]>
Commit | Line | Data |
---|---|---|
b2441318 | 1 | // SPDX-License-Identifier: GPL-2.0 |
1da177e4 | 2 | /* |
1da177e4 LT |
3 | * Copyright (C) 1991, 1992, 1993, 1994 Linus Torvalds |
4 | * | |
5 | * Swap reorganised 29.12.95, Stephen Tweedie. | |
6 | * kswapd added: 7.1.96 sct | |
7 | * Removed kswapd_ctl limits, and swap out as many pages as needed | |
8 | * to bring the system back to freepages.high: 2.4.97, Rik van Riel. | |
9 | * Zone aware kswapd started 02/00, Kanoj Sarcar (kanoj@sgi.com). | |
10 | * Multiqueue VM started 5.8.00, Rik van Riel. | |
11 | */ | |
12 | ||
b1de0d13 MH |
13 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
14 | ||
1da177e4 | 15 | #include <linux/mm.h> |
5b3cc15a | 16 | #include <linux/sched/mm.h> |
1da177e4 | 17 | #include <linux/module.h> |
5a0e3ad6 | 18 | #include <linux/gfp.h> |
1da177e4 LT |
19 | #include <linux/kernel_stat.h> |
20 | #include <linux/swap.h> | |
21 | #include <linux/pagemap.h> | |
22 | #include <linux/init.h> | |
23 | #include <linux/highmem.h> | |
70ddf637 | 24 | #include <linux/vmpressure.h> |
e129b5c2 | 25 | #include <linux/vmstat.h> |
1da177e4 LT |
26 | #include <linux/file.h> |
27 | #include <linux/writeback.h> | |
28 | #include <linux/blkdev.h> | |
29 | #include <linux/buffer_head.h> /* for try_to_release_page(), | |
30 | buffer_heads_over_limit */ | |
31 | #include <linux/mm_inline.h> | |
1da177e4 LT |
32 | #include <linux/backing-dev.h> |
33 | #include <linux/rmap.h> | |
34 | #include <linux/topology.h> | |
35 | #include <linux/cpu.h> | |
36 | #include <linux/cpuset.h> | |
3e7d3449 | 37 | #include <linux/compaction.h> |
1da177e4 LT |
38 | #include <linux/notifier.h> |
39 | #include <linux/rwsem.h> | |
248a0301 | 40 | #include <linux/delay.h> |
3218ae14 | 41 | #include <linux/kthread.h> |
7dfb7103 | 42 | #include <linux/freezer.h> |
66e1707b | 43 | #include <linux/memcontrol.h> |
26aa2d19 | 44 | #include <linux/migrate.h> |
873b4771 | 45 | #include <linux/delayacct.h> |
af936a16 | 46 | #include <linux/sysctl.h> |
929bea7c | 47 | #include <linux/oom.h> |
64e3d12f | 48 | #include <linux/pagevec.h> |
268bb0ce | 49 | #include <linux/prefetch.h> |
b1de0d13 | 50 | #include <linux/printk.h> |
f9fe48be | 51 | #include <linux/dax.h> |
eb414681 | 52 | #include <linux/psi.h> |
1da177e4 LT |
53 | |
54 | #include <asm/tlbflush.h> | |
55 | #include <asm/div64.h> | |
56 | ||
57 | #include <linux/swapops.h> | |
117aad1e | 58 | #include <linux/balloon_compaction.h> |
1da177e4 | 59 | |
0f8053a5 NP |
60 | #include "internal.h" |
61 | ||
33906bc5 MG |
62 | #define CREATE_TRACE_POINTS |
63 | #include <trace/events/vmscan.h> | |
64 | ||
1da177e4 | 65 | struct scan_control { |
22fba335 KM |
66 | /* How many pages shrink_list() should reclaim */ |
67 | unsigned long nr_to_reclaim; | |
68 | ||
ee814fe2 JW |
69 | /* |
70 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
71 | * are scanned. | |
72 | */ | |
73 | nodemask_t *nodemask; | |
9e3b2f8c | 74 | |
f16015fb JW |
75 | /* |
76 | * The memory cgroup that hit its limit and as a result is the | |
77 | * primary target of this reclaim invocation. | |
78 | */ | |
79 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 80 | |
7cf111bc JW |
81 | /* |
82 | * Scan pressure balancing between anon and file LRUs | |
83 | */ | |
84 | unsigned long anon_cost; | |
85 | unsigned long file_cost; | |
86 | ||
b91ac374 JW |
87 | /* Can active pages be deactivated as part of reclaim? */ |
88 | #define DEACTIVATE_ANON 1 | |
89 | #define DEACTIVATE_FILE 2 | |
90 | unsigned int may_deactivate:2; | |
91 | unsigned int force_deactivate:1; | |
92 | unsigned int skipped_deactivate:1; | |
93 | ||
1276ad68 | 94 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
ee814fe2 JW |
95 | unsigned int may_writepage:1; |
96 | ||
97 | /* Can mapped pages be reclaimed? */ | |
98 | unsigned int may_unmap:1; | |
99 | ||
100 | /* Can pages be swapped as part of reclaim? */ | |
101 | unsigned int may_swap:1; | |
102 | ||
d6622f63 | 103 | /* |
f56ce412 JW |
104 | * Cgroup memory below memory.low is protected as long as we |
105 | * don't threaten to OOM. If any cgroup is reclaimed at | |
106 | * reduced force or passed over entirely due to its memory.low | |
107 | * setting (memcg_low_skipped), and nothing is reclaimed as a | |
108 | * result, then go back for one more cycle that reclaims the protected | |
109 | * memory (memcg_low_reclaim) to avert OOM. | |
d6622f63 YX |
110 | */ |
111 | unsigned int memcg_low_reclaim:1; | |
112 | unsigned int memcg_low_skipped:1; | |
241994ed | 113 | |
ee814fe2 JW |
114 | unsigned int hibernation_mode:1; |
115 | ||
116 | /* One of the zones is ready for compaction */ | |
117 | unsigned int compaction_ready:1; | |
118 | ||
b91ac374 JW |
119 | /* There is easily reclaimable cold cache in the current node */ |
120 | unsigned int cache_trim_mode:1; | |
121 | ||
53138cea JW |
122 | /* The file pages on the current node are dangerously low */ |
123 | unsigned int file_is_tiny:1; | |
124 | ||
26aa2d19 DH |
125 | /* Always discard instead of demoting to lower tier memory */ |
126 | unsigned int no_demotion:1; | |
127 | ||
bb451fdf GT |
128 | /* Allocation order */ |
129 | s8 order; | |
130 | ||
131 | /* Scan (total_size >> priority) pages at once */ | |
132 | s8 priority; | |
133 | ||
134 | /* The highest zone to isolate pages for reclaim from */ | |
135 | s8 reclaim_idx; | |
136 | ||
137 | /* This context's GFP mask */ | |
138 | gfp_t gfp_mask; | |
139 | ||
ee814fe2 JW |
140 | /* Incremented by the number of inactive pages that were scanned */ |
141 | unsigned long nr_scanned; | |
142 | ||
143 | /* Number of pages freed so far during a call to shrink_zones() */ | |
144 | unsigned long nr_reclaimed; | |
d108c772 AR |
145 | |
146 | struct { | |
147 | unsigned int dirty; | |
148 | unsigned int unqueued_dirty; | |
149 | unsigned int congested; | |
150 | unsigned int writeback; | |
151 | unsigned int immediate; | |
152 | unsigned int file_taken; | |
153 | unsigned int taken; | |
154 | } nr; | |
e5ca8071 YS |
155 | |
156 | /* for recording the reclaimed slab by now */ | |
157 | struct reclaim_state reclaim_state; | |
1da177e4 LT |
158 | }; |
159 | ||
1da177e4 LT |
160 | #ifdef ARCH_HAS_PREFETCHW |
161 | #define prefetchw_prev_lru_page(_page, _base, _field) \ | |
162 | do { \ | |
163 | if ((_page)->lru.prev != _base) { \ | |
164 | struct page *prev; \ | |
165 | \ | |
166 | prev = lru_to_page(&(_page->lru)); \ | |
167 | prefetchw(&prev->_field); \ | |
168 | } \ | |
169 | } while (0) | |
170 | #else | |
171 | #define prefetchw_prev_lru_page(_page, _base, _field) do { } while (0) | |
172 | #endif | |
173 | ||
174 | /* | |
c843966c | 175 | * From 0 .. 200. Higher means more swappy. |
1da177e4 LT |
176 | */ |
177 | int vm_swappiness = 60; | |
1da177e4 | 178 | |
0a432dcb YS |
179 | static void set_task_reclaim_state(struct task_struct *task, |
180 | struct reclaim_state *rs) | |
181 | { | |
182 | /* Check for an overwrite */ | |
183 | WARN_ON_ONCE(rs && task->reclaim_state); | |
184 | ||
185 | /* Check for the nulling of an already-nulled member */ | |
186 | WARN_ON_ONCE(!rs && !task->reclaim_state); | |
187 | ||
188 | task->reclaim_state = rs; | |
189 | } | |
190 | ||
1da177e4 LT |
191 | static LIST_HEAD(shrinker_list); |
192 | static DECLARE_RWSEM(shrinker_rwsem); | |
193 | ||
0a432dcb | 194 | #ifdef CONFIG_MEMCG |
a2fb1261 | 195 | static int shrinker_nr_max; |
2bfd3637 | 196 | |
3c6f17e6 | 197 | /* The shrinker_info is expanded in a batch of BITS_PER_LONG */ |
a2fb1261 YS |
198 | static inline int shrinker_map_size(int nr_items) |
199 | { | |
200 | return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long)); | |
201 | } | |
2bfd3637 | 202 | |
3c6f17e6 YS |
203 | static inline int shrinker_defer_size(int nr_items) |
204 | { | |
205 | return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t)); | |
206 | } | |
207 | ||
468ab843 YS |
208 | static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg, |
209 | int nid) | |
210 | { | |
211 | return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info, | |
212 | lockdep_is_held(&shrinker_rwsem)); | |
213 | } | |
214 | ||
e4262c4f | 215 | static int expand_one_shrinker_info(struct mem_cgroup *memcg, |
3c6f17e6 YS |
216 | int map_size, int defer_size, |
217 | int old_map_size, int old_defer_size) | |
2bfd3637 | 218 | { |
e4262c4f | 219 | struct shrinker_info *new, *old; |
2bfd3637 YS |
220 | struct mem_cgroup_per_node *pn; |
221 | int nid; | |
3c6f17e6 | 222 | int size = map_size + defer_size; |
2bfd3637 | 223 | |
2bfd3637 YS |
224 | for_each_node(nid) { |
225 | pn = memcg->nodeinfo[nid]; | |
468ab843 | 226 | old = shrinker_info_protected(memcg, nid); |
2bfd3637 YS |
227 | /* Not yet online memcg */ |
228 | if (!old) | |
229 | return 0; | |
230 | ||
231 | new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid); | |
232 | if (!new) | |
233 | return -ENOMEM; | |
234 | ||
3c6f17e6 YS |
235 | new->nr_deferred = (atomic_long_t *)(new + 1); |
236 | new->map = (void *)new->nr_deferred + defer_size; | |
237 | ||
238 | /* map: set all old bits, clear all new bits */ | |
239 | memset(new->map, (int)0xff, old_map_size); | |
240 | memset((void *)new->map + old_map_size, 0, map_size - old_map_size); | |
241 | /* nr_deferred: copy old values, clear all new values */ | |
242 | memcpy(new->nr_deferred, old->nr_deferred, old_defer_size); | |
243 | memset((void *)new->nr_deferred + old_defer_size, 0, | |
244 | defer_size - old_defer_size); | |
2bfd3637 | 245 | |
e4262c4f | 246 | rcu_assign_pointer(pn->shrinker_info, new); |
72673e86 | 247 | kvfree_rcu(old, rcu); |
2bfd3637 YS |
248 | } |
249 | ||
250 | return 0; | |
251 | } | |
252 | ||
e4262c4f | 253 | void free_shrinker_info(struct mem_cgroup *memcg) |
2bfd3637 YS |
254 | { |
255 | struct mem_cgroup_per_node *pn; | |
e4262c4f | 256 | struct shrinker_info *info; |
2bfd3637 YS |
257 | int nid; |
258 | ||
2bfd3637 YS |
259 | for_each_node(nid) { |
260 | pn = memcg->nodeinfo[nid]; | |
e4262c4f YS |
261 | info = rcu_dereference_protected(pn->shrinker_info, true); |
262 | kvfree(info); | |
263 | rcu_assign_pointer(pn->shrinker_info, NULL); | |
2bfd3637 YS |
264 | } |
265 | } | |
266 | ||
e4262c4f | 267 | int alloc_shrinker_info(struct mem_cgroup *memcg) |
2bfd3637 | 268 | { |
e4262c4f | 269 | struct shrinker_info *info; |
2bfd3637 | 270 | int nid, size, ret = 0; |
3c6f17e6 | 271 | int map_size, defer_size = 0; |
2bfd3637 | 272 | |
d27cf2aa | 273 | down_write(&shrinker_rwsem); |
3c6f17e6 YS |
274 | map_size = shrinker_map_size(shrinker_nr_max); |
275 | defer_size = shrinker_defer_size(shrinker_nr_max); | |
276 | size = map_size + defer_size; | |
2bfd3637 | 277 | for_each_node(nid) { |
e4262c4f YS |
278 | info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid); |
279 | if (!info) { | |
280 | free_shrinker_info(memcg); | |
2bfd3637 YS |
281 | ret = -ENOMEM; |
282 | break; | |
283 | } | |
3c6f17e6 YS |
284 | info->nr_deferred = (atomic_long_t *)(info + 1); |
285 | info->map = (void *)info->nr_deferred + defer_size; | |
e4262c4f | 286 | rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info); |
2bfd3637 | 287 | } |
d27cf2aa | 288 | up_write(&shrinker_rwsem); |
2bfd3637 YS |
289 | |
290 | return ret; | |
291 | } | |
292 | ||
3c6f17e6 YS |
293 | static inline bool need_expand(int nr_max) |
294 | { | |
295 | return round_up(nr_max, BITS_PER_LONG) > | |
296 | round_up(shrinker_nr_max, BITS_PER_LONG); | |
297 | } | |
298 | ||
e4262c4f | 299 | static int expand_shrinker_info(int new_id) |
2bfd3637 | 300 | { |
3c6f17e6 | 301 | int ret = 0; |
a2fb1261 | 302 | int new_nr_max = new_id + 1; |
3c6f17e6 YS |
303 | int map_size, defer_size = 0; |
304 | int old_map_size, old_defer_size = 0; | |
2bfd3637 YS |
305 | struct mem_cgroup *memcg; |
306 | ||
3c6f17e6 | 307 | if (!need_expand(new_nr_max)) |
a2fb1261 | 308 | goto out; |
2bfd3637 | 309 | |
2bfd3637 | 310 | if (!root_mem_cgroup) |
d27cf2aa YS |
311 | goto out; |
312 | ||
313 | lockdep_assert_held(&shrinker_rwsem); | |
2bfd3637 | 314 | |
3c6f17e6 YS |
315 | map_size = shrinker_map_size(new_nr_max); |
316 | defer_size = shrinker_defer_size(new_nr_max); | |
317 | old_map_size = shrinker_map_size(shrinker_nr_max); | |
318 | old_defer_size = shrinker_defer_size(shrinker_nr_max); | |
319 | ||
2bfd3637 YS |
320 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
321 | do { | |
3c6f17e6 YS |
322 | ret = expand_one_shrinker_info(memcg, map_size, defer_size, |
323 | old_map_size, old_defer_size); | |
2bfd3637 YS |
324 | if (ret) { |
325 | mem_cgroup_iter_break(NULL, memcg); | |
d27cf2aa | 326 | goto out; |
2bfd3637 YS |
327 | } |
328 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
d27cf2aa | 329 | out: |
2bfd3637 | 330 | if (!ret) |
a2fb1261 | 331 | shrinker_nr_max = new_nr_max; |
d27cf2aa | 332 | |
2bfd3637 YS |
333 | return ret; |
334 | } | |
335 | ||
336 | void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id) | |
337 | { | |
338 | if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) { | |
e4262c4f | 339 | struct shrinker_info *info; |
2bfd3637 YS |
340 | |
341 | rcu_read_lock(); | |
e4262c4f | 342 | info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info); |
2bfd3637 YS |
343 | /* Pairs with smp mb in shrink_slab() */ |
344 | smp_mb__before_atomic(); | |
e4262c4f | 345 | set_bit(shrinker_id, info->map); |
2bfd3637 YS |
346 | rcu_read_unlock(); |
347 | } | |
348 | } | |
349 | ||
b4c2b231 | 350 | static DEFINE_IDR(shrinker_idr); |
b4c2b231 KT |
351 | |
352 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) | |
353 | { | |
354 | int id, ret = -ENOMEM; | |
355 | ||
476b30a0 YS |
356 | if (mem_cgroup_disabled()) |
357 | return -ENOSYS; | |
358 | ||
b4c2b231 KT |
359 | down_write(&shrinker_rwsem); |
360 | /* This may call shrinker, so it must use down_read_trylock() */ | |
41ca668a | 361 | id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL); |
b4c2b231 KT |
362 | if (id < 0) |
363 | goto unlock; | |
364 | ||
0a4465d3 | 365 | if (id >= shrinker_nr_max) { |
e4262c4f | 366 | if (expand_shrinker_info(id)) { |
0a4465d3 KT |
367 | idr_remove(&shrinker_idr, id); |
368 | goto unlock; | |
369 | } | |
0a4465d3 | 370 | } |
b4c2b231 KT |
371 | shrinker->id = id; |
372 | ret = 0; | |
373 | unlock: | |
374 | up_write(&shrinker_rwsem); | |
375 | return ret; | |
376 | } | |
377 | ||
378 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
379 | { | |
380 | int id = shrinker->id; | |
381 | ||
382 | BUG_ON(id < 0); | |
383 | ||
41ca668a YS |
384 | lockdep_assert_held(&shrinker_rwsem); |
385 | ||
b4c2b231 | 386 | idr_remove(&shrinker_idr, id); |
b4c2b231 | 387 | } |
b4c2b231 | 388 | |
86750830 YS |
389 | static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker, |
390 | struct mem_cgroup *memcg) | |
391 | { | |
392 | struct shrinker_info *info; | |
393 | ||
394 | info = shrinker_info_protected(memcg, nid); | |
395 | return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0); | |
396 | } | |
397 | ||
398 | static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker, | |
399 | struct mem_cgroup *memcg) | |
400 | { | |
401 | struct shrinker_info *info; | |
402 | ||
403 | info = shrinker_info_protected(memcg, nid); | |
404 | return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]); | |
405 | } | |
406 | ||
a178015c YS |
407 | void reparent_shrinker_deferred(struct mem_cgroup *memcg) |
408 | { | |
409 | int i, nid; | |
410 | long nr; | |
411 | struct mem_cgroup *parent; | |
412 | struct shrinker_info *child_info, *parent_info; | |
413 | ||
414 | parent = parent_mem_cgroup(memcg); | |
415 | if (!parent) | |
416 | parent = root_mem_cgroup; | |
417 | ||
418 | /* Prevent from concurrent shrinker_info expand */ | |
419 | down_read(&shrinker_rwsem); | |
420 | for_each_node(nid) { | |
421 | child_info = shrinker_info_protected(memcg, nid); | |
422 | parent_info = shrinker_info_protected(parent, nid); | |
423 | for (i = 0; i < shrinker_nr_max; i++) { | |
424 | nr = atomic_long_read(&child_info->nr_deferred[i]); | |
425 | atomic_long_add(nr, &parent_info->nr_deferred[i]); | |
426 | } | |
427 | } | |
428 | up_read(&shrinker_rwsem); | |
429 | } | |
430 | ||
b5ead35e | 431 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 432 | { |
b5ead35e | 433 | return sc->target_mem_cgroup; |
89b5fae5 | 434 | } |
97c9341f TH |
435 | |
436 | /** | |
b5ead35e | 437 | * writeback_throttling_sane - is the usual dirty throttling mechanism available? |
97c9341f TH |
438 | * @sc: scan_control in question |
439 | * | |
440 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
441 | * completely broken with the legacy memcg and direct stalling in | |
442 | * shrink_page_list() is used for throttling instead, which lacks all the | |
443 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
444 | * allocation and configurability. | |
445 | * | |
446 | * This function tests whether the vmscan currently in progress can assume | |
447 | * that the normal dirty throttling mechanism is operational. | |
448 | */ | |
b5ead35e | 449 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f | 450 | { |
b5ead35e | 451 | if (!cgroup_reclaim(sc)) |
97c9341f TH |
452 | return true; |
453 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 454 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
455 | return true; |
456 | #endif | |
457 | return false; | |
458 | } | |
91a45470 | 459 | #else |
0a432dcb YS |
460 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) |
461 | { | |
476b30a0 | 462 | return -ENOSYS; |
0a432dcb YS |
463 | } |
464 | ||
465 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
466 | { | |
467 | } | |
468 | ||
86750830 YS |
469 | static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker, |
470 | struct mem_cgroup *memcg) | |
471 | { | |
472 | return 0; | |
473 | } | |
474 | ||
475 | static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker, | |
476 | struct mem_cgroup *memcg) | |
477 | { | |
478 | return 0; | |
479 | } | |
480 | ||
b5ead35e | 481 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 482 | { |
b5ead35e | 483 | return false; |
89b5fae5 | 484 | } |
97c9341f | 485 | |
b5ead35e | 486 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f TH |
487 | { |
488 | return true; | |
489 | } | |
91a45470 KH |
490 | #endif |
491 | ||
86750830 YS |
492 | static long xchg_nr_deferred(struct shrinker *shrinker, |
493 | struct shrink_control *sc) | |
494 | { | |
495 | int nid = sc->nid; | |
496 | ||
497 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) | |
498 | nid = 0; | |
499 | ||
500 | if (sc->memcg && | |
501 | (shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
502 | return xchg_nr_deferred_memcg(nid, shrinker, | |
503 | sc->memcg); | |
504 | ||
505 | return atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
506 | } | |
507 | ||
508 | ||
509 | static long add_nr_deferred(long nr, struct shrinker *shrinker, | |
510 | struct shrink_control *sc) | |
511 | { | |
512 | int nid = sc->nid; | |
513 | ||
514 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) | |
515 | nid = 0; | |
516 | ||
517 | if (sc->memcg && | |
518 | (shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
519 | return add_nr_deferred_memcg(nr, nid, shrinker, | |
520 | sc->memcg); | |
521 | ||
522 | return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]); | |
523 | } | |
524 | ||
26aa2d19 DH |
525 | static bool can_demote(int nid, struct scan_control *sc) |
526 | { | |
20b51af1 HY |
527 | if (!numa_demotion_enabled) |
528 | return false; | |
3a235693 DH |
529 | if (sc) { |
530 | if (sc->no_demotion) | |
531 | return false; | |
532 | /* It is pointless to do demotion in memcg reclaim */ | |
533 | if (cgroup_reclaim(sc)) | |
534 | return false; | |
535 | } | |
26aa2d19 DH |
536 | if (next_demotion_node(nid) == NUMA_NO_NODE) |
537 | return false; | |
538 | ||
20b51af1 | 539 | return true; |
26aa2d19 DH |
540 | } |
541 | ||
a2a36488 KB |
542 | static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg, |
543 | int nid, | |
544 | struct scan_control *sc) | |
545 | { | |
546 | if (memcg == NULL) { | |
547 | /* | |
548 | * For non-memcg reclaim, is there | |
549 | * space in any swap device? | |
550 | */ | |
551 | if (get_nr_swap_pages() > 0) | |
552 | return true; | |
553 | } else { | |
554 | /* Is the memcg below its swap limit? */ | |
555 | if (mem_cgroup_get_nr_swap_pages(memcg) > 0) | |
556 | return true; | |
557 | } | |
558 | ||
559 | /* | |
560 | * The page can not be swapped. | |
561 | * | |
562 | * Can it be reclaimed from this node via demotion? | |
563 | */ | |
564 | return can_demote(nid, sc); | |
565 | } | |
566 | ||
5a1c84b4 MG |
567 | /* |
568 | * This misses isolated pages which are not accounted for to save counters. | |
569 | * As the data only determines if reclaim or compaction continues, it is | |
570 | * not expected that isolated pages will be a dominating factor. | |
571 | */ | |
572 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
573 | { | |
574 | unsigned long nr; | |
575 | ||
576 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
577 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
a2a36488 | 578 | if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL)) |
5a1c84b4 MG |
579 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + |
580 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
581 | ||
582 | return nr; | |
583 | } | |
584 | ||
fd538803 MH |
585 | /** |
586 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
587 | * @lruvec: lru vector | |
588 | * @lru: lru to use | |
589 | * @zone_idx: zones to consider (use MAX_NR_ZONES for the whole LRU list) | |
590 | */ | |
2091339d YZ |
591 | static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, |
592 | int zone_idx) | |
c9f299d9 | 593 | { |
de3b0150 | 594 | unsigned long size = 0; |
fd538803 MH |
595 | int zid; |
596 | ||
de3b0150 | 597 | for (zid = 0; zid <= zone_idx && zid < MAX_NR_ZONES; zid++) { |
fd538803 | 598 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; |
c9f299d9 | 599 | |
fd538803 MH |
600 | if (!managed_zone(zone)) |
601 | continue; | |
602 | ||
603 | if (!mem_cgroup_disabled()) | |
de3b0150 | 604 | size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid); |
fd538803 | 605 | else |
de3b0150 | 606 | size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru); |
fd538803 | 607 | } |
de3b0150 | 608 | return size; |
b4536f0c MH |
609 | } |
610 | ||
1da177e4 | 611 | /* |
1d3d4437 | 612 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 613 | */ |
8e04944f | 614 | int prealloc_shrinker(struct shrinker *shrinker) |
1da177e4 | 615 | { |
476b30a0 YS |
616 | unsigned int size; |
617 | int err; | |
618 | ||
619 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) { | |
620 | err = prealloc_memcg_shrinker(shrinker); | |
621 | if (err != -ENOSYS) | |
622 | return err; | |
1d3d4437 | 623 | |
476b30a0 YS |
624 | shrinker->flags &= ~SHRINKER_MEMCG_AWARE; |
625 | } | |
626 | ||
627 | size = sizeof(*shrinker->nr_deferred); | |
1d3d4437 GC |
628 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
629 | size *= nr_node_ids; | |
630 | ||
631 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
632 | if (!shrinker->nr_deferred) | |
633 | return -ENOMEM; | |
b4c2b231 | 634 | |
8e04944f TH |
635 | return 0; |
636 | } | |
637 | ||
638 | void free_prealloced_shrinker(struct shrinker *shrinker) | |
639 | { | |
41ca668a YS |
640 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) { |
641 | down_write(&shrinker_rwsem); | |
b4c2b231 | 642 | unregister_memcg_shrinker(shrinker); |
41ca668a | 643 | up_write(&shrinker_rwsem); |
476b30a0 | 644 | return; |
41ca668a | 645 | } |
b4c2b231 | 646 | |
8e04944f TH |
647 | kfree(shrinker->nr_deferred); |
648 | shrinker->nr_deferred = NULL; | |
649 | } | |
1d3d4437 | 650 | |
8e04944f TH |
651 | void register_shrinker_prepared(struct shrinker *shrinker) |
652 | { | |
8e1f936b RR |
653 | down_write(&shrinker_rwsem); |
654 | list_add_tail(&shrinker->list, &shrinker_list); | |
41ca668a | 655 | shrinker->flags |= SHRINKER_REGISTERED; |
8e1f936b | 656 | up_write(&shrinker_rwsem); |
8e04944f TH |
657 | } |
658 | ||
659 | int register_shrinker(struct shrinker *shrinker) | |
660 | { | |
661 | int err = prealloc_shrinker(shrinker); | |
662 | ||
663 | if (err) | |
664 | return err; | |
665 | register_shrinker_prepared(shrinker); | |
1d3d4437 | 666 | return 0; |
1da177e4 | 667 | } |
8e1f936b | 668 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
669 | |
670 | /* | |
671 | * Remove one | |
672 | */ | |
8e1f936b | 673 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 | 674 | { |
41ca668a | 675 | if (!(shrinker->flags & SHRINKER_REGISTERED)) |
bb422a73 | 676 | return; |
41ca668a | 677 | |
1da177e4 LT |
678 | down_write(&shrinker_rwsem); |
679 | list_del(&shrinker->list); | |
41ca668a YS |
680 | shrinker->flags &= ~SHRINKER_REGISTERED; |
681 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) | |
682 | unregister_memcg_shrinker(shrinker); | |
1da177e4 | 683 | up_write(&shrinker_rwsem); |
41ca668a | 684 | |
ae393321 | 685 | kfree(shrinker->nr_deferred); |
bb422a73 | 686 | shrinker->nr_deferred = NULL; |
1da177e4 | 687 | } |
8e1f936b | 688 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 | 689 | |
880121be CK |
690 | /** |
691 | * synchronize_shrinkers - Wait for all running shrinkers to complete. | |
692 | * | |
693 | * This is equivalent to calling unregister_shrink() and register_shrinker(), | |
694 | * but atomically and with less overhead. This is useful to guarantee that all | |
695 | * shrinker invocations have seen an update, before freeing memory, similar to | |
696 | * rcu. | |
697 | */ | |
698 | void synchronize_shrinkers(void) | |
699 | { | |
700 | down_write(&shrinker_rwsem); | |
701 | up_write(&shrinker_rwsem); | |
702 | } | |
703 | EXPORT_SYMBOL(synchronize_shrinkers); | |
704 | ||
1da177e4 | 705 | #define SHRINK_BATCH 128 |
1d3d4437 | 706 | |
cb731d6c | 707 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
9092c71b | 708 | struct shrinker *shrinker, int priority) |
1d3d4437 GC |
709 | { |
710 | unsigned long freed = 0; | |
711 | unsigned long long delta; | |
712 | long total_scan; | |
d5bc5fd3 | 713 | long freeable; |
1d3d4437 GC |
714 | long nr; |
715 | long new_nr; | |
1d3d4437 GC |
716 | long batch_size = shrinker->batch ? shrinker->batch |
717 | : SHRINK_BATCH; | |
5f33a080 | 718 | long scanned = 0, next_deferred; |
1d3d4437 | 719 | |
d5bc5fd3 | 720 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
9b996468 KT |
721 | if (freeable == 0 || freeable == SHRINK_EMPTY) |
722 | return freeable; | |
1d3d4437 GC |
723 | |
724 | /* | |
725 | * copy the current shrinker scan count into a local variable | |
726 | * and zero it so that other concurrent shrinker invocations | |
727 | * don't also do this scanning work. | |
728 | */ | |
86750830 | 729 | nr = xchg_nr_deferred(shrinker, shrinkctl); |
1d3d4437 | 730 | |
4b85afbd JW |
731 | if (shrinker->seeks) { |
732 | delta = freeable >> priority; | |
733 | delta *= 4; | |
734 | do_div(delta, shrinker->seeks); | |
735 | } else { | |
736 | /* | |
737 | * These objects don't require any IO to create. Trim | |
738 | * them aggressively under memory pressure to keep | |
739 | * them from causing refetches in the IO caches. | |
740 | */ | |
741 | delta = freeable / 2; | |
742 | } | |
172b06c3 | 743 | |
18bb473e | 744 | total_scan = nr >> priority; |
1d3d4437 | 745 | total_scan += delta; |
18bb473e | 746 | total_scan = min(total_scan, (2 * freeable)); |
1d3d4437 GC |
747 | |
748 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
9092c71b | 749 | freeable, delta, total_scan, priority); |
1d3d4437 | 750 | |
0b1fb40a VD |
751 | /* |
752 | * Normally, we should not scan less than batch_size objects in one | |
753 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
754 | * than batch_size objects in total and we are really tight on memory, | |
755 | * we will try to reclaim all available objects, otherwise we can end | |
756 | * up failing allocations although there are plenty of reclaimable | |
757 | * objects spread over several slabs with usage less than the | |
758 | * batch_size. | |
759 | * | |
760 | * We detect the "tight on memory" situations by looking at the total | |
761 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 762 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
763 | * scanning at high prio and therefore should try to reclaim as much as |
764 | * possible. | |
765 | */ | |
766 | while (total_scan >= batch_size || | |
d5bc5fd3 | 767 | total_scan >= freeable) { |
a0b02131 | 768 | unsigned long ret; |
0b1fb40a | 769 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 770 | |
0b1fb40a | 771 | shrinkctl->nr_to_scan = nr_to_scan; |
d460acb5 | 772 | shrinkctl->nr_scanned = nr_to_scan; |
a0b02131 DC |
773 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
774 | if (ret == SHRINK_STOP) | |
775 | break; | |
776 | freed += ret; | |
1d3d4437 | 777 | |
d460acb5 CW |
778 | count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); |
779 | total_scan -= shrinkctl->nr_scanned; | |
780 | scanned += shrinkctl->nr_scanned; | |
1d3d4437 GC |
781 | |
782 | cond_resched(); | |
783 | } | |
784 | ||
18bb473e YS |
785 | /* |
786 | * The deferred work is increased by any new work (delta) that wasn't | |
787 | * done, decreased by old deferred work that was done now. | |
788 | * | |
789 | * And it is capped to two times of the freeable items. | |
790 | */ | |
791 | next_deferred = max_t(long, (nr + delta - scanned), 0); | |
792 | next_deferred = min(next_deferred, (2 * freeable)); | |
793 | ||
1d3d4437 GC |
794 | /* |
795 | * move the unused scan count back into the shrinker in a | |
86750830 | 796 | * manner that handles concurrent updates. |
1d3d4437 | 797 | */ |
86750830 | 798 | new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); |
1d3d4437 | 799 | |
8efb4b59 | 800 | trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 801 | return freed; |
1495f230 YH |
802 | } |
803 | ||
0a432dcb | 804 | #ifdef CONFIG_MEMCG |
b0dedc49 KT |
805 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
806 | struct mem_cgroup *memcg, int priority) | |
807 | { | |
e4262c4f | 808 | struct shrinker_info *info; |
b8e57efa KT |
809 | unsigned long ret, freed = 0; |
810 | int i; | |
b0dedc49 | 811 | |
0a432dcb | 812 | if (!mem_cgroup_online(memcg)) |
b0dedc49 KT |
813 | return 0; |
814 | ||
815 | if (!down_read_trylock(&shrinker_rwsem)) | |
816 | return 0; | |
817 | ||
468ab843 | 818 | info = shrinker_info_protected(memcg, nid); |
e4262c4f | 819 | if (unlikely(!info)) |
b0dedc49 KT |
820 | goto unlock; |
821 | ||
e4262c4f | 822 | for_each_set_bit(i, info->map, shrinker_nr_max) { |
b0dedc49 KT |
823 | struct shrink_control sc = { |
824 | .gfp_mask = gfp_mask, | |
825 | .nid = nid, | |
826 | .memcg = memcg, | |
827 | }; | |
828 | struct shrinker *shrinker; | |
829 | ||
830 | shrinker = idr_find(&shrinker_idr, i); | |
41ca668a | 831 | if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) { |
7e010df5 | 832 | if (!shrinker) |
e4262c4f | 833 | clear_bit(i, info->map); |
b0dedc49 KT |
834 | continue; |
835 | } | |
836 | ||
0a432dcb YS |
837 | /* Call non-slab shrinkers even though kmem is disabled */ |
838 | if (!memcg_kmem_enabled() && | |
839 | !(shrinker->flags & SHRINKER_NONSLAB)) | |
840 | continue; | |
841 | ||
b0dedc49 | 842 | ret = do_shrink_slab(&sc, shrinker, priority); |
f90280d6 | 843 | if (ret == SHRINK_EMPTY) { |
e4262c4f | 844 | clear_bit(i, info->map); |
f90280d6 KT |
845 | /* |
846 | * After the shrinker reported that it had no objects to | |
847 | * free, but before we cleared the corresponding bit in | |
848 | * the memcg shrinker map, a new object might have been | |
849 | * added. To make sure, we have the bit set in this | |
850 | * case, we invoke the shrinker one more time and reset | |
851 | * the bit if it reports that it is not empty anymore. | |
852 | * The memory barrier here pairs with the barrier in | |
2bfd3637 | 853 | * set_shrinker_bit(): |
f90280d6 KT |
854 | * |
855 | * list_lru_add() shrink_slab_memcg() | |
856 | * list_add_tail() clear_bit() | |
857 | * <MB> <MB> | |
858 | * set_bit() do_shrink_slab() | |
859 | */ | |
860 | smp_mb__after_atomic(); | |
861 | ret = do_shrink_slab(&sc, shrinker, priority); | |
862 | if (ret == SHRINK_EMPTY) | |
863 | ret = 0; | |
864 | else | |
2bfd3637 | 865 | set_shrinker_bit(memcg, nid, i); |
f90280d6 | 866 | } |
b0dedc49 KT |
867 | freed += ret; |
868 | ||
869 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
870 | freed = freed ? : 1; | |
871 | break; | |
872 | } | |
873 | } | |
874 | unlock: | |
875 | up_read(&shrinker_rwsem); | |
876 | return freed; | |
877 | } | |
0a432dcb | 878 | #else /* CONFIG_MEMCG */ |
b0dedc49 KT |
879 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
880 | struct mem_cgroup *memcg, int priority) | |
881 | { | |
882 | return 0; | |
883 | } | |
0a432dcb | 884 | #endif /* CONFIG_MEMCG */ |
b0dedc49 | 885 | |
6b4f7799 | 886 | /** |
cb731d6c | 887 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
888 | * @gfp_mask: allocation context |
889 | * @nid: node whose slab caches to target | |
cb731d6c | 890 | * @memcg: memory cgroup whose slab caches to target |
9092c71b | 891 | * @priority: the reclaim priority |
1da177e4 | 892 | * |
6b4f7799 | 893 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 894 | * |
6b4f7799 JW |
895 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
896 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 897 | * |
aeed1d32 VD |
898 | * @memcg specifies the memory cgroup to target. Unaware shrinkers |
899 | * are called only if it is the root cgroup. | |
cb731d6c | 900 | * |
9092c71b JB |
901 | * @priority is sc->priority, we take the number of objects and >> by priority |
902 | * in order to get the scan target. | |
b15e0905 | 903 | * |
6b4f7799 | 904 | * Returns the number of reclaimed slab objects. |
1da177e4 | 905 | */ |
cb731d6c VD |
906 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
907 | struct mem_cgroup *memcg, | |
9092c71b | 908 | int priority) |
1da177e4 | 909 | { |
b8e57efa | 910 | unsigned long ret, freed = 0; |
1da177e4 LT |
911 | struct shrinker *shrinker; |
912 | ||
fa1e512f YS |
913 | /* |
914 | * The root memcg might be allocated even though memcg is disabled | |
915 | * via "cgroup_disable=memory" boot parameter. This could make | |
916 | * mem_cgroup_is_root() return false, then just run memcg slab | |
917 | * shrink, but skip global shrink. This may result in premature | |
918 | * oom. | |
919 | */ | |
920 | if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg)) | |
b0dedc49 | 921 | return shrink_slab_memcg(gfp_mask, nid, memcg, priority); |
cb731d6c | 922 | |
e830c63a | 923 | if (!down_read_trylock(&shrinker_rwsem)) |
f06590bd | 924 | goto out; |
1da177e4 LT |
925 | |
926 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
927 | struct shrink_control sc = { |
928 | .gfp_mask = gfp_mask, | |
929 | .nid = nid, | |
cb731d6c | 930 | .memcg = memcg, |
6b4f7799 | 931 | }; |
ec97097b | 932 | |
9b996468 KT |
933 | ret = do_shrink_slab(&sc, shrinker, priority); |
934 | if (ret == SHRINK_EMPTY) | |
935 | ret = 0; | |
936 | freed += ret; | |
e496612c MK |
937 | /* |
938 | * Bail out if someone want to register a new shrinker to | |
55b65a57 | 939 | * prevent the registration from being stalled for long periods |
e496612c MK |
940 | * by parallel ongoing shrinking. |
941 | */ | |
942 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
943 | freed = freed ? : 1; | |
944 | break; | |
945 | } | |
1da177e4 | 946 | } |
6b4f7799 | 947 | |
1da177e4 | 948 | up_read(&shrinker_rwsem); |
f06590bd MK |
949 | out: |
950 | cond_resched(); | |
24f7c6b9 | 951 | return freed; |
1da177e4 LT |
952 | } |
953 | ||
e4b424b7 | 954 | static void drop_slab_node(int nid) |
cb731d6c VD |
955 | { |
956 | unsigned long freed; | |
1399af7e | 957 | int shift = 0; |
cb731d6c VD |
958 | |
959 | do { | |
960 | struct mem_cgroup *memcg = NULL; | |
961 | ||
069c411d CZ |
962 | if (fatal_signal_pending(current)) |
963 | return; | |
964 | ||
cb731d6c | 965 | freed = 0; |
aeed1d32 | 966 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
cb731d6c | 967 | do { |
9092c71b | 968 | freed += shrink_slab(GFP_KERNEL, nid, memcg, 0); |
cb731d6c | 969 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); |
1399af7e | 970 | } while ((freed >> shift++) > 1); |
cb731d6c VD |
971 | } |
972 | ||
973 | void drop_slab(void) | |
974 | { | |
975 | int nid; | |
976 | ||
977 | for_each_online_node(nid) | |
978 | drop_slab_node(nid); | |
979 | } | |
980 | ||
1da177e4 LT |
981 | static inline int is_page_cache_freeable(struct page *page) |
982 | { | |
ceddc3a5 JW |
983 | /* |
984 | * A freeable page cache page is referenced only by the caller | |
67891fff MW |
985 | * that isolated the page, the page cache and optional buffer |
986 | * heads at page->private. | |
ceddc3a5 | 987 | */ |
3efe62e4 | 988 | int page_cache_pins = thp_nr_pages(page); |
67891fff | 989 | return page_count(page) - page_has_private(page) == 1 + page_cache_pins; |
1da177e4 LT |
990 | } |
991 | ||
cb16556d | 992 | static int may_write_to_inode(struct inode *inode) |
1da177e4 | 993 | { |
930d9152 | 994 | if (current->flags & PF_SWAPWRITE) |
1da177e4 | 995 | return 1; |
703c2708 | 996 | if (!inode_write_congested(inode)) |
1da177e4 | 997 | return 1; |
703c2708 | 998 | if (inode_to_bdi(inode) == current->backing_dev_info) |
1da177e4 LT |
999 | return 1; |
1000 | return 0; | |
1001 | } | |
1002 | ||
1003 | /* | |
1004 | * We detected a synchronous write error writing a page out. Probably | |
1005 | * -ENOSPC. We need to propagate that into the address_space for a subsequent | |
1006 | * fsync(), msync() or close(). | |
1007 | * | |
1008 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
1009 | * prevents it from being freed up. But we have a ref on the page and once | |
1010 | * that page is locked, the mapping is pinned. | |
1011 | * | |
1012 | * We're allowed to run sleeping lock_page() here because we know the caller has | |
1013 | * __GFP_FS. | |
1014 | */ | |
1015 | static void handle_write_error(struct address_space *mapping, | |
1016 | struct page *page, int error) | |
1017 | { | |
7eaceacc | 1018 | lock_page(page); |
3e9f45bd GC |
1019 | if (page_mapping(page) == mapping) |
1020 | mapping_set_error(mapping, error); | |
1da177e4 LT |
1021 | unlock_page(page); |
1022 | } | |
1023 | ||
1b4e3f26 MG |
1024 | static bool skip_throttle_noprogress(pg_data_t *pgdat) |
1025 | { | |
1026 | int reclaimable = 0, write_pending = 0; | |
1027 | int i; | |
1028 | ||
1029 | /* | |
1030 | * If kswapd is disabled, reschedule if necessary but do not | |
1031 | * throttle as the system is likely near OOM. | |
1032 | */ | |
1033 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
1034 | return true; | |
1035 | ||
1036 | /* | |
1037 | * If there are a lot of dirty/writeback pages then do not | |
1038 | * throttle as throttling will occur when the pages cycle | |
1039 | * towards the end of the LRU if still under writeback. | |
1040 | */ | |
1041 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
1042 | struct zone *zone = pgdat->node_zones + i; | |
1043 | ||
1044 | if (!populated_zone(zone)) | |
1045 | continue; | |
1046 | ||
1047 | reclaimable += zone_reclaimable_pages(zone); | |
1048 | write_pending += zone_page_state_snapshot(zone, | |
1049 | NR_ZONE_WRITE_PENDING); | |
1050 | } | |
1051 | if (2 * write_pending <= reclaimable) | |
1052 | return true; | |
1053 | ||
1054 | return false; | |
1055 | } | |
1056 | ||
c3f4a9a2 | 1057 | void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason) |
8cd7c588 MG |
1058 | { |
1059 | wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason]; | |
c3f4a9a2 | 1060 | long timeout, ret; |
8cd7c588 MG |
1061 | DEFINE_WAIT(wait); |
1062 | ||
1063 | /* | |
1064 | * Do not throttle IO workers, kthreads other than kswapd or | |
1065 | * workqueues. They may be required for reclaim to make | |
1066 | * forward progress (e.g. journalling workqueues or kthreads). | |
1067 | */ | |
1068 | if (!current_is_kswapd() && | |
b485c6f1 MG |
1069 | current->flags & (PF_IO_WORKER|PF_KTHREAD)) { |
1070 | cond_resched(); | |
8cd7c588 | 1071 | return; |
b485c6f1 | 1072 | } |
8cd7c588 | 1073 | |
c3f4a9a2 MG |
1074 | /* |
1075 | * These figures are pulled out of thin air. | |
1076 | * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many | |
1077 | * parallel reclaimers which is a short-lived event so the timeout is | |
1078 | * short. Failing to make progress or waiting on writeback are | |
1079 | * potentially long-lived events so use a longer timeout. This is shaky | |
1080 | * logic as a failure to make progress could be due to anything from | |
1081 | * writeback to a slow device to excessive references pages at the tail | |
1082 | * of the inactive LRU. | |
1083 | */ | |
1084 | switch(reason) { | |
1085 | case VMSCAN_THROTTLE_WRITEBACK: | |
1086 | timeout = HZ/10; | |
1087 | ||
1088 | if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) { | |
1089 | WRITE_ONCE(pgdat->nr_reclaim_start, | |
1090 | node_page_state(pgdat, NR_THROTTLED_WRITTEN)); | |
1091 | } | |
1092 | ||
1093 | break; | |
1b4e3f26 MG |
1094 | case VMSCAN_THROTTLE_CONGESTED: |
1095 | fallthrough; | |
c3f4a9a2 | 1096 | case VMSCAN_THROTTLE_NOPROGRESS: |
1b4e3f26 MG |
1097 | if (skip_throttle_noprogress(pgdat)) { |
1098 | cond_resched(); | |
1099 | return; | |
1100 | } | |
1101 | ||
1102 | timeout = 1; | |
1103 | ||
c3f4a9a2 MG |
1104 | break; |
1105 | case VMSCAN_THROTTLE_ISOLATED: | |
1106 | timeout = HZ/50; | |
1107 | break; | |
1108 | default: | |
1109 | WARN_ON_ONCE(1); | |
1110 | timeout = HZ; | |
1111 | break; | |
8cd7c588 MG |
1112 | } |
1113 | ||
1114 | prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); | |
1115 | ret = schedule_timeout(timeout); | |
1116 | finish_wait(wqh, &wait); | |
d818fca1 | 1117 | |
c3f4a9a2 | 1118 | if (reason == VMSCAN_THROTTLE_WRITEBACK) |
d818fca1 | 1119 | atomic_dec(&pgdat->nr_writeback_throttled); |
8cd7c588 MG |
1120 | |
1121 | trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout), | |
1122 | jiffies_to_usecs(timeout - ret), | |
1123 | reason); | |
1124 | } | |
1125 | ||
1126 | /* | |
1127 | * Account for pages written if tasks are throttled waiting on dirty | |
1128 | * pages to clean. If enough pages have been cleaned since throttling | |
1129 | * started then wakeup the throttled tasks. | |
1130 | */ | |
512b7931 | 1131 | void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio, |
8cd7c588 MG |
1132 | int nr_throttled) |
1133 | { | |
1134 | unsigned long nr_written; | |
1135 | ||
512b7931 | 1136 | node_stat_add_folio(folio, NR_THROTTLED_WRITTEN); |
8cd7c588 MG |
1137 | |
1138 | /* | |
1139 | * This is an inaccurate read as the per-cpu deltas may not | |
1140 | * be synchronised. However, given that the system is | |
1141 | * writeback throttled, it is not worth taking the penalty | |
1142 | * of getting an accurate count. At worst, the throttle | |
1143 | * timeout guarantees forward progress. | |
1144 | */ | |
1145 | nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) - | |
1146 | READ_ONCE(pgdat->nr_reclaim_start); | |
1147 | ||
1148 | if (nr_written > SWAP_CLUSTER_MAX * nr_throttled) | |
1149 | wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]); | |
1150 | } | |
1151 | ||
04e62a29 CL |
1152 | /* possible outcome of pageout() */ |
1153 | typedef enum { | |
1154 | /* failed to write page out, page is locked */ | |
1155 | PAGE_KEEP, | |
1156 | /* move page to the active list, page is locked */ | |
1157 | PAGE_ACTIVATE, | |
1158 | /* page has been sent to the disk successfully, page is unlocked */ | |
1159 | PAGE_SUCCESS, | |
1160 | /* page is clean and locked */ | |
1161 | PAGE_CLEAN, | |
1162 | } pageout_t; | |
1163 | ||
1da177e4 | 1164 | /* |
1742f19f AM |
1165 | * pageout is called by shrink_page_list() for each dirty page. |
1166 | * Calls ->writepage(). | |
1da177e4 | 1167 | */ |
cb16556d | 1168 | static pageout_t pageout(struct page *page, struct address_space *mapping) |
1da177e4 LT |
1169 | { |
1170 | /* | |
1171 | * If the page is dirty, only perform writeback if that write | |
1172 | * will be non-blocking. To prevent this allocation from being | |
1173 | * stalled by pagecache activity. But note that there may be | |
1174 | * stalls if we need to run get_block(). We could test | |
1175 | * PagePrivate for that. | |
1176 | * | |
8174202b | 1177 | * If this process is currently in __generic_file_write_iter() against |
1da177e4 LT |
1178 | * this page's queue, we can perform writeback even if that |
1179 | * will block. | |
1180 | * | |
1181 | * If the page is swapcache, write it back even if that would | |
1182 | * block, for some throttling. This happens by accident, because | |
1183 | * swap_backing_dev_info is bust: it doesn't reflect the | |
1184 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 LT |
1185 | */ |
1186 | if (!is_page_cache_freeable(page)) | |
1187 | return PAGE_KEEP; | |
1188 | if (!mapping) { | |
1189 | /* | |
1190 | * Some data journaling orphaned pages can have | |
1191 | * page->mapping == NULL while being dirty with clean buffers. | |
1192 | */ | |
266cf658 | 1193 | if (page_has_private(page)) { |
1da177e4 LT |
1194 | if (try_to_free_buffers(page)) { |
1195 | ClearPageDirty(page); | |
b1de0d13 | 1196 | pr_info("%s: orphaned page\n", __func__); |
1da177e4 LT |
1197 | return PAGE_CLEAN; |
1198 | } | |
1199 | } | |
1200 | return PAGE_KEEP; | |
1201 | } | |
1202 | if (mapping->a_ops->writepage == NULL) | |
1203 | return PAGE_ACTIVATE; | |
cb16556d | 1204 | if (!may_write_to_inode(mapping->host)) |
1da177e4 LT |
1205 | return PAGE_KEEP; |
1206 | ||
1207 | if (clear_page_dirty_for_io(page)) { | |
1208 | int res; | |
1209 | struct writeback_control wbc = { | |
1210 | .sync_mode = WB_SYNC_NONE, | |
1211 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
1212 | .range_start = 0, |
1213 | .range_end = LLONG_MAX, | |
1da177e4 LT |
1214 | .for_reclaim = 1, |
1215 | }; | |
1216 | ||
1217 | SetPageReclaim(page); | |
1218 | res = mapping->a_ops->writepage(page, &wbc); | |
1219 | if (res < 0) | |
1220 | handle_write_error(mapping, page, res); | |
994fc28c | 1221 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
1da177e4 LT |
1222 | ClearPageReclaim(page); |
1223 | return PAGE_ACTIVATE; | |
1224 | } | |
c661b078 | 1225 | |
1da177e4 LT |
1226 | if (!PageWriteback(page)) { |
1227 | /* synchronous write or broken a_ops? */ | |
1228 | ClearPageReclaim(page); | |
1229 | } | |
3aa23851 | 1230 | trace_mm_vmscan_writepage(page); |
c4a25635 | 1231 | inc_node_page_state(page, NR_VMSCAN_WRITE); |
1da177e4 LT |
1232 | return PAGE_SUCCESS; |
1233 | } | |
1234 | ||
1235 | return PAGE_CLEAN; | |
1236 | } | |
1237 | ||
a649fd92 | 1238 | /* |
e286781d NP |
1239 | * Same as remove_mapping, but if the page is removed from the mapping, it |
1240 | * gets returned with a refcount of 0. | |
a649fd92 | 1241 | */ |
be7c07d6 | 1242 | static int __remove_mapping(struct address_space *mapping, struct folio *folio, |
b910718a | 1243 | bool reclaimed, struct mem_cgroup *target_memcg) |
49d2e9cc | 1244 | { |
bd4c82c2 | 1245 | int refcount; |
aae466b0 | 1246 | void *shadow = NULL; |
c4843a75 | 1247 | |
be7c07d6 MWO |
1248 | BUG_ON(!folio_test_locked(folio)); |
1249 | BUG_ON(mapping != folio_mapping(folio)); | |
49d2e9cc | 1250 | |
be7c07d6 | 1251 | if (!folio_test_swapcache(folio)) |
51b8c1fe | 1252 | spin_lock(&mapping->host->i_lock); |
30472509 | 1253 | xa_lock_irq(&mapping->i_pages); |
49d2e9cc | 1254 | /* |
0fd0e6b0 NP |
1255 | * The non racy check for a busy page. |
1256 | * | |
1257 | * Must be careful with the order of the tests. When someone has | |
1258 | * a ref to the page, it may be possible that they dirty it then | |
1259 | * drop the reference. So if PageDirty is tested before page_count | |
1260 | * here, then the following race may occur: | |
1261 | * | |
1262 | * get_user_pages(&page); | |
1263 | * [user mapping goes away] | |
1264 | * write_to(page); | |
1265 | * !PageDirty(page) [good] | |
1266 | * SetPageDirty(page); | |
1267 | * put_page(page); | |
1268 | * !page_count(page) [good, discard it] | |
1269 | * | |
1270 | * [oops, our write_to data is lost] | |
1271 | * | |
1272 | * Reversing the order of the tests ensures such a situation cannot | |
1273 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 1274 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
1275 | * |
1276 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
b93b0163 | 1277 | * and thus under the i_pages lock, then this ordering is not required. |
49d2e9cc | 1278 | */ |
be7c07d6 MWO |
1279 | refcount = 1 + folio_nr_pages(folio); |
1280 | if (!folio_ref_freeze(folio, refcount)) | |
49d2e9cc | 1281 | goto cannot_free; |
1c4c3b99 | 1282 | /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */ |
be7c07d6 MWO |
1283 | if (unlikely(folio_test_dirty(folio))) { |
1284 | folio_ref_unfreeze(folio, refcount); | |
49d2e9cc | 1285 | goto cannot_free; |
e286781d | 1286 | } |
49d2e9cc | 1287 | |
be7c07d6 MWO |
1288 | if (folio_test_swapcache(folio)) { |
1289 | swp_entry_t swap = folio_swap_entry(folio); | |
3ecb0087 | 1290 | mem_cgroup_swapout(folio, swap); |
aae466b0 | 1291 | if (reclaimed && !mapping_exiting(mapping)) |
8927f647 | 1292 | shadow = workingset_eviction(folio, target_memcg); |
be7c07d6 | 1293 | __delete_from_swap_cache(&folio->page, swap, shadow); |
30472509 | 1294 | xa_unlock_irq(&mapping->i_pages); |
be7c07d6 | 1295 | put_swap_page(&folio->page, swap); |
e286781d | 1296 | } else { |
6072d13c LT |
1297 | void (*freepage)(struct page *); |
1298 | ||
1299 | freepage = mapping->a_ops->freepage; | |
a528910e JW |
1300 | /* |
1301 | * Remember a shadow entry for reclaimed file cache in | |
1302 | * order to detect refaults, thus thrashing, later on. | |
1303 | * | |
1304 | * But don't store shadows in an address space that is | |
238c3046 | 1305 | * already exiting. This is not just an optimization, |
a528910e JW |
1306 | * inode reclaim needs to empty out the radix tree or |
1307 | * the nodes are lost. Don't plant shadows behind its | |
1308 | * back. | |
f9fe48be RZ |
1309 | * |
1310 | * We also don't store shadows for DAX mappings because the | |
1311 | * only page cache pages found in these are zero pages | |
1312 | * covering holes, and because we don't want to mix DAX | |
1313 | * exceptional entries and shadow exceptional entries in the | |
b93b0163 | 1314 | * same address_space. |
a528910e | 1315 | */ |
be7c07d6 | 1316 | if (reclaimed && folio_is_file_lru(folio) && |
f9fe48be | 1317 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
8927f647 MWO |
1318 | shadow = workingset_eviction(folio, target_memcg); |
1319 | __filemap_remove_folio(folio, shadow); | |
30472509 | 1320 | xa_unlock_irq(&mapping->i_pages); |
51b8c1fe JW |
1321 | if (mapping_shrinkable(mapping)) |
1322 | inode_add_lru(mapping->host); | |
1323 | spin_unlock(&mapping->host->i_lock); | |
6072d13c LT |
1324 | |
1325 | if (freepage != NULL) | |
be7c07d6 | 1326 | freepage(&folio->page); |
49d2e9cc CL |
1327 | } |
1328 | ||
49d2e9cc CL |
1329 | return 1; |
1330 | ||
1331 | cannot_free: | |
30472509 | 1332 | xa_unlock_irq(&mapping->i_pages); |
be7c07d6 | 1333 | if (!folio_test_swapcache(folio)) |
51b8c1fe | 1334 | spin_unlock(&mapping->host->i_lock); |
49d2e9cc CL |
1335 | return 0; |
1336 | } | |
1337 | ||
5100da38 MWO |
1338 | /** |
1339 | * remove_mapping() - Attempt to remove a folio from its mapping. | |
1340 | * @mapping: The address space. | |
1341 | * @folio: The folio to remove. | |
1342 | * | |
1343 | * If the folio is dirty, under writeback or if someone else has a ref | |
1344 | * on it, removal will fail. | |
1345 | * Return: The number of pages removed from the mapping. 0 if the folio | |
1346 | * could not be removed. | |
1347 | * Context: The caller should have a single refcount on the folio and | |
1348 | * hold its lock. | |
e286781d | 1349 | */ |
5100da38 | 1350 | long remove_mapping(struct address_space *mapping, struct folio *folio) |
e286781d | 1351 | { |
be7c07d6 | 1352 | if (__remove_mapping(mapping, folio, false, NULL)) { |
e286781d | 1353 | /* |
5100da38 | 1354 | * Unfreezing the refcount with 1 effectively |
e286781d NP |
1355 | * drops the pagecache ref for us without requiring another |
1356 | * atomic operation. | |
1357 | */ | |
be7c07d6 | 1358 | folio_ref_unfreeze(folio, 1); |
5100da38 | 1359 | return folio_nr_pages(folio); |
e286781d NP |
1360 | } |
1361 | return 0; | |
1362 | } | |
1363 | ||
894bc310 | 1364 | /** |
ca6d60f3 MWO |
1365 | * folio_putback_lru - Put previously isolated folio onto appropriate LRU list. |
1366 | * @folio: Folio to be returned to an LRU list. | |
894bc310 | 1367 | * |
ca6d60f3 MWO |
1368 | * Add previously isolated @folio to appropriate LRU list. |
1369 | * The folio may still be unevictable for other reasons. | |
894bc310 | 1370 | * |
ca6d60f3 | 1371 | * Context: lru_lock must not be held, interrupts must be enabled. |
894bc310 | 1372 | */ |
ca6d60f3 | 1373 | void folio_putback_lru(struct folio *folio) |
894bc310 | 1374 | { |
ca6d60f3 MWO |
1375 | folio_add_lru(folio); |
1376 | folio_put(folio); /* drop ref from isolate */ | |
894bc310 LS |
1377 | } |
1378 | ||
dfc8d636 JW |
1379 | enum page_references { |
1380 | PAGEREF_RECLAIM, | |
1381 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 1382 | PAGEREF_KEEP, |
dfc8d636 JW |
1383 | PAGEREF_ACTIVATE, |
1384 | }; | |
1385 | ||
1386 | static enum page_references page_check_references(struct page *page, | |
1387 | struct scan_control *sc) | |
1388 | { | |
b3ac0413 | 1389 | struct folio *folio = page_folio(page); |
64574746 | 1390 | int referenced_ptes, referenced_page; |
dfc8d636 | 1391 | unsigned long vm_flags; |
dfc8d636 | 1392 | |
b3ac0413 MWO |
1393 | referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup, |
1394 | &vm_flags); | |
64574746 | 1395 | referenced_page = TestClearPageReferenced(page); |
dfc8d636 | 1396 | |
dfc8d636 | 1397 | /* |
47d4f3ee HD |
1398 | * The supposedly reclaimable page was found to be in a VM_LOCKED vma. |
1399 | * Let the page, now marked Mlocked, be moved to the unevictable list. | |
dfc8d636 JW |
1400 | */ |
1401 | if (vm_flags & VM_LOCKED) | |
47d4f3ee | 1402 | return PAGEREF_ACTIVATE; |
dfc8d636 | 1403 | |
64574746 | 1404 | if (referenced_ptes) { |
64574746 JW |
1405 | /* |
1406 | * All mapped pages start out with page table | |
1407 | * references from the instantiating fault, so we need | |
1408 | * to look twice if a mapped file page is used more | |
1409 | * than once. | |
1410 | * | |
1411 | * Mark it and spare it for another trip around the | |
1412 | * inactive list. Another page table reference will | |
1413 | * lead to its activation. | |
1414 | * | |
1415 | * Note: the mark is set for activated pages as well | |
1416 | * so that recently deactivated but used pages are | |
1417 | * quickly recovered. | |
1418 | */ | |
1419 | SetPageReferenced(page); | |
1420 | ||
34dbc67a | 1421 | if (referenced_page || referenced_ptes > 1) |
64574746 JW |
1422 | return PAGEREF_ACTIVATE; |
1423 | ||
c909e993 KK |
1424 | /* |
1425 | * Activate file-backed executable pages after first usage. | |
1426 | */ | |
b518154e | 1427 | if ((vm_flags & VM_EXEC) && !PageSwapBacked(page)) |
c909e993 KK |
1428 | return PAGEREF_ACTIVATE; |
1429 | ||
64574746 JW |
1430 | return PAGEREF_KEEP; |
1431 | } | |
dfc8d636 JW |
1432 | |
1433 | /* Reclaim if clean, defer dirty pages to writeback */ | |
2e30244a | 1434 | if (referenced_page && !PageSwapBacked(page)) |
64574746 JW |
1435 | return PAGEREF_RECLAIM_CLEAN; |
1436 | ||
1437 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
1438 | } |
1439 | ||
e2be15f6 | 1440 | /* Check if a page is dirty or under writeback */ |
e20c41b1 | 1441 | static void folio_check_dirty_writeback(struct folio *folio, |
e2be15f6 MG |
1442 | bool *dirty, bool *writeback) |
1443 | { | |
b4597226 MG |
1444 | struct address_space *mapping; |
1445 | ||
e2be15f6 MG |
1446 | /* |
1447 | * Anonymous pages are not handled by flushers and must be written | |
1448 | * from reclaim context. Do not stall reclaim based on them | |
1449 | */ | |
e20c41b1 MWO |
1450 | if (!folio_is_file_lru(folio) || |
1451 | (folio_test_anon(folio) && !folio_test_swapbacked(folio))) { | |
e2be15f6 MG |
1452 | *dirty = false; |
1453 | *writeback = false; | |
1454 | return; | |
1455 | } | |
1456 | ||
e20c41b1 MWO |
1457 | /* By default assume that the folio flags are accurate */ |
1458 | *dirty = folio_test_dirty(folio); | |
1459 | *writeback = folio_test_writeback(folio); | |
b4597226 MG |
1460 | |
1461 | /* Verify dirty/writeback state if the filesystem supports it */ | |
e20c41b1 | 1462 | if (!folio_test_private(folio)) |
b4597226 MG |
1463 | return; |
1464 | ||
e20c41b1 | 1465 | mapping = folio_mapping(folio); |
b4597226 | 1466 | if (mapping && mapping->a_ops->is_dirty_writeback) |
e20c41b1 | 1467 | mapping->a_ops->is_dirty_writeback(&folio->page, dirty, writeback); |
e2be15f6 MG |
1468 | } |
1469 | ||
26aa2d19 DH |
1470 | static struct page *alloc_demote_page(struct page *page, unsigned long node) |
1471 | { | |
1472 | struct migration_target_control mtc = { | |
1473 | /* | |
1474 | * Allocate from 'node', or fail quickly and quietly. | |
1475 | * When this happens, 'page' will likely just be discarded | |
1476 | * instead of migrated. | |
1477 | */ | |
1478 | .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | | |
1479 | __GFP_THISNODE | __GFP_NOWARN | | |
1480 | __GFP_NOMEMALLOC | GFP_NOWAIT, | |
1481 | .nid = node | |
1482 | }; | |
1483 | ||
1484 | return alloc_migration_target(page, (unsigned long)&mtc); | |
1485 | } | |
1486 | ||
1487 | /* | |
1488 | * Take pages on @demote_list and attempt to demote them to | |
1489 | * another node. Pages which are not demoted are left on | |
1490 | * @demote_pages. | |
1491 | */ | |
1492 | static unsigned int demote_page_list(struct list_head *demote_pages, | |
1493 | struct pglist_data *pgdat) | |
1494 | { | |
1495 | int target_nid = next_demotion_node(pgdat->node_id); | |
1496 | unsigned int nr_succeeded; | |
26aa2d19 DH |
1497 | |
1498 | if (list_empty(demote_pages)) | |
1499 | return 0; | |
1500 | ||
1501 | if (target_nid == NUMA_NO_NODE) | |
1502 | return 0; | |
1503 | ||
1504 | /* Demotion ignores all cpuset and mempolicy settings */ | |
cb75463c | 1505 | migrate_pages(demote_pages, alloc_demote_page, NULL, |
26aa2d19 DH |
1506 | target_nid, MIGRATE_ASYNC, MR_DEMOTION, |
1507 | &nr_succeeded); | |
1508 | ||
668e4147 YS |
1509 | if (current_is_kswapd()) |
1510 | __count_vm_events(PGDEMOTE_KSWAPD, nr_succeeded); | |
1511 | else | |
1512 | __count_vm_events(PGDEMOTE_DIRECT, nr_succeeded); | |
1513 | ||
26aa2d19 DH |
1514 | return nr_succeeded; |
1515 | } | |
1516 | ||
1da177e4 | 1517 | /* |
1742f19f | 1518 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 1519 | */ |
730ec8c0 MS |
1520 | static unsigned int shrink_page_list(struct list_head *page_list, |
1521 | struct pglist_data *pgdat, | |
1522 | struct scan_control *sc, | |
730ec8c0 MS |
1523 | struct reclaim_stat *stat, |
1524 | bool ignore_references) | |
1da177e4 LT |
1525 | { |
1526 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 1527 | LIST_HEAD(free_pages); |
26aa2d19 | 1528 | LIST_HEAD(demote_pages); |
730ec8c0 MS |
1529 | unsigned int nr_reclaimed = 0; |
1530 | unsigned int pgactivate = 0; | |
26aa2d19 | 1531 | bool do_demote_pass; |
1da177e4 | 1532 | |
060f005f | 1533 | memset(stat, 0, sizeof(*stat)); |
1da177e4 | 1534 | cond_resched(); |
26aa2d19 | 1535 | do_demote_pass = can_demote(pgdat->node_id, sc); |
1da177e4 | 1536 | |
26aa2d19 | 1537 | retry: |
1da177e4 LT |
1538 | while (!list_empty(page_list)) { |
1539 | struct address_space *mapping; | |
1540 | struct page *page; | |
be7c07d6 | 1541 | struct folio *folio; |
8940b34a | 1542 | enum page_references references = PAGEREF_RECLAIM; |
4b793062 | 1543 | bool dirty, writeback, may_enter_fs; |
98879b3b | 1544 | unsigned int nr_pages; |
1da177e4 LT |
1545 | |
1546 | cond_resched(); | |
1547 | ||
be7c07d6 MWO |
1548 | folio = lru_to_folio(page_list); |
1549 | list_del(&folio->lru); | |
1550 | page = &folio->page; | |
1da177e4 | 1551 | |
529ae9aa | 1552 | if (!trylock_page(page)) |
1da177e4 LT |
1553 | goto keep; |
1554 | ||
309381fe | 1555 | VM_BUG_ON_PAGE(PageActive(page), page); |
1da177e4 | 1556 | |
d8c6546b | 1557 | nr_pages = compound_nr(page); |
98879b3b YS |
1558 | |
1559 | /* Account the number of base pages even though THP */ | |
1560 | sc->nr_scanned += nr_pages; | |
80e43426 | 1561 | |
39b5f29a | 1562 | if (unlikely(!page_evictable(page))) |
ad6b6704 | 1563 | goto activate_locked; |
894bc310 | 1564 | |
a6dc60f8 | 1565 | if (!sc->may_unmap && page_mapped(page)) |
80e43426 CL |
1566 | goto keep_locked; |
1567 | ||
c661b078 AW |
1568 | may_enter_fs = (sc->gfp_mask & __GFP_FS) || |
1569 | (PageSwapCache(page) && (sc->gfp_mask & __GFP_IO)); | |
1570 | ||
e2be15f6 | 1571 | /* |
894befec | 1572 | * The number of dirty pages determines if a node is marked |
8cd7c588 MG |
1573 | * reclaim_congested. kswapd will stall and start writing |
1574 | * pages if the tail of the LRU is all dirty unqueued pages. | |
e2be15f6 | 1575 | */ |
e20c41b1 | 1576 | folio_check_dirty_writeback(folio, &dirty, &writeback); |
e2be15f6 | 1577 | if (dirty || writeback) |
c79b7b96 | 1578 | stat->nr_dirty += nr_pages; |
e2be15f6 MG |
1579 | |
1580 | if (dirty && !writeback) | |
c79b7b96 | 1581 | stat->nr_unqueued_dirty += nr_pages; |
e2be15f6 | 1582 | |
d04e8acd MG |
1583 | /* |
1584 | * Treat this page as congested if the underlying BDI is or if | |
1585 | * pages are cycling through the LRU so quickly that the | |
1586 | * pages marked for immediate reclaim are making it to the | |
1587 | * end of the LRU a second time. | |
1588 | */ | |
e2be15f6 | 1589 | mapping = page_mapping(page); |
1da58ee2 | 1590 | if (((dirty || writeback) && mapping && |
703c2708 | 1591 | inode_write_congested(mapping->host)) || |
d04e8acd | 1592 | (writeback && PageReclaim(page))) |
c79b7b96 | 1593 | stat->nr_congested += nr_pages; |
e2be15f6 | 1594 | |
283aba9f MG |
1595 | /* |
1596 | * If a page at the tail of the LRU is under writeback, there | |
1597 | * are three cases to consider. | |
1598 | * | |
1599 | * 1) If reclaim is encountering an excessive number of pages | |
1600 | * under writeback and this page is both under writeback and | |
1601 | * PageReclaim then it indicates that pages are being queued | |
1602 | * for IO but are being recycled through the LRU before the | |
1603 | * IO can complete. Waiting on the page itself risks an | |
1604 | * indefinite stall if it is impossible to writeback the | |
1605 | * page due to IO error or disconnected storage so instead | |
b1a6f21e MG |
1606 | * note that the LRU is being scanned too quickly and the |
1607 | * caller can stall after page list has been processed. | |
283aba9f | 1608 | * |
97c9341f | 1609 | * 2) Global or new memcg reclaim encounters a page that is |
ecf5fc6e MH |
1610 | * not marked for immediate reclaim, or the caller does not |
1611 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
1612 | * not to fs). In this case mark the page for immediate | |
97c9341f | 1613 | * reclaim and continue scanning. |
283aba9f | 1614 | * |
ecf5fc6e MH |
1615 | * Require may_enter_fs because we would wait on fs, which |
1616 | * may not have submitted IO yet. And the loop driver might | |
283aba9f MG |
1617 | * enter reclaim, and deadlock if it waits on a page for |
1618 | * which it is needed to do the write (loop masks off | |
1619 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1620 | * would probably show more reasons. | |
1621 | * | |
7fadc820 | 1622 | * 3) Legacy memcg encounters a page that is already marked |
283aba9f MG |
1623 | * PageReclaim. memcg does not have any dirty pages |
1624 | * throttling so we could easily OOM just because too many | |
1625 | * pages are in writeback and there is nothing else to | |
1626 | * reclaim. Wait for the writeback to complete. | |
c55e8d03 JW |
1627 | * |
1628 | * In cases 1) and 2) we activate the pages to get them out of | |
1629 | * the way while we continue scanning for clean pages on the | |
1630 | * inactive list and refilling from the active list. The | |
1631 | * observation here is that waiting for disk writes is more | |
1632 | * expensive than potentially causing reloads down the line. | |
1633 | * Since they're marked for immediate reclaim, they won't put | |
1634 | * memory pressure on the cache working set any longer than it | |
1635 | * takes to write them to disk. | |
283aba9f | 1636 | */ |
c661b078 | 1637 | if (PageWriteback(page)) { |
283aba9f MG |
1638 | /* Case 1 above */ |
1639 | if (current_is_kswapd() && | |
1640 | PageReclaim(page) && | |
599d0c95 | 1641 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
c79b7b96 | 1642 | stat->nr_immediate += nr_pages; |
c55e8d03 | 1643 | goto activate_locked; |
283aba9f MG |
1644 | |
1645 | /* Case 2 above */ | |
b5ead35e | 1646 | } else if (writeback_throttling_sane(sc) || |
ecf5fc6e | 1647 | !PageReclaim(page) || !may_enter_fs) { |
c3b94f44 HD |
1648 | /* |
1649 | * This is slightly racy - end_page_writeback() | |
1650 | * might have just cleared PageReclaim, then | |
1651 | * setting PageReclaim here end up interpreted | |
1652 | * as PageReadahead - but that does not matter | |
1653 | * enough to care. What we do want is for this | |
1654 | * page to have PageReclaim set next time memcg | |
1655 | * reclaim reaches the tests above, so it will | |
1656 | * then wait_on_page_writeback() to avoid OOM; | |
1657 | * and it's also appropriate in global reclaim. | |
1658 | */ | |
1659 | SetPageReclaim(page); | |
c79b7b96 | 1660 | stat->nr_writeback += nr_pages; |
c55e8d03 | 1661 | goto activate_locked; |
283aba9f MG |
1662 | |
1663 | /* Case 3 above */ | |
1664 | } else { | |
7fadc820 | 1665 | unlock_page(page); |
283aba9f | 1666 | wait_on_page_writeback(page); |
7fadc820 HD |
1667 | /* then go back and try same page again */ |
1668 | list_add_tail(&page->lru, page_list); | |
1669 | continue; | |
e62e384e | 1670 | } |
c661b078 | 1671 | } |
1da177e4 | 1672 | |
8940b34a | 1673 | if (!ignore_references) |
02c6de8d MK |
1674 | references = page_check_references(page, sc); |
1675 | ||
dfc8d636 JW |
1676 | switch (references) { |
1677 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1678 | goto activate_locked; |
64574746 | 1679 | case PAGEREF_KEEP: |
98879b3b | 1680 | stat->nr_ref_keep += nr_pages; |
64574746 | 1681 | goto keep_locked; |
dfc8d636 JW |
1682 | case PAGEREF_RECLAIM: |
1683 | case PAGEREF_RECLAIM_CLEAN: | |
1684 | ; /* try to reclaim the page below */ | |
1685 | } | |
1da177e4 | 1686 | |
26aa2d19 DH |
1687 | /* |
1688 | * Before reclaiming the page, try to relocate | |
1689 | * its contents to another node. | |
1690 | */ | |
1691 | if (do_demote_pass && | |
1692 | (thp_migration_supported() || !PageTransHuge(page))) { | |
1693 | list_add(&page->lru, &demote_pages); | |
1694 | unlock_page(page); | |
1695 | continue; | |
1696 | } | |
1697 | ||
1da177e4 LT |
1698 | /* |
1699 | * Anonymous process memory has backing store? | |
1700 | * Try to allocate it some swap space here. | |
802a3a92 | 1701 | * Lazyfree page could be freed directly |
1da177e4 | 1702 | */ |
bd4c82c2 HY |
1703 | if (PageAnon(page) && PageSwapBacked(page)) { |
1704 | if (!PageSwapCache(page)) { | |
1705 | if (!(sc->gfp_mask & __GFP_IO)) | |
1706 | goto keep_locked; | |
feb889fb LT |
1707 | if (page_maybe_dma_pinned(page)) |
1708 | goto keep_locked; | |
bd4c82c2 HY |
1709 | if (PageTransHuge(page)) { |
1710 | /* cannot split THP, skip it */ | |
1711 | if (!can_split_huge_page(page, NULL)) | |
1712 | goto activate_locked; | |
1713 | /* | |
1714 | * Split pages without a PMD map right | |
1715 | * away. Chances are some or all of the | |
1716 | * tail pages can be freed without IO. | |
1717 | */ | |
1718 | if (!compound_mapcount(page) && | |
346cf613 MWO |
1719 | split_folio_to_list(folio, |
1720 | page_list)) | |
bd4c82c2 HY |
1721 | goto activate_locked; |
1722 | } | |
1723 | if (!add_to_swap(page)) { | |
1724 | if (!PageTransHuge(page)) | |
98879b3b | 1725 | goto activate_locked_split; |
bd4c82c2 | 1726 | /* Fallback to swap normal pages */ |
346cf613 MWO |
1727 | if (split_folio_to_list(folio, |
1728 | page_list)) | |
bd4c82c2 | 1729 | goto activate_locked; |
fe490cc0 HY |
1730 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1731 | count_vm_event(THP_SWPOUT_FALLBACK); | |
1732 | #endif | |
bd4c82c2 | 1733 | if (!add_to_swap(page)) |
98879b3b | 1734 | goto activate_locked_split; |
bd4c82c2 | 1735 | } |
0f074658 | 1736 | |
4b793062 | 1737 | may_enter_fs = true; |
1da177e4 | 1738 | |
bd4c82c2 HY |
1739 | /* Adding to swap updated mapping */ |
1740 | mapping = page_mapping(page); | |
1741 | } | |
820c4e2e MWO |
1742 | } else if (PageSwapBacked(page) && PageTransHuge(page)) { |
1743 | /* Split shmem THP */ | |
346cf613 | 1744 | if (split_folio_to_list(folio, page_list)) |
7751b2da | 1745 | goto keep_locked; |
e2be15f6 | 1746 | } |
1da177e4 | 1747 | |
98879b3b YS |
1748 | /* |
1749 | * THP may get split above, need minus tail pages and update | |
1750 | * nr_pages to avoid accounting tail pages twice. | |
1751 | * | |
1752 | * The tail pages that are added into swap cache successfully | |
1753 | * reach here. | |
1754 | */ | |
1755 | if ((nr_pages > 1) && !PageTransHuge(page)) { | |
1756 | sc->nr_scanned -= (nr_pages - 1); | |
1757 | nr_pages = 1; | |
1758 | } | |
1759 | ||
1da177e4 LT |
1760 | /* |
1761 | * The page is mapped into the page tables of one or more | |
1762 | * processes. Try to unmap it here. | |
1763 | */ | |
802a3a92 | 1764 | if (page_mapped(page)) { |
013339df | 1765 | enum ttu_flags flags = TTU_BATCH_FLUSH; |
1f318a9b | 1766 | bool was_swapbacked = PageSwapBacked(page); |
bd4c82c2 | 1767 | |
343b2888 MWO |
1768 | if (PageTransHuge(page) && |
1769 | thp_order(page) >= HPAGE_PMD_ORDER) | |
bd4c82c2 | 1770 | flags |= TTU_SPLIT_HUGE_PMD; |
1f318a9b | 1771 | |
869f7ee6 | 1772 | try_to_unmap(folio, flags); |
1fb08ac6 | 1773 | if (page_mapped(page)) { |
98879b3b | 1774 | stat->nr_unmap_fail += nr_pages; |
1f318a9b JK |
1775 | if (!was_swapbacked && PageSwapBacked(page)) |
1776 | stat->nr_lazyfree_fail += nr_pages; | |
1da177e4 | 1777 | goto activate_locked; |
1da177e4 LT |
1778 | } |
1779 | } | |
1780 | ||
1781 | if (PageDirty(page)) { | |
ee72886d | 1782 | /* |
4eda4823 JW |
1783 | * Only kswapd can writeback filesystem pages |
1784 | * to avoid risk of stack overflow. But avoid | |
1785 | * injecting inefficient single-page IO into | |
1786 | * flusher writeback as much as possible: only | |
1787 | * write pages when we've encountered many | |
1788 | * dirty pages, and when we've already scanned | |
1789 | * the rest of the LRU for clean pages and see | |
1790 | * the same dirty pages again (PageReclaim). | |
ee72886d | 1791 | */ |
9de4f22a | 1792 | if (page_is_file_lru(page) && |
4eda4823 JW |
1793 | (!current_is_kswapd() || !PageReclaim(page) || |
1794 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { | |
49ea7eb6 MG |
1795 | /* |
1796 | * Immediately reclaim when written back. | |
1797 | * Similar in principal to deactivate_page() | |
1798 | * except we already have the page isolated | |
1799 | * and know it's dirty | |
1800 | */ | |
c4a25635 | 1801 | inc_node_page_state(page, NR_VMSCAN_IMMEDIATE); |
49ea7eb6 MG |
1802 | SetPageReclaim(page); |
1803 | ||
c55e8d03 | 1804 | goto activate_locked; |
ee72886d MG |
1805 | } |
1806 | ||
dfc8d636 | 1807 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1808 | goto keep_locked; |
4dd4b920 | 1809 | if (!may_enter_fs) |
1da177e4 | 1810 | goto keep_locked; |
52a8363e | 1811 | if (!sc->may_writepage) |
1da177e4 LT |
1812 | goto keep_locked; |
1813 | ||
d950c947 MG |
1814 | /* |
1815 | * Page is dirty. Flush the TLB if a writable entry | |
1816 | * potentially exists to avoid CPU writes after IO | |
1817 | * starts and then write it out here. | |
1818 | */ | |
1819 | try_to_unmap_flush_dirty(); | |
cb16556d | 1820 | switch (pageout(page, mapping)) { |
1da177e4 LT |
1821 | case PAGE_KEEP: |
1822 | goto keep_locked; | |
1823 | case PAGE_ACTIVATE: | |
1824 | goto activate_locked; | |
1825 | case PAGE_SUCCESS: | |
c79b7b96 | 1826 | stat->nr_pageout += nr_pages; |
96f8bf4f | 1827 | |
7d3579e8 | 1828 | if (PageWriteback(page)) |
41ac1999 | 1829 | goto keep; |
7d3579e8 | 1830 | if (PageDirty(page)) |
1da177e4 | 1831 | goto keep; |
7d3579e8 | 1832 | |
1da177e4 LT |
1833 | /* |
1834 | * A synchronous write - probably a ramdisk. Go | |
1835 | * ahead and try to reclaim the page. | |
1836 | */ | |
529ae9aa | 1837 | if (!trylock_page(page)) |
1da177e4 LT |
1838 | goto keep; |
1839 | if (PageDirty(page) || PageWriteback(page)) | |
1840 | goto keep_locked; | |
1841 | mapping = page_mapping(page); | |
01359eb2 | 1842 | fallthrough; |
1da177e4 LT |
1843 | case PAGE_CLEAN: |
1844 | ; /* try to free the page below */ | |
1845 | } | |
1846 | } | |
1847 | ||
1848 | /* | |
1849 | * If the page has buffers, try to free the buffer mappings | |
1850 | * associated with this page. If we succeed we try to free | |
1851 | * the page as well. | |
1852 | * | |
1853 | * We do this even if the page is PageDirty(). | |
1854 | * try_to_release_page() does not perform I/O, but it is | |
1855 | * possible for a page to have PageDirty set, but it is actually | |
1856 | * clean (all its buffers are clean). This happens if the | |
1857 | * buffers were written out directly, with submit_bh(). ext3 | |
894bc310 | 1858 | * will do this, as well as the blockdev mapping. |
1da177e4 LT |
1859 | * try_to_release_page() will discover that cleanness and will |
1860 | * drop the buffers and mark the page clean - it can be freed. | |
1861 | * | |
1862 | * Rarely, pages can have buffers and no ->mapping. These are | |
1863 | * the pages which were not successfully invalidated in | |
d12b8951 | 1864 | * truncate_cleanup_page(). We try to drop those buffers here |
1da177e4 LT |
1865 | * and if that worked, and the page is no longer mapped into |
1866 | * process address space (page_count == 1) it can be freed. | |
1867 | * Otherwise, leave the page on the LRU so it is swappable. | |
1868 | */ | |
266cf658 | 1869 | if (page_has_private(page)) { |
1da177e4 LT |
1870 | if (!try_to_release_page(page, sc->gfp_mask)) |
1871 | goto activate_locked; | |
e286781d NP |
1872 | if (!mapping && page_count(page) == 1) { |
1873 | unlock_page(page); | |
1874 | if (put_page_testzero(page)) | |
1875 | goto free_it; | |
1876 | else { | |
1877 | /* | |
1878 | * rare race with speculative reference. | |
1879 | * the speculative reference will free | |
1880 | * this page shortly, so we may | |
1881 | * increment nr_reclaimed here (and | |
1882 | * leave it off the LRU). | |
1883 | */ | |
1884 | nr_reclaimed++; | |
1885 | continue; | |
1886 | } | |
1887 | } | |
1da177e4 LT |
1888 | } |
1889 | ||
802a3a92 SL |
1890 | if (PageAnon(page) && !PageSwapBacked(page)) { |
1891 | /* follow __remove_mapping for reference */ | |
1892 | if (!page_ref_freeze(page, 1)) | |
1893 | goto keep_locked; | |
d17be2d9 ML |
1894 | /* |
1895 | * The page has only one reference left, which is | |
1896 | * from the isolation. After the caller puts the | |
1897 | * page back on lru and drops the reference, the | |
1898 | * page will be freed anyway. It doesn't matter | |
1899 | * which lru it goes. So we don't bother checking | |
1900 | * PageDirty here. | |
1901 | */ | |
802a3a92 | 1902 | count_vm_event(PGLAZYFREED); |
2262185c | 1903 | count_memcg_page_event(page, PGLAZYFREED); |
be7c07d6 | 1904 | } else if (!mapping || !__remove_mapping(mapping, folio, true, |
b910718a | 1905 | sc->target_mem_cgroup)) |
802a3a92 | 1906 | goto keep_locked; |
9a1ea439 HD |
1907 | |
1908 | unlock_page(page); | |
e286781d | 1909 | free_it: |
98879b3b YS |
1910 | /* |
1911 | * THP may get swapped out in a whole, need account | |
1912 | * all base pages. | |
1913 | */ | |
1914 | nr_reclaimed += nr_pages; | |
abe4c3b5 MG |
1915 | |
1916 | /* | |
1917 | * Is there need to periodically free_page_list? It would | |
1918 | * appear not as the counts should be low | |
1919 | */ | |
7ae88534 | 1920 | if (unlikely(PageTransHuge(page))) |
ff45fc3c | 1921 | destroy_compound_page(page); |
7ae88534 | 1922 | else |
bd4c82c2 | 1923 | list_add(&page->lru, &free_pages); |
1da177e4 LT |
1924 | continue; |
1925 | ||
98879b3b YS |
1926 | activate_locked_split: |
1927 | /* | |
1928 | * The tail pages that are failed to add into swap cache | |
1929 | * reach here. Fixup nr_scanned and nr_pages. | |
1930 | */ | |
1931 | if (nr_pages > 1) { | |
1932 | sc->nr_scanned -= (nr_pages - 1); | |
1933 | nr_pages = 1; | |
1934 | } | |
1da177e4 | 1935 | activate_locked: |
68a22394 | 1936 | /* Not a candidate for swapping, so reclaim swap space. */ |
ad6b6704 MK |
1937 | if (PageSwapCache(page) && (mem_cgroup_swap_full(page) || |
1938 | PageMlocked(page))) | |
a2c43eed | 1939 | try_to_free_swap(page); |
309381fe | 1940 | VM_BUG_ON_PAGE(PageActive(page), page); |
ad6b6704 | 1941 | if (!PageMlocked(page)) { |
9de4f22a | 1942 | int type = page_is_file_lru(page); |
ad6b6704 | 1943 | SetPageActive(page); |
98879b3b | 1944 | stat->nr_activate[type] += nr_pages; |
2262185c | 1945 | count_memcg_page_event(page, PGACTIVATE); |
ad6b6704 | 1946 | } |
1da177e4 LT |
1947 | keep_locked: |
1948 | unlock_page(page); | |
1949 | keep: | |
1950 | list_add(&page->lru, &ret_pages); | |
309381fe | 1951 | VM_BUG_ON_PAGE(PageLRU(page) || PageUnevictable(page), page); |
1da177e4 | 1952 | } |
26aa2d19 DH |
1953 | /* 'page_list' is always empty here */ |
1954 | ||
1955 | /* Migrate pages selected for demotion */ | |
1956 | nr_reclaimed += demote_page_list(&demote_pages, pgdat); | |
1957 | /* Pages that could not be demoted are still in @demote_pages */ | |
1958 | if (!list_empty(&demote_pages)) { | |
1959 | /* Pages which failed to demoted go back on @page_list for retry: */ | |
1960 | list_splice_init(&demote_pages, page_list); | |
1961 | do_demote_pass = false; | |
1962 | goto retry; | |
1963 | } | |
abe4c3b5 | 1964 | |
98879b3b YS |
1965 | pgactivate = stat->nr_activate[0] + stat->nr_activate[1]; |
1966 | ||
747db954 | 1967 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 1968 | try_to_unmap_flush(); |
2d4894b5 | 1969 | free_unref_page_list(&free_pages); |
abe4c3b5 | 1970 | |
1da177e4 | 1971 | list_splice(&ret_pages, page_list); |
886cf190 | 1972 | count_vm_events(PGACTIVATE, pgactivate); |
060f005f | 1973 | |
05ff5137 | 1974 | return nr_reclaimed; |
1da177e4 LT |
1975 | } |
1976 | ||
730ec8c0 | 1977 | unsigned int reclaim_clean_pages_from_list(struct zone *zone, |
02c6de8d MK |
1978 | struct list_head *page_list) |
1979 | { | |
1980 | struct scan_control sc = { | |
1981 | .gfp_mask = GFP_KERNEL, | |
02c6de8d MK |
1982 | .may_unmap = 1, |
1983 | }; | |
1f318a9b | 1984 | struct reclaim_stat stat; |
730ec8c0 | 1985 | unsigned int nr_reclaimed; |
02c6de8d MK |
1986 | struct page *page, *next; |
1987 | LIST_HEAD(clean_pages); | |
2d2b8d2b | 1988 | unsigned int noreclaim_flag; |
02c6de8d MK |
1989 | |
1990 | list_for_each_entry_safe(page, next, page_list, lru) { | |
ae37c7ff OS |
1991 | if (!PageHuge(page) && page_is_file_lru(page) && |
1992 | !PageDirty(page) && !__PageMovable(page) && | |
1993 | !PageUnevictable(page)) { | |
02c6de8d MK |
1994 | ClearPageActive(page); |
1995 | list_move(&page->lru, &clean_pages); | |
1996 | } | |
1997 | } | |
1998 | ||
2d2b8d2b YZ |
1999 | /* |
2000 | * We should be safe here since we are only dealing with file pages and | |
2001 | * we are not kswapd and therefore cannot write dirty file pages. But | |
2002 | * call memalloc_noreclaim_save() anyway, just in case these conditions | |
2003 | * change in the future. | |
2004 | */ | |
2005 | noreclaim_flag = memalloc_noreclaim_save(); | |
1f318a9b | 2006 | nr_reclaimed = shrink_page_list(&clean_pages, zone->zone_pgdat, &sc, |
013339df | 2007 | &stat, true); |
2d2b8d2b YZ |
2008 | memalloc_noreclaim_restore(noreclaim_flag); |
2009 | ||
02c6de8d | 2010 | list_splice(&clean_pages, page_list); |
2da9f630 NP |
2011 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, |
2012 | -(long)nr_reclaimed); | |
1f318a9b JK |
2013 | /* |
2014 | * Since lazyfree pages are isolated from file LRU from the beginning, | |
2015 | * they will rotate back to anonymous LRU in the end if it failed to | |
2016 | * discard so isolated count will be mismatched. | |
2017 | * Compensate the isolated count for both LRU lists. | |
2018 | */ | |
2019 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON, | |
2020 | stat.nr_lazyfree_fail); | |
2021 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, | |
2da9f630 | 2022 | -(long)stat.nr_lazyfree_fail); |
1f318a9b | 2023 | return nr_reclaimed; |
02c6de8d MK |
2024 | } |
2025 | ||
5ad333eb AW |
2026 | /* |
2027 | * Attempt to remove the specified page from its LRU. Only take this page | |
2028 | * if it is of the appropriate PageActive status. Pages which are being | |
2029 | * freed elsewhere are also ignored. | |
2030 | * | |
2031 | * page: page to consider | |
2032 | * mode: one of the LRU isolation modes defined above | |
2033 | * | |
c2135f7c | 2034 | * returns true on success, false on failure. |
5ad333eb | 2035 | */ |
c2135f7c | 2036 | bool __isolate_lru_page_prepare(struct page *page, isolate_mode_t mode) |
5ad333eb | 2037 | { |
5ad333eb AW |
2038 | /* Only take pages on the LRU. */ |
2039 | if (!PageLRU(page)) | |
c2135f7c | 2040 | return false; |
5ad333eb | 2041 | |
e46a2879 MK |
2042 | /* Compaction should not handle unevictable pages but CMA can do so */ |
2043 | if (PageUnevictable(page) && !(mode & ISOLATE_UNEVICTABLE)) | |
c2135f7c | 2044 | return false; |
894bc310 | 2045 | |
c8244935 MG |
2046 | /* |
2047 | * To minimise LRU disruption, the caller can indicate that it only | |
2048 | * wants to isolate pages it will be able to operate on without | |
2049 | * blocking - clean pages for the most part. | |
2050 | * | |
c8244935 MG |
2051 | * ISOLATE_ASYNC_MIGRATE is used to indicate that it only wants to pages |
2052 | * that it is possible to migrate without blocking | |
2053 | */ | |
1276ad68 | 2054 | if (mode & ISOLATE_ASYNC_MIGRATE) { |
c8244935 MG |
2055 | /* All the caller can do on PageWriteback is block */ |
2056 | if (PageWriteback(page)) | |
c2135f7c | 2057 | return false; |
c8244935 MG |
2058 | |
2059 | if (PageDirty(page)) { | |
2060 | struct address_space *mapping; | |
69d763fc | 2061 | bool migrate_dirty; |
c8244935 | 2062 | |
c8244935 MG |
2063 | /* |
2064 | * Only pages without mappings or that have a | |
2065 | * ->migratepage callback are possible to migrate | |
69d763fc MG |
2066 | * without blocking. However, we can be racing with |
2067 | * truncation so it's necessary to lock the page | |
2068 | * to stabilise the mapping as truncation holds | |
2069 | * the page lock until after the page is removed | |
2070 | * from the page cache. | |
c8244935 | 2071 | */ |
69d763fc | 2072 | if (!trylock_page(page)) |
c2135f7c | 2073 | return false; |
69d763fc | 2074 | |
c8244935 | 2075 | mapping = page_mapping(page); |
145e1a71 | 2076 | migrate_dirty = !mapping || mapping->a_ops->migratepage; |
69d763fc MG |
2077 | unlock_page(page); |
2078 | if (!migrate_dirty) | |
c2135f7c | 2079 | return false; |
c8244935 MG |
2080 | } |
2081 | } | |
39deaf85 | 2082 | |
f80c0673 | 2083 | if ((mode & ISOLATE_UNMAPPED) && page_mapped(page)) |
c2135f7c | 2084 | return false; |
f80c0673 | 2085 | |
c2135f7c | 2086 | return true; |
5ad333eb AW |
2087 | } |
2088 | ||
7ee36a14 MG |
2089 | /* |
2090 | * Update LRU sizes after isolating pages. The LRU size updates must | |
55b65a57 | 2091 | * be complete before mem_cgroup_update_lru_size due to a sanity check. |
7ee36a14 MG |
2092 | */ |
2093 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 2094 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 2095 | { |
7ee36a14 MG |
2096 | int zid; |
2097 | ||
7ee36a14 MG |
2098 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
2099 | if (!nr_zone_taken[zid]) | |
2100 | continue; | |
2101 | ||
a892cb6b | 2102 | update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
b4536f0c MH |
2103 | } |
2104 | ||
7ee36a14 MG |
2105 | } |
2106 | ||
f611fab7 | 2107 | /* |
15b44736 HD |
2108 | * Isolating page from the lruvec to fill in @dst list by nr_to_scan times. |
2109 | * | |
2110 | * lruvec->lru_lock is heavily contended. Some of the functions that | |
1da177e4 LT |
2111 | * shrink the lists perform better by taking out a batch of pages |
2112 | * and working on them outside the LRU lock. | |
2113 | * | |
2114 | * For pagecache intensive workloads, this function is the hottest | |
2115 | * spot in the kernel (apart from copy_*_user functions). | |
2116 | * | |
15b44736 | 2117 | * Lru_lock must be held before calling this function. |
1da177e4 | 2118 | * |
791b48b6 | 2119 | * @nr_to_scan: The number of eligible pages to look through on the list. |
5dc35979 | 2120 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 2121 | * @dst: The temp list to put pages on to. |
f626012d | 2122 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 2123 | * @sc: The scan_control struct for this reclaim session |
3cb99451 | 2124 | * @lru: LRU list id for isolating |
1da177e4 LT |
2125 | * |
2126 | * returns how many pages were moved onto *@dst. | |
2127 | */ | |
69e05944 | 2128 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 2129 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 2130 | unsigned long *nr_scanned, struct scan_control *sc, |
a9e7c39f | 2131 | enum lru_list lru) |
1da177e4 | 2132 | { |
75b00af7 | 2133 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 2134 | unsigned long nr_taken = 0; |
599d0c95 | 2135 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 2136 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
3db65812 | 2137 | unsigned long skipped = 0; |
791b48b6 | 2138 | unsigned long scan, total_scan, nr_pages; |
b2e18757 | 2139 | LIST_HEAD(pages_skipped); |
a9e7c39f | 2140 | isolate_mode_t mode = (sc->may_unmap ? 0 : ISOLATE_UNMAPPED); |
1da177e4 | 2141 | |
98879b3b | 2142 | total_scan = 0; |
791b48b6 | 2143 | scan = 0; |
98879b3b | 2144 | while (scan < nr_to_scan && !list_empty(src)) { |
5ad333eb | 2145 | struct page *page; |
5ad333eb | 2146 | |
1da177e4 LT |
2147 | page = lru_to_page(src); |
2148 | prefetchw_prev_lru_page(page, src, flags); | |
2149 | ||
d8c6546b | 2150 | nr_pages = compound_nr(page); |
98879b3b YS |
2151 | total_scan += nr_pages; |
2152 | ||
b2e18757 MG |
2153 | if (page_zonenum(page) > sc->reclaim_idx) { |
2154 | list_move(&page->lru, &pages_skipped); | |
98879b3b | 2155 | nr_skipped[page_zonenum(page)] += nr_pages; |
b2e18757 MG |
2156 | continue; |
2157 | } | |
2158 | ||
791b48b6 MK |
2159 | /* |
2160 | * Do not count skipped pages because that makes the function | |
2161 | * return with no isolated pages if the LRU mostly contains | |
2162 | * ineligible pages. This causes the VM to not reclaim any | |
2163 | * pages, triggering a premature OOM. | |
98879b3b YS |
2164 | * |
2165 | * Account all tail pages of THP. This would not cause | |
2166 | * premature OOM since __isolate_lru_page() returns -EBUSY | |
2167 | * only when the page is being freed somewhere else. | |
791b48b6 | 2168 | */ |
98879b3b | 2169 | scan += nr_pages; |
c2135f7c AS |
2170 | if (!__isolate_lru_page_prepare(page, mode)) { |
2171 | /* It is being freed elsewhere */ | |
2172 | list_move(&page->lru, src); | |
2173 | continue; | |
2174 | } | |
2175 | /* | |
2176 | * Be careful not to clear PageLRU until after we're | |
2177 | * sure the page is not being freed elsewhere -- the | |
2178 | * page release code relies on it. | |
2179 | */ | |
2180 | if (unlikely(!get_page_unless_zero(page))) { | |
2181 | list_move(&page->lru, src); | |
2182 | continue; | |
2183 | } | |
5ad333eb | 2184 | |
c2135f7c AS |
2185 | if (!TestClearPageLRU(page)) { |
2186 | /* Another thread is already isolating this page */ | |
2187 | put_page(page); | |
5ad333eb | 2188 | list_move(&page->lru, src); |
c2135f7c | 2189 | continue; |
5ad333eb | 2190 | } |
c2135f7c AS |
2191 | |
2192 | nr_taken += nr_pages; | |
2193 | nr_zone_taken[page_zonenum(page)] += nr_pages; | |
2194 | list_move(&page->lru, dst); | |
1da177e4 LT |
2195 | } |
2196 | ||
b2e18757 MG |
2197 | /* |
2198 | * Splice any skipped pages to the start of the LRU list. Note that | |
2199 | * this disrupts the LRU order when reclaiming for lower zones but | |
2200 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
2201 | * scanning would soon rescan the same pages to skip and put the | |
2202 | * system at risk of premature OOM. | |
2203 | */ | |
7cc30fcf MG |
2204 | if (!list_empty(&pages_skipped)) { |
2205 | int zid; | |
2206 | ||
3db65812 | 2207 | list_splice(&pages_skipped, src); |
7cc30fcf MG |
2208 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
2209 | if (!nr_skipped[zid]) | |
2210 | continue; | |
2211 | ||
2212 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 2213 | skipped += nr_skipped[zid]; |
7cc30fcf MG |
2214 | } |
2215 | } | |
791b48b6 | 2216 | *nr_scanned = total_scan; |
1265e3a6 | 2217 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
791b48b6 | 2218 | total_scan, skipped, nr_taken, mode, lru); |
b4536f0c | 2219 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
2220 | return nr_taken; |
2221 | } | |
2222 | ||
62695a84 | 2223 | /** |
d1d8a3b4 MWO |
2224 | * folio_isolate_lru() - Try to isolate a folio from its LRU list. |
2225 | * @folio: Folio to isolate from its LRU list. | |
62695a84 | 2226 | * |
d1d8a3b4 MWO |
2227 | * Isolate a @folio from an LRU list and adjust the vmstat statistic |
2228 | * corresponding to whatever LRU list the folio was on. | |
62695a84 | 2229 | * |
d1d8a3b4 MWO |
2230 | * The folio will have its LRU flag cleared. If it was found on the |
2231 | * active list, it will have the Active flag set. If it was found on the | |
2232 | * unevictable list, it will have the Unevictable flag set. These flags | |
894bc310 | 2233 | * may need to be cleared by the caller before letting the page go. |
62695a84 | 2234 | * |
d1d8a3b4 | 2235 | * Context: |
a5d09bed | 2236 | * |
62695a84 | 2237 | * (1) Must be called with an elevated refcount on the page. This is a |
d1d8a3b4 | 2238 | * fundamental difference from isolate_lru_pages() (which is called |
62695a84 | 2239 | * without a stable reference). |
d1d8a3b4 MWO |
2240 | * (2) The lru_lock must not be held. |
2241 | * (3) Interrupts must be enabled. | |
2242 | * | |
2243 | * Return: 0 if the folio was removed from an LRU list. | |
2244 | * -EBUSY if the folio was not on an LRU list. | |
62695a84 | 2245 | */ |
d1d8a3b4 | 2246 | int folio_isolate_lru(struct folio *folio) |
62695a84 NP |
2247 | { |
2248 | int ret = -EBUSY; | |
2249 | ||
d1d8a3b4 | 2250 | VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio); |
0c917313 | 2251 | |
d1d8a3b4 | 2252 | if (folio_test_clear_lru(folio)) { |
fa9add64 | 2253 | struct lruvec *lruvec; |
62695a84 | 2254 | |
d1d8a3b4 | 2255 | folio_get(folio); |
e809c3fe | 2256 | lruvec = folio_lruvec_lock_irq(folio); |
d1d8a3b4 | 2257 | lruvec_del_folio(lruvec, folio); |
6168d0da | 2258 | unlock_page_lruvec_irq(lruvec); |
d25b5bd8 | 2259 | ret = 0; |
62695a84 | 2260 | } |
d25b5bd8 | 2261 | |
62695a84 NP |
2262 | return ret; |
2263 | } | |
2264 | ||
35cd7815 | 2265 | /* |
d37dd5dc | 2266 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
178821b8 | 2267 | * then get rescheduled. When there are massive number of tasks doing page |
d37dd5dc FW |
2268 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, |
2269 | * the LRU list will go small and be scanned faster than necessary, leading to | |
2270 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 2271 | */ |
599d0c95 | 2272 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
2273 | struct scan_control *sc) |
2274 | { | |
2275 | unsigned long inactive, isolated; | |
d818fca1 | 2276 | bool too_many; |
35cd7815 RR |
2277 | |
2278 | if (current_is_kswapd()) | |
2279 | return 0; | |
2280 | ||
b5ead35e | 2281 | if (!writeback_throttling_sane(sc)) |
35cd7815 RR |
2282 | return 0; |
2283 | ||
2284 | if (file) { | |
599d0c95 MG |
2285 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
2286 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 2287 | } else { |
599d0c95 MG |
2288 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
2289 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
2290 | } |
2291 | ||
3cf23841 FW |
2292 | /* |
2293 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
2294 | * won't get blocked by normal direct-reclaimers, forming a circular | |
2295 | * deadlock. | |
2296 | */ | |
d0164adc | 2297 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
2298 | inactive >>= 3; |
2299 | ||
d818fca1 MG |
2300 | too_many = isolated > inactive; |
2301 | ||
2302 | /* Wake up tasks throttled due to too_many_isolated. */ | |
2303 | if (!too_many) | |
2304 | wake_throttle_isolated(pgdat); | |
2305 | ||
2306 | return too_many; | |
35cd7815 RR |
2307 | } |
2308 | ||
a222f341 | 2309 | /* |
15b44736 HD |
2310 | * move_pages_to_lru() moves pages from private @list to appropriate LRU list. |
2311 | * On return, @list is reused as a list of pages to be freed by the caller. | |
a222f341 KT |
2312 | * |
2313 | * Returns the number of pages moved to the given lruvec. | |
2314 | */ | |
9ef56b78 MS |
2315 | static unsigned int move_pages_to_lru(struct lruvec *lruvec, |
2316 | struct list_head *list) | |
66635629 | 2317 | { |
a222f341 | 2318 | int nr_pages, nr_moved = 0; |
3f79768f | 2319 | LIST_HEAD(pages_to_free); |
a222f341 | 2320 | struct page *page; |
66635629 | 2321 | |
a222f341 KT |
2322 | while (!list_empty(list)) { |
2323 | page = lru_to_page(list); | |
309381fe | 2324 | VM_BUG_ON_PAGE(PageLRU(page), page); |
3d06afab | 2325 | list_del(&page->lru); |
39b5f29a | 2326 | if (unlikely(!page_evictable(page))) { |
6168d0da | 2327 | spin_unlock_irq(&lruvec->lru_lock); |
66635629 | 2328 | putback_lru_page(page); |
6168d0da | 2329 | spin_lock_irq(&lruvec->lru_lock); |
66635629 MG |
2330 | continue; |
2331 | } | |
fa9add64 | 2332 | |
3d06afab AS |
2333 | /* |
2334 | * The SetPageLRU needs to be kept here for list integrity. | |
2335 | * Otherwise: | |
2336 | * #0 move_pages_to_lru #1 release_pages | |
2337 | * if !put_page_testzero | |
2338 | * if (put_page_testzero()) | |
2339 | * !PageLRU //skip lru_lock | |
2340 | * SetPageLRU() | |
2341 | * list_add(&page->lru,) | |
2342 | * list_add(&page->lru,) | |
2343 | */ | |
7a608572 | 2344 | SetPageLRU(page); |
a222f341 | 2345 | |
3d06afab | 2346 | if (unlikely(put_page_testzero(page))) { |
87560179 | 2347 | __clear_page_lru_flags(page); |
2bcf8879 HD |
2348 | |
2349 | if (unlikely(PageCompound(page))) { | |
6168d0da | 2350 | spin_unlock_irq(&lruvec->lru_lock); |
ff45fc3c | 2351 | destroy_compound_page(page); |
6168d0da | 2352 | spin_lock_irq(&lruvec->lru_lock); |
2bcf8879 HD |
2353 | } else |
2354 | list_add(&page->lru, &pages_to_free); | |
3d06afab AS |
2355 | |
2356 | continue; | |
66635629 | 2357 | } |
3d06afab | 2358 | |
afca9157 AS |
2359 | /* |
2360 | * All pages were isolated from the same lruvec (and isolation | |
2361 | * inhibits memcg migration). | |
2362 | */ | |
0de340cb | 2363 | VM_BUG_ON_PAGE(!folio_matches_lruvec(page_folio(page), lruvec), page); |
3a9c9788 | 2364 | add_page_to_lru_list(page, lruvec); |
3d06afab | 2365 | nr_pages = thp_nr_pages(page); |
3d06afab AS |
2366 | nr_moved += nr_pages; |
2367 | if (PageActive(page)) | |
2368 | workingset_age_nonresident(lruvec, nr_pages); | |
66635629 | 2369 | } |
66635629 | 2370 | |
3f79768f HD |
2371 | /* |
2372 | * To save our caller's stack, now use input list for pages to free. | |
2373 | */ | |
a222f341 KT |
2374 | list_splice(&pages_to_free, list); |
2375 | ||
2376 | return nr_moved; | |
66635629 MG |
2377 | } |
2378 | ||
399ba0b9 N |
2379 | /* |
2380 | * If a kernel thread (such as nfsd for loop-back mounts) services | |
a37b0715 | 2381 | * a backing device by writing to the page cache it sets PF_LOCAL_THROTTLE. |
399ba0b9 N |
2382 | * In that case we should only throttle if the backing device it is |
2383 | * writing to is congested. In other cases it is safe to throttle. | |
2384 | */ | |
2385 | static int current_may_throttle(void) | |
2386 | { | |
a37b0715 | 2387 | return !(current->flags & PF_LOCAL_THROTTLE) || |
399ba0b9 N |
2388 | current->backing_dev_info == NULL || |
2389 | bdi_write_congested(current->backing_dev_info); | |
2390 | } | |
2391 | ||
1da177e4 | 2392 | /* |
b2e18757 | 2393 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 2394 | * of reclaimed pages |
1da177e4 | 2395 | */ |
9ef56b78 | 2396 | static unsigned long |
1a93be0e | 2397 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 2398 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
2399 | { |
2400 | LIST_HEAD(page_list); | |
e247dbce | 2401 | unsigned long nr_scanned; |
730ec8c0 | 2402 | unsigned int nr_reclaimed = 0; |
e247dbce | 2403 | unsigned long nr_taken; |
060f005f | 2404 | struct reclaim_stat stat; |
497a6c1b | 2405 | bool file = is_file_lru(lru); |
f46b7912 | 2406 | enum vm_event_item item; |
599d0c95 | 2407 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
db73ee0d | 2408 | bool stalled = false; |
78dc583d | 2409 | |
599d0c95 | 2410 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
db73ee0d MH |
2411 | if (stalled) |
2412 | return 0; | |
2413 | ||
2414 | /* wait a bit for the reclaimer. */ | |
db73ee0d | 2415 | stalled = true; |
c3f4a9a2 | 2416 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED); |
35cd7815 RR |
2417 | |
2418 | /* We are about to die and free our memory. Return now. */ | |
2419 | if (fatal_signal_pending(current)) | |
2420 | return SWAP_CLUSTER_MAX; | |
2421 | } | |
2422 | ||
1da177e4 | 2423 | lru_add_drain(); |
f80c0673 | 2424 | |
6168d0da | 2425 | spin_lock_irq(&lruvec->lru_lock); |
b35ea17b | 2426 | |
5dc35979 | 2427 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
a9e7c39f | 2428 | &nr_scanned, sc, lru); |
95d918fc | 2429 | |
599d0c95 | 2430 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
f46b7912 | 2431 | item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT; |
b5ead35e | 2432 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
2433 | __count_vm_events(item, nr_scanned); |
2434 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned); | |
497a6c1b JW |
2435 | __count_vm_events(PGSCAN_ANON + file, nr_scanned); |
2436 | ||
6168d0da | 2437 | spin_unlock_irq(&lruvec->lru_lock); |
b35ea17b | 2438 | |
d563c050 | 2439 | if (nr_taken == 0) |
66635629 | 2440 | return 0; |
5ad333eb | 2441 | |
013339df | 2442 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false); |
c661b078 | 2443 | |
6168d0da | 2444 | spin_lock_irq(&lruvec->lru_lock); |
497a6c1b JW |
2445 | move_pages_to_lru(lruvec, &page_list); |
2446 | ||
2447 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); | |
f46b7912 | 2448 | item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT; |
b5ead35e | 2449 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
2450 | __count_vm_events(item, nr_reclaimed); |
2451 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed); | |
497a6c1b | 2452 | __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed); |
6168d0da | 2453 | spin_unlock_irq(&lruvec->lru_lock); |
3f79768f | 2454 | |
75cc3c91 | 2455 | lru_note_cost(lruvec, file, stat.nr_pageout); |
747db954 | 2456 | mem_cgroup_uncharge_list(&page_list); |
2d4894b5 | 2457 | free_unref_page_list(&page_list); |
e11da5b4 | 2458 | |
1c610d5f AR |
2459 | /* |
2460 | * If dirty pages are scanned that are not queued for IO, it | |
2461 | * implies that flushers are not doing their job. This can | |
2462 | * happen when memory pressure pushes dirty pages to the end of | |
2463 | * the LRU before the dirty limits are breached and the dirty | |
2464 | * data has expired. It can also happen when the proportion of | |
2465 | * dirty pages grows not through writes but through memory | |
2466 | * pressure reclaiming all the clean cache. And in some cases, | |
2467 | * the flushers simply cannot keep up with the allocation | |
2468 | * rate. Nudge the flusher threads in case they are asleep. | |
2469 | */ | |
2470 | if (stat.nr_unqueued_dirty == nr_taken) | |
2471 | wakeup_flusher_threads(WB_REASON_VMSCAN); | |
2472 | ||
d108c772 AR |
2473 | sc->nr.dirty += stat.nr_dirty; |
2474 | sc->nr.congested += stat.nr_congested; | |
2475 | sc->nr.unqueued_dirty += stat.nr_unqueued_dirty; | |
2476 | sc->nr.writeback += stat.nr_writeback; | |
2477 | sc->nr.immediate += stat.nr_immediate; | |
2478 | sc->nr.taken += nr_taken; | |
2479 | if (file) | |
2480 | sc->nr.file_taken += nr_taken; | |
8e950282 | 2481 | |
599d0c95 | 2482 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
d51d1e64 | 2483 | nr_scanned, nr_reclaimed, &stat, sc->priority, file); |
05ff5137 | 2484 | return nr_reclaimed; |
1da177e4 LT |
2485 | } |
2486 | ||
15b44736 HD |
2487 | /* |
2488 | * shrink_active_list() moves pages from the active LRU to the inactive LRU. | |
2489 | * | |
2490 | * We move them the other way if the page is referenced by one or more | |
2491 | * processes. | |
2492 | * | |
2493 | * If the pages are mostly unmapped, the processing is fast and it is | |
2494 | * appropriate to hold lru_lock across the whole operation. But if | |
b3ac0413 | 2495 | * the pages are mapped, the processing is slow (folio_referenced()), so |
15b44736 HD |
2496 | * we should drop lru_lock around each page. It's impossible to balance |
2497 | * this, so instead we remove the pages from the LRU while processing them. | |
2498 | * It is safe to rely on PG_active against the non-LRU pages in here because | |
2499 | * nobody will play with that bit on a non-LRU page. | |
2500 | * | |
2501 | * The downside is that we have to touch page->_refcount against each page. | |
2502 | * But we had to alter page->flags anyway. | |
2503 | */ | |
f626012d | 2504 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 2505 | struct lruvec *lruvec, |
f16015fb | 2506 | struct scan_control *sc, |
9e3b2f8c | 2507 | enum lru_list lru) |
1da177e4 | 2508 | { |
44c241f1 | 2509 | unsigned long nr_taken; |
f626012d | 2510 | unsigned long nr_scanned; |
6fe6b7e3 | 2511 | unsigned long vm_flags; |
1da177e4 | 2512 | LIST_HEAD(l_hold); /* The pages which were snipped off */ |
8cab4754 | 2513 | LIST_HEAD(l_active); |
b69408e8 | 2514 | LIST_HEAD(l_inactive); |
9d998b4f MH |
2515 | unsigned nr_deactivate, nr_activate; |
2516 | unsigned nr_rotated = 0; | |
3cb99451 | 2517 | int file = is_file_lru(lru); |
599d0c95 | 2518 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
2519 | |
2520 | lru_add_drain(); | |
f80c0673 | 2521 | |
6168d0da | 2522 | spin_lock_irq(&lruvec->lru_lock); |
925b7673 | 2523 | |
5dc35979 | 2524 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
a9e7c39f | 2525 | &nr_scanned, sc, lru); |
89b5fae5 | 2526 | |
599d0c95 | 2527 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
1cfb419b | 2528 | |
912c0572 SB |
2529 | if (!cgroup_reclaim(sc)) |
2530 | __count_vm_events(PGREFILL, nr_scanned); | |
2fa2690c | 2531 | __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); |
9d5e6a9f | 2532 | |
6168d0da | 2533 | spin_unlock_irq(&lruvec->lru_lock); |
1da177e4 | 2534 | |
1da177e4 | 2535 | while (!list_empty(&l_hold)) { |
b3ac0413 MWO |
2536 | struct folio *folio; |
2537 | struct page *page; | |
2538 | ||
1da177e4 | 2539 | cond_resched(); |
b3ac0413 MWO |
2540 | folio = lru_to_folio(&l_hold); |
2541 | list_del(&folio->lru); | |
2542 | page = &folio->page; | |
7e9cd484 | 2543 | |
39b5f29a | 2544 | if (unlikely(!page_evictable(page))) { |
894bc310 LS |
2545 | putback_lru_page(page); |
2546 | continue; | |
2547 | } | |
2548 | ||
cc715d99 MG |
2549 | if (unlikely(buffer_heads_over_limit)) { |
2550 | if (page_has_private(page) && trylock_page(page)) { | |
2551 | if (page_has_private(page)) | |
2552 | try_to_release_page(page, 0); | |
2553 | unlock_page(page); | |
2554 | } | |
2555 | } | |
2556 | ||
b3ac0413 MWO |
2557 | if (folio_referenced(folio, 0, sc->target_mem_cgroup, |
2558 | &vm_flags)) { | |
8cab4754 WF |
2559 | /* |
2560 | * Identify referenced, file-backed active pages and | |
2561 | * give them one more trip around the active list. So | |
2562 | * that executable code get better chances to stay in | |
2563 | * memory under moderate memory pressure. Anon pages | |
2564 | * are not likely to be evicted by use-once streaming | |
2565 | * IO, plus JVM can create lots of anon VM_EXEC pages, | |
2566 | * so we ignore them here. | |
2567 | */ | |
9de4f22a | 2568 | if ((vm_flags & VM_EXEC) && page_is_file_lru(page)) { |
6c357848 | 2569 | nr_rotated += thp_nr_pages(page); |
8cab4754 WF |
2570 | list_add(&page->lru, &l_active); |
2571 | continue; | |
2572 | } | |
2573 | } | |
7e9cd484 | 2574 | |
5205e56e | 2575 | ClearPageActive(page); /* we are de-activating */ |
1899ad18 | 2576 | SetPageWorkingset(page); |
1da177e4 LT |
2577 | list_add(&page->lru, &l_inactive); |
2578 | } | |
2579 | ||
b555749a | 2580 | /* |
8cab4754 | 2581 | * Move pages back to the lru list. |
b555749a | 2582 | */ |
6168d0da | 2583 | spin_lock_irq(&lruvec->lru_lock); |
556adecb | 2584 | |
a222f341 KT |
2585 | nr_activate = move_pages_to_lru(lruvec, &l_active); |
2586 | nr_deactivate = move_pages_to_lru(lruvec, &l_inactive); | |
f372d89e KT |
2587 | /* Keep all free pages in l_active list */ |
2588 | list_splice(&l_inactive, &l_active); | |
9851ac13 KT |
2589 | |
2590 | __count_vm_events(PGDEACTIVATE, nr_deactivate); | |
2591 | __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate); | |
2592 | ||
599d0c95 | 2593 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
6168d0da | 2594 | spin_unlock_irq(&lruvec->lru_lock); |
2bcf8879 | 2595 | |
f372d89e KT |
2596 | mem_cgroup_uncharge_list(&l_active); |
2597 | free_unref_page_list(&l_active); | |
9d998b4f MH |
2598 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
2599 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
2600 | } |
2601 | ||
1a4e58cc MK |
2602 | unsigned long reclaim_pages(struct list_head *page_list) |
2603 | { | |
f661d007 | 2604 | int nid = NUMA_NO_NODE; |
730ec8c0 | 2605 | unsigned int nr_reclaimed = 0; |
1a4e58cc MK |
2606 | LIST_HEAD(node_page_list); |
2607 | struct reclaim_stat dummy_stat; | |
2608 | struct page *page; | |
2d2b8d2b | 2609 | unsigned int noreclaim_flag; |
1a4e58cc MK |
2610 | struct scan_control sc = { |
2611 | .gfp_mask = GFP_KERNEL, | |
1a4e58cc MK |
2612 | .may_writepage = 1, |
2613 | .may_unmap = 1, | |
2614 | .may_swap = 1, | |
26aa2d19 | 2615 | .no_demotion = 1, |
1a4e58cc MK |
2616 | }; |
2617 | ||
2d2b8d2b YZ |
2618 | noreclaim_flag = memalloc_noreclaim_save(); |
2619 | ||
1a4e58cc MK |
2620 | while (!list_empty(page_list)) { |
2621 | page = lru_to_page(page_list); | |
f661d007 | 2622 | if (nid == NUMA_NO_NODE) { |
1a4e58cc MK |
2623 | nid = page_to_nid(page); |
2624 | INIT_LIST_HEAD(&node_page_list); | |
2625 | } | |
2626 | ||
2627 | if (nid == page_to_nid(page)) { | |
2628 | ClearPageActive(page); | |
2629 | list_move(&page->lru, &node_page_list); | |
2630 | continue; | |
2631 | } | |
2632 | ||
2633 | nr_reclaimed += shrink_page_list(&node_page_list, | |
2634 | NODE_DATA(nid), | |
013339df | 2635 | &sc, &dummy_stat, false); |
1a4e58cc MK |
2636 | while (!list_empty(&node_page_list)) { |
2637 | page = lru_to_page(&node_page_list); | |
2638 | list_del(&page->lru); | |
2639 | putback_lru_page(page); | |
2640 | } | |
2641 | ||
f661d007 | 2642 | nid = NUMA_NO_NODE; |
1a4e58cc MK |
2643 | } |
2644 | ||
2645 | if (!list_empty(&node_page_list)) { | |
2646 | nr_reclaimed += shrink_page_list(&node_page_list, | |
2647 | NODE_DATA(nid), | |
013339df | 2648 | &sc, &dummy_stat, false); |
1a4e58cc MK |
2649 | while (!list_empty(&node_page_list)) { |
2650 | page = lru_to_page(&node_page_list); | |
2651 | list_del(&page->lru); | |
2652 | putback_lru_page(page); | |
2653 | } | |
2654 | } | |
2655 | ||
2d2b8d2b YZ |
2656 | memalloc_noreclaim_restore(noreclaim_flag); |
2657 | ||
1a4e58cc MK |
2658 | return nr_reclaimed; |
2659 | } | |
2660 | ||
b91ac374 JW |
2661 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
2662 | struct lruvec *lruvec, struct scan_control *sc) | |
2663 | { | |
2664 | if (is_active_lru(lru)) { | |
2665 | if (sc->may_deactivate & (1 << is_file_lru(lru))) | |
2666 | shrink_active_list(nr_to_scan, lruvec, sc, lru); | |
2667 | else | |
2668 | sc->skipped_deactivate = 1; | |
2669 | return 0; | |
2670 | } | |
2671 | ||
2672 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); | |
2673 | } | |
2674 | ||
59dc76b0 RR |
2675 | /* |
2676 | * The inactive anon list should be small enough that the VM never has | |
2677 | * to do too much work. | |
14797e23 | 2678 | * |
59dc76b0 RR |
2679 | * The inactive file list should be small enough to leave most memory |
2680 | * to the established workingset on the scan-resistant active list, | |
2681 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2682 | * |
59dc76b0 RR |
2683 | * Both inactive lists should also be large enough that each inactive |
2684 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 2685 | * |
2a2e4885 JW |
2686 | * If that fails and refaulting is observed, the inactive list grows. |
2687 | * | |
59dc76b0 | 2688 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
3a50d14d | 2689 | * on this LRU, maintained by the pageout code. An inactive_ratio |
59dc76b0 | 2690 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. |
56e49d21 | 2691 | * |
59dc76b0 RR |
2692 | * total target max |
2693 | * memory ratio inactive | |
2694 | * ------------------------------------- | |
2695 | * 10MB 1 5MB | |
2696 | * 100MB 1 50MB | |
2697 | * 1GB 3 250MB | |
2698 | * 10GB 10 0.9GB | |
2699 | * 100GB 31 3GB | |
2700 | * 1TB 101 10GB | |
2701 | * 10TB 320 32GB | |
56e49d21 | 2702 | */ |
b91ac374 | 2703 | static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru) |
56e49d21 | 2704 | { |
b91ac374 | 2705 | enum lru_list active_lru = inactive_lru + LRU_ACTIVE; |
2a2e4885 JW |
2706 | unsigned long inactive, active; |
2707 | unsigned long inactive_ratio; | |
59dc76b0 | 2708 | unsigned long gb; |
e3790144 | 2709 | |
b91ac374 JW |
2710 | inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru); |
2711 | active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru); | |
f8d1a311 | 2712 | |
b91ac374 | 2713 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
4002570c | 2714 | if (gb) |
b91ac374 JW |
2715 | inactive_ratio = int_sqrt(10 * gb); |
2716 | else | |
2717 | inactive_ratio = 1; | |
fd538803 | 2718 | |
59dc76b0 | 2719 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2720 | } |
2721 | ||
9a265114 JW |
2722 | enum scan_balance { |
2723 | SCAN_EQUAL, | |
2724 | SCAN_FRACT, | |
2725 | SCAN_ANON, | |
2726 | SCAN_FILE, | |
2727 | }; | |
2728 | ||
4f98a2fe RR |
2729 | /* |
2730 | * Determine how aggressively the anon and file LRU lists should be | |
2731 | * scanned. The relative value of each set of LRU lists is determined | |
2732 | * by looking at the fraction of the pages scanned we did rotate back | |
2733 | * onto the active list instead of evict. | |
2734 | * | |
be7bd59d WL |
2735 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2736 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 2737 | */ |
afaf07a6 JW |
2738 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, |
2739 | unsigned long *nr) | |
4f98a2fe | 2740 | { |
a2a36488 | 2741 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
afaf07a6 | 2742 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); |
d483a5dd | 2743 | unsigned long anon_cost, file_cost, total_cost; |
33377678 | 2744 | int swappiness = mem_cgroup_swappiness(memcg); |
ed017373 | 2745 | u64 fraction[ANON_AND_FILE]; |
9a265114 | 2746 | u64 denominator = 0; /* gcc */ |
9a265114 | 2747 | enum scan_balance scan_balance; |
4f98a2fe | 2748 | unsigned long ap, fp; |
4111304d | 2749 | enum lru_list lru; |
76a33fc3 SL |
2750 | |
2751 | /* If we have no swap space, do not bother scanning anon pages. */ | |
a2a36488 | 2752 | if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) { |
9a265114 | 2753 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2754 | goto out; |
2755 | } | |
4f98a2fe | 2756 | |
10316b31 JW |
2757 | /* |
2758 | * Global reclaim will swap to prevent OOM even with no | |
2759 | * swappiness, but memcg users want to use this knob to | |
2760 | * disable swapping for individual groups completely when | |
2761 | * using the memory controller's swap limit feature would be | |
2762 | * too expensive. | |
2763 | */ | |
b5ead35e | 2764 | if (cgroup_reclaim(sc) && !swappiness) { |
9a265114 | 2765 | scan_balance = SCAN_FILE; |
10316b31 JW |
2766 | goto out; |
2767 | } | |
2768 | ||
2769 | /* | |
2770 | * Do not apply any pressure balancing cleverness when the | |
2771 | * system is close to OOM, scan both anon and file equally | |
2772 | * (unless the swappiness setting disagrees with swapping). | |
2773 | */ | |
02695175 | 2774 | if (!sc->priority && swappiness) { |
9a265114 | 2775 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2776 | goto out; |
2777 | } | |
2778 | ||
62376251 | 2779 | /* |
53138cea | 2780 | * If the system is almost out of file pages, force-scan anon. |
62376251 | 2781 | */ |
b91ac374 | 2782 | if (sc->file_is_tiny) { |
53138cea JW |
2783 | scan_balance = SCAN_ANON; |
2784 | goto out; | |
62376251 JW |
2785 | } |
2786 | ||
7c5bd705 | 2787 | /* |
b91ac374 JW |
2788 | * If there is enough inactive page cache, we do not reclaim |
2789 | * anything from the anonymous working right now. | |
7c5bd705 | 2790 | */ |
b91ac374 | 2791 | if (sc->cache_trim_mode) { |
9a265114 | 2792 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2793 | goto out; |
2794 | } | |
2795 | ||
9a265114 | 2796 | scan_balance = SCAN_FRACT; |
58c37f6e | 2797 | /* |
314b57fb JW |
2798 | * Calculate the pressure balance between anon and file pages. |
2799 | * | |
2800 | * The amount of pressure we put on each LRU is inversely | |
2801 | * proportional to the cost of reclaiming each list, as | |
2802 | * determined by the share of pages that are refaulting, times | |
2803 | * the relative IO cost of bringing back a swapped out | |
2804 | * anonymous page vs reloading a filesystem page (swappiness). | |
2805 | * | |
d483a5dd JW |
2806 | * Although we limit that influence to ensure no list gets |
2807 | * left behind completely: at least a third of the pressure is | |
2808 | * applied, before swappiness. | |
2809 | * | |
314b57fb | 2810 | * With swappiness at 100, anon and file have equal IO cost. |
58c37f6e | 2811 | */ |
d483a5dd JW |
2812 | total_cost = sc->anon_cost + sc->file_cost; |
2813 | anon_cost = total_cost + sc->anon_cost; | |
2814 | file_cost = total_cost + sc->file_cost; | |
2815 | total_cost = anon_cost + file_cost; | |
58c37f6e | 2816 | |
d483a5dd JW |
2817 | ap = swappiness * (total_cost + 1); |
2818 | ap /= anon_cost + 1; | |
4f98a2fe | 2819 | |
d483a5dd JW |
2820 | fp = (200 - swappiness) * (total_cost + 1); |
2821 | fp /= file_cost + 1; | |
4f98a2fe | 2822 | |
76a33fc3 SL |
2823 | fraction[0] = ap; |
2824 | fraction[1] = fp; | |
a4fe1631 | 2825 | denominator = ap + fp; |
76a33fc3 | 2826 | out: |
688035f7 JW |
2827 | for_each_evictable_lru(lru) { |
2828 | int file = is_file_lru(lru); | |
9783aa99 | 2829 | unsigned long lruvec_size; |
f56ce412 | 2830 | unsigned long low, min; |
688035f7 | 2831 | unsigned long scan; |
9783aa99 CD |
2832 | |
2833 | lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); | |
f56ce412 JW |
2834 | mem_cgroup_protection(sc->target_mem_cgroup, memcg, |
2835 | &min, &low); | |
9783aa99 | 2836 | |
f56ce412 | 2837 | if (min || low) { |
9783aa99 CD |
2838 | /* |
2839 | * Scale a cgroup's reclaim pressure by proportioning | |
2840 | * its current usage to its memory.low or memory.min | |
2841 | * setting. | |
2842 | * | |
2843 | * This is important, as otherwise scanning aggression | |
2844 | * becomes extremely binary -- from nothing as we | |
2845 | * approach the memory protection threshold, to totally | |
2846 | * nominal as we exceed it. This results in requiring | |
2847 | * setting extremely liberal protection thresholds. It | |
2848 | * also means we simply get no protection at all if we | |
2849 | * set it too low, which is not ideal. | |
1bc63fb1 CD |
2850 | * |
2851 | * If there is any protection in place, we reduce scan | |
2852 | * pressure by how much of the total memory used is | |
2853 | * within protection thresholds. | |
9783aa99 | 2854 | * |
9de7ca46 CD |
2855 | * There is one special case: in the first reclaim pass, |
2856 | * we skip over all groups that are within their low | |
2857 | * protection. If that fails to reclaim enough pages to | |
2858 | * satisfy the reclaim goal, we come back and override | |
2859 | * the best-effort low protection. However, we still | |
2860 | * ideally want to honor how well-behaved groups are in | |
2861 | * that case instead of simply punishing them all | |
2862 | * equally. As such, we reclaim them based on how much | |
1bc63fb1 CD |
2863 | * memory they are using, reducing the scan pressure |
2864 | * again by how much of the total memory used is under | |
2865 | * hard protection. | |
9783aa99 | 2866 | */ |
1bc63fb1 | 2867 | unsigned long cgroup_size = mem_cgroup_size(memcg); |
f56ce412 JW |
2868 | unsigned long protection; |
2869 | ||
2870 | /* memory.low scaling, make sure we retry before OOM */ | |
2871 | if (!sc->memcg_low_reclaim && low > min) { | |
2872 | protection = low; | |
2873 | sc->memcg_low_skipped = 1; | |
2874 | } else { | |
2875 | protection = min; | |
2876 | } | |
1bc63fb1 CD |
2877 | |
2878 | /* Avoid TOCTOU with earlier protection check */ | |
2879 | cgroup_size = max(cgroup_size, protection); | |
2880 | ||
2881 | scan = lruvec_size - lruvec_size * protection / | |
32d4f4b7 | 2882 | (cgroup_size + 1); |
9783aa99 CD |
2883 | |
2884 | /* | |
1bc63fb1 | 2885 | * Minimally target SWAP_CLUSTER_MAX pages to keep |
55b65a57 | 2886 | * reclaim moving forwards, avoiding decrementing |
9de7ca46 | 2887 | * sc->priority further than desirable. |
9783aa99 | 2888 | */ |
1bc63fb1 | 2889 | scan = max(scan, SWAP_CLUSTER_MAX); |
9783aa99 CD |
2890 | } else { |
2891 | scan = lruvec_size; | |
2892 | } | |
2893 | ||
2894 | scan >>= sc->priority; | |
6b4f7799 | 2895 | |
688035f7 JW |
2896 | /* |
2897 | * If the cgroup's already been deleted, make sure to | |
2898 | * scrape out the remaining cache. | |
2899 | */ | |
2900 | if (!scan && !mem_cgroup_online(memcg)) | |
9783aa99 | 2901 | scan = min(lruvec_size, SWAP_CLUSTER_MAX); |
6b4f7799 | 2902 | |
688035f7 JW |
2903 | switch (scan_balance) { |
2904 | case SCAN_EQUAL: | |
2905 | /* Scan lists relative to size */ | |
2906 | break; | |
2907 | case SCAN_FRACT: | |
9a265114 | 2908 | /* |
688035f7 JW |
2909 | * Scan types proportional to swappiness and |
2910 | * their relative recent reclaim efficiency. | |
76073c64 GS |
2911 | * Make sure we don't miss the last page on |
2912 | * the offlined memory cgroups because of a | |
2913 | * round-off error. | |
9a265114 | 2914 | */ |
76073c64 GS |
2915 | scan = mem_cgroup_online(memcg) ? |
2916 | div64_u64(scan * fraction[file], denominator) : | |
2917 | DIV64_U64_ROUND_UP(scan * fraction[file], | |
68600f62 | 2918 | denominator); |
688035f7 JW |
2919 | break; |
2920 | case SCAN_FILE: | |
2921 | case SCAN_ANON: | |
2922 | /* Scan one type exclusively */ | |
e072bff6 | 2923 | if ((scan_balance == SCAN_FILE) != file) |
688035f7 | 2924 | scan = 0; |
688035f7 JW |
2925 | break; |
2926 | default: | |
2927 | /* Look ma, no brain */ | |
2928 | BUG(); | |
9a265114 | 2929 | } |
688035f7 | 2930 | |
688035f7 | 2931 | nr[lru] = scan; |
76a33fc3 | 2932 | } |
6e08a369 | 2933 | } |
4f98a2fe | 2934 | |
2f368a9f DH |
2935 | /* |
2936 | * Anonymous LRU management is a waste if there is | |
2937 | * ultimately no way to reclaim the memory. | |
2938 | */ | |
2939 | static bool can_age_anon_pages(struct pglist_data *pgdat, | |
2940 | struct scan_control *sc) | |
2941 | { | |
2942 | /* Aging the anon LRU is valuable if swap is present: */ | |
2943 | if (total_swap_pages > 0) | |
2944 | return true; | |
2945 | ||
2946 | /* Also valuable if anon pages can be demoted: */ | |
2947 | return can_demote(pgdat->node_id, sc); | |
2948 | } | |
2949 | ||
afaf07a6 | 2950 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) |
9b4f98cd JW |
2951 | { |
2952 | unsigned long nr[NR_LRU_LISTS]; | |
e82e0561 | 2953 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
2954 | unsigned long nr_to_scan; |
2955 | enum lru_list lru; | |
2956 | unsigned long nr_reclaimed = 0; | |
2957 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
2958 | struct blk_plug plug; | |
1a501907 | 2959 | bool scan_adjusted; |
9b4f98cd | 2960 | |
afaf07a6 | 2961 | get_scan_count(lruvec, sc, nr); |
9b4f98cd | 2962 | |
e82e0561 MG |
2963 | /* Record the original scan target for proportional adjustments later */ |
2964 | memcpy(targets, nr, sizeof(nr)); | |
2965 | ||
1a501907 MG |
2966 | /* |
2967 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
2968 | * event that can occur when there is little memory pressure e.g. | |
2969 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
2970 | * when the requested number of pages are reclaimed when scanning at | |
2971 | * DEF_PRIORITY on the assumption that the fact we are direct | |
2972 | * reclaiming implies that kswapd is not keeping up and it is best to | |
2973 | * do a batch of work at once. For memcg reclaim one check is made to | |
2974 | * abort proportional reclaim if either the file or anon lru has already | |
2975 | * dropped to zero at the first pass. | |
2976 | */ | |
b5ead35e | 2977 | scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() && |
1a501907 MG |
2978 | sc->priority == DEF_PRIORITY); |
2979 | ||
9b4f98cd JW |
2980 | blk_start_plug(&plug); |
2981 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
2982 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
2983 | unsigned long nr_anon, nr_file, percentage; |
2984 | unsigned long nr_scanned; | |
2985 | ||
9b4f98cd JW |
2986 | for_each_evictable_lru(lru) { |
2987 | if (nr[lru]) { | |
2988 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
2989 | nr[lru] -= nr_to_scan; | |
2990 | ||
2991 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
3b991208 | 2992 | lruvec, sc); |
9b4f98cd JW |
2993 | } |
2994 | } | |
e82e0561 | 2995 | |
bd041733 MH |
2996 | cond_resched(); |
2997 | ||
e82e0561 MG |
2998 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
2999 | continue; | |
3000 | ||
e82e0561 MG |
3001 | /* |
3002 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 3003 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
3004 | * proportionally what was requested by get_scan_count(). We |
3005 | * stop reclaiming one LRU and reduce the amount scanning | |
3006 | * proportional to the original scan target. | |
3007 | */ | |
3008 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
3009 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
3010 | ||
1a501907 MG |
3011 | /* |
3012 | * It's just vindictive to attack the larger once the smaller | |
3013 | * has gone to zero. And given the way we stop scanning the | |
3014 | * smaller below, this makes sure that we only make one nudge | |
3015 | * towards proportionality once we've got nr_to_reclaim. | |
3016 | */ | |
3017 | if (!nr_file || !nr_anon) | |
3018 | break; | |
3019 | ||
e82e0561 MG |
3020 | if (nr_file > nr_anon) { |
3021 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
3022 | targets[LRU_ACTIVE_ANON] + 1; | |
3023 | lru = LRU_BASE; | |
3024 | percentage = nr_anon * 100 / scan_target; | |
3025 | } else { | |
3026 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
3027 | targets[LRU_ACTIVE_FILE] + 1; | |
3028 | lru = LRU_FILE; | |
3029 | percentage = nr_file * 100 / scan_target; | |
3030 | } | |
3031 | ||
3032 | /* Stop scanning the smaller of the LRU */ | |
3033 | nr[lru] = 0; | |
3034 | nr[lru + LRU_ACTIVE] = 0; | |
3035 | ||
3036 | /* | |
3037 | * Recalculate the other LRU scan count based on its original | |
3038 | * scan target and the percentage scanning already complete | |
3039 | */ | |
3040 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
3041 | nr_scanned = targets[lru] - nr[lru]; | |
3042 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
3043 | nr[lru] -= min(nr[lru], nr_scanned); | |
3044 | ||
3045 | lru += LRU_ACTIVE; | |
3046 | nr_scanned = targets[lru] - nr[lru]; | |
3047 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
3048 | nr[lru] -= min(nr[lru], nr_scanned); | |
3049 | ||
3050 | scan_adjusted = true; | |
9b4f98cd JW |
3051 | } |
3052 | blk_finish_plug(&plug); | |
3053 | sc->nr_reclaimed += nr_reclaimed; | |
3054 | ||
3055 | /* | |
3056 | * Even if we did not try to evict anon pages at all, we want to | |
3057 | * rebalance the anon lru active/inactive ratio. | |
3058 | */ | |
2f368a9f DH |
3059 | if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) && |
3060 | inactive_is_low(lruvec, LRU_INACTIVE_ANON)) | |
9b4f98cd JW |
3061 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
3062 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
3063 | } |
3064 | ||
23b9da55 | 3065 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 3066 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 3067 | { |
d84da3f9 | 3068 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 3069 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 3070 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
3071 | return true; |
3072 | ||
3073 | return false; | |
3074 | } | |
3075 | ||
3e7d3449 | 3076 | /* |
23b9da55 MG |
3077 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
3078 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
3079 | * true if more pages should be reclaimed such that when the page allocator | |
df3a45f9 | 3080 | * calls try_to_compact_pages() that it will have enough free pages to succeed. |
23b9da55 | 3081 | * It will give up earlier than that if there is difficulty reclaiming pages. |
3e7d3449 | 3082 | */ |
a9dd0a83 | 3083 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 | 3084 | unsigned long nr_reclaimed, |
3e7d3449 MG |
3085 | struct scan_control *sc) |
3086 | { | |
3087 | unsigned long pages_for_compaction; | |
3088 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 3089 | int z; |
3e7d3449 MG |
3090 | |
3091 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 3092 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
3093 | return false; |
3094 | ||
5ee04716 VB |
3095 | /* |
3096 | * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX | |
3097 | * number of pages that were scanned. This will return to the caller | |
3098 | * with the risk reclaim/compaction and the resulting allocation attempt | |
3099 | * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL | |
3100 | * allocations through requiring that the full LRU list has been scanned | |
3101 | * first, by assuming that zero delta of sc->nr_scanned means full LRU | |
3102 | * scan, but that approximation was wrong, and there were corner cases | |
3103 | * where always a non-zero amount of pages were scanned. | |
3104 | */ | |
3105 | if (!nr_reclaimed) | |
3106 | return false; | |
3e7d3449 | 3107 | |
3e7d3449 | 3108 | /* If compaction would go ahead or the allocation would succeed, stop */ |
a9dd0a83 MG |
3109 | for (z = 0; z <= sc->reclaim_idx; z++) { |
3110 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 3111 | if (!managed_zone(zone)) |
a9dd0a83 MG |
3112 | continue; |
3113 | ||
3114 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 3115 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
3116 | case COMPACT_CONTINUE: |
3117 | return false; | |
3118 | default: | |
3119 | /* check next zone */ | |
3120 | ; | |
3121 | } | |
3e7d3449 | 3122 | } |
1c6c1597 HD |
3123 | |
3124 | /* | |
3125 | * If we have not reclaimed enough pages for compaction and the | |
3126 | * inactive lists are large enough, continue reclaiming | |
3127 | */ | |
3128 | pages_for_compaction = compact_gap(sc->order); | |
3129 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); | |
a2a36488 | 3130 | if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc)) |
1c6c1597 HD |
3131 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
3132 | ||
5ee04716 | 3133 | return inactive_lru_pages > pages_for_compaction; |
3e7d3449 MG |
3134 | } |
3135 | ||
0f6a5cff | 3136 | static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 3137 | { |
0f6a5cff | 3138 | struct mem_cgroup *target_memcg = sc->target_mem_cgroup; |
d2af3397 | 3139 | struct mem_cgroup *memcg; |
1da177e4 | 3140 | |
0f6a5cff | 3141 | memcg = mem_cgroup_iter(target_memcg, NULL, NULL); |
d2af3397 | 3142 | do { |
afaf07a6 | 3143 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
d2af3397 JW |
3144 | unsigned long reclaimed; |
3145 | unsigned long scanned; | |
5660048c | 3146 | |
e3336cab XP |
3147 | /* |
3148 | * This loop can become CPU-bound when target memcgs | |
3149 | * aren't eligible for reclaim - either because they | |
3150 | * don't have any reclaimable pages, or because their | |
3151 | * memory is explicitly protected. Avoid soft lockups. | |
3152 | */ | |
3153 | cond_resched(); | |
3154 | ||
45c7f7e1 CD |
3155 | mem_cgroup_calculate_protection(target_memcg, memcg); |
3156 | ||
3157 | if (mem_cgroup_below_min(memcg)) { | |
d2af3397 JW |
3158 | /* |
3159 | * Hard protection. | |
3160 | * If there is no reclaimable memory, OOM. | |
3161 | */ | |
3162 | continue; | |
45c7f7e1 | 3163 | } else if (mem_cgroup_below_low(memcg)) { |
d2af3397 JW |
3164 | /* |
3165 | * Soft protection. | |
3166 | * Respect the protection only as long as | |
3167 | * there is an unprotected supply | |
3168 | * of reclaimable memory from other cgroups. | |
3169 | */ | |
3170 | if (!sc->memcg_low_reclaim) { | |
3171 | sc->memcg_low_skipped = 1; | |
bf8d5d52 | 3172 | continue; |
241994ed | 3173 | } |
d2af3397 | 3174 | memcg_memory_event(memcg, MEMCG_LOW); |
d2af3397 | 3175 | } |
241994ed | 3176 | |
d2af3397 JW |
3177 | reclaimed = sc->nr_reclaimed; |
3178 | scanned = sc->nr_scanned; | |
afaf07a6 JW |
3179 | |
3180 | shrink_lruvec(lruvec, sc); | |
70ddf637 | 3181 | |
d2af3397 JW |
3182 | shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, |
3183 | sc->priority); | |
6b4f7799 | 3184 | |
d2af3397 JW |
3185 | /* Record the group's reclaim efficiency */ |
3186 | vmpressure(sc->gfp_mask, memcg, false, | |
3187 | sc->nr_scanned - scanned, | |
3188 | sc->nr_reclaimed - reclaimed); | |
70ddf637 | 3189 | |
0f6a5cff JW |
3190 | } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL))); |
3191 | } | |
3192 | ||
6c9e0907 | 3193 | static void shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
0f6a5cff JW |
3194 | { |
3195 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
0f6a5cff | 3196 | unsigned long nr_reclaimed, nr_scanned; |
1b05117d | 3197 | struct lruvec *target_lruvec; |
0f6a5cff | 3198 | bool reclaimable = false; |
b91ac374 | 3199 | unsigned long file; |
0f6a5cff | 3200 | |
1b05117d JW |
3201 | target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat); |
3202 | ||
0f6a5cff | 3203 | again: |
aa48e47e SB |
3204 | /* |
3205 | * Flush the memory cgroup stats, so that we read accurate per-memcg | |
3206 | * lruvec stats for heuristics. | |
3207 | */ | |
3208 | mem_cgroup_flush_stats(); | |
3209 | ||
0f6a5cff JW |
3210 | memset(&sc->nr, 0, sizeof(sc->nr)); |
3211 | ||
3212 | nr_reclaimed = sc->nr_reclaimed; | |
3213 | nr_scanned = sc->nr_scanned; | |
3214 | ||
7cf111bc JW |
3215 | /* |
3216 | * Determine the scan balance between anon and file LRUs. | |
3217 | */ | |
6168d0da | 3218 | spin_lock_irq(&target_lruvec->lru_lock); |
7cf111bc JW |
3219 | sc->anon_cost = target_lruvec->anon_cost; |
3220 | sc->file_cost = target_lruvec->file_cost; | |
6168d0da | 3221 | spin_unlock_irq(&target_lruvec->lru_lock); |
7cf111bc | 3222 | |
b91ac374 JW |
3223 | /* |
3224 | * Target desirable inactive:active list ratios for the anon | |
3225 | * and file LRU lists. | |
3226 | */ | |
3227 | if (!sc->force_deactivate) { | |
3228 | unsigned long refaults; | |
3229 | ||
170b04b7 JK |
3230 | refaults = lruvec_page_state(target_lruvec, |
3231 | WORKINGSET_ACTIVATE_ANON); | |
3232 | if (refaults != target_lruvec->refaults[0] || | |
3233 | inactive_is_low(target_lruvec, LRU_INACTIVE_ANON)) | |
b91ac374 JW |
3234 | sc->may_deactivate |= DEACTIVATE_ANON; |
3235 | else | |
3236 | sc->may_deactivate &= ~DEACTIVATE_ANON; | |
3237 | ||
3238 | /* | |
3239 | * When refaults are being observed, it means a new | |
3240 | * workingset is being established. Deactivate to get | |
3241 | * rid of any stale active pages quickly. | |
3242 | */ | |
3243 | refaults = lruvec_page_state(target_lruvec, | |
170b04b7 JK |
3244 | WORKINGSET_ACTIVATE_FILE); |
3245 | if (refaults != target_lruvec->refaults[1] || | |
b91ac374 JW |
3246 | inactive_is_low(target_lruvec, LRU_INACTIVE_FILE)) |
3247 | sc->may_deactivate |= DEACTIVATE_FILE; | |
3248 | else | |
3249 | sc->may_deactivate &= ~DEACTIVATE_FILE; | |
3250 | } else | |
3251 | sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE; | |
3252 | ||
3253 | /* | |
3254 | * If we have plenty of inactive file pages that aren't | |
3255 | * thrashing, try to reclaim those first before touching | |
3256 | * anonymous pages. | |
3257 | */ | |
3258 | file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE); | |
3259 | if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE)) | |
3260 | sc->cache_trim_mode = 1; | |
3261 | else | |
3262 | sc->cache_trim_mode = 0; | |
3263 | ||
53138cea JW |
3264 | /* |
3265 | * Prevent the reclaimer from falling into the cache trap: as | |
3266 | * cache pages start out inactive, every cache fault will tip | |
3267 | * the scan balance towards the file LRU. And as the file LRU | |
3268 | * shrinks, so does the window for rotation from references. | |
3269 | * This means we have a runaway feedback loop where a tiny | |
3270 | * thrashing file LRU becomes infinitely more attractive than | |
3271 | * anon pages. Try to detect this based on file LRU size. | |
3272 | */ | |
3273 | if (!cgroup_reclaim(sc)) { | |
53138cea | 3274 | unsigned long total_high_wmark = 0; |
b91ac374 JW |
3275 | unsigned long free, anon; |
3276 | int z; | |
53138cea JW |
3277 | |
3278 | free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); | |
3279 | file = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
3280 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
3281 | ||
3282 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
3283 | struct zone *zone = &pgdat->node_zones[z]; | |
3284 | if (!managed_zone(zone)) | |
3285 | continue; | |
3286 | ||
3287 | total_high_wmark += high_wmark_pages(zone); | |
3288 | } | |
3289 | ||
b91ac374 JW |
3290 | /* |
3291 | * Consider anon: if that's low too, this isn't a | |
3292 | * runaway file reclaim problem, but rather just | |
3293 | * extreme pressure. Reclaim as per usual then. | |
3294 | */ | |
3295 | anon = node_page_state(pgdat, NR_INACTIVE_ANON); | |
3296 | ||
3297 | sc->file_is_tiny = | |
3298 | file + free <= total_high_wmark && | |
3299 | !(sc->may_deactivate & DEACTIVATE_ANON) && | |
3300 | anon >> sc->priority; | |
53138cea JW |
3301 | } |
3302 | ||
0f6a5cff | 3303 | shrink_node_memcgs(pgdat, sc); |
2344d7e4 | 3304 | |
d2af3397 JW |
3305 | if (reclaim_state) { |
3306 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
3307 | reclaim_state->reclaimed_slab = 0; | |
3308 | } | |
d108c772 | 3309 | |
d2af3397 | 3310 | /* Record the subtree's reclaim efficiency */ |
1b05117d | 3311 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, |
d2af3397 JW |
3312 | sc->nr_scanned - nr_scanned, |
3313 | sc->nr_reclaimed - nr_reclaimed); | |
d108c772 | 3314 | |
d2af3397 JW |
3315 | if (sc->nr_reclaimed - nr_reclaimed) |
3316 | reclaimable = true; | |
d108c772 | 3317 | |
d2af3397 JW |
3318 | if (current_is_kswapd()) { |
3319 | /* | |
3320 | * If reclaim is isolating dirty pages under writeback, | |
3321 | * it implies that the long-lived page allocation rate | |
3322 | * is exceeding the page laundering rate. Either the | |
3323 | * global limits are not being effective at throttling | |
3324 | * processes due to the page distribution throughout | |
3325 | * zones or there is heavy usage of a slow backing | |
3326 | * device. The only option is to throttle from reclaim | |
3327 | * context which is not ideal as there is no guarantee | |
3328 | * the dirtying process is throttled in the same way | |
3329 | * balance_dirty_pages() manages. | |
3330 | * | |
3331 | * Once a node is flagged PGDAT_WRITEBACK, kswapd will | |
3332 | * count the number of pages under pages flagged for | |
3333 | * immediate reclaim and stall if any are encountered | |
3334 | * in the nr_immediate check below. | |
3335 | */ | |
3336 | if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken) | |
3337 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
d108c772 | 3338 | |
d2af3397 JW |
3339 | /* Allow kswapd to start writing pages during reclaim.*/ |
3340 | if (sc->nr.unqueued_dirty == sc->nr.file_taken) | |
3341 | set_bit(PGDAT_DIRTY, &pgdat->flags); | |
e3c1ac58 | 3342 | |
d108c772 | 3343 | /* |
1eba09c1 | 3344 | * If kswapd scans pages marked for immediate |
d2af3397 JW |
3345 | * reclaim and under writeback (nr_immediate), it |
3346 | * implies that pages are cycling through the LRU | |
8cd7c588 MG |
3347 | * faster than they are written so forcibly stall |
3348 | * until some pages complete writeback. | |
d108c772 | 3349 | */ |
d2af3397 | 3350 | if (sc->nr.immediate) |
c3f4a9a2 | 3351 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK); |
d2af3397 JW |
3352 | } |
3353 | ||
3354 | /* | |
8cd7c588 MG |
3355 | * Tag a node/memcg as congested if all the dirty pages were marked |
3356 | * for writeback and immediate reclaim (counted in nr.congested). | |
1b05117d | 3357 | * |
d2af3397 | 3358 | * Legacy memcg will stall in page writeback so avoid forcibly |
8cd7c588 | 3359 | * stalling in reclaim_throttle(). |
d2af3397 | 3360 | */ |
1b05117d JW |
3361 | if ((current_is_kswapd() || |
3362 | (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) && | |
d2af3397 | 3363 | sc->nr.dirty && sc->nr.dirty == sc->nr.congested) |
1b05117d | 3364 | set_bit(LRUVEC_CONGESTED, &target_lruvec->flags); |
d2af3397 JW |
3365 | |
3366 | /* | |
8cd7c588 MG |
3367 | * Stall direct reclaim for IO completions if the lruvec is |
3368 | * node is congested. Allow kswapd to continue until it | |
d2af3397 JW |
3369 | * starts encountering unqueued dirty pages or cycling through |
3370 | * the LRU too quickly. | |
3371 | */ | |
1b05117d JW |
3372 | if (!current_is_kswapd() && current_may_throttle() && |
3373 | !sc->hibernation_mode && | |
3374 | test_bit(LRUVEC_CONGESTED, &target_lruvec->flags)) | |
1b4e3f26 | 3375 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED); |
d108c772 | 3376 | |
d2af3397 JW |
3377 | if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
3378 | sc)) | |
3379 | goto again; | |
2344d7e4 | 3380 | |
c73322d0 JW |
3381 | /* |
3382 | * Kswapd gives up on balancing particular nodes after too | |
3383 | * many failures to reclaim anything from them and goes to | |
3384 | * sleep. On reclaim progress, reset the failure counter. A | |
3385 | * successful direct reclaim run will revive a dormant kswapd. | |
3386 | */ | |
3387 | if (reclaimable) | |
3388 | pgdat->kswapd_failures = 0; | |
f16015fb JW |
3389 | } |
3390 | ||
53853e2d | 3391 | /* |
fdd4c614 VB |
3392 | * Returns true if compaction should go ahead for a costly-order request, or |
3393 | * the allocation would already succeed without compaction. Return false if we | |
3394 | * should reclaim first. | |
53853e2d | 3395 | */ |
4f588331 | 3396 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 3397 | { |
31483b6a | 3398 | unsigned long watermark; |
fdd4c614 | 3399 | enum compact_result suitable; |
fe4b1b24 | 3400 | |
fdd4c614 VB |
3401 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
3402 | if (suitable == COMPACT_SUCCESS) | |
3403 | /* Allocation should succeed already. Don't reclaim. */ | |
3404 | return true; | |
3405 | if (suitable == COMPACT_SKIPPED) | |
3406 | /* Compaction cannot yet proceed. Do reclaim. */ | |
3407 | return false; | |
fe4b1b24 | 3408 | |
53853e2d | 3409 | /* |
fdd4c614 VB |
3410 | * Compaction is already possible, but it takes time to run and there |
3411 | * are potentially other callers using the pages just freed. So proceed | |
3412 | * with reclaim to make a buffer of free pages available to give | |
3413 | * compaction a reasonable chance of completing and allocating the page. | |
3414 | * Note that we won't actually reclaim the whole buffer in one attempt | |
3415 | * as the target watermark in should_continue_reclaim() is lower. But if | |
3416 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 3417 | */ |
fdd4c614 | 3418 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 3419 | |
fdd4c614 | 3420 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
3421 | } |
3422 | ||
69392a40 MG |
3423 | static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc) |
3424 | { | |
66ce520b MG |
3425 | /* |
3426 | * If reclaim is making progress greater than 12% efficiency then | |
3427 | * wake all the NOPROGRESS throttled tasks. | |
3428 | */ | |
3429 | if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) { | |
69392a40 MG |
3430 | wait_queue_head_t *wqh; |
3431 | ||
3432 | wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS]; | |
3433 | if (waitqueue_active(wqh)) | |
3434 | wake_up(wqh); | |
3435 | ||
3436 | return; | |
3437 | } | |
3438 | ||
3439 | /* | |
1b4e3f26 MG |
3440 | * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will |
3441 | * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages | |
3442 | * under writeback and marked for immediate reclaim at the tail of the | |
3443 | * LRU. | |
69392a40 | 3444 | */ |
1b4e3f26 | 3445 | if (current_is_kswapd() || cgroup_reclaim(sc)) |
69392a40 MG |
3446 | return; |
3447 | ||
3448 | /* Throttle if making no progress at high prioities. */ | |
1b4e3f26 | 3449 | if (sc->priority == 1 && !sc->nr_reclaimed) |
c3f4a9a2 | 3450 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS); |
69392a40 MG |
3451 | } |
3452 | ||
1da177e4 LT |
3453 | /* |
3454 | * This is the direct reclaim path, for page-allocating processes. We only | |
3455 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
3456 | * request. | |
3457 | * | |
1da177e4 LT |
3458 | * If a zone is deemed to be full of pinned pages then just give it a light |
3459 | * scan then give up on it. | |
3460 | */ | |
0a0337e0 | 3461 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 3462 | { |
dd1a239f | 3463 | struct zoneref *z; |
54a6eb5c | 3464 | struct zone *zone; |
0608f43d AM |
3465 | unsigned long nr_soft_reclaimed; |
3466 | unsigned long nr_soft_scanned; | |
619d0d76 | 3467 | gfp_t orig_mask; |
79dafcdc | 3468 | pg_data_t *last_pgdat = NULL; |
1b4e3f26 | 3469 | pg_data_t *first_pgdat = NULL; |
1cfb419b | 3470 | |
cc715d99 MG |
3471 | /* |
3472 | * If the number of buffer_heads in the machine exceeds the maximum | |
3473 | * allowed level, force direct reclaim to scan the highmem zone as | |
3474 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
3475 | */ | |
619d0d76 | 3476 | orig_mask = sc->gfp_mask; |
b2e18757 | 3477 | if (buffer_heads_over_limit) { |
cc715d99 | 3478 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 3479 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 3480 | } |
cc715d99 | 3481 | |
d4debc66 | 3482 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 3483 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
3484 | /* |
3485 | * Take care memory controller reclaiming has small influence | |
3486 | * to global LRU. | |
3487 | */ | |
b5ead35e | 3488 | if (!cgroup_reclaim(sc)) { |
344736f2 VD |
3489 | if (!cpuset_zone_allowed(zone, |
3490 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 3491 | continue; |
65ec02cb | 3492 | |
0b06496a JW |
3493 | /* |
3494 | * If we already have plenty of memory free for | |
3495 | * compaction in this zone, don't free any more. | |
3496 | * Even though compaction is invoked for any | |
3497 | * non-zero order, only frequent costly order | |
3498 | * reclamation is disruptive enough to become a | |
3499 | * noticeable problem, like transparent huge | |
3500 | * page allocations. | |
3501 | */ | |
3502 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
3503 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 3504 | compaction_ready(zone, sc)) { |
0b06496a JW |
3505 | sc->compaction_ready = true; |
3506 | continue; | |
e0887c19 | 3507 | } |
0b06496a | 3508 | |
79dafcdc MG |
3509 | /* |
3510 | * Shrink each node in the zonelist once. If the | |
3511 | * zonelist is ordered by zone (not the default) then a | |
3512 | * node may be shrunk multiple times but in that case | |
3513 | * the user prefers lower zones being preserved. | |
3514 | */ | |
3515 | if (zone->zone_pgdat == last_pgdat) | |
3516 | continue; | |
3517 | ||
0608f43d AM |
3518 | /* |
3519 | * This steals pages from memory cgroups over softlimit | |
3520 | * and returns the number of reclaimed pages and | |
3521 | * scanned pages. This works for global memory pressure | |
3522 | * and balancing, not for a memcg's limit. | |
3523 | */ | |
3524 | nr_soft_scanned = 0; | |
ef8f2327 | 3525 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
3526 | sc->order, sc->gfp_mask, |
3527 | &nr_soft_scanned); | |
3528 | sc->nr_reclaimed += nr_soft_reclaimed; | |
3529 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 3530 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 3531 | } |
408d8544 | 3532 | |
1b4e3f26 MG |
3533 | if (!first_pgdat) |
3534 | first_pgdat = zone->zone_pgdat; | |
3535 | ||
79dafcdc MG |
3536 | /* See comment about same check for global reclaim above */ |
3537 | if (zone->zone_pgdat == last_pgdat) | |
3538 | continue; | |
3539 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 3540 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 3541 | } |
e0c23279 | 3542 | |
80082938 MG |
3543 | if (first_pgdat) |
3544 | consider_reclaim_throttle(first_pgdat, sc); | |
1b4e3f26 | 3545 | |
619d0d76 WY |
3546 | /* |
3547 | * Restore to original mask to avoid the impact on the caller if we | |
3548 | * promoted it to __GFP_HIGHMEM. | |
3549 | */ | |
3550 | sc->gfp_mask = orig_mask; | |
1da177e4 | 3551 | } |
4f98a2fe | 3552 | |
b910718a | 3553 | static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat) |
2a2e4885 | 3554 | { |
b910718a JW |
3555 | struct lruvec *target_lruvec; |
3556 | unsigned long refaults; | |
2a2e4885 | 3557 | |
b910718a | 3558 | target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat); |
170b04b7 JK |
3559 | refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON); |
3560 | target_lruvec->refaults[0] = refaults; | |
3561 | refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE); | |
3562 | target_lruvec->refaults[1] = refaults; | |
2a2e4885 JW |
3563 | } |
3564 | ||
1da177e4 LT |
3565 | /* |
3566 | * This is the main entry point to direct page reclaim. | |
3567 | * | |
3568 | * If a full scan of the inactive list fails to free enough memory then we | |
3569 | * are "out of memory" and something needs to be killed. | |
3570 | * | |
3571 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
3572 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
3573 | * caller can't do much about. We kick the writeback threads and take explicit |
3574 | * naps in the hope that some of these pages can be written. But if the | |
3575 | * allocating task holds filesystem locks which prevent writeout this might not | |
3576 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
3577 | * |
3578 | * returns: 0, if no pages reclaimed | |
3579 | * else, the number of pages reclaimed | |
1da177e4 | 3580 | */ |
dac1d27b | 3581 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 3582 | struct scan_control *sc) |
1da177e4 | 3583 | { |
241994ed | 3584 | int initial_priority = sc->priority; |
2a2e4885 JW |
3585 | pg_data_t *last_pgdat; |
3586 | struct zoneref *z; | |
3587 | struct zone *zone; | |
241994ed | 3588 | retry: |
873b4771 KK |
3589 | delayacct_freepages_start(); |
3590 | ||
b5ead35e | 3591 | if (!cgroup_reclaim(sc)) |
7cc30fcf | 3592 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 3593 | |
9e3b2f8c | 3594 | do { |
70ddf637 AV |
3595 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, |
3596 | sc->priority); | |
66e1707b | 3597 | sc->nr_scanned = 0; |
0a0337e0 | 3598 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 3599 | |
bb21c7ce | 3600 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
3601 | break; |
3602 | ||
3603 | if (sc->compaction_ready) | |
3604 | break; | |
1da177e4 | 3605 | |
0e50ce3b MK |
3606 | /* |
3607 | * If we're getting trouble reclaiming, start doing | |
3608 | * writepage even in laptop mode. | |
3609 | */ | |
3610 | if (sc->priority < DEF_PRIORITY - 2) | |
3611 | sc->may_writepage = 1; | |
0b06496a | 3612 | } while (--sc->priority >= 0); |
bb21c7ce | 3613 | |
2a2e4885 JW |
3614 | last_pgdat = NULL; |
3615 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
3616 | sc->nodemask) { | |
3617 | if (zone->zone_pgdat == last_pgdat) | |
3618 | continue; | |
3619 | last_pgdat = zone->zone_pgdat; | |
1b05117d | 3620 | |
2a2e4885 | 3621 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); |
1b05117d JW |
3622 | |
3623 | if (cgroup_reclaim(sc)) { | |
3624 | struct lruvec *lruvec; | |
3625 | ||
3626 | lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, | |
3627 | zone->zone_pgdat); | |
3628 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
3629 | } | |
2a2e4885 JW |
3630 | } |
3631 | ||
873b4771 KK |
3632 | delayacct_freepages_end(); |
3633 | ||
bb21c7ce KM |
3634 | if (sc->nr_reclaimed) |
3635 | return sc->nr_reclaimed; | |
3636 | ||
0cee34fd | 3637 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 3638 | if (sc->compaction_ready) |
7335084d MG |
3639 | return 1; |
3640 | ||
b91ac374 JW |
3641 | /* |
3642 | * We make inactive:active ratio decisions based on the node's | |
3643 | * composition of memory, but a restrictive reclaim_idx or a | |
3644 | * memory.low cgroup setting can exempt large amounts of | |
3645 | * memory from reclaim. Neither of which are very common, so | |
3646 | * instead of doing costly eligibility calculations of the | |
3647 | * entire cgroup subtree up front, we assume the estimates are | |
3648 | * good, and retry with forcible deactivation if that fails. | |
3649 | */ | |
3650 | if (sc->skipped_deactivate) { | |
3651 | sc->priority = initial_priority; | |
3652 | sc->force_deactivate = 1; | |
3653 | sc->skipped_deactivate = 0; | |
3654 | goto retry; | |
3655 | } | |
3656 | ||
241994ed | 3657 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
d6622f63 | 3658 | if (sc->memcg_low_skipped) { |
241994ed | 3659 | sc->priority = initial_priority; |
b91ac374 | 3660 | sc->force_deactivate = 0; |
d6622f63 YX |
3661 | sc->memcg_low_reclaim = 1; |
3662 | sc->memcg_low_skipped = 0; | |
241994ed JW |
3663 | goto retry; |
3664 | } | |
3665 | ||
bb21c7ce | 3666 | return 0; |
1da177e4 LT |
3667 | } |
3668 | ||
c73322d0 | 3669 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
5515061d MG |
3670 | { |
3671 | struct zone *zone; | |
3672 | unsigned long pfmemalloc_reserve = 0; | |
3673 | unsigned long free_pages = 0; | |
3674 | int i; | |
3675 | bool wmark_ok; | |
3676 | ||
c73322d0 JW |
3677 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
3678 | return true; | |
3679 | ||
5515061d MG |
3680 | for (i = 0; i <= ZONE_NORMAL; i++) { |
3681 | zone = &pgdat->node_zones[i]; | |
d450abd8 JW |
3682 | if (!managed_zone(zone)) |
3683 | continue; | |
3684 | ||
3685 | if (!zone_reclaimable_pages(zone)) | |
675becce MG |
3686 | continue; |
3687 | ||
5515061d MG |
3688 | pfmemalloc_reserve += min_wmark_pages(zone); |
3689 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
3690 | } | |
3691 | ||
675becce MG |
3692 | /* If there are no reserves (unexpected config) then do not throttle */ |
3693 | if (!pfmemalloc_reserve) | |
3694 | return true; | |
3695 | ||
5515061d MG |
3696 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
3697 | ||
3698 | /* kswapd must be awake if processes are being throttled */ | |
3699 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
97a225e6 JK |
3700 | if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL) |
3701 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL); | |
5644e1fb | 3702 | |
5515061d MG |
3703 | wake_up_interruptible(&pgdat->kswapd_wait); |
3704 | } | |
3705 | ||
3706 | return wmark_ok; | |
3707 | } | |
3708 | ||
3709 | /* | |
3710 | * Throttle direct reclaimers if backing storage is backed by the network | |
3711 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
3712 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
3713 | * when the low watermark is reached. |
3714 | * | |
3715 | * Returns true if a fatal signal was delivered during throttling. If this | |
3716 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 3717 | */ |
50694c28 | 3718 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
3719 | nodemask_t *nodemask) |
3720 | { | |
675becce | 3721 | struct zoneref *z; |
5515061d | 3722 | struct zone *zone; |
675becce | 3723 | pg_data_t *pgdat = NULL; |
5515061d MG |
3724 | |
3725 | /* | |
3726 | * Kernel threads should not be throttled as they may be indirectly | |
3727 | * responsible for cleaning pages necessary for reclaim to make forward | |
3728 | * progress. kjournald for example may enter direct reclaim while | |
3729 | * committing a transaction where throttling it could forcing other | |
3730 | * processes to block on log_wait_commit(). | |
3731 | */ | |
3732 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
3733 | goto out; |
3734 | ||
3735 | /* | |
3736 | * If a fatal signal is pending, this process should not throttle. | |
3737 | * It should return quickly so it can exit and free its memory | |
3738 | */ | |
3739 | if (fatal_signal_pending(current)) | |
3740 | goto out; | |
5515061d | 3741 | |
675becce MG |
3742 | /* |
3743 | * Check if the pfmemalloc reserves are ok by finding the first node | |
3744 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
3745 | * GFP_KERNEL will be required for allocating network buffers when | |
3746 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
3747 | * | |
3748 | * Throttling is based on the first usable node and throttled processes | |
3749 | * wait on a queue until kswapd makes progress and wakes them. There | |
3750 | * is an affinity then between processes waking up and where reclaim | |
3751 | * progress has been made assuming the process wakes on the same node. | |
3752 | * More importantly, processes running on remote nodes will not compete | |
3753 | * for remote pfmemalloc reserves and processes on different nodes | |
3754 | * should make reasonable progress. | |
3755 | */ | |
3756 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 3757 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
3758 | if (zone_idx(zone) > ZONE_NORMAL) |
3759 | continue; | |
3760 | ||
3761 | /* Throttle based on the first usable node */ | |
3762 | pgdat = zone->zone_pgdat; | |
c73322d0 | 3763 | if (allow_direct_reclaim(pgdat)) |
675becce MG |
3764 | goto out; |
3765 | break; | |
3766 | } | |
3767 | ||
3768 | /* If no zone was usable by the allocation flags then do not throttle */ | |
3769 | if (!pgdat) | |
50694c28 | 3770 | goto out; |
5515061d | 3771 | |
68243e76 MG |
3772 | /* Account for the throttling */ |
3773 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
3774 | ||
5515061d MG |
3775 | /* |
3776 | * If the caller cannot enter the filesystem, it's possible that it | |
3777 | * is due to the caller holding an FS lock or performing a journal | |
3778 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
3779 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
3780 | * blocked waiting on the same lock. Instead, throttle for up to a | |
3781 | * second before continuing. | |
3782 | */ | |
2e786d9e | 3783 | if (!(gfp_mask & __GFP_FS)) |
5515061d | 3784 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, |
c73322d0 | 3785 | allow_direct_reclaim(pgdat), HZ); |
2e786d9e ML |
3786 | else |
3787 | /* Throttle until kswapd wakes the process */ | |
3788 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
3789 | allow_direct_reclaim(pgdat)); | |
50694c28 | 3790 | |
50694c28 MG |
3791 | if (fatal_signal_pending(current)) |
3792 | return true; | |
3793 | ||
3794 | out: | |
3795 | return false; | |
5515061d MG |
3796 | } |
3797 | ||
dac1d27b | 3798 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 3799 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 3800 | { |
33906bc5 | 3801 | unsigned long nr_reclaimed; |
66e1707b | 3802 | struct scan_control sc = { |
ee814fe2 | 3803 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
f2f43e56 | 3804 | .gfp_mask = current_gfp_context(gfp_mask), |
b2e18757 | 3805 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
3806 | .order = order, |
3807 | .nodemask = nodemask, | |
3808 | .priority = DEF_PRIORITY, | |
66e1707b | 3809 | .may_writepage = !laptop_mode, |
a6dc60f8 | 3810 | .may_unmap = 1, |
2e2e4259 | 3811 | .may_swap = 1, |
66e1707b BS |
3812 | }; |
3813 | ||
bb451fdf GT |
3814 | /* |
3815 | * scan_control uses s8 fields for order, priority, and reclaim_idx. | |
3816 | * Confirm they are large enough for max values. | |
3817 | */ | |
3818 | BUILD_BUG_ON(MAX_ORDER > S8_MAX); | |
3819 | BUILD_BUG_ON(DEF_PRIORITY > S8_MAX); | |
3820 | BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX); | |
3821 | ||
5515061d | 3822 | /* |
50694c28 MG |
3823 | * Do not enter reclaim if fatal signal was delivered while throttled. |
3824 | * 1 is returned so that the page allocator does not OOM kill at this | |
3825 | * point. | |
5515061d | 3826 | */ |
f2f43e56 | 3827 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
5515061d MG |
3828 | return 1; |
3829 | ||
1732d2b0 | 3830 | set_task_reclaim_state(current, &sc.reclaim_state); |
3481c37f | 3831 | trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask); |
33906bc5 | 3832 | |
3115cd91 | 3833 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
3834 | |
3835 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
1732d2b0 | 3836 | set_task_reclaim_state(current, NULL); |
33906bc5 MG |
3837 | |
3838 | return nr_reclaimed; | |
66e1707b BS |
3839 | } |
3840 | ||
c255a458 | 3841 | #ifdef CONFIG_MEMCG |
66e1707b | 3842 | |
d2e5fb92 | 3843 | /* Only used by soft limit reclaim. Do not reuse for anything else. */ |
a9dd0a83 | 3844 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 3845 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 3846 | pg_data_t *pgdat, |
0ae5e89c | 3847 | unsigned long *nr_scanned) |
4e416953 | 3848 | { |
afaf07a6 | 3849 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
4e416953 | 3850 | struct scan_control sc = { |
b8f5c566 | 3851 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 3852 | .target_mem_cgroup = memcg, |
4e416953 BS |
3853 | .may_writepage = !laptop_mode, |
3854 | .may_unmap = 1, | |
b2e18757 | 3855 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 3856 | .may_swap = !noswap, |
4e416953 | 3857 | }; |
0ae5e89c | 3858 | |
d2e5fb92 MH |
3859 | WARN_ON_ONCE(!current->reclaim_state); |
3860 | ||
4e416953 BS |
3861 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
3862 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 3863 | |
9e3b2f8c | 3864 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
3481c37f | 3865 | sc.gfp_mask); |
bdce6d9e | 3866 | |
4e416953 BS |
3867 | /* |
3868 | * NOTE: Although we can get the priority field, using it | |
3869 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 3870 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
3871 | * will pick up pages from other mem cgroup's as well. We hack |
3872 | * the priority and make it zero. | |
3873 | */ | |
afaf07a6 | 3874 | shrink_lruvec(lruvec, &sc); |
bdce6d9e KM |
3875 | |
3876 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
3877 | ||
0ae5e89c | 3878 | *nr_scanned = sc.nr_scanned; |
0308f7cf | 3879 | |
4e416953 BS |
3880 | return sc.nr_reclaimed; |
3881 | } | |
3882 | ||
72835c86 | 3883 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 3884 | unsigned long nr_pages, |
a7885eb8 | 3885 | gfp_t gfp_mask, |
b70a2a21 | 3886 | bool may_swap) |
66e1707b | 3887 | { |
bdce6d9e | 3888 | unsigned long nr_reclaimed; |
499118e9 | 3889 | unsigned int noreclaim_flag; |
66e1707b | 3890 | struct scan_control sc = { |
b70a2a21 | 3891 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 3892 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
a09ed5e0 | 3893 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
b2e18757 | 3894 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
3895 | .target_mem_cgroup = memcg, |
3896 | .priority = DEF_PRIORITY, | |
3897 | .may_writepage = !laptop_mode, | |
3898 | .may_unmap = 1, | |
b70a2a21 | 3899 | .may_swap = may_swap, |
a09ed5e0 | 3900 | }; |
889976db | 3901 | /* |
fa40d1ee SB |
3902 | * Traverse the ZONELIST_FALLBACK zonelist of the current node to put |
3903 | * equal pressure on all the nodes. This is based on the assumption that | |
3904 | * the reclaim does not bail out early. | |
889976db | 3905 | */ |
fa40d1ee | 3906 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
889976db | 3907 | |
fa40d1ee | 3908 | set_task_reclaim_state(current, &sc.reclaim_state); |
3481c37f | 3909 | trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask); |
499118e9 | 3910 | noreclaim_flag = memalloc_noreclaim_save(); |
eb414681 | 3911 | |
3115cd91 | 3912 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
eb414681 | 3913 | |
499118e9 | 3914 | memalloc_noreclaim_restore(noreclaim_flag); |
bdce6d9e | 3915 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); |
1732d2b0 | 3916 | set_task_reclaim_state(current, NULL); |
bdce6d9e KM |
3917 | |
3918 | return nr_reclaimed; | |
66e1707b BS |
3919 | } |
3920 | #endif | |
3921 | ||
1d82de61 | 3922 | static void age_active_anon(struct pglist_data *pgdat, |
ef8f2327 | 3923 | struct scan_control *sc) |
f16015fb | 3924 | { |
b95a2f2d | 3925 | struct mem_cgroup *memcg; |
b91ac374 | 3926 | struct lruvec *lruvec; |
f16015fb | 3927 | |
2f368a9f | 3928 | if (!can_age_anon_pages(pgdat, sc)) |
b95a2f2d JW |
3929 | return; |
3930 | ||
b91ac374 JW |
3931 | lruvec = mem_cgroup_lruvec(NULL, pgdat); |
3932 | if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON)) | |
3933 | return; | |
3934 | ||
b95a2f2d JW |
3935 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
3936 | do { | |
b91ac374 JW |
3937 | lruvec = mem_cgroup_lruvec(memcg, pgdat); |
3938 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
3939 | sc, LRU_ACTIVE_ANON); | |
b95a2f2d JW |
3940 | memcg = mem_cgroup_iter(NULL, memcg, NULL); |
3941 | } while (memcg); | |
f16015fb JW |
3942 | } |
3943 | ||
97a225e6 | 3944 | static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx) |
1c30844d MG |
3945 | { |
3946 | int i; | |
3947 | struct zone *zone; | |
3948 | ||
3949 | /* | |
3950 | * Check for watermark boosts top-down as the higher zones | |
3951 | * are more likely to be boosted. Both watermarks and boosts | |
1eba09c1 | 3952 | * should not be checked at the same time as reclaim would |
1c30844d MG |
3953 | * start prematurely when there is no boosting and a lower |
3954 | * zone is balanced. | |
3955 | */ | |
97a225e6 | 3956 | for (i = highest_zoneidx; i >= 0; i--) { |
1c30844d MG |
3957 | zone = pgdat->node_zones + i; |
3958 | if (!managed_zone(zone)) | |
3959 | continue; | |
3960 | ||
3961 | if (zone->watermark_boost) | |
3962 | return true; | |
3963 | } | |
3964 | ||
3965 | return false; | |
3966 | } | |
3967 | ||
e716f2eb MG |
3968 | /* |
3969 | * Returns true if there is an eligible zone balanced for the request order | |
97a225e6 | 3970 | * and highest_zoneidx |
e716f2eb | 3971 | */ |
97a225e6 | 3972 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx) |
60cefed4 | 3973 | { |
e716f2eb MG |
3974 | int i; |
3975 | unsigned long mark = -1; | |
3976 | struct zone *zone; | |
60cefed4 | 3977 | |
1c30844d MG |
3978 | /* |
3979 | * Check watermarks bottom-up as lower zones are more likely to | |
3980 | * meet watermarks. | |
3981 | */ | |
97a225e6 | 3982 | for (i = 0; i <= highest_zoneidx; i++) { |
e716f2eb | 3983 | zone = pgdat->node_zones + i; |
6256c6b4 | 3984 | |
e716f2eb MG |
3985 | if (!managed_zone(zone)) |
3986 | continue; | |
3987 | ||
3988 | mark = high_wmark_pages(zone); | |
97a225e6 | 3989 | if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx)) |
e716f2eb MG |
3990 | return true; |
3991 | } | |
3992 | ||
3993 | /* | |
97a225e6 | 3994 | * If a node has no populated zone within highest_zoneidx, it does not |
e716f2eb MG |
3995 | * need balancing by definition. This can happen if a zone-restricted |
3996 | * allocation tries to wake a remote kswapd. | |
3997 | */ | |
3998 | if (mark == -1) | |
3999 | return true; | |
4000 | ||
4001 | return false; | |
60cefed4 JW |
4002 | } |
4003 | ||
631b6e08 MG |
4004 | /* Clear pgdat state for congested, dirty or under writeback. */ |
4005 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
4006 | { | |
1b05117d JW |
4007 | struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat); |
4008 | ||
4009 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
631b6e08 MG |
4010 | clear_bit(PGDAT_DIRTY, &pgdat->flags); |
4011 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
4012 | } | |
4013 | ||
5515061d MG |
4014 | /* |
4015 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
4016 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
4017 | * | |
4018 | * Returns true if kswapd is ready to sleep | |
4019 | */ | |
97a225e6 JK |
4020 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, |
4021 | int highest_zoneidx) | |
f50de2d3 | 4022 | { |
5515061d | 4023 | /* |
9e5e3661 | 4024 | * The throttled processes are normally woken up in balance_pgdat() as |
c73322d0 | 4025 | * soon as allow_direct_reclaim() is true. But there is a potential |
9e5e3661 VB |
4026 | * race between when kswapd checks the watermarks and a process gets |
4027 | * throttled. There is also a potential race if processes get | |
4028 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
4029 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
4030 | * the wake up checks. If kswapd is going to sleep, no process should | |
4031 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
4032 | * the wake up is premature, processes will wake kswapd and get | |
4033 | * throttled again. The difference from wake ups in balance_pgdat() is | |
4034 | * that here we are under prepare_to_wait(). | |
5515061d | 4035 | */ |
9e5e3661 VB |
4036 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
4037 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 4038 | |
c73322d0 JW |
4039 | /* Hopeless node, leave it to direct reclaim */ |
4040 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
4041 | return true; | |
4042 | ||
97a225e6 | 4043 | if (pgdat_balanced(pgdat, order, highest_zoneidx)) { |
e716f2eb MG |
4044 | clear_pgdat_congested(pgdat); |
4045 | return true; | |
1d82de61 MG |
4046 | } |
4047 | ||
333b0a45 | 4048 | return false; |
f50de2d3 MG |
4049 | } |
4050 | ||
75485363 | 4051 | /* |
1d82de61 MG |
4052 | * kswapd shrinks a node of pages that are at or below the highest usable |
4053 | * zone that is currently unbalanced. | |
b8e83b94 MG |
4054 | * |
4055 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
4056 | * reclaim or if the lack of progress was due to pages under writeback. |
4057 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 4058 | */ |
1d82de61 | 4059 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 4060 | struct scan_control *sc) |
75485363 | 4061 | { |
1d82de61 MG |
4062 | struct zone *zone; |
4063 | int z; | |
75485363 | 4064 | |
1d82de61 MG |
4065 | /* Reclaim a number of pages proportional to the number of zones */ |
4066 | sc->nr_to_reclaim = 0; | |
970a39a3 | 4067 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 4068 | zone = pgdat->node_zones + z; |
6aa303de | 4069 | if (!managed_zone(zone)) |
1d82de61 | 4070 | continue; |
7c954f6d | 4071 | |
1d82de61 MG |
4072 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
4073 | } | |
7c954f6d MG |
4074 | |
4075 | /* | |
1d82de61 MG |
4076 | * Historically care was taken to put equal pressure on all zones but |
4077 | * now pressure is applied based on node LRU order. | |
7c954f6d | 4078 | */ |
970a39a3 | 4079 | shrink_node(pgdat, sc); |
283aba9f | 4080 | |
7c954f6d | 4081 | /* |
1d82de61 MG |
4082 | * Fragmentation may mean that the system cannot be rebalanced for |
4083 | * high-order allocations. If twice the allocation size has been | |
4084 | * reclaimed then recheck watermarks only at order-0 to prevent | |
4085 | * excessive reclaim. Assume that a process requested a high-order | |
4086 | * can direct reclaim/compact. | |
7c954f6d | 4087 | */ |
9861a62c | 4088 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 4089 | sc->order = 0; |
7c954f6d | 4090 | |
b8e83b94 | 4091 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
4092 | } |
4093 | ||
c49c2c47 MG |
4094 | /* Page allocator PCP high watermark is lowered if reclaim is active. */ |
4095 | static inline void | |
4096 | update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active) | |
4097 | { | |
4098 | int i; | |
4099 | struct zone *zone; | |
4100 | ||
4101 | for (i = 0; i <= highest_zoneidx; i++) { | |
4102 | zone = pgdat->node_zones + i; | |
4103 | ||
4104 | if (!managed_zone(zone)) | |
4105 | continue; | |
4106 | ||
4107 | if (active) | |
4108 | set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags); | |
4109 | else | |
4110 | clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags); | |
4111 | } | |
4112 | } | |
4113 | ||
4114 | static inline void | |
4115 | set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx) | |
4116 | { | |
4117 | update_reclaim_active(pgdat, highest_zoneidx, true); | |
4118 | } | |
4119 | ||
4120 | static inline void | |
4121 | clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx) | |
4122 | { | |
4123 | update_reclaim_active(pgdat, highest_zoneidx, false); | |
4124 | } | |
4125 | ||
1da177e4 | 4126 | /* |
1d82de61 MG |
4127 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
4128 | * that are eligible for use by the caller until at least one zone is | |
4129 | * balanced. | |
1da177e4 | 4130 | * |
1d82de61 | 4131 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
4132 | * |
4133 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 4134 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
8bb4e7a2 | 4135 | * found to have free_pages <= high_wmark_pages(zone), any page in that zone |
1d82de61 MG |
4136 | * or lower is eligible for reclaim until at least one usable zone is |
4137 | * balanced. | |
1da177e4 | 4138 | */ |
97a225e6 | 4139 | static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx) |
1da177e4 | 4140 | { |
1da177e4 | 4141 | int i; |
0608f43d AM |
4142 | unsigned long nr_soft_reclaimed; |
4143 | unsigned long nr_soft_scanned; | |
eb414681 | 4144 | unsigned long pflags; |
1c30844d MG |
4145 | unsigned long nr_boost_reclaim; |
4146 | unsigned long zone_boosts[MAX_NR_ZONES] = { 0, }; | |
4147 | bool boosted; | |
1d82de61 | 4148 | struct zone *zone; |
179e9639 AM |
4149 | struct scan_control sc = { |
4150 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 4151 | .order = order, |
a6dc60f8 | 4152 | .may_unmap = 1, |
179e9639 | 4153 | }; |
93781325 | 4154 | |
1732d2b0 | 4155 | set_task_reclaim_state(current, &sc.reclaim_state); |
eb414681 | 4156 | psi_memstall_enter(&pflags); |
4f3eaf45 | 4157 | __fs_reclaim_acquire(_THIS_IP_); |
93781325 | 4158 | |
f8891e5e | 4159 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 4160 | |
1c30844d MG |
4161 | /* |
4162 | * Account for the reclaim boost. Note that the zone boost is left in | |
4163 | * place so that parallel allocations that are near the watermark will | |
4164 | * stall or direct reclaim until kswapd is finished. | |
4165 | */ | |
4166 | nr_boost_reclaim = 0; | |
97a225e6 | 4167 | for (i = 0; i <= highest_zoneidx; i++) { |
1c30844d MG |
4168 | zone = pgdat->node_zones + i; |
4169 | if (!managed_zone(zone)) | |
4170 | continue; | |
4171 | ||
4172 | nr_boost_reclaim += zone->watermark_boost; | |
4173 | zone_boosts[i] = zone->watermark_boost; | |
4174 | } | |
4175 | boosted = nr_boost_reclaim; | |
4176 | ||
4177 | restart: | |
c49c2c47 | 4178 | set_reclaim_active(pgdat, highest_zoneidx); |
1c30844d | 4179 | sc.priority = DEF_PRIORITY; |
9e3b2f8c | 4180 | do { |
c73322d0 | 4181 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
b8e83b94 | 4182 | bool raise_priority = true; |
1c30844d | 4183 | bool balanced; |
93781325 | 4184 | bool ret; |
b8e83b94 | 4185 | |
97a225e6 | 4186 | sc.reclaim_idx = highest_zoneidx; |
1da177e4 | 4187 | |
86c79f6b | 4188 | /* |
84c7a777 MG |
4189 | * If the number of buffer_heads exceeds the maximum allowed |
4190 | * then consider reclaiming from all zones. This has a dual | |
4191 | * purpose -- on 64-bit systems it is expected that | |
4192 | * buffer_heads are stripped during active rotation. On 32-bit | |
4193 | * systems, highmem pages can pin lowmem memory and shrinking | |
4194 | * buffers can relieve lowmem pressure. Reclaim may still not | |
4195 | * go ahead if all eligible zones for the original allocation | |
4196 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
4197 | */ |
4198 | if (buffer_heads_over_limit) { | |
4199 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
4200 | zone = pgdat->node_zones + i; | |
6aa303de | 4201 | if (!managed_zone(zone)) |
86c79f6b | 4202 | continue; |
cc715d99 | 4203 | |
970a39a3 | 4204 | sc.reclaim_idx = i; |
e1dbeda6 | 4205 | break; |
1da177e4 | 4206 | } |
1da177e4 | 4207 | } |
dafcb73e | 4208 | |
86c79f6b | 4209 | /* |
1c30844d MG |
4210 | * If the pgdat is imbalanced then ignore boosting and preserve |
4211 | * the watermarks for a later time and restart. Note that the | |
4212 | * zone watermarks will be still reset at the end of balancing | |
4213 | * on the grounds that the normal reclaim should be enough to | |
4214 | * re-evaluate if boosting is required when kswapd next wakes. | |
4215 | */ | |
97a225e6 | 4216 | balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx); |
1c30844d MG |
4217 | if (!balanced && nr_boost_reclaim) { |
4218 | nr_boost_reclaim = 0; | |
4219 | goto restart; | |
4220 | } | |
4221 | ||
4222 | /* | |
4223 | * If boosting is not active then only reclaim if there are no | |
4224 | * eligible zones. Note that sc.reclaim_idx is not used as | |
4225 | * buffer_heads_over_limit may have adjusted it. | |
86c79f6b | 4226 | */ |
1c30844d | 4227 | if (!nr_boost_reclaim && balanced) |
e716f2eb | 4228 | goto out; |
e1dbeda6 | 4229 | |
1c30844d MG |
4230 | /* Limit the priority of boosting to avoid reclaim writeback */ |
4231 | if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2) | |
4232 | raise_priority = false; | |
4233 | ||
4234 | /* | |
4235 | * Do not writeback or swap pages for boosted reclaim. The | |
4236 | * intent is to relieve pressure not issue sub-optimal IO | |
4237 | * from reclaim context. If no pages are reclaimed, the | |
4238 | * reclaim will be aborted. | |
4239 | */ | |
4240 | sc.may_writepage = !laptop_mode && !nr_boost_reclaim; | |
4241 | sc.may_swap = !nr_boost_reclaim; | |
1c30844d | 4242 | |
1d82de61 MG |
4243 | /* |
4244 | * Do some background aging of the anon list, to give | |
4245 | * pages a chance to be referenced before reclaiming. All | |
4246 | * pages are rotated regardless of classzone as this is | |
4247 | * about consistent aging. | |
4248 | */ | |
ef8f2327 | 4249 | age_active_anon(pgdat, &sc); |
1d82de61 | 4250 | |
b7ea3c41 MG |
4251 | /* |
4252 | * If we're getting trouble reclaiming, start doing writepage | |
4253 | * even in laptop mode. | |
4254 | */ | |
047d72c3 | 4255 | if (sc.priority < DEF_PRIORITY - 2) |
b7ea3c41 MG |
4256 | sc.may_writepage = 1; |
4257 | ||
1d82de61 MG |
4258 | /* Call soft limit reclaim before calling shrink_node. */ |
4259 | sc.nr_scanned = 0; | |
4260 | nr_soft_scanned = 0; | |
ef8f2327 | 4261 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
4262 | sc.gfp_mask, &nr_soft_scanned); |
4263 | sc.nr_reclaimed += nr_soft_reclaimed; | |
4264 | ||
1da177e4 | 4265 | /* |
1d82de61 MG |
4266 | * There should be no need to raise the scanning priority if |
4267 | * enough pages are already being scanned that that high | |
4268 | * watermark would be met at 100% efficiency. | |
1da177e4 | 4269 | */ |
970a39a3 | 4270 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 4271 | raise_priority = false; |
5515061d MG |
4272 | |
4273 | /* | |
4274 | * If the low watermark is met there is no need for processes | |
4275 | * to be throttled on pfmemalloc_wait as they should not be | |
4276 | * able to safely make forward progress. Wake them | |
4277 | */ | |
4278 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
c73322d0 | 4279 | allow_direct_reclaim(pgdat)) |
cfc51155 | 4280 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 4281 | |
b8e83b94 | 4282 | /* Check if kswapd should be suspending */ |
4f3eaf45 | 4283 | __fs_reclaim_release(_THIS_IP_); |
93781325 | 4284 | ret = try_to_freeze(); |
4f3eaf45 | 4285 | __fs_reclaim_acquire(_THIS_IP_); |
93781325 | 4286 | if (ret || kthread_should_stop()) |
b8e83b94 | 4287 | break; |
8357376d | 4288 | |
73ce02e9 | 4289 | /* |
b8e83b94 MG |
4290 | * Raise priority if scanning rate is too low or there was no |
4291 | * progress in reclaiming pages | |
73ce02e9 | 4292 | */ |
c73322d0 | 4293 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
1c30844d MG |
4294 | nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed); |
4295 | ||
4296 | /* | |
4297 | * If reclaim made no progress for a boost, stop reclaim as | |
4298 | * IO cannot be queued and it could be an infinite loop in | |
4299 | * extreme circumstances. | |
4300 | */ | |
4301 | if (nr_boost_reclaim && !nr_reclaimed) | |
4302 | break; | |
4303 | ||
c73322d0 | 4304 | if (raise_priority || !nr_reclaimed) |
b8e83b94 | 4305 | sc.priority--; |
1d82de61 | 4306 | } while (sc.priority >= 1); |
1da177e4 | 4307 | |
c73322d0 JW |
4308 | if (!sc.nr_reclaimed) |
4309 | pgdat->kswapd_failures++; | |
4310 | ||
b8e83b94 | 4311 | out: |
c49c2c47 MG |
4312 | clear_reclaim_active(pgdat, highest_zoneidx); |
4313 | ||
1c30844d MG |
4314 | /* If reclaim was boosted, account for the reclaim done in this pass */ |
4315 | if (boosted) { | |
4316 | unsigned long flags; | |
4317 | ||
97a225e6 | 4318 | for (i = 0; i <= highest_zoneidx; i++) { |
1c30844d MG |
4319 | if (!zone_boosts[i]) |
4320 | continue; | |
4321 | ||
4322 | /* Increments are under the zone lock */ | |
4323 | zone = pgdat->node_zones + i; | |
4324 | spin_lock_irqsave(&zone->lock, flags); | |
4325 | zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]); | |
4326 | spin_unlock_irqrestore(&zone->lock, flags); | |
4327 | } | |
4328 | ||
4329 | /* | |
4330 | * As there is now likely space, wakeup kcompact to defragment | |
4331 | * pageblocks. | |
4332 | */ | |
97a225e6 | 4333 | wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx); |
1c30844d MG |
4334 | } |
4335 | ||
2a2e4885 | 4336 | snapshot_refaults(NULL, pgdat); |
4f3eaf45 | 4337 | __fs_reclaim_release(_THIS_IP_); |
eb414681 | 4338 | psi_memstall_leave(&pflags); |
1732d2b0 | 4339 | set_task_reclaim_state(current, NULL); |
e5ca8071 | 4340 | |
0abdee2b | 4341 | /* |
1d82de61 MG |
4342 | * Return the order kswapd stopped reclaiming at as |
4343 | * prepare_kswapd_sleep() takes it into account. If another caller | |
4344 | * entered the allocator slow path while kswapd was awake, order will | |
4345 | * remain at the higher level. | |
0abdee2b | 4346 | */ |
1d82de61 | 4347 | return sc.order; |
1da177e4 LT |
4348 | } |
4349 | ||
e716f2eb | 4350 | /* |
97a225e6 JK |
4351 | * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to |
4352 | * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is | |
4353 | * not a valid index then either kswapd runs for first time or kswapd couldn't | |
4354 | * sleep after previous reclaim attempt (node is still unbalanced). In that | |
4355 | * case return the zone index of the previous kswapd reclaim cycle. | |
e716f2eb | 4356 | */ |
97a225e6 JK |
4357 | static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat, |
4358 | enum zone_type prev_highest_zoneidx) | |
e716f2eb | 4359 | { |
97a225e6 | 4360 | enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx); |
5644e1fb | 4361 | |
97a225e6 | 4362 | return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx; |
e716f2eb MG |
4363 | } |
4364 | ||
38087d9b | 4365 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
97a225e6 | 4366 | unsigned int highest_zoneidx) |
f0bc0a60 KM |
4367 | { |
4368 | long remaining = 0; | |
4369 | DEFINE_WAIT(wait); | |
4370 | ||
4371 | if (freezing(current) || kthread_should_stop()) | |
4372 | return; | |
4373 | ||
4374 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
4375 | ||
333b0a45 SG |
4376 | /* |
4377 | * Try to sleep for a short interval. Note that kcompactd will only be | |
4378 | * woken if it is possible to sleep for a short interval. This is | |
4379 | * deliberate on the assumption that if reclaim cannot keep an | |
4380 | * eligible zone balanced that it's also unlikely that compaction will | |
4381 | * succeed. | |
4382 | */ | |
97a225e6 | 4383 | if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) { |
fd901c95 VB |
4384 | /* |
4385 | * Compaction records what page blocks it recently failed to | |
4386 | * isolate pages from and skips them in the future scanning. | |
4387 | * When kswapd is going to sleep, it is reasonable to assume | |
4388 | * that pages and compaction may succeed so reset the cache. | |
4389 | */ | |
4390 | reset_isolation_suitable(pgdat); | |
4391 | ||
4392 | /* | |
4393 | * We have freed the memory, now we should compact it to make | |
4394 | * allocation of the requested order possible. | |
4395 | */ | |
97a225e6 | 4396 | wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx); |
fd901c95 | 4397 | |
f0bc0a60 | 4398 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
4399 | |
4400 | /* | |
97a225e6 | 4401 | * If woken prematurely then reset kswapd_highest_zoneidx and |
38087d9b MG |
4402 | * order. The values will either be from a wakeup request or |
4403 | * the previous request that slept prematurely. | |
4404 | */ | |
4405 | if (remaining) { | |
97a225e6 JK |
4406 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, |
4407 | kswapd_highest_zoneidx(pgdat, | |
4408 | highest_zoneidx)); | |
5644e1fb QC |
4409 | |
4410 | if (READ_ONCE(pgdat->kswapd_order) < reclaim_order) | |
4411 | WRITE_ONCE(pgdat->kswapd_order, reclaim_order); | |
38087d9b MG |
4412 | } |
4413 | ||
f0bc0a60 KM |
4414 | finish_wait(&pgdat->kswapd_wait, &wait); |
4415 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
4416 | } | |
4417 | ||
4418 | /* | |
4419 | * After a short sleep, check if it was a premature sleep. If not, then | |
4420 | * go fully to sleep until explicitly woken up. | |
4421 | */ | |
d9f21d42 | 4422 | if (!remaining && |
97a225e6 | 4423 | prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) { |
f0bc0a60 KM |
4424 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
4425 | ||
4426 | /* | |
4427 | * vmstat counters are not perfectly accurate and the estimated | |
4428 | * value for counters such as NR_FREE_PAGES can deviate from the | |
4429 | * true value by nr_online_cpus * threshold. To avoid the zone | |
4430 | * watermarks being breached while under pressure, we reduce the | |
4431 | * per-cpu vmstat threshold while kswapd is awake and restore | |
4432 | * them before going back to sleep. | |
4433 | */ | |
4434 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
4435 | |
4436 | if (!kthread_should_stop()) | |
4437 | schedule(); | |
4438 | ||
f0bc0a60 KM |
4439 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
4440 | } else { | |
4441 | if (remaining) | |
4442 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
4443 | else | |
4444 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
4445 | } | |
4446 | finish_wait(&pgdat->kswapd_wait, &wait); | |
4447 | } | |
4448 | ||
1da177e4 LT |
4449 | /* |
4450 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 4451 | * from the init process. |
1da177e4 LT |
4452 | * |
4453 | * This basically trickles out pages so that we have _some_ | |
4454 | * free memory available even if there is no other activity | |
4455 | * that frees anything up. This is needed for things like routing | |
4456 | * etc, where we otherwise might have all activity going on in | |
4457 | * asynchronous contexts that cannot page things out. | |
4458 | * | |
4459 | * If there are applications that are active memory-allocators | |
4460 | * (most normal use), this basically shouldn't matter. | |
4461 | */ | |
4462 | static int kswapd(void *p) | |
4463 | { | |
e716f2eb | 4464 | unsigned int alloc_order, reclaim_order; |
97a225e6 | 4465 | unsigned int highest_zoneidx = MAX_NR_ZONES - 1; |
68d68ff6 | 4466 | pg_data_t *pgdat = (pg_data_t *)p; |
1da177e4 | 4467 | struct task_struct *tsk = current; |
a70f7302 | 4468 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 4469 | |
174596a0 | 4470 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 4471 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
4472 | |
4473 | /* | |
4474 | * Tell the memory management that we're a "memory allocator", | |
4475 | * and that if we need more memory we should get access to it | |
4476 | * regardless (see "__alloc_pages()"). "kswapd" should | |
4477 | * never get caught in the normal page freeing logic. | |
4478 | * | |
4479 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
4480 | * you need a small amount of memory in order to be able to | |
4481 | * page out something else, and this flag essentially protects | |
4482 | * us from recursively trying to free more memory as we're | |
4483 | * trying to free the first piece of memory in the first place). | |
4484 | */ | |
930d9152 | 4485 | tsk->flags |= PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD; |
83144186 | 4486 | set_freezable(); |
1da177e4 | 4487 | |
5644e1fb | 4488 | WRITE_ONCE(pgdat->kswapd_order, 0); |
97a225e6 | 4489 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES); |
8cd7c588 | 4490 | atomic_set(&pgdat->nr_writeback_throttled, 0); |
1da177e4 | 4491 | for ( ; ; ) { |
6f6313d4 | 4492 | bool ret; |
3e1d1d28 | 4493 | |
5644e1fb | 4494 | alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order); |
97a225e6 JK |
4495 | highest_zoneidx = kswapd_highest_zoneidx(pgdat, |
4496 | highest_zoneidx); | |
e716f2eb | 4497 | |
38087d9b MG |
4498 | kswapd_try_sleep: |
4499 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
97a225e6 | 4500 | highest_zoneidx); |
215ddd66 | 4501 | |
97a225e6 | 4502 | /* Read the new order and highest_zoneidx */ |
2b47a24c | 4503 | alloc_order = READ_ONCE(pgdat->kswapd_order); |
97a225e6 JK |
4504 | highest_zoneidx = kswapd_highest_zoneidx(pgdat, |
4505 | highest_zoneidx); | |
5644e1fb | 4506 | WRITE_ONCE(pgdat->kswapd_order, 0); |
97a225e6 | 4507 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES); |
1da177e4 | 4508 | |
8fe23e05 DR |
4509 | ret = try_to_freeze(); |
4510 | if (kthread_should_stop()) | |
4511 | break; | |
4512 | ||
4513 | /* | |
4514 | * We can speed up thawing tasks if we don't call balance_pgdat | |
4515 | * after returning from the refrigerator | |
4516 | */ | |
38087d9b MG |
4517 | if (ret) |
4518 | continue; | |
4519 | ||
4520 | /* | |
4521 | * Reclaim begins at the requested order but if a high-order | |
4522 | * reclaim fails then kswapd falls back to reclaiming for | |
4523 | * order-0. If that happens, kswapd will consider sleeping | |
4524 | * for the order it finished reclaiming at (reclaim_order) | |
4525 | * but kcompactd is woken to compact for the original | |
4526 | * request (alloc_order). | |
4527 | */ | |
97a225e6 | 4528 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx, |
e5146b12 | 4529 | alloc_order); |
97a225e6 JK |
4530 | reclaim_order = balance_pgdat(pgdat, alloc_order, |
4531 | highest_zoneidx); | |
38087d9b MG |
4532 | if (reclaim_order < alloc_order) |
4533 | goto kswapd_try_sleep; | |
1da177e4 | 4534 | } |
b0a8cc58 | 4535 | |
71abdc15 | 4536 | tsk->flags &= ~(PF_MEMALLOC | PF_SWAPWRITE | PF_KSWAPD); |
71abdc15 | 4537 | |
1da177e4 LT |
4538 | return 0; |
4539 | } | |
4540 | ||
4541 | /* | |
5ecd9d40 DR |
4542 | * A zone is low on free memory or too fragmented for high-order memory. If |
4543 | * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's | |
4544 | * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim | |
4545 | * has failed or is not needed, still wake up kcompactd if only compaction is | |
4546 | * needed. | |
1da177e4 | 4547 | */ |
5ecd9d40 | 4548 | void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, |
97a225e6 | 4549 | enum zone_type highest_zoneidx) |
1da177e4 LT |
4550 | { |
4551 | pg_data_t *pgdat; | |
5644e1fb | 4552 | enum zone_type curr_idx; |
1da177e4 | 4553 | |
6aa303de | 4554 | if (!managed_zone(zone)) |
1da177e4 LT |
4555 | return; |
4556 | ||
5ecd9d40 | 4557 | if (!cpuset_zone_allowed(zone, gfp_flags)) |
1da177e4 | 4558 | return; |
5644e1fb | 4559 | |
88f5acf8 | 4560 | pgdat = zone->zone_pgdat; |
97a225e6 | 4561 | curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx); |
5644e1fb | 4562 | |
97a225e6 JK |
4563 | if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx) |
4564 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx); | |
5644e1fb QC |
4565 | |
4566 | if (READ_ONCE(pgdat->kswapd_order) < order) | |
4567 | WRITE_ONCE(pgdat->kswapd_order, order); | |
dffcac2c | 4568 | |
8d0986e2 | 4569 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 4570 | return; |
e1a55637 | 4571 | |
5ecd9d40 DR |
4572 | /* Hopeless node, leave it to direct reclaim if possible */ |
4573 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES || | |
97a225e6 JK |
4574 | (pgdat_balanced(pgdat, order, highest_zoneidx) && |
4575 | !pgdat_watermark_boosted(pgdat, highest_zoneidx))) { | |
5ecd9d40 DR |
4576 | /* |
4577 | * There may be plenty of free memory available, but it's too | |
4578 | * fragmented for high-order allocations. Wake up kcompactd | |
4579 | * and rely on compaction_suitable() to determine if it's | |
4580 | * needed. If it fails, it will defer subsequent attempts to | |
4581 | * ratelimit its work. | |
4582 | */ | |
4583 | if (!(gfp_flags & __GFP_DIRECT_RECLAIM)) | |
97a225e6 | 4584 | wakeup_kcompactd(pgdat, order, highest_zoneidx); |
e716f2eb | 4585 | return; |
5ecd9d40 | 4586 | } |
88f5acf8 | 4587 | |
97a225e6 | 4588 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order, |
5ecd9d40 | 4589 | gfp_flags); |
8d0986e2 | 4590 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
4591 | } |
4592 | ||
c6f37f12 | 4593 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 4594 | /* |
7b51755c | 4595 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
4596 | * freed pages. |
4597 | * | |
4598 | * Rather than trying to age LRUs the aim is to preserve the overall | |
4599 | * LRU order by reclaiming preferentially | |
4600 | * inactive > active > active referenced > active mapped | |
1da177e4 | 4601 | */ |
7b51755c | 4602 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 4603 | { |
d6277db4 | 4604 | struct scan_control sc = { |
ee814fe2 | 4605 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 4606 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 4607 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 4608 | .priority = DEF_PRIORITY, |
d6277db4 | 4609 | .may_writepage = 1, |
ee814fe2 JW |
4610 | .may_unmap = 1, |
4611 | .may_swap = 1, | |
7b51755c | 4612 | .hibernation_mode = 1, |
1da177e4 | 4613 | }; |
a09ed5e0 | 4614 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c | 4615 | unsigned long nr_reclaimed; |
499118e9 | 4616 | unsigned int noreclaim_flag; |
1da177e4 | 4617 | |
d92a8cfc | 4618 | fs_reclaim_acquire(sc.gfp_mask); |
93781325 | 4619 | noreclaim_flag = memalloc_noreclaim_save(); |
1732d2b0 | 4620 | set_task_reclaim_state(current, &sc.reclaim_state); |
d6277db4 | 4621 | |
3115cd91 | 4622 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 4623 | |
1732d2b0 | 4624 | set_task_reclaim_state(current, NULL); |
499118e9 | 4625 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 4626 | fs_reclaim_release(sc.gfp_mask); |
d6277db4 | 4627 | |
7b51755c | 4628 | return nr_reclaimed; |
1da177e4 | 4629 | } |
c6f37f12 | 4630 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 4631 | |
3218ae14 YG |
4632 | /* |
4633 | * This kswapd start function will be called by init and node-hot-add. | |
4634 | * On node-hot-add, kswapd will moved to proper cpus if cpus are hot-added. | |
4635 | */ | |
b87c517a | 4636 | void kswapd_run(int nid) |
3218ae14 YG |
4637 | { |
4638 | pg_data_t *pgdat = NODE_DATA(nid); | |
3218ae14 YG |
4639 | |
4640 | if (pgdat->kswapd) | |
b87c517a | 4641 | return; |
3218ae14 YG |
4642 | |
4643 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
4644 | if (IS_ERR(pgdat->kswapd)) { | |
4645 | /* failure at boot is fatal */ | |
c6202adf | 4646 | BUG_ON(system_state < SYSTEM_RUNNING); |
d5dc0ad9 | 4647 | pr_err("Failed to start kswapd on node %d\n", nid); |
d72515b8 | 4648 | pgdat->kswapd = NULL; |
3218ae14 | 4649 | } |
3218ae14 YG |
4650 | } |
4651 | ||
8fe23e05 | 4652 | /* |
d8adde17 | 4653 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
bfc8c901 | 4654 | * hold mem_hotplug_begin/end(). |
8fe23e05 DR |
4655 | */ |
4656 | void kswapd_stop(int nid) | |
4657 | { | |
4658 | struct task_struct *kswapd = NODE_DATA(nid)->kswapd; | |
4659 | ||
d8adde17 | 4660 | if (kswapd) { |
8fe23e05 | 4661 | kthread_stop(kswapd); |
d8adde17 JL |
4662 | NODE_DATA(nid)->kswapd = NULL; |
4663 | } | |
8fe23e05 DR |
4664 | } |
4665 | ||
1da177e4 LT |
4666 | static int __init kswapd_init(void) |
4667 | { | |
6b700b5b | 4668 | int nid; |
69e05944 | 4669 | |
1da177e4 | 4670 | swap_setup(); |
48fb2e24 | 4671 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 4672 | kswapd_run(nid); |
1da177e4 LT |
4673 | return 0; |
4674 | } | |
4675 | ||
4676 | module_init(kswapd_init) | |
9eeff239 CL |
4677 | |
4678 | #ifdef CONFIG_NUMA | |
4679 | /* | |
a5f5f91d | 4680 | * Node reclaim mode |
9eeff239 | 4681 | * |
a5f5f91d | 4682 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 4683 | * the watermarks. |
9eeff239 | 4684 | */ |
a5f5f91d | 4685 | int node_reclaim_mode __read_mostly; |
9eeff239 | 4686 | |
a92f7126 | 4687 | /* |
a5f5f91d | 4688 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
4689 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
4690 | * a zone. | |
4691 | */ | |
a5f5f91d | 4692 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 4693 | |
9614634f | 4694 | /* |
a5f5f91d | 4695 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
4696 | * occur. |
4697 | */ | |
4698 | int sysctl_min_unmapped_ratio = 1; | |
4699 | ||
0ff38490 CL |
4700 | /* |
4701 | * If the number of slab pages in a zone grows beyond this percentage then | |
4702 | * slab reclaim needs to occur. | |
4703 | */ | |
4704 | int sysctl_min_slab_ratio = 5; | |
4705 | ||
11fb9989 | 4706 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 4707 | { |
11fb9989 MG |
4708 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
4709 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
4710 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
4711 | |
4712 | /* | |
4713 | * It's possible for there to be more file mapped pages than | |
4714 | * accounted for by the pages on the file LRU lists because | |
4715 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
4716 | */ | |
4717 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
4718 | } | |
4719 | ||
4720 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 4721 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 4722 | { |
d031a157 AM |
4723 | unsigned long nr_pagecache_reclaimable; |
4724 | unsigned long delta = 0; | |
90afa5de MG |
4725 | |
4726 | /* | |
95bbc0c7 | 4727 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 4728 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 4729 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
4730 | * a better estimate |
4731 | */ | |
a5f5f91d MG |
4732 | if (node_reclaim_mode & RECLAIM_UNMAP) |
4733 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 4734 | else |
a5f5f91d | 4735 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
4736 | |
4737 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
4738 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
4739 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
4740 | |
4741 | /* Watch for any possible underflows due to delta */ | |
4742 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
4743 | delta = nr_pagecache_reclaimable; | |
4744 | ||
4745 | return nr_pagecache_reclaimable - delta; | |
4746 | } | |
4747 | ||
9eeff239 | 4748 | /* |
a5f5f91d | 4749 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 4750 | */ |
a5f5f91d | 4751 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 4752 | { |
7fb2d46d | 4753 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 4754 | const unsigned long nr_pages = 1 << order; |
9eeff239 | 4755 | struct task_struct *p = current; |
499118e9 | 4756 | unsigned int noreclaim_flag; |
179e9639 | 4757 | struct scan_control sc = { |
62b726c1 | 4758 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
f2f43e56 | 4759 | .gfp_mask = current_gfp_context(gfp_mask), |
bd2f6199 | 4760 | .order = order, |
a5f5f91d MG |
4761 | .priority = NODE_RECLAIM_PRIORITY, |
4762 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
4763 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 4764 | .may_swap = 1, |
f2f43e56 | 4765 | .reclaim_idx = gfp_zone(gfp_mask), |
179e9639 | 4766 | }; |
57f29762 | 4767 | unsigned long pflags; |
9eeff239 | 4768 | |
132bb8cf YS |
4769 | trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order, |
4770 | sc.gfp_mask); | |
4771 | ||
9eeff239 | 4772 | cond_resched(); |
57f29762 | 4773 | psi_memstall_enter(&pflags); |
93781325 | 4774 | fs_reclaim_acquire(sc.gfp_mask); |
d4f7796e | 4775 | /* |
95bbc0c7 | 4776 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 4777 | * and we also need to be able to write out pages for RECLAIM_WRITE |
95bbc0c7 | 4778 | * and RECLAIM_UNMAP. |
d4f7796e | 4779 | */ |
499118e9 VB |
4780 | noreclaim_flag = memalloc_noreclaim_save(); |
4781 | p->flags |= PF_SWAPWRITE; | |
1732d2b0 | 4782 | set_task_reclaim_state(p, &sc.reclaim_state); |
c84db23c | 4783 | |
a5f5f91d | 4784 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages) { |
0ff38490 | 4785 | /* |
894befec | 4786 | * Free memory by calling shrink node with increasing |
0ff38490 CL |
4787 | * priorities until we have enough memory freed. |
4788 | */ | |
0ff38490 | 4789 | do { |
970a39a3 | 4790 | shrink_node(pgdat, &sc); |
9e3b2f8c | 4791 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 4792 | } |
c84db23c | 4793 | |
1732d2b0 | 4794 | set_task_reclaim_state(p, NULL); |
499118e9 VB |
4795 | current->flags &= ~PF_SWAPWRITE; |
4796 | memalloc_noreclaim_restore(noreclaim_flag); | |
93781325 | 4797 | fs_reclaim_release(sc.gfp_mask); |
57f29762 | 4798 | psi_memstall_leave(&pflags); |
132bb8cf YS |
4799 | |
4800 | trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed); | |
4801 | ||
a79311c1 | 4802 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 4803 | } |
179e9639 | 4804 | |
a5f5f91d | 4805 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 4806 | { |
d773ed6b | 4807 | int ret; |
179e9639 AM |
4808 | |
4809 | /* | |
a5f5f91d | 4810 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 4811 | * slab pages if we are over the defined limits. |
34aa1330 | 4812 | * |
9614634f CL |
4813 | * A small portion of unmapped file backed pages is needed for |
4814 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
4815 | * thrown out if the node is overallocated. So we do not reclaim |
4816 | * if less than a specified percentage of the node is used by | |
9614634f | 4817 | * unmapped file backed pages. |
179e9639 | 4818 | */ |
a5f5f91d | 4819 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
d42f3245 RG |
4820 | node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <= |
4821 | pgdat->min_slab_pages) | |
a5f5f91d | 4822 | return NODE_RECLAIM_FULL; |
179e9639 AM |
4823 | |
4824 | /* | |
d773ed6b | 4825 | * Do not scan if the allocation should not be delayed. |
179e9639 | 4826 | */ |
d0164adc | 4827 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 4828 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
4829 | |
4830 | /* | |
a5f5f91d | 4831 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
4832 | * have associated processors. This will favor the local processor |
4833 | * over remote processors and spread off node memory allocations | |
4834 | * as wide as possible. | |
4835 | */ | |
a5f5f91d MG |
4836 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
4837 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 4838 | |
a5f5f91d MG |
4839 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
4840 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 4841 | |
a5f5f91d MG |
4842 | ret = __node_reclaim(pgdat, gfp_mask, order); |
4843 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 4844 | |
24cf7251 MG |
4845 | if (!ret) |
4846 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
4847 | ||
d773ed6b | 4848 | return ret; |
179e9639 | 4849 | } |
9eeff239 | 4850 | #endif |
894bc310 | 4851 | |
89e004ea | 4852 | /** |
64e3d12f KHY |
4853 | * check_move_unevictable_pages - check pages for evictability and move to |
4854 | * appropriate zone lru list | |
4855 | * @pvec: pagevec with lru pages to check | |
89e004ea | 4856 | * |
64e3d12f KHY |
4857 | * Checks pages for evictability, if an evictable page is in the unevictable |
4858 | * lru list, moves it to the appropriate evictable lru list. This function | |
4859 | * should be only used for lru pages. | |
89e004ea | 4860 | */ |
64e3d12f | 4861 | void check_move_unevictable_pages(struct pagevec *pvec) |
89e004ea | 4862 | { |
6168d0da | 4863 | struct lruvec *lruvec = NULL; |
24513264 HD |
4864 | int pgscanned = 0; |
4865 | int pgrescued = 0; | |
4866 | int i; | |
89e004ea | 4867 | |
64e3d12f KHY |
4868 | for (i = 0; i < pvec->nr; i++) { |
4869 | struct page *page = pvec->pages[i]; | |
0de340cb | 4870 | struct folio *folio = page_folio(page); |
8d8869ca HD |
4871 | int nr_pages; |
4872 | ||
4873 | if (PageTransTail(page)) | |
4874 | continue; | |
4875 | ||
4876 | nr_pages = thp_nr_pages(page); | |
4877 | pgscanned += nr_pages; | |
89e004ea | 4878 | |
d25b5bd8 AS |
4879 | /* block memcg migration during page moving between lru */ |
4880 | if (!TestClearPageLRU(page)) | |
4881 | continue; | |
4882 | ||
0de340cb | 4883 | lruvec = folio_lruvec_relock_irq(folio, lruvec); |
d25b5bd8 | 4884 | if (page_evictable(page) && PageUnevictable(page)) { |
46ae6b2c | 4885 | del_page_from_lru_list(page, lruvec); |
24513264 | 4886 | ClearPageUnevictable(page); |
3a9c9788 | 4887 | add_page_to_lru_list(page, lruvec); |
8d8869ca | 4888 | pgrescued += nr_pages; |
89e004ea | 4889 | } |
d25b5bd8 | 4890 | SetPageLRU(page); |
24513264 | 4891 | } |
89e004ea | 4892 | |
6168d0da | 4893 | if (lruvec) { |
24513264 HD |
4894 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
4895 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
6168d0da | 4896 | unlock_page_lruvec_irq(lruvec); |
d25b5bd8 AS |
4897 | } else if (pgscanned) { |
4898 | count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
89e004ea | 4899 | } |
89e004ea | 4900 | } |
64e3d12f | 4901 | EXPORT_SYMBOL_GPL(check_move_unevictable_pages); |