS. | |
753 | */ | |
754 | static void handle_write_error(struct address_space *mapping, | |
755 | struct page *page, int error) | |
756 | { | |
7eaceacc | 757 | lock_page(page); |
3e9f45bd GC |
758 | if (page_mapping(page) == mapping) |
759 | mapping_set_error(mapping, error); | |
1da177e4 LT |
760 | unlock_page(page); |
761 | } | |
762 | ||
04e62a29 CL |
763 | /* possible outcome of pageout() */ |
764 | typedef enum { | |
765 | /* failed to write page out, page is locked */ | |
766 | PAGE_KEEP, | |
767 | /* move page to the active list, page is locked */ | |
768 | PAGE_ACTIVATE, | |
769 | /* page has been sent to the disk successfully, page is unlocked */ | |
770 | PAGE_SUCCESS, | |
771 | /* page is clean and locked */ | |
772 | PAGE_CLEAN, | |
773 | } pageout_t; | |
774 | ||
1da177e4 | 775 | /* |
1742f19f AM |
776 | * pageout is called by shrink_page_list() for each dirty page. |
777 | * Calls ->writepage(). | |
1da177e4 | 778 | */ |
cb16556d | 779 | static pageout_t pageout(struct page *page, struct address_space *mapping) |
1da177e4 LT |
780 | { |
781 | /* | |
782 | * If the page is dirty, only perform writeback if that write | |
783 | * will be non-blocking. To prevent this allocation from being | |
784 | * stalled by pagecache activity. But note that there may be | |
785 | * stalls if we need to run get_block(). We could test | |
786 | * PagePrivate for that. | |
787 | * | |
8174202b | 788 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
789 | * this page's queue, we can perform writeback even if that |
790 | * will block. | |
791 | * | |
792 | * If the page is swapcache, write it back even if that would | |
793 | * block, for some throttling. This happens by accident, because | |
794 | * swap_backing_dev_info is bust: it doesn't reflect the | |
795 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
796 | */ |
797 | if (!is_page_cache_freeable(page)) | |
798 | return PAGE_KEEP; | |
799 | if (!mapping) { | |
800 | /* | |
801 | * Some data journaling orphaned pages can have | |
802 | * page->mapping == NULL while being dirty with clean buffers. | |
803 | */ | |
266cf658 | 804 | if (page_has_private(page)) { |
1da177e4 LT |
805 | if (try_to_free_buffers(page)) { |
806 | ClearPageDirty(page); | |
b1de0d13 | 807 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
808 | return PAGE_CLEAN; |
809 | } | |
810 | } | |
811 | return PAGE_KEEP; | |
812 | } | |
813 | if (mapping->a_ops->writepage == NULL) | |
814 | return PAGE_ACTIVATE; | |
cb16556d | 815 | if (!may_write_to_inode(mapping->host)) |
1da177e4 LT |
816 | return PAGE_KEEP; |
817 | ||
818 | if (clear_page_dirty_for_io(page)) { | |
819 | int res; | |
820 | struct writeback_control wbc = { | |
821 | .sync_mode = WB_SYNC_NONE, | |
822 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
823 | .range_start = 0, |
824 | .range_end = LLONG_MAX, | |
1da177e4 LT |
825 | .for_reclaim = 1, |
826 | }; | |
827 | ||
828 | SetPageReclaim(page); | |
829 | res = mapping->a_ops->writepage(page, &wbc); | |
830 | if (res < 0) | |
831 | handle_write_error(mapping, page, res); | |
994fc28c | 832 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
833 | ClearPageReclaim(page); |
834 | return PAGE_ACTIVATE; | |
835 | } | |
c661b078 | 836 | |
1da177e4 LT |
837 | if (!PageWriteback(page)) { |
838 | /* synchronous write or broken a_ops? */ | |
839 | ClearPageReclaim(page); | |
840 | } | |
3aa23851 | 841 | trace_mm_vmscan_writepage(page); |
c4a25635 | 842 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
843 | return PAGE_SUCCESS; |
844 | } | |
845 | ||
846 | return PAGE_CLEAN; | |
847 | } | |
848 | ||
a649fd92 | 849 | /* |
e286781d NP |
850 | * Same as remove_mapping, but if the page is removed from the mapping, it |
851 | * gets returned with a refcount of 0. | |
a649fd92 | 852 | */ |
a528910e | 853 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
b910718a | 854 | bool reclaimed, struct mem_cgroup *target_memcg) |
49d2e9cc | 855 | { |
c4843a75 | 856 | unsigned long flags; |
bd4c82c2 | 857 | int refcount; |
aae466b0 | 858 | void *shadow = NULL; |
c4843a75 | 859 | |
28e4d965 NP |
860 | BUG_ON(!PageLocked(page)); |
861 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 862 | |
b93b0163 | 863 | xa_lock_irqsave(&mapping->i_pages, flags); |
49d2e9cc | 864 | /* |
0fd0e6b0 NP |
865 | * The non racy check for a busy page. |
866 | * | |
867 | * Must be careful with the order of the tests. When someone has | |
868 | * a ref to the page, it may be possible that they dirty it then | |
869 | * drop the reference. So if PageDirty is tested before page_count | |
870 | * here, then the following race may occur: | |
871 | * | |
872 | * get_user_pages(&page); | |
873 | * [user mapping goes away] | |
874 | * write_to(page); | |
875 | * !PageDirty(page) [good] | |
876 | * SetPageDirty(page); | |
877 | * put_page(page); | |
878 | * !page_count(page) [good, discard it] | |
879 | * | |
880 | * [oops, our write_to data is lost] | |
881 | * | |
882 | * Reversing the order of the tests ensures such a situation cannot | |
883 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 884 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
885 | * |
886 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
b93b0163 | 887 | * and thus under the i_pages lock, then this ordering is not required. |
49d2e9cc | 888 | */ |
906d278d | 889 | refcount = 1 + compound_nr(page); |
bd4c82c2 | 890 | if (!page_ref_freeze(page, refcount)) |
49d2e9cc | 891 | goto cannot_free; |
1c4c3b99 | 892 | /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */ |
e286781d | 893 | if (unlikely(PageDirty(page))) { |
bd4c82c2 | 894 | page_ref_unfreeze(page, refcount); |
49d2e9cc | 895 | goto cannot_free; |
e286781d | 896 | } |
49d2e9cc CL |
897 | |
898 | if (PageSwapCache(page)) { | |
899 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 900 | mem_cgroup_swapout(page, swap); |
aae466b0 JK |
901 | if (reclaimed && !mapping_exiting(mapping)) |
902 | shadow = workingset_eviction(page, target_memcg); | |
903 | __delete_from_swap_cache(page, swap, shadow); | |
b93b0163 | 904 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
75f6d6d2 | 905 | put_swap_page(page, swap); |
e286781d | 906 | } else { |
6072d13c LT |
907 | void (*freepage)(struct page *); |
908 | ||
909 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
910 | /* |
911 | * Remember a shadow entry for reclaimed file cache in | |
912 | * order to detect refaults, thus thrashing, later on. | |
913 | * | |
914 | * But don't store shadows in an address space that is | |
238c3046 | 915 | * already exiting. This is not just an optimization, |
a528910e JW |
916 | * inode reclaim needs to empty out the radix tree or |
917 | * the nodes are lost. Don't plant shadows behind its | |
918 | * back. | |
f9fe48be RZ |
919 | * |
920 | * We also don't store shadows for DAX mappings because the | |
921 | * only page cache pages found in these are zero pages | |
922 | * covering holes, and because we don't want to mix DAX | |
923 | * exceptional entries and shadow exceptional entries in the | |
b93b0163 | 924 | * same address_space. |
a528910e | 925 | */ |
9de4f22a | 926 | if (reclaimed && page_is_file_lru(page) && |
f9fe48be | 927 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
b910718a | 928 | shadow = workingset_eviction(page, target_memcg); |
62cccb8c | 929 | __delete_from_page_cache(page, shadow); |
b93b0163 | 930 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
6072d13c LT |
931 | |
932 | if (freepage != NULL) | |
933 | freepage(page); | |
49d2e9cc CL |
934 | } |
935 | ||
49d2e9cc CL |
936 | return 1; |
937 | ||
938 | cannot_free: | |
b93b0163 | 939 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
49d2e9cc CL |
940 | return 0; |
941 | } | |
942 | ||
e286781d NP |
943 | /* |
944 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
945 | * someone else has a ref on the page, abort and return 0. If it was | |
946 | * successfully detached, return 1. Assumes the caller has a single ref on | |
947 | * this page. | |
948 | */ | |
949 | int remove_mapping(struct address_space *mapping, struct page *page) | |
950 | { | |
b910718a | 951 | if (__remove_mapping(mapping, page, false, NULL)) { |
e286781d NP |
952 | /* |
953 | * Unfreezing the refcount with 1 rather than 2 effectively | |
954 | * drops the pagecache ref for us without requiring another | |
955 | * atomic operation. | |
956 | */ | |
fe896d18 | 957 | page_ref_unfreeze(page, 1); |
e286781d NP |
958 | return 1; |
959 | } | |
960 | return 0; | |
961 | } | |
962 | ||
894bc310 LS |
963 | /** |
964 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
965 | * @page: page to be put back to appropriate lru list | |
966 | * | |
967 | * Add previously isolated @page to appropriate LRU list. | |
968 | * Page may still be unevictable for other reasons. | |
969 | * | |
970 | * lru_lock must not be held, interrupts must be enabled. | |
971 | */ | |
894bc310 LS |
972 | void putback_lru_page(struct page *page) |
973 | { | |
9c4e6b1a | 974 | lru_cache_add(page); |
894bc310 LS |
975 | put_page(page); /* drop ref from isolate */ |
976 | } | |
977 | ||
dfc8d636 JW |
978 | enum page_references { |
979 | PAGEREF_RECLAIM, | |
980 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 981 | PAGEREF_KEEP, |
dfc8d636 JW |
982 | PAGEREF_ACTIVATE, |
983 | }; | |
984 | ||
985 | static enum page_references page_check_references(struct page *page, | |
986 | struct scan_control *sc) | |
987 | { | |
64574746 | 988 | int referenced_ptes, referenced_page; |
dfc8d636 | 989 | unsigned long vm_flags; |
dfc8d636 | 990 | |
c3ac9a8a JW |
991 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
992 | &vm_flags); | |
64574746 | 993 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 994 | |
dfc8d636 JW |
995 | /* |
996 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
997 | * move the page to the unevictable list. | |
998 | */ | |
999 | if (vm_flags & VM_LOCKED) | |
1000 | return PAGEREF_RECLAIM; | |
1001 | ||
64574746 | 1002 | if (referenced_ptes) { |
64574746 JW |
1003 | /* |
1004 | * All mapped pages start out with page table | |
1005 | * references from the instantiating fault, so we need | |
1006 | * to look twice if a mapped file page is used more | |
1007 | * than once. | |
1008 | * | |
1009 | * Mark it and spare it for another trip around the | |
1010 | * inactive list. Another page table reference will | |
1011 | * lead to its activation. | |
1012 | * | |
1013 | * Note: the mark is set for activated pages as well | |
1014 | * so that recently deactivated but used pages are | |
1015 | * quickly recovered. | |
1016 | */ | |
1017 | SetPageReferenced(page); | |
1018 | ||
34dbc67a | 1019 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
1020 | return PAGEREF_ACTIVATE; |
1021 | ||
c909e993 KK |
1022 | /* |
1023 | * Activate file-backed executable pages after first usage. | |
1024 | */ | |
b518154e | 1025 | if ((vm_flags & VM_EXEC) && !PageSwapBacked(page)) |
c909e993 KK |
1026 | return PAGEREF_ACTIVATE; |
1027 | ||
64574746 JW |
1028 | return PAGEREF_KEEP; |
1029 | } | |
dfc8d636 JW |
1030 | |
1031 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 1032 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
1033 | return PAGEREF_RECLAIM_CLEAN; |
1034 | ||
1035 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
1036 | } |
1037 | ||
e2be15f6 MG |
1038 | /* Check if a page is dirty or under writeback */ |
1039 | static void page_check_dirty_writeback(struct page *page, | |
1040 | bool *dirty, bool *writeback) | |
1041 | { | |
b4597226 MG |
1042 | struct address_space *mapping; |
1043 | ||
e2be15f6 MG |
1044 | /* |
1045 | * Anonymous pages are not handled by flushers and must be written | |
1046 | * from reclaim context. Do not stall reclaim based on them | |
1047 | */ | |
9de4f22a | 1048 | if (!page_is_file_lru(page) || |
802a3a92 | 1049 | (PageAnon(page) && !PageSwapBacked(page))) { |
e2be15f6 MG |
1050 | *dirty = false; |
1051 | *writeback = false; | |
1052 | return; | |
1053 | } | |
1054 | ||
1055 | /* By default assume that the page flags are accurate */ | |
1056 | *dirty = PageDirty(page); | |
1057 | *writeback = PageWriteback(page); | |
b4597226 MG |
1058 | |
1059 | /* Verify dirty/writeback state if the filesystem supports it */ | |
1060 | if (!page_has_private(page)) | |
1061 | return; | |
1062 | ||
1063 | mapping = page_mapping(page); | |
1064 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
1065 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
1066 | } |
1067 | ||
1da177e4 | 1068 | /* |
1742f19f | 1069 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 1070 | */ |
730ec8c0 MS |
1071 | static unsigned int shrink_page_list(struct list_head *page_list, |
1072 | struct pglist_data *pgdat, | |
1073 | struct scan_control *sc, | |
730ec8c0 MS |
1074 | struct reclaim_stat *stat, |
1075 | bool ignore_references) | |
1da177e4 LT |
1076 | { |
1077 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 1078 | LIST_HEAD(free_pages); |
730ec8c0 MS |
1079 | unsigned int nr_reclaimed = 0; |
1080 | unsigned int pgactivate = 0; | |
1da177e4 | 1081 | |
060f005f | 1082 | memset(stat, 0, sizeof(*stat)); |
1da177e4 LT |
1083 | cond_resched(); |
1084 | ||
1da177e4 LT |
1085 | while (!list_empty(page_list)) { |
1086 | struct address_space *mapping; | |
1087 | struct page *page; | |
8940b34a | 1088 | enum page_references references = PAGEREF_RECLAIM; |
4b793062 | 1089 | bool dirty, writeback, may_enter_fs; |
98879b3b | 1090 | unsigned int nr_pages; |
1da177e4 LT |
1091 | |
1092 | cond_resched(); | |
1093 | ||
1094 | page = lru_to_page(page_list); | |
1095 | list_del(&page->lru); | |
1096 | ||
529ae9aa | 1097 | if (!trylock_page(page)) |
1da177e4 LT |
1098 | goto keep; |
1099 | ||
309381fe | 1100 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 | 1101 | |
d8c6546b | 1102 | nr_pages = compound_nr(page); |
98879b3b YS |
1103 | |
1104 | /* Account the number of base pages even though THP */ | |
1105 | sc->nr_scanned += nr_pages; | |
80e43426 | 1106 | |
39b5f29a | 1107 | if (unlikely(!page_evictable(page))) |
ad6b6704 | 1108 | goto activate_locked; |
894bc310 | 1109 | |
a6dc60f8 | 1110 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
1111 | goto keep_locked; |
1112 | ||
c661b078 AW |
1113 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
1114 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
1115 | ||
e2be15f6 | 1116 | /* |
894befec | 1117 | * The number of dirty pages determines if a node is marked |
e2be15f6 MG |
1118 | * reclaim_congested which affects wait_iff_congested. kswapd |
1119 | * will stall and start writing pages if the tail of the LRU | |
1120 | * is all dirty unqueued pages. | |
1121 | */ | |
1122 | page_check_dirty_writeback(page, &dirty, &writeback); | |
1123 | if (dirty || writeback) | |
060f005f | 1124 | stat->nr_dirty++; |
e2be15f6 MG |
1125 | |
1126 | if (dirty && !writeback) | |
060f005f | 1127 | stat->nr_unqueued_dirty++; |
e2be15f6 | 1128 | |
d04e8acd MG |
1129 | /* |
1130 | * Treat this page as congested if the underlying BDI is or if | |
1131 | * pages are cycling through the LRU so quickly that the | |
1132 | * pages marked for immediate reclaim are making it to the | |
1133 | * end of the LRU a second time. | |
1134 | */ | |
e2be15f6 | 1135 | mapping = page_mapping(page); |
1da58ee2 | 1136 | if (((dirty || writeback) && mapping && |
703c2708 | 1137 | inode_write_congested(mapping->host)) || |
d04e8acd | 1138 | (writeback && PageReclaim(page))) |
060f005f | 1139 | stat->nr_congested++; |
e2be15f6 | 1140 | |
283aba9f MG |
1141 | /* |
1142 | * If a page at the tail of the LRU is under writeback, there | |
1143 | * are three cases to consider. | |
1144 | * | |
1145 | * 1) If reclaim is encountering an excessive number of pages | |
1146 | * under writeback and this page is both under writeback and | |
1147 | * PageReclaim then it indicates that pages are being queued | |
1148 | * for IO but are being recycled through the LRU before the | |
1149 | * IO can complete. Waiting on the page itself risks an | |
1150 | * indefinite stall if it is impossible to writeback the | |
1151 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
1152 | * note that the LRU is being scanned too quickly and the |
1153 | * caller can stall after page list has been processed. | |
283aba9f | 1154 | * |
97c9341f | 1155 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
1156 | * not marked for immediate reclaim, or the caller does not |
1157 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
1158 | * not to fs). In this case mark the page for immediate | |
97c9341f | 1159 | * reclaim and continue scanning. |
283aba9f | 1160 | * |
ecf5fc6e MH |
1161 | * Require may_enter_fs because we would wait on fs, which |
1162 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
1163 | * enter reclaim, and deadlock if it waits on a page for |
1164 | * which it is needed to do the write (loop masks off | |
1165 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1166 | * would probably show more reasons. | |
1167 | * | |
7fadc820 | 1168 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1169 | * PageReclaim. memcg does not have any dirty pages |
1170 | * throttling so we could easily OOM just because too many | |
1171 | * pages are in writeback and there is nothing else to | |
1172 | * reclaim. Wait for the writeback to complete. | |
c55e8d03 JW |
1173 | * |
1174 | * In cases 1) and 2) we activate the pages to get them out of | |
1175 | * the way while we continue scanning for clean pages on the | |
1176 | * inactive list and refilling from the active list. The | |
1177 | * observation here is that waiting for disk writes is more | |
1178 | * expensive than potentially causing reloads down the line. | |
1179 | * Since they're marked for immediate reclaim, they won't put | |
1180 | * memory pressure on the cache working set any longer than it | |
1181 | * takes to write them to disk. | |
283aba9f | 1182 | */ |
c661b078 | 1183 | if (PageWriteback(page)) { |
283aba9f MG |
1184 | /* Case 1 above */ |
1185 | if (current_is_kswapd() && | |
1186 | PageReclaim(page) && | |
599d0c95 | 1187 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
060f005f | 1188 | stat->nr_immediate++; |
c55e8d03 | 1189 | goto activate_locked; |
283aba9f MG |
1190 | |
1191 | /* Case 2 above */ | |
b5ead35e | 1192 | } else if (writeback_throttling_sane(sc) || |
ecf5fc6e | 1193 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1194 | /* |
1195 | * This is slightly racy - end_page_writeback() | |
1196 | * might have just cleared PageReclaim, then | |
1197 | * setting PageReclaim here end up interpreted | |
1198 | * as PageReadahead - but that does not matter | |
1199 | * enough to care. What we do want is for this | |
1200 | * page to have PageReclaim set next time memcg | |
1201 | * reclaim reaches the tests above, so it will | |
1202 | * then wait_on_page_writeback() to avoid OOM; | |
1203 | * and it's also appropriate in global reclaim. | |
1204 | */ | |
1205 | SetPageReclaim(page); | |
060f005f | 1206 | stat->nr_writeback++; |
c55e8d03 | 1207 | goto activate_locked; |
283aba9f MG |
1208 | |
1209 | /* Case 3 above */ | |
1210 | } else { | |
7fadc820 | 1211 | unlock_page(page); |
283aba9f | 1212 | wait_on_page_writeback(page); |
7fadc820 HD |
1213 | /* then go back and try same page again */ |
1214 | list_add_tail(&page->lru, page_list); | |
1215 | continue; | |
e62e384e | 1216 | } |
c661b078 | 1217 | } |
1da177e4 | 1218 | |
8940b34a | 1219 | if (!ignore_references) |
02c6de8d MK |
1220 | references = page_check_references(page, sc); |
1221 | ||
dfc8d636 JW |
1222 | switch (references) { |
1223 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1224 | goto activate_locked; |
64574746 | 1225 | case PAGEREF_KEEP: |
98879b3b | 1226 | stat->nr_ref_keep += nr_pages; |
64574746 | 1227 | goto keep_locked; |
dfc8d636 JW |
1228 | case PAGEREF_RECLAIM: |
1229 | case PAGEREF_RECLAIM_CLEAN: | |
1230 | ; /* try to reclaim the page below */ | |
1231 | } | |
1da177e4 | 1232 | |
1da177e4 LT |
1233 | /* |
1234 | * Anonymous process memory has backing store? | |
1235 | * Try to allocate it some swap space here. | |
802a3a92 | 1236 | * Lazyfree page could be freed directly |
1da177e4 | 1237 | */ |
bd4c82c2 HY |
1238 | if (PageAnon(page) && PageSwapBacked(page)) { |
1239 | if (!PageSwapCache(page)) { | |
1240 | if (!(sc->gfp_mask & __GFP_IO)) | |
1241 | goto keep_locked; | |
feb889fb LT |
1242 | if (page_maybe_dma_pinned(page)) |
1243 | goto keep_locked; | |
bd4c82c2 HY |
1244 | if (PageTransHuge(page)) { |
1245 | /* cannot split THP, skip it */ | |
1246 | if (!can_split_huge_page(page, NULL)) | |
1247 | goto activate_locked; | |
1248 | /* | |
1249 | * Split pages without a PMD map right | |
1250 | * away. Chances are some or all of the | |
1251 | * tail pages can be freed without IO. | |
1252 | */ | |
1253 | if (!compound_mapcount(page) && | |
1254 | split_huge_page_to_list(page, | |
1255 | page_list)) | |
1256 | goto activate_locked; | |
1257 | } | |
1258 | if (!add_to_swap(page)) { | |
1259 | if (!PageTransHuge(page)) | |
98879b3b | 1260 | goto activate_locked_split; |
bd4c82c2 HY |
1261 | /* Fallback to swap normal pages */ |
1262 | if (split_huge_page_to_list(page, | |
1263 | page_list)) | |
1264 | goto activate_locked; | |
fe490cc0 HY |
1265 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1266 | count_vm_event(THP_SWPOUT_FALLBACK); | |
1267 | #endif | |
bd4c82c2 | 1268 | if (!add_to_swap(page)) |
98879b3b | 1269 | goto activate_locked_split; |
bd4c82c2 | 1270 | } |
0f074658 | 1271 | |
4b793062 | 1272 | may_enter_fs = true; |
1da177e4 | 1273 | |
bd4c82c2 HY |
1274 | /* Adding to swap updated mapping */ |
1275 | mapping = page_mapping(page); | |
1276 | } | |
7751b2da KS |
1277 | } else if (unlikely(PageTransHuge(page))) { |
1278 | /* Split file THP */ | |
1279 | if (split_huge_page_to_list(page, page_list)) | |
1280 | goto keep_locked; | |
e2be15f6 | 1281 | } |
1da177e4 | 1282 | |
98879b3b YS |
1283 | /* |
1284 | * THP may get split above, need minus tail pages and update | |
1285 | * nr_pages to avoid accounting tail pages twice. | |
1286 | * | |
1287 | * The tail pages that are added into swap cache successfully | |
1288 | * reach here. | |
1289 | */ | |
1290 | if ((nr_pages > 1) && !PageTransHuge(page)) { | |
1291 | sc->nr_scanned -= (nr_pages - 1); | |
1292 | nr_pages = 1; | |
1293 | } | |
1294 | ||
1da177e4 LT |
1295 | /* |
1296 | * The page is mapped into the page tables of one or more | |
1297 | * processes. Try to unmap it here. | |
1298 | */ | |
802a3a92 | 1299 | if (page_mapped(page)) { |
013339df | 1300 | enum ttu_flags flags = TTU_BATCH_FLUSH; |
1f318a9b | 1301 | bool was_swapbacked = PageSwapBacked(page); |
bd4c82c2 HY |
1302 | |
1303 | if (unlikely(PageTransHuge(page))) | |
1304 | flags |= TTU_SPLIT_HUGE_PMD; | |
1f318a9b | 1305 | |
bd4c82c2 | 1306 | if (!try_to_unmap(page, flags)) { |
98879b3b | 1307 | stat->nr_unmap_fail += nr_pages; |
1f318a9b JK |
1308 | if (!was_swapbacked && PageSwapBacked(page)) |
1309 | stat->nr_lazyfree_fail += nr_pages; | |
1da177e4 | 1310 | goto activate_locked; |
1da177e4 LT |
1311 | } |
1312 | } | |
1313 | ||
1314 | if (PageDirty(page)) { | |
ee72886d | 1315 | /* |
4eda4823 JW |
1316 | * Only kswapd can writeback filesystem pages |
1317 | * to avoid risk of stack overflow. But avoid | |
1318 | * injecting inefficient single-page IO into | |
1319 | * flusher writeback as much as possible: only | |
1320 | * write pages when we've encountered many | |
1321 | * dirty pages, and when we've already scanned | |
1322 | * the rest of the LRU for clean pages and see | |
1323 | * the same dirty pages again (PageReclaim). | |
ee72886d | 1324 | */ |
9de4f22a | 1325 | if (page_is_file_lru(page) && |
4eda4823 JW |
1326 | (!current_is_kswapd() || !PageReclaim(page) || |
1327 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { | |
49ea7eb6 MG |
1328 | /* |
1329 | * Immediately reclaim when written back. | |
1330 | * Similar in principal to deactivate_page() | |
1331 | * except we already have the page isolated | |
1332 | * and know it's dirty | |
1333 | */ | |
c4a25635 | 1334 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
49ea7eb6 MG |
1335 | SetPageReclaim(page); |
1336 | ||
c55e8d03 | 1337 | goto activate_locked; |
ee72886d MG |
1338 | } |
1339 | ||
dfc8d636 | 1340 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1341 | goto keep_locked; |
4dd4b920 | 1342 | if (!may_enter_fs) |
1da177e4 | 1343 | goto keep_locked; |
52a8363e | 1344 | if (!sc->may_writepage) |
1da177e4 LT |
1345 | goto keep_locked; |
1346 | ||
d950c947 MG |
1347 | /* |
1348 | * Page is dirty. Flush the TLB if a writable entry | |
1349 | * potentially exists to avoid CPU writes after IO | |
1350 | * starts and then write it out here. | |
1351 | */ | |
1352 | try_to_unmap_flush_dirty(); | |
cb16556d | 1353 | switch (pageout(page, mapping)) { |
1da177e4 LT |
1354 | case PAGE_KEEP: |
1355 | goto keep_locked; | |
1356 | case PAGE_ACTIVATE: | |
1357 | goto activate_locked; | |
1358 | case PAGE_SUCCESS: | |
6c357848 | 1359 | stat->nr_pageout += thp_nr_pages(page); |
96f8bf4f | 1360 | |
7d3579e8 | 1361 | if (PageWriteback(page)) |
41ac1999 | 1362 | goto keep; |
7d3579e8 | 1363 | if (PageDirty(page)) |
1da177e4 | 1364 | goto keep; |
7d3579e8 | 1365 | |
1da177e4 LT |
1366 | /* |
1367 | * A synchronous write - probably a ramdisk. Go | |
1368 | * ahead and try to reclaim the page. | |
1369 | */ | |
529ae9aa | 1370 | if (!trylock_page(page)) |
1da177e4 LT |
1371 | goto keep; |
1372 | if (PageDirty(page) || PageWriteback(page)) | |
1373 | goto keep_locked; | |
1374 | mapping = page_mapping(page); | |
01359eb2 | 1375 | fallthrough; |
1da177e4 LT |
1376 | case PAGE_CLEAN: |
1377 | ; /* try to free the page below */ | |
1378 | } | |
1379 | } | |
1380 | ||
1381 | /* | |
1382 | * If the page has buffers, try to free the buffer mappings | |
1383 | * associated with this page. If we succeed we try to free | |
1384 | * the page as well. | |
1385 | * | |
1386 | * We do this even if the page is PageDirty(). | |
1387 | * try_to_release_page() does not perform I/O, but it is | |
1388 | * possible for a page to have PageDirty set, but it is actually | |
1389 | * clean (all its buffers are clean). This happens if the | |
1390 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1391 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1392 | * try_to_release_page() will discover that cleanness and will |
1393 | * drop the buffers and mark the page clean - it can be freed. | |
1394 | * | |
1395 | * Rarely, pages can have buffers and no ->mapping. These are | |
1396 | * the pages which were not successfully invalidated in | |
d12b8951 | 1397 | * truncate_cleanup_page(). We try to drop those buffers here |
1da177e4 LT |
1398 | * and if that worked, and the page is no longer mapped into |
1399 | * process address space (page_count == 1) it can be freed. | |
1400 | * Otherwise, leave the page on the LRU so it is swappable. | |
1401 | */ | |
266cf658 | 1402 | if (page_has_private(page)) { |
1da177e4 LT |
1403 | if (!try_to_release_page(page, sc->gfp_mask)) |
1404 | goto activate_locked; | |
e286781d NP |
1405 | if (!mapping && page_count(page) == 1) { |
1406 | unlock_page(page); | |
1407 | if (put_page_testzero(page)) | |
1408 | goto free_it; | |
1409 | else { | |
1410 | /* | |
1411 | * rare race with speculative reference. | |
1412 | * the speculative reference will free | |
1413 | * this page shortly, so we may | |
1414 | * increment nr_reclaimed here (and | |
1415 | * leave it off the LRU). | |
1416 | */ | |
1417 | nr_reclaimed++; | |
1418 | continue; | |
1419 | } | |
1420 | } | |
1da177e4 LT |
1421 | } |
1422 | ||
802a3a92 SL |
1423 | if (PageAnon(page) && !PageSwapBacked(page)) { |
1424 | /* follow __remove_mapping for reference */ | |
1425 | if (!page_ref_freeze(page, 1)) | |
1426 | goto keep_locked; | |
1427 | if (PageDirty(page)) { | |
1428 | page_ref_unfreeze(page, 1); | |
1429 | goto keep_locked; | |
1430 | } | |
1da177e4 | 1431 | |
802a3a92 | 1432 | count_vm_event(PGLAZYFREED); |
2262185c | 1433 | count_memcg_page_event(page, PGLAZYFREED); |
b910718a JW |
1434 | } else if (!mapping || !__remove_mapping(mapping, page, true, |
1435 | sc->target_mem_cgroup)) | |
802a3a92 | 1436 | goto keep_locked; |
9a1ea439 HD |
1437 | |
1438 | unlock_page(page); | |
e286781d | 1439 | free_it: |
98879b3b YS |
1440 | /* |
1441 | * THP may get swapped out in a whole, need account | |
1442 | * all base pages. | |
1443 | */ | |
1444 | nr_reclaimed += nr_pages; | |
abe4c3b5 MG |
1445 | |
1446 | /* | |
1447 | * Is there need to periodically free_page_list? It would | |
1448 | * appear not as the counts should be low | |
1449 | */ | |
7ae88534 | 1450 | if (unlikely(PageTransHuge(page))) |
ff45fc3c | 1451 | destroy_compound_page(page); |
7ae88534 | 1452 | else |
bd4c82c2 | 1453 | list_add(&page->lru, &free_pages); |
1da177e4 LT |
1454 | continue; |
1455 | ||
98879b3b YS |
1456 | activate_locked_split: |
1457 | /* | |
1458 | * The tail pages that are failed to add into swap cache | |
1459 | * reach here. Fixup nr_scanned and nr_pages. | |
1460 | */ | |
1461 | if (nr_pages > 1) { | |
1462 | sc->nr_scanned -= (nr_pages - 1); | |
1463 | nr_pages = 1; | |
1464 | } | |
1da177e4 | 1465 | activate_locked: |
68a22394 | 1466 | /* Not a candidate for swapping, so reclaim swap space. */ |
ad6b6704 MK |
1467 | if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || |
1468 | PageMlocked(page))) | |
a2c43eed | 1469 | try_to_free_swap(page); |
309381fe | 1470 | VM_BUG_ON_PAGE(PageActive(page), page); |
ad6b6704 | 1471 | if (!PageMlocked(page)) { |
9de4f22a | 1472 | int type = page_is_file_lru(page); |
ad6b6704 | 1473 | SetPageActive(page); |
98879b3b | 1474 | stat->nr_activate[type] += nr_pages; |
2262185c | 1475 | count_memcg_page_event(page, PGACTIVATE); |
ad6b6704 | 1476 | } |
1da177e4 LT |
1477 | keep_locked: |
1478 | unlock_page(page); | |
1479 | keep: | |
1480 | list_add(&page->lru, &ret_pages); | |
309381fe | 1481 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1482 | } |
abe4c3b5 | 1483 | |
98879b3b YS |
1484 | pgactivate = stat->nr_activate[0] + stat->nr_activate[1]; |
1485 | ||
747db954 | 1486 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1487 | try_to_unmap_flush(); |
2d4894b5 | 1488 | free_unref_page_list(&free_pages); |
abe4c3b5 | 1489 | |
1da177e4 | 1490 | list_splice(&ret_pages, page_list); |
886cf190 | 1491 | count_vm_events(PGACTIVATE, pgactivate); |
060f005f | 1492 | |
05ff5137 | 1493 | return nr_reclaimed; |
1da177e4 LT |
1494 | } |
1495 | ||
730ec8c0 | 1496 | unsigned int reclaim_clean_pages_from_list(struct zone *zone, |
02c6de8d MK |
1497 | struct list_head *page_list) |
1498 | { | |
1499 | struct scan_control sc = { | |
1500 | .gfp_mask = GFP_KERNEL, | |
1501 | .priority = DEF_PRIORITY, | |
1502 | .may_unmap = 1, | |
1503 | }; | |
1f318a9b | 1504 | struct reclaim_stat stat; |
730ec8c0 | 1505 | unsigned int nr_reclaimed; |
02c6de8d MK |
1506 | struct page *page, *next; |
1507 | LIST_HEAD(clean_pages); | |
1508 | ||
1509 | list_for_each_entry_safe(page, next, page_list, lru) { | |
ae37c7ff OS |
1510 | if (!PageHuge(page) && page_is_file_lru(page) && |
1511 | !PageDirty(page) && !__PageMovable(page) && | |
1512 | !PageUnevictable(page)) { | |
02c6de8d MK |
1513 | ClearPageActive(page); |
1514 | list_move(&page->lru, &clean_pages); | |
1515 | } | |
1516 | } | |
1517 | ||
1f318a9b | 1518 | nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
013339df | 1519 | &stat, true); |
02c6de8d | 1520 | list_splice(&clean_pages, page_list); |
2da9f630 NP |
1521 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, |
1522 | -(long)nr_reclaimed); | |
1f318a9b JK |
1523 | /* |
1524 | * Since lazyfree pages are isolated from file LRU from the beginning, | |
1525 | * they will rotate back to anonymous LRU in the end if it failed to | |
1526 | * discard so isolated count will be mismatched. | |
1527 | * Compensate the isolated count for both LRU lists. | |
1528 | */ | |
1529 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON, | |
1530 | stat.nr_lazyfree_fail); | |
1531 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, | |
2da9f630 | 1532 | -(long)stat.nr_lazyfree_fail); |
1f318a9b | 1533 | return nr_reclaimed; |
02c6de8d MK |
1534 | } |
1535 | ||
5ad333eb AW |
1536 | /* |
1537 | * Attempt to remove the specified page from its LRU. Only take this page | |
1538 | * if it is of the appropriate PageActive status. Pages which are being | |
1539 | * freed elsewhere are also ignored. | |
1540 | * | |
1541 | * page: page to consider | |
1542 | * mode: one of the LRU isolation modes defined above | |
1543 | * | |
c2135f7c | 1544 | * returns true on success, false on failure. |
5ad333eb | 1545 | */ |
c2135f7c | 1546 | bool __isolate_lru_page_prepare(struct page *page, isolate_mode_t mode) |
5ad333eb | 1547 | { |
5ad333eb AW |
1548 | /* Only take pages on the LRU. */ |
1549 | if (!PageLRU(page)) | |
c2135f7c | 1550 | return false; |
5ad333eb | 1551 | |
e46a2879 MK |
1552 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1553 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
c2135f7c | 1554 | return false; |
894bc310 | 1555 | |
c8244935 MG |
1556 | /* |
1557 | * To minimise LRU disruption, the caller can indicate that it only | |
1558 | * wants to isolate pages it will be able to operate on without | |
1559 | * blocking - clean pages for the most part. | |
1560 | * | |
c8244935 MG |
1561 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
1562 | * that it is possible to migrate without blocking | |
1563 | */ | |
1276ad68 | 1564 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
c8244935 MG |
1565 | /* All the caller can do on PageWriteback is block */ |
1566 | if (PageWriteback(page)) | |
c2135f7c | 1567 | return false; |
c8244935 MG |
1568 | |
1569 | if (PageDirty(page)) { | |
1570 | struct address_space *mapping; | |
69d763fc | 1571 | bool migrate_dirty; |
c8244935 | 1572 | |
c8244935 MG |
1573 | /* |
1574 | * Only pages without mappings or that have a | |
1575 | * ->migratepage callback are possible to migrate | |
69d763fc MG |
1576 | * without blocking. However, we can be racing with |
1577 | * truncation so it's necessary to lock the page | |
1578 | * to stabilise the mapping as truncation holds | |
1579 | * the page lock until after the page is removed | |
1580 | * from the page cache. | |
c8244935 | 1581 | */ |
69d763fc | 1582 | if (!trylock_page(page)) |
c2135f7c | 1583 | return false; |
69d763fc | 1584 | |
c8244935 | 1585 | mapping = page_mapping(page); |
145e1a71 | 1586 | migrate_dirty = !mapping || mapping->a_ops->migratepage; |
69d763fc MG |
1587 | unlock_page(page); |
1588 | if (!migrate_dirty) | |
c2135f7c | 1589 | return false; |
c8244935 MG |
1590 | } |
1591 | } | |
39deaf85 | 1592 | |
f80c0673 | 1593 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
c2135f7c | 1594 | return false; |
f80c0673 | 1595 | |
c2135f7c | 1596 | return true; |
5ad333eb AW |
1597 | } |
1598 | ||
7ee36a14 MG |
1599 | /* |
1600 | * Update LRU sizes after isolating pages. The LRU size updates must | |
55b65a57 | 1601 | * be complete before mem_cgroup_update_lru_size due to a sanity check. |
7ee36a14 MG |
1602 | */ |
1603 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 1604 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 1605 | { |
7ee36a14 MG |
1606 | int zid; |
1607 | ||
7ee36a14 MG |
1608 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1609 | if (!nr_zone_taken[zid]) | |
1610 | continue; | |
1611 | ||
a892cb6b | 1612 | update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
b4536f0c MH |
1613 | } |
1614 | ||
7ee36a14 MG |
1615 | } |
1616 | ||
f4b7e272 | 1617 | /** |
15b44736 HD |
1618 | * Isolating page from the lruvec to fill in @dst list by nr_to_scan times. |
1619 | * | |
1620 | * lruvec->lru_lock is heavily contended. Some of the functions that | |
1da177e4 LT |
1621 | * shrink the lists perform better by taking out a batch of pages |
1622 | * and working on them outside the LRU lock. | |
1623 | * | |
1624 | * For pagecache intensive workloads, this function is the hottest | |
1625 | * spot in the kernel (apart from copy_*_user functions). | |
1626 | * | |
15b44736 | 1627 | * Lru_lock must be held before calling this function. |
1da177e4 | 1628 | * |
791b48b6 | 1629 | * @nr_to_scan: The number of eligible pages to look through on the list. |
5dc35979 | 1630 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1631 | * @dst: The temp list to put pages on to. |
f626012d | 1632 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1633 | * @sc: The scan_control struct for this reclaim session |
3cb99451 | 1634 | * @lru: LRU list id for isolating |
1da177e4 LT |
1635 | * |
1636 | * returns how many pages were moved onto *@dst. | |
1637 | */ | |
69e05944 | 1638 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1639 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1640 | unsigned long *nr_scanned, struct scan_control *sc, |
a9e7c39f | 1641 | enum lru_list lru) |
1da177e4 | 1642 | { |
75b00af7 | 1643 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1644 | unsigned long nr_taken = 0; |
599d0c95 | 1645 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 1646 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
3db65812 | 1647 | unsigned long skipped = 0; |
791b48b6 | 1648 | unsigned long scan, total_scan, nr_pages; |
b2e18757 | 1649 | LIST_HEAD(pages_skipped); |
a9e7c39f | 1650 | isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED); |
1da177e4 | 1651 | |
98879b3b | 1652 | total_scan = 0; |
791b48b6 | 1653 | scan = 0; |
98879b3b | 1654 | while (scan < nr_to_scan && !list_empty(src)) { |
5ad333eb | 1655 | struct page *page; |
5ad333eb | 1656 | |
1da177e4 LT |
1657 | page = lru_to_page(src); |
1658 | prefetchw_prev_lru_page(page, src, flags); | |
1659 | ||
d8c6546b | 1660 | nr_pages = compound_nr(page); |
98879b3b YS |
1661 | total_scan += nr_pages; |
1662 | ||
b2e18757 MG |
1663 | if (page_zonenum(page) > sc->reclaim_idx) { |
1664 | list_move(&page->lru, &pages_skipped); | |
98879b3b | 1665 | nr_skipped[page_zonenum(page)] += nr_pages; |
b2e18757 MG |
1666 | continue; |
1667 | } | |
1668 | ||
791b48b6 MK |
1669 | /* |
1670 | * Do not count skipped pages because that makes the function | |
1671 | * return with no isolated pages if the LRU mostly contains | |
1672 | * ineligible pages. This causes the VM to not reclaim any | |
1673 | * pages, triggering a premature OOM. | |
98879b3b YS |
1674 | * |
1675 | * Account all tail pages of THP. This would not cause | |
1676 | * premature OOM since __isolate_lru_page() returns -EBUSY | |
1677 | * only when the page is being freed somewhere else. | |
791b48b6 | 1678 | */ |
98879b3b | 1679 | scan += nr_pages; |
c2135f7c AS |
1680 | if (!__isolate_lru_page_prepare(page, mode)) { |
1681 | /* It is being freed elsewhere */ | |
1682 | list_move(&page->lru, src); | |
1683 | continue; | |
1684 | } | |
1685 | /* | |
1686 | * Be careful not to clear PageLRU until after we're | |
1687 | * sure the page is not being freed elsewhere -- the | |
1688 | * page release code relies on it. | |
1689 | */ | |
1690 | if (unlikely(!get_page_unless_zero(page))) { | |
1691 | list_move(&page->lru, src); | |
1692 | continue; | |
1693 | } | |
5ad333eb | 1694 | |
c2135f7c AS |
1695 | if (!TestClearPageLRU(page)) { |
1696 | /* Another thread is already isolating this page */ | |
1697 | put_page(page); | |
5ad333eb | 1698 | list_move(&page->lru, src); |
c2135f7c | 1699 | continue; |
5ad333eb | 1700 | } |
c2135f7c AS |
1701 | |
1702 | nr_taken += nr_pages; | |
1703 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
1704 | list_move(&page->lru, dst); | |
1da177e4 LT |
1705 | } |
1706 | ||
b2e18757 MG |
1707 | /* |
1708 | * Splice any skipped pages to the start of the LRU list. Note that | |
1709 | * this disrupts the LRU order when reclaiming for lower zones but | |
1710 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
1711 | * scanning would soon rescan the same pages to skip and put the | |
1712 | * system at risk of premature OOM. | |
1713 | */ | |
7cc30fcf MG |
1714 | if (!list_empty(&pages_skipped)) { |
1715 | int zid; | |
1716 | ||
3db65812 | 1717 | list_splice(&pages_skipped, src); |
7cc30fcf MG |
1718 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1719 | if (!nr_skipped[zid]) | |
1720 | continue; | |
1721 | ||
1722 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 1723 | skipped += nr_skipped[zid]; |
7cc30fcf MG |
1724 | } |
1725 | } | |
791b48b6 | 1726 | *nr_scanned = total_scan; |
1265e3a6 | 1727 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
791b48b6 | 1728 | total_scan, skipped, nr_taken, mode, lru); |
b4536f0c | 1729 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
1730 | return nr_taken; |
1731 | } | |
1732 | ||
62695a84 NP |
1733 | /** |
1734 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1735 | * @page: page to isolate from its LRU list | |
1736 | * | |
1737 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1738 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1739 | * | |
1740 | * Returns 0 if the page was removed from an LRU list. | |
1741 | * Returns -EBUSY if the page was not on an LRU list. | |
1742 | * | |
1743 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1744 | * the active list, it will have PageActive set. If it was found on |
1745 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1746 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1747 | * |
1748 | * The vmstat statistic corresponding to the list on which the page was | |
1749 | * found will be decremented. | |
1750 | * | |
1751 | * Restrictions: | |
a5d09bed | 1752 | * |
62695a84 | 1753 | * (1) Must be called with an elevated refcount on the page. This is a |
01c4776b | 1754 | * fundamental difference from isolate_lru_pages (which is called |
62695a84 NP |
1755 | * without a stable reference). |
1756 | * (2) the lru_lock must not be held. | |
1757 | * (3) interrupts must be enabled. | |
1758 | */ | |
1759 | int isolate_lru_page(struct page *page) | |
1760 | { | |
1761 | int ret = -EBUSY; | |
1762 | ||
309381fe | 1763 | VM_BUG_ON_PAGE(!page_count(page), page); |
cf2a82ee | 1764 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
0c917313 | 1765 | |
d25b5bd8 | 1766 | if (TestClearPageLRU(page)) { |
fa9add64 | 1767 | struct lruvec *lruvec; |
62695a84 | 1768 | |
d25b5bd8 | 1769 | get_page(page); |
6168d0da | 1770 | lruvec = lock_page_lruvec_irq(page); |
46ae6b2c | 1771 | del_page_from_lru_list(page, lruvec); |
6168d0da | 1772 | unlock_page_lruvec_irq(lruvec); |
d25b5bd8 | 1773 | ret = 0; |
62695a84 | 1774 | } |
d25b5bd8 | 1775 | |
62695a84 NP |
1776 | return ret; |
1777 | } | |
1778 | ||
35cd7815 | 1779 | /* |
d37dd5dc | 1780 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
178821b8 | 1781 | * then get rescheduled. When there are massive number of tasks doing page |
d37dd5dc FW |
1782 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, |
1783 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1784 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 1785 | */ |
599d0c95 | 1786 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
1787 | struct scan_control *sc) |
1788 | { | |
1789 | unsigned long inactive, isolated; | |
1790 | ||
1791 | if (current_is_kswapd()) | |
1792 | return 0; | |
1793 | ||
b5ead35e | 1794 | if (!writeback_throttling_sane(sc)) |
35cd7815 RR |
1795 | return 0; |
1796 | ||
1797 | if (file) { | |
599d0c95 MG |
1798 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1799 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 1800 | } else { |
599d0c95 MG |
1801 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1802 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
1803 | } |
1804 | ||
3cf23841 FW |
1805 | /* |
1806 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1807 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1808 | * deadlock. | |
1809 | */ | |
d0164adc | 1810 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1811 | inactive >>= 3; |
1812 | ||
35cd7815 RR |
1813 | return isolated > inactive; |
1814 | } | |
1815 | ||
a222f341 | 1816 | /* |
15b44736 HD |
1817 | * move_pages_to_lru() moves pages from private @list to appropriate LRU list. |
1818 | * On return, @list is reused as a list of pages to be freed by the caller. | |
a222f341 KT |
1819 | * |
1820 | * Returns the number of pages moved to the given lruvec. | |
1821 | */ | |
a222f341 KT |
1822 | static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec, |
1823 | struct list_head *list) | |
66635629 | 1824 | { |
a222f341 | 1825 | int nr_pages, nr_moved = 0; |
3f79768f | 1826 | LIST_HEAD(pages_to_free); |
a222f341 | 1827 | struct page *page; |
66635629 | 1828 | |
a222f341 KT |
1829 | while (!list_empty(list)) { |
1830 | page = lru_to_page(list); | |
309381fe | 1831 | VM_BUG_ON_PAGE(PageLRU(page), page); |
3d06afab | 1832 | list_del(&page->lru); |
39b5f29a | 1833 | if (unlikely(!page_evictable(page))) { |
6168d0da | 1834 | spin_unlock_irq(&lruvec->lru_lock); |
66635629 | 1835 | putback_lru_page(page); |
6168d0da | 1836 | spin_lock_irq(&lruvec->lru_lock); |
66635629 MG |
1837 | continue; |
1838 | } | |
fa9add64 | 1839 | |
3d06afab AS |
1840 | /* |
1841 | * The SetPageLRU needs to be kept here for list integrity. | |
1842 | * Otherwise: | |
1843 | * #0 move_pages_to_lru #1 release_pages | |
1844 | * if !put_page_testzero | |
1845 | * if (put_page_testzero()) | |
1846 | * !PageLRU //skip lru_lock | |
1847 | * SetPageLRU() | |
1848 | * list_add(&page->lru,) | |
1849 | * list_add(&page->lru,) | |
1850 | */ | |
7a608572 | 1851 | SetPageLRU(page); |
a222f341 | 1852 | |
3d06afab | 1853 | if (unlikely(put_page_testzero(page))) { |
87560179 | 1854 | __clear_page_lru_flags(page); |
2bcf8879 HD |
1855 | |
1856 | if (unlikely(PageCompound(page))) { | |
6168d0da | 1857 | spin_unlock_irq(&lruvec->lru_lock); |
ff45fc3c | 1858 | destroy_compound_page(page); |
6168d0da | 1859 | spin_lock_irq(&lruvec->lru_lock); |
2bcf8879 HD |
1860 | } else |
1861 | list_add(&page->lru, &pages_to_free); | |
3d06afab AS |
1862 | |
1863 | continue; | |
66635629 | 1864 | } |
3d06afab | 1865 | |
afca9157 AS |
1866 | /* |
1867 | * All pages were isolated from the same lruvec (and isolation | |
1868 | * inhibits memcg migration). | |
1869 | */ | |
2a5e4e34 | 1870 | VM_BUG_ON_PAGE(!lruvec_holds_page_lru_lock(page, lruvec), page); |
3a9c9788 | 1871 | add_page_to_lru_list(page, lruvec); |
3d06afab | 1872 | nr_pages = thp_nr_pages(page); |
3d06afab AS |
1873 | nr_moved += nr_pages; |
1874 | if (PageActive(page)) | |
1875 | workingset_age_nonresident(lruvec, nr_pages); | |
66635629 | 1876 | } |
66635629 | 1877 | |
3f79768f HD |
1878 | /* |
1879 | * To save our caller's stack, now use input list for pages to free. | |
1880 | */ | |
a222f341 KT |
1881 | list_splice(&pages_to_free, list); |
1882 | ||
1883 | return nr_moved; | |
66635629 MG |
1884 | } |
1885 | ||
399ba0b9 N |
1886 | /* |
1887 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
a37b0715 | 1888 | * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE. |
399ba0b9 N |
1889 | * In that case we should only throttle if the backing device it is |
1890 | * writing to is congested. In other cases it is safe to throttle. | |
1891 | */ | |
1892 | static int current_may_throttle(void) | |
1893 | { | |
a37b0715 | 1894 | return !(current->flags & PF_LOCAL_THROTTLE) || |
399ba0b9 N |
1895 | current->backing_dev_info == NULL || |
1896 | bdi_write_congested(current->backing_dev_info); | |
1897 | } | |
1898 | ||
1da177e4 | 1899 | /* |
b2e18757 | 1900 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 1901 | * of reclaimed pages |
1da177e4 | 1902 | */ |
66635629 | 1903 | static noinline_for_stack unsigned long |
1a93be0e | 1904 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1905 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1906 | { |
1907 | LIST_HEAD(page_list); | |
e247dbce | 1908 | unsigned long nr_scanned; |
730ec8c0 | 1909 | unsigned int nr_reclaimed = 0; |
e247dbce | 1910 | unsigned long nr_taken; |
060f005f | 1911 | struct reclaim_stat stat; |
497a6c1b | 1912 | bool file = is_file_lru(lru); |
f46b7912 | 1913 | enum vm_event_item item; |
599d0c95 | 1914 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
db73ee0d | 1915 | bool stalled = false; |
78dc583d | 1916 | |
599d0c95 | 1917 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
db73ee0d MH |
1918 | if (stalled) |
1919 | return 0; | |
1920 | ||
1921 | /* wait a bit for the reclaimer. */ | |
1922 | msleep(100); | |
1923 | stalled = true; | |
35cd7815 RR |
1924 | |
1925 | /* We are about to die and free our memory. Return now. */ | |
1926 | if (fatal_signal_pending(current)) | |
1927 | return SWAP_CLUSTER_MAX; | |
1928 | } | |
1929 | ||
1da177e4 | 1930 | lru_add_drain(); |
f80c0673 | 1931 | |
6168d0da | 1932 | spin_lock_irq(&lruvec->lru_lock); |
b35ea17b | 1933 | |
5dc35979 | 1934 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
a9e7c39f | 1935 | &nr_scanned, sc, lru); |
95d918fc | 1936 | |
599d0c95 | 1937 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
f46b7912 | 1938 | item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT; |
b5ead35e | 1939 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
1940 | __count_vm_events(item, nr_scanned); |
1941 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned); | |
497a6c1b JW |
1942 | __count_vm_events(PGSCAN_ANON + file, nr_scanned); |
1943 | ||
6168d0da | 1944 | spin_unlock_irq(&lruvec->lru_lock); |
b35ea17b | 1945 | |
d563c050 | 1946 | if (nr_taken == 0) |
66635629 | 1947 | return 0; |
5ad333eb | 1948 | |
013339df | 1949 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false); |
c661b078 | 1950 | |
6168d0da | 1951 | spin_lock_irq(&lruvec->lru_lock); |
497a6c1b JW |
1952 | move_pages_to_lru(lruvec, &page_list); |
1953 | ||
1954 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); | |
f46b7912 | 1955 | item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT; |
b5ead35e | 1956 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
1957 | __count_vm_events(item, nr_reclaimed); |
1958 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed); | |
497a6c1b | 1959 | __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed); |
6168d0da | 1960 | spin_unlock_irq(&lruvec->lru_lock); |
3f79768f | 1961 | |
75cc3c91 | 1962 | lru_note_cost(lruvec, file, stat.nr_pageout); |
747db954 | 1963 | mem_cgroup_uncharge_list(&page_list); |
2d4894b5 | 1964 | free_unref_page_list(&page_list); |
e11da5b4 | 1965 | |
1c610d5f AR |
1966 | /* |
1967 | * If dirty pages are scanned that are not queued for IO, it | |
1968 | * implies that flushers are not doing their job. This can | |
1969 | * happen when memory pressure pushes dirty pages to the end of | |
1970 | * the LRU before the dirty limits are breached and the dirty | |
1971 | * data has expired. It can also happen when the proportion of | |
1972 | * dirty pages grows not through writes but through memory | |
1973 | * pressure reclaiming all the clean cache. And in some cases, | |
1974 | * the flushers simply cannot keep up with the allocation | |
1975 | * rate. Nudge the flusher threads in case they are asleep. | |
1976 | */ | |
1977 | if (stat.nr_unqueued_dirty == nr_taken) | |
1978 | wakeup_flusher_threads(WB_REASON_VMSCAN); | |
1979 | ||
d108c772 AR |
1980 | sc->nr.dirty += stat.nr_dirty; |
1981 | sc->nr.congested += stat.nr_congested; | |
1982 | sc->nr.unqueued_dirty += stat.nr_unqueued_dirty; | |
1983 | sc->nr.writeback += stat.nr_writeback; | |
1984 | sc->nr.immediate += stat.nr_immediate; | |
1985 | sc->nr.taken += nr_taken; | |
1986 | if (file) | |
1987 | sc->nr.file_taken += nr_taken; | |
8e950282 | 1988 | |
599d0c95 | 1989 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
d51d1e64 | 1990 | nr_scanned, nr_reclaimed, &stat, sc->priority, file); |
05ff5137 | 1991 | return nr_reclaimed; |
1da177e4 LT |
1992 | } |
1993 | ||
15b44736 HD |
1994 | /* |
1995 | * shrink_active_list() moves pages from the active LRU to the inactive LRU. | |
1996 | * | |
1997 | * We move them the other way if the page is referenced by one or more | |
1998 | * processes. | |
1999 | * | |
2000 | * If the pages are mostly unmapped, the processing is fast and it is | |
2001 | * appropriate to hold lru_lock across the whole operation. But if | |
2002 | * the pages are mapped, the processing is slow (page_referenced()), so | |
2003 | * we should drop lru_lock around each page. It's impossible to balance | |
2004 | * this, so instead we remove the pages from the LRU while processing them. | |
2005 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
2006 | * nobody will play with that bit on a non-LRU page. | |
2007 | * | |
2008 | * The downside is that we have to touch page->_refcount against each page. | |
2009 | * But we had to alter page->flags anyway. | |
2010 | */ | |
f626012d | 2011 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 2012 | struct lruvec *lruvec, |
f16015fb | 2013 | struct scan_control *sc, |
9e3b2f8c | 2014 | enum lru_list lru) |
1da177e4 | 2015 | { |
44c241f1 | 2016 | unsigned long nr_taken; |
f626012d | 2017 | unsigned long nr_scanned; |
6fe6b7e3 | 2018 | unsigned long vm_flags; |
1da177e4 | 2019 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 2020 | LIST_HEAD(l_active); |
b69408e8 | 2021 | LIST_HEAD(l_inactive); |
1da177e4 | 2022 | struct page *page; |
9d998b4f MH |
2023 | unsigned nr_deactivate, nr_activate; |
2024 | unsigned nr_rotated = 0; | |
3cb99451 | 2025 | int file = is_file_lru(lru); |
599d0c95 | 2026 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
2027 | |
2028 | lru_add_drain(); | |
f80c0673 | 2029 | |
6168d0da | 2030 | spin_lock_irq(&lruvec->lru_lock); |
925b7673 | 2031 | |
5dc35979 | 2032 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
a9e7c39f | 2033 | &nr_scanned, sc, lru); |
89b5fae5 | 2034 | |
599d0c95 | 2035 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
1cfb419b | 2036 | |
912c0572 SB |
2037 | if (!cgroup_reclaim(sc)) |
2038 | __count_vm_events(PGREFILL, nr_scanned); | |
2fa2690c | 2039 | __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); |
9d5e6a9f | 2040 | |
6168d0da | 2041 | spin_unlock_irq(&lruvec->lru_lock); |
1da177e4 | 2042 | |
1da177e4 LT |
2043 | while (!list_empty(&l_hold)) { |
2044 | cond_resched(); | |
2045 | page = lru_to_page(&l_hold); | |
2046 | list_del(&page->lru); | |
7e9cd484 | 2047 | |
39b5f29a | 2048 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
2049 | putback_lru_page(page); |
2050 | continue; | |
2051 | } | |
2052 | ||
cc715d99 MG |
2053 | if (unlikely(buffer_heads_over_limit)) { |
2054 | if (page_has_private(page) && trylock_page(page)) { | |
2055 | if (page_has_private(page)) | |
2056 | try_to_release_page(page, 0); | |
2057 | unlock_page(page); | |
2058 | } | |
2059 | } | |
2060 | ||
c3ac9a8a JW |
2061 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
2062 | &vm_flags)) { | |
8cab4754 WF |
2063 | /* |
2064 | * Identify referenced, file-backed active pages and | |
2065 | * give them one more trip around the active list. So | |
2066 | * that executable code get better chances to stay in | |
2067 | * memory under moderate memory pressure. Anon pages | |
2068 | * are not likely to be evicted by use-once streaming | |
2069 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
2070 | * so we ignore them here. | |
2071 | */ | |
9de4f22a | 2072 | if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) { |
6c357848 | 2073 | nr_rotated += thp_nr_pages(page); |
8cab4754 WF |
2074 | list_add(&page->lru, &l_active); |
2075 | continue; | |
2076 | } | |
2077 | } | |
7e9cd484 | 2078 | |
5205e56e | 2079 | ClearPageActive(page); /* we are de-activating */ |
1899ad18 | 2080 | SetPageWorkingset(page); |
1da177e4 LT |
2081 | list_add(&page->lru, &l_inactive); |
2082 | } | |
2083 | ||
b555749a | 2084 | /* |
8cab4754 | 2085 | * Move pages back to the lru list. |
b555749a | 2086 | */ |
6168d0da | 2087 | spin_lock_irq(&lruvec->lru_lock); |
556adecb | 2088 | |
a222f341 KT |
2089 | nr_activate = move_pages_to_lru(lruvec, &l_active); |
2090 | nr_deactivate = move_pages_to_lru(lruvec, &l_inactive); | |
f372d89e KT |
2091 | /* Keep all free pages in l_active list */ |
2092 | list_splice(&l_inactive, &l_active); | |
9851ac13 KT |
2093 | |
2094 | __count_vm_events(PGDEACTIVATE, nr_deactivate); | |
2095 | __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate); | |
2096 | ||
599d0c95 | 2097 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
6168d0da | 2098 | spin_unlock_irq(&lruvec->lru_lock); |
2bcf8879 | 2099 | |
f372d89e KT |
2100 | mem_cgroup_uncharge_list(&l_active); |
2101 | free_unref_page_list(&l_active); | |
9d998b4f MH |
2102 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
2103 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
2104 | } |
2105 | ||
1a4e58cc MK |
2106 | unsigned long reclaim_pages(struct list_head *page_list) |
2107 | { | |
f661d007 | 2108 | int nid = NUMA_NO_NODE; |
730ec8c0 | 2109 | unsigned int nr_reclaimed = 0; |
1a4e58cc MK |
2110 | LIST_HEAD(node_page_list); |
2111 | struct reclaim_stat dummy_stat; | |
2112 | struct page *page; | |
2113 | struct scan_control sc = { | |
2114 | .gfp_mask = GFP_KERNEL, | |
2115 | .priority = DEF_PRIORITY, | |
2116 | .may_writepage = 1, | |
2117 | .may_unmap = 1, | |
2118 | .may_swap = 1, | |
2119 | }; | |
2120 | ||
2121 | while (!list_empty(page_list)) { | |
2122 | page = lru_to_page(page_list); | |
f661d007 | 2123 | if (nid == NUMA_NO_NODE) { |
1a4e58cc MK |
2124 | nid = page_to_nid(page); |
2125 | INIT_LIST_HEAD(&node_page_list); | |
2126 | } | |
2127 | ||
2128 | if (nid == page_to_nid(page)) { | |
2129 | ClearPageActive(page); | |
2130 | list_move(&page->lru, &node_page_list); | |
2131 | continue; | |
2132 | } | |
2133 | ||
2134 | nr_reclaimed += shrink_page_list(&node_page_list, | |
2135 | NODE_DATA(nid), | |
013339df | 2136 | &sc, &dummy_stat, false); |
1a4e58cc MK |
2137 | while (!list_empty(&node_page_list)) { |
2138 | page = lru_to_page(&node_page_list); | |
2139 | list_del(&page->lru); | |
2140 | putback_lru_page(page); | |
2141 | } | |
2142 | ||
f661d007 | 2143 | nid = NUMA_NO_NODE; |
1a4e58cc MK |
2144 | } |
2145 | ||
2146 | if (!list_empty(&node_page_list)) { | |
2147 | nr_reclaimed += shrink_page_list(&node_page_list, | |
2148 | NODE_DATA(nid), | |
013339df | 2149 | &sc, &dummy_stat, false); |
1a4e58cc MK |
2150 | while (!list_empty(&node_page_list)) { |
2151 | page = lru_to_page(&node_page_list); | |
2152 | list_del(&page->lru); | |
2153 | putback_lru_page(page); | |
2154 | } | |
2155 | } | |
2156 | ||
2157 | return nr_reclaimed; | |
2158 | } | |
2159 | ||
b91ac374 JW |
2160 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
2161 | struct lruvec *lruvec, struct scan_control *sc) | |
2162 | { | |
2163 | if (is_active_lru(lru)) { | |
2164 | if (sc->may_deactivate & (1 << is_file_lru(lru))) | |
2165 | shrink_active_list(nr_to_scan, lruvec, sc, lru); | |
2166 | else | |
2167 | sc->skipped_deactivate = 1; | |
2168 | return 0; | |
2169 | } | |
2170 | ||
2171 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); | |
2172 | } | |
2173 | ||
59dc76b0 RR |
2174 | /* |
2175 | * The inactive anon list should be small enough that the VM never has | |
2176 | * to do too much work. | |
14797e23 | 2177 | * |
59dc76b0 RR |
2178 | * The inactive file list should be small enough to leave most memory |
2179 | * to the established workingset on the scan-resistant active list, | |
2180 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2181 | * |
59dc76b0 RR |
2182 | * Both inactive lists should also be large enough that each inactive |
2183 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 2184 | * |
2a2e4885 JW |
2185 | * If that fails and refaulting is observed, the inactive list grows. |
2186 | * | |
59dc76b0 | 2187 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
3a50d14d | 2188 | * on this LRU, maintained by the pageout code. An inactive_ratio |
59dc76b0 | 2189 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. |
56e49d21 | 2190 | * |
59dc76b0 RR |
2191 | * total target max |
2192 | * memory ratio inactive | |
2193 | * ------------------------------------- | |
2194 | * 10MB 1 5MB | |
2195 | * 100MB 1 50MB | |
2196 | * 1GB 3 250MB | |
2197 | * 10GB 10 0.9GB | |
2198 | * 100GB 31 3GB | |
2199 | * 1TB 101 10GB | |
2200 | * 10TB 320 32GB | |
56e49d21 | 2201 | */ |
b91ac374 | 2202 | static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru) |
56e49d21 | 2203 | { |
b91ac374 | 2204 | enum lru_list active_lru = inactive_lru + LRU_ACTIVE; |
2a2e4885 JW |
2205 | unsigned long inactive, active; |
2206 | unsigned long inactive_ratio; | |
59dc76b0 | 2207 | unsigned long gb; |
e3790144 | 2208 | |
b91ac374 JW |
2209 | inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru); |
2210 | active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru); | |
f8d1a311 | 2211 | |
b91ac374 | 2212 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
4002570c | 2213 | if (gb) |
b91ac374 JW |
2214 | inactive_ratio = int_sqrt(10 * gb); |
2215 | else | |
2216 | inactive_ratio = 1; | |
fd538803 | 2217 | |
59dc76b0 | 2218 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2219 | } |
2220 | ||
9a265114 JW |
2221 | enum scan_balance { |
2222 | SCAN_EQUAL, | |
2223 | SCAN_FRACT, | |
2224 | SCAN_ANON, | |
2225 | SCAN_FILE, | |
2226 | }; | |
2227 | ||
4f98a2fe RR |
2228 | /* |
2229 | * Determine how aggressively the anon and file LRU lists should be | |
2230 | * scanned. The relative value of each set of LRU lists is determined | |
2231 | * by looking at the fraction of the pages scanned we did rotate back | |
2232 | * onto the active list instead of evict. | |
2233 | * | |
be7bd59d WL |
2234 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2235 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 2236 | */ |
afaf07a6 JW |
2237 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, |
2238 | unsigned long *nr) | |
4f98a2fe | 2239 | { |
afaf07a6 | 2240 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); |
d483a5dd | 2241 | unsigned long anon_cost, file_cost, total_cost; |
33377678 | 2242 | int swappiness = mem_cgroup_swappiness(memcg); |
ed017373 | 2243 | u64 fraction[ANON_AND_FILE]; |
9a265114 | 2244 | u64 denominator = 0; /* gcc */ |
9a265114 | 2245 | enum scan_balance scan_balance; |
4f98a2fe | 2246 | unsigned long ap, fp; |
4111304d | 2247 | enum lru_list lru; |
76a33fc3 SL |
2248 | |
2249 | /* If we have no swap space, do not bother scanning anon pages. */ | |
d8b38438 | 2250 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
9a265114 | 2251 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2252 | goto out; |
2253 | } | |
4f98a2fe | 2254 | |
10316b31 JW |
2255 | /* |
2256 | * Global reclaim will swap to prevent OOM even with no | |
2257 | * swappiness, but memcg users want to use this knob to | |
2258 | * disable swapping for individual groups completely when | |
2259 | * using the memory controller's swap limit feature would be | |
2260 | * too expensive. | |
2261 | */ | |
b5ead35e | 2262 | if (cgroup_reclaim(sc) && !swappiness) { |
9a265114 | 2263 | scan_balance = SCAN_FILE; |
10316b31 JW |
2264 | goto out; |
2265 | } | |
2266 | ||
2267 | /* | |
2268 | * Do not apply any pressure balancing cleverness when the | |
2269 | * system is close to OOM, scan both anon and file equally | |
2270 | * (unless the swappiness setting disagrees with swapping). | |
2271 | */ | |
02695175 | 2272 | if (!sc->priority && swappiness) { |
9a265114 | 2273 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2274 | goto out; |
2275 | } | |
2276 | ||
62376251 | 2277 | /* |
53138cea | 2278 | * If the system is almost out of file pages, force-scan anon. |
62376251 | 2279 | */ |
b91ac374 | 2280 | if (sc->file_is_tiny) { |
53138cea JW |
2281 | scan_balance = SCAN_ANON; |
2282 | goto out; | |
62376251 JW |
2283 | } |
2284 | ||
7c5bd705 | 2285 | /* |
b91ac374 JW |
2286 | * If there is enough inactive page cache, we do not reclaim |
2287 | * anything from the anonymous working right now. | |
7c5bd705 | 2288 | */ |
b91ac374 | 2289 | if (sc->cache_trim_mode) { |
9a265114 | 2290 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2291 | goto out; |
2292 | } | |
2293 | ||
9a265114 | 2294 | scan_balance = SCAN_FRACT; |
58c37f6e | 2295 | /* |
314b57fb JW |
2296 | * Calculate the pressure balance between anon and file pages. |
2297 | * | |
2298 | * The amount of pressure we put on each LRU is inversely | |
2299 | * proportional to the cost of reclaiming each list, as | |
2300 | * determined by the share of pages that are refaulting, times | |
2301 | * the relative IO cost of bringing back a swapped out | |
2302 | * anonymous page vs reloading a filesystem page (swappiness). | |
2303 | * | |
d483a5dd JW |
2304 | * Although we limit that influence to ensure no list gets |
2305 | * left behind completely: at least a third of the pressure is | |
2306 | * applied, before swappiness. | |
2307 | * | |
314b57fb | 2308 | * With swappiness at 100, anon and file have equal IO cost. |
58c37f6e | 2309 | */ |
d483a5dd JW |
2310 | total_cost = sc->anon_cost + sc->file_cost; |
2311 | anon_cost = total_cost + sc->anon_cost; | |
2312 | file_cost = total_cost + sc->file_cost; | |
2313 | total_cost = anon_cost + file_cost; | |
58c37f6e | 2314 | |
d483a5dd JW |
2315 | ap = swappiness * (total_cost + 1); |
2316 | ap /= anon_cost + 1; | |
4f98a2fe | 2317 | |
d483a5dd JW |
2318 | fp = (200 - swappiness) * (total_cost + 1); |
2319 | fp /= file_cost + 1; | |
4f98a2fe | 2320 | |
76a33fc3 SL |
2321 | fraction[0] = ap; |
2322 | fraction[1] = fp; | |
a4fe1631 | 2323 | denominator = ap + fp; |
76a33fc3 | 2324 | out: |
688035f7 JW |
2325 | for_each_evictable_lru(lru) { |
2326 | int file = is_file_lru(lru); | |
9783aa99 | 2327 | unsigned long lruvec_size; |
688035f7 | 2328 | unsigned long scan; |
1bc63fb1 | 2329 | unsigned long protection; |
9783aa99 CD |
2330 | |
2331 | lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); | |
22f7496f YS |
2332 | protection = mem_cgroup_protection(sc->target_mem_cgroup, |
2333 | memcg, | |
1bc63fb1 | 2334 | sc->memcg_low_reclaim); |
9783aa99 | 2335 | |
1bc63fb1 | 2336 | if (protection) { |
9783aa99 CD |
2337 | /* |
2338 | * Scale a cgroup's reclaim pressure by proportioning | |
2339 | * its current usage to its memory.low or memory.min | |
2340 | * setting. | |
2341 | * | |
2342 | * This is important, as otherwise scanning aggression | |
2343 | * becomes extremely binary -- from nothing as we | |
2344 | * approach the memory protection threshold, to totally | |
2345 | * nominal as we exceed it. This results in requiring | |
2346 | * setting extremely liberal protection thresholds. It | |
2347 | * also means we simply get no protection at all if we | |
2348 | * set it too low, which is not ideal. | |
1bc63fb1 CD |
2349 | * |
2350 | * If there is any protection in place, we reduce scan | |
2351 | * pressure by how much of the total memory used is | |
2352 | * within protection thresholds. | |
9783aa99 | 2353 | * |
9de7ca46 CD |
2354 | * There is one special case: in the first reclaim pass, |
2355 | * we skip over all groups that are within their low | |
2356 | * protection. If that fails to reclaim enough pages to | |
2357 | * satisfy the reclaim goal, we come back and override | |
2358 | * the best-effort low protection. However, we still | |
2359 | * ideally want to honor how well-behaved groups are in | |
2360 | * that case instead of simply punishing them all | |
2361 | * equally. As such, we reclaim them based on how much | |
1bc63fb1 CD |
2362 | * memory they are using, reducing the scan pressure |
2363 | * again by how much of the total memory used is under | |
2364 | * hard protection. | |
9783aa99 | 2365 | */ |
1bc63fb1 CD |
2366 | unsigned long cgroup_size = mem_cgroup_size(memcg); |
2367 | ||
2368 | /* Avoid TOCTOU with earlier protection check */ | |
2369 | cgroup_size = max(cgroup_size, protection); | |
2370 | ||
2371 | scan = lruvec_size - lruvec_size * protection / | |
2372 | cgroup_size; | |
9783aa99 CD |
2373 | |
2374 | /* | |
1bc63fb1 | 2375 | * Minimally target SWAP_CLUSTER_MAX pages to keep |
55b65a57 | 2376 | * reclaim moving forwards, avoiding decrementing |
9de7ca46 | 2377 | * sc->priority further than desirable. |
9783aa99 | 2378 | */ |
1bc63fb1 | 2379 | scan = max(scan, SWAP_CLUSTER_MAX); |
9783aa99 CD |
2380 | } else { |
2381 | scan = lruvec_size; | |
2382 | } | |
2383 | ||
2384 | scan >>= sc->priority; | |
6b4f7799 | 2385 | |
688035f7 JW |
2386 | /* |
2387 | * If the cgroup's already been deleted, make sure to | |
2388 | * scrape out the remaining cache. | |
2389 | */ | |
2390 | if (!scan && !mem_cgroup_online(memcg)) | |
9783aa99 | 2391 | scan = min(lruvec_size, SWAP_CLUSTER_MAX); |
6b4f7799 | 2392 | |
688035f7 JW |
2393 | switch (scan_balance) { |
2394 | case SCAN_EQUAL: | |
2395 | /* Scan lists relative to size */ | |
2396 | break; | |
2397 | case SCAN_FRACT: | |
9a265114 | 2398 | /* |
688035f7 JW |
2399 | * Scan types proportional to swappiness and |
2400 | * their relative recent reclaim efficiency. | |
76073c64 GS |
2401 | * Make sure we don't miss the last page on |
2402 | * the offlined memory cgroups because of a | |
2403 | * round-off error. | |
9a265114 | 2404 | */ |
76073c64 GS |
2405 | scan = mem_cgroup_online(memcg) ? |
2406 | div64_u64(scan * fraction[file], denominator) : | |
2407 | DIV64_U64_ROUND_UP(scan * fraction[file], | |
68600f62 | 2408 | denominator); |
688035f7 JW |
2409 | break; |
2410 | case SCAN_FILE: | |
2411 | case SCAN_ANON: | |
2412 | /* Scan one type exclusively */ | |
e072bff6 | 2413 | if ((scan_balance == SCAN_FILE) != file) |
688035f7 | 2414 | scan = 0; |
688035f7 JW |
2415 | break; |
2416 | default: | |
2417 | /* Look ma, no brain */ | |
2418 | BUG(); | |
9a265114 | 2419 | } |
688035f7 | 2420 | |
688035f7 | 2421 | nr[lru] = scan; |
76a33fc3 | 2422 | } |
6e08a369 | 2423 | } |
4f98a2fe | 2424 | |
afaf07a6 | 2425 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) |
9b4f98cd JW |
2426 | { |
2427 | unsigned long nr[NR_LRU_LISTS]; | |
e82e0561 | 2428 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2429 | unsigned long nr_to_scan; |
2430 | enum lru_list lru; | |
2431 | unsigned long nr_reclaimed = 0; | |
2432 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2433 | struct blk_plug plug; | |
1a501907 | 2434 | bool scan_adjusted; |
9b4f98cd | 2435 | |
afaf07a6 | 2436 | get_scan_count(lruvec, sc, nr); |
9b4f98cd | 2437 | |
e82e0561 MG |
2438 | /* Record the original scan target for proportional adjustments later */ |
2439 | memcpy(targets, nr, sizeof(nr)); | |
2440 | ||
1a501907 MG |
2441 | /* |
2442 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2443 | * event that can occur when there is little memory pressure e.g. | |
2444 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2445 | * when the requested number of pages are reclaimed when scanning at | |
2446 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2447 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2448 | * do a batch of work at once. For memcg reclaim one check is made to | |
2449 | * abort proportional reclaim if either the file or anon lru has already | |
2450 | * dropped to zero at the first pass. | |
2451 | */ | |
b5ead35e | 2452 | scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() && |
1a501907 MG |
2453 | sc->priority == DEF_PRIORITY); |
2454 | ||
9b4f98cd JW |
2455 | blk_start_plug(&plug); |
2456 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2457 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2458 | unsigned long nr_anon, nr_file, percentage; |
2459 | unsigned long nr_scanned; | |
2460 | ||
9b4f98cd JW |
2461 | for_each_evictable_lru(lru) { |
2462 | if (nr[lru]) { | |
2463 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2464 | nr[lru] -= nr_to_scan; | |
2465 | ||
2466 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
3b991208 | 2467 | lruvec, sc); |
9b4f98cd JW |
2468 | } |
2469 | } | |
e82e0561 | 2470 | |
bd041733 MH |
2471 | cond_resched(); |
2472 | ||
e82e0561 MG |
2473 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
2474 | continue; | |
2475 | ||
e82e0561 MG |
2476 | /* |
2477 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2478 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2479 | * proportionally what was requested by get_scan_count(). We |
2480 | * stop reclaiming one LRU and reduce the amount scanning | |
2481 | * proportional to the original scan target. | |
2482 | */ | |
2483 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2484 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2485 | ||
1a501907 MG |
2486 | /* |
2487 | * It's just vindictive to attack the larger once the smaller | |
2488 | * has gone to zero. And given the way we stop scanning the | |
2489 | * smaller below, this makes sure that we only make one nudge | |
2490 | * towards proportionality once we've got nr_to_reclaim. | |
2491 | */ | |
2492 | if (!nr_file || !nr_anon) | |
2493 | break; | |
2494 | ||
e82e0561 MG |
2495 | if (nr_file > nr_anon) { |
2496 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2497 | targets[LRU_ACTIVE_ANON] + 1; | |
2498 | lru = LRU_BASE; | |
2499 | percentage = nr_anon * 100 / scan_target; | |
2500 | } else { | |
2501 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2502 | targets[LRU_ACTIVE_FILE] + 1; | |
2503 | lru = LRU_FILE; | |
2504 | percentage = nr_file * 100 / scan_target; | |
2505 | } | |
2506 | ||
2507 | /* Stop scanning the smaller of the LRU */ | |
2508 | nr[lru] = 0; | |
2509 | nr[lru + LRU_ACTIVE] = 0; | |
2510 | ||
2511 | /* | |
2512 | * Recalculate the other LRU scan count based on its original | |
2513 | * scan target and the percentage scanning already complete | |
2514 | */ | |
2515 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2516 | nr_scanned = targets[lru] - nr[lru]; | |
2517 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2518 | nr[lru] -= min(nr[lru], nr_scanned); | |
2519 | ||
2520 | lru += LRU_ACTIVE; | |
2521 | nr_scanned = targets[lru] - nr[lru]; | |
2522 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2523 | nr[lru] -= min(nr[lru], nr_scanned); | |
2524 | ||
2525 | scan_adjusted = true; | |
9b4f98cd JW |
2526 | } |
2527 | blk_finish_plug(&plug); | |
2528 | sc->nr_reclaimed += nr_reclaimed; | |
2529 | ||
2530 | /* | |
2531 | * Even if we did not try to evict anon pages at all, we want to | |
2532 | * rebalance the anon lru active/inactive ratio. | |
2533 | */ | |
b91ac374 | 2534 | if (total_swap_pages && inactive_is_low(lruvec, LRU_INACTIVE_ANON)) |
9b4f98cd JW |
2535 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2536 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
2537 | } |
2538 | ||
23b9da55 | 2539 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2540 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2541 | { |
d84da3f9 | 2542 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2543 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2544 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2545 | return true; |
2546 | ||
2547 | return false; | |
2548 | } | |
2549 | ||
3e7d3449 | 2550 | /* |
23b9da55 MG |
2551 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2552 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2553 | * true if more pages should be reclaimed such that when the page allocator | |
df3a45f9 | 2554 | * calls try_to_compact_pages() that it will have enough free pages to succeed. |
23b9da55 | 2555 | * It will give up earlier than that if there is difficulty reclaiming pages. |
3e7d3449 | 2556 | */ |
a9dd0a83 | 2557 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 | 2558 | unsigned long nr_reclaimed, |
3e7d3449 MG |
2559 | struct scan_control *sc) |
2560 | { | |
2561 | unsigned long pages_for_compaction; | |
2562 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 2563 | int z; |
3e7d3449 MG |
2564 | |
2565 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2566 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2567 | return false; |
2568 | ||
5ee04716 VB |
2569 | /* |
2570 | * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX | |
2571 | * number of pages that were scanned. This will return to the caller | |
2572 | * with the risk reclaim/compaction and the resulting allocation attempt | |
2573 | * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL | |
2574 | * allocations through requiring that the full LRU list has been scanned | |
2575 | * first, by assuming that zero delta of sc->nr_scanned means full LRU | |
2576 | * scan, but that approximation was wrong, and there were corner cases | |
2577 | * where always a non-zero amount of pages were scanned. | |
2578 | */ | |
2579 | if (!nr_reclaimed) | |
2580 | return false; | |
3e7d3449 | 2581 | |
3e7d3449 | 2582 | /* If compaction would go ahead or the allocation would succeed, stop */ |
a9dd0a83 MG |
2583 | for (z = 0; z <= sc->reclaim_idx; z++) { |
2584 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2585 | if (!managed_zone(zone)) |
a9dd0a83 MG |
2586 | continue; |
2587 | ||
2588 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 2589 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
2590 | case COMPACT_CONTINUE: |
2591 | return false; | |
2592 | default: | |
2593 | /* check next zone */ | |
2594 | ; | |
2595 | } | |
3e7d3449 | 2596 | } |
1c6c1597 HD |
2597 | |
2598 | /* | |
2599 | * If we have not reclaimed enough pages for compaction and the | |
2600 | * inactive lists are large enough, continue reclaiming | |
2601 | */ | |
2602 | pages_for_compaction = compact_gap(sc->order); | |
2603 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); | |
2604 | if (get_nr_swap_pages() > 0) | |
2605 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); | |
2606 | ||
5ee04716 | 2607 | return inactive_lru_pages > pages_for_compaction; |
3e7d3449 MG |
2608 | } |
2609 | ||
0f6a5cff | 2610 | static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 2611 | { |
0f6a5cff | 2612 | struct mem_cgroup *target_memcg = sc->target_mem_cgroup; |
d2af3397 | 2613 | struct mem_cgroup *memcg; |
1da177e4 | 2614 | |
0f6a5cff | 2615 | memcg = mem_cgroup_iter(target_memcg, NULL, NULL); |
d2af3397 | 2616 | do { |
afaf07a6 | 2617 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
d2af3397 JW |
2618 | unsigned long reclaimed; |
2619 | unsigned long scanned; | |
5660048c | 2620 | |
e3336cab XP |
2621 | /* |
2622 | * This loop can become CPU-bound when target memcgs | |
2623 | * aren't eligible for reclaim - either because they | |
2624 | * don't have any reclaimable pages, or because their | |
2625 | * memory is explicitly protected. Avoid soft lockups. | |
2626 | */ | |
2627 | cond_resched(); | |
2628 | ||
45c7f7e1 CD |
2629 | mem_cgroup_calculate_protection(target_memcg, memcg); |
2630 | ||
2631 | if (mem_cgroup_below_min(memcg)) { | |
d2af3397 JW |
2632 | /* |
2633 | * Hard protection. | |
2634 | * If there is no reclaimable memory, OOM. | |
2635 | */ | |
2636 | continue; | |
45c7f7e1 | 2637 | } else if (mem_cgroup_below_low(memcg)) { |
d2af3397 JW |
2638 | /* |
2639 | * Soft protection. | |
2640 | * Respect the protection only as long as | |
2641 | * there is an unprotected supply | |
2642 | * of reclaimable memory from other cgroups. | |
2643 | */ | |
2644 | if (!sc->memcg_low_reclaim) { | |
2645 | sc->memcg_low_skipped = 1; | |
bf8d5d52 | 2646 | continue; |
241994ed | 2647 | } |
d2af3397 | 2648 | memcg_memory_event(memcg, MEMCG_LOW); |
d2af3397 | 2649 | } |
241994ed | 2650 | |
d2af3397 JW |
2651 | reclaimed = sc->nr_reclaimed; |
2652 | scanned = sc->nr_scanned; | |
afaf07a6 JW |
2653 | |
2654 | shrink_lruvec(lruvec, sc); | |
70ddf637 | 2655 | |
d2af3397 JW |
2656 | shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, |
2657 | sc->priority); | |
6b4f7799 | 2658 | |
d2af3397 JW |
2659 | /* Record the group's reclaim efficiency */ |
2660 | vmpressure(sc->gfp_mask, memcg, false, | |
2661 | sc->nr_scanned - scanned, | |
2662 | sc->nr_reclaimed - reclaimed); | |
70ddf637 | 2663 | |
0f6a5cff JW |
2664 | } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL))); |
2665 | } | |
2666 | ||
6c9e0907 | 2667 | static void shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
0f6a5cff JW |
2668 | { |
2669 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
0f6a5cff | 2670 | unsigned long nr_reclaimed, nr_scanned; |
1b05117d | 2671 | struct lruvec *target_lruvec; |
0f6a5cff | 2672 | bool reclaimable = false; |
b91ac374 | 2673 | unsigned long file; |
0f6a5cff | 2674 | |
1b05117d JW |
2675 | target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat); |
2676 | ||
0f6a5cff JW |
2677 | again: |
2678 | memset(&sc->nr, 0, sizeof(sc->nr)); | |
2679 | ||
2680 | nr_reclaimed = sc->nr_reclaimed; | |
2681 | nr_scanned = sc->nr_scanned; | |
2682 | ||
7cf111bc JW |
2683 | /* |
2684 | * Determine the scan balance between anon and file LRUs. | |
2685 | */ | |
6168d0da | 2686 | spin_lock_irq(&target_lruvec->lru_lock); |
7cf111bc JW |
2687 | sc->anon_cost = target_lruvec->anon_cost; |
2688 | sc->file_cost = target_lruvec->file_cost; | |
6168d0da | 2689 | spin_unlock_irq(&target_lruvec->lru_lock); |
7cf111bc | 2690 | |
b91ac374 JW |
2691 | /* |
2692 | * Target desirable inactive:active list ratios for the anon | |
2693 | * and file LRU lists. | |
2694 | */ | |
2695 | if (!sc->force_deactivate) { | |
2696 | unsigned long refaults; | |
2697 | ||
170b04b7 JK |
2698 | refaults = lruvec_page_state(target_lruvec, |
2699 | WORKINGSET_ACTIVATE_ANON); | |
2700 | if (refaults != target_lruvec->refaults[0] || | |
2701 | inactive_is_low(target_lruvec, LRU_INACTIVE_ANON)) | |
b91ac374 JW |
2702 | sc->may_deactivate |= DEACTIVATE_ANON; |
2703 | else | |
2704 | sc->may_deactivate &= ~DEACTIVATE_ANON; | |
2705 | ||
2706 | /* | |
2707 | * When refaults are being observed, it means a new | |
2708 | * workingset is being established. Deactivate to get | |
2709 | * rid of any stale active pages quickly. | |
2710 | */ | |
2711 | refaults = lruvec_page_state(target_lruvec, | |
170b04b7 JK |
2712 | WORKINGSET_ACTIVATE_FILE); |
2713 | if (refaults != target_lruvec->refaults[1] || | |
b91ac374 JW |
2714 | inactive_is_low(target_lruvec, LRU_INACTIVE_FILE)) |
2715 | sc->may_deactivate |= DEACTIVATE_FILE; | |
2716 | else | |
2717 | sc->may_deactivate &= ~DEACTIVATE_FILE; | |
2718 | } else | |
2719 | sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE; | |
2720 | ||
2721 | /* | |
2722 | * If we have plenty of inactive file pages that aren't | |
2723 | * thrashing, try to reclaim those first before touching | |
2724 | * anonymous pages. | |
2725 | */ | |
2726 | file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE); | |
2727 | if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE)) | |
2728 | sc->cache_trim_mode = 1; | |
2729 | else | |
2730 | sc->cache_trim_mode = 0; | |
2731 | ||
53138cea JW |
2732 | /* |
2733 | * Prevent the reclaimer from falling into the cache trap: as | |
2734 | * cache pages start out inactive, every cache fault will tip | |
2735 | * the scan balance towards the file LRU. And as the file LRU | |
2736 | * shrinks, so does the window for rotation from references. | |
2737 | * This means we have a runaway feedback loop where a tiny | |
2738 | * thrashing file LRU becomes infinitely more attractive than | |
2739 | * anon pages. Try to detect this based on file LRU size. | |
2740 | */ | |
2741 | if (!cgroup_reclaim(sc)) { | |
53138cea | 2742 | unsigned long total_high_wmark = 0; |
b91ac374 JW |
2743 | unsigned long free, anon; |
2744 | int z; | |
53138cea JW |
2745 | |
2746 | free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); | |
2747 | file = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2748 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2749 | ||
2750 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2751 | struct zone *zone = &pgdat->node_zones[z]; | |
2752 | if (!managed_zone(zone)) | |
2753 | continue; | |
2754 | ||
2755 | total_high_wmark += high_wmark_pages(zone); | |
2756 | } | |
2757 | ||
b91ac374 JW |
2758 | /* |
2759 | * Consider anon: if that's low too, this isn't a | |
2760 | * runaway file reclaim problem, but rather just | |
2761 | * extreme pressure. Reclaim as per usual then. | |
2762 | */ | |
2763 | anon = node_page_state(pgdat, NR_INACTIVE_ANON); | |
2764 | ||
2765 | sc->file_is_tiny = | |
2766 | file + free <= total_high_wmark && | |
2767 | !(sc->may_deactivate & DEACTIVATE_ANON) && | |
2768 | anon >> sc->priority; | |
53138cea JW |
2769 | } |
2770 | ||
0f6a5cff | 2771 | shrink_node_memcgs(pgdat, sc); |
2344d7e4 | 2772 | |
d2af3397 JW |
2773 | if (reclaim_state) { |
2774 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2775 | reclaim_state->reclaimed_slab = 0; | |
2776 | } | |
d108c772 | 2777 | |
d2af3397 | 2778 | /* Record the subtree's reclaim efficiency */ |
1b05117d | 2779 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, |
d2af3397 JW |
2780 | sc->nr_scanned - nr_scanned, |
2781 | sc->nr_reclaimed - nr_reclaimed); | |
d108c772 | 2782 | |
d2af3397 JW |
2783 | if (sc->nr_reclaimed - nr_reclaimed) |
2784 | reclaimable = true; | |
d108c772 | 2785 | |
d2af3397 JW |
2786 | if (current_is_kswapd()) { |
2787 | /* | |
2788 | * If reclaim is isolating dirty pages under writeback, | |
2789 | * it implies that the long-lived page allocation rate | |
2790 | * is exceeding the page laundering rate. Either the | |
2791 | * global limits are not being effective at throttling | |
2792 | * processes due to the page distribution throughout | |
2793 | * zones or there is heavy usage of a slow backing | |
2794 | * device. The only option is to throttle from reclaim | |
2795 | * context which is not ideal as there is no guarantee | |
2796 | * the dirtying process is throttled in the same way | |
2797 | * balance_dirty_pages() manages. | |
2798 | * | |
2799 | * Once a node is flagged PGDAT_WRITEBACK, kswapd will | |
2800 | * count the number of pages under pages flagged for | |
2801 | * immediate reclaim and stall if any are encountered | |
2802 | * in the nr_immediate check below. | |
2803 | */ | |
2804 | if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken) | |
2805 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
d108c772 | 2806 | |
d2af3397 JW |
2807 | /* Allow kswapd to start writing pages during reclaim.*/ |
2808 | if (sc->nr.unqueued_dirty == sc->nr.file_taken) | |
2809 | set_bit(PGDAT_DIRTY, &pgdat->flags); | |
e3c1ac58 | 2810 | |
d108c772 | 2811 | /* |
1eba09c1 | 2812 | * If kswapd scans pages marked for immediate |
d2af3397 JW |
2813 | * reclaim and under writeback (nr_immediate), it |
2814 | * implies that pages are cycling through the LRU | |
2815 | * faster than they are written so also forcibly stall. | |
d108c772 | 2816 | */ |
d2af3397 JW |
2817 | if (sc->nr.immediate) |
2818 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
2819 | } | |
2820 | ||
2821 | /* | |
1b05117d JW |
2822 | * Tag a node/memcg as congested if all the dirty pages |
2823 | * scanned were backed by a congested BDI and | |
2824 | * wait_iff_congested will stall. | |
2825 | * | |
d2af3397 JW |
2826 | * Legacy memcg will stall in page writeback so avoid forcibly |
2827 | * stalling in wait_iff_congested(). | |
2828 | */ | |
1b05117d JW |
2829 | if ((current_is_kswapd() || |
2830 | (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) && | |
d2af3397 | 2831 | sc->nr.dirty && sc->nr.dirty == sc->nr.congested) |
1b05117d | 2832 | set_bit(LRUVEC_CONGESTED, &target_lruvec->flags); |
d2af3397 JW |
2833 | |
2834 | /* | |
2835 | * Stall direct reclaim for IO completions if underlying BDIs | |
2836 | * and node is congested. Allow kswapd to continue until it | |
2837 | * starts encountering unqueued dirty pages or cycling through | |
2838 | * the LRU too quickly. | |
2839 | */ | |
1b05117d JW |
2840 | if (!current_is_kswapd() && current_may_throttle() && |
2841 | !sc->hibernation_mode && | |
2842 | test_bit(LRUVEC_CONGESTED, &target_lruvec->flags)) | |
d2af3397 | 2843 | wait_iff_congested(BLK_RW_ASYNC, HZ/10); |
d108c772 | 2844 | |
d2af3397 JW |
2845 | if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
2846 | sc)) | |
2847 | goto again; | |
2344d7e4 | 2848 | |
c73322d0 JW |
2849 | /* |
2850 | * Kswapd gives up on balancing particular nodes after too | |
2851 | * many failures to reclaim anything from them and goes to | |
2852 | * sleep. On reclaim progress, reset the failure counter. A | |
2853 | * successful direct reclaim run will revive a dormant kswapd. | |
2854 | */ | |
2855 | if (reclaimable) | |
2856 | pgdat->kswapd_failures = 0; | |
f16015fb JW |
2857 | } |
2858 | ||
53853e2d | 2859 | /* |
fdd4c614 VB |
2860 | * Returns true if compaction should go ahead for a costly-order request, or |
2861 | * the allocation would already succeed without compaction. Return false if we | |
2862 | * should reclaim first. | |
53853e2d | 2863 | */ |
4f588331 | 2864 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 2865 | { |
31483b6a | 2866 | unsigned long watermark; |
fdd4c614 | 2867 | enum compact_result suitable; |
fe4b1b24 | 2868 | |
fdd4c614 VB |
2869 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
2870 | if (suitable == COMPACT_SUCCESS) | |
2871 | /* Allocation should succeed already. Don't reclaim. */ | |
2872 | return true; | |
2873 | if (suitable == COMPACT_SKIPPED) | |
2874 | /* Compaction cannot yet proceed. Do reclaim. */ | |
2875 | return false; | |
fe4b1b24 | 2876 | |
53853e2d | 2877 | /* |
fdd4c614 VB |
2878 | * Compaction is already possible, but it takes time to run and there |
2879 | * are potentially other callers using the pages just freed. So proceed | |
2880 | * with reclaim to make a buffer of free pages available to give | |
2881 | * compaction a reasonable chance of completing and allocating the page. | |
2882 | * Note that we won't actually reclaim the whole buffer in one attempt | |
2883 | * as the target watermark in should_continue_reclaim() is lower. But if | |
2884 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 2885 | */ |
fdd4c614 | 2886 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 2887 | |
fdd4c614 | 2888 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
2889 | } |
2890 | ||
1da177e4 LT |
2891 | /* |
2892 | * This is the direct reclaim path, for page-allocating processes. We only | |
2893 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2894 | * request. | |
2895 | * | |
1da177e4 LT |
2896 | * If a zone is deemed to be full of pinned pages then just give it a light |
2897 | * scan then give up on it. | |
2898 | */ | |
0a0337e0 | 2899 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2900 | { |
dd1a239f | 2901 | struct zoneref *z; |
54a6eb5c | 2902 | struct zone *zone; |
0608f43d AM |
2903 | unsigned long nr_soft_reclaimed; |
2904 | unsigned long nr_soft_scanned; | |
619d0d76 | 2905 | gfp_t orig_mask; |
79dafcdc | 2906 | pg_data_t *last_pgdat = NULL; |
1cfb419b | 2907 | |
cc715d99 MG |
2908 | /* |
2909 | * If the number of buffer_heads in the machine exceeds the maximum | |
2910 | * allowed level, force direct reclaim to scan the highmem zone as | |
2911 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2912 | */ | |
619d0d76 | 2913 | orig_mask = sc->gfp_mask; |
b2e18757 | 2914 | if (buffer_heads_over_limit) { |
cc715d99 | 2915 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 2916 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 2917 | } |
cc715d99 | 2918 | |
d4debc66 | 2919 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 2920 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
2921 | /* |
2922 | * Take care memory controller reclaiming has small influence | |
2923 | * to global LRU. | |
2924 | */ | |
b5ead35e | 2925 | if (!cgroup_reclaim(sc)) { |
344736f2 VD |
2926 | if (!cpuset_zone_allowed(zone, |
2927 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2928 | continue; |
65ec02cb | 2929 | |
0b06496a JW |
2930 | /* |
2931 | * If we already have plenty of memory free for | |
2932 | * compaction in this zone, don't free any more. | |
2933 | * Even though compaction is invoked for any | |
2934 | * non-zero order, only frequent costly order | |
2935 | * reclamation is disruptive enough to become a | |
2936 | * noticeable problem, like transparent huge | |
2937 | * page allocations. | |
2938 | */ | |
2939 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2940 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 2941 | compaction_ready(zone, sc)) { |
0b06496a JW |
2942 | sc->compaction_ready = true; |
2943 | continue; | |
e0887c19 | 2944 | } |
0b06496a | 2945 | |
79dafcdc MG |
2946 | /* |
2947 | * Shrink each node in the zonelist once. If the | |
2948 | * zonelist is ordered by zone (not the default) then a | |
2949 | * node may be shrunk multiple times but in that case | |
2950 | * the user prefers lower zones being preserved. | |
2951 | */ | |
2952 | if (zone->zone_pgdat == last_pgdat) | |
2953 | continue; | |
2954 | ||
0608f43d AM |
2955 | /* |
2956 | * This steals pages from memory cgroups over softlimit | |
2957 | * and returns the number of reclaimed pages and | |
2958 | * scanned pages. This works for global memory pressure | |
2959 | * and balancing, not for a memcg's limit. | |
2960 | */ | |
2961 | nr_soft_scanned = 0; | |
ef8f2327 | 2962 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
2963 | sc->order, sc->gfp_mask, |
2964 | &nr_soft_scanned); | |
2965 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2966 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2967 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2968 | } |
408d8544 | 2969 | |
79dafcdc MG |
2970 | /* See comment about same check for global reclaim above */ |
2971 | if (zone->zone_pgdat == last_pgdat) | |
2972 | continue; | |
2973 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 2974 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 2975 | } |
e0c23279 | 2976 | |
619d0d76 WY |
2977 | /* |
2978 | * Restore to original mask to avoid the impact on the caller if we | |
2979 | * promoted it to __GFP_HIGHMEM. | |
2980 | */ | |
2981 | sc->gfp_mask = orig_mask; | |
1da177e4 | 2982 | } |
4f98a2fe | 2983 | |
b910718a | 2984 | static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat) |
2a2e4885 | 2985 | { |
b910718a JW |
2986 | struct lruvec *target_lruvec; |
2987 | unsigned long refaults; | |
2a2e4885 | 2988 | |
b910718a | 2989 | target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat); |
170b04b7 JK |
2990 | refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON); |
2991 | target_lruvec->refaults[0] = refaults; | |
2992 | refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE); | |
2993 | target_lruvec->refaults[1] = refaults; | |
2a2e4885 JW |
2994 | } |
2995 | ||
1da177e4 LT |
2996 | /* |
2997 | * This is the main entry point to direct page reclaim. | |
2998 | * | |
2999 | * If a full scan of the inactive list fails to free enough memory then we | |
3000 | * are "out of memory" and something needs to be killed. | |
3001 | * | |
3002 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
3003 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
3004 | * caller can't do much about. We kick the writeback threads and take explicit |
3005 | * naps in the hope that some of these pages can be written. But if the | |
3006 | * allocating task holds filesystem locks which prevent writeout this might not | |
3007 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
3008 | * |
3009 | * returns: 0, if no pages reclaimed | |
3010 | * else, the number of pages reclaimed | |
1da177e4 | 3011 | */ |
dac1d27b | 3012 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 3013 | struct scan_control *sc) |
1da177e4 | 3014 | { |
241994ed | 3015 | int initial_priority = sc->priority; |
2a2e4885 JW |
3016 | pg_data_t *last_pgdat; |
3017 | struct zoneref *z; | |
3018 | struct zone *zone; | |
241994ed | 3019 | retry: |
873b4771 KK |
3020 | delayacct_freepages_start(); |
3021 | ||
b5ead35e | 3022 | if (!cgroup_reclaim(sc)) |
7cc30fcf | 3023 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 3024 | |
9e3b2f8c | 3025 | do { |
70ddf637 AV |
3026 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
3027 | sc->priority); | |
66e1707b | 3028 | sc->nr_scanned = 0; |
0a0337e0 | 3029 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 3030 | |
bb21c7ce | 3031 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
3032 | break; |
3033 | ||
3034 | if (sc->compaction_ready) | |
3035 | break; | |
1da177e4 | 3036 | |
0e50ce3b MK |
3037 | /* |
3038 | * If we're getting trouble reclaiming, start doing | |
3039 | * writepage even in laptop mode. | |
3040 | */ | |
3041 | if (sc->priority < DEF_PRIORITY - 2) | |
3042 | sc->may_writepage = 1; | |
0b06496a | 3043 | } while (--sc->priority >= 0); |
bb21c7ce | 3044 | |
2a2e4885 JW |
3045 | last_pgdat = NULL; |
3046 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
3047 | sc->nodemask) { | |
3048 | if (zone->zone_pgdat == last_pgdat) | |
3049 | continue; | |
3050 | last_pgdat = zone->zone_pgdat; | |
1b05117d | 3051 | |
2a2e4885 | 3052 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); |
1b05117d JW |
3053 | |
3054 | if (cgroup_reclaim(sc)) { | |
3055 | struct lruvec *lruvec; | |
3056 | ||
3057 | lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, | |
3058 | zone->zone_pgdat); | |
3059 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
3060 | } | |
2a2e4885 JW |
3061 | } |
3062 | ||
873b4771 KK |
3063 | delayacct_freepages_end(); |
3064 | ||
bb21c7ce KM |
3065 | if (sc->nr_reclaimed) |
3066 | return sc->nr_reclaimed; | |
3067 | ||
0cee34fd | 3068 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 3069 | if (sc->compaction_ready) |
7335084d MG |
3070 | return 1; |
3071 | ||
b91ac374 JW |
3072 | /* |
3073 | * We make inactive:active ratio decisions based on the node's | |
3074 | * composition of memory, but a restrictive reclaim_idx or a | |
3075 | * memory.low cgroup setting can exempt large amounts of | |
3076 | * memory from reclaim. Neither of which are very common, so | |
3077 | * instead of doing costly eligibility calculations of the | |
3078 | * entire cgroup subtree up front, we assume the estimates are | |
3079 | * good, and retry with forcible deactivation if that fails. | |
3080 | */ | |
3081 | if (sc->skipped_deactivate) { | |
3082 | sc->priority = initial_priority; | |
3083 | sc->force_deactivate = 1; | |
3084 | sc->skipped_deactivate = 0; | |
3085 | goto retry; | |
3086 | } | |
3087 | ||
241994ed | 3088 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
d6622f63 | 3089 | if (sc->memcg_low_skipped) { |
241994ed | 3090 | sc->priority = initial_priority; |
b91ac374 | 3091 | sc->force_deactivate = 0; |
d6622f63 YX |
3092 | sc->memcg_low_reclaim = 1; |
3093 | sc->memcg_low_skipped = 0; | |
241994ed JW |
3094 | goto retry; |
3095 | } | |
3096 | ||
bb21c7ce | 3097 | return 0; |
1da177e4 LT |
3098 | } |
3099 | ||
c73322d0 | 3100 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
5515061d MG |
3101 | { |
3102 | struct zone *zone; | |
3103 | unsigned long pfmemalloc_reserve = 0; | |
3104 | unsigned long free_pages = 0; | |
3105 | int i; | |
3106 | bool wmark_ok; | |
3107 | ||
c73322d0 JW |
3108 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
3109 | return true; | |
3110 | ||
5515061d MG |
3111 | for (i = 0; i <= ZONE_NORMAL; i++) { |
3112 | zone = &pgdat->node_zones[i]; | |
d450abd8 JW |
3113 | if (!managed_zone(zone)) |
3114 | continue; | |
3115 | ||
3116 | if (!zone_reclaimable_pages(zone)) | |
675becce MG |
3117 | continue; |
3118 | ||
5515061d MG |
3119 | pfmemalloc_reserve += min_wmark_pages(zone); |
3120 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
3121 | } | |
3122 | ||
675becce MG |
3123 | /* If there are no reserves (unexpected config) then do not throttle */ |
3124 | if (!pfmemalloc_reserve) | |
3125 | return true; | |
3126 | ||
5515061d MG |
3127 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
3128 | ||
3129 | /* kswapd must be awake if processes are being throttled */ | |
3130 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
97a225e6 JK |
3131 | if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL) |
3132 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL); | |
5644e1fb | 3133 | |
5515061d MG |
3134 | wake_up_interruptible(&pgdat->kswapd_wait); |
3135 | } | |
3136 | ||
3137 | return wmark_ok; | |
3138 | } | |
3139 | ||
3140 | /* | |
3141 | * Throttle direct reclaimers if backing storage is backed by the network | |
3142 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
3143 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
3144 | * when the low watermark is reached. |
3145 | * | |
3146 | * Returns true if a fatal signal was delivered during throttling. If this | |
3147 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 3148 | */ |
50694c28 | 3149 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
3150 | nodemask_t *nodemask) |
3151 | { | |
675becce | 3152 | struct zoneref *z; |
5515061d | 3153 | struct zone *zone; |
675becce | 3154 | pg_data_t *pgdat = NULL; |
5515061d MG |
3155 | |
3156 | /* | |
3157 | * Kernel threads should not be throttled as they may be indirectly | |
3158 | * responsible for cleaning pages necessary for reclaim to make forward | |
3159 | * progress. kjournald for example may enter direct reclaim while | |
3160 | * committing a transaction where throttling it could forcing other | |
3161 | * processes to block on log_wait_commit(). | |
3162 | */ | |
3163 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
3164 | goto out; |
3165 | ||
3166 | /* | |
3167 | * If a fatal signal is pending, this process should not throttle. | |
3168 | * It should return quickly so it can exit and free its memory | |
3169 | */ | |
3170 | if (fatal_signal_pending(current)) | |
3171 | goto out; | |
5515061d | 3172 | |
675becce MG |
3173 | /* |
3174 | * Check if the pfmemalloc reserves are ok by finding the first node | |
3175 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
3176 | * GFP_KERNEL will be required for allocating network buffers when | |
3177 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
3178 | * | |
3179 | * Throttling is based on the first usable node and throttled processes | |
3180 | * wait on a queue until kswapd makes progress and wakes them. There | |
3181 | * is an affinity then between processes waking up and where reclaim | |
3182 | * progress has been made assuming the process wakes on the same node. | |
3183 | * More importantly, processes running on remote nodes will not compete | |
3184 | * for remote pfmemalloc reserves and processes on different nodes | |
3185 | * should make reasonable progress. | |
3186 | */ | |
3187 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 3188 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
3189 | if (zone_idx(zone) > ZONE_NORMAL) |
3190 | continue; | |
3191 | ||
3192 | /* Throttle based on the first usable node */ | |
3193 | pgdat = zone->zone_pgdat; | |
c73322d0 | 3194 | if (allow_direct_reclaim(pgdat)) |
675becce MG |
3195 | goto out; |
3196 | break; | |
3197 | } | |
3198 | ||
3199 | /* If no zone was usable by the allocation flags then do not throttle */ | |
3200 | if (!pgdat) | |
50694c28 | 3201 | goto out; |
5515061d | 3202 | |
68243e76 MG |
3203 | /* Account for the throttling */ |
3204 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
3205 | ||
5515061d MG |
3206 | /* |
3207 | * If the caller cannot enter the filesystem, it's possible that it | |
3208 | * is due to the caller holding an FS lock or performing a journal | |
3209 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
3210 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
3211 | * blocked waiting on the same lock. Instead, throttle for up to a | |
3212 | * second before continuing. | |
3213 | */ | |
3214 | if (!(gfp_mask & __GFP_FS)) { | |
3215 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
c73322d0 | 3216 | allow_direct_reclaim(pgdat), HZ); |
50694c28 MG |
3217 | |
3218 | goto check_pending; | |
5515061d MG |
3219 | } |
3220 | ||
3221 | /* Throttle until kswapd wakes the process */ | |
3222 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
c73322d0 | 3223 | allow_direct_reclaim(pgdat)); |
50694c28 MG |
3224 | |
3225 | check_pending: | |
3226 | if (fatal_signal_pending(current)) | |
3227 | return true; | |
3228 | ||
3229 | out: | |
3230 | return false; | |
5515061d MG |
3231 | } |
3232 | ||
dac1d27b | 3233 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 3234 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 3235 | { |
33906bc5 | 3236 | unsigned long nr_reclaimed; |
66e1707b | 3237 | struct scan_control sc = { |
ee814fe2 | 3238 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
f2f43e56 | 3239 | .gfp_mask = current_gfp_context(gfp_mask), |
b2e18757 | 3240 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
3241 | .order = order, |
3242 | .nodemask = nodemask, | |
3243 | .priority = DEF_PRIORITY, | |
66e1707b | 3244 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3245 | .may_unmap = 1, |
2e2e4259 | 3246 | .may_swap = 1, |
66e1707b BS |
3247 | }; |
3248 | ||
bb451fdf GT |
3249 | /* |
3250 | * scan_control uses s8 fields for order, priority, and reclaim_idx. | |
3251 | * Confirm they are large enough for max values. | |
3252 | */ | |
3253 | BUILD_BUG_ON(MAX_ORDER > S8_MAX); | |
3254 | BUILD_BUG_ON(DEF_PRIORITY > S8_MAX); | |
3255 | BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX); | |
3256 | ||
5515061d | 3257 | /* |
50694c28 MG |
3258 | * Do not enter reclaim if fatal signal was delivered while throttled. |
3259 | * 1 is returned so that the page allocator does not OOM kill at this | |
3260 | * point. | |
5515061d | 3261 | */ |
f2f43e56 | 3262 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
5515061d MG |
3263 | return 1; |
3264 | ||
1732d2b0 | 3265 | set_task_reclaim_state(current, &sc.reclaim_state); |
3481c37f | 3266 | trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask); |
33906bc5 | 3267 | |
3115cd91 | 3268 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
3269 | |
3270 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
1732d2b0 | 3271 | set_task_reclaim_state(current, NULL); |
33906bc5 MG |
3272 | |
3273 | return nr_reclaimed; | |
66e1707b BS |
3274 | } |
3275 | ||
c255a458 | 3276 | #ifdef CONFIG_MEMCG |
66e1707b | 3277 | |
d2e5fb92 | 3278 | /* Only used by soft limit reclaim. Do not reuse for anything else. */ |
a9dd0a83 | 3279 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 3280 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 3281 | pg_data_t *pgdat, |
0ae5e89c | 3282 | unsigned long *nr_scanned) |
4e416953 | 3283 | { |
afaf07a6 | 3284 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
4e416953 | 3285 | struct scan_control sc = { |
b8f5c566 | 3286 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 3287 | .target_mem_cgroup = memcg, |
4e416953 BS |
3288 | .may_writepage = !laptop_mode, |
3289 | .may_unmap = 1, | |
b2e18757 | 3290 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 3291 | .may_swap = !noswap, |
4e416953 | 3292 | }; |
0ae5e89c | 3293 | |
d2e5fb92 MH |
3294 | WARN_ON_ONCE(!current->reclaim_state); |
3295 | ||
4e416953 BS |
3296 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3297 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 3298 | |
9e3b2f8c | 3299 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
3481c37f | 3300 | sc.gfp_mask); |
bdce6d9e | 3301 | |
4e416953 BS |
3302 | /* |
3303 | * NOTE: Although we can get the priority field, using it | |
3304 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 3305 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
3306 | * will pick up pages from other mem cgroup's as well. We hack |
3307 | * the priority and make it zero. | |
3308 | */ | |
afaf07a6 | 3309 | shrink_lruvec(lruvec, &sc); |
bdce6d9e KM |
3310 | |
3311 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
3312 | ||
0ae5e89c | 3313 | *nr_scanned = sc.nr_scanned; |
0308f7cf | 3314 | |
4e416953 BS |
3315 | return sc.nr_reclaimed; |
3316 | } | |
3317 | ||
72835c86 | 3318 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 3319 | unsigned long nr_pages, |
a7885eb8 | 3320 | gfp_t gfp_mask, |
b70a2a21 | 3321 | bool may_swap) |
66e1707b | 3322 | { |
bdce6d9e | 3323 | unsigned long nr_reclaimed; |
499118e9 | 3324 | unsigned int noreclaim_flag; |
66e1707b | 3325 | struct scan_control sc = { |
b70a2a21 | 3326 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 3327 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
a09ed5e0 | 3328 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
b2e18757 | 3329 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
3330 | .target_mem_cgroup = memcg, |
3331 | .priority = DEF_PRIORITY, | |
3332 | .may_writepage = !laptop_mode, | |
3333 | .may_unmap = 1, | |
b70a2a21 | 3334 | .may_swap = may_swap, |
a09ed5e0 | 3335 | }; |
889976db | 3336 | /* |
fa40d1ee SB |
3337 | * Traverse the ZONELIST_FALLBACK zonelist of the current node to put |
3338 | * equal pressure on all the nodes. This is based on the assumption that | |
3339 | * the reclaim does not bail out early. | |
889976db | 3340 | */ |
fa40d1ee | 3341 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
889976db | 3342 | |
fa40d1ee | 3343 | set_task_reclaim_state(current, &sc.reclaim_state); |
3481c37f | 3344 | trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask); |
499118e9 | 3345 | noreclaim_flag = memalloc_noreclaim_save(); |
eb414681 | 3346 | |
3115cd91 | 3347 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
eb414681 | 3348 | |
499118e9 | 3349 | memalloc_noreclaim_restore(noreclaim_flag); |
bdce6d9e | 3350 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); |
1732d2b0 | 3351 | set_task_reclaim_state(current, NULL); |
bdce6d9e KM |
3352 | |
3353 | return nr_reclaimed; | |
66e1707b BS |
3354 | } |
3355 | #endif | |
3356 | ||
1d82de61 | 3357 | static void age_active_anon(struct pglist_data *pgdat, |
ef8f2327 | 3358 | struct scan_control *sc) |
f16015fb | 3359 | { |
b95a2f2d | 3360 | struct mem_cgroup *memcg; |
b91ac374 | 3361 | struct lruvec *lruvec; |
f16015fb | 3362 | |
b95a2f2d JW |
3363 | if (!total_swap_pages) |
3364 | return; | |
3365 | ||
b91ac374 JW |
3366 | lruvec = mem_cgroup_lruvec(NULL, pgdat); |
3367 | if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON)) | |
3368 | return; | |
3369 | ||
b95a2f2d JW |
3370 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
3371 | do { | |
b91ac374 JW |
3372 | lruvec = mem_cgroup_lruvec(memcg, pgdat); |
3373 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
3374 | sc, LRU_ACTIVE_ANON); | |
b95a2f2d JW |
3375 | memcg = mem_cgroup_iter(NULL, memcg, NULL); |
3376 | } while (memcg); | |
f16015fb JW |
3377 | } |
3378 | ||
97a225e6 | 3379 | static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx) |
1c30844d MG |
3380 | { |
3381 | int i; | |
3382 | struct zone *zone; | |
3383 | ||
3384 | /* | |
3385 | * Check for watermark boosts top-down as the higher zones | |
3386 | * are more likely to be boosted. Both watermarks and boosts | |
1eba09c1 | 3387 | * should not be checked at the same time as reclaim would |
1c30844d MG |
3388 | * start prematurely when there is no boosting and a lower |
3389 | * zone is balanced. | |
3390 | */ | |
97a225e6 | 3391 | for (i = highest_zoneidx; i >= 0; i--) { |
1c30844d MG |
3392 | zone = pgdat->node_zones + i; |
3393 | if (!managed_zone(zone)) | |
3394 | continue; | |
3395 | ||
3396 | if (zone->watermark_boost) | |
3397 | return true; | |
3398 | } | |
3399 | ||
3400 | return false; | |
3401 | } | |
3402 | ||
e716f2eb MG |
3403 | /* |
3404 | * Returns true if there is an eligible zone balanced for the request order | |
97a225e6 | 3405 | * and highest_zoneidx |
e716f2eb | 3406 | */ |
97a225e6 | 3407 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx) |
60cefed4 | 3408 | { |
e716f2eb MG |
3409 | int i; |
3410 | unsigned long mark = -1; | |
3411 | struct zone *zone; | |
60cefed4 | 3412 | |
1c30844d MG |
3413 | /* |
3414 | * Check watermarks bottom-up as lower zones are more likely to | |
3415 | * meet watermarks. | |
3416 | */ | |
97a225e6 | 3417 | for (i = 0; i <= highest_zoneidx; i++) { |
e716f2eb | 3418 | zone = pgdat->node_zones + i; |
6256c6b4 | 3419 | |
e716f2eb MG |
3420 | if (!managed_zone(zone)) |
3421 | continue; | |
3422 | ||
3423 | mark = high_wmark_pages(zone); | |
97a225e6 | 3424 | if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx)) |
e716f2eb MG |
3425 | return true; |
3426 | } | |
3427 | ||
3428 | /* | |
97a225e6 | 3429 | * If a node has no populated zone within highest_zoneidx, it does not |
e716f2eb MG |
3430 | * need balancing by definition. This can happen if a zone-restricted |
3431 | * allocation tries to wake a remote kswapd. | |
3432 | */ | |
3433 | if (mark == -1) | |
3434 | return true; | |
3435 | ||
3436 | return false; | |
60cefed4 JW |
3437 | } |
3438 | ||
631b6e08 MG |
3439 | /* Clear pgdat state for congested, dirty or under writeback. */ |
3440 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
3441 | { | |
1b05117d JW |
3442 | struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat); |
3443 | ||
3444 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
631b6e08 MG |
3445 | clear_bit(PGDAT_DIRTY, &pgdat->flags); |
3446 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
3447 | } | |
3448 | ||
5515061d MG |
3449 | /* |
3450 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3451 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3452 | * | |
3453 | * Returns true if kswapd is ready to sleep | |
3454 | */ | |
97a225e6 JK |
3455 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, |
3456 | int highest_zoneidx) | |
f50de2d3 | 3457 | { |
5515061d | 3458 | /* |
9e5e3661 | 3459 | * The throttled processes are normally woken up in balance_pgdat() as |
c73322d0 | 3460 | * soon as allow_direct_reclaim() is true. But there is a potential |
9e5e3661 VB |
3461 | * race between when kswapd checks the watermarks and a process gets |
3462 | * throttled. There is also a potential race if processes get | |
3463 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3464 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3465 | * the wake up checks. If kswapd is going to sleep, no process should | |
3466 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3467 | * the wake up is premature, processes will wake kswapd and get | |
3468 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3469 | * that here we are under prepare_to_wait(). | |
5515061d | 3470 | */ |
9e5e3661 VB |
3471 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3472 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3473 | |
c73322d0 JW |
3474 | /* Hopeless node, leave it to direct reclaim */ |
3475 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3476 | return true; | |
3477 | ||
97a225e6 | 3478 | if (pgdat_balanced(pgdat, order, highest_zoneidx)) { |
e716f2eb MG |
3479 | clear_pgdat_congested(pgdat); |
3480 | return true; | |
1d82de61 MG |
3481 | } |
3482 | ||
333b0a45 | 3483 | return false; |
f50de2d3 MG |
3484 | } |
3485 | ||
75485363 | 3486 | /* |
1d82de61 MG |
3487 | * kswapd shrinks a node of pages that are at or below the highest usable |
3488 | * zone that is currently unbalanced. | |
b8e83b94 MG |
3489 | * |
3490 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3491 | * reclaim or if the lack of progress was due to pages under writeback. |
3492 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3493 | */ |
1d82de61 | 3494 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 3495 | struct scan_control *sc) |
75485363 | 3496 | { |
1d82de61 MG |
3497 | struct zone *zone; |
3498 | int z; | |
75485363 | 3499 | |
1d82de61 MG |
3500 | /* Reclaim a number of pages proportional to the number of zones */ |
3501 | sc->nr_to_reclaim = 0; | |
970a39a3 | 3502 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 3503 | zone = pgdat->node_zones + z; |
6aa303de | 3504 | if (!managed_zone(zone)) |
1d82de61 | 3505 | continue; |
7c954f6d | 3506 | |
1d82de61 MG |
3507 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
3508 | } | |
7c954f6d MG |
3509 | |
3510 | /* | |
1d82de61 MG |
3511 | * Historically care was taken to put equal pressure on all zones but |
3512 | * now pressure is applied based on node LRU order. | |
7c954f6d | 3513 | */ |
970a39a3 | 3514 | shrink_node(pgdat, sc); |
283aba9f | 3515 | |
7c954f6d | 3516 | /* |
1d82de61 MG |
3517 | * Fragmentation may mean that the system cannot be rebalanced for |
3518 | * high-order allocations. If twice the allocation size has been | |
3519 | * reclaimed then recheck watermarks only at order-0 to prevent | |
3520 | * excessive reclaim. Assume that a process requested a high-order | |
3521 | * can direct reclaim/compact. | |
7c954f6d | 3522 | */ |
9861a62c | 3523 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 3524 | sc->order = 0; |
7c954f6d | 3525 | |
b8e83b94 | 3526 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3527 | } |
3528 | ||
1da177e4 | 3529 | /* |
1d82de61 MG |
3530 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
3531 | * that are eligible for use by the caller until at least one zone is | |
3532 | * balanced. | |
1da177e4 | 3533 | * |
1d82de61 | 3534 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
3535 | * |
3536 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 3537 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
8bb4e7a2 | 3538 | * found to have free_pages <= high_wmark_pages(zone), any page in that zone |
1d82de61 MG |
3539 | * or lower is eligible for reclaim until at least one usable zone is |
3540 | * balanced. | |
1da177e4 | 3541 | */ |
97a225e6 | 3542 | static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx) |
1da177e4 | 3543 | { |
1da177e4 | 3544 | int i; |
0608f43d AM |
3545 | unsigned long nr_soft_reclaimed; |
3546 | unsigned long nr_soft_scanned; | |
eb414681 | 3547 | unsigned long pflags; |
1c30844d MG |
3548 | unsigned long nr_boost_reclaim; |
3549 | unsigned long zone_boosts[MAX_NR_ZONES] = { 0, }; | |
3550 | bool boosted; | |
1d82de61 | 3551 | struct zone *zone; |
179e9639 AM |
3552 | struct scan_control sc = { |
3553 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3554 | .order = order, |
a6dc60f8 | 3555 | .may_unmap = 1, |
179e9639 | 3556 | }; |
93781325 | 3557 | |
1732d2b0 | 3558 | set_task_reclaim_state(current, &sc.reclaim_state); |
eb414681 | 3559 | psi_memstall_enter(&pflags); |
93781325 OS |
3560 | __fs_reclaim_acquire(); |
3561 | ||
f8891e5e | 3562 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3563 | |
1c30844d MG |
3564 | /* |
3565 | * Account for the reclaim boost. Note that the zone boost is left in | |
3566 | * place so that parallel allocations that are near the watermark will | |
3567 | * stall or direct reclaim until kswapd is finished. | |
3568 | */ | |
3569 | nr_boost_reclaim = 0; | |
97a225e6 | 3570 | for (i = 0; i <= highest_zoneidx; i++) { |
1c30844d MG |
3571 | zone = pgdat->node_zones + i; |
3572 | if (!managed_zone(zone)) | |
3573 | continue; | |
3574 | ||
3575 | nr_boost_reclaim += zone->watermark_boost; | |
3576 | zone_boosts[i] = zone->watermark_boost; | |
3577 | } | |
3578 | boosted = nr_boost_reclaim; | |
3579 | ||
3580 | restart: | |
3581 | sc.priority = DEF_PRIORITY; | |
9e3b2f8c | 3582 | do { |
c73322d0 | 3583 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
b8e83b94 | 3584 | bool raise_priority = true; |
1c30844d | 3585 | bool balanced; |
93781325 | 3586 | bool ret; |
b8e83b94 | 3587 | |
97a225e6 | 3588 | sc.reclaim_idx = highest_zoneidx; |
1da177e4 | 3589 | |
86c79f6b | 3590 | /* |
84c7a777 MG |
3591 | * If the number of buffer_heads exceeds the maximum allowed |
3592 | * then consider reclaiming from all zones. This has a dual | |
3593 | * purpose -- on 64-bit systems it is expected that | |
3594 | * buffer_heads are stripped during active rotation. On 32-bit | |
3595 | * systems, highmem pages can pin lowmem memory and shrinking | |
3596 | * buffers can relieve lowmem pressure. Reclaim may still not | |
3597 | * go ahead if all eligible zones for the original allocation | |
3598 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
3599 | */ |
3600 | if (buffer_heads_over_limit) { | |
3601 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
3602 | zone = pgdat->node_zones + i; | |
6aa303de | 3603 | if (!managed_zone(zone)) |
86c79f6b | 3604 | continue; |
cc715d99 | 3605 | |
970a39a3 | 3606 | sc.reclaim_idx = i; |
e1dbeda6 | 3607 | break; |
1da177e4 | 3608 | } |
1da177e4 | 3609 | } |
dafcb73e | 3610 | |
86c79f6b | 3611 | /* |
1c30844d MG |
3612 | * If the pgdat is imbalanced then ignore boosting and preserve |
3613 | * the watermarks for a later time and restart. Note that the | |
3614 | * zone watermarks will be still reset at the end of balancing | |
3615 | * on the grounds that the normal reclaim should be enough to | |
3616 | * re-evaluate if boosting is required when kswapd next wakes. | |
3617 | */ | |
97a225e6 | 3618 | balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx); |
1c30844d MG |
3619 | if (!balanced && nr_boost_reclaim) { |
3620 | nr_boost_reclaim = 0; | |
3621 | goto restart; | |
3622 | } | |
3623 | ||
3624 | /* | |
3625 | * If boosting is not active then only reclaim if there are no | |
3626 | * eligible zones. Note that sc.reclaim_idx is not used as | |
3627 | * buffer_heads_over_limit may have adjusted it. | |
86c79f6b | 3628 | */ |
1c30844d | 3629 | if (!nr_boost_reclaim && balanced) |
e716f2eb | 3630 | goto out; |
e1dbeda6 | 3631 | |
1c30844d MG |
3632 | /* Limit the priority of boosting to avoid reclaim writeback */ |
3633 | if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2) | |
3634 | raise_priority = false; | |
3635 | ||
3636 | /* | |
3637 | * Do not writeback or swap pages for boosted reclaim. The | |
3638 | * intent is to relieve pressure not issue sub-optimal IO | |
3639 | * from reclaim context. If no pages are reclaimed, the | |
3640 | * reclaim will be aborted. | |
3641 | */ | |
3642 | sc.may_writepage = !laptop_mode && !nr_boost_reclaim; | |
3643 | sc.may_swap = !nr_boost_reclaim; | |
1c30844d | 3644 | |
1d82de61 MG |
3645 | /* |
3646 | * Do some background aging of the anon list, to give | |
3647 | * pages a chance to be referenced before reclaiming. All | |
3648 | * pages are rotated regardless of classzone as this is | |
3649 | * about consistent aging. | |
3650 | */ | |
ef8f2327 | 3651 | age_active_anon(pgdat, &sc); |
1d82de61 | 3652 | |
b7ea3c41 MG |
3653 | /* |
3654 | * If we're getting trouble reclaiming, start doing writepage | |
3655 | * even in laptop mode. | |
3656 | */ | |
047d72c3 | 3657 | if (sc.priority < DEF_PRIORITY - 2) |
b7ea3c41 MG |
3658 | sc.may_writepage = 1; |
3659 | ||
1d82de61 MG |
3660 | /* Call soft limit reclaim before calling shrink_node. */ |
3661 | sc.nr_scanned = 0; | |
3662 | nr_soft_scanned = 0; | |
ef8f2327 | 3663 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
3664 | sc.gfp_mask, &nr_soft_scanned); |
3665 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3666 | ||
1da177e4 | 3667 | /* |
1d82de61 MG |
3668 | * There should be no need to raise the scanning priority if |
3669 | * enough pages are already being scanned that that high | |
3670 | * watermark would be met at 100% efficiency. | |
1da177e4 | 3671 | */ |
970a39a3 | 3672 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 3673 | raise_priority = false; |
5515061d MG |
3674 | |
3675 | /* | |
3676 | * If the low watermark is met there is no need for processes | |
3677 | * to be throttled on pfmemalloc_wait as they should not be | |
3678 | * able to safely make forward progress. Wake them | |
3679 | */ | |
3680 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
c73322d0 | 3681 | allow_direct_reclaim(pgdat)) |
cfc51155 | 3682 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3683 | |
b8e83b94 | 3684 | /* Check if kswapd should be suspending */ |
93781325 OS |
3685 | __fs_reclaim_release(); |
3686 | ret = try_to_freeze(); | |
3687 | __fs_reclaim_acquire(); | |
3688 | if (ret || kthread_should_stop()) | |
b8e83b94 | 3689 | break; |
8357376d | 3690 | |
73ce02e9 | 3691 | /* |
b8e83b94 MG |
3692 | * Raise priority if scanning rate is too low or there was no |
3693 | * progress in reclaiming pages | |
73ce02e9 | 3694 | */ |
c73322d0 | 3695 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
1c30844d MG |
3696 | nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed); |
3697 | ||
3698 | /* | |
3699 | * If reclaim made no progress for a boost, stop reclaim as | |
3700 | * IO cannot be queued and it could be an infinite loop in | |
3701 | * extreme circumstances. | |
3702 | */ | |
3703 | if (nr_boost_reclaim && !nr_reclaimed) | |
3704 | break; | |
3705 | ||
c73322d0 | 3706 | if (raise_priority || !nr_reclaimed) |
b8e83b94 | 3707 | sc.priority--; |
1d82de61 | 3708 | } while (sc.priority >= 1); |
1da177e4 | 3709 | |
c73322d0 JW |
3710 | if (!sc.nr_reclaimed) |
3711 | pgdat->kswapd_failures++; | |
3712 | ||
b8e83b94 | 3713 | out: |
1c30844d MG |
3714 | /* If reclaim was boosted, account for the reclaim done in this pass */ |
3715 | if (boosted) { | |
3716 | unsigned long flags; | |
3717 | ||
97a225e6 | 3718 | for (i = 0; i <= highest_zoneidx; i++) { |
1c30844d MG |
3719 | if (!zone_boosts[i]) |
3720 | continue; | |
3721 | ||
3722 | /* Increments are under the zone lock */ | |
3723 | zone = pgdat->node_zones + i; | |
3724 | spin_lock_irqsave(&zone->lock, flags); | |
3725 | zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]); | |
3726 | spin_unlock_irqrestore(&zone->lock, flags); | |
3727 | } | |
3728 | ||
3729 | /* | |
3730 | * As there is now likely space, wakeup kcompact to defragment | |
3731 | * pageblocks. | |
3732 | */ | |
97a225e6 | 3733 | wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx); |
1c30844d MG |
3734 | } |
3735 | ||
2a2e4885 | 3736 | snapshot_refaults(NULL, pgdat); |
93781325 | 3737 | __fs_reclaim_release(); |
eb414681 | 3738 | psi_memstall_leave(&pflags); |
1732d2b0 | 3739 | set_task_reclaim_state(current, NULL); |
e5ca8071 | 3740 | |
0abdee2b | 3741 | /* |
1d82de61 MG |
3742 | * Return the order kswapd stopped reclaiming at as |
3743 | * prepare_kswapd_sleep() takes it into account. If another caller | |
3744 | * entered the allocator slow path while kswapd was awake, order will | |
3745 | * remain at the higher level. | |
0abdee2b | 3746 | */ |
1d82de61 | 3747 | return sc.order; |
1da177e4 LT |
3748 | } |
3749 | ||
e716f2eb | 3750 | /* |
97a225e6 JK |
3751 | * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to |
3752 | * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is | |
3753 | * not a valid index then either kswapd runs for first time or kswapd couldn't | |
3754 | * sleep after previous reclaim attempt (node is still unbalanced). In that | |
3755 | * case return the zone index of the previous kswapd reclaim cycle. | |
e716f2eb | 3756 | */ |
97a225e6 JK |
3757 | static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat, |
3758 | enum zone_type prev_highest_zoneidx) | |
e716f2eb | 3759 | { |
97a225e6 | 3760 | enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx); |
5644e1fb | 3761 | |
97a225e6 | 3762 | return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx; |
e716f2eb MG |
3763 | } |
3764 | ||
38087d9b | 3765 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
97a225e6 | 3766 | unsigned int highest_zoneidx) |
f0bc0a60 KM |
3767 | { |
3768 | long remaining = 0; | |
3769 | DEFINE_WAIT(wait); | |
3770 | ||
3771 | if (freezing(current) || kthread_should_stop()) | |
3772 | return; | |
3773 | ||
3774 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3775 | ||
333b0a45 SG |
3776 | /* |
3777 | * Try to sleep for a short interval. Note that kcompactd will only be | |
3778 | * woken if it is possible to sleep for a short interval. This is | |
3779 | * deliberate on the assumption that if reclaim cannot keep an | |
3780 | * eligible zone balanced that it's also unlikely that compaction will | |
3781 | * succeed. | |
3782 | */ | |
97a225e6 | 3783 | if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) { |
fd901c95 VB |
3784 | /* |
3785 | * Compaction records what page blocks it recently failed to | |
3786 | * isolate pages from and skips them in the future scanning. | |
3787 | * When kswapd is going to sleep, it is reasonable to assume | |
3788 | * that pages and compaction may succeed so reset the cache. | |
3789 | */ | |
3790 | reset_isolation_suitable(pgdat); | |
3791 | ||
3792 | /* | |
3793 | * We have freed the memory, now we should compact it to make | |
3794 | * allocation of the requested order possible. | |
3795 | */ | |
97a225e6 | 3796 | wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx); |
fd901c95 | 3797 | |
f0bc0a60 | 3798 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
3799 | |
3800 | /* | |
97a225e6 | 3801 | * If woken prematurely then reset kswapd_highest_zoneidx and |
38087d9b MG |
3802 | * order. The values will either be from a wakeup request or |
3803 | * the previous request that slept prematurely. | |
3804 | */ | |
3805 | if (remaining) { | |
97a225e6 JK |
3806 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, |
3807 | kswapd_highest_zoneidx(pgdat, | |
3808 | highest_zoneidx)); | |
5644e1fb QC |
3809 | |
3810 | if (READ_ONCE(pgdat->kswapd_order) < reclaim_order) | |
3811 | WRITE_ONCE(pgdat->kswapd_order, reclaim_order); | |
38087d9b MG |
3812 | } |
3813 | ||
f0bc0a60 KM |
3814 | finish_wait(&pgdat->kswapd_wait, &wait); |
3815 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3816 | } | |
3817 | ||
3818 | /* | |
3819 | * After a short sleep, check if it was a premature sleep. If not, then | |
3820 | * go fully to sleep until explicitly woken up. | |
3821 | */ | |
d9f21d42 | 3822 | if (!remaining && |
97a225e6 | 3823 | prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) { |
f0bc0a60 KM |
3824 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3825 | ||
3826 | /* | |
3827 | * vmstat counters are not perfectly accurate and the estimated | |
3828 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3829 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3830 | * watermarks being breached while under pressure, we reduce the | |
3831 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3832 | * them before going back to sleep. | |
3833 | */ | |
3834 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
3835 | |
3836 | if (!kthread_should_stop()) | |
3837 | schedule(); | |
3838 | ||
f0bc0a60 KM |
3839 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3840 | } else { | |
3841 | if (remaining) | |
3842 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3843 | else | |
3844 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3845 | } | |
3846 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3847 | } | |
3848 | ||
1da177e4 LT |
3849 | /* |
3850 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3851 | * from the init process. |
1da177e4 LT |
3852 | * |
3853 | * This basically trickles out pages so that we have _some_ | |
3854 | * free memory available even if there is no other activity | |
3855 | * that frees anything up. This is needed for things like routing | |
3856 | * etc, where we otherwise might have all activity going on in | |
3857 | * asynchronous contexts that cannot page things out. | |
3858 | * | |
3859 | * If there are applications that are active memory-allocators | |
3860 | * (most normal use), this basically shouldn't matter. | |
3861 | */ | |
3862 | static int kswapd(void *p) | |
3863 | { | |
e716f2eb | 3864 | unsigned int alloc_order, reclaim_order; |
97a225e6 | 3865 | unsigned int highest_zoneidx = MAX_NR_ZONES - 1; |
1da177e4 LT |
3866 | pg_data_t *pgdat = (pg_data_t*)p; |
3867 | struct task_struct *tsk = current; | |
a70f7302 | 3868 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3869 | |
174596a0 | 3870 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3871 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3872 | |
3873 | /* | |
3874 | * Tell the memory management that we're a "memory allocator", | |
3875 | * and that if we need more memory we should get access to it | |
3876 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3877 | * never get caught in the normal page freeing logic. | |
3878 | * | |
3879 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3880 | * you need a small amount of memory in order to be able to | |
3881 | * page out something else, and this flag essentially protects | |
3882 | * us from recursively trying to free more memory as we're | |
3883 | * trying to free the first piece of memory in the first place). | |
3884 | */ | |
930d9152 | 3885 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3886 | set_freezable(); |
1da177e4 | 3887 | |
5644e1fb | 3888 | WRITE_ONCE(pgdat->kswapd_order, 0); |
97a225e6 | 3889 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES); |
1da177e4 | 3890 | for ( ; ; ) { |
6f6313d4 | 3891 | bool ret; |
3e1d1d28 | 3892 | |
5644e1fb | 3893 | alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order); |
97a225e6 JK |
3894 | highest_zoneidx = kswapd_highest_zoneidx(pgdat, |
3895 | highest_zoneidx); | |
e716f2eb | 3896 | |
38087d9b MG |
3897 | kswapd_try_sleep: |
3898 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
97a225e6 | 3899 | highest_zoneidx); |
215ddd66 | 3900 | |
97a225e6 | 3901 | /* Read the new order and highest_zoneidx */ |
2b47a24c | 3902 | alloc_order = READ_ONCE(pgdat->kswapd_order); |
97a225e6 JK |
3903 | highest_zoneidx = kswapd_highest_zoneidx(pgdat, |
3904 | highest_zoneidx); | |
5644e1fb | 3905 | WRITE_ONCE(pgdat->kswapd_order, 0); |
97a225e6 | 3906 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES); |
1da177e4 | 3907 | |
8fe23e05 DR |
3908 | ret = try_to_freeze(); |
3909 | if (kthread_should_stop()) | |
3910 | break; | |
3911 | ||
3912 | /* | |
3913 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3914 | * after returning from the refrigerator | |
3915 | */ | |
38087d9b MG |
3916 | if (ret) |
3917 | continue; | |
3918 | ||
3919 | /* | |
3920 | * Reclaim begins at the requested order but if a high-order | |
3921 | * reclaim fails then kswapd falls back to reclaiming for | |
3922 | * order-0. If that happens, kswapd will consider sleeping | |
3923 | * for the order it finished reclaiming at (reclaim_order) | |
3924 | * but kcompactd is woken to compact for the original | |
3925 | * request (alloc_order). | |
3926 | */ | |
97a225e6 | 3927 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx, |
e5146b12 | 3928 | alloc_order); |
97a225e6 JK |
3929 | reclaim_order = balance_pgdat(pgdat, alloc_order, |
3930 | highest_zoneidx); | |
38087d9b MG |
3931 | if (reclaim_order < alloc_order) |
3932 | goto kswapd_try_sleep; | |
1da177e4 | 3933 | } |
b0a8cc58 | 3934 | |
71abdc15 | 3935 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
71abdc15 | 3936 | |
1da177e4 LT |
3937 | return 0; |
3938 | } | |
3939 | ||
3940 | /* | |
5ecd9d40 DR |
3941 | * A zone is low on free memory or too fragmented for high-order memory. If |
3942 | * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's | |
3943 | * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim | |
3944 | * has failed or is not needed, still wake up kcompactd if only compaction is | |
3945 | * needed. | |
1da177e4 | 3946 | */ |
5ecd9d40 | 3947 | void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, |
97a225e6 | 3948 | enum zone_type highest_zoneidx) |
1da177e4 LT |
3949 | { |
3950 | pg_data_t *pgdat; | |
5644e1fb | 3951 | enum zone_type curr_idx; |
1da177e4 | 3952 | |
6aa303de | 3953 | if (!managed_zone(zone)) |
1da177e4 LT |
3954 | return; |
3955 | ||
5ecd9d40 | 3956 | if (!cpuset_zone_allowed(zone, gfp_flags)) |
1da177e4 | 3957 | return; |
5644e1fb | 3958 | |
88f5acf8 | 3959 | pgdat = zone->zone_pgdat; |
97a225e6 | 3960 | curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx); |
5644e1fb | 3961 | |
97a225e6 JK |
3962 | if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx) |
3963 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx); | |
5644e1fb QC |
3964 | |
3965 | if (READ_ONCE(pgdat->kswapd_order) < order) | |
3966 | WRITE_ONCE(pgdat->kswapd_order, order); | |
dffcac2c | 3967 | |
8d0986e2 | 3968 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3969 | return; |
e1a55637 | 3970 | |
5ecd9d40 DR |
3971 | /* Hopeless node, leave it to direct reclaim if possible */ |
3972 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES || | |
97a225e6 JK |
3973 | (pgdat_balanced(pgdat, order, highest_zoneidx) && |
3974 | !pgdat_watermark_boosted(pgdat, highest_zoneidx))) { | |
5ecd9d40 DR |
3975 | /* |
3976 | * There may be plenty of free memory available, but it's too | |
3977 | * fragmented for high-order allocations. Wake up kcompactd | |
3978 | * and rely on compaction_suitable() to determine if it's | |
3979 | * needed. If it fails, it will defer subsequent attempts to | |
3980 | * ratelimit its work. | |
3981 | */ | |
3982 | if (!(gfp_flags & __GFP_DIRECT_RECLAIM)) | |
97a225e6 | 3983 | wakeup_kcompactd(pgdat, order, highest_zoneidx); |
e716f2eb | 3984 | return; |
5ecd9d40 | 3985 | } |
88f5acf8 | 3986 | |
97a225e6 | 3987 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order, |
5ecd9d40 | 3988 | gfp_flags); |
8d0986e2 | 3989 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3990 | } |
3991 | ||
c6f37f12 | 3992 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3993 | /* |
7b51755c | 3994 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3995 | * freed pages. |
3996 | * | |
3997 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3998 | * LRU order by reclaiming preferentially | |
3999 | * inactive > active > active referenced > active mapped | |
1da177e4 | 4000 | */ |
7b51755c | 4001 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 4002 | { |
d6277db4 | 4003 | struct scan_control sc = { |
ee814fe2 | 4004 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 4005 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 4006 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 4007 | .priority = DEF_PRIORITY, |
d6277db4 | 4008 | .may_writepage = 1, |
ee814fe2 JW |
4009 | .may_unmap = 1, |
4010 | .may_swap = 1, | |
7b51755c | 4011 | .hibernation_mode = 1, |
1da177e4 | 4012 | }; |
a09ed5e0 | 4013 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c | 4014 | unsigned long nr_reclaimed; |
499118e9 | 4015 | unsigned int noreclaim_flag; |
1da177e4 | 4016 | |
d92a8cfc | 4017 | fs_reclaim_acquire(sc.gfp_mask); |
93781325 | 4018 | noreclaim_flag = memalloc_noreclaim_save(); |
1732d2b0 | 4019 | set_task_reclaim_state(current, &sc.reclaim_state); |
d6277db4 | 4020 | |
3115cd91 | 4021 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 4022 | |
1732d2b0 | 4023 | set_task_reclaim_state(current, NULL); |
499118e9 | 4024 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 4025 | fs_reclaim_release(sc.gfp_mask); |
d6277db4 | 4026 | |
7b51755c | 4027 | return nr_reclaimed; |
1da177e4 | 4028 | } |
c6f37f12 | 4029 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 4030 | |
3218ae14 YG |
4031 | /* |
4032 | * This kswapd start function will be called by init and node-hot-add. | |
4033 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
4034 | */ | |
4035 | int kswapd_run(int nid) | |
4036 | { | |
4037 | pg_data_t *pgdat = NODE_DATA(nid); | |
4038 | int ret = 0; | |
4039 | ||
4040 | if (pgdat->kswapd) | |
4041 | return 0; | |
4042 | ||
4043 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
4044 | if (IS_ERR(pgdat->kswapd)) { | |
4045 | /* failure at boot is fatal */ | |
c6202adf | 4046 | BUG_ON(system_state < SYSTEM_RUNNING); |
d5dc0ad9 GS |
4047 | pr_err("Failed to start kswapd on node %d\n", nid); |
4048 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 4049 | pgdat->kswapd = NULL; |
3218ae14 YG |
4050 | } |
4051 | return ret; | |
4052 | } | |
4053 | ||
8fe23e05 | 4054 | /* |
d8adde17 | 4055 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 4056 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
4057 | */ |
4058 | void kswapd_stop(int nid) | |
4059 | { | |
4060 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
4061 | ||
d8adde17 | 4062 | if (kswapd) { |
8fe23e05 | 4063 | kthread_stop(kswapd); |
d8adde17 JL |
4064 | NODE_DATA(nid)->kswapd = NULL; |
4065 | } | |
8fe23e05 DR |
4066 | } |
4067 | ||
1da177e4 LT |
4068 | static int __init kswapd_init(void) |
4069 | { | |
6b700b5b | 4070 | int nid; |
69e05944 | 4071 | |
1da177e4 | 4072 | swap_setup(); |
48fb2e24 | 4073 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 4074 | kswapd_run(nid); |
1da177e4 LT |
4075 | return 0; |
4076 | } | |
4077 | ||
4078 | module_init(kswapd_init) | |
9eeff239 CL |
4079 | |
4080 | #ifdef CONFIG_NUMA | |
4081 | /* | |
a5f5f91d | 4082 | * Node reclaim mode |
9eeff239 | 4083 | * |
a5f5f91d | 4084 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 4085 | * the watermarks. |
9eeff239 | 4086 | */ |
a5f5f91d | 4087 | int node_reclaim_mode __read_mostly; |
9eeff239 | 4088 | |
a92f7126 | 4089 | /* |
a5f5f91d | 4090 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
4091 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
4092 | * a zone. | |
4093 | */ | |
a5f5f91d | 4094 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 4095 | |
9614634f | 4096 | /* |
a5f5f91d | 4097 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
4098 | * occur. |
4099 | */ | |
4100 | int sysctl_min_unmapped_ratio = 1; | |
4101 | ||
0ff38490 CL |
4102 | /* |
4103 | * If the number of slab pages in a zone grows beyond this percentage then | |
4104 | * slab reclaim needs to occur. | |
4105 | */ | |
4106 | int sysctl_min_slab_ratio = 5; | |
4107 | ||
11fb9989 | 4108 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 4109 | { |
11fb9989 MG |
4110 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
4111 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
4112 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
4113 | |
4114 | /* | |
4115 | * It's possible for there to be more file mapped pages than | |
4116 | * accounted for by the pages on the file LRU lists because | |
4117 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
4118 | */ | |
4119 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
4120 | } | |
4121 | ||
4122 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 4123 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 4124 | { |
d031a157 AM |
4125 | unsigned long nr_pagecache_reclaimable; |
4126 | unsigned long delta = 0; | |
90afa5de MG |
4127 | |
4128 | /* | |
95bbc0c7 | 4129 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 4130 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 4131 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
4132 | * a better estimate |
4133 | */ | |
a5f5f91d MG |
4134 | if (node_reclaim_mode & RECLAIM_UNMAP) |
4135 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 4136 | else |
a5f5f91d | 4137 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
4138 | |
4139 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
4140 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
4141 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
4142 | |
4143 | /* Watch for any possible underflows due to delta */ | |
4144 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
4145 | delta = nr_pagecache_reclaimable; | |
4146 | ||
4147 | return nr_pagecache_reclaimable - delta; | |
4148 | } | |
4149 | ||
9eeff239 | 4150 | /* |
a5f5f91d | 4151 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 4152 | */ |
a5f5f91d | 4153 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 4154 | { |
7fb2d46d | 4155 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 4156 | const unsigned long nr_pages = 1 << order; |
9eeff239 | 4157 | struct task_struct *p = current; |
499118e9 | 4158 | unsigned int noreclaim_flag; |
179e9639 | 4159 | struct scan_control sc = { |
62b726c1 | 4160 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
f2f43e56 | 4161 | .gfp_mask = current_gfp_context(gfp_mask), |
bd2f6199 | 4162 | .order = order, |
a5f5f91d MG |
4163 | .priority = NODE_RECLAIM_PRIORITY, |
4164 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
4165 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 4166 | .may_swap = 1, |
f2f43e56 | 4167 | .reclaim_idx = gfp_zone(gfp_mask), |
179e9639 | 4168 | }; |
9eeff239 | 4169 | |
132bb8cf YS |
4170 | trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order, |
4171 | sc.gfp_mask); | |
4172 | ||
9eeff239 | 4173 | cond_resched(); |
93781325 | 4174 | fs_reclaim_acquire(sc.gfp_mask); |
d4f7796e | 4175 | /* |
95bbc0c7 | 4176 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 4177 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 4178 | * and RECLAIM_UNMAP. |
d4f7796e | 4179 | */ |
499118e9 VB |
4180 | noreclaim_flag = memalloc_noreclaim_save(); |
4181 | p->flags |= PF_SWAPWRITE; | |
1732d2b0 | 4182 | set_task_reclaim_state(p, &sc.reclaim_state); |
c84db23c | 4183 | |
a5f5f91d | 4184 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
0ff38490 | 4185 | /* |
894befec | 4186 | * Free memory by calling shrink node with increasing |
0ff38490 CL |
4187 | * priorities until we have enough memory freed. |
4188 | */ | |
0ff38490 | 4189 | do { |
970a39a3 | 4190 | shrink_node(pgdat, &sc); |
9e3b2f8c | 4191 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 4192 | } |
c84db23c | 4193 | |
1732d2b0 | 4194 | set_task_reclaim_state(p, NULL); |
499118e9 VB |
4195 | current->flags &= ~PF_SWAPWRITE; |
4196 | memalloc_noreclaim_restore(noreclaim_flag); | |
93781325 | 4197 | fs_reclaim_release(sc.gfp_mask); |
132bb8cf YS |
4198 | |
4199 | trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed); | |
4200 | ||
a79311c1 | 4201 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 4202 | } |
179e9639 | 4203 | |
a5f5f91d | 4204 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 4205 | { |
d773ed6b | 4206 | int ret; |
179e9639 AM |
4207 | |
4208 | /* | |
a5f5f91d | 4209 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 4210 | * slab pages if we are over the defined limits. |
34aa1330 | 4211 | * |
9614634f CL |
4212 | * A small portion of unmapped file backed pages is needed for |
4213 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
4214 | * thrown out if the node is overallocated. So we do not reclaim |
4215 | * if less than a specified percentage of the node is used by | |
9614634f | 4216 | * unmapped file backed pages. |
179e9639 | 4217 | */ |
a5f5f91d | 4218 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
d42f3245 RG |
4219 | node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <= |
4220 | pgdat->min_slab_pages) | |
a5f5f91d | 4221 | return NODE_RECLAIM_FULL; |
179e9639 AM |
4222 | |
4223 | /* | |
d773ed6b | 4224 | * Do not scan if the allocation should not be delayed. |
179e9639 | 4225 | */ |
d0164adc | 4226 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 4227 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
4228 | |
4229 | /* | |
a5f5f91d | 4230 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
4231 | * have associated processors. This will favor the local processor |
4232 | * over remote processors and spread off node memory allocations | |
4233 | * as wide as possible. | |
4234 | */ | |
a5f5f91d MG |
4235 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
4236 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 4237 | |
a5f5f91d MG |
4238 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
4239 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 4240 | |
a5f5f91d MG |
4241 | ret = __node_reclaim(pgdat, gfp_mask, order); |
4242 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 4243 | |
24cf7251 MG |
4244 | if (!ret) |
4245 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
4246 | ||
d773ed6b | 4247 | return ret; |
179e9639 | 4248 | } |
9eeff239 | 4249 | #endif |
894bc310 | 4250 | |
89e004ea | 4251 | /** |
64e3d12f KHY |
4252 | * check_move_unevictable_pages - check pages for evictability and move to |
4253 | * appropriate zone lru list | |
4254 | * @pvec: pagevec with lru pages to check | |
89e004ea | 4255 | * |
64e3d12f KHY |
4256 | * Checks pages for evictability, if an evictable page is in the unevictable |
4257 | * lru list, moves it to the appropriate evictable lru list. This function | |
4258 | * should be only used for lru pages. | |
89e004ea | 4259 | */ |
64e3d12f | 4260 | void check_move_unevictable_pages(struct pagevec *pvec) |
89e004ea | 4261 | { |
6168d0da | 4262 | struct lruvec *lruvec = NULL; |
24513264 HD |
4263 | int pgscanned = 0; |
4264 | int pgrescued = 0; | |
4265 | int i; | |
89e004ea | 4266 | |
64e3d12f KHY |
4267 | for (i = 0; i < pvec->nr; i++) { |
4268 | struct page *page = pvec->pages[i]; | |
8d8869ca HD |
4269 | int nr_pages; |
4270 | ||
4271 | if (PageTransTail(page)) | |
4272 | continue; | |
4273 | ||
4274 | nr_pages = thp_nr_pages(page); | |
4275 | pgscanned += nr_pages; | |
89e004ea | 4276 | |
d25b5bd8 AS |
4277 | /* block memcg migration during page moving between lru */ |
4278 | if (!TestClearPageLRU(page)) | |
4279 | continue; | |
4280 | ||
2a5e4e34 | 4281 | lruvec = relock_page_lruvec_irq(page, lruvec); |
d25b5bd8 | 4282 | if (page_evictable(page) && PageUnevictable(page)) { |
46ae6b2c | 4283 | del_page_from_lru_list(page, lruvec); |
24513264 | 4284 | ClearPageUnevictable(page); |
3a9c9788 | 4285 | add_page_to_lru_list(page, lruvec); |
8d8869ca | 4286 | pgrescued += nr_pages; |
89e004ea | 4287 | } |
d25b5bd8 | 4288 | SetPageLRU(page); |
24513264 | 4289 | } |
89e004ea | 4290 | |
6168d0da | 4291 | if (lruvec) { |
24513264 HD |
4292 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
4293 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
6168d0da | 4294 | unlock_page_lruvec_irq(lruvec); |
d25b5bd8 AS |
4295 | } else if (pgscanned) { |
4296 | count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
89e004ea | 4297 | } |
89e004ea | 4298 | } |
64e3d12f | 4299 | EXPORT_SYMBOL_GPL(check_move_unevictable_pages); |