]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
1da177e4 LT |
2 | /* |
3 | * linux/mm/vmscan.c | |
4 | * | |
5 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds | |
6 | * | |
7 | * Swap reorganised 29.12.95, Stephen Tweedie. | |
8 | * kswapd added: 7.1.96 sct | |
9 | * Removed kswapd_ctl limits, and swap out as many pages as needed | |
10 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. | |
11 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). | |
12 | * Multiqueue VM started 5.8.00, Rik van Riel. | |
13 | */ | |
14 | ||
b1de0d13 MH |
15 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
16 | ||
1da177e4 | 17 | #include <linux/mm.h> |
5b3cc15a | 18 | #include <linux/sched/mm.h> |
1da177e4 | 19 | #include <linux/module.h> |
5a0e3ad6 | 20 | #include <linux/gfp.h> |
1da177e4 LT |
21 | #include <linux/kernel_stat.h> |
22 | #include <linux/swap.h> | |
23 | #include <linux/pagemap.h> | |
24 | #include <linux/init.h> | |
25 | #include <linux/highmem.h> | |
70ddf637 | 26 | #include <linux/vmpressure.h> |
e129b5c2 | 27 | #include <linux/vmstat.h> |
1da177e4 LT |
28 | #include <linux/file.h> |
29 | #include <linux/writeback.h> | |
30 | #include <linux/blkdev.h> | |
31 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
32 | buffer_heads_over_limit */ | |
33 | #include <linux/mm_inline.h> | |
1da177e4 LT |
34 | #include <linux/backing-dev.h> |
35 | #include <linux/rmap.h> | |
36 | #include <linux/topology.h> | |
37 | #include <linux/cpu.h> | |
38 | #include <linux/cpuset.h> | |
3e7d3449 | 39 | #include <linux/compaction.h> |
1da177e4 LT |
40 | #include <linux/notifier.h> |
41 | #include <linux/rwsem.h> | |
248a0301 | 42 | #include <linux/delay.h> |
3218ae14 | 43 | #include <linux/kthread.h> |
7dfb7103 | 44 | #include <linux/freezer.h> |
66e1707b | 45 | #include <linux/memcontrol.h> |
873b4771 | 46 | #include <linux/delayacct.h> |
af936a16 | 47 | #include <linux/sysctl.h> |
929bea7c | 48 | #include <linux/oom.h> |
64e3d12f | 49 | #include <linux/pagevec.h> |
268bb0ce | 50 | #include <linux/prefetch.h> |
b1de0d13 | 51 | #include <linux/printk.h> |
f9fe48be | 52 | #include <linux/dax.h> |
eb414681 | 53 | #include <linux/psi.h> |
1da177e4 LT |
54 | |
55 | #include <asm/tlbflush.h> | |
56 | #include <asm/div64.h> | |
57 | ||
58 | #include <linux/swapops.h> | |
117aad1e | 59 | #include <linux/balloon_compaction.h> |
1da177e4 | 60 | |
0f8053a5 NP |
61 | #include "internal.h" |
62 | ||
33906bc5 MG |
63 | #define CREATE_TRACE_POINTS |
64 | #include <trace/events/vmscan.h> | |
65 | ||
1da177e4 | 66 | struct scan_control { |
22fba335 KM |
67 | /* How many pages shrink_list() should reclaim */ |
68 | unsigned long nr_to_reclaim; | |
69 | ||
ee814fe2 JW |
70 | /* |
71 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
72 | * are scanned. | |
73 | */ | |
74 | nodemask_t *nodemask; | |
9e3b2f8c | 75 | |
f16015fb JW |
76 | /* |
77 | * The memory cgroup that hit its limit and as a result is the | |
78 | * primary target of this reclaim invocation. | |
79 | */ | |
80 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 81 | |
1276ad68 | 82 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
ee814fe2 JW |
83 | unsigned int may_writepage:1; |
84 | ||
85 | /* Can mapped pages be reclaimed? */ | |
86 | unsigned int may_unmap:1; | |
87 | ||
88 | /* Can pages be swapped as part of reclaim? */ | |
89 | unsigned int may_swap:1; | |
90 | ||
d6622f63 YX |
91 | /* |
92 | * Cgroups are not reclaimed below their configured memory.low, | |
93 | * unless we threaten to OOM. If any cgroups are skipped due to | |
94 | * memory.low and nothing was reclaimed, go back for memory.low. | |
95 | */ | |
96 | unsigned int memcg_low_reclaim:1; | |
97 | unsigned int memcg_low_skipped:1; | |
241994ed | 98 | |
ee814fe2 JW |
99 | unsigned int hibernation_mode:1; |
100 | ||
101 | /* One of the zones is ready for compaction */ | |
102 | unsigned int compaction_ready:1; | |
103 | ||
bb451fdf GT |
104 | /* Allocation order */ |
105 | s8 order; | |
106 | ||
107 | /* Scan (total_size >> priority) pages at once */ | |
108 | s8 priority; | |
109 | ||
110 | /* The highest zone to isolate pages for reclaim from */ | |
111 | s8 reclaim_idx; | |
112 | ||
113 | /* This context's GFP mask */ | |
114 | gfp_t gfp_mask; | |
115 | ||
ee814fe2 JW |
116 | /* Incremented by the number of inactive pages that were scanned */ |
117 | unsigned long nr_scanned; | |
118 | ||
119 | /* Number of pages freed so far during a call to shrink_zones() */ | |
120 | unsigned long nr_reclaimed; | |
d108c772 AR |
121 | |
122 | struct { | |
123 | unsigned int dirty; | |
124 | unsigned int unqueued_dirty; | |
125 | unsigned int congested; | |
126 | unsigned int writeback; | |
127 | unsigned int immediate; | |
128 | unsigned int file_taken; | |
129 | unsigned int taken; | |
130 | } nr; | |
e5ca8071 YS |
131 | |
132 | /* for recording the reclaimed slab by now */ | |
133 | struct reclaim_state reclaim_state; | |
1da177e4 LT |
134 | }; |
135 | ||
1da177e4 LT |
136 | #ifdef ARCH_HAS_PREFETCH |
137 | #define prefetch_prev_lru_page(_page, _base, _field) \ | |
138 | do { \ | |
139 | if ((_page)->lru.prev != _base) { \ | |
140 | struct page *prev; \ | |
141 | \ | |
142 | prev = lru_to_page(&(_page->lru)); \ | |
143 | prefetch(&prev->_field); \ | |
144 | } \ | |
145 | } while (0) | |
146 | #else | |
147 | #define prefetch_prev_lru_page(_page, _base, _field) do { } while (0) | |
148 | #endif | |
149 | ||
150 | #ifdef ARCH_HAS_PREFETCHW | |
151 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
152 | do { \ | |
153 | if ((_page)->lru.prev != _base) { \ | |
154 | struct page *prev; \ | |
155 | \ | |
156 | prev = lru_to_page(&(_page->lru)); \ | |
157 | prefetchw(&prev->_field); \ | |
158 | } \ | |
159 | } while (0) | |
160 | #else | |
161 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
162 | #endif | |
163 | ||
164 | /* | |
165 | * From 0 .. 100. Higher means more swappy. | |
166 | */ | |
167 | int vm_swappiness = 60; | |
d0480be4 WSH |
168 | /* |
169 | * The total number of pages which are beyond the high watermark within all | |
170 | * zones. | |
171 | */ | |
172 | unsigned long vm_total_pages; | |
1da177e4 | 173 | |
0a432dcb YS |
174 | static void set_task_reclaim_state(struct task_struct *task, |
175 | struct reclaim_state *rs) | |
176 | { | |
177 | /* Check for an overwrite */ | |
178 | WARN_ON_ONCE(rs && task->reclaim_state); | |
179 | ||
180 | /* Check for the nulling of an already-nulled member */ | |
181 | WARN_ON_ONCE(!rs && !task->reclaim_state); | |
182 | ||
183 | task->reclaim_state = rs; | |
184 | } | |
185 | ||
1da177e4 LT |
186 | static LIST_HEAD(shrinker_list); |
187 | static DECLARE_RWSEM(shrinker_rwsem); | |
188 | ||
0a432dcb | 189 | #ifdef CONFIG_MEMCG |
7e010df5 KT |
190 | /* |
191 | * We allow subsystems to populate their shrinker-related | |
192 | * LRU lists before register_shrinker_prepared() is called | |
193 | * for the shrinker, since we don't want to impose | |
194 | * restrictions on their internal registration order. | |
195 | * In this case shrink_slab_memcg() may find corresponding | |
196 | * bit is set in the shrinkers map. | |
197 | * | |
198 | * This value is used by the function to detect registering | |
199 | * shrinkers and to skip do_shrink_slab() calls for them. | |
200 | */ | |
201 | #define SHRINKER_REGISTERING ((struct shrinker *)~0UL) | |
202 | ||
b4c2b231 KT |
203 | static DEFINE_IDR(shrinker_idr); |
204 | static int shrinker_nr_max; | |
205 | ||
206 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) | |
207 | { | |
208 | int id, ret = -ENOMEM; | |
209 | ||
210 | down_write(&shrinker_rwsem); | |
211 | /* This may call shrinker, so it must use down_read_trylock() */ | |
7e010df5 | 212 | id = idr_alloc(&shrinker_idr, SHRINKER_REGISTERING, 0, 0, GFP_KERNEL); |
b4c2b231 KT |
213 | if (id < 0) |
214 | goto unlock; | |
215 | ||
0a4465d3 KT |
216 | if (id >= shrinker_nr_max) { |
217 | if (memcg_expand_shrinker_maps(id)) { | |
218 | idr_remove(&shrinker_idr, id); | |
219 | goto unlock; | |
220 | } | |
221 | ||
b4c2b231 | 222 | shrinker_nr_max = id + 1; |
0a4465d3 | 223 | } |
b4c2b231 KT |
224 | shrinker->id = id; |
225 | ret = 0; | |
226 | unlock: | |
227 | up_write(&shrinker_rwsem); | |
228 | return ret; | |
229 | } | |
230 | ||
231 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
232 | { | |
233 | int id = shrinker->id; | |
234 | ||
235 | BUG_ON(id < 0); | |
236 | ||
237 | down_write(&shrinker_rwsem); | |
238 | idr_remove(&shrinker_idr, id); | |
239 | up_write(&shrinker_rwsem); | |
240 | } | |
b4c2b231 | 241 | |
b5ead35e | 242 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 243 | { |
b5ead35e | 244 | return sc->target_mem_cgroup; |
89b5fae5 | 245 | } |
97c9341f TH |
246 | |
247 | /** | |
b5ead35e | 248 | * writeback_throttling_sane - is the usual dirty throttling mechanism available? |
97c9341f TH |
249 | * @sc: scan_control in question |
250 | * | |
251 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
252 | * completely broken with the legacy memcg and direct stalling in | |
253 | * shrink_page_list() is used for throttling instead, which lacks all the | |
254 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
255 | * allocation and configurability. | |
256 | * | |
257 | * This function tests whether the vmscan currently in progress can assume | |
258 | * that the normal dirty throttling mechanism is operational. | |
259 | */ | |
b5ead35e | 260 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f | 261 | { |
b5ead35e | 262 | if (!cgroup_reclaim(sc)) |
97c9341f TH |
263 | return true; |
264 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 265 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
266 | return true; |
267 | #endif | |
268 | return false; | |
269 | } | |
91a45470 | 270 | #else |
0a432dcb YS |
271 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) |
272 | { | |
273 | return 0; | |
274 | } | |
275 | ||
276 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
277 | { | |
278 | } | |
279 | ||
b5ead35e | 280 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 281 | { |
b5ead35e | 282 | return false; |
89b5fae5 | 283 | } |
97c9341f | 284 | |
b5ead35e | 285 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f TH |
286 | { |
287 | return true; | |
288 | } | |
91a45470 KH |
289 | #endif |
290 | ||
5a1c84b4 MG |
291 | /* |
292 | * This misses isolated pages which are not accounted for to save counters. | |
293 | * As the data only determines if reclaim or compaction continues, it is | |
294 | * not expected that isolated pages will be a dominating factor. | |
295 | */ | |
296 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
297 | { | |
298 | unsigned long nr; | |
299 | ||
300 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
301 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
302 | if (get_nr_swap_pages() > 0) | |
303 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + | |
304 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
305 | ||
306 | return nr; | |
307 | } | |
308 | ||
fd538803 MH |
309 | /** |
310 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
311 | * @lruvec: lru vector | |
312 | * @lru: lru to use | |
313 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) | |
314 | */ | |
315 | unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, int zone_idx) | |
c9f299d9 | 316 | { |
de3b0150 | 317 | unsigned long size = 0; |
fd538803 MH |
318 | int zid; |
319 | ||
de3b0150 | 320 | for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) { |
fd538803 | 321 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; |
c9f299d9 | 322 | |
fd538803 MH |
323 | if (!managed_zone(zone)) |
324 | continue; | |
325 | ||
326 | if (!mem_cgroup_disabled()) | |
de3b0150 | 327 | size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid); |
fd538803 | 328 | else |
de3b0150 | 329 | size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru); |
fd538803 | 330 | } |
de3b0150 | 331 | return size; |
b4536f0c MH |
332 | } |
333 | ||
1da177e4 | 334 | /* |
1d3d4437 | 335 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 336 | */ |
8e04944f | 337 | int prealloc_shrinker(struct shrinker *shrinker) |
1da177e4 | 338 | { |
b9726c26 | 339 | unsigned int size = sizeof(*shrinker->nr_deferred); |
1d3d4437 | 340 | |
1d3d4437 GC |
341 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
342 | size *= nr_node_ids; | |
343 | ||
344 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
345 | if (!shrinker->nr_deferred) | |
346 | return -ENOMEM; | |
b4c2b231 KT |
347 | |
348 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) { | |
349 | if (prealloc_memcg_shrinker(shrinker)) | |
350 | goto free_deferred; | |
351 | } | |
352 | ||
8e04944f | 353 | return 0; |
b4c2b231 KT |
354 | |
355 | free_deferred: | |
356 | kfree(shrinker->nr_deferred); | |
357 | shrinker->nr_deferred = NULL; | |
358 | return -ENOMEM; | |
8e04944f TH |
359 | } |
360 | ||
361 | void free_prealloced_shrinker(struct shrinker *shrinker) | |
362 | { | |
b4c2b231 KT |
363 | if (!shrinker->nr_deferred) |
364 | return; | |
365 | ||
366 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) | |
367 | unregister_memcg_shrinker(shrinker); | |
368 | ||
8e04944f TH |
369 | kfree(shrinker->nr_deferred); |
370 | shrinker->nr_deferred = NULL; | |
371 | } | |
1d3d4437 | 372 | |
8e04944f TH |
373 | void register_shrinker_prepared(struct shrinker *shrinker) |
374 | { | |
8e1f936b RR |
375 | down_write(&shrinker_rwsem); |
376 | list_add_tail(&shrinker->list, &shrinker_list); | |
7e010df5 | 377 | #ifdef CONFIG_MEMCG_KMEM |
8df4a44c KT |
378 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) |
379 | idr_replace(&shrinker_idr, shrinker, shrinker->id); | |
7e010df5 | 380 | #endif |
8e1f936b | 381 | up_write(&shrinker_rwsem); |
8e04944f TH |
382 | } |
383 | ||
384 | int register_shrinker(struct shrinker *shrinker) | |
385 | { | |
386 | int err = prealloc_shrinker(shrinker); | |
387 | ||
388 | if (err) | |
389 | return err; | |
390 | register_shrinker_prepared(shrinker); | |
1d3d4437 | 391 | return 0; |
1da177e4 | 392 | } |
8e1f936b | 393 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
394 | |
395 | /* | |
396 | * Remove one | |
397 | */ | |
8e1f936b | 398 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 | 399 | { |
bb422a73 TH |
400 | if (!shrinker->nr_deferred) |
401 | return; | |
b4c2b231 KT |
402 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) |
403 | unregister_memcg_shrinker(shrinker); | |
1da177e4 LT |
404 | down_write(&shrinker_rwsem); |
405 | list_del(&shrinker->list); | |
406 | up_write(&shrinker_rwsem); | |
ae393321 | 407 | kfree(shrinker->nr_deferred); |
bb422a73 | 408 | shrinker->nr_deferred = NULL; |
1da177e4 | 409 | } |
8e1f936b | 410 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 LT |
411 | |
412 | #define SHRINK_BATCH 128 | |
1d3d4437 | 413 | |
cb731d6c | 414 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
9092c71b | 415 | struct shrinker *shrinker, int priority) |
1d3d4437 GC |
416 | { |
417 | unsigned long freed = 0; | |
418 | unsigned long long delta; | |
419 | long total_scan; | |
d5bc5fd3 | 420 | long freeable; |
1d3d4437 GC |
421 | long nr; |
422 | long new_nr; | |
423 | int nid = shrinkctl->nid; | |
424 | long batch_size = shrinker->batch ? shrinker->batch | |
425 | : SHRINK_BATCH; | |
5f33a080 | 426 | long scanned = 0, next_deferred; |
1d3d4437 | 427 | |
ac7fb3ad KT |
428 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) |
429 | nid = 0; | |
430 | ||
d5bc5fd3 | 431 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
9b996468 KT |
432 | if (freeable == 0 || freeable == SHRINK_EMPTY) |
433 | return freeable; | |
1d3d4437 GC |
434 | |
435 | /* | |
436 | * copy the current shrinker scan count into a local variable | |
437 | * and zero it so that other concurrent shrinker invocations | |
438 | * don't also do this scanning work. | |
439 | */ | |
440 | nr = atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
441 | ||
442 | total_scan = nr; | |
4b85afbd JW |
443 | if (shrinker->seeks) { |
444 | delta = freeable >> priority; | |
445 | delta *= 4; | |
446 | do_div(delta, shrinker->seeks); | |
447 | } else { | |
448 | /* | |
449 | * These objects don't require any IO to create. Trim | |
450 | * them aggressively under memory pressure to keep | |
451 | * them from causing refetches in the IO caches. | |
452 | */ | |
453 | delta = freeable / 2; | |
454 | } | |
172b06c3 | 455 | |
1d3d4437 GC |
456 | total_scan += delta; |
457 | if (total_scan < 0) { | |
d75f773c | 458 | pr_err("shrink_slab: %pS negative objects to delete nr=%ld\n", |
a0b02131 | 459 | shrinker->scan_objects, total_scan); |
d5bc5fd3 | 460 | total_scan = freeable; |
5f33a080 SL |
461 | next_deferred = nr; |
462 | } else | |
463 | next_deferred = total_scan; | |
1d3d4437 GC |
464 | |
465 | /* | |
466 | * We need to avoid excessive windup on filesystem shrinkers | |
467 | * due to large numbers of GFP_NOFS allocations causing the | |
468 | * shrinkers to return -1 all the time. This results in a large | |
469 | * nr being built up so when a shrink that can do some work | |
470 | * comes along it empties the entire cache due to nr >>> | |
d5bc5fd3 | 471 | * freeable. This is bad for sustaining a working set in |
1d3d4437 GC |
472 | * memory. |
473 | * | |
474 | * Hence only allow the shrinker to scan the entire cache when | |
475 | * a large delta change is calculated directly. | |
476 | */ | |
d5bc5fd3 VD |
477 | if (delta < freeable / 4) |
478 | total_scan = min(total_scan, freeable / 2); | |
1d3d4437 GC |
479 | |
480 | /* | |
481 | * Avoid risking looping forever due to too large nr value: | |
482 | * never try to free more than twice the estimate number of | |
483 | * freeable entries. | |
484 | */ | |
d5bc5fd3 VD |
485 | if (total_scan > freeable * 2) |
486 | total_scan = freeable * 2; | |
1d3d4437 GC |
487 | |
488 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
9092c71b | 489 | freeable, delta, total_scan, priority); |
1d3d4437 | 490 | |
0b1fb40a VD |
491 | /* |
492 | * Normally, we should not scan less than batch_size objects in one | |
493 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
494 | * than batch_size objects in total and we are really tight on memory, | |
495 | * we will try to reclaim all available objects, otherwise we can end | |
496 | * up failing allocations although there are plenty of reclaimable | |
497 | * objects spread over several slabs with usage less than the | |
498 | * batch_size. | |
499 | * | |
500 | * We detect the "tight on memory" situations by looking at the total | |
501 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 502 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
503 | * scanning at high prio and therefore should try to reclaim as much as |
504 | * possible. | |
505 | */ | |
506 | while (total_scan >= batch_size || | |
d5bc5fd3 | 507 | total_scan >= freeable) { |
a0b02131 | 508 | unsigned long ret; |
0b1fb40a | 509 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 510 | |
0b1fb40a | 511 | shrinkctl->nr_to_scan = nr_to_scan; |
d460acb5 | 512 | shrinkctl->nr_scanned = nr_to_scan; |
a0b02131 DC |
513 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
514 | if (ret == SHRINK_STOP) | |
515 | break; | |
516 | freed += ret; | |
1d3d4437 | 517 | |
d460acb5 CW |
518 | count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); |
519 | total_scan -= shrinkctl->nr_scanned; | |
520 | scanned += shrinkctl->nr_scanned; | |
1d3d4437 GC |
521 | |
522 | cond_resched(); | |
523 | } | |
524 | ||
5f33a080 SL |
525 | if (next_deferred >= scanned) |
526 | next_deferred -= scanned; | |
527 | else | |
528 | next_deferred = 0; | |
1d3d4437 GC |
529 | /* |
530 | * move the unused scan count back into the shrinker in a | |
531 | * manner that handles concurrent updates. If we exhausted the | |
532 | * scan, there is no need to do an update. | |
533 | */ | |
5f33a080 SL |
534 | if (next_deferred > 0) |
535 | new_nr = atomic_long_add_return(next_deferred, | |
1d3d4437 GC |
536 | &shrinker->nr_deferred[nid]); |
537 | else | |
538 | new_nr = atomic_long_read(&shrinker->nr_deferred[nid]); | |
539 | ||
df9024a8 | 540 | trace_mm_shrink_slab_end(shrinker, nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 541 | return freed; |
1495f230 YH |
542 | } |
543 | ||
0a432dcb | 544 | #ifdef CONFIG_MEMCG |
b0dedc49 KT |
545 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
546 | struct mem_cgroup *memcg, int priority) | |
547 | { | |
548 | struct memcg_shrinker_map *map; | |
b8e57efa KT |
549 | unsigned long ret, freed = 0; |
550 | int i; | |
b0dedc49 | 551 | |
0a432dcb | 552 | if (!mem_cgroup_online(memcg)) |
b0dedc49 KT |
553 | return 0; |
554 | ||
555 | if (!down_read_trylock(&shrinker_rwsem)) | |
556 | return 0; | |
557 | ||
558 | map = rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_map, | |
559 | true); | |
560 | if (unlikely(!map)) | |
561 | goto unlock; | |
562 | ||
563 | for_each_set_bit(i, map->map, shrinker_nr_max) { | |
564 | struct shrink_control sc = { | |
565 | .gfp_mask = gfp_mask, | |
566 | .nid = nid, | |
567 | .memcg = memcg, | |
568 | }; | |
569 | struct shrinker *shrinker; | |
570 | ||
571 | shrinker = idr_find(&shrinker_idr, i); | |
7e010df5 KT |
572 | if (unlikely(!shrinker || shrinker == SHRINKER_REGISTERING)) { |
573 | if (!shrinker) | |
574 | clear_bit(i, map->map); | |
b0dedc49 KT |
575 | continue; |
576 | } | |
577 | ||
0a432dcb YS |
578 | /* Call non-slab shrinkers even though kmem is disabled */ |
579 | if (!memcg_kmem_enabled() && | |
580 | !(shrinker->flags & SHRINKER_NONSLAB)) | |
581 | continue; | |
582 | ||
b0dedc49 | 583 | ret = do_shrink_slab(&sc, shrinker, priority); |
f90280d6 KT |
584 | if (ret == SHRINK_EMPTY) { |
585 | clear_bit(i, map->map); | |
586 | /* | |
587 | * After the shrinker reported that it had no objects to | |
588 | * free, but before we cleared the corresponding bit in | |
589 | * the memcg shrinker map, a new object might have been | |
590 | * added. To make sure, we have the bit set in this | |
591 | * case, we invoke the shrinker one more time and reset | |
592 | * the bit if it reports that it is not empty anymore. | |
593 | * The memory barrier here pairs with the barrier in | |
594 | * memcg_set_shrinker_bit(): | |
595 | * | |
596 | * list_lru_add() shrink_slab_memcg() | |
597 | * list_add_tail() clear_bit() | |
598 | * <MB> <MB> | |
599 | * set_bit() do_shrink_slab() | |
600 | */ | |
601 | smp_mb__after_atomic(); | |
602 | ret = do_shrink_slab(&sc, shrinker, priority); | |
603 | if (ret == SHRINK_EMPTY) | |
604 | ret = 0; | |
605 | else | |
606 | memcg_set_shrinker_bit(memcg, nid, i); | |
607 | } | |
b0dedc49 KT |
608 | freed += ret; |
609 | ||
610 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
611 | freed = freed ? : 1; | |
612 | break; | |
613 | } | |
614 | } | |
615 | unlock: | |
616 | up_read(&shrinker_rwsem); | |
617 | return freed; | |
618 | } | |
0a432dcb | 619 | #else /* CONFIG_MEMCG */ |
b0dedc49 KT |
620 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
621 | struct mem_cgroup *memcg, int priority) | |
622 | { | |
623 | return 0; | |
624 | } | |
0a432dcb | 625 | #endif /* CONFIG_MEMCG */ |
b0dedc49 | 626 | |
6b4f7799 | 627 | /** |
cb731d6c | 628 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
629 | * @gfp_mask: allocation context |
630 | * @nid: node whose slab caches to target | |
cb731d6c | 631 | * @memcg: memory cgroup whose slab caches to target |
9092c71b | 632 | * @priority: the reclaim priority |
1da177e4 | 633 | * |
6b4f7799 | 634 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 635 | * |
6b4f7799 JW |
636 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
637 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 638 | * |
aeed1d32 VD |
639 | * @memcg specifies the memory cgroup to target. Unaware shrinkers |
640 | * are called only if it is the root cgroup. | |
cb731d6c | 641 | * |
9092c71b JB |
642 | * @priority is sc->priority, we take the number of objects and >> by priority |
643 | * in order to get the scan target. | |
b15e0905 | 644 | * |
6b4f7799 | 645 | * Returns the number of reclaimed slab objects. |
1da177e4 | 646 | */ |
cb731d6c VD |
647 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
648 | struct mem_cgroup *memcg, | |
9092c71b | 649 | int priority) |
1da177e4 | 650 | { |
b8e57efa | 651 | unsigned long ret, freed = 0; |
1da177e4 LT |
652 | struct shrinker *shrinker; |
653 | ||
fa1e512f YS |
654 | /* |
655 | * The root memcg might be allocated even though memcg is disabled | |
656 | * via "cgroup_disable=memory" boot parameter. This could make | |
657 | * mem_cgroup_is_root() return false, then just run memcg slab | |
658 | * shrink, but skip global shrink. This may result in premature | |
659 | * oom. | |
660 | */ | |
661 | if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg)) | |
b0dedc49 | 662 | return shrink_slab_memcg(gfp_mask, nid, memcg, priority); |
cb731d6c | 663 | |
e830c63a | 664 | if (!down_read_trylock(&shrinker_rwsem)) |
f06590bd | 665 | goto out; |
1da177e4 LT |
666 | |
667 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
668 | struct shrink_control sc = { |
669 | .gfp_mask = gfp_mask, | |
670 | .nid = nid, | |
cb731d6c | 671 | .memcg = memcg, |
6b4f7799 | 672 | }; |
ec97097b | 673 | |
9b996468 KT |
674 | ret = do_shrink_slab(&sc, shrinker, priority); |
675 | if (ret == SHRINK_EMPTY) | |
676 | ret = 0; | |
677 | freed += ret; | |
e496612c MK |
678 | /* |
679 | * Bail out if someone want to register a new shrinker to | |
680 | * prevent the regsitration from being stalled for long periods | |
681 | * by parallel ongoing shrinking. | |
682 | */ | |
683 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
684 | freed = freed ? : 1; | |
685 | break; | |
686 | } | |
1da177e4 | 687 | } |
6b4f7799 | 688 | |
1da177e4 | 689 | up_read(&shrinker_rwsem); |
f06590bd MK |
690 | out: |
691 | cond_resched(); | |
24f7c6b9 | 692 | return freed; |
1da177e4 LT |
693 | } |
694 | ||
cb731d6c VD |
695 | void drop_slab_node(int nid) |
696 | { | |
697 | unsigned long freed; | |
698 | ||
699 | do { | |
700 | struct mem_cgroup *memcg = NULL; | |
701 | ||
702 | freed = 0; | |
aeed1d32 | 703 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
cb731d6c | 704 | do { |
9092c71b | 705 | freed += shrink_slab(GFP_KERNEL, nid, memcg, 0); |
cb731d6c VD |
706 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); |
707 | } while (freed > 10); | |
708 | } | |
709 | ||
710 | void drop_slab(void) | |
711 | { | |
712 | int nid; | |
713 | ||
714 | for_each_online_node(nid) | |
715 | drop_slab_node(nid); | |
716 | } | |
717 | ||
1da177e4 LT |
718 | static inline int is_page_cache_freeable(struct page *page) |
719 | { | |
ceddc3a5 JW |
720 | /* |
721 | * A freeable page cache page is referenced only by the caller | |
67891fff MW |
722 | * that isolated the page, the page cache and optional buffer |
723 | * heads at page->private. | |
ceddc3a5 | 724 | */ |
67891fff | 725 | int page_cache_pins = PageTransHuge(page) && PageSwapCache(page) ? |
bd4c82c2 | 726 | HPAGE_PMD_NR : 1; |
67891fff | 727 | return page_count(page) - page_has_private(page) == 1 + page_cache_pins; |
1da177e4 LT |
728 | } |
729 | ||
cb16556d | 730 | static int may_write_to_inode(struct inode *inode) |
1da177e4 | 731 | { |
930d9152 | 732 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 733 | return 1; |
703c2708 | 734 | if (!inode_write_congested(inode)) |
1da177e4 | 735 | return 1; |
703c2708 | 736 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
737 | return 1; |
738 | return 0; | |
739 | } | |
740 | ||
741 | /* | |
742 | * We detected a synchronous write error writing a page out. Probably | |
743 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
744 | * fsync(), msync() or close(). | |
745 | * | |
746 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
747 | * prevents it from being freed up. But we have a ref on the page and once | |
748 | * that page is locked, the mapping is pinned. | |
749 | * | |
750 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
751 | * __GFP_FS. | |
752 | */ | |
753 | static void handle_write_error(struct address_space *mapping, | |
754 | struct page *page, int error) | |
755 | { | |
7eaceacc | 756 | lock_page(page); |
3e9f45bd GC |
757 | if (page_mapping(page) == mapping) |
758 | mapping_set_error(mapping, error); | |
1da177e4 LT |
759 | unlock_page(page); |
760 | } | |
761 | ||
04e62a29 CL |
762 | /* possible outcome of pageout() */ |
763 | typedef enum { | |
764 | /* failed to write page out, page is locked */ | |
765 | PAGE_KEEP, | |
766 | /* move page to the active list, page is locked */ | |
767 | PAGE_ACTIVATE, | |
768 | /* page has been sent to the disk successfully, page is unlocked */ | |
769 | PAGE_SUCCESS, | |
770 | /* page is clean and locked */ | |
771 | PAGE_CLEAN, | |
772 | } pageout_t; | |
773 | ||
1da177e4 | 774 | /* |
1742f19f AM |
775 | * pageout is called by shrink_page_list() for each dirty page. |
776 | * Calls ->writepage(). | |
1da177e4 | 777 | */ |
cb16556d | 778 | static pageout_t pageout(struct page *page, struct address_space *mapping) |
1da177e4 LT |
779 | { |
780 | /* | |
781 | * If the page is dirty, only perform writeback if that write | |
782 | * will be non-blocking. To prevent this allocation from being | |
783 | * stalled by pagecache activity. But note that there may be | |
784 | * stalls if we need to run get_block(). We could test | |
785 | * PagePrivate for that. | |
786 | * | |
8174202b | 787 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
788 | * this page's queue, we can perform writeback even if that |
789 | * will block. | |
790 | * | |
791 | * If the page is swapcache, write it back even if that would | |
792 | * block, for some throttling. This happens by accident, because | |
793 | * swap_backing_dev_info is bust: it doesn't reflect the | |
794 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
795 | */ |
796 | if (!is_page_cache_freeable(page)) | |
797 | return PAGE_KEEP; | |
798 | if (!mapping) { | |
799 | /* | |
800 | * Some data journaling orphaned pages can have | |
801 | * page->mapping == NULL while being dirty with clean buffers. | |
802 | */ | |
266cf658 | 803 | if (page_has_private(page)) { |
1da177e4 LT |
804 | if (try_to_free_buffers(page)) { |
805 | ClearPageDirty(page); | |
b1de0d13 | 806 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
807 | return PAGE_CLEAN; |
808 | } | |
809 | } | |
810 | return PAGE_KEEP; | |
811 | } | |
812 | if (mapping->a_ops->writepage == NULL) | |
813 | return PAGE_ACTIVATE; | |
cb16556d | 814 | if (!may_write_to_inode(mapping->host)) |
1da177e4 LT |
815 | return PAGE_KEEP; |
816 | ||
817 | if (clear_page_dirty_for_io(page)) { | |
818 | int res; | |
819 | struct writeback_control wbc = { | |
820 | .sync_mode = WB_SYNC_NONE, | |
821 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
822 | .range_start = 0, |
823 | .range_end = LLONG_MAX, | |
1da177e4 LT |
824 | .for_reclaim = 1, |
825 | }; | |
826 | ||
827 | SetPageReclaim(page); | |
828 | res = mapping->a_ops->writepage(page, &wbc); | |
829 | if (res < 0) | |
830 | handle_write_error(mapping, page, res); | |
994fc28c | 831 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
832 | ClearPageReclaim(page); |
833 | return PAGE_ACTIVATE; | |
834 | } | |
c661b078 | 835 | |
1da177e4 LT |
836 | if (!PageWriteback(page)) { |
837 | /* synchronous write or broken a_ops? */ | |
838 | ClearPageReclaim(page); | |
839 | } | |
3aa23851 | 840 | trace_mm_vmscan_writepage(page); |
c4a25635 | 841 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
842 | return PAGE_SUCCESS; |
843 | } | |
844 | ||
845 | return PAGE_CLEAN; | |
846 | } | |
847 | ||
a649fd92 | 848 | /* |
e286781d NP |
849 | * Same as remove_mapping, but if the page is removed from the mapping, it |
850 | * gets returned with a refcount of 0. | |
a649fd92 | 851 | */ |
a528910e JW |
852 | static int __remove_mapping(struct address_space *mapping, struct page *page, |
853 | bool reclaimed) | |
49d2e9cc | 854 | { |
c4843a75 | 855 | unsigned long flags; |
bd4c82c2 | 856 | int refcount; |
c4843a75 | 857 | |
28e4d965 NP |
858 | BUG_ON(!PageLocked(page)); |
859 | BUG_ON(mapping != page_mapping(page)); | |
49d2e9cc | 860 | |
b93b0163 | 861 | xa_lock_irqsave(&mapping->i_pages, flags); |
49d2e9cc | 862 | /* |
0fd0e6b0 NP |
863 | * The non racy check for a busy page. |
864 | * | |
865 | * Must be careful with the order of the tests. When someone has | |
866 | * a ref to the page, it may be possible that they dirty it then | |
867 | * drop the reference. So if PageDirty is tested before page_count | |
868 | * here, then the following race may occur: | |
869 | * | |
870 | * get_user_pages(&page); | |
871 | * [user mapping goes away] | |
872 | * write_to(page); | |
873 | * !PageDirty(page) [good] | |
874 | * SetPageDirty(page); | |
875 | * put_page(page); | |
876 | * !page_count(page) [good, discard it] | |
877 | * | |
878 | * [oops, our write_to data is lost] | |
879 | * | |
880 | * Reversing the order of the tests ensures such a situation cannot | |
881 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 882 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
883 | * |
884 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
b93b0163 | 885 | * and thus under the i_pages lock, then this ordering is not required. |
49d2e9cc | 886 | */ |
906d278d | 887 | refcount = 1 + compound_nr(page); |
bd4c82c2 | 888 | if (!page_ref_freeze(page, refcount)) |
49d2e9cc | 889 | goto cannot_free; |
1c4c3b99 | 890 | /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */ |
e286781d | 891 | if (unlikely(PageDirty(page))) { |
bd4c82c2 | 892 | page_ref_unfreeze(page, refcount); |
49d2e9cc | 893 | goto cannot_free; |
e286781d | 894 | } |
49d2e9cc CL |
895 | |
896 | if (PageSwapCache(page)) { | |
897 | swp_entry_t swap = { .val = page_private(page) }; | |
0a31bc97 | 898 | mem_cgroup_swapout(page, swap); |
4e17ec25 | 899 | __delete_from_swap_cache(page, swap); |
b93b0163 | 900 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
75f6d6d2 | 901 | put_swap_page(page, swap); |
e286781d | 902 | } else { |
6072d13c | 903 | void (*freepage)(struct page *); |
a528910e | 904 | void *shadow = NULL; |
6072d13c LT |
905 | |
906 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
907 | /* |
908 | * Remember a shadow entry for reclaimed file cache in | |
909 | * order to detect refaults, thus thrashing, later on. | |
910 | * | |
911 | * But don't store shadows in an address space that is | |
912 | * already exiting. This is not just an optizimation, | |
913 | * inode reclaim needs to empty out the radix tree or | |
914 | * the nodes are lost. Don't plant shadows behind its | |
915 | * back. | |
f9fe48be RZ |
916 | * |
917 | * We also don't store shadows for DAX mappings because the | |
918 | * only page cache pages found in these are zero pages | |
919 | * covering holes, and because we don't want to mix DAX | |
920 | * exceptional entries and shadow exceptional entries in the | |
b93b0163 | 921 | * same address_space. |
a528910e JW |
922 | */ |
923 | if (reclaimed && page_is_file_cache(page) && | |
f9fe48be | 924 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
a7ca12f9 | 925 | shadow = workingset_eviction(page); |
62cccb8c | 926 | __delete_from_page_cache(page, shadow); |
b93b0163 | 927 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
6072d13c LT |
928 | |
929 | if (freepage != NULL) | |
930 | freepage(page); | |
49d2e9cc CL |
931 | } |
932 | ||
49d2e9cc CL |
933 | return 1; |
934 | ||
935 | cannot_free: | |
b93b0163 | 936 | xa_unlock_irqrestore(&mapping->i_pages, flags); |
49d2e9cc CL |
937 | return 0; |
938 | } | |
939 | ||
e286781d NP |
940 | /* |
941 | * Attempt to detach a locked page from its ->mapping. If it is dirty or if | |
942 | * someone else has a ref on the page, abort and return 0. If it was | |
943 | * successfully detached, return 1. Assumes the caller has a single ref on | |
944 | * this page. | |
945 | */ | |
946 | int remove_mapping(struct address_space *mapping, struct page *page) | |
947 | { | |
a528910e | 948 | if (__remove_mapping(mapping, page, false)) { |
e286781d NP |
949 | /* |
950 | * Unfreezing the refcount with 1 rather than 2 effectively | |
951 | * drops the pagecache ref for us without requiring another | |
952 | * atomic operation. | |
953 | */ | |
fe896d18 | 954 | page_ref_unfreeze(page, 1); |
e286781d NP |
955 | return 1; |
956 | } | |
957 | return 0; | |
958 | } | |
959 | ||
894bc310 LS |
960 | /** |
961 | * putback_lru_page - put previously isolated page onto appropriate LRU list | |
962 | * @page: page to be put back to appropriate lru list | |
963 | * | |
964 | * Add previously isolated @page to appropriate LRU list. | |
965 | * Page may still be unevictable for other reasons. | |
966 | * | |
967 | * lru_lock must not be held, interrupts must be enabled. | |
968 | */ | |
894bc310 LS |
969 | void putback_lru_page(struct page *page) |
970 | { | |
9c4e6b1a | 971 | lru_cache_add(page); |
894bc310 LS |
972 | put_page(page); /* drop ref from isolate */ |
973 | } | |
974 | ||
dfc8d636 JW |
975 | enum page_references { |
976 | PAGEREF_RECLAIM, | |
977 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 978 | PAGEREF_KEEP, |
dfc8d636 JW |
979 | PAGEREF_ACTIVATE, |
980 | }; | |
981 | ||
982 | static enum page_references page_check_references(struct page *page, | |
983 | struct scan_control *sc) | |
984 | { | |
64574746 | 985 | int referenced_ptes, referenced_page; |
dfc8d636 | 986 | unsigned long vm_flags; |
dfc8d636 | 987 | |
c3ac9a8a JW |
988 | referenced_ptes = page_referenced(page, 1, sc->target_mem_cgroup, |
989 | &vm_flags); | |
64574746 | 990 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 991 | |
dfc8d636 JW |
992 | /* |
993 | * Mlock lost the isolation race with us. Let try_to_unmap() | |
994 | * move the page to the unevictable list. | |
995 | */ | |
996 | if (vm_flags & VM_LOCKED) | |
997 | return PAGEREF_RECLAIM; | |
998 | ||
64574746 | 999 | if (referenced_ptes) { |
e4898273 | 1000 | if (PageSwapBacked(page)) |
64574746 JW |
1001 | return PAGEREF_ACTIVATE; |
1002 | /* | |
1003 | * All mapped pages start out with page table | |
1004 | * references from the instantiating fault, so we need | |
1005 | * to look twice if a mapped file page is used more | |
1006 | * than once. | |
1007 | * | |
1008 | * Mark it and spare it for another trip around the | |
1009 | * inactive list. Another page table reference will | |
1010 | * lead to its activation. | |
1011 | * | |
1012 | * Note: the mark is set for activated pages as well | |
1013 | * so that recently deactivated but used pages are | |
1014 | * quickly recovered. | |
1015 | */ | |
1016 | SetPageReferenced(page); | |
1017 | ||
34dbc67a | 1018 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
1019 | return PAGEREF_ACTIVATE; |
1020 | ||
c909e993 KK |
1021 | /* |
1022 | * Activate file-backed executable pages after first usage. | |
1023 | */ | |
1024 | if (vm_flags & VM_EXEC) | |
1025 | return PAGEREF_ACTIVATE; | |
1026 | ||
64574746 JW |
1027 | return PAGEREF_KEEP; |
1028 | } | |
dfc8d636 JW |
1029 | |
1030 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 1031 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
1032 | return PAGEREF_RECLAIM_CLEAN; |
1033 | ||
1034 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
1035 | } |
1036 | ||
e2be15f6 MG |
1037 | /* Check if a page is dirty or under writeback */ |
1038 | static void page_check_dirty_writeback(struct page *page, | |
1039 | bool *dirty, bool *writeback) | |
1040 | { | |
b4597226 MG |
1041 | struct address_space *mapping; |
1042 | ||
e2be15f6 MG |
1043 | /* |
1044 | * Anonymous pages are not handled by flushers and must be written | |
1045 | * from reclaim context. Do not stall reclaim based on them | |
1046 | */ | |
802a3a92 SL |
1047 | if (!page_is_file_cache(page) || |
1048 | (PageAnon(page) && !PageSwapBacked(page))) { | |
e2be15f6 MG |
1049 | *dirty = false; |
1050 | *writeback = false; | |
1051 | return; | |
1052 | } | |
1053 | ||
1054 | /* By default assume that the page flags are accurate */ | |
1055 | *dirty = PageDirty(page); | |
1056 | *writeback = PageWriteback(page); | |
b4597226 MG |
1057 | |
1058 | /* Verify dirty/writeback state if the filesystem supports it */ | |
1059 | if (!page_has_private(page)) | |
1060 | return; | |
1061 | ||
1062 | mapping = page_mapping(page); | |
1063 | if (mapping && mapping->a_ops->is_dirty_writeback) | |
1064 | mapping->a_ops->is_dirty_writeback(page, dirty, writeback); | |
e2be15f6 MG |
1065 | } |
1066 | ||
1da177e4 | 1067 | /* |
1742f19f | 1068 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 1069 | */ |
1742f19f | 1070 | static unsigned long shrink_page_list(struct list_head *page_list, |
599d0c95 | 1071 | struct pglist_data *pgdat, |
f84f6e2b | 1072 | struct scan_control *sc, |
02c6de8d | 1073 | enum ttu_flags ttu_flags, |
3c710c1a | 1074 | struct reclaim_stat *stat, |
8940b34a | 1075 | bool ignore_references) |
1da177e4 LT |
1076 | { |
1077 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 1078 | LIST_HEAD(free_pages); |
3c710c1a | 1079 | unsigned nr_reclaimed = 0; |
886cf190 | 1080 | unsigned 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; | |
1088 | int may_enter_fs; | |
8940b34a | 1089 | enum page_references references = PAGEREF_RECLAIM; |
e2be15f6 | 1090 | bool dirty, writeback; |
98879b3b | 1091 | unsigned int nr_pages; |
1da177e4 LT |
1092 | |
1093 | cond_resched(); | |
1094 | ||
1095 | page = lru_to_page(page_list); | |
1096 | list_del(&page->lru); | |
1097 | ||
529ae9aa | 1098 | if (!trylock_page(page)) |
1da177e4 LT |
1099 | goto keep; |
1100 | ||
309381fe | 1101 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 | 1102 | |
d8c6546b | 1103 | nr_pages = compound_nr(page); |
98879b3b YS |
1104 | |
1105 | /* Account the number of base pages even though THP */ | |
1106 | sc->nr_scanned += nr_pages; | |
80e43426 | 1107 | |
39b5f29a | 1108 | if (unlikely(!page_evictable(page))) |
ad6b6704 | 1109 | goto activate_locked; |
894bc310 | 1110 | |
a6dc60f8 | 1111 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
1112 | goto keep_locked; |
1113 | ||
c661b078 AW |
1114 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
1115 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
1116 | ||
e2be15f6 | 1117 | /* |
894befec | 1118 | * The number of dirty pages determines if a node is marked |
e2be15f6 MG |
1119 | * reclaim_congested which affects wait_iff_congested. kswapd |
1120 | * will stall and start writing pages if the tail of the LRU | |
1121 | * is all dirty unqueued pages. | |
1122 | */ | |
1123 | page_check_dirty_writeback(page, &dirty, &writeback); | |
1124 | if (dirty || writeback) | |
060f005f | 1125 | stat->nr_dirty++; |
e2be15f6 MG |
1126 | |
1127 | if (dirty && !writeback) | |
060f005f | 1128 | stat->nr_unqueued_dirty++; |
e2be15f6 | 1129 | |
d04e8acd MG |
1130 | /* |
1131 | * Treat this page as congested if the underlying BDI is or if | |
1132 | * pages are cycling through the LRU so quickly that the | |
1133 | * pages marked for immediate reclaim are making it to the | |
1134 | * end of the LRU a second time. | |
1135 | */ | |
e2be15f6 | 1136 | mapping = page_mapping(page); |
1da58ee2 | 1137 | if (((dirty || writeback) && mapping && |
703c2708 | 1138 | inode_write_congested(mapping->host)) || |
d04e8acd | 1139 | (writeback && PageReclaim(page))) |
060f005f | 1140 | stat->nr_congested++; |
e2be15f6 | 1141 | |
283aba9f MG |
1142 | /* |
1143 | * If a page at the tail of the LRU is under writeback, there | |
1144 | * are three cases to consider. | |
1145 | * | |
1146 | * 1) If reclaim is encountering an excessive number of pages | |
1147 | * under writeback and this page is both under writeback and | |
1148 | * PageReclaim then it indicates that pages are being queued | |
1149 | * for IO but are being recycled through the LRU before the | |
1150 | * IO can complete. Waiting on the page itself risks an | |
1151 | * indefinite stall if it is impossible to writeback the | |
1152 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
1153 | * note that the LRU is being scanned too quickly and the |
1154 | * caller can stall after page list has been processed. | |
283aba9f | 1155 | * |
97c9341f | 1156 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
1157 | * not marked for immediate reclaim, or the caller does not |
1158 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
1159 | * not to fs). In this case mark the page for immediate | |
97c9341f | 1160 | * reclaim and continue scanning. |
283aba9f | 1161 | * |
ecf5fc6e MH |
1162 | * Require may_enter_fs because we would wait on fs, which |
1163 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
1164 | * enter reclaim, and deadlock if it waits on a page for |
1165 | * which it is needed to do the write (loop masks off | |
1166 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1167 | * would probably show more reasons. | |
1168 | * | |
7fadc820 | 1169 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1170 | * PageReclaim. memcg does not have any dirty pages |
1171 | * throttling so we could easily OOM just because too many | |
1172 | * pages are in writeback and there is nothing else to | |
1173 | * reclaim. Wait for the writeback to complete. | |
c55e8d03 JW |
1174 | * |
1175 | * In cases 1) and 2) we activate the pages to get them out of | |
1176 | * the way while we continue scanning for clean pages on the | |
1177 | * inactive list and refilling from the active list. The | |
1178 | * observation here is that waiting for disk writes is more | |
1179 | * expensive than potentially causing reloads down the line. | |
1180 | * Since they're marked for immediate reclaim, they won't put | |
1181 | * memory pressure on the cache working set any longer than it | |
1182 | * takes to write them to disk. | |
283aba9f | 1183 | */ |
c661b078 | 1184 | if (PageWriteback(page)) { |
283aba9f MG |
1185 | /* Case 1 above */ |
1186 | if (current_is_kswapd() && | |
1187 | PageReclaim(page) && | |
599d0c95 | 1188 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
060f005f | 1189 | stat->nr_immediate++; |
c55e8d03 | 1190 | goto activate_locked; |
283aba9f MG |
1191 | |
1192 | /* Case 2 above */ | |
b5ead35e | 1193 | } else if (writeback_throttling_sane(sc) || |
ecf5fc6e | 1194 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1195 | /* |
1196 | * This is slightly racy - end_page_writeback() | |
1197 | * might have just cleared PageReclaim, then | |
1198 | * setting PageReclaim here end up interpreted | |
1199 | * as PageReadahead - but that does not matter | |
1200 | * enough to care. What we do want is for this | |
1201 | * page to have PageReclaim set next time memcg | |
1202 | * reclaim reaches the tests above, so it will | |
1203 | * then wait_on_page_writeback() to avoid OOM; | |
1204 | * and it's also appropriate in global reclaim. | |
1205 | */ | |
1206 | SetPageReclaim(page); | |
060f005f | 1207 | stat->nr_writeback++; |
c55e8d03 | 1208 | goto activate_locked; |
283aba9f MG |
1209 | |
1210 | /* Case 3 above */ | |
1211 | } else { | |
7fadc820 | 1212 | unlock_page(page); |
283aba9f | 1213 | wait_on_page_writeback(page); |
7fadc820 HD |
1214 | /* then go back and try same page again */ |
1215 | list_add_tail(&page->lru, page_list); | |
1216 | continue; | |
e62e384e | 1217 | } |
c661b078 | 1218 | } |
1da177e4 | 1219 | |
8940b34a | 1220 | if (!ignore_references) |
02c6de8d MK |
1221 | references = page_check_references(page, sc); |
1222 | ||
dfc8d636 JW |
1223 | switch (references) { |
1224 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1225 | goto activate_locked; |
64574746 | 1226 | case PAGEREF_KEEP: |
98879b3b | 1227 | stat->nr_ref_keep += nr_pages; |
64574746 | 1228 | goto keep_locked; |
dfc8d636 JW |
1229 | case PAGEREF_RECLAIM: |
1230 | case PAGEREF_RECLAIM_CLEAN: | |
1231 | ; /* try to reclaim the page below */ | |
1232 | } | |
1da177e4 | 1233 | |
1da177e4 LT |
1234 | /* |
1235 | * Anonymous process memory has backing store? | |
1236 | * Try to allocate it some swap space here. | |
802a3a92 | 1237 | * Lazyfree page could be freed directly |
1da177e4 | 1238 | */ |
bd4c82c2 HY |
1239 | if (PageAnon(page) && PageSwapBacked(page)) { |
1240 | if (!PageSwapCache(page)) { | |
1241 | if (!(sc->gfp_mask & __GFP_IO)) | |
1242 | goto keep_locked; | |
1243 | if (PageTransHuge(page)) { | |
1244 | /* cannot split THP, skip it */ | |
1245 | if (!can_split_huge_page(page, NULL)) | |
1246 | goto activate_locked; | |
1247 | /* | |
1248 | * Split pages without a PMD map right | |
1249 | * away. Chances are some or all of the | |
1250 | * tail pages can be freed without IO. | |
1251 | */ | |
1252 | if (!compound_mapcount(page) && | |
1253 | split_huge_page_to_list(page, | |
1254 | page_list)) | |
1255 | goto activate_locked; | |
1256 | } | |
1257 | if (!add_to_swap(page)) { | |
1258 | if (!PageTransHuge(page)) | |
98879b3b | 1259 | goto activate_locked_split; |
bd4c82c2 HY |
1260 | /* Fallback to swap normal pages */ |
1261 | if (split_huge_page_to_list(page, | |
1262 | page_list)) | |
1263 | goto activate_locked; | |
fe490cc0 HY |
1264 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1265 | count_vm_event(THP_SWPOUT_FALLBACK); | |
1266 | #endif | |
bd4c82c2 | 1267 | if (!add_to_swap(page)) |
98879b3b | 1268 | goto activate_locked_split; |
bd4c82c2 | 1269 | } |
0f074658 | 1270 | |
bd4c82c2 | 1271 | may_enter_fs = 1; |
1da177e4 | 1272 | |
bd4c82c2 HY |
1273 | /* Adding to swap updated mapping */ |
1274 | mapping = page_mapping(page); | |
1275 | } | |
7751b2da KS |
1276 | } else if (unlikely(PageTransHuge(page))) { |
1277 | /* Split file THP */ | |
1278 | if (split_huge_page_to_list(page, page_list)) | |
1279 | goto keep_locked; | |
e2be15f6 | 1280 | } |
1da177e4 | 1281 | |
98879b3b YS |
1282 | /* |
1283 | * THP may get split above, need minus tail pages and update | |
1284 | * nr_pages to avoid accounting tail pages twice. | |
1285 | * | |
1286 | * The tail pages that are added into swap cache successfully | |
1287 | * reach here. | |
1288 | */ | |
1289 | if ((nr_pages > 1) && !PageTransHuge(page)) { | |
1290 | sc->nr_scanned -= (nr_pages - 1); | |
1291 | nr_pages = 1; | |
1292 | } | |
1293 | ||
1da177e4 LT |
1294 | /* |
1295 | * The page is mapped into the page tables of one or more | |
1296 | * processes. Try to unmap it here. | |
1297 | */ | |
802a3a92 | 1298 | if (page_mapped(page)) { |
bd4c82c2 HY |
1299 | enum ttu_flags flags = ttu_flags | TTU_BATCH_FLUSH; |
1300 | ||
1301 | if (unlikely(PageTransHuge(page))) | |
1302 | flags |= TTU_SPLIT_HUGE_PMD; | |
1303 | if (!try_to_unmap(page, flags)) { | |
98879b3b | 1304 | stat->nr_unmap_fail += nr_pages; |
1da177e4 | 1305 | goto activate_locked; |
1da177e4 LT |
1306 | } |
1307 | } | |
1308 | ||
1309 | if (PageDirty(page)) { | |
ee72886d | 1310 | /* |
4eda4823 JW |
1311 | * Only kswapd can writeback filesystem pages |
1312 | * to avoid risk of stack overflow. But avoid | |
1313 | * injecting inefficient single-page IO into | |
1314 | * flusher writeback as much as possible: only | |
1315 | * write pages when we've encountered many | |
1316 | * dirty pages, and when we've already scanned | |
1317 | * the rest of the LRU for clean pages and see | |
1318 | * the same dirty pages again (PageReclaim). | |
ee72886d | 1319 | */ |
f84f6e2b | 1320 | if (page_is_file_cache(page) && |
4eda4823 JW |
1321 | (!current_is_kswapd() || !PageReclaim(page) || |
1322 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { | |
49ea7eb6 MG |
1323 | /* |
1324 | * Immediately reclaim when written back. | |
1325 | * Similar in principal to deactivate_page() | |
1326 | * except we already have the page isolated | |
1327 | * and know it's dirty | |
1328 | */ | |
c4a25635 | 1329 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
49ea7eb6 MG |
1330 | SetPageReclaim(page); |
1331 | ||
c55e8d03 | 1332 | goto activate_locked; |
ee72886d MG |
1333 | } |
1334 | ||
dfc8d636 | 1335 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1336 | goto keep_locked; |
4dd4b920 | 1337 | if (!may_enter_fs) |
1da177e4 | 1338 | goto keep_locked; |
52a8363e | 1339 | if (!sc->may_writepage) |
1da177e4 LT |
1340 | goto keep_locked; |
1341 | ||
d950c947 MG |
1342 | /* |
1343 | * Page is dirty. Flush the TLB if a writable entry | |
1344 | * potentially exists to avoid CPU writes after IO | |
1345 | * starts and then write it out here. | |
1346 | */ | |
1347 | try_to_unmap_flush_dirty(); | |
cb16556d | 1348 | switch (pageout(page, mapping)) { |
1da177e4 LT |
1349 | case PAGE_KEEP: |
1350 | goto keep_locked; | |
1351 | case PAGE_ACTIVATE: | |
1352 | goto activate_locked; | |
1353 | case PAGE_SUCCESS: | |
7d3579e8 | 1354 | if (PageWriteback(page)) |
41ac1999 | 1355 | goto keep; |
7d3579e8 | 1356 | if (PageDirty(page)) |
1da177e4 | 1357 | goto keep; |
7d3579e8 | 1358 | |
1da177e4 LT |
1359 | /* |
1360 | * A synchronous write - probably a ramdisk. Go | |
1361 | * ahead and try to reclaim the page. | |
1362 | */ | |
529ae9aa | 1363 | if (!trylock_page(page)) |
1da177e4 LT |
1364 | goto keep; |
1365 | if (PageDirty(page) || PageWriteback(page)) | |
1366 | goto keep_locked; | |
1367 | mapping = page_mapping(page); | |
1368 | case PAGE_CLEAN: | |
1369 | ; /* try to free the page below */ | |
1370 | } | |
1371 | } | |
1372 | ||
1373 | /* | |
1374 | * If the page has buffers, try to free the buffer mappings | |
1375 | * associated with this page. If we succeed we try to free | |
1376 | * the page as well. | |
1377 | * | |
1378 | * We do this even if the page is PageDirty(). | |
1379 | * try_to_release_page() does not perform I/O, but it is | |
1380 | * possible for a page to have PageDirty set, but it is actually | |
1381 | * clean (all its buffers are clean). This happens if the | |
1382 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1383 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1384 | * try_to_release_page() will discover that cleanness and will |
1385 | * drop the buffers and mark the page clean - it can be freed. | |
1386 | * | |
1387 | * Rarely, pages can have buffers and no ->mapping. These are | |
1388 | * the pages which were not successfully invalidated in | |
1389 | * truncate_complete_page(). We try to drop those buffers here | |
1390 | * and if that worked, and the page is no longer mapped into | |
1391 | * process address space (page_count == 1) it can be freed. | |
1392 | * Otherwise, leave the page on the LRU so it is swappable. | |
1393 | */ | |
266cf658 | 1394 | if (page_has_private(page)) { |
1da177e4 LT |
1395 | if (!try_to_release_page(page, sc->gfp_mask)) |
1396 | goto activate_locked; | |
e286781d NP |
1397 | if (!mapping && page_count(page) == 1) { |
1398 | unlock_page(page); | |
1399 | if (put_page_testzero(page)) | |
1400 | goto free_it; | |
1401 | else { | |
1402 | /* | |
1403 | * rare race with speculative reference. | |
1404 | * the speculative reference will free | |
1405 | * this page shortly, so we may | |
1406 | * increment nr_reclaimed here (and | |
1407 | * leave it off the LRU). | |
1408 | */ | |
1409 | nr_reclaimed++; | |
1410 | continue; | |
1411 | } | |
1412 | } | |
1da177e4 LT |
1413 | } |
1414 | ||
802a3a92 SL |
1415 | if (PageAnon(page) && !PageSwapBacked(page)) { |
1416 | /* follow __remove_mapping for reference */ | |
1417 | if (!page_ref_freeze(page, 1)) | |
1418 | goto keep_locked; | |
1419 | if (PageDirty(page)) { | |
1420 | page_ref_unfreeze(page, 1); | |
1421 | goto keep_locked; | |
1422 | } | |
1da177e4 | 1423 | |
802a3a92 | 1424 | count_vm_event(PGLAZYFREED); |
2262185c | 1425 | count_memcg_page_event(page, PGLAZYFREED); |
802a3a92 SL |
1426 | } else if (!mapping || !__remove_mapping(mapping, page, true)) |
1427 | goto keep_locked; | |
9a1ea439 HD |
1428 | |
1429 | unlock_page(page); | |
e286781d | 1430 | free_it: |
98879b3b YS |
1431 | /* |
1432 | * THP may get swapped out in a whole, need account | |
1433 | * all base pages. | |
1434 | */ | |
1435 | nr_reclaimed += nr_pages; | |
abe4c3b5 MG |
1436 | |
1437 | /* | |
1438 | * Is there need to periodically free_page_list? It would | |
1439 | * appear not as the counts should be low | |
1440 | */ | |
7ae88534 | 1441 | if (unlikely(PageTransHuge(page))) |
bd4c82c2 | 1442 | (*get_compound_page_dtor(page))(page); |
7ae88534 | 1443 | else |
bd4c82c2 | 1444 | list_add(&page->lru, &free_pages); |
1da177e4 LT |
1445 | continue; |
1446 | ||
98879b3b YS |
1447 | activate_locked_split: |
1448 | /* | |
1449 | * The tail pages that are failed to add into swap cache | |
1450 | * reach here. Fixup nr_scanned and nr_pages. | |
1451 | */ | |
1452 | if (nr_pages > 1) { | |
1453 | sc->nr_scanned -= (nr_pages - 1); | |
1454 | nr_pages = 1; | |
1455 | } | |
1da177e4 | 1456 | activate_locked: |
68a22394 | 1457 | /* Not a candidate for swapping, so reclaim swap space. */ |
ad6b6704 MK |
1458 | if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || |
1459 | PageMlocked(page))) | |
a2c43eed | 1460 | try_to_free_swap(page); |
309381fe | 1461 | VM_BUG_ON_PAGE(PageActive(page), page); |
ad6b6704 | 1462 | if (!PageMlocked(page)) { |
886cf190 | 1463 | int type = page_is_file_cache(page); |
ad6b6704 | 1464 | SetPageActive(page); |
98879b3b | 1465 | stat->nr_activate[type] += nr_pages; |
2262185c | 1466 | count_memcg_page_event(page, PGACTIVATE); |
ad6b6704 | 1467 | } |
1da177e4 LT |
1468 | keep_locked: |
1469 | unlock_page(page); | |
1470 | keep: | |
1471 | list_add(&page->lru, &ret_pages); | |
309381fe | 1472 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1473 | } |
abe4c3b5 | 1474 | |
98879b3b YS |
1475 | pgactivate = stat->nr_activate[0] + stat->nr_activate[1]; |
1476 | ||
747db954 | 1477 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1478 | try_to_unmap_flush(); |
2d4894b5 | 1479 | free_unref_page_list(&free_pages); |
abe4c3b5 | 1480 | |
1da177e4 | 1481 | list_splice(&ret_pages, page_list); |
886cf190 | 1482 | count_vm_events(PGACTIVATE, pgactivate); |
060f005f | 1483 | |
05ff5137 | 1484 | return nr_reclaimed; |
1da177e4 LT |
1485 | } |
1486 | ||
02c6de8d MK |
1487 | unsigned long reclaim_clean_pages_from_list(struct zone *zone, |
1488 | struct list_head *page_list) | |
1489 | { | |
1490 | struct scan_control sc = { | |
1491 | .gfp_mask = GFP_KERNEL, | |
1492 | .priority = DEF_PRIORITY, | |
1493 | .may_unmap = 1, | |
1494 | }; | |
060f005f | 1495 | struct reclaim_stat dummy_stat; |
3c710c1a | 1496 | unsigned long ret; |
02c6de8d MK |
1497 | struct page *page, *next; |
1498 | LIST_HEAD(clean_pages); | |
1499 | ||
1500 | list_for_each_entry_safe(page, next, page_list, lru) { | |
117aad1e | 1501 | if (page_is_file_cache(page) && !PageDirty(page) && |
a58f2cef | 1502 | !__PageMovable(page) && !PageUnevictable(page)) { |
02c6de8d MK |
1503 | ClearPageActive(page); |
1504 | list_move(&page->lru, &clean_pages); | |
1505 | } | |
1506 | } | |
1507 | ||
599d0c95 | 1508 | ret = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
060f005f | 1509 | TTU_IGNORE_ACCESS, &dummy_stat, true); |
02c6de8d | 1510 | list_splice(&clean_pages, page_list); |
599d0c95 | 1511 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, -ret); |
02c6de8d MK |
1512 | return ret; |
1513 | } | |
1514 | ||
5ad333eb AW |
1515 | /* |
1516 | * Attempt to remove the specified page from its LRU. Only take this page | |
1517 | * if it is of the appropriate PageActive status. Pages which are being | |
1518 | * freed elsewhere are also ignored. | |
1519 | * | |
1520 | * page: page to consider | |
1521 | * mode: one of the LRU isolation modes defined above | |
1522 | * | |
1523 | * returns 0 on success, -ve errno on failure. | |
1524 | */ | |
f3fd4a61 | 1525 | int __isolate_lru_page(struct page *page, isolate_mode_t mode) |
5ad333eb AW |
1526 | { |
1527 | int ret = -EINVAL; | |
1528 | ||
1529 | /* Only take pages on the LRU. */ | |
1530 | if (!PageLRU(page)) | |
1531 | return ret; | |
1532 | ||
e46a2879 MK |
1533 | /* Compaction should not handle unevictable pages but CMA can do so */ |
1534 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
894bc310 LS |
1535 | return ret; |
1536 | ||
5ad333eb | 1537 | ret = -EBUSY; |
08e552c6 | 1538 | |
c8244935 MG |
1539 | /* |
1540 | * To minimise LRU disruption, the caller can indicate that it only | |
1541 | * wants to isolate pages it will be able to operate on without | |
1542 | * blocking - clean pages for the most part. | |
1543 | * | |
c8244935 MG |
1544 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
1545 | * that it is possible to migrate without blocking | |
1546 | */ | |
1276ad68 | 1547 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
c8244935 MG |
1548 | /* All the caller can do on PageWriteback is block */ |
1549 | if (PageWriteback(page)) | |
1550 | return ret; | |
1551 | ||
1552 | if (PageDirty(page)) { | |
1553 | struct address_space *mapping; | |
69d763fc | 1554 | bool migrate_dirty; |
c8244935 | 1555 | |
c8244935 MG |
1556 | /* |
1557 | * Only pages without mappings or that have a | |
1558 | * ->migratepage callback are possible to migrate | |
69d763fc MG |
1559 | * without blocking. However, we can be racing with |
1560 | * truncation so it's necessary to lock the page | |
1561 | * to stabilise the mapping as truncation holds | |
1562 | * the page lock until after the page is removed | |
1563 | * from the page cache. | |
c8244935 | 1564 | */ |
69d763fc MG |
1565 | if (!trylock_page(page)) |
1566 | return ret; | |
1567 | ||
c8244935 | 1568 | mapping = page_mapping(page); |
145e1a71 | 1569 | migrate_dirty = !mapping || mapping->a_ops->migratepage; |
69d763fc MG |
1570 | unlock_page(page); |
1571 | if (!migrate_dirty) | |
c8244935 MG |
1572 | return ret; |
1573 | } | |
1574 | } | |
39deaf85 | 1575 | |
f80c0673 MK |
1576 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
1577 | return ret; | |
1578 | ||
5ad333eb AW |
1579 | if (likely(get_page_unless_zero(page))) { |
1580 | /* | |
1581 | * Be careful not to clear PageLRU until after we're | |
1582 | * sure the page is not being freed elsewhere -- the | |
1583 | * page release code relies on it. | |
1584 | */ | |
1585 | ClearPageLRU(page); | |
1586 | ret = 0; | |
1587 | } | |
1588 | ||
1589 | return ret; | |
1590 | } | |
1591 | ||
7ee36a14 MG |
1592 | |
1593 | /* | |
1594 | * Update LRU sizes after isolating pages. The LRU size updates must | |
1595 | * be complete before mem_cgroup_update_lru_size due to a santity check. | |
1596 | */ | |
1597 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 1598 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 1599 | { |
7ee36a14 MG |
1600 | int zid; |
1601 | ||
7ee36a14 MG |
1602 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1603 | if (!nr_zone_taken[zid]) | |
1604 | continue; | |
1605 | ||
1606 | __update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); | |
7ee36a14 | 1607 | #ifdef CONFIG_MEMCG |
b4536f0c | 1608 | mem_cgroup_update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
7ee36a14 | 1609 | #endif |
b4536f0c MH |
1610 | } |
1611 | ||
7ee36a14 MG |
1612 | } |
1613 | ||
f4b7e272 AR |
1614 | /** |
1615 | * pgdat->lru_lock is heavily contended. Some of the functions that | |
1da177e4 LT |
1616 | * shrink the lists perform better by taking out a batch of pages |
1617 | * and working on them outside the LRU lock. | |
1618 | * | |
1619 | * For pagecache intensive workloads, this function is the hottest | |
1620 | * spot in the kernel (apart from copy_*_user functions). | |
1621 | * | |
1622 | * Appropriate locks must be held before calling this function. | |
1623 | * | |
791b48b6 | 1624 | * @nr_to_scan: The number of eligible pages to look through on the list. |
5dc35979 | 1625 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 1626 | * @dst: The temp list to put pages on to. |
f626012d | 1627 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 1628 | * @sc: The scan_control struct for this reclaim session |
5ad333eb | 1629 | * @mode: One of the LRU isolation modes |
3cb99451 | 1630 | * @lru: LRU list id for isolating |
1da177e4 LT |
1631 | * |
1632 | * returns how many pages were moved onto *@dst. | |
1633 | */ | |
69e05944 | 1634 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 1635 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 1636 | unsigned long *nr_scanned, struct scan_control *sc, |
a9e7c39f | 1637 | enum lru_list lru) |
1da177e4 | 1638 | { |
75b00af7 | 1639 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 1640 | unsigned long nr_taken = 0; |
599d0c95 | 1641 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 1642 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
3db65812 | 1643 | unsigned long skipped = 0; |
791b48b6 | 1644 | unsigned long scan, total_scan, nr_pages; |
b2e18757 | 1645 | LIST_HEAD(pages_skipped); |
a9e7c39f | 1646 | isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED); |
1da177e4 | 1647 | |
98879b3b | 1648 | total_scan = 0; |
791b48b6 | 1649 | scan = 0; |
98879b3b | 1650 | while (scan < nr_to_scan && !list_empty(src)) { |
5ad333eb | 1651 | struct page *page; |
5ad333eb | 1652 | |
1da177e4 LT |
1653 | page = lru_to_page(src); |
1654 | prefetchw_prev_lru_page(page, src, flags); | |
1655 | ||
309381fe | 1656 | VM_BUG_ON_PAGE(!PageLRU(page), page); |
8d438f96 | 1657 | |
d8c6546b | 1658 | nr_pages = compound_nr(page); |
98879b3b YS |
1659 | total_scan += nr_pages; |
1660 | ||
b2e18757 MG |
1661 | if (page_zonenum(page) > sc->reclaim_idx) { |
1662 | list_move(&page->lru, &pages_skipped); | |
98879b3b | 1663 | nr_skipped[page_zonenum(page)] += nr_pages; |
b2e18757 MG |
1664 | continue; |
1665 | } | |
1666 | ||
791b48b6 MK |
1667 | /* |
1668 | * Do not count skipped pages because that makes the function | |
1669 | * return with no isolated pages if the LRU mostly contains | |
1670 | * ineligible pages. This causes the VM to not reclaim any | |
1671 | * pages, triggering a premature OOM. | |
98879b3b YS |
1672 | * |
1673 | * Account all tail pages of THP. This would not cause | |
1674 | * premature OOM since __isolate_lru_page() returns -EBUSY | |
1675 | * only when the page is being freed somewhere else. | |
791b48b6 | 1676 | */ |
98879b3b | 1677 | scan += nr_pages; |
f3fd4a61 | 1678 | switch (__isolate_lru_page(page, mode)) { |
5ad333eb | 1679 | case 0: |
599d0c95 MG |
1680 | nr_taken += nr_pages; |
1681 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
5ad333eb | 1682 | list_move(&page->lru, dst); |
5ad333eb AW |
1683 | break; |
1684 | ||
1685 | case -EBUSY: | |
1686 | /* else it is being freed elsewhere */ | |
1687 | list_move(&page->lru, src); | |
1688 | continue; | |
46453a6e | 1689 | |
5ad333eb AW |
1690 | default: |
1691 | BUG(); | |
1692 | } | |
1da177e4 LT |
1693 | } |
1694 | ||
b2e18757 MG |
1695 | /* |
1696 | * Splice any skipped pages to the start of the LRU list. Note that | |
1697 | * this disrupts the LRU order when reclaiming for lower zones but | |
1698 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
1699 | * scanning would soon rescan the same pages to skip and put the | |
1700 | * system at risk of premature OOM. | |
1701 | */ | |
7cc30fcf MG |
1702 | if (!list_empty(&pages_skipped)) { |
1703 | int zid; | |
1704 | ||
3db65812 | 1705 | list_splice(&pages_skipped, src); |
7cc30fcf MG |
1706 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
1707 | if (!nr_skipped[zid]) | |
1708 | continue; | |
1709 | ||
1710 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 1711 | skipped += nr_skipped[zid]; |
7cc30fcf MG |
1712 | } |
1713 | } | |
791b48b6 | 1714 | *nr_scanned = total_scan; |
1265e3a6 | 1715 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
791b48b6 | 1716 | total_scan, skipped, nr_taken, mode, lru); |
b4536f0c | 1717 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
1718 | return nr_taken; |
1719 | } | |
1720 | ||
62695a84 NP |
1721 | /** |
1722 | * isolate_lru_page - tries to isolate a page from its LRU list | |
1723 | * @page: page to isolate from its LRU list | |
1724 | * | |
1725 | * Isolates a @page from an LRU list, clears PageLRU and adjusts the | |
1726 | * vmstat statistic corresponding to whatever LRU list the page was on. | |
1727 | * | |
1728 | * Returns 0 if the page was removed from an LRU list. | |
1729 | * Returns -EBUSY if the page was not on an LRU list. | |
1730 | * | |
1731 | * The returned page will have PageLRU() cleared. If it was found on | |
894bc310 LS |
1732 | * the active list, it will have PageActive set. If it was found on |
1733 | * the unevictable list, it will have the PageUnevictable bit set. That flag | |
1734 | * may need to be cleared by the caller before letting the page go. | |
62695a84 NP |
1735 | * |
1736 | * The vmstat statistic corresponding to the list on which the page was | |
1737 | * found will be decremented. | |
1738 | * | |
1739 | * Restrictions: | |
a5d09bed | 1740 | * |
62695a84 NP |
1741 | * (1) Must be called with an elevated refcount on the page. This is a |
1742 | * fundamentnal difference from isolate_lru_pages (which is called | |
1743 | * without a stable reference). | |
1744 | * (2) the lru_lock must not be held. | |
1745 | * (3) interrupts must be enabled. | |
1746 | */ | |
1747 | int isolate_lru_page(struct page *page) | |
1748 | { | |
1749 | int ret = -EBUSY; | |
1750 | ||
309381fe | 1751 | VM_BUG_ON_PAGE(!page_count(page), page); |
cf2a82ee | 1752 | WARN_RATELIMIT(PageTail(page), "trying to isolate tail page"); |
0c917313 | 1753 | |
62695a84 | 1754 | if (PageLRU(page)) { |
f4b7e272 | 1755 | pg_data_t *pgdat = page_pgdat(page); |
fa9add64 | 1756 | struct lruvec *lruvec; |
62695a84 | 1757 | |
f4b7e272 AR |
1758 | spin_lock_irq(&pgdat->lru_lock); |
1759 | lruvec = mem_cgroup_page_lruvec(page, pgdat); | |
0c917313 | 1760 | if (PageLRU(page)) { |
894bc310 | 1761 | int lru = page_lru(page); |
0c917313 | 1762 | get_page(page); |
62695a84 | 1763 | ClearPageLRU(page); |
fa9add64 HD |
1764 | del_page_from_lru_list(page, lruvec, lru); |
1765 | ret = 0; | |
62695a84 | 1766 | } |
f4b7e272 | 1767 | spin_unlock_irq(&pgdat->lru_lock); |
62695a84 NP |
1768 | } |
1769 | return ret; | |
1770 | } | |
1771 | ||
35cd7815 | 1772 | /* |
d37dd5dc FW |
1773 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
1774 | * then get resheduled. When there are massive number of tasks doing page | |
1775 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, | |
1776 | * the LRU list will go small and be scanned faster than necessary, leading to | |
1777 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 1778 | */ |
599d0c95 | 1779 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
1780 | struct scan_control *sc) |
1781 | { | |
1782 | unsigned long inactive, isolated; | |
1783 | ||
1784 | if (current_is_kswapd()) | |
1785 | return 0; | |
1786 | ||
b5ead35e | 1787 | if (!writeback_throttling_sane(sc)) |
35cd7815 RR |
1788 | return 0; |
1789 | ||
1790 | if (file) { | |
599d0c95 MG |
1791 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
1792 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 1793 | } else { |
599d0c95 MG |
1794 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
1795 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
1796 | } |
1797 | ||
3cf23841 FW |
1798 | /* |
1799 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
1800 | * won't get blocked by normal direct-reclaimers, forming a circular | |
1801 | * deadlock. | |
1802 | */ | |
d0164adc | 1803 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
1804 | inactive >>= 3; |
1805 | ||
35cd7815 RR |
1806 | return isolated > inactive; |
1807 | } | |
1808 | ||
a222f341 KT |
1809 | /* |
1810 | * This moves pages from @list to corresponding LRU list. | |
1811 | * | |
1812 | * We move them the other way if the page is referenced by one or more | |
1813 | * processes, from rmap. | |
1814 | * | |
1815 | * If the pages are mostly unmapped, the processing is fast and it is | |
1816 | * appropriate to hold zone_lru_lock across the whole operation. But if | |
1817 | * the pages are mapped, the processing is slow (page_referenced()) so we | |
1818 | * should drop zone_lru_lock around each page. It's impossible to balance | |
1819 | * this, so instead we remove the pages from the LRU while processing them. | |
1820 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
1821 | * nobody will play with that bit on a non-LRU page. | |
1822 | * | |
1823 | * The downside is that we have to touch page->_refcount against each page. | |
1824 | * But we had to alter page->flags anyway. | |
1825 | * | |
1826 | * Returns the number of pages moved to the given lruvec. | |
1827 | */ | |
1828 | ||
1829 | static unsigned noinline_for_stack move_pages_to_lru(struct lruvec *lruvec, | |
1830 | struct list_head *list) | |
66635629 | 1831 | { |
599d0c95 | 1832 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
a222f341 | 1833 | int nr_pages, nr_moved = 0; |
3f79768f | 1834 | LIST_HEAD(pages_to_free); |
a222f341 KT |
1835 | struct page *page; |
1836 | enum lru_list lru; | |
66635629 | 1837 | |
a222f341 KT |
1838 | while (!list_empty(list)) { |
1839 | page = lru_to_page(list); | |
309381fe | 1840 | VM_BUG_ON_PAGE(PageLRU(page), page); |
39b5f29a | 1841 | if (unlikely(!page_evictable(page))) { |
a222f341 | 1842 | list_del(&page->lru); |
599d0c95 | 1843 | spin_unlock_irq(&pgdat->lru_lock); |
66635629 | 1844 | putback_lru_page(page); |
599d0c95 | 1845 | spin_lock_irq(&pgdat->lru_lock); |
66635629 MG |
1846 | continue; |
1847 | } | |
599d0c95 | 1848 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
fa9add64 | 1849 | |
7a608572 | 1850 | SetPageLRU(page); |
66635629 | 1851 | lru = page_lru(page); |
a222f341 KT |
1852 | |
1853 | nr_pages = hpage_nr_pages(page); | |
1854 | update_lru_size(lruvec, lru, page_zonenum(page), nr_pages); | |
1855 | list_move(&page->lru, &lruvec->lists[lru]); | |
fa9add64 | 1856 | |
2bcf8879 HD |
1857 | if (put_page_testzero(page)) { |
1858 | __ClearPageLRU(page); | |
1859 | __ClearPageActive(page); | |
fa9add64 | 1860 | del_page_from_lru_list(page, lruvec, lru); |
2bcf8879 HD |
1861 | |
1862 | if (unlikely(PageCompound(page))) { | |
599d0c95 | 1863 | spin_unlock_irq(&pgdat->lru_lock); |
2bcf8879 | 1864 | (*get_compound_page_dtor(page))(page); |
599d0c95 | 1865 | spin_lock_irq(&pgdat->lru_lock); |
2bcf8879 HD |
1866 | } else |
1867 | list_add(&page->lru, &pages_to_free); | |
a222f341 KT |
1868 | } else { |
1869 | nr_moved += nr_pages; | |
66635629 MG |
1870 | } |
1871 | } | |
66635629 | 1872 | |
3f79768f HD |
1873 | /* |
1874 | * To save our caller's stack, now use input list for pages to free. | |
1875 | */ | |
a222f341 KT |
1876 | list_splice(&pages_to_free, list); |
1877 | ||
1878 | return nr_moved; | |
66635629 MG |
1879 | } |
1880 | ||
399ba0b9 N |
1881 | /* |
1882 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
1883 | * a backing device by writing to the page cache it sets PF_LESS_THROTTLE. | |
1884 | * In that case we should only throttle if the backing device it is | |
1885 | * writing to is congested. In other cases it is safe to throttle. | |
1886 | */ | |
1887 | static int current_may_throttle(void) | |
1888 | { | |
1889 | return !(current->flags & PF_LESS_THROTTLE) || | |
1890 | current->backing_dev_info == NULL || | |
1891 | bdi_write_congested(current->backing_dev_info); | |
1892 | } | |
1893 | ||
1da177e4 | 1894 | /* |
b2e18757 | 1895 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 1896 | * of reclaimed pages |
1da177e4 | 1897 | */ |
66635629 | 1898 | static noinline_for_stack unsigned long |
1a93be0e | 1899 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 1900 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
1901 | { |
1902 | LIST_HEAD(page_list); | |
e247dbce | 1903 | unsigned long nr_scanned; |
05ff5137 | 1904 | unsigned long nr_reclaimed = 0; |
e247dbce | 1905 | unsigned long nr_taken; |
060f005f | 1906 | struct reclaim_stat stat; |
3cb99451 | 1907 | int file = is_file_lru(lru); |
f46b7912 | 1908 | enum vm_event_item item; |
599d0c95 | 1909 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1a93be0e | 1910 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
db73ee0d | 1911 | bool stalled = false; |
78dc583d | 1912 | |
599d0c95 | 1913 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
db73ee0d MH |
1914 | if (stalled) |
1915 | return 0; | |
1916 | ||
1917 | /* wait a bit for the reclaimer. */ | |
1918 | msleep(100); | |
1919 | stalled = true; | |
35cd7815 RR |
1920 | |
1921 | /* We are about to die and free our memory. Return now. */ | |
1922 | if (fatal_signal_pending(current)) | |
1923 | return SWAP_CLUSTER_MAX; | |
1924 | } | |
1925 | ||
1da177e4 | 1926 | lru_add_drain(); |
f80c0673 | 1927 | |
599d0c95 | 1928 | spin_lock_irq(&pgdat->lru_lock); |
b35ea17b | 1929 | |
5dc35979 | 1930 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
a9e7c39f | 1931 | &nr_scanned, sc, lru); |
95d918fc | 1932 | |
599d0c95 | 1933 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
9d5e6a9f | 1934 | reclaim_stat->recent_scanned[file] += nr_taken; |
95d918fc | 1935 | |
f46b7912 | 1936 | item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT; |
b5ead35e | 1937 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
1938 | __count_vm_events(item, nr_scanned); |
1939 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned); | |
599d0c95 | 1940 | spin_unlock_irq(&pgdat->lru_lock); |
b35ea17b | 1941 | |
d563c050 | 1942 | if (nr_taken == 0) |
66635629 | 1943 | return 0; |
5ad333eb | 1944 | |
a128ca71 | 1945 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, 0, |
3c710c1a | 1946 | &stat, false); |
c661b078 | 1947 | |
599d0c95 | 1948 | spin_lock_irq(&pgdat->lru_lock); |
3f79768f | 1949 | |
f46b7912 | 1950 | item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT; |
b5ead35e | 1951 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
1952 | __count_vm_events(item, nr_reclaimed); |
1953 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed); | |
b17f18af KT |
1954 | reclaim_stat->recent_rotated[0] += stat.nr_activate[0]; |
1955 | reclaim_stat->recent_rotated[1] += stat.nr_activate[1]; | |
a74609fa | 1956 | |
a222f341 | 1957 | move_pages_to_lru(lruvec, &page_list); |
3f79768f | 1958 | |
599d0c95 | 1959 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
3f79768f | 1960 | |
599d0c95 | 1961 | spin_unlock_irq(&pgdat->lru_lock); |
3f79768f | 1962 | |
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 | ||
f626012d | 1994 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 1995 | struct lruvec *lruvec, |
f16015fb | 1996 | struct scan_control *sc, |
9e3b2f8c | 1997 | enum lru_list lru) |
1da177e4 | 1998 | { |
44c241f1 | 1999 | unsigned long nr_taken; |
f626012d | 2000 | unsigned long nr_scanned; |
6fe6b7e3 | 2001 | unsigned long vm_flags; |
1da177e4 | 2002 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 2003 | LIST_HEAD(l_active); |
b69408e8 | 2004 | LIST_HEAD(l_inactive); |
1da177e4 | 2005 | struct page *page; |
1a93be0e | 2006 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
9d998b4f MH |
2007 | unsigned nr_deactivate, nr_activate; |
2008 | unsigned nr_rotated = 0; | |
3cb99451 | 2009 | int file = is_file_lru(lru); |
599d0c95 | 2010 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
2011 | |
2012 | lru_add_drain(); | |
f80c0673 | 2013 | |
599d0c95 | 2014 | spin_lock_irq(&pgdat->lru_lock); |
925b7673 | 2015 | |
5dc35979 | 2016 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
a9e7c39f | 2017 | &nr_scanned, sc, lru); |
89b5fae5 | 2018 | |
599d0c95 | 2019 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
b7c46d15 | 2020 | reclaim_stat->recent_scanned[file] += nr_taken; |
1cfb419b | 2021 | |
599d0c95 | 2022 | __count_vm_events(PGREFILL, nr_scanned); |
2fa2690c | 2023 | __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); |
9d5e6a9f | 2024 | |
599d0c95 | 2025 | spin_unlock_irq(&pgdat->lru_lock); |
1da177e4 | 2026 | |
1da177e4 LT |
2027 | while (!list_empty(&l_hold)) { |
2028 | cond_resched(); | |
2029 | page = lru_to_page(&l_hold); | |
2030 | list_del(&page->lru); | |
7e9cd484 | 2031 | |
39b5f29a | 2032 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
2033 | putback_lru_page(page); |
2034 | continue; | |
2035 | } | |
2036 | ||
cc715d99 MG |
2037 | if (unlikely(buffer_heads_over_limit)) { |
2038 | if (page_has_private(page) && trylock_page(page)) { | |
2039 | if (page_has_private(page)) | |
2040 | try_to_release_page(page, 0); | |
2041 | unlock_page(page); | |
2042 | } | |
2043 | } | |
2044 | ||
c3ac9a8a JW |
2045 | if (page_referenced(page, 0, sc->target_mem_cgroup, |
2046 | &vm_flags)) { | |
9992af10 | 2047 | nr_rotated += hpage_nr_pages(page); |
8cab4754 WF |
2048 | /* |
2049 | * Identify referenced, file-backed active pages and | |
2050 | * give them one more trip around the active list. So | |
2051 | * that executable code get better chances to stay in | |
2052 | * memory under moderate memory pressure. Anon pages | |
2053 | * are not likely to be evicted by use-once streaming | |
2054 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
2055 | * so we ignore them here. | |
2056 | */ | |
41e20983 | 2057 | if ((vm_flags & VM_EXEC) && page_is_file_cache(page)) { |
8cab4754 WF |
2058 | list_add(&page->lru, &l_active); |
2059 | continue; | |
2060 | } | |
2061 | } | |
7e9cd484 | 2062 | |
5205e56e | 2063 | ClearPageActive(page); /* we are de-activating */ |
1899ad18 | 2064 | SetPageWorkingset(page); |
1da177e4 LT |
2065 | list_add(&page->lru, &l_inactive); |
2066 | } | |
2067 | ||
b555749a | 2068 | /* |
8cab4754 | 2069 | * Move pages back to the lru list. |
b555749a | 2070 | */ |
599d0c95 | 2071 | spin_lock_irq(&pgdat->lru_lock); |
556adecb | 2072 | /* |
8cab4754 WF |
2073 | * Count referenced pages from currently used mappings as rotated, |
2074 | * even though only some of them are actually re-activated. This | |
2075 | * helps balance scan pressure between file and anonymous pages in | |
7c0db9e9 | 2076 | * get_scan_count. |
7e9cd484 | 2077 | */ |
b7c46d15 | 2078 | reclaim_stat->recent_rotated[file] += nr_rotated; |
556adecb | 2079 | |
a222f341 KT |
2080 | nr_activate = move_pages_to_lru(lruvec, &l_active); |
2081 | nr_deactivate = move_pages_to_lru(lruvec, &l_inactive); | |
f372d89e KT |
2082 | /* Keep all free pages in l_active list */ |
2083 | list_splice(&l_inactive, &l_active); | |
9851ac13 KT |
2084 | |
2085 | __count_vm_events(PGDEACTIVATE, nr_deactivate); | |
2086 | __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate); | |
2087 | ||
599d0c95 MG |
2088 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
2089 | spin_unlock_irq(&pgdat->lru_lock); | |
2bcf8879 | 2090 | |
f372d89e KT |
2091 | mem_cgroup_uncharge_list(&l_active); |
2092 | free_unref_page_list(&l_active); | |
9d998b4f MH |
2093 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
2094 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
2095 | } |
2096 | ||
1a4e58cc MK |
2097 | unsigned long reclaim_pages(struct list_head *page_list) |
2098 | { | |
2099 | int nid = -1; | |
2100 | unsigned long nr_reclaimed = 0; | |
2101 | LIST_HEAD(node_page_list); | |
2102 | struct reclaim_stat dummy_stat; | |
2103 | struct page *page; | |
2104 | struct scan_control sc = { | |
2105 | .gfp_mask = GFP_KERNEL, | |
2106 | .priority = DEF_PRIORITY, | |
2107 | .may_writepage = 1, | |
2108 | .may_unmap = 1, | |
2109 | .may_swap = 1, | |
2110 | }; | |
2111 | ||
2112 | while (!list_empty(page_list)) { | |
2113 | page = lru_to_page(page_list); | |
2114 | if (nid == -1) { | |
2115 | nid = page_to_nid(page); | |
2116 | INIT_LIST_HEAD(&node_page_list); | |
2117 | } | |
2118 | ||
2119 | if (nid == page_to_nid(page)) { | |
2120 | ClearPageActive(page); | |
2121 | list_move(&page->lru, &node_page_list); | |
2122 | continue; | |
2123 | } | |
2124 | ||
2125 | nr_reclaimed += shrink_page_list(&node_page_list, | |
2126 | NODE_DATA(nid), | |
2127 | &sc, 0, | |
2128 | &dummy_stat, false); | |
2129 | while (!list_empty(&node_page_list)) { | |
2130 | page = lru_to_page(&node_page_list); | |
2131 | list_del(&page->lru); | |
2132 | putback_lru_page(page); | |
2133 | } | |
2134 | ||
2135 | nid = -1; | |
2136 | } | |
2137 | ||
2138 | if (!list_empty(&node_page_list)) { | |
2139 | nr_reclaimed += shrink_page_list(&node_page_list, | |
2140 | NODE_DATA(nid), | |
2141 | &sc, 0, | |
2142 | &dummy_stat, false); | |
2143 | while (!list_empty(&node_page_list)) { | |
2144 | page = lru_to_page(&node_page_list); | |
2145 | list_del(&page->lru); | |
2146 | putback_lru_page(page); | |
2147 | } | |
2148 | } | |
2149 | ||
2150 | return nr_reclaimed; | |
2151 | } | |
2152 | ||
59dc76b0 RR |
2153 | /* |
2154 | * The inactive anon list should be small enough that the VM never has | |
2155 | * to do too much work. | |
14797e23 | 2156 | * |
59dc76b0 RR |
2157 | * The inactive file list should be small enough to leave most memory |
2158 | * to the established workingset on the scan-resistant active list, | |
2159 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2160 | * |
59dc76b0 RR |
2161 | * Both inactive lists should also be large enough that each inactive |
2162 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 2163 | * |
2a2e4885 JW |
2164 | * If that fails and refaulting is observed, the inactive list grows. |
2165 | * | |
59dc76b0 | 2166 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
3a50d14d | 2167 | * on this LRU, maintained by the pageout code. An inactive_ratio |
59dc76b0 | 2168 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. |
56e49d21 | 2169 | * |
59dc76b0 RR |
2170 | * total target max |
2171 | * memory ratio inactive | |
2172 | * ------------------------------------- | |
2173 | * 10MB 1 5MB | |
2174 | * 100MB 1 50MB | |
2175 | * 1GB 3 250MB | |
2176 | * 10GB 10 0.9GB | |
2177 | * 100GB 31 3GB | |
2178 | * 1TB 101 10GB | |
2179 | * 10TB 320 32GB | |
56e49d21 | 2180 | */ |
f8d1a311 | 2181 | static bool inactive_list_is_low(struct lruvec *lruvec, bool file, |
2c012a4a | 2182 | struct scan_control *sc, bool trace) |
56e49d21 | 2183 | { |
fd538803 | 2184 | enum lru_list active_lru = file * LRU_FILE + LRU_ACTIVE; |
2a2e4885 JW |
2185 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
2186 | enum lru_list inactive_lru = file * LRU_FILE; | |
2187 | unsigned long inactive, active; | |
2188 | unsigned long inactive_ratio; | |
2189 | unsigned long refaults; | |
59dc76b0 | 2190 | unsigned long gb; |
e3790144 | 2191 | |
fd538803 MH |
2192 | inactive = lruvec_lru_size(lruvec, inactive_lru, sc->reclaim_idx); |
2193 | active = lruvec_lru_size(lruvec, active_lru, sc->reclaim_idx); | |
f8d1a311 | 2194 | |
2a2e4885 JW |
2195 | /* |
2196 | * When refaults are being observed, it means a new workingset | |
2197 | * is being established. Disable active list protection to get | |
2198 | * rid of the stale workingset quickly. | |
2199 | */ | |
205b20cc | 2200 | refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE); |
2c012a4a | 2201 | if (file && lruvec->refaults != refaults) { |
2a2e4885 JW |
2202 | inactive_ratio = 0; |
2203 | } else { | |
2204 | gb = (inactive + active) >> (30 - PAGE_SHIFT); | |
2205 | if (gb) | |
2206 | inactive_ratio = int_sqrt(10 * gb); | |
2207 | else | |
2208 | inactive_ratio = 1; | |
2209 | } | |
59dc76b0 | 2210 | |
2c012a4a | 2211 | if (trace) |
2a2e4885 JW |
2212 | trace_mm_vmscan_inactive_list_is_low(pgdat->node_id, sc->reclaim_idx, |
2213 | lruvec_lru_size(lruvec, inactive_lru, MAX_NR_ZONES), inactive, | |
2214 | lruvec_lru_size(lruvec, active_lru, MAX_NR_ZONES), active, | |
2215 | inactive_ratio, file); | |
fd538803 | 2216 | |
59dc76b0 | 2217 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2218 | } |
2219 | ||
4f98a2fe | 2220 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
3b991208 | 2221 | struct lruvec *lruvec, struct scan_control *sc) |
b69408e8 | 2222 | { |
b39415b2 | 2223 | if (is_active_lru(lru)) { |
3b991208 | 2224 | if (inactive_list_is_low(lruvec, is_file_lru(lru), sc, true)) |
1a93be0e | 2225 | shrink_active_list(nr_to_scan, lruvec, sc, lru); |
556adecb RR |
2226 | return 0; |
2227 | } | |
2228 | ||
1a93be0e | 2229 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); |
4f98a2fe RR |
2230 | } |
2231 | ||
9a265114 JW |
2232 | enum scan_balance { |
2233 | SCAN_EQUAL, | |
2234 | SCAN_FRACT, | |
2235 | SCAN_ANON, | |
2236 | SCAN_FILE, | |
2237 | }; | |
2238 | ||
4f98a2fe RR |
2239 | /* |
2240 | * Determine how aggressively the anon and file LRU lists should be | |
2241 | * scanned. The relative value of each set of LRU lists is determined | |
2242 | * by looking at the fraction of the pages scanned we did rotate back | |
2243 | * onto the active list instead of evict. | |
2244 | * | |
be7bd59d WL |
2245 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2246 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 2247 | */ |
afaf07a6 JW |
2248 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, |
2249 | unsigned long *nr) | |
4f98a2fe | 2250 | { |
afaf07a6 | 2251 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); |
33377678 | 2252 | int swappiness = mem_cgroup_swappiness(memcg); |
9a265114 JW |
2253 | struct zone_reclaim_stat *reclaim_stat = &lruvec->reclaim_stat; |
2254 | u64 fraction[2]; | |
2255 | u64 denominator = 0; /* gcc */ | |
599d0c95 | 2256 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
4f98a2fe | 2257 | unsigned long anon_prio, file_prio; |
9a265114 | 2258 | enum scan_balance scan_balance; |
0bf1457f | 2259 | unsigned long anon, file; |
4f98a2fe | 2260 | unsigned long ap, fp; |
4111304d | 2261 | enum lru_list lru; |
76a33fc3 SL |
2262 | |
2263 | /* If we have no swap space, do not bother scanning anon pages. */ | |
d8b38438 | 2264 | if (!sc->may_swap || mem_cgroup_get_nr_swap_pages(memcg) <= 0) { |
9a265114 | 2265 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2266 | goto out; |
2267 | } | |
4f98a2fe | 2268 | |
10316b31 JW |
2269 | /* |
2270 | * Global reclaim will swap to prevent OOM even with no | |
2271 | * swappiness, but memcg users want to use this knob to | |
2272 | * disable swapping for individual groups completely when | |
2273 | * using the memory controller's swap limit feature would be | |
2274 | * too expensive. | |
2275 | */ | |
b5ead35e | 2276 | if (cgroup_reclaim(sc) && !swappiness) { |
9a265114 | 2277 | scan_balance = SCAN_FILE; |
10316b31 JW |
2278 | goto out; |
2279 | } | |
2280 | ||
2281 | /* | |
2282 | * Do not apply any pressure balancing cleverness when the | |
2283 | * system is close to OOM, scan both anon and file equally | |
2284 | * (unless the swappiness setting disagrees with swapping). | |
2285 | */ | |
02695175 | 2286 | if (!sc->priority && swappiness) { |
9a265114 | 2287 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2288 | goto out; |
2289 | } | |
2290 | ||
62376251 JW |
2291 | /* |
2292 | * Prevent the reclaimer from falling into the cache trap: as | |
2293 | * cache pages start out inactive, every cache fault will tip | |
2294 | * the scan balance towards the file LRU. And as the file LRU | |
2295 | * shrinks, so does the window for rotation from references. | |
2296 | * This means we have a runaway feedback loop where a tiny | |
2297 | * thrashing file LRU becomes infinitely more attractive than | |
2298 | * anon pages. Try to detect this based on file LRU size. | |
2299 | */ | |
b5ead35e | 2300 | if (!cgroup_reclaim(sc)) { |
599d0c95 MG |
2301 | unsigned long pgdatfile; |
2302 | unsigned long pgdatfree; | |
2303 | int z; | |
2304 | unsigned long total_high_wmark = 0; | |
2ab051e1 | 2305 | |
599d0c95 MG |
2306 | pgdatfree = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); |
2307 | pgdatfile = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2308 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2309 | ||
2310 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2311 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2312 | if (!managed_zone(zone)) |
599d0c95 MG |
2313 | continue; |
2314 | ||
2315 | total_high_wmark += high_wmark_pages(zone); | |
2316 | } | |
62376251 | 2317 | |
599d0c95 | 2318 | if (unlikely(pgdatfile + pgdatfree <= total_high_wmark)) { |
06226226 DR |
2319 | /* |
2320 | * Force SCAN_ANON if there are enough inactive | |
2321 | * anonymous pages on the LRU in eligible zones. | |
2322 | * Otherwise, the small LRU gets thrashed. | |
2323 | */ | |
3b991208 | 2324 | if (!inactive_list_is_low(lruvec, false, sc, false) && |
06226226 DR |
2325 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, sc->reclaim_idx) |
2326 | >> sc->priority) { | |
2327 | scan_balance = SCAN_ANON; | |
2328 | goto out; | |
2329 | } | |
62376251 JW |
2330 | } |
2331 | } | |
2332 | ||
7c5bd705 | 2333 | /* |
316bda0e VD |
2334 | * If there is enough inactive page cache, i.e. if the size of the |
2335 | * inactive list is greater than that of the active list *and* the | |
2336 | * inactive list actually has some pages to scan on this priority, we | |
2337 | * do not reclaim anything from the anonymous working set right now. | |
2338 | * Without the second condition we could end up never scanning an | |
2339 | * lruvec even if it has plenty of old anonymous pages unless the | |
2340 | * system is under heavy pressure. | |
7c5bd705 | 2341 | */ |
3b991208 | 2342 | if (!inactive_list_is_low(lruvec, true, sc, false) && |
71ab6cfe | 2343 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, sc->reclaim_idx) >> sc->priority) { |
9a265114 | 2344 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2345 | goto out; |
2346 | } | |
2347 | ||
9a265114 JW |
2348 | scan_balance = SCAN_FRACT; |
2349 | ||
58c37f6e KM |
2350 | /* |
2351 | * With swappiness at 100, anonymous and file have the same priority. | |
2352 | * This scanning priority is essentially the inverse of IO cost. | |
2353 | */ | |
02695175 | 2354 | anon_prio = swappiness; |
75b00af7 | 2355 | file_prio = 200 - anon_prio; |
58c37f6e | 2356 | |
4f98a2fe RR |
2357 | /* |
2358 | * OK, so we have swap space and a fair amount of page cache | |
2359 | * pages. We use the recently rotated / recently scanned | |
2360 | * ratios to determine how valuable each cache is. | |
2361 | * | |
2362 | * Because workloads change over time (and to avoid overflow) | |
2363 | * we keep these statistics as a floating average, which ends | |
2364 | * up weighing recent references more than old ones. | |
2365 | * | |
2366 | * anon in [0], file in [1] | |
2367 | */ | |
2ab051e1 | 2368 | |
fd538803 MH |
2369 | anon = lruvec_lru_size(lruvec, LRU_ACTIVE_ANON, MAX_NR_ZONES) + |
2370 | lruvec_lru_size(lruvec, LRU_INACTIVE_ANON, MAX_NR_ZONES); | |
2371 | file = lruvec_lru_size(lruvec, LRU_ACTIVE_FILE, MAX_NR_ZONES) + | |
2372 | lruvec_lru_size(lruvec, LRU_INACTIVE_FILE, MAX_NR_ZONES); | |
2ab051e1 | 2373 | |
599d0c95 | 2374 | spin_lock_irq(&pgdat->lru_lock); |
6e901571 | 2375 | if (unlikely(reclaim_stat->recent_scanned[0] > anon / 4)) { |
6e901571 KM |
2376 | reclaim_stat->recent_scanned[0] /= 2; |
2377 | reclaim_stat->recent_rotated[0] /= 2; | |
4f98a2fe RR |
2378 | } |
2379 | ||
6e901571 | 2380 | if (unlikely(reclaim_stat->recent_scanned[1] > file / 4)) { |
6e901571 KM |
2381 | reclaim_stat->recent_scanned[1] /= 2; |
2382 | reclaim_stat->recent_rotated[1] /= 2; | |
4f98a2fe RR |
2383 | } |
2384 | ||
4f98a2fe | 2385 | /* |
00d8089c RR |
2386 | * The amount of pressure on anon vs file pages is inversely |
2387 | * proportional to the fraction of recently scanned pages on | |
2388 | * each list that were recently referenced and in active use. | |
4f98a2fe | 2389 | */ |
fe35004f | 2390 | ap = anon_prio * (reclaim_stat->recent_scanned[0] + 1); |
6e901571 | 2391 | ap /= reclaim_stat->recent_rotated[0] + 1; |
4f98a2fe | 2392 | |
fe35004f | 2393 | fp = file_prio * (reclaim_stat->recent_scanned[1] + 1); |
6e901571 | 2394 | fp /= reclaim_stat->recent_rotated[1] + 1; |
599d0c95 | 2395 | spin_unlock_irq(&pgdat->lru_lock); |
4f98a2fe | 2396 | |
76a33fc3 SL |
2397 | fraction[0] = ap; |
2398 | fraction[1] = fp; | |
2399 | denominator = ap + fp + 1; | |
2400 | out: | |
688035f7 JW |
2401 | for_each_evictable_lru(lru) { |
2402 | int file = is_file_lru(lru); | |
9783aa99 | 2403 | unsigned long lruvec_size; |
688035f7 | 2404 | unsigned long scan; |
1bc63fb1 | 2405 | unsigned long protection; |
9783aa99 CD |
2406 | |
2407 | lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); | |
1bc63fb1 CD |
2408 | protection = mem_cgroup_protection(memcg, |
2409 | sc->memcg_low_reclaim); | |
9783aa99 | 2410 | |
1bc63fb1 | 2411 | if (protection) { |
9783aa99 CD |
2412 | /* |
2413 | * Scale a cgroup's reclaim pressure by proportioning | |
2414 | * its current usage to its memory.low or memory.min | |
2415 | * setting. | |
2416 | * | |
2417 | * This is important, as otherwise scanning aggression | |
2418 | * becomes extremely binary -- from nothing as we | |
2419 | * approach the memory protection threshold, to totally | |
2420 | * nominal as we exceed it. This results in requiring | |
2421 | * setting extremely liberal protection thresholds. It | |
2422 | * also means we simply get no protection at all if we | |
2423 | * set it too low, which is not ideal. | |
1bc63fb1 CD |
2424 | * |
2425 | * If there is any protection in place, we reduce scan | |
2426 | * pressure by how much of the total memory used is | |
2427 | * within protection thresholds. | |
9783aa99 | 2428 | * |
9de7ca46 CD |
2429 | * There is one special case: in the first reclaim pass, |
2430 | * we skip over all groups that are within their low | |
2431 | * protection. If that fails to reclaim enough pages to | |
2432 | * satisfy the reclaim goal, we come back and override | |
2433 | * the best-effort low protection. However, we still | |
2434 | * ideally want to honor how well-behaved groups are in | |
2435 | * that case instead of simply punishing them all | |
2436 | * equally. As such, we reclaim them based on how much | |
1bc63fb1 CD |
2437 | * memory they are using, reducing the scan pressure |
2438 | * again by how much of the total memory used is under | |
2439 | * hard protection. | |
9783aa99 | 2440 | */ |
1bc63fb1 CD |
2441 | unsigned long cgroup_size = mem_cgroup_size(memcg); |
2442 | ||
2443 | /* Avoid TOCTOU with earlier protection check */ | |
2444 | cgroup_size = max(cgroup_size, protection); | |
2445 | ||
2446 | scan = lruvec_size - lruvec_size * protection / | |
2447 | cgroup_size; | |
9783aa99 CD |
2448 | |
2449 | /* | |
1bc63fb1 | 2450 | * Minimally target SWAP_CLUSTER_MAX pages to keep |
9de7ca46 CD |
2451 | * reclaim moving forwards, avoiding decremeting |
2452 | * sc->priority further than desirable. | |
9783aa99 | 2453 | */ |
1bc63fb1 | 2454 | scan = max(scan, SWAP_CLUSTER_MAX); |
9783aa99 CD |
2455 | } else { |
2456 | scan = lruvec_size; | |
2457 | } | |
2458 | ||
2459 | scan >>= sc->priority; | |
6b4f7799 | 2460 | |
688035f7 JW |
2461 | /* |
2462 | * If the cgroup's already been deleted, make sure to | |
2463 | * scrape out the remaining cache. | |
2464 | */ | |
2465 | if (!scan && !mem_cgroup_online(memcg)) | |
9783aa99 | 2466 | scan = min(lruvec_size, SWAP_CLUSTER_MAX); |
6b4f7799 | 2467 | |
688035f7 JW |
2468 | switch (scan_balance) { |
2469 | case SCAN_EQUAL: | |
2470 | /* Scan lists relative to size */ | |
2471 | break; | |
2472 | case SCAN_FRACT: | |
9a265114 | 2473 | /* |
688035f7 JW |
2474 | * Scan types proportional to swappiness and |
2475 | * their relative recent reclaim efficiency. | |
68600f62 RG |
2476 | * Make sure we don't miss the last page |
2477 | * because of a round-off error. | |
9a265114 | 2478 | */ |
68600f62 RG |
2479 | scan = DIV64_U64_ROUND_UP(scan * fraction[file], |
2480 | denominator); | |
688035f7 JW |
2481 | break; |
2482 | case SCAN_FILE: | |
2483 | case SCAN_ANON: | |
2484 | /* Scan one type exclusively */ | |
2485 | if ((scan_balance == SCAN_FILE) != file) { | |
9783aa99 | 2486 | lruvec_size = 0; |
688035f7 JW |
2487 | scan = 0; |
2488 | } | |
2489 | break; | |
2490 | default: | |
2491 | /* Look ma, no brain */ | |
2492 | BUG(); | |
9a265114 | 2493 | } |
688035f7 | 2494 | |
688035f7 | 2495 | nr[lru] = scan; |
76a33fc3 | 2496 | } |
6e08a369 | 2497 | } |
4f98a2fe | 2498 | |
afaf07a6 | 2499 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) |
9b4f98cd JW |
2500 | { |
2501 | unsigned long nr[NR_LRU_LISTS]; | |
e82e0561 | 2502 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2503 | unsigned long nr_to_scan; |
2504 | enum lru_list lru; | |
2505 | unsigned long nr_reclaimed = 0; | |
2506 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2507 | struct blk_plug plug; | |
1a501907 | 2508 | bool scan_adjusted; |
9b4f98cd | 2509 | |
afaf07a6 | 2510 | get_scan_count(lruvec, sc, nr); |
9b4f98cd | 2511 | |
e82e0561 MG |
2512 | /* Record the original scan target for proportional adjustments later */ |
2513 | memcpy(targets, nr, sizeof(nr)); | |
2514 | ||
1a501907 MG |
2515 | /* |
2516 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2517 | * event that can occur when there is little memory pressure e.g. | |
2518 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2519 | * when the requested number of pages are reclaimed when scanning at | |
2520 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2521 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2522 | * do a batch of work at once. For memcg reclaim one check is made to | |
2523 | * abort proportional reclaim if either the file or anon lru has already | |
2524 | * dropped to zero at the first pass. | |
2525 | */ | |
b5ead35e | 2526 | scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() && |
1a501907 MG |
2527 | sc->priority == DEF_PRIORITY); |
2528 | ||
9b4f98cd JW |
2529 | blk_start_plug(&plug); |
2530 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2531 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2532 | unsigned long nr_anon, nr_file, percentage; |
2533 | unsigned long nr_scanned; | |
2534 | ||
9b4f98cd JW |
2535 | for_each_evictable_lru(lru) { |
2536 | if (nr[lru]) { | |
2537 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2538 | nr[lru] -= nr_to_scan; | |
2539 | ||
2540 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
3b991208 | 2541 | lruvec, sc); |
9b4f98cd JW |
2542 | } |
2543 | } | |
e82e0561 | 2544 | |
bd041733 MH |
2545 | cond_resched(); |
2546 | ||
e82e0561 MG |
2547 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
2548 | continue; | |
2549 | ||
e82e0561 MG |
2550 | /* |
2551 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 2552 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
2553 | * proportionally what was requested by get_scan_count(). We |
2554 | * stop reclaiming one LRU and reduce the amount scanning | |
2555 | * proportional to the original scan target. | |
2556 | */ | |
2557 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
2558 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
2559 | ||
1a501907 MG |
2560 | /* |
2561 | * It's just vindictive to attack the larger once the smaller | |
2562 | * has gone to zero. And given the way we stop scanning the | |
2563 | * smaller below, this makes sure that we only make one nudge | |
2564 | * towards proportionality once we've got nr_to_reclaim. | |
2565 | */ | |
2566 | if (!nr_file || !nr_anon) | |
2567 | break; | |
2568 | ||
e82e0561 MG |
2569 | if (nr_file > nr_anon) { |
2570 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
2571 | targets[LRU_ACTIVE_ANON] + 1; | |
2572 | lru = LRU_BASE; | |
2573 | percentage = nr_anon * 100 / scan_target; | |
2574 | } else { | |
2575 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
2576 | targets[LRU_ACTIVE_FILE] + 1; | |
2577 | lru = LRU_FILE; | |
2578 | percentage = nr_file * 100 / scan_target; | |
2579 | } | |
2580 | ||
2581 | /* Stop scanning the smaller of the LRU */ | |
2582 | nr[lru] = 0; | |
2583 | nr[lru + LRU_ACTIVE] = 0; | |
2584 | ||
2585 | /* | |
2586 | * Recalculate the other LRU scan count based on its original | |
2587 | * scan target and the percentage scanning already complete | |
2588 | */ | |
2589 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
2590 | nr_scanned = targets[lru] - nr[lru]; | |
2591 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2592 | nr[lru] -= min(nr[lru], nr_scanned); | |
2593 | ||
2594 | lru += LRU_ACTIVE; | |
2595 | nr_scanned = targets[lru] - nr[lru]; | |
2596 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
2597 | nr[lru] -= min(nr[lru], nr_scanned); | |
2598 | ||
2599 | scan_adjusted = true; | |
9b4f98cd JW |
2600 | } |
2601 | blk_finish_plug(&plug); | |
2602 | sc->nr_reclaimed += nr_reclaimed; | |
2603 | ||
2604 | /* | |
2605 | * Even if we did not try to evict anon pages at all, we want to | |
2606 | * rebalance the anon lru active/inactive ratio. | |
2607 | */ | |
a1086291 | 2608 | if (total_swap_pages && inactive_list_is_low(lruvec, false, sc, true)) |
9b4f98cd JW |
2609 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
2610 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
2611 | } |
2612 | ||
23b9da55 | 2613 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 2614 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 2615 | { |
d84da3f9 | 2616 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 2617 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 2618 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
2619 | return true; |
2620 | ||
2621 | return false; | |
2622 | } | |
2623 | ||
3e7d3449 | 2624 | /* |
23b9da55 MG |
2625 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
2626 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
2627 | * true if more pages should be reclaimed such that when the page allocator | |
2628 | * calls try_to_compact_zone() that it will have enough free pages to succeed. | |
2629 | * It will give up earlier than that if there is difficulty reclaiming pages. | |
3e7d3449 | 2630 | */ |
a9dd0a83 | 2631 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 | 2632 | unsigned long nr_reclaimed, |
3e7d3449 MG |
2633 | struct scan_control *sc) |
2634 | { | |
2635 | unsigned long pages_for_compaction; | |
2636 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 2637 | int z; |
3e7d3449 MG |
2638 | |
2639 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 2640 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
2641 | return false; |
2642 | ||
5ee04716 VB |
2643 | /* |
2644 | * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX | |
2645 | * number of pages that were scanned. This will return to the caller | |
2646 | * with the risk reclaim/compaction and the resulting allocation attempt | |
2647 | * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL | |
2648 | * allocations through requiring that the full LRU list has been scanned | |
2649 | * first, by assuming that zero delta of sc->nr_scanned means full LRU | |
2650 | * scan, but that approximation was wrong, and there were corner cases | |
2651 | * where always a non-zero amount of pages were scanned. | |
2652 | */ | |
2653 | if (!nr_reclaimed) | |
2654 | return false; | |
3e7d3449 | 2655 | |
3e7d3449 | 2656 | /* If compaction would go ahead or the allocation would succeed, stop */ |
a9dd0a83 MG |
2657 | for (z = 0; z <= sc->reclaim_idx; z++) { |
2658 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 2659 | if (!managed_zone(zone)) |
a9dd0a83 MG |
2660 | continue; |
2661 | ||
2662 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 2663 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
2664 | case COMPACT_CONTINUE: |
2665 | return false; | |
2666 | default: | |
2667 | /* check next zone */ | |
2668 | ; | |
2669 | } | |
3e7d3449 | 2670 | } |
1c6c1597 HD |
2671 | |
2672 | /* | |
2673 | * If we have not reclaimed enough pages for compaction and the | |
2674 | * inactive lists are large enough, continue reclaiming | |
2675 | */ | |
2676 | pages_for_compaction = compact_gap(sc->order); | |
2677 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); | |
2678 | if (get_nr_swap_pages() > 0) | |
2679 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); | |
2680 | ||
5ee04716 | 2681 | return inactive_lru_pages > pages_for_compaction; |
3e7d3449 MG |
2682 | } |
2683 | ||
0f6a5cff | 2684 | static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 2685 | { |
0f6a5cff | 2686 | struct mem_cgroup *target_memcg = sc->target_mem_cgroup; |
d2af3397 | 2687 | struct mem_cgroup *memcg; |
1da177e4 | 2688 | |
0f6a5cff | 2689 | memcg = mem_cgroup_iter(target_memcg, NULL, NULL); |
d2af3397 | 2690 | do { |
afaf07a6 | 2691 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
d2af3397 JW |
2692 | unsigned long reclaimed; |
2693 | unsigned long scanned; | |
5660048c | 2694 | |
0f6a5cff | 2695 | switch (mem_cgroup_protected(target_memcg, memcg)) { |
d2af3397 JW |
2696 | case MEMCG_PROT_MIN: |
2697 | /* | |
2698 | * Hard protection. | |
2699 | * If there is no reclaimable memory, OOM. | |
2700 | */ | |
2701 | continue; | |
2702 | case MEMCG_PROT_LOW: | |
2703 | /* | |
2704 | * Soft protection. | |
2705 | * Respect the protection only as long as | |
2706 | * there is an unprotected supply | |
2707 | * of reclaimable memory from other cgroups. | |
2708 | */ | |
2709 | if (!sc->memcg_low_reclaim) { | |
2710 | sc->memcg_low_skipped = 1; | |
bf8d5d52 | 2711 | continue; |
241994ed | 2712 | } |
d2af3397 JW |
2713 | memcg_memory_event(memcg, MEMCG_LOW); |
2714 | break; | |
2715 | case MEMCG_PROT_NONE: | |
2716 | /* | |
2717 | * All protection thresholds breached. We may | |
2718 | * still choose to vary the scan pressure | |
2719 | * applied based on by how much the cgroup in | |
2720 | * question has exceeded its protection | |
2721 | * thresholds (see get_scan_count). | |
2722 | */ | |
2723 | break; | |
2724 | } | |
241994ed | 2725 | |
d2af3397 JW |
2726 | reclaimed = sc->nr_reclaimed; |
2727 | scanned = sc->nr_scanned; | |
afaf07a6 JW |
2728 | |
2729 | shrink_lruvec(lruvec, sc); | |
70ddf637 | 2730 | |
d2af3397 JW |
2731 | shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, |
2732 | sc->priority); | |
6b4f7799 | 2733 | |
d2af3397 JW |
2734 | /* Record the group's reclaim efficiency */ |
2735 | vmpressure(sc->gfp_mask, memcg, false, | |
2736 | sc->nr_scanned - scanned, | |
2737 | sc->nr_reclaimed - reclaimed); | |
70ddf637 | 2738 | |
0f6a5cff JW |
2739 | } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL))); |
2740 | } | |
2741 | ||
2742 | static bool shrink_node(pg_data_t *pgdat, struct scan_control *sc) | |
2743 | { | |
2744 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
0f6a5cff | 2745 | unsigned long nr_reclaimed, nr_scanned; |
1b05117d | 2746 | struct lruvec *target_lruvec; |
0f6a5cff JW |
2747 | bool reclaimable = false; |
2748 | ||
1b05117d JW |
2749 | target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat); |
2750 | ||
0f6a5cff JW |
2751 | again: |
2752 | memset(&sc->nr, 0, sizeof(sc->nr)); | |
2753 | ||
2754 | nr_reclaimed = sc->nr_reclaimed; | |
2755 | nr_scanned = sc->nr_scanned; | |
2756 | ||
2757 | shrink_node_memcgs(pgdat, sc); | |
2344d7e4 | 2758 | |
d2af3397 JW |
2759 | if (reclaim_state) { |
2760 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
2761 | reclaim_state->reclaimed_slab = 0; | |
2762 | } | |
d108c772 | 2763 | |
d2af3397 | 2764 | /* Record the subtree's reclaim efficiency */ |
1b05117d | 2765 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, |
d2af3397 JW |
2766 | sc->nr_scanned - nr_scanned, |
2767 | sc->nr_reclaimed - nr_reclaimed); | |
d108c772 | 2768 | |
d2af3397 JW |
2769 | if (sc->nr_reclaimed - nr_reclaimed) |
2770 | reclaimable = true; | |
d108c772 | 2771 | |
d2af3397 JW |
2772 | if (current_is_kswapd()) { |
2773 | /* | |
2774 | * If reclaim is isolating dirty pages under writeback, | |
2775 | * it implies that the long-lived page allocation rate | |
2776 | * is exceeding the page laundering rate. Either the | |
2777 | * global limits are not being effective at throttling | |
2778 | * processes due to the page distribution throughout | |
2779 | * zones or there is heavy usage of a slow backing | |
2780 | * device. The only option is to throttle from reclaim | |
2781 | * context which is not ideal as there is no guarantee | |
2782 | * the dirtying process is throttled in the same way | |
2783 | * balance_dirty_pages() manages. | |
2784 | * | |
2785 | * Once a node is flagged PGDAT_WRITEBACK, kswapd will | |
2786 | * count the number of pages under pages flagged for | |
2787 | * immediate reclaim and stall if any are encountered | |
2788 | * in the nr_immediate check below. | |
2789 | */ | |
2790 | if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken) | |
2791 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
d108c772 | 2792 | |
d2af3397 JW |
2793 | /* Allow kswapd to start writing pages during reclaim.*/ |
2794 | if (sc->nr.unqueued_dirty == sc->nr.file_taken) | |
2795 | set_bit(PGDAT_DIRTY, &pgdat->flags); | |
e3c1ac58 | 2796 | |
d108c772 | 2797 | /* |
d2af3397 JW |
2798 | * If kswapd scans pages marked marked for immediate |
2799 | * reclaim and under writeback (nr_immediate), it | |
2800 | * implies that pages are cycling through the LRU | |
2801 | * faster than they are written so also forcibly stall. | |
d108c772 | 2802 | */ |
d2af3397 JW |
2803 | if (sc->nr.immediate) |
2804 | congestion_wait(BLK_RW_ASYNC, HZ/10); | |
2805 | } | |
2806 | ||
2807 | /* | |
1b05117d JW |
2808 | * Tag a node/memcg as congested if all the dirty pages |
2809 | * scanned were backed by a congested BDI and | |
2810 | * wait_iff_congested will stall. | |
2811 | * | |
d2af3397 JW |
2812 | * Legacy memcg will stall in page writeback so avoid forcibly |
2813 | * stalling in wait_iff_congested(). | |
2814 | */ | |
1b05117d JW |
2815 | if ((current_is_kswapd() || |
2816 | (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) && | |
d2af3397 | 2817 | sc->nr.dirty && sc->nr.dirty == sc->nr.congested) |
1b05117d | 2818 | set_bit(LRUVEC_CONGESTED, &target_lruvec->flags); |
d2af3397 JW |
2819 | |
2820 | /* | |
2821 | * Stall direct reclaim for IO completions if underlying BDIs | |
2822 | * and node is congested. Allow kswapd to continue until it | |
2823 | * starts encountering unqueued dirty pages or cycling through | |
2824 | * the LRU too quickly. | |
2825 | */ | |
1b05117d JW |
2826 | if (!current_is_kswapd() && current_may_throttle() && |
2827 | !sc->hibernation_mode && | |
2828 | test_bit(LRUVEC_CONGESTED, &target_lruvec->flags)) | |
d2af3397 | 2829 | wait_iff_congested(BLK_RW_ASYNC, HZ/10); |
d108c772 | 2830 | |
d2af3397 JW |
2831 | if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
2832 | sc)) | |
2833 | goto again; | |
2344d7e4 | 2834 | |
c73322d0 JW |
2835 | /* |
2836 | * Kswapd gives up on balancing particular nodes after too | |
2837 | * many failures to reclaim anything from them and goes to | |
2838 | * sleep. On reclaim progress, reset the failure counter. A | |
2839 | * successful direct reclaim run will revive a dormant kswapd. | |
2840 | */ | |
2841 | if (reclaimable) | |
2842 | pgdat->kswapd_failures = 0; | |
2843 | ||
2344d7e4 | 2844 | return reclaimable; |
f16015fb JW |
2845 | } |
2846 | ||
53853e2d | 2847 | /* |
fdd4c614 VB |
2848 | * Returns true if compaction should go ahead for a costly-order request, or |
2849 | * the allocation would already succeed without compaction. Return false if we | |
2850 | * should reclaim first. | |
53853e2d | 2851 | */ |
4f588331 | 2852 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 2853 | { |
31483b6a | 2854 | unsigned long watermark; |
fdd4c614 | 2855 | enum compact_result suitable; |
fe4b1b24 | 2856 | |
fdd4c614 VB |
2857 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
2858 | if (suitable == COMPACT_SUCCESS) | |
2859 | /* Allocation should succeed already. Don't reclaim. */ | |
2860 | return true; | |
2861 | if (suitable == COMPACT_SKIPPED) | |
2862 | /* Compaction cannot yet proceed. Do reclaim. */ | |
2863 | return false; | |
fe4b1b24 | 2864 | |
53853e2d | 2865 | /* |
fdd4c614 VB |
2866 | * Compaction is already possible, but it takes time to run and there |
2867 | * are potentially other callers using the pages just freed. So proceed | |
2868 | * with reclaim to make a buffer of free pages available to give | |
2869 | * compaction a reasonable chance of completing and allocating the page. | |
2870 | * Note that we won't actually reclaim the whole buffer in one attempt | |
2871 | * as the target watermark in should_continue_reclaim() is lower. But if | |
2872 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 2873 | */ |
fdd4c614 | 2874 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 2875 | |
fdd4c614 | 2876 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
2877 | } |
2878 | ||
1da177e4 LT |
2879 | /* |
2880 | * This is the direct reclaim path, for page-allocating processes. We only | |
2881 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
2882 | * request. | |
2883 | * | |
1da177e4 LT |
2884 | * If a zone is deemed to be full of pinned pages then just give it a light |
2885 | * scan then give up on it. | |
2886 | */ | |
0a0337e0 | 2887 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 2888 | { |
dd1a239f | 2889 | struct zoneref *z; |
54a6eb5c | 2890 | struct zone *zone; |
0608f43d AM |
2891 | unsigned long nr_soft_reclaimed; |
2892 | unsigned long nr_soft_scanned; | |
619d0d76 | 2893 | gfp_t orig_mask; |
79dafcdc | 2894 | pg_data_t *last_pgdat = NULL; |
1cfb419b | 2895 | |
cc715d99 MG |
2896 | /* |
2897 | * If the number of buffer_heads in the machine exceeds the maximum | |
2898 | * allowed level, force direct reclaim to scan the highmem zone as | |
2899 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
2900 | */ | |
619d0d76 | 2901 | orig_mask = sc->gfp_mask; |
b2e18757 | 2902 | if (buffer_heads_over_limit) { |
cc715d99 | 2903 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 2904 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 2905 | } |
cc715d99 | 2906 | |
d4debc66 | 2907 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 2908 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
2909 | /* |
2910 | * Take care memory controller reclaiming has small influence | |
2911 | * to global LRU. | |
2912 | */ | |
b5ead35e | 2913 | if (!cgroup_reclaim(sc)) { |
344736f2 VD |
2914 | if (!cpuset_zone_allowed(zone, |
2915 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 2916 | continue; |
65ec02cb | 2917 | |
0b06496a JW |
2918 | /* |
2919 | * If we already have plenty of memory free for | |
2920 | * compaction in this zone, don't free any more. | |
2921 | * Even though compaction is invoked for any | |
2922 | * non-zero order, only frequent costly order | |
2923 | * reclamation is disruptive enough to become a | |
2924 | * noticeable problem, like transparent huge | |
2925 | * page allocations. | |
2926 | */ | |
2927 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
2928 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 2929 | compaction_ready(zone, sc)) { |
0b06496a JW |
2930 | sc->compaction_ready = true; |
2931 | continue; | |
e0887c19 | 2932 | } |
0b06496a | 2933 | |
79dafcdc MG |
2934 | /* |
2935 | * Shrink each node in the zonelist once. If the | |
2936 | * zonelist is ordered by zone (not the default) then a | |
2937 | * node may be shrunk multiple times but in that case | |
2938 | * the user prefers lower zones being preserved. | |
2939 | */ | |
2940 | if (zone->zone_pgdat == last_pgdat) | |
2941 | continue; | |
2942 | ||
0608f43d AM |
2943 | /* |
2944 | * This steals pages from memory cgroups over softlimit | |
2945 | * and returns the number of reclaimed pages and | |
2946 | * scanned pages. This works for global memory pressure | |
2947 | * and balancing, not for a memcg's limit. | |
2948 | */ | |
2949 | nr_soft_scanned = 0; | |
ef8f2327 | 2950 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
2951 | sc->order, sc->gfp_mask, |
2952 | &nr_soft_scanned); | |
2953 | sc->nr_reclaimed += nr_soft_reclaimed; | |
2954 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 2955 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 2956 | } |
408d8544 | 2957 | |
79dafcdc MG |
2958 | /* See comment about same check for global reclaim above */ |
2959 | if (zone->zone_pgdat == last_pgdat) | |
2960 | continue; | |
2961 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 2962 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 2963 | } |
e0c23279 | 2964 | |
619d0d76 WY |
2965 | /* |
2966 | * Restore to original mask to avoid the impact on the caller if we | |
2967 | * promoted it to __GFP_HIGHMEM. | |
2968 | */ | |
2969 | sc->gfp_mask = orig_mask; | |
1da177e4 | 2970 | } |
4f98a2fe | 2971 | |
2a2e4885 JW |
2972 | static void snapshot_refaults(struct mem_cgroup *root_memcg, pg_data_t *pgdat) |
2973 | { | |
2974 | struct mem_cgroup *memcg; | |
2975 | ||
2976 | memcg = mem_cgroup_iter(root_memcg, NULL, NULL); | |
2977 | do { | |
2978 | unsigned long refaults; | |
2979 | struct lruvec *lruvec; | |
2980 | ||
867e5e1d | 2981 | lruvec = mem_cgroup_lruvec(memcg, pgdat); |
205b20cc | 2982 | refaults = lruvec_page_state_local(lruvec, WORKINGSET_ACTIVATE); |
2a2e4885 JW |
2983 | lruvec->refaults = refaults; |
2984 | } while ((memcg = mem_cgroup_iter(root_memcg, memcg, NULL))); | |
2985 | } | |
2986 | ||
1da177e4 LT |
2987 | /* |
2988 | * This is the main entry point to direct page reclaim. | |
2989 | * | |
2990 | * If a full scan of the inactive list fails to free enough memory then we | |
2991 | * are "out of memory" and something needs to be killed. | |
2992 | * | |
2993 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
2994 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
2995 | * caller can't do much about. We kick the writeback threads and take explicit |
2996 | * naps in the hope that some of these pages can be written. But if the | |
2997 | * allocating task holds filesystem locks which prevent writeout this might not | |
2998 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
2999 | * |
3000 | * returns: 0, if no pages reclaimed | |
3001 | * else, the number of pages reclaimed | |
1da177e4 | 3002 | */ |
dac1d27b | 3003 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 3004 | struct scan_control *sc) |
1da177e4 | 3005 | { |
241994ed | 3006 | int initial_priority = sc->priority; |
2a2e4885 JW |
3007 | pg_data_t *last_pgdat; |
3008 | struct zoneref *z; | |
3009 | struct zone *zone; | |
241994ed | 3010 | retry: |
873b4771 KK |
3011 | delayacct_freepages_start(); |
3012 | ||
b5ead35e | 3013 | if (!cgroup_reclaim(sc)) |
7cc30fcf | 3014 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 3015 | |
9e3b2f8c | 3016 | do { |
70ddf637 AV |
3017 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
3018 | sc->priority); | |
66e1707b | 3019 | sc->nr_scanned = 0; |
0a0337e0 | 3020 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 3021 | |
bb21c7ce | 3022 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
3023 | break; |
3024 | ||
3025 | if (sc->compaction_ready) | |
3026 | break; | |
1da177e4 | 3027 | |
0e50ce3b MK |
3028 | /* |
3029 | * If we're getting trouble reclaiming, start doing | |
3030 | * writepage even in laptop mode. | |
3031 | */ | |
3032 | if (sc->priority < DEF_PRIORITY - 2) | |
3033 | sc->may_writepage = 1; | |
0b06496a | 3034 | } while (--sc->priority >= 0); |
bb21c7ce | 3035 | |
2a2e4885 JW |
3036 | last_pgdat = NULL; |
3037 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
3038 | sc->nodemask) { | |
3039 | if (zone->zone_pgdat == last_pgdat) | |
3040 | continue; | |
3041 | last_pgdat = zone->zone_pgdat; | |
1b05117d | 3042 | |
2a2e4885 | 3043 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); |
1b05117d JW |
3044 | |
3045 | if (cgroup_reclaim(sc)) { | |
3046 | struct lruvec *lruvec; | |
3047 | ||
3048 | lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, | |
3049 | zone->zone_pgdat); | |
3050 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
3051 | } | |
2a2e4885 JW |
3052 | } |
3053 | ||
873b4771 KK |
3054 | delayacct_freepages_end(); |
3055 | ||
bb21c7ce KM |
3056 | if (sc->nr_reclaimed) |
3057 | return sc->nr_reclaimed; | |
3058 | ||
0cee34fd | 3059 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 3060 | if (sc->compaction_ready) |
7335084d MG |
3061 | return 1; |
3062 | ||
241994ed | 3063 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
d6622f63 | 3064 | if (sc->memcg_low_skipped) { |
241994ed | 3065 | sc->priority = initial_priority; |
d6622f63 YX |
3066 | sc->memcg_low_reclaim = 1; |
3067 | sc->memcg_low_skipped = 0; | |
241994ed JW |
3068 | goto retry; |
3069 | } | |
3070 | ||
bb21c7ce | 3071 | return 0; |
1da177e4 LT |
3072 | } |
3073 | ||
c73322d0 | 3074 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
5515061d MG |
3075 | { |
3076 | struct zone *zone; | |
3077 | unsigned long pfmemalloc_reserve = 0; | |
3078 | unsigned long free_pages = 0; | |
3079 | int i; | |
3080 | bool wmark_ok; | |
3081 | ||
c73322d0 JW |
3082 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
3083 | return true; | |
3084 | ||
5515061d MG |
3085 | for (i = 0; i <= ZONE_NORMAL; i++) { |
3086 | zone = &pgdat->node_zones[i]; | |
d450abd8 JW |
3087 | if (!managed_zone(zone)) |
3088 | continue; | |
3089 | ||
3090 | if (!zone_reclaimable_pages(zone)) | |
675becce MG |
3091 | continue; |
3092 | ||
5515061d MG |
3093 | pfmemalloc_reserve += min_wmark_pages(zone); |
3094 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
3095 | } | |
3096 | ||
675becce MG |
3097 | /* If there are no reserves (unexpected config) then do not throttle */ |
3098 | if (!pfmemalloc_reserve) | |
3099 | return true; | |
3100 | ||
5515061d MG |
3101 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
3102 | ||
3103 | /* kswapd must be awake if processes are being throttled */ | |
3104 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
38087d9b | 3105 | pgdat->kswapd_classzone_idx = min(pgdat->kswapd_classzone_idx, |
5515061d MG |
3106 | (enum zone_type)ZONE_NORMAL); |
3107 | wake_up_interruptible(&pgdat->kswapd_wait); | |
3108 | } | |
3109 | ||
3110 | return wmark_ok; | |
3111 | } | |
3112 | ||
3113 | /* | |
3114 | * Throttle direct reclaimers if backing storage is backed by the network | |
3115 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
3116 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
3117 | * when the low watermark is reached. |
3118 | * | |
3119 | * Returns true if a fatal signal was delivered during throttling. If this | |
3120 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 3121 | */ |
50694c28 | 3122 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
3123 | nodemask_t *nodemask) |
3124 | { | |
675becce | 3125 | struct zoneref *z; |
5515061d | 3126 | struct zone *zone; |
675becce | 3127 | pg_data_t *pgdat = NULL; |
5515061d MG |
3128 | |
3129 | /* | |
3130 | * Kernel threads should not be throttled as they may be indirectly | |
3131 | * responsible for cleaning pages necessary for reclaim to make forward | |
3132 | * progress. kjournald for example may enter direct reclaim while | |
3133 | * committing a transaction where throttling it could forcing other | |
3134 | * processes to block on log_wait_commit(). | |
3135 | */ | |
3136 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
3137 | goto out; |
3138 | ||
3139 | /* | |
3140 | * If a fatal signal is pending, this process should not throttle. | |
3141 | * It should return quickly so it can exit and free its memory | |
3142 | */ | |
3143 | if (fatal_signal_pending(current)) | |
3144 | goto out; | |
5515061d | 3145 | |
675becce MG |
3146 | /* |
3147 | * Check if the pfmemalloc reserves are ok by finding the first node | |
3148 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
3149 | * GFP_KERNEL will be required for allocating network buffers when | |
3150 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
3151 | * | |
3152 | * Throttling is based on the first usable node and throttled processes | |
3153 | * wait on a queue until kswapd makes progress and wakes them. There | |
3154 | * is an affinity then between processes waking up and where reclaim | |
3155 | * progress has been made assuming the process wakes on the same node. | |
3156 | * More importantly, processes running on remote nodes will not compete | |
3157 | * for remote pfmemalloc reserves and processes on different nodes | |
3158 | * should make reasonable progress. | |
3159 | */ | |
3160 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 3161 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
3162 | if (zone_idx(zone) > ZONE_NORMAL) |
3163 | continue; | |
3164 | ||
3165 | /* Throttle based on the first usable node */ | |
3166 | pgdat = zone->zone_pgdat; | |
c73322d0 | 3167 | if (allow_direct_reclaim(pgdat)) |
675becce MG |
3168 | goto out; |
3169 | break; | |
3170 | } | |
3171 | ||
3172 | /* If no zone was usable by the allocation flags then do not throttle */ | |
3173 | if (!pgdat) | |
50694c28 | 3174 | goto out; |
5515061d | 3175 | |
68243e76 MG |
3176 | /* Account for the throttling */ |
3177 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
3178 | ||
5515061d MG |
3179 | /* |
3180 | * If the caller cannot enter the filesystem, it's possible that it | |
3181 | * is due to the caller holding an FS lock or performing a journal | |
3182 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
3183 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
3184 | * blocked waiting on the same lock. Instead, throttle for up to a | |
3185 | * second before continuing. | |
3186 | */ | |
3187 | if (!(gfp_mask & __GFP_FS)) { | |
3188 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, | |
c73322d0 | 3189 | allow_direct_reclaim(pgdat), HZ); |
50694c28 MG |
3190 | |
3191 | goto check_pending; | |
5515061d MG |
3192 | } |
3193 | ||
3194 | /* Throttle until kswapd wakes the process */ | |
3195 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
c73322d0 | 3196 | allow_direct_reclaim(pgdat)); |
50694c28 MG |
3197 | |
3198 | check_pending: | |
3199 | if (fatal_signal_pending(current)) | |
3200 | return true; | |
3201 | ||
3202 | out: | |
3203 | return false; | |
5515061d MG |
3204 | } |
3205 | ||
dac1d27b | 3206 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 3207 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 3208 | { |
33906bc5 | 3209 | unsigned long nr_reclaimed; |
66e1707b | 3210 | struct scan_control sc = { |
ee814fe2 | 3211 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
f2f43e56 | 3212 | .gfp_mask = current_gfp_context(gfp_mask), |
b2e18757 | 3213 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
3214 | .order = order, |
3215 | .nodemask = nodemask, | |
3216 | .priority = DEF_PRIORITY, | |
66e1707b | 3217 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3218 | .may_unmap = 1, |
2e2e4259 | 3219 | .may_swap = 1, |
66e1707b BS |
3220 | }; |
3221 | ||
bb451fdf GT |
3222 | /* |
3223 | * scan_control uses s8 fields for order, priority, and reclaim_idx. | |
3224 | * Confirm they are large enough for max values. | |
3225 | */ | |
3226 | BUILD_BUG_ON(MAX_ORDER > S8_MAX); | |
3227 | BUILD_BUG_ON(DEF_PRIORITY > S8_MAX); | |
3228 | BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX); | |
3229 | ||
5515061d | 3230 | /* |
50694c28 MG |
3231 | * Do not enter reclaim if fatal signal was delivered while throttled. |
3232 | * 1 is returned so that the page allocator does not OOM kill at this | |
3233 | * point. | |
5515061d | 3234 | */ |
f2f43e56 | 3235 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
5515061d MG |
3236 | return 1; |
3237 | ||
1732d2b0 | 3238 | set_task_reclaim_state(current, &sc.reclaim_state); |
3481c37f | 3239 | trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask); |
33906bc5 | 3240 | |
3115cd91 | 3241 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
3242 | |
3243 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
1732d2b0 | 3244 | set_task_reclaim_state(current, NULL); |
33906bc5 MG |
3245 | |
3246 | return nr_reclaimed; | |
66e1707b BS |
3247 | } |
3248 | ||
c255a458 | 3249 | #ifdef CONFIG_MEMCG |
66e1707b | 3250 | |
d2e5fb92 | 3251 | /* Only used by soft limit reclaim. Do not reuse for anything else. */ |
a9dd0a83 | 3252 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 3253 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 3254 | pg_data_t *pgdat, |
0ae5e89c | 3255 | unsigned long *nr_scanned) |
4e416953 | 3256 | { |
afaf07a6 | 3257 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
4e416953 | 3258 | struct scan_control sc = { |
b8f5c566 | 3259 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 3260 | .target_mem_cgroup = memcg, |
4e416953 BS |
3261 | .may_writepage = !laptop_mode, |
3262 | .may_unmap = 1, | |
b2e18757 | 3263 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 3264 | .may_swap = !noswap, |
4e416953 | 3265 | }; |
0ae5e89c | 3266 | |
d2e5fb92 MH |
3267 | WARN_ON_ONCE(!current->reclaim_state); |
3268 | ||
4e416953 BS |
3269 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3270 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 3271 | |
9e3b2f8c | 3272 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
3481c37f | 3273 | sc.gfp_mask); |
bdce6d9e | 3274 | |
4e416953 BS |
3275 | /* |
3276 | * NOTE: Although we can get the priority field, using it | |
3277 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 3278 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
3279 | * will pick up pages from other mem cgroup's as well. We hack |
3280 | * the priority and make it zero. | |
3281 | */ | |
afaf07a6 | 3282 | shrink_lruvec(lruvec, &sc); |
bdce6d9e KM |
3283 | |
3284 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
3285 | ||
0ae5e89c | 3286 | *nr_scanned = sc.nr_scanned; |
0308f7cf | 3287 | |
4e416953 BS |
3288 | return sc.nr_reclaimed; |
3289 | } | |
3290 | ||
72835c86 | 3291 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 3292 | unsigned long nr_pages, |
a7885eb8 | 3293 | gfp_t gfp_mask, |
b70a2a21 | 3294 | bool may_swap) |
66e1707b | 3295 | { |
bdce6d9e | 3296 | unsigned long nr_reclaimed; |
eb414681 | 3297 | unsigned long pflags; |
499118e9 | 3298 | unsigned int noreclaim_flag; |
66e1707b | 3299 | struct scan_control sc = { |
b70a2a21 | 3300 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 3301 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
a09ed5e0 | 3302 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
b2e18757 | 3303 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
3304 | .target_mem_cgroup = memcg, |
3305 | .priority = DEF_PRIORITY, | |
3306 | .may_writepage = !laptop_mode, | |
3307 | .may_unmap = 1, | |
b70a2a21 | 3308 | .may_swap = may_swap, |
a09ed5e0 | 3309 | }; |
889976db | 3310 | /* |
fa40d1ee SB |
3311 | * Traverse the ZONELIST_FALLBACK zonelist of the current node to put |
3312 | * equal pressure on all the nodes. This is based on the assumption that | |
3313 | * the reclaim does not bail out early. | |
889976db | 3314 | */ |
fa40d1ee | 3315 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
889976db | 3316 | |
fa40d1ee | 3317 | set_task_reclaim_state(current, &sc.reclaim_state); |
bdce6d9e | 3318 | |
3481c37f | 3319 | trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask); |
bdce6d9e | 3320 | |
eb414681 | 3321 | psi_memstall_enter(&pflags); |
499118e9 | 3322 | noreclaim_flag = memalloc_noreclaim_save(); |
eb414681 | 3323 | |
3115cd91 | 3324 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
eb414681 | 3325 | |
499118e9 | 3326 | memalloc_noreclaim_restore(noreclaim_flag); |
eb414681 | 3327 | psi_memstall_leave(&pflags); |
bdce6d9e KM |
3328 | |
3329 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); | |
1732d2b0 | 3330 | set_task_reclaim_state(current, NULL); |
bdce6d9e KM |
3331 | |
3332 | return nr_reclaimed; | |
66e1707b BS |
3333 | } |
3334 | #endif | |
3335 | ||
1d82de61 | 3336 | static void age_active_anon(struct pglist_data *pgdat, |
ef8f2327 | 3337 | struct scan_control *sc) |
f16015fb | 3338 | { |
b95a2f2d | 3339 | struct mem_cgroup *memcg; |
f16015fb | 3340 | |
b95a2f2d JW |
3341 | if (!total_swap_pages) |
3342 | return; | |
3343 | ||
3344 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
3345 | do { | |
867e5e1d | 3346 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
b95a2f2d | 3347 | |
3b991208 | 3348 | if (inactive_list_is_low(lruvec, false, sc, true)) |
1a93be0e | 3349 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
9e3b2f8c | 3350 | sc, LRU_ACTIVE_ANON); |
b95a2f2d JW |
3351 | |
3352 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
3353 | } while (memcg); | |
f16015fb JW |
3354 | } |
3355 | ||
1c30844d MG |
3356 | static bool pgdat_watermark_boosted(pg_data_t *pgdat, int classzone_idx) |
3357 | { | |
3358 | int i; | |
3359 | struct zone *zone; | |
3360 | ||
3361 | /* | |
3362 | * Check for watermark boosts top-down as the higher zones | |
3363 | * are more likely to be boosted. Both watermarks and boosts | |
3364 | * should not be checked at the time time as reclaim would | |
3365 | * start prematurely when there is no boosting and a lower | |
3366 | * zone is balanced. | |
3367 | */ | |
3368 | for (i = classzone_idx; i >= 0; i--) { | |
3369 | zone = pgdat->node_zones + i; | |
3370 | if (!managed_zone(zone)) | |
3371 | continue; | |
3372 | ||
3373 | if (zone->watermark_boost) | |
3374 | return true; | |
3375 | } | |
3376 | ||
3377 | return false; | |
3378 | } | |
3379 | ||
e716f2eb MG |
3380 | /* |
3381 | * Returns true if there is an eligible zone balanced for the request order | |
3382 | * and classzone_idx | |
3383 | */ | |
3384 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int classzone_idx) | |
60cefed4 | 3385 | { |
e716f2eb MG |
3386 | int i; |
3387 | unsigned long mark = -1; | |
3388 | struct zone *zone; | |
60cefed4 | 3389 | |
1c30844d MG |
3390 | /* |
3391 | * Check watermarks bottom-up as lower zones are more likely to | |
3392 | * meet watermarks. | |
3393 | */ | |
e716f2eb MG |
3394 | for (i = 0; i <= classzone_idx; i++) { |
3395 | zone = pgdat->node_zones + i; | |
6256c6b4 | 3396 | |
e716f2eb MG |
3397 | if (!managed_zone(zone)) |
3398 | continue; | |
3399 | ||
3400 | mark = high_wmark_pages(zone); | |
3401 | if (zone_watermark_ok_safe(zone, order, mark, classzone_idx)) | |
3402 | return true; | |
3403 | } | |
3404 | ||
3405 | /* | |
3406 | * If a node has no populated zone within classzone_idx, it does not | |
3407 | * need balancing by definition. This can happen if a zone-restricted | |
3408 | * allocation tries to wake a remote kswapd. | |
3409 | */ | |
3410 | if (mark == -1) | |
3411 | return true; | |
3412 | ||
3413 | return false; | |
60cefed4 JW |
3414 | } |
3415 | ||
631b6e08 MG |
3416 | /* Clear pgdat state for congested, dirty or under writeback. */ |
3417 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
3418 | { | |
1b05117d JW |
3419 | struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat); |
3420 | ||
3421 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
631b6e08 MG |
3422 | clear_bit(PGDAT_DIRTY, &pgdat->flags); |
3423 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
3424 | } | |
3425 | ||
5515061d MG |
3426 | /* |
3427 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
3428 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
3429 | * | |
3430 | * Returns true if kswapd is ready to sleep | |
3431 | */ | |
d9f21d42 | 3432 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, int classzone_idx) |
f50de2d3 | 3433 | { |
5515061d | 3434 | /* |
9e5e3661 | 3435 | * The throttled processes are normally woken up in balance_pgdat() as |
c73322d0 | 3436 | * soon as allow_direct_reclaim() is true. But there is a potential |
9e5e3661 VB |
3437 | * race between when kswapd checks the watermarks and a process gets |
3438 | * throttled. There is also a potential race if processes get | |
3439 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
3440 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
3441 | * the wake up checks. If kswapd is going to sleep, no process should | |
3442 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
3443 | * the wake up is premature, processes will wake kswapd and get | |
3444 | * throttled again. The difference from wake ups in balance_pgdat() is | |
3445 | * that here we are under prepare_to_wait(). | |
5515061d | 3446 | */ |
9e5e3661 VB |
3447 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
3448 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 3449 | |
c73322d0 JW |
3450 | /* Hopeless node, leave it to direct reclaim */ |
3451 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
3452 | return true; | |
3453 | ||
e716f2eb MG |
3454 | if (pgdat_balanced(pgdat, order, classzone_idx)) { |
3455 | clear_pgdat_congested(pgdat); | |
3456 | return true; | |
1d82de61 MG |
3457 | } |
3458 | ||
333b0a45 | 3459 | return false; |
f50de2d3 MG |
3460 | } |
3461 | ||
75485363 | 3462 | /* |
1d82de61 MG |
3463 | * kswapd shrinks a node of pages that are at or below the highest usable |
3464 | * zone that is currently unbalanced. | |
b8e83b94 MG |
3465 | * |
3466 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
3467 | * reclaim or if the lack of progress was due to pages under writeback. |
3468 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 3469 | */ |
1d82de61 | 3470 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 3471 | struct scan_control *sc) |
75485363 | 3472 | { |
1d82de61 MG |
3473 | struct zone *zone; |
3474 | int z; | |
75485363 | 3475 | |
1d82de61 MG |
3476 | /* Reclaim a number of pages proportional to the number of zones */ |
3477 | sc->nr_to_reclaim = 0; | |
970a39a3 | 3478 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 3479 | zone = pgdat->node_zones + z; |
6aa303de | 3480 | if (!managed_zone(zone)) |
1d82de61 | 3481 | continue; |
7c954f6d | 3482 | |
1d82de61 MG |
3483 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
3484 | } | |
7c954f6d MG |
3485 | |
3486 | /* | |
1d82de61 MG |
3487 | * Historically care was taken to put equal pressure on all zones but |
3488 | * now pressure is applied based on node LRU order. | |
7c954f6d | 3489 | */ |
970a39a3 | 3490 | shrink_node(pgdat, sc); |
283aba9f | 3491 | |
7c954f6d | 3492 | /* |
1d82de61 MG |
3493 | * Fragmentation may mean that the system cannot be rebalanced for |
3494 | * high-order allocations. If twice the allocation size has been | |
3495 | * reclaimed then recheck watermarks only at order-0 to prevent | |
3496 | * excessive reclaim. Assume that a process requested a high-order | |
3497 | * can direct reclaim/compact. | |
7c954f6d | 3498 | */ |
9861a62c | 3499 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 3500 | sc->order = 0; |
7c954f6d | 3501 | |
b8e83b94 | 3502 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
3503 | } |
3504 | ||
1da177e4 | 3505 | /* |
1d82de61 MG |
3506 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
3507 | * that are eligible for use by the caller until at least one zone is | |
3508 | * balanced. | |
1da177e4 | 3509 | * |
1d82de61 | 3510 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
3511 | * |
3512 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 3513 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
8bb4e7a2 | 3514 | * found to have free_pages <= high_wmark_pages(zone), any page in that zone |
1d82de61 MG |
3515 | * or lower is eligible for reclaim until at least one usable zone is |
3516 | * balanced. | |
1da177e4 | 3517 | */ |
accf6242 | 3518 | static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx) |
1da177e4 | 3519 | { |
1da177e4 | 3520 | int i; |
0608f43d AM |
3521 | unsigned long nr_soft_reclaimed; |
3522 | unsigned long nr_soft_scanned; | |
eb414681 | 3523 | unsigned long pflags; |
1c30844d MG |
3524 | unsigned long nr_boost_reclaim; |
3525 | unsigned long zone_boosts[MAX_NR_ZONES] = { 0, }; | |
3526 | bool boosted; | |
1d82de61 | 3527 | struct zone *zone; |
179e9639 AM |
3528 | struct scan_control sc = { |
3529 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 3530 | .order = order, |
a6dc60f8 | 3531 | .may_unmap = 1, |
179e9639 | 3532 | }; |
93781325 | 3533 | |
1732d2b0 | 3534 | set_task_reclaim_state(current, &sc.reclaim_state); |
eb414681 | 3535 | psi_memstall_enter(&pflags); |
93781325 OS |
3536 | __fs_reclaim_acquire(); |
3537 | ||
f8891e5e | 3538 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 3539 | |
1c30844d MG |
3540 | /* |
3541 | * Account for the reclaim boost. Note that the zone boost is left in | |
3542 | * place so that parallel allocations that are near the watermark will | |
3543 | * stall or direct reclaim until kswapd is finished. | |
3544 | */ | |
3545 | nr_boost_reclaim = 0; | |
3546 | for (i = 0; i <= classzone_idx; i++) { | |
3547 | zone = pgdat->node_zones + i; | |
3548 | if (!managed_zone(zone)) | |
3549 | continue; | |
3550 | ||
3551 | nr_boost_reclaim += zone->watermark_boost; | |
3552 | zone_boosts[i] = zone->watermark_boost; | |
3553 | } | |
3554 | boosted = nr_boost_reclaim; | |
3555 | ||
3556 | restart: | |
3557 | sc.priority = DEF_PRIORITY; | |
9e3b2f8c | 3558 | do { |
c73322d0 | 3559 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
b8e83b94 | 3560 | bool raise_priority = true; |
1c30844d | 3561 | bool balanced; |
93781325 | 3562 | bool ret; |
b8e83b94 | 3563 | |
84c7a777 | 3564 | sc.reclaim_idx = classzone_idx; |
1da177e4 | 3565 | |
86c79f6b | 3566 | /* |
84c7a777 MG |
3567 | * If the number of buffer_heads exceeds the maximum allowed |
3568 | * then consider reclaiming from all zones. This has a dual | |
3569 | * purpose -- on 64-bit systems it is expected that | |
3570 | * buffer_heads are stripped during active rotation. On 32-bit | |
3571 | * systems, highmem pages can pin lowmem memory and shrinking | |
3572 | * buffers can relieve lowmem pressure. Reclaim may still not | |
3573 | * go ahead if all eligible zones for the original allocation | |
3574 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
3575 | */ |
3576 | if (buffer_heads_over_limit) { | |
3577 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
3578 | zone = pgdat->node_zones + i; | |
6aa303de | 3579 | if (!managed_zone(zone)) |
86c79f6b | 3580 | continue; |
cc715d99 | 3581 | |
970a39a3 | 3582 | sc.reclaim_idx = i; |
e1dbeda6 | 3583 | break; |
1da177e4 | 3584 | } |
1da177e4 | 3585 | } |
dafcb73e | 3586 | |
86c79f6b | 3587 | /* |
1c30844d MG |
3588 | * If the pgdat is imbalanced then ignore boosting and preserve |
3589 | * the watermarks for a later time and restart. Note that the | |
3590 | * zone watermarks will be still reset at the end of balancing | |
3591 | * on the grounds that the normal reclaim should be enough to | |
3592 | * re-evaluate if boosting is required when kswapd next wakes. | |
3593 | */ | |
3594 | balanced = pgdat_balanced(pgdat, sc.order, classzone_idx); | |
3595 | if (!balanced && nr_boost_reclaim) { | |
3596 | nr_boost_reclaim = 0; | |
3597 | goto restart; | |
3598 | } | |
3599 | ||
3600 | /* | |
3601 | * If boosting is not active then only reclaim if there are no | |
3602 | * eligible zones. Note that sc.reclaim_idx is not used as | |
3603 | * buffer_heads_over_limit may have adjusted it. | |
86c79f6b | 3604 | */ |
1c30844d | 3605 | if (!nr_boost_reclaim && balanced) |
e716f2eb | 3606 | goto out; |
e1dbeda6 | 3607 | |
1c30844d MG |
3608 | /* Limit the priority of boosting to avoid reclaim writeback */ |
3609 | if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2) | |
3610 | raise_priority = false; | |
3611 | ||
3612 | /* | |
3613 | * Do not writeback or swap pages for boosted reclaim. The | |
3614 | * intent is to relieve pressure not issue sub-optimal IO | |
3615 | * from reclaim context. If no pages are reclaimed, the | |
3616 | * reclaim will be aborted. | |
3617 | */ | |
3618 | sc.may_writepage = !laptop_mode && !nr_boost_reclaim; | |
3619 | sc.may_swap = !nr_boost_reclaim; | |
1c30844d | 3620 | |
1d82de61 MG |
3621 | /* |
3622 | * Do some background aging of the anon list, to give | |
3623 | * pages a chance to be referenced before reclaiming. All | |
3624 | * pages are rotated regardless of classzone as this is | |
3625 | * about consistent aging. | |
3626 | */ | |
ef8f2327 | 3627 | age_active_anon(pgdat, &sc); |
1d82de61 | 3628 | |
b7ea3c41 MG |
3629 | /* |
3630 | * If we're getting trouble reclaiming, start doing writepage | |
3631 | * even in laptop mode. | |
3632 | */ | |
047d72c3 | 3633 | if (sc.priority < DEF_PRIORITY - 2) |
b7ea3c41 MG |
3634 | sc.may_writepage = 1; |
3635 | ||
1d82de61 MG |
3636 | /* Call soft limit reclaim before calling shrink_node. */ |
3637 | sc.nr_scanned = 0; | |
3638 | nr_soft_scanned = 0; | |
ef8f2327 | 3639 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
3640 | sc.gfp_mask, &nr_soft_scanned); |
3641 | sc.nr_reclaimed += nr_soft_reclaimed; | |
3642 | ||
1da177e4 | 3643 | /* |
1d82de61 MG |
3644 | * There should be no need to raise the scanning priority if |
3645 | * enough pages are already being scanned that that high | |
3646 | * watermark would be met at 100% efficiency. | |
1da177e4 | 3647 | */ |
970a39a3 | 3648 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 3649 | raise_priority = false; |
5515061d MG |
3650 | |
3651 | /* | |
3652 | * If the low watermark is met there is no need for processes | |
3653 | * to be throttled on pfmemalloc_wait as they should not be | |
3654 | * able to safely make forward progress. Wake them | |
3655 | */ | |
3656 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
c73322d0 | 3657 | allow_direct_reclaim(pgdat)) |
cfc51155 | 3658 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 3659 | |
b8e83b94 | 3660 | /* Check if kswapd should be suspending */ |
93781325 OS |
3661 | __fs_reclaim_release(); |
3662 | ret = try_to_freeze(); | |
3663 | __fs_reclaim_acquire(); | |
3664 | if (ret || kthread_should_stop()) | |
b8e83b94 | 3665 | break; |
8357376d | 3666 | |
73ce02e9 | 3667 | /* |
b8e83b94 MG |
3668 | * Raise priority if scanning rate is too low or there was no |
3669 | * progress in reclaiming pages | |
73ce02e9 | 3670 | */ |
c73322d0 | 3671 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
1c30844d MG |
3672 | nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed); |
3673 | ||
3674 | /* | |
3675 | * If reclaim made no progress for a boost, stop reclaim as | |
3676 | * IO cannot be queued and it could be an infinite loop in | |
3677 | * extreme circumstances. | |
3678 | */ | |
3679 | if (nr_boost_reclaim && !nr_reclaimed) | |
3680 | break; | |
3681 | ||
c73322d0 | 3682 | if (raise_priority || !nr_reclaimed) |
b8e83b94 | 3683 | sc.priority--; |
1d82de61 | 3684 | } while (sc.priority >= 1); |
1da177e4 | 3685 | |
c73322d0 JW |
3686 | if (!sc.nr_reclaimed) |
3687 | pgdat->kswapd_failures++; | |
3688 | ||
b8e83b94 | 3689 | out: |
1c30844d MG |
3690 | /* If reclaim was boosted, account for the reclaim done in this pass */ |
3691 | if (boosted) { | |
3692 | unsigned long flags; | |
3693 | ||
3694 | for (i = 0; i <= classzone_idx; i++) { | |
3695 | if (!zone_boosts[i]) | |
3696 | continue; | |
3697 | ||
3698 | /* Increments are under the zone lock */ | |
3699 | zone = pgdat->node_zones + i; | |
3700 | spin_lock_irqsave(&zone->lock, flags); | |
3701 | zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]); | |
3702 | spin_unlock_irqrestore(&zone->lock, flags); | |
3703 | } | |
3704 | ||
3705 | /* | |
3706 | * As there is now likely space, wakeup kcompact to defragment | |
3707 | * pageblocks. | |
3708 | */ | |
3709 | wakeup_kcompactd(pgdat, pageblock_order, classzone_idx); | |
3710 | } | |
3711 | ||
2a2e4885 | 3712 | snapshot_refaults(NULL, pgdat); |
93781325 | 3713 | __fs_reclaim_release(); |
eb414681 | 3714 | psi_memstall_leave(&pflags); |
1732d2b0 | 3715 | set_task_reclaim_state(current, NULL); |
e5ca8071 | 3716 | |
0abdee2b | 3717 | /* |
1d82de61 MG |
3718 | * Return the order kswapd stopped reclaiming at as |
3719 | * prepare_kswapd_sleep() takes it into account. If another caller | |
3720 | * entered the allocator slow path while kswapd was awake, order will | |
3721 | * remain at the higher level. | |
0abdee2b | 3722 | */ |
1d82de61 | 3723 | return sc.order; |
1da177e4 LT |
3724 | } |
3725 | ||
e716f2eb | 3726 | /* |
dffcac2c SB |
3727 | * The pgdat->kswapd_classzone_idx is used to pass the highest zone index to be |
3728 | * reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is not | |
3729 | * a valid index then either kswapd runs for first time or kswapd couldn't sleep | |
3730 | * after previous reclaim attempt (node is still unbalanced). In that case | |
3731 | * return the zone index of the previous kswapd reclaim cycle. | |
e716f2eb MG |
3732 | */ |
3733 | static enum zone_type kswapd_classzone_idx(pg_data_t *pgdat, | |
dffcac2c | 3734 | enum zone_type prev_classzone_idx) |
e716f2eb MG |
3735 | { |
3736 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
dffcac2c SB |
3737 | return prev_classzone_idx; |
3738 | return pgdat->kswapd_classzone_idx; | |
e716f2eb MG |
3739 | } |
3740 | ||
38087d9b MG |
3741 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
3742 | unsigned int classzone_idx) | |
f0bc0a60 KM |
3743 | { |
3744 | long remaining = 0; | |
3745 | DEFINE_WAIT(wait); | |
3746 | ||
3747 | if (freezing(current) || kthread_should_stop()) | |
3748 | return; | |
3749 | ||
3750 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3751 | ||
333b0a45 SG |
3752 | /* |
3753 | * Try to sleep for a short interval. Note that kcompactd will only be | |
3754 | * woken if it is possible to sleep for a short interval. This is | |
3755 | * deliberate on the assumption that if reclaim cannot keep an | |
3756 | * eligible zone balanced that it's also unlikely that compaction will | |
3757 | * succeed. | |
3758 | */ | |
d9f21d42 | 3759 | if (prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { |
fd901c95 VB |
3760 | /* |
3761 | * Compaction records what page blocks it recently failed to | |
3762 | * isolate pages from and skips them in the future scanning. | |
3763 | * When kswapd is going to sleep, it is reasonable to assume | |
3764 | * that pages and compaction may succeed so reset the cache. | |
3765 | */ | |
3766 | reset_isolation_suitable(pgdat); | |
3767 | ||
3768 | /* | |
3769 | * We have freed the memory, now we should compact it to make | |
3770 | * allocation of the requested order possible. | |
3771 | */ | |
38087d9b | 3772 | wakeup_kcompactd(pgdat, alloc_order, classzone_idx); |
fd901c95 | 3773 | |
f0bc0a60 | 3774 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
3775 | |
3776 | /* | |
3777 | * If woken prematurely then reset kswapd_classzone_idx and | |
3778 | * order. The values will either be from a wakeup request or | |
3779 | * the previous request that slept prematurely. | |
3780 | */ | |
3781 | if (remaining) { | |
e716f2eb | 3782 | pgdat->kswapd_classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
38087d9b MG |
3783 | pgdat->kswapd_order = max(pgdat->kswapd_order, reclaim_order); |
3784 | } | |
3785 | ||
f0bc0a60 KM |
3786 | finish_wait(&pgdat->kswapd_wait, &wait); |
3787 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
3788 | } | |
3789 | ||
3790 | /* | |
3791 | * After a short sleep, check if it was a premature sleep. If not, then | |
3792 | * go fully to sleep until explicitly woken up. | |
3793 | */ | |
d9f21d42 MG |
3794 | if (!remaining && |
3795 | prepare_kswapd_sleep(pgdat, reclaim_order, classzone_idx)) { | |
f0bc0a60 KM |
3796 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
3797 | ||
3798 | /* | |
3799 | * vmstat counters are not perfectly accurate and the estimated | |
3800 | * value for counters such as NR_FREE_PAGES can deviate from the | |
3801 | * true value by nr_online_cpus * threshold. To avoid the zone | |
3802 | * watermarks being breached while under pressure, we reduce the | |
3803 | * per-cpu vmstat threshold while kswapd is awake and restore | |
3804 | * them before going back to sleep. | |
3805 | */ | |
3806 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
3807 | |
3808 | if (!kthread_should_stop()) | |
3809 | schedule(); | |
3810 | ||
f0bc0a60 KM |
3811 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
3812 | } else { | |
3813 | if (remaining) | |
3814 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
3815 | else | |
3816 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
3817 | } | |
3818 | finish_wait(&pgdat->kswapd_wait, &wait); | |
3819 | } | |
3820 | ||
1da177e4 LT |
3821 | /* |
3822 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 3823 | * from the init process. |
1da177e4 LT |
3824 | * |
3825 | * This basically trickles out pages so that we have _some_ | |
3826 | * free memory available even if there is no other activity | |
3827 | * that frees anything up. This is needed for things like routing | |
3828 | * etc, where we otherwise might have all activity going on in | |
3829 | * asynchronous contexts that cannot page things out. | |
3830 | * | |
3831 | * If there are applications that are active memory-allocators | |
3832 | * (most normal use), this basically shouldn't matter. | |
3833 | */ | |
3834 | static int kswapd(void *p) | |
3835 | { | |
e716f2eb MG |
3836 | unsigned int alloc_order, reclaim_order; |
3837 | unsigned int classzone_idx = MAX_NR_ZONES - 1; | |
1da177e4 LT |
3838 | pg_data_t *pgdat = (pg_data_t*)p; |
3839 | struct task_struct *tsk = current; | |
a70f7302 | 3840 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 3841 | |
174596a0 | 3842 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 3843 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
3844 | |
3845 | /* | |
3846 | * Tell the memory management that we're a "memory allocator", | |
3847 | * and that if we need more memory we should get access to it | |
3848 | * regardless (see "__alloc_pages()"). "kswapd" should | |
3849 | * never get caught in the normal page freeing logic. | |
3850 | * | |
3851 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
3852 | * you need a small amount of memory in order to be able to | |
3853 | * page out something else, and this flag essentially protects | |
3854 | * us from recursively trying to free more memory as we're | |
3855 | * trying to free the first piece of memory in the first place). | |
3856 | */ | |
930d9152 | 3857 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 3858 | set_freezable(); |
1da177e4 | 3859 | |
e716f2eb MG |
3860 | pgdat->kswapd_order = 0; |
3861 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; | |
1da177e4 | 3862 | for ( ; ; ) { |
6f6313d4 | 3863 | bool ret; |
3e1d1d28 | 3864 | |
e716f2eb MG |
3865 | alloc_order = reclaim_order = pgdat->kswapd_order; |
3866 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); | |
3867 | ||
38087d9b MG |
3868 | kswapd_try_sleep: |
3869 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
3870 | classzone_idx); | |
215ddd66 | 3871 | |
38087d9b MG |
3872 | /* Read the new order and classzone_idx */ |
3873 | alloc_order = reclaim_order = pgdat->kswapd_order; | |
dffcac2c | 3874 | classzone_idx = kswapd_classzone_idx(pgdat, classzone_idx); |
38087d9b | 3875 | pgdat->kswapd_order = 0; |
e716f2eb | 3876 | pgdat->kswapd_classzone_idx = MAX_NR_ZONES; |
1da177e4 | 3877 | |
8fe23e05 DR |
3878 | ret = try_to_freeze(); |
3879 | if (kthread_should_stop()) | |
3880 | break; | |
3881 | ||
3882 | /* | |
3883 | * We can speed up thawing tasks if we don't call balance_pgdat | |
3884 | * after returning from the refrigerator | |
3885 | */ | |
38087d9b MG |
3886 | if (ret) |
3887 | continue; | |
3888 | ||
3889 | /* | |
3890 | * Reclaim begins at the requested order but if a high-order | |
3891 | * reclaim fails then kswapd falls back to reclaiming for | |
3892 | * order-0. If that happens, kswapd will consider sleeping | |
3893 | * for the order it finished reclaiming at (reclaim_order) | |
3894 | * but kcompactd is woken to compact for the original | |
3895 | * request (alloc_order). | |
3896 | */ | |
e5146b12 MG |
3897 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, classzone_idx, |
3898 | alloc_order); | |
38087d9b MG |
3899 | reclaim_order = balance_pgdat(pgdat, alloc_order, classzone_idx); |
3900 | if (reclaim_order < alloc_order) | |
3901 | goto kswapd_try_sleep; | |
1da177e4 | 3902 | } |
b0a8cc58 | 3903 | |
71abdc15 | 3904 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
71abdc15 | 3905 | |
1da177e4 LT |
3906 | return 0; |
3907 | } | |
3908 | ||
3909 | /* | |
5ecd9d40 DR |
3910 | * A zone is low on free memory or too fragmented for high-order memory. If |
3911 | * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's | |
3912 | * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim | |
3913 | * has failed or is not needed, still wake up kcompactd if only compaction is | |
3914 | * needed. | |
1da177e4 | 3915 | */ |
5ecd9d40 DR |
3916 | void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, |
3917 | enum zone_type classzone_idx) | |
1da177e4 LT |
3918 | { |
3919 | pg_data_t *pgdat; | |
3920 | ||
6aa303de | 3921 | if (!managed_zone(zone)) |
1da177e4 LT |
3922 | return; |
3923 | ||
5ecd9d40 | 3924 | if (!cpuset_zone_allowed(zone, gfp_flags)) |
1da177e4 | 3925 | return; |
88f5acf8 | 3926 | pgdat = zone->zone_pgdat; |
dffcac2c SB |
3927 | |
3928 | if (pgdat->kswapd_classzone_idx == MAX_NR_ZONES) | |
3929 | pgdat->kswapd_classzone_idx = classzone_idx; | |
3930 | else | |
3931 | pgdat->kswapd_classzone_idx = max(pgdat->kswapd_classzone_idx, | |
3932 | classzone_idx); | |
38087d9b | 3933 | pgdat->kswapd_order = max(pgdat->kswapd_order, order); |
8d0986e2 | 3934 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 3935 | return; |
e1a55637 | 3936 | |
5ecd9d40 DR |
3937 | /* Hopeless node, leave it to direct reclaim if possible */ |
3938 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES || | |
1c30844d MG |
3939 | (pgdat_balanced(pgdat, order, classzone_idx) && |
3940 | !pgdat_watermark_boosted(pgdat, classzone_idx))) { | |
5ecd9d40 DR |
3941 | /* |
3942 | * There may be plenty of free memory available, but it's too | |
3943 | * fragmented for high-order allocations. Wake up kcompactd | |
3944 | * and rely on compaction_suitable() to determine if it's | |
3945 | * needed. If it fails, it will defer subsequent attempts to | |
3946 | * ratelimit its work. | |
3947 | */ | |
3948 | if (!(gfp_flags & __GFP_DIRECT_RECLAIM)) | |
3949 | wakeup_kcompactd(pgdat, order, classzone_idx); | |
e716f2eb | 3950 | return; |
5ecd9d40 | 3951 | } |
88f5acf8 | 3952 | |
5ecd9d40 DR |
3953 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, classzone_idx, order, |
3954 | gfp_flags); | |
8d0986e2 | 3955 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
3956 | } |
3957 | ||
c6f37f12 | 3958 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 3959 | /* |
7b51755c | 3960 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
3961 | * freed pages. |
3962 | * | |
3963 | * Rather than trying to age LRUs the aim is to preserve the overall | |
3964 | * LRU order by reclaiming preferentially | |
3965 | * inactive > active > active referenced > active mapped | |
1da177e4 | 3966 | */ |
7b51755c | 3967 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 3968 | { |
d6277db4 | 3969 | struct scan_control sc = { |
ee814fe2 | 3970 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 3971 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 3972 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 3973 | .priority = DEF_PRIORITY, |
d6277db4 | 3974 | .may_writepage = 1, |
ee814fe2 JW |
3975 | .may_unmap = 1, |
3976 | .may_swap = 1, | |
7b51755c | 3977 | .hibernation_mode = 1, |
1da177e4 | 3978 | }; |
a09ed5e0 | 3979 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c | 3980 | unsigned long nr_reclaimed; |
499118e9 | 3981 | unsigned int noreclaim_flag; |
1da177e4 | 3982 | |
d92a8cfc | 3983 | fs_reclaim_acquire(sc.gfp_mask); |
93781325 | 3984 | noreclaim_flag = memalloc_noreclaim_save(); |
1732d2b0 | 3985 | set_task_reclaim_state(current, &sc.reclaim_state); |
d6277db4 | 3986 | |
3115cd91 | 3987 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 3988 | |
1732d2b0 | 3989 | set_task_reclaim_state(current, NULL); |
499118e9 | 3990 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 3991 | fs_reclaim_release(sc.gfp_mask); |
d6277db4 | 3992 | |
7b51755c | 3993 | return nr_reclaimed; |
1da177e4 | 3994 | } |
c6f37f12 | 3995 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 3996 | |
1da177e4 LT |
3997 | /* It's optimal to keep kswapds on the same CPUs as their memory, but |
3998 | not required for correctness. So if the last cpu in a node goes | |
3999 | away, we get changed to run anywhere: as the first one comes back, | |
4000 | restore their cpu bindings. */ | |
517bbed9 | 4001 | static int kswapd_cpu_online(unsigned int cpu) |
1da177e4 | 4002 | { |
58c0a4a7 | 4003 | int nid; |
1da177e4 | 4004 | |
517bbed9 SAS |
4005 | for_each_node_state(nid, N_MEMORY) { |
4006 | pg_data_t *pgdat = NODE_DATA(nid); | |
4007 | const struct cpumask *mask; | |
a70f7302 | 4008 | |
517bbed9 | 4009 | mask = cpumask_of_node(pgdat->node_id); |
c5f59f08 | 4010 | |
517bbed9 SAS |
4011 | if (cpumask_any_and(cpu_online_mask, mask) < nr_cpu_ids) |
4012 | /* One of our CPUs online: restore mask */ | |
4013 | set_cpus_allowed_ptr(pgdat->kswapd, mask); | |
1da177e4 | 4014 | } |
517bbed9 | 4015 | return 0; |
1da177e4 | 4016 | } |
1da177e4 | 4017 | |
3218ae14 YG |
4018 | /* |
4019 | * This kswapd start function will be called by init and node-hot-add. | |
4020 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
4021 | */ | |
4022 | int kswapd_run(int nid) | |
4023 | { | |
4024 | pg_data_t *pgdat = NODE_DATA(nid); | |
4025 | int ret = 0; | |
4026 | ||
4027 | if (pgdat->kswapd) | |
4028 | return 0; | |
4029 | ||
4030 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
4031 | if (IS_ERR(pgdat->kswapd)) { | |
4032 | /* failure at boot is fatal */ | |
c6202adf | 4033 | BUG_ON(system_state < SYSTEM_RUNNING); |
d5dc0ad9 GS |
4034 | pr_err("Failed to start kswapd on node %d\n", nid); |
4035 | ret = PTR_ERR(pgdat->kswapd); | |
d72515b8 | 4036 | pgdat->kswapd = NULL; |
3218ae14 YG |
4037 | } |
4038 | return ret; | |
4039 | } | |
4040 | ||
8fe23e05 | 4041 | /* |
d8adde17 | 4042 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 4043 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
4044 | */ |
4045 | void kswapd_stop(int nid) | |
4046 | { | |
4047 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
4048 | ||
d8adde17 | 4049 | if (kswapd) { |
8fe23e05 | 4050 | kthread_stop(kswapd); |
d8adde17 JL |
4051 | NODE_DATA(nid)->kswapd = NULL; |
4052 | } | |
8fe23e05 DR |
4053 | } |
4054 | ||
1da177e4 LT |
4055 | static int __init kswapd_init(void) |
4056 | { | |
517bbed9 | 4057 | int nid, ret; |
69e05944 | 4058 | |
1da177e4 | 4059 | swap_setup(); |
48fb2e24 | 4060 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 4061 | kswapd_run(nid); |
517bbed9 SAS |
4062 | ret = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
4063 | "mm/vmscan:online", kswapd_cpu_online, | |
4064 | NULL); | |
4065 | WARN_ON(ret < 0); | |
1da177e4 LT |
4066 | return 0; |
4067 | } | |
4068 | ||
4069 | module_init(kswapd_init) | |
9eeff239 CL |
4070 | |
4071 | #ifdef CONFIG_NUMA | |
4072 | /* | |
a5f5f91d | 4073 | * Node reclaim mode |
9eeff239 | 4074 | * |
a5f5f91d | 4075 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 4076 | * the watermarks. |
9eeff239 | 4077 | */ |
a5f5f91d | 4078 | int node_reclaim_mode __read_mostly; |
9eeff239 | 4079 | |
1b2ffb78 | 4080 | #define RECLAIM_OFF 0 |
7d03431c | 4081 | #define RECLAIM_ZONE (1<<0) /* Run shrink_inactive_list on the zone */ |
1b2ffb78 | 4082 | #define RECLAIM_WRITE (1<<1) /* Writeout pages during reclaim */ |
95bbc0c7 | 4083 | #define RECLAIM_UNMAP (1<<2) /* Unmap pages during reclaim */ |
1b2ffb78 | 4084 | |
a92f7126 | 4085 | /* |
a5f5f91d | 4086 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
4087 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
4088 | * a zone. | |
4089 | */ | |
a5f5f91d | 4090 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 4091 | |
9614634f | 4092 | /* |
a5f5f91d | 4093 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
4094 | * occur. |
4095 | */ | |
4096 | int sysctl_min_unmapped_ratio = 1; | |
4097 | ||
0ff38490 CL |
4098 | /* |
4099 | * If the number of slab pages in a zone grows beyond this percentage then | |
4100 | * slab reclaim needs to occur. | |
4101 | */ | |
4102 | int sysctl_min_slab_ratio = 5; | |
4103 | ||
11fb9989 | 4104 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 4105 | { |
11fb9989 MG |
4106 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
4107 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
4108 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
4109 | |
4110 | /* | |
4111 | * It's possible for there to be more file mapped pages than | |
4112 | * accounted for by the pages on the file LRU lists because | |
4113 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
4114 | */ | |
4115 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
4116 | } | |
4117 | ||
4118 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 4119 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 4120 | { |
d031a157 AM |
4121 | unsigned long nr_pagecache_reclaimable; |
4122 | unsigned long delta = 0; | |
90afa5de MG |
4123 | |
4124 | /* | |
95bbc0c7 | 4125 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 4126 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 4127 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
4128 | * a better estimate |
4129 | */ | |
a5f5f91d MG |
4130 | if (node_reclaim_mode & RECLAIM_UNMAP) |
4131 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 4132 | else |
a5f5f91d | 4133 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
4134 | |
4135 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
4136 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
4137 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
4138 | |
4139 | /* Watch for any possible underflows due to delta */ | |
4140 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
4141 | delta = nr_pagecache_reclaimable; | |
4142 | ||
4143 | return nr_pagecache_reclaimable - delta; | |
4144 | } | |
4145 | ||
9eeff239 | 4146 | /* |
a5f5f91d | 4147 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 4148 | */ |
a5f5f91d | 4149 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 4150 | { |
7fb2d46d | 4151 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 4152 | const unsigned long nr_pages = 1 << order; |
9eeff239 | 4153 | struct task_struct *p = current; |
499118e9 | 4154 | unsigned int noreclaim_flag; |
179e9639 | 4155 | struct scan_control sc = { |
62b726c1 | 4156 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
f2f43e56 | 4157 | .gfp_mask = current_gfp_context(gfp_mask), |
bd2f6199 | 4158 | .order = order, |
a5f5f91d MG |
4159 | .priority = NODE_RECLAIM_PRIORITY, |
4160 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
4161 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 4162 | .may_swap = 1, |
f2f43e56 | 4163 | .reclaim_idx = gfp_zone(gfp_mask), |
179e9639 | 4164 | }; |
9eeff239 | 4165 | |
132bb8cf YS |
4166 | trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order, |
4167 | sc.gfp_mask); | |
4168 | ||
9eeff239 | 4169 | cond_resched(); |
93781325 | 4170 | fs_reclaim_acquire(sc.gfp_mask); |
d4f7796e | 4171 | /* |
95bbc0c7 | 4172 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 4173 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 4174 | * and RECLAIM_UNMAP. |
d4f7796e | 4175 | */ |
499118e9 VB |
4176 | noreclaim_flag = memalloc_noreclaim_save(); |
4177 | p->flags |= PF_SWAPWRITE; | |
1732d2b0 | 4178 | set_task_reclaim_state(p, &sc.reclaim_state); |
c84db23c | 4179 | |
a5f5f91d | 4180 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
0ff38490 | 4181 | /* |
894befec | 4182 | * Free memory by calling shrink node with increasing |
0ff38490 CL |
4183 | * priorities until we have enough memory freed. |
4184 | */ | |
0ff38490 | 4185 | do { |
970a39a3 | 4186 | shrink_node(pgdat, &sc); |
9e3b2f8c | 4187 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 4188 | } |
c84db23c | 4189 | |
1732d2b0 | 4190 | set_task_reclaim_state(p, NULL); |
499118e9 VB |
4191 | current->flags &= ~PF_SWAPWRITE; |
4192 | memalloc_noreclaim_restore(noreclaim_flag); | |
93781325 | 4193 | fs_reclaim_release(sc.gfp_mask); |
132bb8cf YS |
4194 | |
4195 | trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed); | |
4196 | ||
a79311c1 | 4197 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 4198 | } |
179e9639 | 4199 | |
a5f5f91d | 4200 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 4201 | { |
d773ed6b | 4202 | int ret; |
179e9639 AM |
4203 | |
4204 | /* | |
a5f5f91d | 4205 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 4206 | * slab pages if we are over the defined limits. |
34aa1330 | 4207 | * |
9614634f CL |
4208 | * A small portion of unmapped file backed pages is needed for |
4209 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
4210 | * thrown out if the node is overallocated. So we do not reclaim |
4211 | * if less than a specified percentage of the node is used by | |
9614634f | 4212 | * unmapped file backed pages. |
179e9639 | 4213 | */ |
a5f5f91d | 4214 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
385386cf | 4215 | node_page_state(pgdat, NR_SLAB_RECLAIMABLE) <= pgdat->min_slab_pages) |
a5f5f91d | 4216 | return NODE_RECLAIM_FULL; |
179e9639 AM |
4217 | |
4218 | /* | |
d773ed6b | 4219 | * Do not scan if the allocation should not be delayed. |
179e9639 | 4220 | */ |
d0164adc | 4221 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 4222 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
4223 | |
4224 | /* | |
a5f5f91d | 4225 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
4226 | * have associated processors. This will favor the local processor |
4227 | * over remote processors and spread off node memory allocations | |
4228 | * as wide as possible. | |
4229 | */ | |
a5f5f91d MG |
4230 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
4231 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 4232 | |
a5f5f91d MG |
4233 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
4234 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 4235 | |
a5f5f91d MG |
4236 | ret = __node_reclaim(pgdat, gfp_mask, order); |
4237 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 4238 | |
24cf7251 MG |
4239 | if (!ret) |
4240 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
4241 | ||
d773ed6b | 4242 | return ret; |
179e9639 | 4243 | } |
9eeff239 | 4244 | #endif |
894bc310 | 4245 | |
894bc310 LS |
4246 | /* |
4247 | * page_evictable - test whether a page is evictable | |
4248 | * @page: the page to test | |
894bc310 LS |
4249 | * |
4250 | * Test whether page is evictable--i.e., should be placed on active/inactive | |
39b5f29a | 4251 | * lists vs unevictable list. |
894bc310 LS |
4252 | * |
4253 | * Reasons page might not be evictable: | |
ba9ddf49 | 4254 | * (1) page's mapping marked unevictable |
b291f000 | 4255 | * (2) page is part of an mlocked VMA |
ba9ddf49 | 4256 | * |
894bc310 | 4257 | */ |
39b5f29a | 4258 | int page_evictable(struct page *page) |
894bc310 | 4259 | { |
e92bb4dd HY |
4260 | int ret; |
4261 | ||
4262 | /* Prevent address_space of inode and swap cache from being freed */ | |
4263 | rcu_read_lock(); | |
4264 | ret = !mapping_unevictable(page_mapping(page)) && !PageMlocked(page); | |
4265 | rcu_read_unlock(); | |
4266 | return ret; | |
894bc310 | 4267 | } |
89e004ea LS |
4268 | |
4269 | /** | |
64e3d12f KHY |
4270 | * check_move_unevictable_pages - check pages for evictability and move to |
4271 | * appropriate zone lru list | |
4272 | * @pvec: pagevec with lru pages to check | |
89e004ea | 4273 | * |
64e3d12f KHY |
4274 | * Checks pages for evictability, if an evictable page is in the unevictable |
4275 | * lru list, moves it to the appropriate evictable lru list. This function | |
4276 | * should be only used for lru pages. | |
89e004ea | 4277 | */ |
64e3d12f | 4278 | void check_move_unevictable_pages(struct pagevec *pvec) |
89e004ea | 4279 | { |
925b7673 | 4280 | struct lruvec *lruvec; |
785b99fe | 4281 | struct pglist_data *pgdat = NULL; |
24513264 HD |
4282 | int pgscanned = 0; |
4283 | int pgrescued = 0; | |
4284 | int i; | |
89e004ea | 4285 | |
64e3d12f KHY |
4286 | for (i = 0; i < pvec->nr; i++) { |
4287 | struct page *page = pvec->pages[i]; | |
785b99fe | 4288 | struct pglist_data *pagepgdat = page_pgdat(page); |
89e004ea | 4289 | |
24513264 | 4290 | pgscanned++; |
785b99fe MG |
4291 | if (pagepgdat != pgdat) { |
4292 | if (pgdat) | |
4293 | spin_unlock_irq(&pgdat->lru_lock); | |
4294 | pgdat = pagepgdat; | |
4295 | spin_lock_irq(&pgdat->lru_lock); | |
24513264 | 4296 | } |
785b99fe | 4297 | lruvec = mem_cgroup_page_lruvec(page, pgdat); |
89e004ea | 4298 | |
24513264 HD |
4299 | if (!PageLRU(page) || !PageUnevictable(page)) |
4300 | continue; | |
89e004ea | 4301 | |
39b5f29a | 4302 | if (page_evictable(page)) { |
24513264 HD |
4303 | enum lru_list lru = page_lru_base_type(page); |
4304 | ||
309381fe | 4305 | VM_BUG_ON_PAGE(PageActive(page), page); |
24513264 | 4306 | ClearPageUnevictable(page); |
fa9add64 HD |
4307 | del_page_from_lru_list(page, lruvec, LRU_UNEVICTABLE); |
4308 | add_page_to_lru_list(page, lruvec, lru); | |
24513264 | 4309 | pgrescued++; |
89e004ea | 4310 | } |
24513264 | 4311 | } |
89e004ea | 4312 | |
785b99fe | 4313 | if (pgdat) { |
24513264 HD |
4314 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
4315 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
785b99fe | 4316 | spin_unlock_irq(&pgdat->lru_lock); |
89e004ea | 4317 | } |
89e004ea | 4318 | } |
64e3d12f | 4319 | EXPORT_SYMBOL_GPL(check_move_unevictable_pages); |