]>
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> | |
07f67a8d | 29 | #include <linux/buffer_head.h> /* for buffer_heads_over_limit */ |
1da177e4 | 30 | #include <linux/mm_inline.h> |
1da177e4 LT |
31 | #include <linux/backing-dev.h> |
32 | #include <linux/rmap.h> | |
33 | #include <linux/topology.h> | |
34 | #include <linux/cpu.h> | |
35 | #include <linux/cpuset.h> | |
3e7d3449 | 36 | #include <linux/compaction.h> |
1da177e4 LT |
37 | #include <linux/notifier.h> |
38 | #include <linux/rwsem.h> | |
248a0301 | 39 | #include <linux/delay.h> |
3218ae14 | 40 | #include <linux/kthread.h> |
7dfb7103 | 41 | #include <linux/freezer.h> |
66e1707b | 42 | #include <linux/memcontrol.h> |
26aa2d19 | 43 | #include <linux/migrate.h> |
873b4771 | 44 | #include <linux/delayacct.h> |
af936a16 | 45 | #include <linux/sysctl.h> |
91952440 | 46 | #include <linux/memory-tiers.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> |
bd74fdae YZ |
53 | #include <linux/pagewalk.h> |
54 | #include <linux/shmem_fs.h> | |
354ed597 | 55 | #include <linux/ctype.h> |
d6c3af7d | 56 | #include <linux/debugfs.h> |
57e9cc50 | 57 | #include <linux/khugepaged.h> |
1da177e4 LT |
58 | |
59 | #include <asm/tlbflush.h> | |
60 | #include <asm/div64.h> | |
61 | ||
62 | #include <linux/swapops.h> | |
117aad1e | 63 | #include <linux/balloon_compaction.h> |
c574bbe9 | 64 | #include <linux/sched/sysctl.h> |
1da177e4 | 65 | |
0f8053a5 | 66 | #include "internal.h" |
014bb1de | 67 | #include "swap.h" |
0f8053a5 | 68 | |
33906bc5 MG |
69 | #define CREATE_TRACE_POINTS |
70 | #include <trace/events/vmscan.h> | |
71 | ||
1da177e4 | 72 | struct scan_control { |
22fba335 KM |
73 | /* How many pages shrink_list() should reclaim */ |
74 | unsigned long nr_to_reclaim; | |
75 | ||
ee814fe2 JW |
76 | /* |
77 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
78 | * are scanned. | |
79 | */ | |
80 | nodemask_t *nodemask; | |
9e3b2f8c | 81 | |
f16015fb JW |
82 | /* |
83 | * The memory cgroup that hit its limit and as a result is the | |
84 | * primary target of this reclaim invocation. | |
85 | */ | |
86 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 87 | |
7cf111bc JW |
88 | /* |
89 | * Scan pressure balancing between anon and file LRUs | |
90 | */ | |
91 | unsigned long anon_cost; | |
92 | unsigned long file_cost; | |
93 | ||
49fd9b6d | 94 | /* Can active folios be deactivated as part of reclaim? */ |
b91ac374 JW |
95 | #define DEACTIVATE_ANON 1 |
96 | #define DEACTIVATE_FILE 2 | |
97 | unsigned int may_deactivate:2; | |
98 | unsigned int force_deactivate:1; | |
99 | unsigned int skipped_deactivate:1; | |
100 | ||
1276ad68 | 101 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
ee814fe2 JW |
102 | unsigned int may_writepage:1; |
103 | ||
49fd9b6d | 104 | /* Can mapped folios be reclaimed? */ |
ee814fe2 JW |
105 | unsigned int may_unmap:1; |
106 | ||
49fd9b6d | 107 | /* Can folios be swapped as part of reclaim? */ |
ee814fe2 JW |
108 | unsigned int may_swap:1; |
109 | ||
73b73bac YA |
110 | /* Proactive reclaim invoked by userspace through memory.reclaim */ |
111 | unsigned int proactive:1; | |
112 | ||
d6622f63 | 113 | /* |
f56ce412 JW |
114 | * Cgroup memory below memory.low is protected as long as we |
115 | * don't threaten to OOM. If any cgroup is reclaimed at | |
116 | * reduced force or passed over entirely due to its memory.low | |
117 | * setting (memcg_low_skipped), and nothing is reclaimed as a | |
118 | * result, then go back for one more cycle that reclaims the protected | |
119 | * memory (memcg_low_reclaim) to avert OOM. | |
d6622f63 YX |
120 | */ |
121 | unsigned int memcg_low_reclaim:1; | |
122 | unsigned int memcg_low_skipped:1; | |
241994ed | 123 | |
ee814fe2 JW |
124 | unsigned int hibernation_mode:1; |
125 | ||
126 | /* One of the zones is ready for compaction */ | |
127 | unsigned int compaction_ready:1; | |
128 | ||
b91ac374 JW |
129 | /* There is easily reclaimable cold cache in the current node */ |
130 | unsigned int cache_trim_mode:1; | |
131 | ||
49fd9b6d | 132 | /* The file folios on the current node are dangerously low */ |
53138cea JW |
133 | unsigned int file_is_tiny:1; |
134 | ||
26aa2d19 DH |
135 | /* Always discard instead of demoting to lower tier memory */ |
136 | unsigned int no_demotion:1; | |
137 | ||
f76c8337 YZ |
138 | #ifdef CONFIG_LRU_GEN |
139 | /* help kswapd make better choices among multiple memcgs */ | |
140 | unsigned int memcgs_need_aging:1; | |
141 | unsigned long last_reclaimed; | |
142 | #endif | |
143 | ||
bb451fdf GT |
144 | /* Allocation order */ |
145 | s8 order; | |
146 | ||
147 | /* Scan (total_size >> priority) pages at once */ | |
148 | s8 priority; | |
149 | ||
49fd9b6d | 150 | /* The highest zone to isolate folios for reclaim from */ |
bb451fdf GT |
151 | s8 reclaim_idx; |
152 | ||
153 | /* This context's GFP mask */ | |
154 | gfp_t gfp_mask; | |
155 | ||
ee814fe2 JW |
156 | /* Incremented by the number of inactive pages that were scanned */ |
157 | unsigned long nr_scanned; | |
158 | ||
159 | /* Number of pages freed so far during a call to shrink_zones() */ | |
160 | unsigned long nr_reclaimed; | |
d108c772 AR |
161 | |
162 | struct { | |
163 | unsigned int dirty; | |
164 | unsigned int unqueued_dirty; | |
165 | unsigned int congested; | |
166 | unsigned int writeback; | |
167 | unsigned int immediate; | |
168 | unsigned int file_taken; | |
169 | unsigned int taken; | |
170 | } nr; | |
e5ca8071 YS |
171 | |
172 | /* for recording the reclaimed slab by now */ | |
173 | struct reclaim_state reclaim_state; | |
1da177e4 LT |
174 | }; |
175 | ||
1da177e4 | 176 | #ifdef ARCH_HAS_PREFETCHW |
166e3d32 | 177 | #define prefetchw_prev_lru_folio(_folio, _base, _field) \ |
1da177e4 | 178 | do { \ |
166e3d32 MWO |
179 | if ((_folio)->lru.prev != _base) { \ |
180 | struct folio *prev; \ | |
1da177e4 | 181 | \ |
166e3d32 | 182 | prev = lru_to_folio(&(_folio->lru)); \ |
1da177e4 LT |
183 | prefetchw(&prev->_field); \ |
184 | } \ | |
185 | } while (0) | |
186 | #else | |
166e3d32 | 187 | #define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0) |
1da177e4 LT |
188 | #endif |
189 | ||
190 | /* | |
c843966c | 191 | * From 0 .. 200. Higher means more swappy. |
1da177e4 LT |
192 | */ |
193 | int vm_swappiness = 60; | |
1da177e4 | 194 | |
0a432dcb YS |
195 | static void set_task_reclaim_state(struct task_struct *task, |
196 | struct reclaim_state *rs) | |
197 | { | |
198 | /* Check for an overwrite */ | |
199 | WARN_ON_ONCE(rs && task->reclaim_state); | |
200 | ||
201 | /* Check for the nulling of an already-nulled member */ | |
202 | WARN_ON_ONCE(!rs && !task->reclaim_state); | |
203 | ||
204 | task->reclaim_state = rs; | |
205 | } | |
206 | ||
5035ebc6 RG |
207 | LIST_HEAD(shrinker_list); |
208 | DECLARE_RWSEM(shrinker_rwsem); | |
1da177e4 | 209 | |
0a432dcb | 210 | #ifdef CONFIG_MEMCG |
a2fb1261 | 211 | static int shrinker_nr_max; |
2bfd3637 | 212 | |
3c6f17e6 | 213 | /* The shrinker_info is expanded in a batch of BITS_PER_LONG */ |
a2fb1261 YS |
214 | static inline int shrinker_map_size(int nr_items) |
215 | { | |
216 | return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long)); | |
217 | } | |
2bfd3637 | 218 | |
3c6f17e6 YS |
219 | static inline int shrinker_defer_size(int nr_items) |
220 | { | |
221 | return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t)); | |
222 | } | |
223 | ||
468ab843 YS |
224 | static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg, |
225 | int nid) | |
226 | { | |
227 | return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info, | |
228 | lockdep_is_held(&shrinker_rwsem)); | |
229 | } | |
230 | ||
e4262c4f | 231 | static int expand_one_shrinker_info(struct mem_cgroup *memcg, |
3c6f17e6 YS |
232 | int map_size, int defer_size, |
233 | int old_map_size, int old_defer_size) | |
2bfd3637 | 234 | { |
e4262c4f | 235 | struct shrinker_info *new, *old; |
2bfd3637 YS |
236 | struct mem_cgroup_per_node *pn; |
237 | int nid; | |
3c6f17e6 | 238 | int size = map_size + defer_size; |
2bfd3637 | 239 | |
2bfd3637 YS |
240 | for_each_node(nid) { |
241 | pn = memcg->nodeinfo[nid]; | |
468ab843 | 242 | old = shrinker_info_protected(memcg, nid); |
2bfd3637 YS |
243 | /* Not yet online memcg */ |
244 | if (!old) | |
245 | return 0; | |
246 | ||
247 | new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid); | |
248 | if (!new) | |
249 | return -ENOMEM; | |
250 | ||
3c6f17e6 YS |
251 | new->nr_deferred = (atomic_long_t *)(new + 1); |
252 | new->map = (void *)new->nr_deferred + defer_size; | |
253 | ||
254 | /* map: set all old bits, clear all new bits */ | |
255 | memset(new->map, (int)0xff, old_map_size); | |
256 | memset((void *)new->map + old_map_size, 0, map_size - old_map_size); | |
257 | /* nr_deferred: copy old values, clear all new values */ | |
258 | memcpy(new->nr_deferred, old->nr_deferred, old_defer_size); | |
259 | memset((void *)new->nr_deferred + old_defer_size, 0, | |
260 | defer_size - old_defer_size); | |
2bfd3637 | 261 | |
e4262c4f | 262 | rcu_assign_pointer(pn->shrinker_info, new); |
72673e86 | 263 | kvfree_rcu(old, rcu); |
2bfd3637 YS |
264 | } |
265 | ||
266 | return 0; | |
267 | } | |
268 | ||
e4262c4f | 269 | void free_shrinker_info(struct mem_cgroup *memcg) |
2bfd3637 YS |
270 | { |
271 | struct mem_cgroup_per_node *pn; | |
e4262c4f | 272 | struct shrinker_info *info; |
2bfd3637 YS |
273 | int nid; |
274 | ||
2bfd3637 YS |
275 | for_each_node(nid) { |
276 | pn = memcg->nodeinfo[nid]; | |
e4262c4f YS |
277 | info = rcu_dereference_protected(pn->shrinker_info, true); |
278 | kvfree(info); | |
279 | rcu_assign_pointer(pn->shrinker_info, NULL); | |
2bfd3637 YS |
280 | } |
281 | } | |
282 | ||
e4262c4f | 283 | int alloc_shrinker_info(struct mem_cgroup *memcg) |
2bfd3637 | 284 | { |
e4262c4f | 285 | struct shrinker_info *info; |
2bfd3637 | 286 | int nid, size, ret = 0; |
3c6f17e6 | 287 | int map_size, defer_size = 0; |
2bfd3637 | 288 | |
d27cf2aa | 289 | down_write(&shrinker_rwsem); |
3c6f17e6 YS |
290 | map_size = shrinker_map_size(shrinker_nr_max); |
291 | defer_size = shrinker_defer_size(shrinker_nr_max); | |
292 | size = map_size + defer_size; | |
2bfd3637 | 293 | for_each_node(nid) { |
e4262c4f YS |
294 | info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid); |
295 | if (!info) { | |
296 | free_shrinker_info(memcg); | |
2bfd3637 YS |
297 | ret = -ENOMEM; |
298 | break; | |
299 | } | |
3c6f17e6 YS |
300 | info->nr_deferred = (atomic_long_t *)(info + 1); |
301 | info->map = (void *)info->nr_deferred + defer_size; | |
e4262c4f | 302 | rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info); |
2bfd3637 | 303 | } |
d27cf2aa | 304 | up_write(&shrinker_rwsem); |
2bfd3637 YS |
305 | |
306 | return ret; | |
307 | } | |
308 | ||
3c6f17e6 YS |
309 | static inline bool need_expand(int nr_max) |
310 | { | |
311 | return round_up(nr_max, BITS_PER_LONG) > | |
312 | round_up(shrinker_nr_max, BITS_PER_LONG); | |
313 | } | |
314 | ||
e4262c4f | 315 | static int expand_shrinker_info(int new_id) |
2bfd3637 | 316 | { |
3c6f17e6 | 317 | int ret = 0; |
a2fb1261 | 318 | int new_nr_max = new_id + 1; |
3c6f17e6 YS |
319 | int map_size, defer_size = 0; |
320 | int old_map_size, old_defer_size = 0; | |
2bfd3637 YS |
321 | struct mem_cgroup *memcg; |
322 | ||
3c6f17e6 | 323 | if (!need_expand(new_nr_max)) |
a2fb1261 | 324 | goto out; |
2bfd3637 | 325 | |
2bfd3637 | 326 | if (!root_mem_cgroup) |
d27cf2aa YS |
327 | goto out; |
328 | ||
329 | lockdep_assert_held(&shrinker_rwsem); | |
2bfd3637 | 330 | |
3c6f17e6 YS |
331 | map_size = shrinker_map_size(new_nr_max); |
332 | defer_size = shrinker_defer_size(new_nr_max); | |
333 | old_map_size = shrinker_map_size(shrinker_nr_max); | |
334 | old_defer_size = shrinker_defer_size(shrinker_nr_max); | |
335 | ||
2bfd3637 YS |
336 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
337 | do { | |
3c6f17e6 YS |
338 | ret = expand_one_shrinker_info(memcg, map_size, defer_size, |
339 | old_map_size, old_defer_size); | |
2bfd3637 YS |
340 | if (ret) { |
341 | mem_cgroup_iter_break(NULL, memcg); | |
d27cf2aa | 342 | goto out; |
2bfd3637 YS |
343 | } |
344 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
d27cf2aa | 345 | out: |
2bfd3637 | 346 | if (!ret) |
a2fb1261 | 347 | shrinker_nr_max = new_nr_max; |
d27cf2aa | 348 | |
2bfd3637 YS |
349 | return ret; |
350 | } | |
351 | ||
352 | void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id) | |
353 | { | |
354 | if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) { | |
e4262c4f | 355 | struct shrinker_info *info; |
2bfd3637 YS |
356 | |
357 | rcu_read_lock(); | |
e4262c4f | 358 | info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info); |
2bfd3637 YS |
359 | /* Pairs with smp mb in shrink_slab() */ |
360 | smp_mb__before_atomic(); | |
e4262c4f | 361 | set_bit(shrinker_id, info->map); |
2bfd3637 YS |
362 | rcu_read_unlock(); |
363 | } | |
364 | } | |
365 | ||
b4c2b231 | 366 | static DEFINE_IDR(shrinker_idr); |
b4c2b231 KT |
367 | |
368 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) | |
369 | { | |
370 | int id, ret = -ENOMEM; | |
371 | ||
476b30a0 YS |
372 | if (mem_cgroup_disabled()) |
373 | return -ENOSYS; | |
374 | ||
b4c2b231 KT |
375 | down_write(&shrinker_rwsem); |
376 | /* This may call shrinker, so it must use down_read_trylock() */ | |
41ca668a | 377 | id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL); |
b4c2b231 KT |
378 | if (id < 0) |
379 | goto unlock; | |
380 | ||
0a4465d3 | 381 | if (id >= shrinker_nr_max) { |
e4262c4f | 382 | if (expand_shrinker_info(id)) { |
0a4465d3 KT |
383 | idr_remove(&shrinker_idr, id); |
384 | goto unlock; | |
385 | } | |
0a4465d3 | 386 | } |
b4c2b231 KT |
387 | shrinker->id = id; |
388 | ret = 0; | |
389 | unlock: | |
390 | up_write(&shrinker_rwsem); | |
391 | return ret; | |
392 | } | |
393 | ||
394 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
395 | { | |
396 | int id = shrinker->id; | |
397 | ||
398 | BUG_ON(id < 0); | |
399 | ||
41ca668a YS |
400 | lockdep_assert_held(&shrinker_rwsem); |
401 | ||
b4c2b231 | 402 | idr_remove(&shrinker_idr, id); |
b4c2b231 | 403 | } |
b4c2b231 | 404 | |
86750830 YS |
405 | static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker, |
406 | struct mem_cgroup *memcg) | |
407 | { | |
408 | struct shrinker_info *info; | |
409 | ||
410 | info = shrinker_info_protected(memcg, nid); | |
411 | return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0); | |
412 | } | |
413 | ||
414 | static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker, | |
415 | struct mem_cgroup *memcg) | |
416 | { | |
417 | struct shrinker_info *info; | |
418 | ||
419 | info = shrinker_info_protected(memcg, nid); | |
420 | return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]); | |
421 | } | |
422 | ||
a178015c YS |
423 | void reparent_shrinker_deferred(struct mem_cgroup *memcg) |
424 | { | |
425 | int i, nid; | |
426 | long nr; | |
427 | struct mem_cgroup *parent; | |
428 | struct shrinker_info *child_info, *parent_info; | |
429 | ||
430 | parent = parent_mem_cgroup(memcg); | |
431 | if (!parent) | |
432 | parent = root_mem_cgroup; | |
433 | ||
434 | /* Prevent from concurrent shrinker_info expand */ | |
435 | down_read(&shrinker_rwsem); | |
436 | for_each_node(nid) { | |
437 | child_info = shrinker_info_protected(memcg, nid); | |
438 | parent_info = shrinker_info_protected(parent, nid); | |
439 | for (i = 0; i < shrinker_nr_max; i++) { | |
440 | nr = atomic_long_read(&child_info->nr_deferred[i]); | |
441 | atomic_long_add(nr, &parent_info->nr_deferred[i]); | |
442 | } | |
443 | } | |
444 | up_read(&shrinker_rwsem); | |
445 | } | |
446 | ||
b5ead35e | 447 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 448 | { |
b5ead35e | 449 | return sc->target_mem_cgroup; |
89b5fae5 | 450 | } |
97c9341f | 451 | |
a579086c YZ |
452 | static bool global_reclaim(struct scan_control *sc) |
453 | { | |
454 | return !sc->target_mem_cgroup || mem_cgroup_is_root(sc->target_mem_cgroup); | |
455 | } | |
456 | ||
97c9341f | 457 | /** |
b5ead35e | 458 | * writeback_throttling_sane - is the usual dirty throttling mechanism available? |
97c9341f TH |
459 | * @sc: scan_control in question |
460 | * | |
461 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
462 | * completely broken with the legacy memcg and direct stalling in | |
49fd9b6d | 463 | * shrink_folio_list() is used for throttling instead, which lacks all the |
97c9341f TH |
464 | * niceties such as fairness, adaptive pausing, bandwidth proportional |
465 | * allocation and configurability. | |
466 | * | |
467 | * This function tests whether the vmscan currently in progress can assume | |
468 | * that the normal dirty throttling mechanism is operational. | |
469 | */ | |
b5ead35e | 470 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f | 471 | { |
b5ead35e | 472 | if (!cgroup_reclaim(sc)) |
97c9341f TH |
473 | return true; |
474 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 475 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
476 | return true; |
477 | #endif | |
478 | return false; | |
479 | } | |
91a45470 | 480 | #else |
0a432dcb YS |
481 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) |
482 | { | |
476b30a0 | 483 | return -ENOSYS; |
0a432dcb YS |
484 | } |
485 | ||
486 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
487 | { | |
488 | } | |
489 | ||
86750830 YS |
490 | static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker, |
491 | struct mem_cgroup *memcg) | |
492 | { | |
493 | return 0; | |
494 | } | |
495 | ||
496 | static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker, | |
497 | struct mem_cgroup *memcg) | |
498 | { | |
499 | return 0; | |
500 | } | |
501 | ||
b5ead35e | 502 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 503 | { |
b5ead35e | 504 | return false; |
89b5fae5 | 505 | } |
97c9341f | 506 | |
a579086c YZ |
507 | static bool global_reclaim(struct scan_control *sc) |
508 | { | |
509 | return true; | |
510 | } | |
511 | ||
b5ead35e | 512 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f TH |
513 | { |
514 | return true; | |
515 | } | |
91a45470 KH |
516 | #endif |
517 | ||
86750830 YS |
518 | static long xchg_nr_deferred(struct shrinker *shrinker, |
519 | struct shrink_control *sc) | |
520 | { | |
521 | int nid = sc->nid; | |
522 | ||
523 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) | |
524 | nid = 0; | |
525 | ||
526 | if (sc->memcg && | |
527 | (shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
528 | return xchg_nr_deferred_memcg(nid, shrinker, | |
529 | sc->memcg); | |
530 | ||
531 | return atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
532 | } | |
533 | ||
534 | ||
535 | static long add_nr_deferred(long nr, struct shrinker *shrinker, | |
536 | struct shrink_control *sc) | |
537 | { | |
538 | int nid = sc->nid; | |
539 | ||
540 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) | |
541 | nid = 0; | |
542 | ||
543 | if (sc->memcg && | |
544 | (shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
545 | return add_nr_deferred_memcg(nr, nid, shrinker, | |
546 | sc->memcg); | |
547 | ||
548 | return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]); | |
549 | } | |
550 | ||
26aa2d19 DH |
551 | static bool can_demote(int nid, struct scan_control *sc) |
552 | { | |
20b51af1 HY |
553 | if (!numa_demotion_enabled) |
554 | return false; | |
3f1509c5 JW |
555 | if (sc && sc->no_demotion) |
556 | return false; | |
26aa2d19 DH |
557 | if (next_demotion_node(nid) == NUMA_NO_NODE) |
558 | return false; | |
559 | ||
20b51af1 | 560 | return true; |
26aa2d19 DH |
561 | } |
562 | ||
a2a36488 KB |
563 | static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg, |
564 | int nid, | |
565 | struct scan_control *sc) | |
566 | { | |
567 | if (memcg == NULL) { | |
568 | /* | |
569 | * For non-memcg reclaim, is there | |
570 | * space in any swap device? | |
571 | */ | |
572 | if (get_nr_swap_pages() > 0) | |
573 | return true; | |
574 | } else { | |
575 | /* Is the memcg below its swap limit? */ | |
576 | if (mem_cgroup_get_nr_swap_pages(memcg) > 0) | |
577 | return true; | |
578 | } | |
579 | ||
580 | /* | |
581 | * The page can not be swapped. | |
582 | * | |
583 | * Can it be reclaimed from this node via demotion? | |
584 | */ | |
585 | return can_demote(nid, sc); | |
586 | } | |
587 | ||
5a1c84b4 | 588 | /* |
49fd9b6d | 589 | * This misses isolated folios which are not accounted for to save counters. |
5a1c84b4 | 590 | * As the data only determines if reclaim or compaction continues, it is |
49fd9b6d | 591 | * not expected that isolated folios will be a dominating factor. |
5a1c84b4 MG |
592 | */ |
593 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
594 | { | |
595 | unsigned long nr; | |
596 | ||
597 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
598 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
a2a36488 | 599 | if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL)) |
5a1c84b4 MG |
600 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + |
601 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
602 | ||
603 | return nr; | |
604 | } | |
605 | ||
fd538803 MH |
606 | /** |
607 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
608 | * @lruvec: lru vector | |
609 | * @lru: lru to use | |
8b3a899a | 610 | * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list) |
fd538803 | 611 | */ |
2091339d YZ |
612 | static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, |
613 | int zone_idx) | |
c9f299d9 | 614 | { |
de3b0150 | 615 | unsigned long size = 0; |
fd538803 MH |
616 | int zid; |
617 | ||
8b3a899a | 618 | for (zid = 0; zid <= zone_idx; zid++) { |
fd538803 | 619 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; |
c9f299d9 | 620 | |
fd538803 MH |
621 | if (!managed_zone(zone)) |
622 | continue; | |
623 | ||
624 | if (!mem_cgroup_disabled()) | |
de3b0150 | 625 | size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid); |
fd538803 | 626 | else |
de3b0150 | 627 | size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru); |
fd538803 | 628 | } |
de3b0150 | 629 | return size; |
b4536f0c MH |
630 | } |
631 | ||
1da177e4 | 632 | /* |
1d3d4437 | 633 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 634 | */ |
e33c267a | 635 | static int __prealloc_shrinker(struct shrinker *shrinker) |
1da177e4 | 636 | { |
476b30a0 YS |
637 | unsigned int size; |
638 | int err; | |
639 | ||
640 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) { | |
641 | err = prealloc_memcg_shrinker(shrinker); | |
642 | if (err != -ENOSYS) | |
643 | return err; | |
1d3d4437 | 644 | |
476b30a0 YS |
645 | shrinker->flags &= ~SHRINKER_MEMCG_AWARE; |
646 | } | |
647 | ||
648 | size = sizeof(*shrinker->nr_deferred); | |
1d3d4437 GC |
649 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
650 | size *= nr_node_ids; | |
651 | ||
652 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
653 | if (!shrinker->nr_deferred) | |
654 | return -ENOMEM; | |
b4c2b231 | 655 | |
8e04944f TH |
656 | return 0; |
657 | } | |
658 | ||
e33c267a RG |
659 | #ifdef CONFIG_SHRINKER_DEBUG |
660 | int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...) | |
661 | { | |
662 | va_list ap; | |
663 | int err; | |
664 | ||
665 | va_start(ap, fmt); | |
666 | shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap); | |
667 | va_end(ap); | |
668 | if (!shrinker->name) | |
669 | return -ENOMEM; | |
670 | ||
671 | err = __prealloc_shrinker(shrinker); | |
14773bfa | 672 | if (err) { |
e33c267a | 673 | kfree_const(shrinker->name); |
14773bfa TH |
674 | shrinker->name = NULL; |
675 | } | |
e33c267a RG |
676 | |
677 | return err; | |
678 | } | |
679 | #else | |
680 | int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...) | |
681 | { | |
682 | return __prealloc_shrinker(shrinker); | |
683 | } | |
684 | #endif | |
685 | ||
8e04944f TH |
686 | void free_prealloced_shrinker(struct shrinker *shrinker) |
687 | { | |
e33c267a RG |
688 | #ifdef CONFIG_SHRINKER_DEBUG |
689 | kfree_const(shrinker->name); | |
14773bfa | 690 | shrinker->name = NULL; |
e33c267a | 691 | #endif |
41ca668a YS |
692 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) { |
693 | down_write(&shrinker_rwsem); | |
b4c2b231 | 694 | unregister_memcg_shrinker(shrinker); |
41ca668a | 695 | up_write(&shrinker_rwsem); |
476b30a0 | 696 | return; |
41ca668a | 697 | } |
b4c2b231 | 698 | |
8e04944f TH |
699 | kfree(shrinker->nr_deferred); |
700 | shrinker->nr_deferred = NULL; | |
701 | } | |
1d3d4437 | 702 | |
8e04944f TH |
703 | void register_shrinker_prepared(struct shrinker *shrinker) |
704 | { | |
8e1f936b RR |
705 | down_write(&shrinker_rwsem); |
706 | list_add_tail(&shrinker->list, &shrinker_list); | |
41ca668a | 707 | shrinker->flags |= SHRINKER_REGISTERED; |
5035ebc6 | 708 | shrinker_debugfs_add(shrinker); |
8e1f936b | 709 | up_write(&shrinker_rwsem); |
8e04944f TH |
710 | } |
711 | ||
e33c267a | 712 | static int __register_shrinker(struct shrinker *shrinker) |
8e04944f | 713 | { |
e33c267a | 714 | int err = __prealloc_shrinker(shrinker); |
8e04944f TH |
715 | |
716 | if (err) | |
717 | return err; | |
718 | register_shrinker_prepared(shrinker); | |
1d3d4437 | 719 | return 0; |
1da177e4 | 720 | } |
e33c267a RG |
721 | |
722 | #ifdef CONFIG_SHRINKER_DEBUG | |
723 | int register_shrinker(struct shrinker *shrinker, const char *fmt, ...) | |
724 | { | |
725 | va_list ap; | |
726 | int err; | |
727 | ||
728 | va_start(ap, fmt); | |
729 | shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap); | |
730 | va_end(ap); | |
731 | if (!shrinker->name) | |
732 | return -ENOMEM; | |
733 | ||
734 | err = __register_shrinker(shrinker); | |
14773bfa | 735 | if (err) { |
e33c267a | 736 | kfree_const(shrinker->name); |
14773bfa TH |
737 | shrinker->name = NULL; |
738 | } | |
e33c267a RG |
739 | return err; |
740 | } | |
741 | #else | |
742 | int register_shrinker(struct shrinker *shrinker, const char *fmt, ...) | |
743 | { | |
744 | return __register_shrinker(shrinker); | |
745 | } | |
746 | #endif | |
8e1f936b | 747 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
748 | |
749 | /* | |
750 | * Remove one | |
751 | */ | |
8e1f936b | 752 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 | 753 | { |
41ca668a | 754 | if (!(shrinker->flags & SHRINKER_REGISTERED)) |
bb422a73 | 755 | return; |
41ca668a | 756 | |
1da177e4 LT |
757 | down_write(&shrinker_rwsem); |
758 | list_del(&shrinker->list); | |
41ca668a YS |
759 | shrinker->flags &= ~SHRINKER_REGISTERED; |
760 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) | |
761 | unregister_memcg_shrinker(shrinker); | |
5035ebc6 | 762 | shrinker_debugfs_remove(shrinker); |
1da177e4 | 763 | up_write(&shrinker_rwsem); |
41ca668a | 764 | |
ae393321 | 765 | kfree(shrinker->nr_deferred); |
bb422a73 | 766 | shrinker->nr_deferred = NULL; |
1da177e4 | 767 | } |
8e1f936b | 768 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 | 769 | |
880121be CK |
770 | /** |
771 | * synchronize_shrinkers - Wait for all running shrinkers to complete. | |
772 | * | |
773 | * This is equivalent to calling unregister_shrink() and register_shrinker(), | |
774 | * but atomically and with less overhead. This is useful to guarantee that all | |
775 | * shrinker invocations have seen an update, before freeing memory, similar to | |
776 | * rcu. | |
777 | */ | |
778 | void synchronize_shrinkers(void) | |
779 | { | |
780 | down_write(&shrinker_rwsem); | |
781 | up_write(&shrinker_rwsem); | |
782 | } | |
783 | EXPORT_SYMBOL(synchronize_shrinkers); | |
784 | ||
1da177e4 | 785 | #define SHRINK_BATCH 128 |
1d3d4437 | 786 | |
cb731d6c | 787 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
9092c71b | 788 | struct shrinker *shrinker, int priority) |
1d3d4437 GC |
789 | { |
790 | unsigned long freed = 0; | |
791 | unsigned long long delta; | |
792 | long total_scan; | |
d5bc5fd3 | 793 | long freeable; |
1d3d4437 GC |
794 | long nr; |
795 | long new_nr; | |
1d3d4437 GC |
796 | long batch_size = shrinker->batch ? shrinker->batch |
797 | : SHRINK_BATCH; | |
5f33a080 | 798 | long scanned = 0, next_deferred; |
1d3d4437 | 799 | |
d5bc5fd3 | 800 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
9b996468 KT |
801 | if (freeable == 0 || freeable == SHRINK_EMPTY) |
802 | return freeable; | |
1d3d4437 GC |
803 | |
804 | /* | |
805 | * copy the current shrinker scan count into a local variable | |
806 | * and zero it so that other concurrent shrinker invocations | |
807 | * don't also do this scanning work. | |
808 | */ | |
86750830 | 809 | nr = xchg_nr_deferred(shrinker, shrinkctl); |
1d3d4437 | 810 | |
4b85afbd JW |
811 | if (shrinker->seeks) { |
812 | delta = freeable >> priority; | |
813 | delta *= 4; | |
814 | do_div(delta, shrinker->seeks); | |
815 | } else { | |
816 | /* | |
817 | * These objects don't require any IO to create. Trim | |
818 | * them aggressively under memory pressure to keep | |
819 | * them from causing refetches in the IO caches. | |
820 | */ | |
821 | delta = freeable / 2; | |
822 | } | |
172b06c3 | 823 | |
18bb473e | 824 | total_scan = nr >> priority; |
1d3d4437 | 825 | total_scan += delta; |
18bb473e | 826 | total_scan = min(total_scan, (2 * freeable)); |
1d3d4437 GC |
827 | |
828 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
9092c71b | 829 | freeable, delta, total_scan, priority); |
1d3d4437 | 830 | |
0b1fb40a VD |
831 | /* |
832 | * Normally, we should not scan less than batch_size objects in one | |
833 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
834 | * than batch_size objects in total and we are really tight on memory, | |
835 | * we will try to reclaim all available objects, otherwise we can end | |
836 | * up failing allocations although there are plenty of reclaimable | |
837 | * objects spread over several slabs with usage less than the | |
838 | * batch_size. | |
839 | * | |
840 | * We detect the "tight on memory" situations by looking at the total | |
841 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 842 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
843 | * scanning at high prio and therefore should try to reclaim as much as |
844 | * possible. | |
845 | */ | |
846 | while (total_scan >= batch_size || | |
d5bc5fd3 | 847 | total_scan >= freeable) { |
a0b02131 | 848 | unsigned long ret; |
0b1fb40a | 849 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 850 | |
0b1fb40a | 851 | shrinkctl->nr_to_scan = nr_to_scan; |
d460acb5 | 852 | shrinkctl->nr_scanned = nr_to_scan; |
a0b02131 DC |
853 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
854 | if (ret == SHRINK_STOP) | |
855 | break; | |
856 | freed += ret; | |
1d3d4437 | 857 | |
d460acb5 CW |
858 | count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); |
859 | total_scan -= shrinkctl->nr_scanned; | |
860 | scanned += shrinkctl->nr_scanned; | |
1d3d4437 GC |
861 | |
862 | cond_resched(); | |
863 | } | |
864 | ||
18bb473e YS |
865 | /* |
866 | * The deferred work is increased by any new work (delta) that wasn't | |
867 | * done, decreased by old deferred work that was done now. | |
868 | * | |
869 | * And it is capped to two times of the freeable items. | |
870 | */ | |
871 | next_deferred = max_t(long, (nr + delta - scanned), 0); | |
872 | next_deferred = min(next_deferred, (2 * freeable)); | |
873 | ||
1d3d4437 GC |
874 | /* |
875 | * move the unused scan count back into the shrinker in a | |
86750830 | 876 | * manner that handles concurrent updates. |
1d3d4437 | 877 | */ |
86750830 | 878 | new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); |
1d3d4437 | 879 | |
8efb4b59 | 880 | trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 881 | return freed; |
1495f230 YH |
882 | } |
883 | ||
0a432dcb | 884 | #ifdef CONFIG_MEMCG |
b0dedc49 KT |
885 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
886 | struct mem_cgroup *memcg, int priority) | |
887 | { | |
e4262c4f | 888 | struct shrinker_info *info; |
b8e57efa KT |
889 | unsigned long ret, freed = 0; |
890 | int i; | |
b0dedc49 | 891 | |
0a432dcb | 892 | if (!mem_cgroup_online(memcg)) |
b0dedc49 KT |
893 | return 0; |
894 | ||
895 | if (!down_read_trylock(&shrinker_rwsem)) | |
896 | return 0; | |
897 | ||
468ab843 | 898 | info = shrinker_info_protected(memcg, nid); |
e4262c4f | 899 | if (unlikely(!info)) |
b0dedc49 KT |
900 | goto unlock; |
901 | ||
e4262c4f | 902 | for_each_set_bit(i, info->map, shrinker_nr_max) { |
b0dedc49 KT |
903 | struct shrink_control sc = { |
904 | .gfp_mask = gfp_mask, | |
905 | .nid = nid, | |
906 | .memcg = memcg, | |
907 | }; | |
908 | struct shrinker *shrinker; | |
909 | ||
910 | shrinker = idr_find(&shrinker_idr, i); | |
41ca668a | 911 | if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) { |
7e010df5 | 912 | if (!shrinker) |
e4262c4f | 913 | clear_bit(i, info->map); |
b0dedc49 KT |
914 | continue; |
915 | } | |
916 | ||
0a432dcb YS |
917 | /* Call non-slab shrinkers even though kmem is disabled */ |
918 | if (!memcg_kmem_enabled() && | |
919 | !(shrinker->flags & SHRINKER_NONSLAB)) | |
920 | continue; | |
921 | ||
b0dedc49 | 922 | ret = do_shrink_slab(&sc, shrinker, priority); |
f90280d6 | 923 | if (ret == SHRINK_EMPTY) { |
e4262c4f | 924 | clear_bit(i, info->map); |
f90280d6 KT |
925 | /* |
926 | * After the shrinker reported that it had no objects to | |
927 | * free, but before we cleared the corresponding bit in | |
928 | * the memcg shrinker map, a new object might have been | |
929 | * added. To make sure, we have the bit set in this | |
930 | * case, we invoke the shrinker one more time and reset | |
931 | * the bit if it reports that it is not empty anymore. | |
932 | * The memory barrier here pairs with the barrier in | |
2bfd3637 | 933 | * set_shrinker_bit(): |
f90280d6 KT |
934 | * |
935 | * list_lru_add() shrink_slab_memcg() | |
936 | * list_add_tail() clear_bit() | |
937 | * <MB> <MB> | |
938 | * set_bit() do_shrink_slab() | |
939 | */ | |
940 | smp_mb__after_atomic(); | |
941 | ret = do_shrink_slab(&sc, shrinker, priority); | |
942 | if (ret == SHRINK_EMPTY) | |
943 | ret = 0; | |
944 | else | |
2bfd3637 | 945 | set_shrinker_bit(memcg, nid, i); |
f90280d6 | 946 | } |
b0dedc49 KT |
947 | freed += ret; |
948 | ||
949 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
950 | freed = freed ? : 1; | |
951 | break; | |
952 | } | |
953 | } | |
954 | unlock: | |
955 | up_read(&shrinker_rwsem); | |
956 | return freed; | |
957 | } | |
0a432dcb | 958 | #else /* CONFIG_MEMCG */ |
b0dedc49 KT |
959 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
960 | struct mem_cgroup *memcg, int priority) | |
961 | { | |
962 | return 0; | |
963 | } | |
0a432dcb | 964 | #endif /* CONFIG_MEMCG */ |
b0dedc49 | 965 | |
6b4f7799 | 966 | /** |
cb731d6c | 967 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
968 | * @gfp_mask: allocation context |
969 | * @nid: node whose slab caches to target | |
cb731d6c | 970 | * @memcg: memory cgroup whose slab caches to target |
9092c71b | 971 | * @priority: the reclaim priority |
1da177e4 | 972 | * |
6b4f7799 | 973 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 974 | * |
6b4f7799 JW |
975 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
976 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 977 | * |
aeed1d32 VD |
978 | * @memcg specifies the memory cgroup to target. Unaware shrinkers |
979 | * are called only if it is the root cgroup. | |
cb731d6c | 980 | * |
9092c71b JB |
981 | * @priority is sc->priority, we take the number of objects and >> by priority |
982 | * in order to get the scan target. | |
b15e0905 | 983 | * |
6b4f7799 | 984 | * Returns the number of reclaimed slab objects. |
1da177e4 | 985 | */ |
cb731d6c VD |
986 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
987 | struct mem_cgroup *memcg, | |
9092c71b | 988 | int priority) |
1da177e4 | 989 | { |
b8e57efa | 990 | unsigned long ret, freed = 0; |
1da177e4 LT |
991 | struct shrinker *shrinker; |
992 | ||
fa1e512f YS |
993 | /* |
994 | * The root memcg might be allocated even though memcg is disabled | |
995 | * via "cgroup_disable=memory" boot parameter. This could make | |
996 | * mem_cgroup_is_root() return false, then just run memcg slab | |
997 | * shrink, but skip global shrink. This may result in premature | |
998 | * oom. | |
999 | */ | |
1000 | if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg)) | |
b0dedc49 | 1001 | return shrink_slab_memcg(gfp_mask, nid, memcg, priority); |
cb731d6c | 1002 | |
e830c63a | 1003 | if (!down_read_trylock(&shrinker_rwsem)) |
f06590bd | 1004 | goto out; |
1da177e4 LT |
1005 | |
1006 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
1007 | struct shrink_control sc = { |
1008 | .gfp_mask = gfp_mask, | |
1009 | .nid = nid, | |
cb731d6c | 1010 | .memcg = memcg, |
6b4f7799 | 1011 | }; |
ec97097b | 1012 | |
9b996468 KT |
1013 | ret = do_shrink_slab(&sc, shrinker, priority); |
1014 | if (ret == SHRINK_EMPTY) | |
1015 | ret = 0; | |
1016 | freed += ret; | |
e496612c MK |
1017 | /* |
1018 | * Bail out if someone want to register a new shrinker to | |
55b65a57 | 1019 | * prevent the registration from being stalled for long periods |
e496612c MK |
1020 | * by parallel ongoing shrinking. |
1021 | */ | |
1022 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
1023 | freed = freed ? : 1; | |
1024 | break; | |
1025 | } | |
1da177e4 | 1026 | } |
6b4f7799 | 1027 | |
1da177e4 | 1028 | up_read(&shrinker_rwsem); |
f06590bd MK |
1029 | out: |
1030 | cond_resched(); | |
24f7c6b9 | 1031 | return freed; |
1da177e4 LT |
1032 | } |
1033 | ||
e83b39d6 | 1034 | static unsigned long drop_slab_node(int nid) |
cb731d6c | 1035 | { |
e83b39d6 JK |
1036 | unsigned long freed = 0; |
1037 | struct mem_cgroup *memcg = NULL; | |
cb731d6c | 1038 | |
e83b39d6 | 1039 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
cb731d6c | 1040 | do { |
e83b39d6 JK |
1041 | freed += shrink_slab(GFP_KERNEL, nid, memcg, 0); |
1042 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
069c411d | 1043 | |
e83b39d6 | 1044 | return freed; |
cb731d6c VD |
1045 | } |
1046 | ||
1047 | void drop_slab(void) | |
1048 | { | |
1049 | int nid; | |
e83b39d6 JK |
1050 | int shift = 0; |
1051 | unsigned long freed; | |
1052 | ||
1053 | do { | |
1054 | freed = 0; | |
1055 | for_each_online_node(nid) { | |
1056 | if (fatal_signal_pending(current)) | |
1057 | return; | |
cb731d6c | 1058 | |
e83b39d6 JK |
1059 | freed += drop_slab_node(nid); |
1060 | } | |
1061 | } while ((freed >> shift++) > 1); | |
cb731d6c VD |
1062 | } |
1063 | ||
57e9cc50 JW |
1064 | static int reclaimer_offset(void) |
1065 | { | |
1066 | BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD != | |
1067 | PGDEMOTE_DIRECT - PGDEMOTE_KSWAPD); | |
1068 | BUILD_BUG_ON(PGSTEAL_DIRECT - PGSTEAL_KSWAPD != | |
1069 | PGSCAN_DIRECT - PGSCAN_KSWAPD); | |
1070 | BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD != | |
1071 | PGDEMOTE_KHUGEPAGED - PGDEMOTE_KSWAPD); | |
1072 | BUILD_BUG_ON(PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD != | |
1073 | PGSCAN_KHUGEPAGED - PGSCAN_KSWAPD); | |
1074 | ||
1075 | if (current_is_kswapd()) | |
1076 | return 0; | |
1077 | if (current_is_khugepaged()) | |
1078 | return PGSTEAL_KHUGEPAGED - PGSTEAL_KSWAPD; | |
1079 | return PGSTEAL_DIRECT - PGSTEAL_KSWAPD; | |
1080 | } | |
1081 | ||
e0cd5e7f | 1082 | static inline int is_page_cache_freeable(struct folio *folio) |
1da177e4 | 1083 | { |
ceddc3a5 | 1084 | /* |
49fd9b6d MWO |
1085 | * A freeable page cache folio is referenced only by the caller |
1086 | * that isolated the folio, the page cache and optional filesystem | |
1087 | * private data at folio->private. | |
ceddc3a5 | 1088 | */ |
e0cd5e7f MWO |
1089 | return folio_ref_count(folio) - folio_test_private(folio) == |
1090 | 1 + folio_nr_pages(folio); | |
1da177e4 LT |
1091 | } |
1092 | ||
1da177e4 | 1093 | /* |
e0cd5e7f | 1094 | * We detected a synchronous write error writing a folio out. Probably |
1da177e4 LT |
1095 | * -ENOSPC. We need to propagate that into the address_space for a subsequent |
1096 | * fsync(), msync() or close(). | |
1097 | * | |
1098 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
e0cd5e7f MWO |
1099 | * prevents it from being freed up. But we have a ref on the folio and once |
1100 | * that folio is locked, the mapping is pinned. | |
1da177e4 | 1101 | * |
e0cd5e7f | 1102 | * We're allowed to run sleeping folio_lock() here because we know the caller has |
1da177e4 LT |
1103 | * __GFP_FS. |
1104 | */ | |
1105 | static void handle_write_error(struct address_space *mapping, | |
e0cd5e7f | 1106 | struct folio *folio, int error) |
1da177e4 | 1107 | { |
e0cd5e7f MWO |
1108 | folio_lock(folio); |
1109 | if (folio_mapping(folio) == mapping) | |
3e9f45bd | 1110 | mapping_set_error(mapping, error); |
e0cd5e7f | 1111 | folio_unlock(folio); |
1da177e4 LT |
1112 | } |
1113 | ||
1b4e3f26 MG |
1114 | static bool skip_throttle_noprogress(pg_data_t *pgdat) |
1115 | { | |
1116 | int reclaimable = 0, write_pending = 0; | |
1117 | int i; | |
1118 | ||
1119 | /* | |
1120 | * If kswapd is disabled, reschedule if necessary but do not | |
1121 | * throttle as the system is likely near OOM. | |
1122 | */ | |
1123 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
1124 | return true; | |
1125 | ||
1126 | /* | |
49fd9b6d MWO |
1127 | * If there are a lot of dirty/writeback folios then do not |
1128 | * throttle as throttling will occur when the folios cycle | |
1b4e3f26 MG |
1129 | * towards the end of the LRU if still under writeback. |
1130 | */ | |
1131 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
1132 | struct zone *zone = pgdat->node_zones + i; | |
1133 | ||
36c26128 | 1134 | if (!managed_zone(zone)) |
1b4e3f26 MG |
1135 | continue; |
1136 | ||
1137 | reclaimable += zone_reclaimable_pages(zone); | |
1138 | write_pending += zone_page_state_snapshot(zone, | |
1139 | NR_ZONE_WRITE_PENDING); | |
1140 | } | |
1141 | if (2 * write_pending <= reclaimable) | |
1142 | return true; | |
1143 | ||
1144 | return false; | |
1145 | } | |
1146 | ||
c3f4a9a2 | 1147 | void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason) |
8cd7c588 MG |
1148 | { |
1149 | wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason]; | |
c3f4a9a2 | 1150 | long timeout, ret; |
8cd7c588 MG |
1151 | DEFINE_WAIT(wait); |
1152 | ||
1153 | /* | |
1154 | * Do not throttle IO workers, kthreads other than kswapd or | |
1155 | * workqueues. They may be required for reclaim to make | |
1156 | * forward progress (e.g. journalling workqueues or kthreads). | |
1157 | */ | |
1158 | if (!current_is_kswapd() && | |
b485c6f1 MG |
1159 | current->flags & (PF_IO_WORKER|PF_KTHREAD)) { |
1160 | cond_resched(); | |
8cd7c588 | 1161 | return; |
b485c6f1 | 1162 | } |
8cd7c588 | 1163 | |
c3f4a9a2 MG |
1164 | /* |
1165 | * These figures are pulled out of thin air. | |
1166 | * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many | |
1167 | * parallel reclaimers which is a short-lived event so the timeout is | |
1168 | * short. Failing to make progress or waiting on writeback are | |
1169 | * potentially long-lived events so use a longer timeout. This is shaky | |
1170 | * logic as a failure to make progress could be due to anything from | |
49fd9b6d | 1171 | * writeback to a slow device to excessive referenced folios at the tail |
c3f4a9a2 MG |
1172 | * of the inactive LRU. |
1173 | */ | |
1174 | switch(reason) { | |
1175 | case VMSCAN_THROTTLE_WRITEBACK: | |
1176 | timeout = HZ/10; | |
1177 | ||
1178 | if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) { | |
1179 | WRITE_ONCE(pgdat->nr_reclaim_start, | |
1180 | node_page_state(pgdat, NR_THROTTLED_WRITTEN)); | |
1181 | } | |
1182 | ||
1183 | break; | |
1b4e3f26 MG |
1184 | case VMSCAN_THROTTLE_CONGESTED: |
1185 | fallthrough; | |
c3f4a9a2 | 1186 | case VMSCAN_THROTTLE_NOPROGRESS: |
1b4e3f26 MG |
1187 | if (skip_throttle_noprogress(pgdat)) { |
1188 | cond_resched(); | |
1189 | return; | |
1190 | } | |
1191 | ||
1192 | timeout = 1; | |
1193 | ||
c3f4a9a2 MG |
1194 | break; |
1195 | case VMSCAN_THROTTLE_ISOLATED: | |
1196 | timeout = HZ/50; | |
1197 | break; | |
1198 | default: | |
1199 | WARN_ON_ONCE(1); | |
1200 | timeout = HZ; | |
1201 | break; | |
8cd7c588 MG |
1202 | } |
1203 | ||
1204 | prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); | |
1205 | ret = schedule_timeout(timeout); | |
1206 | finish_wait(wqh, &wait); | |
d818fca1 | 1207 | |
c3f4a9a2 | 1208 | if (reason == VMSCAN_THROTTLE_WRITEBACK) |
d818fca1 | 1209 | atomic_dec(&pgdat->nr_writeback_throttled); |
8cd7c588 MG |
1210 | |
1211 | trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout), | |
1212 | jiffies_to_usecs(timeout - ret), | |
1213 | reason); | |
1214 | } | |
1215 | ||
1216 | /* | |
49fd9b6d MWO |
1217 | * Account for folios written if tasks are throttled waiting on dirty |
1218 | * folios to clean. If enough folios have been cleaned since throttling | |
8cd7c588 MG |
1219 | * started then wakeup the throttled tasks. |
1220 | */ | |
512b7931 | 1221 | void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio, |
8cd7c588 MG |
1222 | int nr_throttled) |
1223 | { | |
1224 | unsigned long nr_written; | |
1225 | ||
512b7931 | 1226 | node_stat_add_folio(folio, NR_THROTTLED_WRITTEN); |
8cd7c588 MG |
1227 | |
1228 | /* | |
1229 | * This is an inaccurate read as the per-cpu deltas may not | |
1230 | * be synchronised. However, given that the system is | |
1231 | * writeback throttled, it is not worth taking the penalty | |
1232 | * of getting an accurate count. At worst, the throttle | |
1233 | * timeout guarantees forward progress. | |
1234 | */ | |
1235 | nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) - | |
1236 | READ_ONCE(pgdat->nr_reclaim_start); | |
1237 | ||
1238 | if (nr_written > SWAP_CLUSTER_MAX * nr_throttled) | |
1239 | wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]); | |
1240 | } | |
1241 | ||
04e62a29 CL |
1242 | /* possible outcome of pageout() */ |
1243 | typedef enum { | |
49fd9b6d | 1244 | /* failed to write folio out, folio is locked */ |
04e62a29 | 1245 | PAGE_KEEP, |
49fd9b6d | 1246 | /* move folio to the active list, folio is locked */ |
04e62a29 | 1247 | PAGE_ACTIVATE, |
49fd9b6d | 1248 | /* folio has been sent to the disk successfully, folio is unlocked */ |
04e62a29 | 1249 | PAGE_SUCCESS, |
49fd9b6d | 1250 | /* folio is clean and locked */ |
04e62a29 CL |
1251 | PAGE_CLEAN, |
1252 | } pageout_t; | |
1253 | ||
1da177e4 | 1254 | /* |
49fd9b6d | 1255 | * pageout is called by shrink_folio_list() for each dirty folio. |
1742f19f | 1256 | * Calls ->writepage(). |
1da177e4 | 1257 | */ |
2282679f N |
1258 | static pageout_t pageout(struct folio *folio, struct address_space *mapping, |
1259 | struct swap_iocb **plug) | |
1da177e4 LT |
1260 | { |
1261 | /* | |
e0cd5e7f | 1262 | * If the folio is dirty, only perform writeback if that write |
1da177e4 LT |
1263 | * will be non-blocking. To prevent this allocation from being |
1264 | * stalled by pagecache activity. But note that there may be | |
1265 | * stalls if we need to run get_block(). We could test | |
1266 | * PagePrivate for that. | |
1267 | * | |
8174202b | 1268 | * If this process is currently in __generic_file_write_iter() against |
e0cd5e7f | 1269 | * this folio's queue, we can perform writeback even if that |
1da177e4 LT |
1270 | * will block. |
1271 | * | |
e0cd5e7f | 1272 | * If the folio is swapcache, write it back even if that would |
1da177e4 LT |
1273 | * block, for some throttling. This happens by accident, because |
1274 | * swap_backing_dev_info is bust: it doesn't reflect the | |
1275 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 | 1276 | */ |
e0cd5e7f | 1277 | if (!is_page_cache_freeable(folio)) |
1da177e4 LT |
1278 | return PAGE_KEEP; |
1279 | if (!mapping) { | |
1280 | /* | |
e0cd5e7f MWO |
1281 | * Some data journaling orphaned folios can have |
1282 | * folio->mapping == NULL while being dirty with clean buffers. | |
1da177e4 | 1283 | */ |
e0cd5e7f | 1284 | if (folio_test_private(folio)) { |
68189fef | 1285 | if (try_to_free_buffers(folio)) { |
e0cd5e7f MWO |
1286 | folio_clear_dirty(folio); |
1287 | pr_info("%s: orphaned folio\n", __func__); | |
1da177e4 LT |
1288 | return PAGE_CLEAN; |
1289 | } | |
1290 | } | |
1291 | return PAGE_KEEP; | |
1292 | } | |
1293 | if (mapping->a_ops->writepage == NULL) | |
1294 | return PAGE_ACTIVATE; | |
1da177e4 | 1295 | |
e0cd5e7f | 1296 | if (folio_clear_dirty_for_io(folio)) { |
1da177e4 LT |
1297 | int res; |
1298 | struct writeback_control wbc = { | |
1299 | .sync_mode = WB_SYNC_NONE, | |
1300 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
1301 | .range_start = 0, |
1302 | .range_end = LLONG_MAX, | |
1da177e4 | 1303 | .for_reclaim = 1, |
2282679f | 1304 | .swap_plug = plug, |
1da177e4 LT |
1305 | }; |
1306 | ||
e0cd5e7f MWO |
1307 | folio_set_reclaim(folio); |
1308 | res = mapping->a_ops->writepage(&folio->page, &wbc); | |
1da177e4 | 1309 | if (res < 0) |
e0cd5e7f | 1310 | handle_write_error(mapping, folio, res); |
994fc28c | 1311 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
e0cd5e7f | 1312 | folio_clear_reclaim(folio); |
1da177e4 LT |
1313 | return PAGE_ACTIVATE; |
1314 | } | |
c661b078 | 1315 | |
e0cd5e7f | 1316 | if (!folio_test_writeback(folio)) { |
1da177e4 | 1317 | /* synchronous write or broken a_ops? */ |
e0cd5e7f | 1318 | folio_clear_reclaim(folio); |
1da177e4 | 1319 | } |
e0cd5e7f MWO |
1320 | trace_mm_vmscan_write_folio(folio); |
1321 | node_stat_add_folio(folio, NR_VMSCAN_WRITE); | |
1da177e4 LT |
1322 | return PAGE_SUCCESS; |
1323 | } | |
1324 | ||
1325 | return PAGE_CLEAN; | |
1326 | } | |
1327 | ||
a649fd92 | 1328 | /* |
49fd9b6d | 1329 | * Same as remove_mapping, but if the folio is removed from the mapping, it |
e286781d | 1330 | * gets returned with a refcount of 0. |
a649fd92 | 1331 | */ |
be7c07d6 | 1332 | static int __remove_mapping(struct address_space *mapping, struct folio *folio, |
b910718a | 1333 | bool reclaimed, struct mem_cgroup *target_memcg) |
49d2e9cc | 1334 | { |
bd4c82c2 | 1335 | int refcount; |
aae466b0 | 1336 | void *shadow = NULL; |
c4843a75 | 1337 | |
be7c07d6 MWO |
1338 | BUG_ON(!folio_test_locked(folio)); |
1339 | BUG_ON(mapping != folio_mapping(folio)); | |
49d2e9cc | 1340 | |
be7c07d6 | 1341 | if (!folio_test_swapcache(folio)) |
51b8c1fe | 1342 | spin_lock(&mapping->host->i_lock); |
30472509 | 1343 | xa_lock_irq(&mapping->i_pages); |
49d2e9cc | 1344 | /* |
49fd9b6d | 1345 | * The non racy check for a busy folio. |
0fd0e6b0 NP |
1346 | * |
1347 | * Must be careful with the order of the tests. When someone has | |
49fd9b6d MWO |
1348 | * a ref to the folio, it may be possible that they dirty it then |
1349 | * drop the reference. So if the dirty flag is tested before the | |
1350 | * refcount here, then the following race may occur: | |
0fd0e6b0 NP |
1351 | * |
1352 | * get_user_pages(&page); | |
1353 | * [user mapping goes away] | |
1354 | * write_to(page); | |
49fd9b6d MWO |
1355 | * !folio_test_dirty(folio) [good] |
1356 | * folio_set_dirty(folio); | |
1357 | * folio_put(folio); | |
1358 | * !refcount(folio) [good, discard it] | |
0fd0e6b0 NP |
1359 | * |
1360 | * [oops, our write_to data is lost] | |
1361 | * | |
1362 | * Reversing the order of the tests ensures such a situation cannot | |
49fd9b6d MWO |
1363 | * escape unnoticed. The smp_rmb is needed to ensure the folio->flags |
1364 | * load is not satisfied before that of folio->_refcount. | |
0fd0e6b0 | 1365 | * |
49fd9b6d | 1366 | * Note that if the dirty flag is always set via folio_mark_dirty, |
b93b0163 | 1367 | * and thus under the i_pages lock, then this ordering is not required. |
49d2e9cc | 1368 | */ |
be7c07d6 MWO |
1369 | refcount = 1 + folio_nr_pages(folio); |
1370 | if (!folio_ref_freeze(folio, refcount)) | |
49d2e9cc | 1371 | goto cannot_free; |
49fd9b6d | 1372 | /* note: atomic_cmpxchg in folio_ref_freeze provides the smp_rmb */ |
be7c07d6 MWO |
1373 | if (unlikely(folio_test_dirty(folio))) { |
1374 | folio_ref_unfreeze(folio, refcount); | |
49d2e9cc | 1375 | goto cannot_free; |
e286781d | 1376 | } |
49d2e9cc | 1377 | |
be7c07d6 MWO |
1378 | if (folio_test_swapcache(folio)) { |
1379 | swp_entry_t swap = folio_swap_entry(folio); | |
ac35a490 | 1380 | |
aae466b0 | 1381 | if (reclaimed && !mapping_exiting(mapping)) |
8927f647 | 1382 | shadow = workingset_eviction(folio, target_memcg); |
ceff9d33 | 1383 | __delete_from_swap_cache(folio, swap, shadow); |
c449deb2 | 1384 | mem_cgroup_swapout(folio, swap); |
30472509 | 1385 | xa_unlock_irq(&mapping->i_pages); |
4081f744 | 1386 | put_swap_folio(folio, swap); |
e286781d | 1387 | } else { |
d2329aa0 | 1388 | void (*free_folio)(struct folio *); |
6072d13c | 1389 | |
d2329aa0 | 1390 | free_folio = mapping->a_ops->free_folio; |
a528910e JW |
1391 | /* |
1392 | * Remember a shadow entry for reclaimed file cache in | |
1393 | * order to detect refaults, thus thrashing, later on. | |
1394 | * | |
1395 | * But don't store shadows in an address space that is | |
238c3046 | 1396 | * already exiting. This is not just an optimization, |
a528910e JW |
1397 | * inode reclaim needs to empty out the radix tree or |
1398 | * the nodes are lost. Don't plant shadows behind its | |
1399 | * back. | |
f9fe48be RZ |
1400 | * |
1401 | * We also don't store shadows for DAX mappings because the | |
49fd9b6d | 1402 | * only page cache folios found in these are zero pages |
f9fe48be RZ |
1403 | * covering holes, and because we don't want to mix DAX |
1404 | * exceptional entries and shadow exceptional entries in the | |
b93b0163 | 1405 | * same address_space. |
a528910e | 1406 | */ |
be7c07d6 | 1407 | if (reclaimed && folio_is_file_lru(folio) && |
f9fe48be | 1408 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
8927f647 MWO |
1409 | shadow = workingset_eviction(folio, target_memcg); |
1410 | __filemap_remove_folio(folio, shadow); | |
30472509 | 1411 | xa_unlock_irq(&mapping->i_pages); |
51b8c1fe JW |
1412 | if (mapping_shrinkable(mapping)) |
1413 | inode_add_lru(mapping->host); | |
1414 | spin_unlock(&mapping->host->i_lock); | |
6072d13c | 1415 | |
d2329aa0 MWO |
1416 | if (free_folio) |
1417 | free_folio(folio); | |
49d2e9cc CL |
1418 | } |
1419 | ||
49d2e9cc CL |
1420 | return 1; |
1421 | ||
1422 | cannot_free: | |
30472509 | 1423 | xa_unlock_irq(&mapping->i_pages); |
be7c07d6 | 1424 | if (!folio_test_swapcache(folio)) |
51b8c1fe | 1425 | spin_unlock(&mapping->host->i_lock); |
49d2e9cc CL |
1426 | return 0; |
1427 | } | |
1428 | ||
5100da38 MWO |
1429 | /** |
1430 | * remove_mapping() - Attempt to remove a folio from its mapping. | |
1431 | * @mapping: The address space. | |
1432 | * @folio: The folio to remove. | |
1433 | * | |
1434 | * If the folio is dirty, under writeback or if someone else has a ref | |
1435 | * on it, removal will fail. | |
1436 | * Return: The number of pages removed from the mapping. 0 if the folio | |
1437 | * could not be removed. | |
1438 | * Context: The caller should have a single refcount on the folio and | |
1439 | * hold its lock. | |
e286781d | 1440 | */ |
5100da38 | 1441 | long remove_mapping(struct address_space *mapping, struct folio *folio) |
e286781d | 1442 | { |
be7c07d6 | 1443 | if (__remove_mapping(mapping, folio, false, NULL)) { |
e286781d | 1444 | /* |
5100da38 | 1445 | * Unfreezing the refcount with 1 effectively |
e286781d NP |
1446 | * drops the pagecache ref for us without requiring another |
1447 | * atomic operation. | |
1448 | */ | |
be7c07d6 | 1449 | folio_ref_unfreeze(folio, 1); |
5100da38 | 1450 | return folio_nr_pages(folio); |
e286781d NP |
1451 | } |
1452 | return 0; | |
1453 | } | |
1454 | ||
894bc310 | 1455 | /** |
ca6d60f3 MWO |
1456 | * folio_putback_lru - Put previously isolated folio onto appropriate LRU list. |
1457 | * @folio: Folio to be returned to an LRU list. | |
894bc310 | 1458 | * |
ca6d60f3 MWO |
1459 | * Add previously isolated @folio to appropriate LRU list. |
1460 | * The folio may still be unevictable for other reasons. | |
894bc310 | 1461 | * |
ca6d60f3 | 1462 | * Context: lru_lock must not be held, interrupts must be enabled. |
894bc310 | 1463 | */ |
ca6d60f3 | 1464 | void folio_putback_lru(struct folio *folio) |
894bc310 | 1465 | { |
ca6d60f3 MWO |
1466 | folio_add_lru(folio); |
1467 | folio_put(folio); /* drop ref from isolate */ | |
894bc310 LS |
1468 | } |
1469 | ||
49fd9b6d MWO |
1470 | enum folio_references { |
1471 | FOLIOREF_RECLAIM, | |
1472 | FOLIOREF_RECLAIM_CLEAN, | |
1473 | FOLIOREF_KEEP, | |
1474 | FOLIOREF_ACTIVATE, | |
dfc8d636 JW |
1475 | }; |
1476 | ||
49fd9b6d | 1477 | static enum folio_references folio_check_references(struct folio *folio, |
dfc8d636 JW |
1478 | struct scan_control *sc) |
1479 | { | |
d92013d1 | 1480 | int referenced_ptes, referenced_folio; |
dfc8d636 | 1481 | unsigned long vm_flags; |
dfc8d636 | 1482 | |
b3ac0413 MWO |
1483 | referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup, |
1484 | &vm_flags); | |
d92013d1 | 1485 | referenced_folio = folio_test_clear_referenced(folio); |
dfc8d636 | 1486 | |
dfc8d636 | 1487 | /* |
d92013d1 MWO |
1488 | * The supposedly reclaimable folio was found to be in a VM_LOCKED vma. |
1489 | * Let the folio, now marked Mlocked, be moved to the unevictable list. | |
dfc8d636 JW |
1490 | */ |
1491 | if (vm_flags & VM_LOCKED) | |
49fd9b6d | 1492 | return FOLIOREF_ACTIVATE; |
dfc8d636 | 1493 | |
6d4675e6 MK |
1494 | /* rmap lock contention: rotate */ |
1495 | if (referenced_ptes == -1) | |
49fd9b6d | 1496 | return FOLIOREF_KEEP; |
6d4675e6 | 1497 | |
64574746 | 1498 | if (referenced_ptes) { |
64574746 | 1499 | /* |
d92013d1 | 1500 | * All mapped folios start out with page table |
64574746 | 1501 | * references from the instantiating fault, so we need |
9030fb0b | 1502 | * to look twice if a mapped file/anon folio is used more |
64574746 JW |
1503 | * than once. |
1504 | * | |
1505 | * Mark it and spare it for another trip around the | |
1506 | * inactive list. Another page table reference will | |
1507 | * lead to its activation. | |
1508 | * | |
d92013d1 MWO |
1509 | * Note: the mark is set for activated folios as well |
1510 | * so that recently deactivated but used folios are | |
64574746 JW |
1511 | * quickly recovered. |
1512 | */ | |
d92013d1 | 1513 | folio_set_referenced(folio); |
64574746 | 1514 | |
d92013d1 | 1515 | if (referenced_folio || referenced_ptes > 1) |
49fd9b6d | 1516 | return FOLIOREF_ACTIVATE; |
64574746 | 1517 | |
c909e993 | 1518 | /* |
d92013d1 | 1519 | * Activate file-backed executable folios after first usage. |
c909e993 | 1520 | */ |
f19a27e3 | 1521 | if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) |
49fd9b6d | 1522 | return FOLIOREF_ACTIVATE; |
c909e993 | 1523 | |
49fd9b6d | 1524 | return FOLIOREF_KEEP; |
64574746 | 1525 | } |
dfc8d636 | 1526 | |
d92013d1 | 1527 | /* Reclaim if clean, defer dirty folios to writeback */ |
f19a27e3 | 1528 | if (referenced_folio && folio_is_file_lru(folio)) |
49fd9b6d | 1529 | return FOLIOREF_RECLAIM_CLEAN; |
64574746 | 1530 | |
49fd9b6d | 1531 | return FOLIOREF_RECLAIM; |
dfc8d636 JW |
1532 | } |
1533 | ||
49fd9b6d | 1534 | /* Check if a folio is dirty or under writeback */ |
e20c41b1 | 1535 | static void folio_check_dirty_writeback(struct folio *folio, |
e2be15f6 MG |
1536 | bool *dirty, bool *writeback) |
1537 | { | |
b4597226 MG |
1538 | struct address_space *mapping; |
1539 | ||
e2be15f6 | 1540 | /* |
49fd9b6d | 1541 | * Anonymous folios are not handled by flushers and must be written |
32a331a7 | 1542 | * from reclaim context. Do not stall reclaim based on them. |
49fd9b6d | 1543 | * MADV_FREE anonymous folios are put into inactive file list too. |
32a331a7 ML |
1544 | * They could be mistakenly treated as file lru. So further anon |
1545 | * test is needed. | |
e2be15f6 | 1546 | */ |
e20c41b1 MWO |
1547 | if (!folio_is_file_lru(folio) || |
1548 | (folio_test_anon(folio) && !folio_test_swapbacked(folio))) { | |
e2be15f6 MG |
1549 | *dirty = false; |
1550 | *writeback = false; | |
1551 | return; | |
1552 | } | |
1553 | ||
e20c41b1 MWO |
1554 | /* By default assume that the folio flags are accurate */ |
1555 | *dirty = folio_test_dirty(folio); | |
1556 | *writeback = folio_test_writeback(folio); | |
b4597226 MG |
1557 | |
1558 | /* Verify dirty/writeback state if the filesystem supports it */ | |
e20c41b1 | 1559 | if (!folio_test_private(folio)) |
b4597226 MG |
1560 | return; |
1561 | ||
e20c41b1 | 1562 | mapping = folio_mapping(folio); |
b4597226 | 1563 | if (mapping && mapping->a_ops->is_dirty_writeback) |
520f301c | 1564 | mapping->a_ops->is_dirty_writeback(folio, dirty, writeback); |
e2be15f6 MG |
1565 | } |
1566 | ||
32008027 | 1567 | static struct page *alloc_demote_page(struct page *page, unsigned long private) |
26aa2d19 | 1568 | { |
32008027 JG |
1569 | struct page *target_page; |
1570 | nodemask_t *allowed_mask; | |
1571 | struct migration_target_control *mtc; | |
1572 | ||
1573 | mtc = (struct migration_target_control *)private; | |
1574 | ||
1575 | allowed_mask = mtc->nmask; | |
1576 | /* | |
1577 | * make sure we allocate from the target node first also trying to | |
1578 | * demote or reclaim pages from the target node via kswapd if we are | |
1579 | * low on free memory on target node. If we don't do this and if | |
1580 | * we have free memory on the slower(lower) memtier, we would start | |
1581 | * allocating pages from slower(lower) memory tiers without even forcing | |
1582 | * a demotion of cold pages from the target memtier. This can result | |
1583 | * in the kernel placing hot pages in slower(lower) memory tiers. | |
1584 | */ | |
1585 | mtc->nmask = NULL; | |
1586 | mtc->gfp_mask |= __GFP_THISNODE; | |
1587 | target_page = alloc_migration_target(page, (unsigned long)mtc); | |
1588 | if (target_page) | |
1589 | return target_page; | |
26aa2d19 | 1590 | |
32008027 JG |
1591 | mtc->gfp_mask &= ~__GFP_THISNODE; |
1592 | mtc->nmask = allowed_mask; | |
1593 | ||
1594 | return alloc_migration_target(page, (unsigned long)mtc); | |
26aa2d19 DH |
1595 | } |
1596 | ||
1597 | /* | |
49fd9b6d MWO |
1598 | * Take folios on @demote_folios and attempt to demote them to another node. |
1599 | * Folios which are not demoted are left on @demote_folios. | |
26aa2d19 | 1600 | */ |
49fd9b6d | 1601 | static unsigned int demote_folio_list(struct list_head *demote_folios, |
26aa2d19 DH |
1602 | struct pglist_data *pgdat) |
1603 | { | |
1604 | int target_nid = next_demotion_node(pgdat->node_id); | |
1605 | unsigned int nr_succeeded; | |
32008027 JG |
1606 | nodemask_t allowed_mask; |
1607 | ||
1608 | struct migration_target_control mtc = { | |
1609 | /* | |
1610 | * Allocate from 'node', or fail quickly and quietly. | |
1611 | * When this happens, 'page' will likely just be discarded | |
1612 | * instead of migrated. | |
1613 | */ | |
1614 | .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | __GFP_NOWARN | | |
1615 | __GFP_NOMEMALLOC | GFP_NOWAIT, | |
1616 | .nid = target_nid, | |
1617 | .nmask = &allowed_mask | |
1618 | }; | |
26aa2d19 | 1619 | |
49fd9b6d | 1620 | if (list_empty(demote_folios)) |
26aa2d19 DH |
1621 | return 0; |
1622 | ||
1623 | if (target_nid == NUMA_NO_NODE) | |
1624 | return 0; | |
1625 | ||
32008027 JG |
1626 | node_get_allowed_targets(pgdat, &allowed_mask); |
1627 | ||
26aa2d19 | 1628 | /* Demotion ignores all cpuset and mempolicy settings */ |
49fd9b6d | 1629 | migrate_pages(demote_folios, alloc_demote_page, NULL, |
32008027 JG |
1630 | (unsigned long)&mtc, MIGRATE_ASYNC, MR_DEMOTION, |
1631 | &nr_succeeded); | |
26aa2d19 | 1632 | |
57e9cc50 | 1633 | __count_vm_events(PGDEMOTE_KSWAPD + reclaimer_offset(), nr_succeeded); |
668e4147 | 1634 | |
26aa2d19 DH |
1635 | return nr_succeeded; |
1636 | } | |
1637 | ||
c28a0e96 | 1638 | static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask) |
d791ea67 N |
1639 | { |
1640 | if (gfp_mask & __GFP_FS) | |
1641 | return true; | |
c28a0e96 | 1642 | if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO)) |
d791ea67 N |
1643 | return false; |
1644 | /* | |
1645 | * We can "enter_fs" for swap-cache with only __GFP_IO | |
1646 | * providing this isn't SWP_FS_OPS. | |
1647 | * ->flags can be updated non-atomicially (scan_swap_map_slots), | |
1648 | * but that will never affect SWP_FS_OPS, so the data_race | |
1649 | * is safe. | |
1650 | */ | |
b98c359f | 1651 | return !data_race(folio_swap_flags(folio) & SWP_FS_OPS); |
d791ea67 N |
1652 | } |
1653 | ||
1da177e4 | 1654 | /* |
49fd9b6d | 1655 | * shrink_folio_list() returns the number of reclaimed pages |
1da177e4 | 1656 | */ |
49fd9b6d MWO |
1657 | static unsigned int shrink_folio_list(struct list_head *folio_list, |
1658 | struct pglist_data *pgdat, struct scan_control *sc, | |
1659 | struct reclaim_stat *stat, bool ignore_references) | |
1660 | { | |
1661 | LIST_HEAD(ret_folios); | |
1662 | LIST_HEAD(free_folios); | |
1663 | LIST_HEAD(demote_folios); | |
730ec8c0 MS |
1664 | unsigned int nr_reclaimed = 0; |
1665 | unsigned int pgactivate = 0; | |
26aa2d19 | 1666 | bool do_demote_pass; |
2282679f | 1667 | struct swap_iocb *plug = NULL; |
1da177e4 | 1668 | |
060f005f | 1669 | memset(stat, 0, sizeof(*stat)); |
1da177e4 | 1670 | cond_resched(); |
26aa2d19 | 1671 | do_demote_pass = can_demote(pgdat->node_id, sc); |
1da177e4 | 1672 | |
26aa2d19 | 1673 | retry: |
49fd9b6d | 1674 | while (!list_empty(folio_list)) { |
1da177e4 | 1675 | struct address_space *mapping; |
be7c07d6 | 1676 | struct folio *folio; |
49fd9b6d | 1677 | enum folio_references references = FOLIOREF_RECLAIM; |
d791ea67 | 1678 | bool dirty, writeback; |
98879b3b | 1679 | unsigned int nr_pages; |
1da177e4 LT |
1680 | |
1681 | cond_resched(); | |
1682 | ||
49fd9b6d | 1683 | folio = lru_to_folio(folio_list); |
be7c07d6 | 1684 | list_del(&folio->lru); |
1da177e4 | 1685 | |
c28a0e96 | 1686 | if (!folio_trylock(folio)) |
1da177e4 LT |
1687 | goto keep; |
1688 | ||
c28a0e96 | 1689 | VM_BUG_ON_FOLIO(folio_test_active(folio), folio); |
1da177e4 | 1690 | |
c28a0e96 | 1691 | nr_pages = folio_nr_pages(folio); |
98879b3b | 1692 | |
c28a0e96 | 1693 | /* Account the number of base pages */ |
98879b3b | 1694 | sc->nr_scanned += nr_pages; |
80e43426 | 1695 | |
c28a0e96 | 1696 | if (unlikely(!folio_evictable(folio))) |
ad6b6704 | 1697 | goto activate_locked; |
894bc310 | 1698 | |
1bee2c16 | 1699 | if (!sc->may_unmap && folio_mapped(folio)) |
80e43426 CL |
1700 | goto keep_locked; |
1701 | ||
018ee47f YZ |
1702 | /* folio_update_gen() tried to promote this page? */ |
1703 | if (lru_gen_enabled() && !ignore_references && | |
1704 | folio_mapped(folio) && folio_test_referenced(folio)) | |
1705 | goto keep_locked; | |
1706 | ||
e2be15f6 | 1707 | /* |
894befec | 1708 | * The number of dirty pages determines if a node is marked |
8cd7c588 | 1709 | * reclaim_congested. kswapd will stall and start writing |
c28a0e96 | 1710 | * folios if the tail of the LRU is all dirty unqueued folios. |
e2be15f6 | 1711 | */ |
e20c41b1 | 1712 | folio_check_dirty_writeback(folio, &dirty, &writeback); |
e2be15f6 | 1713 | if (dirty || writeback) |
c79b7b96 | 1714 | stat->nr_dirty += nr_pages; |
e2be15f6 MG |
1715 | |
1716 | if (dirty && !writeback) | |
c79b7b96 | 1717 | stat->nr_unqueued_dirty += nr_pages; |
e2be15f6 | 1718 | |
d04e8acd | 1719 | /* |
c28a0e96 MWO |
1720 | * Treat this folio as congested if folios are cycling |
1721 | * through the LRU so quickly that the folios marked | |
1722 | * for immediate reclaim are making it to the end of | |
1723 | * the LRU a second time. | |
d04e8acd | 1724 | */ |
c28a0e96 | 1725 | if (writeback && folio_test_reclaim(folio)) |
c79b7b96 | 1726 | stat->nr_congested += nr_pages; |
e2be15f6 | 1727 | |
283aba9f | 1728 | /* |
d33e4e14 | 1729 | * If a folio at the tail of the LRU is under writeback, there |
283aba9f MG |
1730 | * are three cases to consider. |
1731 | * | |
c28a0e96 MWO |
1732 | * 1) If reclaim is encountering an excessive number |
1733 | * of folios under writeback and this folio has both | |
1734 | * the writeback and reclaim flags set, then it | |
d33e4e14 MWO |
1735 | * indicates that folios are being queued for I/O but |
1736 | * are being recycled through the LRU before the I/O | |
1737 | * can complete. Waiting on the folio itself risks an | |
1738 | * indefinite stall if it is impossible to writeback | |
1739 | * the folio due to I/O error or disconnected storage | |
1740 | * so instead note that the LRU is being scanned too | |
1741 | * quickly and the caller can stall after the folio | |
1742 | * list has been processed. | |
283aba9f | 1743 | * |
d33e4e14 | 1744 | * 2) Global or new memcg reclaim encounters a folio that is |
ecf5fc6e MH |
1745 | * not marked for immediate reclaim, or the caller does not |
1746 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
d33e4e14 | 1747 | * not to fs). In this case mark the folio for immediate |
97c9341f | 1748 | * reclaim and continue scanning. |
283aba9f | 1749 | * |
d791ea67 | 1750 | * Require may_enter_fs() because we would wait on fs, which |
d33e4e14 MWO |
1751 | * may not have submitted I/O yet. And the loop driver might |
1752 | * enter reclaim, and deadlock if it waits on a folio for | |
283aba9f MG |
1753 | * which it is needed to do the write (loop masks off |
1754 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1755 | * would probably show more reasons. | |
1756 | * | |
d33e4e14 MWO |
1757 | * 3) Legacy memcg encounters a folio that already has the |
1758 | * reclaim flag set. memcg does not have any dirty folio | |
283aba9f | 1759 | * throttling so we could easily OOM just because too many |
d33e4e14 | 1760 | * folios are in writeback and there is nothing else to |
283aba9f | 1761 | * reclaim. Wait for the writeback to complete. |
c55e8d03 | 1762 | * |
d33e4e14 MWO |
1763 | * In cases 1) and 2) we activate the folios to get them out of |
1764 | * the way while we continue scanning for clean folios on the | |
c55e8d03 JW |
1765 | * inactive list and refilling from the active list. The |
1766 | * observation here is that waiting for disk writes is more | |
1767 | * expensive than potentially causing reloads down the line. | |
1768 | * Since they're marked for immediate reclaim, they won't put | |
1769 | * memory pressure on the cache working set any longer than it | |
1770 | * takes to write them to disk. | |
283aba9f | 1771 | */ |
d33e4e14 | 1772 | if (folio_test_writeback(folio)) { |
283aba9f MG |
1773 | /* Case 1 above */ |
1774 | if (current_is_kswapd() && | |
d33e4e14 | 1775 | folio_test_reclaim(folio) && |
599d0c95 | 1776 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
c79b7b96 | 1777 | stat->nr_immediate += nr_pages; |
c55e8d03 | 1778 | goto activate_locked; |
283aba9f MG |
1779 | |
1780 | /* Case 2 above */ | |
b5ead35e | 1781 | } else if (writeback_throttling_sane(sc) || |
d33e4e14 | 1782 | !folio_test_reclaim(folio) || |
c28a0e96 | 1783 | !may_enter_fs(folio, sc->gfp_mask)) { |
c3b94f44 | 1784 | /* |
d33e4e14 | 1785 | * This is slightly racy - |
c28a0e96 MWO |
1786 | * folio_end_writeback() might have |
1787 | * just cleared the reclaim flag, then | |
1788 | * setting the reclaim flag here ends up | |
1789 | * interpreted as the readahead flag - but | |
1790 | * that does not matter enough to care. | |
1791 | * What we do want is for this folio to | |
1792 | * have the reclaim flag set next time | |
1793 | * memcg reclaim reaches the tests above, | |
1794 | * so it will then wait for writeback to | |
1795 | * avoid OOM; and it's also appropriate | |
d33e4e14 | 1796 | * in global reclaim. |
c3b94f44 | 1797 | */ |
d33e4e14 | 1798 | folio_set_reclaim(folio); |
c79b7b96 | 1799 | stat->nr_writeback += nr_pages; |
c55e8d03 | 1800 | goto activate_locked; |
283aba9f MG |
1801 | |
1802 | /* Case 3 above */ | |
1803 | } else { | |
d33e4e14 MWO |
1804 | folio_unlock(folio); |
1805 | folio_wait_writeback(folio); | |
1806 | /* then go back and try same folio again */ | |
49fd9b6d | 1807 | list_add_tail(&folio->lru, folio_list); |
7fadc820 | 1808 | continue; |
e62e384e | 1809 | } |
c661b078 | 1810 | } |
1da177e4 | 1811 | |
8940b34a | 1812 | if (!ignore_references) |
d92013d1 | 1813 | references = folio_check_references(folio, sc); |
02c6de8d | 1814 | |
dfc8d636 | 1815 | switch (references) { |
49fd9b6d | 1816 | case FOLIOREF_ACTIVATE: |
1da177e4 | 1817 | goto activate_locked; |
49fd9b6d | 1818 | case FOLIOREF_KEEP: |
98879b3b | 1819 | stat->nr_ref_keep += nr_pages; |
64574746 | 1820 | goto keep_locked; |
49fd9b6d MWO |
1821 | case FOLIOREF_RECLAIM: |
1822 | case FOLIOREF_RECLAIM_CLEAN: | |
c28a0e96 | 1823 | ; /* try to reclaim the folio below */ |
dfc8d636 | 1824 | } |
1da177e4 | 1825 | |
26aa2d19 | 1826 | /* |
c28a0e96 | 1827 | * Before reclaiming the folio, try to relocate |
26aa2d19 DH |
1828 | * its contents to another node. |
1829 | */ | |
1830 | if (do_demote_pass && | |
c28a0e96 | 1831 | (thp_migration_supported() || !folio_test_large(folio))) { |
49fd9b6d | 1832 | list_add(&folio->lru, &demote_folios); |
c28a0e96 | 1833 | folio_unlock(folio); |
26aa2d19 DH |
1834 | continue; |
1835 | } | |
1836 | ||
1da177e4 LT |
1837 | /* |
1838 | * Anonymous process memory has backing store? | |
1839 | * Try to allocate it some swap space here. | |
c28a0e96 | 1840 | * Lazyfree folio could be freed directly |
1da177e4 | 1841 | */ |
c28a0e96 MWO |
1842 | if (folio_test_anon(folio) && folio_test_swapbacked(folio)) { |
1843 | if (!folio_test_swapcache(folio)) { | |
bd4c82c2 HY |
1844 | if (!(sc->gfp_mask & __GFP_IO)) |
1845 | goto keep_locked; | |
d4b4084a | 1846 | if (folio_maybe_dma_pinned(folio)) |
feb889fb | 1847 | goto keep_locked; |
c28a0e96 MWO |
1848 | if (folio_test_large(folio)) { |
1849 | /* cannot split folio, skip it */ | |
d4b4084a | 1850 | if (!can_split_folio(folio, NULL)) |
bd4c82c2 HY |
1851 | goto activate_locked; |
1852 | /* | |
c28a0e96 | 1853 | * Split folios without a PMD map right |
bd4c82c2 HY |
1854 | * away. Chances are some or all of the |
1855 | * tail pages can be freed without IO. | |
1856 | */ | |
d4b4084a | 1857 | if (!folio_entire_mapcount(folio) && |
346cf613 | 1858 | split_folio_to_list(folio, |
49fd9b6d | 1859 | folio_list)) |
bd4c82c2 HY |
1860 | goto activate_locked; |
1861 | } | |
09c02e56 MWO |
1862 | if (!add_to_swap(folio)) { |
1863 | if (!folio_test_large(folio)) | |
98879b3b | 1864 | goto activate_locked_split; |
bd4c82c2 | 1865 | /* Fallback to swap normal pages */ |
346cf613 | 1866 | if (split_folio_to_list(folio, |
49fd9b6d | 1867 | folio_list)) |
bd4c82c2 | 1868 | goto activate_locked; |
fe490cc0 HY |
1869 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1870 | count_vm_event(THP_SWPOUT_FALLBACK); | |
1871 | #endif | |
09c02e56 | 1872 | if (!add_to_swap(folio)) |
98879b3b | 1873 | goto activate_locked_split; |
bd4c82c2 | 1874 | } |
bd4c82c2 | 1875 | } |
c28a0e96 MWO |
1876 | } else if (folio_test_swapbacked(folio) && |
1877 | folio_test_large(folio)) { | |
1878 | /* Split shmem folio */ | |
49fd9b6d | 1879 | if (split_folio_to_list(folio, folio_list)) |
7751b2da | 1880 | goto keep_locked; |
e2be15f6 | 1881 | } |
1da177e4 | 1882 | |
98879b3b | 1883 | /* |
c28a0e96 MWO |
1884 | * If the folio was split above, the tail pages will make |
1885 | * their own pass through this function and be accounted | |
1886 | * then. | |
98879b3b | 1887 | */ |
c28a0e96 | 1888 | if ((nr_pages > 1) && !folio_test_large(folio)) { |
98879b3b YS |
1889 | sc->nr_scanned -= (nr_pages - 1); |
1890 | nr_pages = 1; | |
1891 | } | |
1892 | ||
1da177e4 | 1893 | /* |
1bee2c16 | 1894 | * The folio is mapped into the page tables of one or more |
1da177e4 LT |
1895 | * processes. Try to unmap it here. |
1896 | */ | |
1bee2c16 | 1897 | if (folio_mapped(folio)) { |
013339df | 1898 | enum ttu_flags flags = TTU_BATCH_FLUSH; |
1bee2c16 | 1899 | bool was_swapbacked = folio_test_swapbacked(folio); |
bd4c82c2 | 1900 | |
1bee2c16 | 1901 | if (folio_test_pmd_mappable(folio)) |
bd4c82c2 | 1902 | flags |= TTU_SPLIT_HUGE_PMD; |
1f318a9b | 1903 | |
869f7ee6 | 1904 | try_to_unmap(folio, flags); |
1bee2c16 | 1905 | if (folio_mapped(folio)) { |
98879b3b | 1906 | stat->nr_unmap_fail += nr_pages; |
1bee2c16 MWO |
1907 | if (!was_swapbacked && |
1908 | folio_test_swapbacked(folio)) | |
1f318a9b | 1909 | stat->nr_lazyfree_fail += nr_pages; |
1da177e4 | 1910 | goto activate_locked; |
1da177e4 LT |
1911 | } |
1912 | } | |
1913 | ||
5441d490 | 1914 | mapping = folio_mapping(folio); |
49bd2bf9 | 1915 | if (folio_test_dirty(folio)) { |
ee72886d | 1916 | /* |
49bd2bf9 | 1917 | * Only kswapd can writeback filesystem folios |
4eda4823 | 1918 | * to avoid risk of stack overflow. But avoid |
49bd2bf9 | 1919 | * injecting inefficient single-folio I/O into |
4eda4823 | 1920 | * flusher writeback as much as possible: only |
49bd2bf9 MWO |
1921 | * write folios when we've encountered many |
1922 | * dirty folios, and when we've already scanned | |
1923 | * the rest of the LRU for clean folios and see | |
1924 | * the same dirty folios again (with the reclaim | |
1925 | * flag set). | |
ee72886d | 1926 | */ |
49bd2bf9 MWO |
1927 | if (folio_is_file_lru(folio) && |
1928 | (!current_is_kswapd() || | |
1929 | !folio_test_reclaim(folio) || | |
4eda4823 | 1930 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { |
49ea7eb6 MG |
1931 | /* |
1932 | * Immediately reclaim when written back. | |
5a9e3474 | 1933 | * Similar in principle to folio_deactivate() |
49bd2bf9 | 1934 | * except we already have the folio isolated |
49ea7eb6 MG |
1935 | * and know it's dirty |
1936 | */ | |
49bd2bf9 MWO |
1937 | node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE, |
1938 | nr_pages); | |
1939 | folio_set_reclaim(folio); | |
49ea7eb6 | 1940 | |
c55e8d03 | 1941 | goto activate_locked; |
ee72886d MG |
1942 | } |
1943 | ||
49fd9b6d | 1944 | if (references == FOLIOREF_RECLAIM_CLEAN) |
1da177e4 | 1945 | goto keep_locked; |
c28a0e96 | 1946 | if (!may_enter_fs(folio, sc->gfp_mask)) |
1da177e4 | 1947 | goto keep_locked; |
52a8363e | 1948 | if (!sc->may_writepage) |
1da177e4 LT |
1949 | goto keep_locked; |
1950 | ||
d950c947 | 1951 | /* |
49bd2bf9 MWO |
1952 | * Folio is dirty. Flush the TLB if a writable entry |
1953 | * potentially exists to avoid CPU writes after I/O | |
d950c947 MG |
1954 | * starts and then write it out here. |
1955 | */ | |
1956 | try_to_unmap_flush_dirty(); | |
2282679f | 1957 | switch (pageout(folio, mapping, &plug)) { |
1da177e4 LT |
1958 | case PAGE_KEEP: |
1959 | goto keep_locked; | |
1960 | case PAGE_ACTIVATE: | |
1961 | goto activate_locked; | |
1962 | case PAGE_SUCCESS: | |
c79b7b96 | 1963 | stat->nr_pageout += nr_pages; |
96f8bf4f | 1964 | |
49bd2bf9 | 1965 | if (folio_test_writeback(folio)) |
41ac1999 | 1966 | goto keep; |
49bd2bf9 | 1967 | if (folio_test_dirty(folio)) |
1da177e4 | 1968 | goto keep; |
7d3579e8 | 1969 | |
1da177e4 LT |
1970 | /* |
1971 | * A synchronous write - probably a ramdisk. Go | |
49bd2bf9 | 1972 | * ahead and try to reclaim the folio. |
1da177e4 | 1973 | */ |
49bd2bf9 | 1974 | if (!folio_trylock(folio)) |
1da177e4 | 1975 | goto keep; |
49bd2bf9 MWO |
1976 | if (folio_test_dirty(folio) || |
1977 | folio_test_writeback(folio)) | |
1da177e4 | 1978 | goto keep_locked; |
49bd2bf9 | 1979 | mapping = folio_mapping(folio); |
01359eb2 | 1980 | fallthrough; |
1da177e4 | 1981 | case PAGE_CLEAN: |
49bd2bf9 | 1982 | ; /* try to free the folio below */ |
1da177e4 LT |
1983 | } |
1984 | } | |
1985 | ||
1986 | /* | |
0a36111c MWO |
1987 | * If the folio has buffers, try to free the buffer |
1988 | * mappings associated with this folio. If we succeed | |
1989 | * we try to free the folio as well. | |
1da177e4 | 1990 | * |
0a36111c MWO |
1991 | * We do this even if the folio is dirty. |
1992 | * filemap_release_folio() does not perform I/O, but it | |
1993 | * is possible for a folio to have the dirty flag set, | |
1994 | * but it is actually clean (all its buffers are clean). | |
1995 | * This happens if the buffers were written out directly, | |
1996 | * with submit_bh(). ext3 will do this, as well as | |
1997 | * the blockdev mapping. filemap_release_folio() will | |
1998 | * discover that cleanness and will drop the buffers | |
1999 | * and mark the folio clean - it can be freed. | |
1da177e4 | 2000 | * |
0a36111c MWO |
2001 | * Rarely, folios can have buffers and no ->mapping. |
2002 | * These are the folios which were not successfully | |
2003 | * invalidated in truncate_cleanup_folio(). We try to | |
2004 | * drop those buffers here and if that worked, and the | |
2005 | * folio is no longer mapped into process address space | |
2006 | * (refcount == 1) it can be freed. Otherwise, leave | |
2007 | * the folio on the LRU so it is swappable. | |
1da177e4 | 2008 | */ |
0a36111c MWO |
2009 | if (folio_has_private(folio)) { |
2010 | if (!filemap_release_folio(folio, sc->gfp_mask)) | |
1da177e4 | 2011 | goto activate_locked; |
0a36111c MWO |
2012 | if (!mapping && folio_ref_count(folio) == 1) { |
2013 | folio_unlock(folio); | |
2014 | if (folio_put_testzero(folio)) | |
e286781d NP |
2015 | goto free_it; |
2016 | else { | |
2017 | /* | |
2018 | * rare race with speculative reference. | |
2019 | * the speculative reference will free | |
0a36111c | 2020 | * this folio shortly, so we may |
e286781d NP |
2021 | * increment nr_reclaimed here (and |
2022 | * leave it off the LRU). | |
2023 | */ | |
9aafcffc | 2024 | nr_reclaimed += nr_pages; |
e286781d NP |
2025 | continue; |
2026 | } | |
2027 | } | |
1da177e4 LT |
2028 | } |
2029 | ||
64daa5d8 | 2030 | if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) { |
802a3a92 | 2031 | /* follow __remove_mapping for reference */ |
64daa5d8 | 2032 | if (!folio_ref_freeze(folio, 1)) |
802a3a92 | 2033 | goto keep_locked; |
d17be2d9 | 2034 | /* |
64daa5d8 | 2035 | * The folio has only one reference left, which is |
d17be2d9 | 2036 | * from the isolation. After the caller puts the |
64daa5d8 MWO |
2037 | * folio back on the lru and drops the reference, the |
2038 | * folio will be freed anyway. It doesn't matter | |
2039 | * which lru it goes on. So we don't bother checking | |
2040 | * the dirty flag here. | |
d17be2d9 | 2041 | */ |
64daa5d8 MWO |
2042 | count_vm_events(PGLAZYFREED, nr_pages); |
2043 | count_memcg_folio_events(folio, PGLAZYFREED, nr_pages); | |
be7c07d6 | 2044 | } else if (!mapping || !__remove_mapping(mapping, folio, true, |
b910718a | 2045 | sc->target_mem_cgroup)) |
802a3a92 | 2046 | goto keep_locked; |
9a1ea439 | 2047 | |
c28a0e96 | 2048 | folio_unlock(folio); |
e286781d | 2049 | free_it: |
98879b3b | 2050 | /* |
c28a0e96 MWO |
2051 | * Folio may get swapped out as a whole, need to account |
2052 | * all pages in it. | |
98879b3b YS |
2053 | */ |
2054 | nr_reclaimed += nr_pages; | |
abe4c3b5 MG |
2055 | |
2056 | /* | |
49fd9b6d | 2057 | * Is there need to periodically free_folio_list? It would |
abe4c3b5 MG |
2058 | * appear not as the counts should be low |
2059 | */ | |
c28a0e96 | 2060 | if (unlikely(folio_test_large(folio))) |
5375336c | 2061 | destroy_large_folio(folio); |
7ae88534 | 2062 | else |
49fd9b6d | 2063 | list_add(&folio->lru, &free_folios); |
1da177e4 LT |
2064 | continue; |
2065 | ||
98879b3b YS |
2066 | activate_locked_split: |
2067 | /* | |
2068 | * The tail pages that are failed to add into swap cache | |
2069 | * reach here. Fixup nr_scanned and nr_pages. | |
2070 | */ | |
2071 | if (nr_pages > 1) { | |
2072 | sc->nr_scanned -= (nr_pages - 1); | |
2073 | nr_pages = 1; | |
2074 | } | |
1da177e4 | 2075 | activate_locked: |
68a22394 | 2076 | /* Not a candidate for swapping, so reclaim swap space. */ |
246b6480 | 2077 | if (folio_test_swapcache(folio) && |
9202d527 | 2078 | (mem_cgroup_swap_full(folio) || folio_test_mlocked(folio))) |
bdb0ed54 | 2079 | folio_free_swap(folio); |
246b6480 MWO |
2080 | VM_BUG_ON_FOLIO(folio_test_active(folio), folio); |
2081 | if (!folio_test_mlocked(folio)) { | |
2082 | int type = folio_is_file_lru(folio); | |
2083 | folio_set_active(folio); | |
98879b3b | 2084 | stat->nr_activate[type] += nr_pages; |
246b6480 | 2085 | count_memcg_folio_events(folio, PGACTIVATE, nr_pages); |
ad6b6704 | 2086 | } |
1da177e4 | 2087 | keep_locked: |
c28a0e96 | 2088 | folio_unlock(folio); |
1da177e4 | 2089 | keep: |
49fd9b6d | 2090 | list_add(&folio->lru, &ret_folios); |
c28a0e96 MWO |
2091 | VM_BUG_ON_FOLIO(folio_test_lru(folio) || |
2092 | folio_test_unevictable(folio), folio); | |
1da177e4 | 2093 | } |
49fd9b6d | 2094 | /* 'folio_list' is always empty here */ |
26aa2d19 | 2095 | |
c28a0e96 | 2096 | /* Migrate folios selected for demotion */ |
49fd9b6d MWO |
2097 | nr_reclaimed += demote_folio_list(&demote_folios, pgdat); |
2098 | /* Folios that could not be demoted are still in @demote_folios */ | |
2099 | if (!list_empty(&demote_folios)) { | |
6b426d07 | 2100 | /* Folios which weren't demoted go back on @folio_list */ |
49fd9b6d | 2101 | list_splice_init(&demote_folios, folio_list); |
6b426d07 MA |
2102 | |
2103 | /* | |
2104 | * goto retry to reclaim the undemoted folios in folio_list if | |
2105 | * desired. | |
2106 | * | |
2107 | * Reclaiming directly from top tier nodes is not often desired | |
2108 | * due to it breaking the LRU ordering: in general memory | |
2109 | * should be reclaimed from lower tier nodes and demoted from | |
2110 | * top tier nodes. | |
2111 | * | |
2112 | * However, disabling reclaim from top tier nodes entirely | |
2113 | * would cause ooms in edge scenarios where lower tier memory | |
2114 | * is unreclaimable for whatever reason, eg memory being | |
2115 | * mlocked or too hot to reclaim. We can disable reclaim | |
2116 | * from top tier nodes in proactive reclaim though as that is | |
2117 | * not real memory pressure. | |
2118 | */ | |
2119 | if (!sc->proactive) { | |
2120 | do_demote_pass = false; | |
2121 | goto retry; | |
2122 | } | |
26aa2d19 | 2123 | } |
abe4c3b5 | 2124 | |
98879b3b YS |
2125 | pgactivate = stat->nr_activate[0] + stat->nr_activate[1]; |
2126 | ||
49fd9b6d | 2127 | mem_cgroup_uncharge_list(&free_folios); |
72b252ae | 2128 | try_to_unmap_flush(); |
49fd9b6d | 2129 | free_unref_page_list(&free_folios); |
abe4c3b5 | 2130 | |
49fd9b6d | 2131 | list_splice(&ret_folios, folio_list); |
886cf190 | 2132 | count_vm_events(PGACTIVATE, pgactivate); |
060f005f | 2133 | |
2282679f N |
2134 | if (plug) |
2135 | swap_write_unplug(plug); | |
05ff5137 | 2136 | return nr_reclaimed; |
1da177e4 LT |
2137 | } |
2138 | ||
730ec8c0 | 2139 | unsigned int reclaim_clean_pages_from_list(struct zone *zone, |
49fd9b6d | 2140 | struct list_head *folio_list) |
02c6de8d MK |
2141 | { |
2142 | struct scan_control sc = { | |
2143 | .gfp_mask = GFP_KERNEL, | |
02c6de8d MK |
2144 | .may_unmap = 1, |
2145 | }; | |
1f318a9b | 2146 | struct reclaim_stat stat; |
730ec8c0 | 2147 | unsigned int nr_reclaimed; |
b8cecb93 MWO |
2148 | struct folio *folio, *next; |
2149 | LIST_HEAD(clean_folios); | |
2d2b8d2b | 2150 | unsigned int noreclaim_flag; |
02c6de8d | 2151 | |
b8cecb93 MWO |
2152 | list_for_each_entry_safe(folio, next, folio_list, lru) { |
2153 | if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) && | |
2154 | !folio_test_dirty(folio) && !__folio_test_movable(folio) && | |
2155 | !folio_test_unevictable(folio)) { | |
2156 | folio_clear_active(folio); | |
2157 | list_move(&folio->lru, &clean_folios); | |
02c6de8d MK |
2158 | } |
2159 | } | |
2160 | ||
2d2b8d2b YZ |
2161 | /* |
2162 | * We should be safe here since we are only dealing with file pages and | |
2163 | * we are not kswapd and therefore cannot write dirty file pages. But | |
2164 | * call memalloc_noreclaim_save() anyway, just in case these conditions | |
2165 | * change in the future. | |
2166 | */ | |
2167 | noreclaim_flag = memalloc_noreclaim_save(); | |
49fd9b6d | 2168 | nr_reclaimed = shrink_folio_list(&clean_folios, zone->zone_pgdat, &sc, |
013339df | 2169 | &stat, true); |
2d2b8d2b YZ |
2170 | memalloc_noreclaim_restore(noreclaim_flag); |
2171 | ||
b8cecb93 | 2172 | list_splice(&clean_folios, folio_list); |
2da9f630 NP |
2173 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, |
2174 | -(long)nr_reclaimed); | |
1f318a9b JK |
2175 | /* |
2176 | * Since lazyfree pages are isolated from file LRU from the beginning, | |
2177 | * they will rotate back to anonymous LRU in the end if it failed to | |
2178 | * discard so isolated count will be mismatched. | |
2179 | * Compensate the isolated count for both LRU lists. | |
2180 | */ | |
2181 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON, | |
2182 | stat.nr_lazyfree_fail); | |
2183 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, | |
2da9f630 | 2184 | -(long)stat.nr_lazyfree_fail); |
1f318a9b | 2185 | return nr_reclaimed; |
02c6de8d MK |
2186 | } |
2187 | ||
7ee36a14 MG |
2188 | /* |
2189 | * Update LRU sizes after isolating pages. The LRU size updates must | |
55b65a57 | 2190 | * be complete before mem_cgroup_update_lru_size due to a sanity check. |
7ee36a14 MG |
2191 | */ |
2192 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 2193 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 2194 | { |
7ee36a14 MG |
2195 | int zid; |
2196 | ||
7ee36a14 MG |
2197 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
2198 | if (!nr_zone_taken[zid]) | |
2199 | continue; | |
2200 | ||
a892cb6b | 2201 | update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
b4536f0c MH |
2202 | } |
2203 | ||
7ee36a14 MG |
2204 | } |
2205 | ||
f611fab7 | 2206 | /* |
15b44736 HD |
2207 | * Isolating page from the lruvec to fill in @dst list by nr_to_scan times. |
2208 | * | |
2209 | * lruvec->lru_lock is heavily contended. Some of the functions that | |
1da177e4 LT |
2210 | * shrink the lists perform better by taking out a batch of pages |
2211 | * and working on them outside the LRU lock. | |
2212 | * | |
2213 | * For pagecache intensive workloads, this function is the hottest | |
2214 | * spot in the kernel (apart from copy_*_user functions). | |
2215 | * | |
15b44736 | 2216 | * Lru_lock must be held before calling this function. |
1da177e4 | 2217 | * |
791b48b6 | 2218 | * @nr_to_scan: The number of eligible pages to look through on the list. |
5dc35979 | 2219 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 2220 | * @dst: The temp list to put pages on to. |
f626012d | 2221 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 2222 | * @sc: The scan_control struct for this reclaim session |
3cb99451 | 2223 | * @lru: LRU list id for isolating |
1da177e4 LT |
2224 | * |
2225 | * returns how many pages were moved onto *@dst. | |
2226 | */ | |
49fd9b6d | 2227 | static unsigned long isolate_lru_folios(unsigned long nr_to_scan, |
5dc35979 | 2228 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 2229 | unsigned long *nr_scanned, struct scan_control *sc, |
a9e7c39f | 2230 | enum lru_list lru) |
1da177e4 | 2231 | { |
75b00af7 | 2232 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 2233 | unsigned long nr_taken = 0; |
599d0c95 | 2234 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 2235 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
3db65812 | 2236 | unsigned long skipped = 0; |
791b48b6 | 2237 | unsigned long scan, total_scan, nr_pages; |
166e3d32 | 2238 | LIST_HEAD(folios_skipped); |
1da177e4 | 2239 | |
98879b3b | 2240 | total_scan = 0; |
791b48b6 | 2241 | scan = 0; |
98879b3b | 2242 | while (scan < nr_to_scan && !list_empty(src)) { |
89f6c88a | 2243 | struct list_head *move_to = src; |
166e3d32 | 2244 | struct folio *folio; |
5ad333eb | 2245 | |
166e3d32 MWO |
2246 | folio = lru_to_folio(src); |
2247 | prefetchw_prev_lru_folio(folio, src, flags); | |
1da177e4 | 2248 | |
166e3d32 | 2249 | nr_pages = folio_nr_pages(folio); |
98879b3b YS |
2250 | total_scan += nr_pages; |
2251 | ||
166e3d32 MWO |
2252 | if (folio_zonenum(folio) > sc->reclaim_idx) { |
2253 | nr_skipped[folio_zonenum(folio)] += nr_pages; | |
2254 | move_to = &folios_skipped; | |
89f6c88a | 2255 | goto move; |
b2e18757 MG |
2256 | } |
2257 | ||
791b48b6 | 2258 | /* |
166e3d32 MWO |
2259 | * Do not count skipped folios because that makes the function |
2260 | * return with no isolated folios if the LRU mostly contains | |
2261 | * ineligible folios. This causes the VM to not reclaim any | |
2262 | * folios, triggering a premature OOM. | |
2263 | * Account all pages in a folio. | |
791b48b6 | 2264 | */ |
98879b3b | 2265 | scan += nr_pages; |
89f6c88a | 2266 | |
166e3d32 | 2267 | if (!folio_test_lru(folio)) |
89f6c88a | 2268 | goto move; |
166e3d32 | 2269 | if (!sc->may_unmap && folio_mapped(folio)) |
89f6c88a HD |
2270 | goto move; |
2271 | ||
c2135f7c | 2272 | /* |
166e3d32 MWO |
2273 | * Be careful not to clear the lru flag until after we're |
2274 | * sure the folio is not being freed elsewhere -- the | |
2275 | * folio release code relies on it. | |
c2135f7c | 2276 | */ |
166e3d32 | 2277 | if (unlikely(!folio_try_get(folio))) |
89f6c88a | 2278 | goto move; |
5ad333eb | 2279 | |
166e3d32 MWO |
2280 | if (!folio_test_clear_lru(folio)) { |
2281 | /* Another thread is already isolating this folio */ | |
2282 | folio_put(folio); | |
89f6c88a | 2283 | goto move; |
5ad333eb | 2284 | } |
c2135f7c AS |
2285 | |
2286 | nr_taken += nr_pages; | |
166e3d32 | 2287 | nr_zone_taken[folio_zonenum(folio)] += nr_pages; |
89f6c88a HD |
2288 | move_to = dst; |
2289 | move: | |
166e3d32 | 2290 | list_move(&folio->lru, move_to); |
1da177e4 LT |
2291 | } |
2292 | ||
b2e18757 | 2293 | /* |
166e3d32 | 2294 | * Splice any skipped folios to the start of the LRU list. Note that |
b2e18757 MG |
2295 | * this disrupts the LRU order when reclaiming for lower zones but |
2296 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
166e3d32 | 2297 | * scanning would soon rescan the same folios to skip and waste lots |
b2cb6826 | 2298 | * of cpu cycles. |
b2e18757 | 2299 | */ |
166e3d32 | 2300 | if (!list_empty(&folios_skipped)) { |
7cc30fcf MG |
2301 | int zid; |
2302 | ||
166e3d32 | 2303 | list_splice(&folios_skipped, src); |
7cc30fcf MG |
2304 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
2305 | if (!nr_skipped[zid]) | |
2306 | continue; | |
2307 | ||
2308 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 2309 | skipped += nr_skipped[zid]; |
7cc30fcf MG |
2310 | } |
2311 | } | |
791b48b6 | 2312 | *nr_scanned = total_scan; |
1265e3a6 | 2313 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
89f6c88a HD |
2314 | total_scan, skipped, nr_taken, |
2315 | sc->may_unmap ? 0 : ISOLATE_UNMAPPED, lru); | |
b4536f0c | 2316 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
2317 | return nr_taken; |
2318 | } | |
2319 | ||
62695a84 | 2320 | /** |
d1d8a3b4 MWO |
2321 | * folio_isolate_lru() - Try to isolate a folio from its LRU list. |
2322 | * @folio: Folio to isolate from its LRU list. | |
62695a84 | 2323 | * |
d1d8a3b4 MWO |
2324 | * Isolate a @folio from an LRU list and adjust the vmstat statistic |
2325 | * corresponding to whatever LRU list the folio was on. | |
62695a84 | 2326 | * |
d1d8a3b4 MWO |
2327 | * The folio will have its LRU flag cleared. If it was found on the |
2328 | * active list, it will have the Active flag set. If it was found on the | |
2329 | * unevictable list, it will have the Unevictable flag set. These flags | |
894bc310 | 2330 | * may need to be cleared by the caller before letting the page go. |
62695a84 | 2331 | * |
d1d8a3b4 | 2332 | * Context: |
a5d09bed | 2333 | * |
49fd9b6d MWO |
2334 | * (1) Must be called with an elevated refcount on the folio. This is a |
2335 | * fundamental difference from isolate_lru_folios() (which is called | |
62695a84 | 2336 | * without a stable reference). |
d1d8a3b4 MWO |
2337 | * (2) The lru_lock must not be held. |
2338 | * (3) Interrupts must be enabled. | |
2339 | * | |
2340 | * Return: 0 if the folio was removed from an LRU list. | |
2341 | * -EBUSY if the folio was not on an LRU list. | |
62695a84 | 2342 | */ |
d1d8a3b4 | 2343 | int folio_isolate_lru(struct folio *folio) |
62695a84 NP |
2344 | { |
2345 | int ret = -EBUSY; | |
2346 | ||
d1d8a3b4 | 2347 | VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio); |
0c917313 | 2348 | |
d1d8a3b4 | 2349 | if (folio_test_clear_lru(folio)) { |
fa9add64 | 2350 | struct lruvec *lruvec; |
62695a84 | 2351 | |
d1d8a3b4 | 2352 | folio_get(folio); |
e809c3fe | 2353 | lruvec = folio_lruvec_lock_irq(folio); |
d1d8a3b4 | 2354 | lruvec_del_folio(lruvec, folio); |
6168d0da | 2355 | unlock_page_lruvec_irq(lruvec); |
d25b5bd8 | 2356 | ret = 0; |
62695a84 | 2357 | } |
d25b5bd8 | 2358 | |
62695a84 NP |
2359 | return ret; |
2360 | } | |
2361 | ||
35cd7815 | 2362 | /* |
d37dd5dc | 2363 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
178821b8 | 2364 | * then get rescheduled. When there are massive number of tasks doing page |
d37dd5dc FW |
2365 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, |
2366 | * the LRU list will go small and be scanned faster than necessary, leading to | |
2367 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 2368 | */ |
599d0c95 | 2369 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
2370 | struct scan_control *sc) |
2371 | { | |
2372 | unsigned long inactive, isolated; | |
d818fca1 | 2373 | bool too_many; |
35cd7815 RR |
2374 | |
2375 | if (current_is_kswapd()) | |
2376 | return 0; | |
2377 | ||
b5ead35e | 2378 | if (!writeback_throttling_sane(sc)) |
35cd7815 RR |
2379 | return 0; |
2380 | ||
2381 | if (file) { | |
599d0c95 MG |
2382 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
2383 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 2384 | } else { |
599d0c95 MG |
2385 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
2386 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
2387 | } |
2388 | ||
3cf23841 FW |
2389 | /* |
2390 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
2391 | * won't get blocked by normal direct-reclaimers, forming a circular | |
2392 | * deadlock. | |
2393 | */ | |
d0164adc | 2394 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
2395 | inactive >>= 3; |
2396 | ||
d818fca1 MG |
2397 | too_many = isolated > inactive; |
2398 | ||
2399 | /* Wake up tasks throttled due to too_many_isolated. */ | |
2400 | if (!too_many) | |
2401 | wake_throttle_isolated(pgdat); | |
2402 | ||
2403 | return too_many; | |
35cd7815 RR |
2404 | } |
2405 | ||
a222f341 | 2406 | /* |
49fd9b6d | 2407 | * move_folios_to_lru() moves folios from private @list to appropriate LRU list. |
ff00a170 | 2408 | * On return, @list is reused as a list of folios to be freed by the caller. |
a222f341 KT |
2409 | * |
2410 | * Returns the number of pages moved to the given lruvec. | |
2411 | */ | |
49fd9b6d MWO |
2412 | static unsigned int move_folios_to_lru(struct lruvec *lruvec, |
2413 | struct list_head *list) | |
66635629 | 2414 | { |
a222f341 | 2415 | int nr_pages, nr_moved = 0; |
ff00a170 | 2416 | LIST_HEAD(folios_to_free); |
66635629 | 2417 | |
a222f341 | 2418 | while (!list_empty(list)) { |
ff00a170 MWO |
2419 | struct folio *folio = lru_to_folio(list); |
2420 | ||
2421 | VM_BUG_ON_FOLIO(folio_test_lru(folio), folio); | |
2422 | list_del(&folio->lru); | |
2423 | if (unlikely(!folio_evictable(folio))) { | |
6168d0da | 2424 | spin_unlock_irq(&lruvec->lru_lock); |
ff00a170 | 2425 | folio_putback_lru(folio); |
6168d0da | 2426 | spin_lock_irq(&lruvec->lru_lock); |
66635629 MG |
2427 | continue; |
2428 | } | |
fa9add64 | 2429 | |
3d06afab | 2430 | /* |
ff00a170 | 2431 | * The folio_set_lru needs to be kept here for list integrity. |
3d06afab | 2432 | * Otherwise: |
49fd9b6d | 2433 | * #0 move_folios_to_lru #1 release_pages |
ff00a170 MWO |
2434 | * if (!folio_put_testzero()) |
2435 | * if (folio_put_testzero()) | |
2436 | * !lru //skip lru_lock | |
2437 | * folio_set_lru() | |
2438 | * list_add(&folio->lru,) | |
2439 | * list_add(&folio->lru,) | |
3d06afab | 2440 | */ |
ff00a170 | 2441 | folio_set_lru(folio); |
a222f341 | 2442 | |
ff00a170 MWO |
2443 | if (unlikely(folio_put_testzero(folio))) { |
2444 | __folio_clear_lru_flags(folio); | |
2bcf8879 | 2445 | |
ff00a170 | 2446 | if (unlikely(folio_test_large(folio))) { |
6168d0da | 2447 | spin_unlock_irq(&lruvec->lru_lock); |
5375336c | 2448 | destroy_large_folio(folio); |
6168d0da | 2449 | spin_lock_irq(&lruvec->lru_lock); |
2bcf8879 | 2450 | } else |
ff00a170 | 2451 | list_add(&folio->lru, &folios_to_free); |
3d06afab AS |
2452 | |
2453 | continue; | |
66635629 | 2454 | } |
3d06afab | 2455 | |
afca9157 AS |
2456 | /* |
2457 | * All pages were isolated from the same lruvec (and isolation | |
2458 | * inhibits memcg migration). | |
2459 | */ | |
ff00a170 MWO |
2460 | VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio); |
2461 | lruvec_add_folio(lruvec, folio); | |
2462 | nr_pages = folio_nr_pages(folio); | |
3d06afab | 2463 | nr_moved += nr_pages; |
ff00a170 | 2464 | if (folio_test_active(folio)) |
3d06afab | 2465 | workingset_age_nonresident(lruvec, nr_pages); |
66635629 | 2466 | } |
66635629 | 2467 | |
3f79768f HD |
2468 | /* |
2469 | * To save our caller's stack, now use input list for pages to free. | |
2470 | */ | |
ff00a170 | 2471 | list_splice(&folios_to_free, list); |
a222f341 KT |
2472 | |
2473 | return nr_moved; | |
66635629 MG |
2474 | } |
2475 | ||
399ba0b9 | 2476 | /* |
5829f7db ML |
2477 | * If a kernel thread (such as nfsd for loop-back mounts) services a backing |
2478 | * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case | |
2479 | * we should not throttle. Otherwise it is safe to do so. | |
399ba0b9 N |
2480 | */ |
2481 | static int current_may_throttle(void) | |
2482 | { | |
b9b1335e | 2483 | return !(current->flags & PF_LOCAL_THROTTLE); |
399ba0b9 N |
2484 | } |
2485 | ||
1da177e4 | 2486 | /* |
b2e18757 | 2487 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 2488 | * of reclaimed pages |
1da177e4 | 2489 | */ |
49fd9b6d MWO |
2490 | static unsigned long shrink_inactive_list(unsigned long nr_to_scan, |
2491 | struct lruvec *lruvec, struct scan_control *sc, | |
2492 | enum lru_list lru) | |
1da177e4 | 2493 | { |
49fd9b6d | 2494 | LIST_HEAD(folio_list); |
e247dbce | 2495 | unsigned long nr_scanned; |
730ec8c0 | 2496 | unsigned int nr_reclaimed = 0; |
e247dbce | 2497 | unsigned long nr_taken; |
060f005f | 2498 | struct reclaim_stat stat; |
497a6c1b | 2499 | bool file = is_file_lru(lru); |
f46b7912 | 2500 | enum vm_event_item item; |
599d0c95 | 2501 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
db73ee0d | 2502 | bool stalled = false; |
78dc583d | 2503 | |
599d0c95 | 2504 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
db73ee0d MH |
2505 | if (stalled) |
2506 | return 0; | |
2507 | ||
2508 | /* wait a bit for the reclaimer. */ | |
db73ee0d | 2509 | stalled = true; |
c3f4a9a2 | 2510 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED); |
35cd7815 RR |
2511 | |
2512 | /* We are about to die and free our memory. Return now. */ | |
2513 | if (fatal_signal_pending(current)) | |
2514 | return SWAP_CLUSTER_MAX; | |
2515 | } | |
2516 | ||
1da177e4 | 2517 | lru_add_drain(); |
f80c0673 | 2518 | |
6168d0da | 2519 | spin_lock_irq(&lruvec->lru_lock); |
b35ea17b | 2520 | |
49fd9b6d | 2521 | nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &folio_list, |
a9e7c39f | 2522 | &nr_scanned, sc, lru); |
95d918fc | 2523 | |
599d0c95 | 2524 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
57e9cc50 | 2525 | item = PGSCAN_KSWAPD + reclaimer_offset(); |
b5ead35e | 2526 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
2527 | __count_vm_events(item, nr_scanned); |
2528 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned); | |
497a6c1b JW |
2529 | __count_vm_events(PGSCAN_ANON + file, nr_scanned); |
2530 | ||
6168d0da | 2531 | spin_unlock_irq(&lruvec->lru_lock); |
b35ea17b | 2532 | |
d563c050 | 2533 | if (nr_taken == 0) |
66635629 | 2534 | return 0; |
5ad333eb | 2535 | |
49fd9b6d | 2536 | nr_reclaimed = shrink_folio_list(&folio_list, pgdat, sc, &stat, false); |
c661b078 | 2537 | |
6168d0da | 2538 | spin_lock_irq(&lruvec->lru_lock); |
49fd9b6d | 2539 | move_folios_to_lru(lruvec, &folio_list); |
497a6c1b JW |
2540 | |
2541 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); | |
57e9cc50 | 2542 | item = PGSTEAL_KSWAPD + reclaimer_offset(); |
b5ead35e | 2543 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
2544 | __count_vm_events(item, nr_reclaimed); |
2545 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed); | |
497a6c1b | 2546 | __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed); |
6168d0da | 2547 | spin_unlock_irq(&lruvec->lru_lock); |
3f79768f | 2548 | |
0538a82c | 2549 | lru_note_cost(lruvec, file, stat.nr_pageout, nr_scanned - nr_reclaimed); |
49fd9b6d MWO |
2550 | mem_cgroup_uncharge_list(&folio_list); |
2551 | free_unref_page_list(&folio_list); | |
e11da5b4 | 2552 | |
1c610d5f | 2553 | /* |
49fd9b6d | 2554 | * If dirty folios are scanned that are not queued for IO, it |
1c610d5f | 2555 | * implies that flushers are not doing their job. This can |
49fd9b6d | 2556 | * happen when memory pressure pushes dirty folios to the end of |
1c610d5f AR |
2557 | * the LRU before the dirty limits are breached and the dirty |
2558 | * data has expired. It can also happen when the proportion of | |
49fd9b6d | 2559 | * dirty folios grows not through writes but through memory |
1c610d5f AR |
2560 | * pressure reclaiming all the clean cache. And in some cases, |
2561 | * the flushers simply cannot keep up with the allocation | |
2562 | * rate. Nudge the flusher threads in case they are asleep. | |
2563 | */ | |
81a70c21 | 2564 | if (stat.nr_unqueued_dirty == nr_taken) { |
1c610d5f | 2565 | wakeup_flusher_threads(WB_REASON_VMSCAN); |
81a70c21 AK |
2566 | /* |
2567 | * For cgroupv1 dirty throttling is achieved by waking up | |
2568 | * the kernel flusher here and later waiting on folios | |
2569 | * which are in writeback to finish (see shrink_folio_list()). | |
2570 | * | |
2571 | * Flusher may not be able to issue writeback quickly | |
2572 | * enough for cgroupv1 writeback throttling to work | |
2573 | * on a large system. | |
2574 | */ | |
2575 | if (!writeback_throttling_sane(sc)) | |
2576 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK); | |
2577 | } | |
1c610d5f | 2578 | |
d108c772 AR |
2579 | sc->nr.dirty += stat.nr_dirty; |
2580 | sc->nr.congested += stat.nr_congested; | |
2581 | sc->nr.unqueued_dirty += stat.nr_unqueued_dirty; | |
2582 | sc->nr.writeback += stat.nr_writeback; | |
2583 | sc->nr.immediate += stat.nr_immediate; | |
2584 | sc->nr.taken += nr_taken; | |
2585 | if (file) | |
2586 | sc->nr.file_taken += nr_taken; | |
8e950282 | 2587 | |
599d0c95 | 2588 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
d51d1e64 | 2589 | nr_scanned, nr_reclaimed, &stat, sc->priority, file); |
05ff5137 | 2590 | return nr_reclaimed; |
1da177e4 LT |
2591 | } |
2592 | ||
15b44736 | 2593 | /* |
07f67a8d | 2594 | * shrink_active_list() moves folios from the active LRU to the inactive LRU. |
15b44736 | 2595 | * |
07f67a8d | 2596 | * We move them the other way if the folio is referenced by one or more |
15b44736 HD |
2597 | * processes. |
2598 | * | |
07f67a8d | 2599 | * If the folios are mostly unmapped, the processing is fast and it is |
15b44736 | 2600 | * appropriate to hold lru_lock across the whole operation. But if |
07f67a8d MWO |
2601 | * the folios are mapped, the processing is slow (folio_referenced()), so |
2602 | * we should drop lru_lock around each folio. It's impossible to balance | |
2603 | * this, so instead we remove the folios from the LRU while processing them. | |
2604 | * It is safe to rely on the active flag against the non-LRU folios in here | |
2605 | * because nobody will play with that bit on a non-LRU folio. | |
15b44736 | 2606 | * |
07f67a8d MWO |
2607 | * The downside is that we have to touch folio->_refcount against each folio. |
2608 | * But we had to alter folio->flags anyway. | |
15b44736 | 2609 | */ |
f626012d | 2610 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 2611 | struct lruvec *lruvec, |
f16015fb | 2612 | struct scan_control *sc, |
9e3b2f8c | 2613 | enum lru_list lru) |
1da177e4 | 2614 | { |
44c241f1 | 2615 | unsigned long nr_taken; |
f626012d | 2616 | unsigned long nr_scanned; |
6fe6b7e3 | 2617 | unsigned long vm_flags; |
07f67a8d | 2618 | LIST_HEAD(l_hold); /* The folios which were snipped off */ |
8cab4754 | 2619 | LIST_HEAD(l_active); |
b69408e8 | 2620 | LIST_HEAD(l_inactive); |
9d998b4f MH |
2621 | unsigned nr_deactivate, nr_activate; |
2622 | unsigned nr_rotated = 0; | |
3cb99451 | 2623 | int file = is_file_lru(lru); |
599d0c95 | 2624 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
2625 | |
2626 | lru_add_drain(); | |
f80c0673 | 2627 | |
6168d0da | 2628 | spin_lock_irq(&lruvec->lru_lock); |
925b7673 | 2629 | |
49fd9b6d | 2630 | nr_taken = isolate_lru_folios(nr_to_scan, lruvec, &l_hold, |
a9e7c39f | 2631 | &nr_scanned, sc, lru); |
89b5fae5 | 2632 | |
599d0c95 | 2633 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
1cfb419b | 2634 | |
912c0572 SB |
2635 | if (!cgroup_reclaim(sc)) |
2636 | __count_vm_events(PGREFILL, nr_scanned); | |
2fa2690c | 2637 | __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); |
9d5e6a9f | 2638 | |
6168d0da | 2639 | spin_unlock_irq(&lruvec->lru_lock); |
1da177e4 | 2640 | |
1da177e4 | 2641 | while (!list_empty(&l_hold)) { |
b3ac0413 | 2642 | struct folio *folio; |
b3ac0413 | 2643 | |
1da177e4 | 2644 | cond_resched(); |
b3ac0413 MWO |
2645 | folio = lru_to_folio(&l_hold); |
2646 | list_del(&folio->lru); | |
7e9cd484 | 2647 | |
07f67a8d MWO |
2648 | if (unlikely(!folio_evictable(folio))) { |
2649 | folio_putback_lru(folio); | |
894bc310 LS |
2650 | continue; |
2651 | } | |
2652 | ||
cc715d99 | 2653 | if (unlikely(buffer_heads_over_limit)) { |
36a3b14b MWO |
2654 | if (folio_test_private(folio) && folio_trylock(folio)) { |
2655 | if (folio_test_private(folio)) | |
07f67a8d MWO |
2656 | filemap_release_folio(folio, 0); |
2657 | folio_unlock(folio); | |
cc715d99 MG |
2658 | } |
2659 | } | |
2660 | ||
6d4675e6 | 2661 | /* Referenced or rmap lock contention: rotate */ |
b3ac0413 | 2662 | if (folio_referenced(folio, 0, sc->target_mem_cgroup, |
6d4675e6 | 2663 | &vm_flags) != 0) { |
8cab4754 | 2664 | /* |
07f67a8d | 2665 | * Identify referenced, file-backed active folios and |
8cab4754 WF |
2666 | * give them one more trip around the active list. So |
2667 | * that executable code get better chances to stay in | |
07f67a8d | 2668 | * memory under moderate memory pressure. Anon folios |
8cab4754 | 2669 | * are not likely to be evicted by use-once streaming |
07f67a8d | 2670 | * IO, plus JVM can create lots of anon VM_EXEC folios, |
8cab4754 WF |
2671 | * so we ignore them here. |
2672 | */ | |
07f67a8d MWO |
2673 | if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) { |
2674 | nr_rotated += folio_nr_pages(folio); | |
2675 | list_add(&folio->lru, &l_active); | |
8cab4754 WF |
2676 | continue; |
2677 | } | |
2678 | } | |
7e9cd484 | 2679 | |
07f67a8d MWO |
2680 | folio_clear_active(folio); /* we are de-activating */ |
2681 | folio_set_workingset(folio); | |
2682 | list_add(&folio->lru, &l_inactive); | |
1da177e4 LT |
2683 | } |
2684 | ||
b555749a | 2685 | /* |
07f67a8d | 2686 | * Move folios back to the lru list. |
b555749a | 2687 | */ |
6168d0da | 2688 | spin_lock_irq(&lruvec->lru_lock); |
556adecb | 2689 | |
49fd9b6d MWO |
2690 | nr_activate = move_folios_to_lru(lruvec, &l_active); |
2691 | nr_deactivate = move_folios_to_lru(lruvec, &l_inactive); | |
07f67a8d | 2692 | /* Keep all free folios in l_active list */ |
f372d89e | 2693 | list_splice(&l_inactive, &l_active); |
9851ac13 KT |
2694 | |
2695 | __count_vm_events(PGDEACTIVATE, nr_deactivate); | |
2696 | __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate); | |
2697 | ||
599d0c95 | 2698 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
6168d0da | 2699 | spin_unlock_irq(&lruvec->lru_lock); |
2bcf8879 | 2700 | |
0538a82c JW |
2701 | if (nr_rotated) |
2702 | lru_note_cost(lruvec, file, 0, nr_rotated); | |
f372d89e KT |
2703 | mem_cgroup_uncharge_list(&l_active); |
2704 | free_unref_page_list(&l_active); | |
9d998b4f MH |
2705 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
2706 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
2707 | } |
2708 | ||
49fd9b6d | 2709 | static unsigned int reclaim_folio_list(struct list_head *folio_list, |
1fe47c0b | 2710 | struct pglist_data *pgdat) |
1a4e58cc | 2711 | { |
1a4e58cc | 2712 | struct reclaim_stat dummy_stat; |
1fe47c0b ML |
2713 | unsigned int nr_reclaimed; |
2714 | struct folio *folio; | |
1a4e58cc MK |
2715 | struct scan_control sc = { |
2716 | .gfp_mask = GFP_KERNEL, | |
1a4e58cc MK |
2717 | .may_writepage = 1, |
2718 | .may_unmap = 1, | |
2719 | .may_swap = 1, | |
26aa2d19 | 2720 | .no_demotion = 1, |
1a4e58cc MK |
2721 | }; |
2722 | ||
49fd9b6d MWO |
2723 | nr_reclaimed = shrink_folio_list(folio_list, pgdat, &sc, &dummy_stat, false); |
2724 | while (!list_empty(folio_list)) { | |
2725 | folio = lru_to_folio(folio_list); | |
1fe47c0b ML |
2726 | list_del(&folio->lru); |
2727 | folio_putback_lru(folio); | |
2728 | } | |
2729 | ||
2730 | return nr_reclaimed; | |
2731 | } | |
2732 | ||
a83f0551 | 2733 | unsigned long reclaim_pages(struct list_head *folio_list) |
1fe47c0b | 2734 | { |
ed657e55 | 2735 | int nid; |
1fe47c0b | 2736 | unsigned int nr_reclaimed = 0; |
a83f0551 | 2737 | LIST_HEAD(node_folio_list); |
1fe47c0b ML |
2738 | unsigned int noreclaim_flag; |
2739 | ||
a83f0551 | 2740 | if (list_empty(folio_list)) |
1ae65e27 WY |
2741 | return nr_reclaimed; |
2742 | ||
2d2b8d2b YZ |
2743 | noreclaim_flag = memalloc_noreclaim_save(); |
2744 | ||
a83f0551 | 2745 | nid = folio_nid(lru_to_folio(folio_list)); |
1ae65e27 | 2746 | do { |
a83f0551 | 2747 | struct folio *folio = lru_to_folio(folio_list); |
1a4e58cc | 2748 | |
a83f0551 MWO |
2749 | if (nid == folio_nid(folio)) { |
2750 | folio_clear_active(folio); | |
2751 | list_move(&folio->lru, &node_folio_list); | |
1a4e58cc MK |
2752 | continue; |
2753 | } | |
2754 | ||
49fd9b6d | 2755 | nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid)); |
a83f0551 MWO |
2756 | nid = folio_nid(lru_to_folio(folio_list)); |
2757 | } while (!list_empty(folio_list)); | |
1a4e58cc | 2758 | |
49fd9b6d | 2759 | nr_reclaimed += reclaim_folio_list(&node_folio_list, NODE_DATA(nid)); |
1a4e58cc | 2760 | |
2d2b8d2b YZ |
2761 | memalloc_noreclaim_restore(noreclaim_flag); |
2762 | ||
1a4e58cc MK |
2763 | return nr_reclaimed; |
2764 | } | |
2765 | ||
b91ac374 JW |
2766 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
2767 | struct lruvec *lruvec, struct scan_control *sc) | |
2768 | { | |
2769 | if (is_active_lru(lru)) { | |
2770 | if (sc->may_deactivate & (1 << is_file_lru(lru))) | |
2771 | shrink_active_list(nr_to_scan, lruvec, sc, lru); | |
2772 | else | |
2773 | sc->skipped_deactivate = 1; | |
2774 | return 0; | |
2775 | } | |
2776 | ||
2777 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); | |
2778 | } | |
2779 | ||
59dc76b0 RR |
2780 | /* |
2781 | * The inactive anon list should be small enough that the VM never has | |
2782 | * to do too much work. | |
14797e23 | 2783 | * |
59dc76b0 RR |
2784 | * The inactive file list should be small enough to leave most memory |
2785 | * to the established workingset on the scan-resistant active list, | |
2786 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2787 | * |
59dc76b0 | 2788 | * Both inactive lists should also be large enough that each inactive |
49fd9b6d | 2789 | * folio has a chance to be referenced again before it is reclaimed. |
56e49d21 | 2790 | * |
2a2e4885 JW |
2791 | * If that fails and refaulting is observed, the inactive list grows. |
2792 | * | |
49fd9b6d | 2793 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE folios |
3a50d14d | 2794 | * on this LRU, maintained by the pageout code. An inactive_ratio |
49fd9b6d | 2795 | * of 3 means 3:1 or 25% of the folios are kept on the inactive list. |
56e49d21 | 2796 | * |
59dc76b0 RR |
2797 | * total target max |
2798 | * memory ratio inactive | |
2799 | * ------------------------------------- | |
2800 | * 10MB 1 5MB | |
2801 | * 100MB 1 50MB | |
2802 | * 1GB 3 250MB | |
2803 | * 10GB 10 0.9GB | |
2804 | * 100GB 31 3GB | |
2805 | * 1TB 101 10GB | |
2806 | * 10TB 320 32GB | |
56e49d21 | 2807 | */ |
b91ac374 | 2808 | static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru) |
56e49d21 | 2809 | { |
b91ac374 | 2810 | enum lru_list active_lru = inactive_lru + LRU_ACTIVE; |
2a2e4885 JW |
2811 | unsigned long inactive, active; |
2812 | unsigned long inactive_ratio; | |
59dc76b0 | 2813 | unsigned long gb; |
e3790144 | 2814 | |
b91ac374 JW |
2815 | inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru); |
2816 | active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru); | |
f8d1a311 | 2817 | |
b91ac374 | 2818 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
4002570c | 2819 | if (gb) |
b91ac374 JW |
2820 | inactive_ratio = int_sqrt(10 * gb); |
2821 | else | |
2822 | inactive_ratio = 1; | |
fd538803 | 2823 | |
59dc76b0 | 2824 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2825 | } |
2826 | ||
9a265114 JW |
2827 | enum scan_balance { |
2828 | SCAN_EQUAL, | |
2829 | SCAN_FRACT, | |
2830 | SCAN_ANON, | |
2831 | SCAN_FILE, | |
2832 | }; | |
2833 | ||
f1e1a7be YZ |
2834 | static void prepare_scan_count(pg_data_t *pgdat, struct scan_control *sc) |
2835 | { | |
2836 | unsigned long file; | |
2837 | struct lruvec *target_lruvec; | |
2838 | ||
ac35a490 YZ |
2839 | if (lru_gen_enabled()) |
2840 | return; | |
2841 | ||
f1e1a7be YZ |
2842 | target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat); |
2843 | ||
2844 | /* | |
2845 | * Flush the memory cgroup stats, so that we read accurate per-memcg | |
2846 | * lruvec stats for heuristics. | |
2847 | */ | |
2848 | mem_cgroup_flush_stats(); | |
2849 | ||
2850 | /* | |
2851 | * Determine the scan balance between anon and file LRUs. | |
2852 | */ | |
2853 | spin_lock_irq(&target_lruvec->lru_lock); | |
2854 | sc->anon_cost = target_lruvec->anon_cost; | |
2855 | sc->file_cost = target_lruvec->file_cost; | |
2856 | spin_unlock_irq(&target_lruvec->lru_lock); | |
2857 | ||
2858 | /* | |
2859 | * Target desirable inactive:active list ratios for the anon | |
2860 | * and file LRU lists. | |
2861 | */ | |
2862 | if (!sc->force_deactivate) { | |
2863 | unsigned long refaults; | |
2864 | ||
2865 | /* | |
2866 | * When refaults are being observed, it means a new | |
2867 | * workingset is being established. Deactivate to get | |
2868 | * rid of any stale active pages quickly. | |
2869 | */ | |
2870 | refaults = lruvec_page_state(target_lruvec, | |
2871 | WORKINGSET_ACTIVATE_ANON); | |
2872 | if (refaults != target_lruvec->refaults[WORKINGSET_ANON] || | |
2873 | inactive_is_low(target_lruvec, LRU_INACTIVE_ANON)) | |
2874 | sc->may_deactivate |= DEACTIVATE_ANON; | |
2875 | else | |
2876 | sc->may_deactivate &= ~DEACTIVATE_ANON; | |
2877 | ||
2878 | refaults = lruvec_page_state(target_lruvec, | |
2879 | WORKINGSET_ACTIVATE_FILE); | |
2880 | if (refaults != target_lruvec->refaults[WORKINGSET_FILE] || | |
2881 | inactive_is_low(target_lruvec, LRU_INACTIVE_FILE)) | |
2882 | sc->may_deactivate |= DEACTIVATE_FILE; | |
2883 | else | |
2884 | sc->may_deactivate &= ~DEACTIVATE_FILE; | |
2885 | } else | |
2886 | sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE; | |
2887 | ||
2888 | /* | |
2889 | * If we have plenty of inactive file pages that aren't | |
2890 | * thrashing, try to reclaim those first before touching | |
2891 | * anonymous pages. | |
2892 | */ | |
2893 | file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE); | |
2894 | if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE)) | |
2895 | sc->cache_trim_mode = 1; | |
2896 | else | |
2897 | sc->cache_trim_mode = 0; | |
2898 | ||
2899 | /* | |
2900 | * Prevent the reclaimer from falling into the cache trap: as | |
2901 | * cache pages start out inactive, every cache fault will tip | |
2902 | * the scan balance towards the file LRU. And as the file LRU | |
2903 | * shrinks, so does the window for rotation from references. | |
2904 | * This means we have a runaway feedback loop where a tiny | |
2905 | * thrashing file LRU becomes infinitely more attractive than | |
2906 | * anon pages. Try to detect this based on file LRU size. | |
2907 | */ | |
2908 | if (!cgroup_reclaim(sc)) { | |
2909 | unsigned long total_high_wmark = 0; | |
2910 | unsigned long free, anon; | |
2911 | int z; | |
2912 | ||
2913 | free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); | |
2914 | file = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2915 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2916 | ||
2917 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2918 | struct zone *zone = &pgdat->node_zones[z]; | |
2919 | ||
2920 | if (!managed_zone(zone)) | |
2921 | continue; | |
2922 | ||
2923 | total_high_wmark += high_wmark_pages(zone); | |
2924 | } | |
2925 | ||
2926 | /* | |
2927 | * Consider anon: if that's low too, this isn't a | |
2928 | * runaway file reclaim problem, but rather just | |
2929 | * extreme pressure. Reclaim as per usual then. | |
2930 | */ | |
2931 | anon = node_page_state(pgdat, NR_INACTIVE_ANON); | |
2932 | ||
2933 | sc->file_is_tiny = | |
2934 | file + free <= total_high_wmark && | |
2935 | !(sc->may_deactivate & DEACTIVATE_ANON) && | |
2936 | anon >> sc->priority; | |
2937 | } | |
2938 | } | |
2939 | ||
4f98a2fe RR |
2940 | /* |
2941 | * Determine how aggressively the anon and file LRU lists should be | |
02e458d8 | 2942 | * scanned. |
4f98a2fe | 2943 | * |
49fd9b6d MWO |
2944 | * nr[0] = anon inactive folios to scan; nr[1] = anon active folios to scan |
2945 | * nr[2] = file inactive folios to scan; nr[3] = file active folios to scan | |
4f98a2fe | 2946 | */ |
afaf07a6 JW |
2947 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, |
2948 | unsigned long *nr) | |
4f98a2fe | 2949 | { |
a2a36488 | 2950 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
afaf07a6 | 2951 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); |
d483a5dd | 2952 | unsigned long anon_cost, file_cost, total_cost; |
33377678 | 2953 | int swappiness = mem_cgroup_swappiness(memcg); |
ed017373 | 2954 | u64 fraction[ANON_AND_FILE]; |
9a265114 | 2955 | u64 denominator = 0; /* gcc */ |
9a265114 | 2956 | enum scan_balance scan_balance; |
4f98a2fe | 2957 | unsigned long ap, fp; |
4111304d | 2958 | enum lru_list lru; |
76a33fc3 | 2959 | |
49fd9b6d | 2960 | /* If we have no swap space, do not bother scanning anon folios. */ |
a2a36488 | 2961 | if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) { |
9a265114 | 2962 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2963 | goto out; |
2964 | } | |
4f98a2fe | 2965 | |
10316b31 JW |
2966 | /* |
2967 | * Global reclaim will swap to prevent OOM even with no | |
2968 | * swappiness, but memcg users want to use this knob to | |
2969 | * disable swapping for individual groups completely when | |
2970 | * using the memory controller's swap limit feature would be | |
2971 | * too expensive. | |
2972 | */ | |
b5ead35e | 2973 | if (cgroup_reclaim(sc) && !swappiness) { |
9a265114 | 2974 | scan_balance = SCAN_FILE; |
10316b31 JW |
2975 | goto out; |
2976 | } | |
2977 | ||
2978 | /* | |
2979 | * Do not apply any pressure balancing cleverness when the | |
2980 | * system is close to OOM, scan both anon and file equally | |
2981 | * (unless the swappiness setting disagrees with swapping). | |
2982 | */ | |
02695175 | 2983 | if (!sc->priority && swappiness) { |
9a265114 | 2984 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2985 | goto out; |
2986 | } | |
2987 | ||
62376251 | 2988 | /* |
53138cea | 2989 | * If the system is almost out of file pages, force-scan anon. |
62376251 | 2990 | */ |
b91ac374 | 2991 | if (sc->file_is_tiny) { |
53138cea JW |
2992 | scan_balance = SCAN_ANON; |
2993 | goto out; | |
62376251 JW |
2994 | } |
2995 | ||
7c5bd705 | 2996 | /* |
b91ac374 JW |
2997 | * If there is enough inactive page cache, we do not reclaim |
2998 | * anything from the anonymous working right now. | |
7c5bd705 | 2999 | */ |
b91ac374 | 3000 | if (sc->cache_trim_mode) { |
9a265114 | 3001 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
3002 | goto out; |
3003 | } | |
3004 | ||
9a265114 | 3005 | scan_balance = SCAN_FRACT; |
58c37f6e | 3006 | /* |
314b57fb JW |
3007 | * Calculate the pressure balance between anon and file pages. |
3008 | * | |
3009 | * The amount of pressure we put on each LRU is inversely | |
3010 | * proportional to the cost of reclaiming each list, as | |
3011 | * determined by the share of pages that are refaulting, times | |
3012 | * the relative IO cost of bringing back a swapped out | |
3013 | * anonymous page vs reloading a filesystem page (swappiness). | |
3014 | * | |
d483a5dd JW |
3015 | * Although we limit that influence to ensure no list gets |
3016 | * left behind completely: at least a third of the pressure is | |
3017 | * applied, before swappiness. | |
3018 | * | |
314b57fb | 3019 | * With swappiness at 100, anon and file have equal IO cost. |
58c37f6e | 3020 | */ |
d483a5dd JW |
3021 | total_cost = sc->anon_cost + sc->file_cost; |
3022 | anon_cost = total_cost + sc->anon_cost; | |
3023 | file_cost = total_cost + sc->file_cost; | |
3024 | total_cost = anon_cost + file_cost; | |
58c37f6e | 3025 | |
d483a5dd JW |
3026 | ap = swappiness * (total_cost + 1); |
3027 | ap /= anon_cost + 1; | |
4f98a2fe | 3028 | |
d483a5dd JW |
3029 | fp = (200 - swappiness) * (total_cost + 1); |
3030 | fp /= file_cost + 1; | |
4f98a2fe | 3031 | |
76a33fc3 SL |
3032 | fraction[0] = ap; |
3033 | fraction[1] = fp; | |
a4fe1631 | 3034 | denominator = ap + fp; |
76a33fc3 | 3035 | out: |
688035f7 JW |
3036 | for_each_evictable_lru(lru) { |
3037 | int file = is_file_lru(lru); | |
9783aa99 | 3038 | unsigned long lruvec_size; |
f56ce412 | 3039 | unsigned long low, min; |
688035f7 | 3040 | unsigned long scan; |
9783aa99 CD |
3041 | |
3042 | lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); | |
f56ce412 JW |
3043 | mem_cgroup_protection(sc->target_mem_cgroup, memcg, |
3044 | &min, &low); | |
9783aa99 | 3045 | |
f56ce412 | 3046 | if (min || low) { |
9783aa99 CD |
3047 | /* |
3048 | * Scale a cgroup's reclaim pressure by proportioning | |
3049 | * its current usage to its memory.low or memory.min | |
3050 | * setting. | |
3051 | * | |
3052 | * This is important, as otherwise scanning aggression | |
3053 | * becomes extremely binary -- from nothing as we | |
3054 | * approach the memory protection threshold, to totally | |
3055 | * nominal as we exceed it. This results in requiring | |
3056 | * setting extremely liberal protection thresholds. It | |
3057 | * also means we simply get no protection at all if we | |
3058 | * set it too low, which is not ideal. | |
1bc63fb1 CD |
3059 | * |
3060 | * If there is any protection in place, we reduce scan | |
3061 | * pressure by how much of the total memory used is | |
3062 | * within protection thresholds. | |
9783aa99 | 3063 | * |
9de7ca46 CD |
3064 | * There is one special case: in the first reclaim pass, |
3065 | * we skip over all groups that are within their low | |
3066 | * protection. If that fails to reclaim enough pages to | |
3067 | * satisfy the reclaim goal, we come back and override | |
3068 | * the best-effort low protection. However, we still | |
3069 | * ideally want to honor how well-behaved groups are in | |
3070 | * that case instead of simply punishing them all | |
3071 | * equally. As such, we reclaim them based on how much | |
1bc63fb1 CD |
3072 | * memory they are using, reducing the scan pressure |
3073 | * again by how much of the total memory used is under | |
3074 | * hard protection. | |
9783aa99 | 3075 | */ |
1bc63fb1 | 3076 | unsigned long cgroup_size = mem_cgroup_size(memcg); |
f56ce412 JW |
3077 | unsigned long protection; |
3078 | ||
3079 | /* memory.low scaling, make sure we retry before OOM */ | |
3080 | if (!sc->memcg_low_reclaim && low > min) { | |
3081 | protection = low; | |
3082 | sc->memcg_low_skipped = 1; | |
3083 | } else { | |
3084 | protection = min; | |
3085 | } | |
1bc63fb1 CD |
3086 | |
3087 | /* Avoid TOCTOU with earlier protection check */ | |
3088 | cgroup_size = max(cgroup_size, protection); | |
3089 | ||
3090 | scan = lruvec_size - lruvec_size * protection / | |
32d4f4b7 | 3091 | (cgroup_size + 1); |
9783aa99 CD |
3092 | |
3093 | /* | |
1bc63fb1 | 3094 | * Minimally target SWAP_CLUSTER_MAX pages to keep |
55b65a57 | 3095 | * reclaim moving forwards, avoiding decrementing |
9de7ca46 | 3096 | * sc->priority further than desirable. |
9783aa99 | 3097 | */ |
1bc63fb1 | 3098 | scan = max(scan, SWAP_CLUSTER_MAX); |
9783aa99 CD |
3099 | } else { |
3100 | scan = lruvec_size; | |
3101 | } | |
3102 | ||
3103 | scan >>= sc->priority; | |
6b4f7799 | 3104 | |
688035f7 JW |
3105 | /* |
3106 | * If the cgroup's already been deleted, make sure to | |
3107 | * scrape out the remaining cache. | |
3108 | */ | |
3109 | if (!scan && !mem_cgroup_online(memcg)) | |
9783aa99 | 3110 | scan = min(lruvec_size, SWAP_CLUSTER_MAX); |
6b4f7799 | 3111 | |
688035f7 JW |
3112 | switch (scan_balance) { |
3113 | case SCAN_EQUAL: | |
3114 | /* Scan lists relative to size */ | |
3115 | break; | |
3116 | case SCAN_FRACT: | |
9a265114 | 3117 | /* |
688035f7 JW |
3118 | * Scan types proportional to swappiness and |
3119 | * their relative recent reclaim efficiency. | |
76073c64 GS |
3120 | * Make sure we don't miss the last page on |
3121 | * the offlined memory cgroups because of a | |
3122 | * round-off error. | |
9a265114 | 3123 | */ |
76073c64 GS |
3124 | scan = mem_cgroup_online(memcg) ? |
3125 | div64_u64(scan * fraction[file], denominator) : | |
3126 | DIV64_U64_ROUND_UP(scan * fraction[file], | |
68600f62 | 3127 | denominator); |
688035f7 JW |
3128 | break; |
3129 | case SCAN_FILE: | |
3130 | case SCAN_ANON: | |
3131 | /* Scan one type exclusively */ | |
e072bff6 | 3132 | if ((scan_balance == SCAN_FILE) != file) |
688035f7 | 3133 | scan = 0; |
688035f7 JW |
3134 | break; |
3135 | default: | |
3136 | /* Look ma, no brain */ | |
3137 | BUG(); | |
9a265114 | 3138 | } |
688035f7 | 3139 | |
688035f7 | 3140 | nr[lru] = scan; |
76a33fc3 | 3141 | } |
6e08a369 | 3142 | } |
4f98a2fe | 3143 | |
2f368a9f DH |
3144 | /* |
3145 | * Anonymous LRU management is a waste if there is | |
3146 | * ultimately no way to reclaim the memory. | |
3147 | */ | |
3148 | static bool can_age_anon_pages(struct pglist_data *pgdat, | |
3149 | struct scan_control *sc) | |
3150 | { | |
3151 | /* Aging the anon LRU is valuable if swap is present: */ | |
3152 | if (total_swap_pages > 0) | |
3153 | return true; | |
3154 | ||
3155 | /* Also valuable if anon pages can be demoted: */ | |
3156 | return can_demote(pgdat->node_id, sc); | |
3157 | } | |
3158 | ||
ec1c86b2 YZ |
3159 | #ifdef CONFIG_LRU_GEN |
3160 | ||
354ed597 YZ |
3161 | #ifdef CONFIG_LRU_GEN_ENABLED |
3162 | DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS); | |
3163 | #define get_cap(cap) static_branch_likely(&lru_gen_caps[cap]) | |
3164 | #else | |
3165 | DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS); | |
3166 | #define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap]) | |
3167 | #endif | |
3168 | ||
ec1c86b2 YZ |
3169 | /****************************************************************************** |
3170 | * shorthand helpers | |
3171 | ******************************************************************************/ | |
3172 | ||
ac35a490 YZ |
3173 | #define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset)) |
3174 | ||
3175 | #define DEFINE_MAX_SEQ(lruvec) \ | |
3176 | unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq) | |
3177 | ||
3178 | #define DEFINE_MIN_SEQ(lruvec) \ | |
3179 | unsigned long min_seq[ANON_AND_FILE] = { \ | |
3180 | READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \ | |
3181 | READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \ | |
3182 | } | |
3183 | ||
ec1c86b2 YZ |
3184 | #define for_each_gen_type_zone(gen, type, zone) \ |
3185 | for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \ | |
3186 | for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \ | |
3187 | for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++) | |
3188 | ||
bd74fdae | 3189 | static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid) |
ec1c86b2 YZ |
3190 | { |
3191 | struct pglist_data *pgdat = NODE_DATA(nid); | |
3192 | ||
3193 | #ifdef CONFIG_MEMCG | |
3194 | if (memcg) { | |
3195 | struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec; | |
3196 | ||
931b6a8b | 3197 | /* see the comment in mem_cgroup_lruvec() */ |
ec1c86b2 YZ |
3198 | if (!lruvec->pgdat) |
3199 | lruvec->pgdat = pgdat; | |
3200 | ||
3201 | return lruvec; | |
3202 | } | |
3203 | #endif | |
3204 | VM_WARN_ON_ONCE(!mem_cgroup_disabled()); | |
3205 | ||
931b6a8b | 3206 | return &pgdat->__lruvec; |
ec1c86b2 YZ |
3207 | } |
3208 | ||
ac35a490 YZ |
3209 | static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc) |
3210 | { | |
3211 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
3212 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); | |
3213 | ||
3214 | if (!can_demote(pgdat->node_id, sc) && | |
3215 | mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH) | |
3216 | return 0; | |
3217 | ||
3218 | return mem_cgroup_swappiness(memcg); | |
3219 | } | |
3220 | ||
3221 | static int get_nr_gens(struct lruvec *lruvec, int type) | |
3222 | { | |
3223 | return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1; | |
3224 | } | |
3225 | ||
3226 | static bool __maybe_unused seq_is_valid(struct lruvec *lruvec) | |
3227 | { | |
391655fe | 3228 | /* see the comment on lru_gen_folio */ |
ac35a490 YZ |
3229 | return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS && |
3230 | get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) && | |
3231 | get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS; | |
3232 | } | |
3233 | ||
bd74fdae YZ |
3234 | /****************************************************************************** |
3235 | * mm_struct list | |
3236 | ******************************************************************************/ | |
3237 | ||
3238 | static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg) | |
3239 | { | |
3240 | static struct lru_gen_mm_list mm_list = { | |
3241 | .fifo = LIST_HEAD_INIT(mm_list.fifo), | |
3242 | .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock), | |
3243 | }; | |
3244 | ||
3245 | #ifdef CONFIG_MEMCG | |
3246 | if (memcg) | |
3247 | return &memcg->mm_list; | |
3248 | #endif | |
3249 | VM_WARN_ON_ONCE(!mem_cgroup_disabled()); | |
3250 | ||
3251 | return &mm_list; | |
3252 | } | |
3253 | ||
3254 | void lru_gen_add_mm(struct mm_struct *mm) | |
3255 | { | |
3256 | int nid; | |
3257 | struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm); | |
3258 | struct lru_gen_mm_list *mm_list = get_mm_list(memcg); | |
3259 | ||
3260 | VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list)); | |
3261 | #ifdef CONFIG_MEMCG | |
3262 | VM_WARN_ON_ONCE(mm->lru_gen.memcg); | |
3263 | mm->lru_gen.memcg = memcg; | |
3264 | #endif | |
3265 | spin_lock(&mm_list->lock); | |
3266 | ||
3267 | for_each_node_state(nid, N_MEMORY) { | |
3268 | struct lruvec *lruvec = get_lruvec(memcg, nid); | |
3269 | ||
bd74fdae YZ |
3270 | /* the first addition since the last iteration */ |
3271 | if (lruvec->mm_state.tail == &mm_list->fifo) | |
3272 | lruvec->mm_state.tail = &mm->lru_gen.list; | |
3273 | } | |
3274 | ||
3275 | list_add_tail(&mm->lru_gen.list, &mm_list->fifo); | |
3276 | ||
3277 | spin_unlock(&mm_list->lock); | |
3278 | } | |
3279 | ||
3280 | void lru_gen_del_mm(struct mm_struct *mm) | |
3281 | { | |
3282 | int nid; | |
3283 | struct lru_gen_mm_list *mm_list; | |
3284 | struct mem_cgroup *memcg = NULL; | |
3285 | ||
3286 | if (list_empty(&mm->lru_gen.list)) | |
3287 | return; | |
3288 | ||
3289 | #ifdef CONFIG_MEMCG | |
3290 | memcg = mm->lru_gen.memcg; | |
3291 | #endif | |
3292 | mm_list = get_mm_list(memcg); | |
3293 | ||
3294 | spin_lock(&mm_list->lock); | |
3295 | ||
3296 | for_each_node(nid) { | |
3297 | struct lruvec *lruvec = get_lruvec(memcg, nid); | |
3298 | ||
bd74fdae YZ |
3299 | /* where the last iteration ended (exclusive) */ |
3300 | if (lruvec->mm_state.tail == &mm->lru_gen.list) | |
3301 | lruvec->mm_state.tail = lruvec->mm_state.tail->next; | |
3302 | ||
3303 | /* where the current iteration continues (inclusive) */ | |
3304 | if (lruvec->mm_state.head != &mm->lru_gen.list) | |
3305 | continue; | |
3306 | ||
3307 | lruvec->mm_state.head = lruvec->mm_state.head->next; | |
3308 | /* the deletion ends the current iteration */ | |
3309 | if (lruvec->mm_state.head == &mm_list->fifo) | |
3310 | WRITE_ONCE(lruvec->mm_state.seq, lruvec->mm_state.seq + 1); | |
3311 | } | |
3312 | ||
3313 | list_del_init(&mm->lru_gen.list); | |
3314 | ||
3315 | spin_unlock(&mm_list->lock); | |
3316 | ||
3317 | #ifdef CONFIG_MEMCG | |
3318 | mem_cgroup_put(mm->lru_gen.memcg); | |
3319 | mm->lru_gen.memcg = NULL; | |
3320 | #endif | |
3321 | } | |
3322 | ||
3323 | #ifdef CONFIG_MEMCG | |
3324 | void lru_gen_migrate_mm(struct mm_struct *mm) | |
3325 | { | |
3326 | struct mem_cgroup *memcg; | |
3327 | struct task_struct *task = rcu_dereference_protected(mm->owner, true); | |
3328 | ||
3329 | VM_WARN_ON_ONCE(task->mm != mm); | |
3330 | lockdep_assert_held(&task->alloc_lock); | |
3331 | ||
3332 | /* for mm_update_next_owner() */ | |
3333 | if (mem_cgroup_disabled()) | |
3334 | return; | |
3335 | ||
3336 | rcu_read_lock(); | |
3337 | memcg = mem_cgroup_from_task(task); | |
3338 | rcu_read_unlock(); | |
3339 | if (memcg == mm->lru_gen.memcg) | |
3340 | return; | |
3341 | ||
3342 | VM_WARN_ON_ONCE(!mm->lru_gen.memcg); | |
3343 | VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list)); | |
3344 | ||
3345 | lru_gen_del_mm(mm); | |
3346 | lru_gen_add_mm(mm); | |
3347 | } | |
3348 | #endif | |
3349 | ||
3350 | /* | |
3351 | * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when | |
3352 | * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of | |
3353 | * bits in a bitmap, k is the number of hash functions and n is the number of | |
3354 | * inserted items. | |
3355 | * | |
3356 | * Page table walkers use one of the two filters to reduce their search space. | |
3357 | * To get rid of non-leaf entries that no longer have enough leaf entries, the | |
3358 | * aging uses the double-buffering technique to flip to the other filter each | |
3359 | * time it produces a new generation. For non-leaf entries that have enough | |
3360 | * leaf entries, the aging carries them over to the next generation in | |
3361 | * walk_pmd_range(); the eviction also report them when walking the rmap | |
3362 | * in lru_gen_look_around(). | |
3363 | * | |
3364 | * For future optimizations: | |
3365 | * 1. It's not necessary to keep both filters all the time. The spare one can be | |
3366 | * freed after the RCU grace period and reallocated if needed again. | |
3367 | * 2. And when reallocating, it's worth scaling its size according to the number | |
3368 | * of inserted entries in the other filter, to reduce the memory overhead on | |
3369 | * small systems and false positives on large systems. | |
3370 | * 3. Jenkins' hash function is an alternative to Knuth's. | |
3371 | */ | |
3372 | #define BLOOM_FILTER_SHIFT 15 | |
3373 | ||
3374 | static inline int filter_gen_from_seq(unsigned long seq) | |
3375 | { | |
3376 | return seq % NR_BLOOM_FILTERS; | |
3377 | } | |
3378 | ||
3379 | static void get_item_key(void *item, int *key) | |
3380 | { | |
3381 | u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2); | |
3382 | ||
3383 | BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32)); | |
3384 | ||
3385 | key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1); | |
3386 | key[1] = hash >> BLOOM_FILTER_SHIFT; | |
3387 | } | |
3388 | ||
3389 | static void reset_bloom_filter(struct lruvec *lruvec, unsigned long seq) | |
3390 | { | |
3391 | unsigned long *filter; | |
3392 | int gen = filter_gen_from_seq(seq); | |
3393 | ||
3394 | filter = lruvec->mm_state.filters[gen]; | |
3395 | if (filter) { | |
3396 | bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT)); | |
3397 | return; | |
3398 | } | |
3399 | ||
3400 | filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT), | |
3401 | __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN); | |
3402 | WRITE_ONCE(lruvec->mm_state.filters[gen], filter); | |
3403 | } | |
3404 | ||
3405 | static void update_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item) | |
3406 | { | |
3407 | int key[2]; | |
3408 | unsigned long *filter; | |
3409 | int gen = filter_gen_from_seq(seq); | |
3410 | ||
3411 | filter = READ_ONCE(lruvec->mm_state.filters[gen]); | |
3412 | if (!filter) | |
3413 | return; | |
3414 | ||
3415 | get_item_key(item, key); | |
3416 | ||
3417 | if (!test_bit(key[0], filter)) | |
3418 | set_bit(key[0], filter); | |
3419 | if (!test_bit(key[1], filter)) | |
3420 | set_bit(key[1], filter); | |
3421 | } | |
3422 | ||
3423 | static bool test_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item) | |
3424 | { | |
3425 | int key[2]; | |
3426 | unsigned long *filter; | |
3427 | int gen = filter_gen_from_seq(seq); | |
3428 | ||
3429 | filter = READ_ONCE(lruvec->mm_state.filters[gen]); | |
3430 | if (!filter) | |
3431 | return true; | |
3432 | ||
3433 | get_item_key(item, key); | |
3434 | ||
3435 | return test_bit(key[0], filter) && test_bit(key[1], filter); | |
3436 | } | |
3437 | ||
3438 | static void reset_mm_stats(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, bool last) | |
3439 | { | |
3440 | int i; | |
3441 | int hist; | |
3442 | ||
3443 | lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock); | |
3444 | ||
3445 | if (walk) { | |
3446 | hist = lru_hist_from_seq(walk->max_seq); | |
3447 | ||
3448 | for (i = 0; i < NR_MM_STATS; i++) { | |
3449 | WRITE_ONCE(lruvec->mm_state.stats[hist][i], | |
3450 | lruvec->mm_state.stats[hist][i] + walk->mm_stats[i]); | |
3451 | walk->mm_stats[i] = 0; | |
3452 | } | |
3453 | } | |
3454 | ||
3455 | if (NR_HIST_GENS > 1 && last) { | |
3456 | hist = lru_hist_from_seq(lruvec->mm_state.seq + 1); | |
3457 | ||
3458 | for (i = 0; i < NR_MM_STATS; i++) | |
3459 | WRITE_ONCE(lruvec->mm_state.stats[hist][i], 0); | |
3460 | } | |
3461 | } | |
3462 | ||
3463 | static bool should_skip_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk) | |
3464 | { | |
3465 | int type; | |
3466 | unsigned long size = 0; | |
3467 | struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec); | |
3468 | int key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap); | |
3469 | ||
3470 | if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap)) | |
3471 | return true; | |
3472 | ||
3473 | clear_bit(key, &mm->lru_gen.bitmap); | |
3474 | ||
3475 | for (type = !walk->can_swap; type < ANON_AND_FILE; type++) { | |
3476 | size += type ? get_mm_counter(mm, MM_FILEPAGES) : | |
3477 | get_mm_counter(mm, MM_ANONPAGES) + | |
3478 | get_mm_counter(mm, MM_SHMEMPAGES); | |
3479 | } | |
3480 | ||
3481 | if (size < MIN_LRU_BATCH) | |
3482 | return true; | |
3483 | ||
3484 | return !mmget_not_zero(mm); | |
3485 | } | |
3486 | ||
3487 | static bool iterate_mm_list(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, | |
3488 | struct mm_struct **iter) | |
3489 | { | |
3490 | bool first = false; | |
3491 | bool last = true; | |
3492 | struct mm_struct *mm = NULL; | |
3493 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
3494 | struct lru_gen_mm_list *mm_list = get_mm_list(memcg); | |
3495 | struct lru_gen_mm_state *mm_state = &lruvec->mm_state; | |
3496 | ||
3497 | /* | |
3498 | * There are four interesting cases for this page table walker: | |
3499 | * 1. It tries to start a new iteration of mm_list with a stale max_seq; | |
3500 | * there is nothing left to do. | |
3501 | * 2. It's the first of the current generation, and it needs to reset | |
3502 | * the Bloom filter for the next generation. | |
3503 | * 3. It reaches the end of mm_list, and it needs to increment | |
3504 | * mm_state->seq; the iteration is done. | |
3505 | * 4. It's the last of the current generation, and it needs to reset the | |
3506 | * mm stats counters for the next generation. | |
3507 | */ | |
3508 | spin_lock(&mm_list->lock); | |
3509 | ||
3510 | VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->max_seq); | |
3511 | VM_WARN_ON_ONCE(*iter && mm_state->seq > walk->max_seq); | |
3512 | VM_WARN_ON_ONCE(*iter && !mm_state->nr_walkers); | |
3513 | ||
3514 | if (walk->max_seq <= mm_state->seq) { | |
3515 | if (!*iter) | |
3516 | last = false; | |
3517 | goto done; | |
3518 | } | |
3519 | ||
3520 | if (!mm_state->nr_walkers) { | |
3521 | VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo); | |
3522 | ||
3523 | mm_state->head = mm_list->fifo.next; | |
3524 | first = true; | |
3525 | } | |
3526 | ||
3527 | while (!mm && mm_state->head != &mm_list->fifo) { | |
3528 | mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list); | |
3529 | ||
3530 | mm_state->head = mm_state->head->next; | |
3531 | ||
3532 | /* force scan for those added after the last iteration */ | |
3533 | if (!mm_state->tail || mm_state->tail == &mm->lru_gen.list) { | |
3534 | mm_state->tail = mm_state->head; | |
3535 | walk->force_scan = true; | |
3536 | } | |
3537 | ||
3538 | if (should_skip_mm(mm, walk)) | |
3539 | mm = NULL; | |
3540 | } | |
3541 | ||
3542 | if (mm_state->head == &mm_list->fifo) | |
3543 | WRITE_ONCE(mm_state->seq, mm_state->seq + 1); | |
3544 | done: | |
3545 | if (*iter && !mm) | |
3546 | mm_state->nr_walkers--; | |
3547 | if (!*iter && mm) | |
3548 | mm_state->nr_walkers++; | |
3549 | ||
3550 | if (mm_state->nr_walkers) | |
3551 | last = false; | |
3552 | ||
3553 | if (*iter || last) | |
3554 | reset_mm_stats(lruvec, walk, last); | |
3555 | ||
3556 | spin_unlock(&mm_list->lock); | |
3557 | ||
3558 | if (mm && first) | |
3559 | reset_bloom_filter(lruvec, walk->max_seq + 1); | |
3560 | ||
3561 | if (*iter) | |
3562 | mmput_async(*iter); | |
3563 | ||
3564 | *iter = mm; | |
3565 | ||
3566 | return last; | |
3567 | } | |
3568 | ||
3569 | static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long max_seq) | |
3570 | { | |
3571 | bool success = false; | |
3572 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
3573 | struct lru_gen_mm_list *mm_list = get_mm_list(memcg); | |
3574 | struct lru_gen_mm_state *mm_state = &lruvec->mm_state; | |
3575 | ||
3576 | spin_lock(&mm_list->lock); | |
3577 | ||
3578 | VM_WARN_ON_ONCE(mm_state->seq + 1 < max_seq); | |
3579 | ||
3580 | if (max_seq > mm_state->seq && !mm_state->nr_walkers) { | |
3581 | VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo); | |
3582 | ||
3583 | WRITE_ONCE(mm_state->seq, mm_state->seq + 1); | |
3584 | reset_mm_stats(lruvec, NULL, true); | |
3585 | success = true; | |
3586 | } | |
3587 | ||
3588 | spin_unlock(&mm_list->lock); | |
3589 | ||
3590 | return success; | |
3591 | } | |
3592 | ||
ac35a490 YZ |
3593 | /****************************************************************************** |
3594 | * refault feedback loop | |
3595 | ******************************************************************************/ | |
3596 | ||
3597 | /* | |
3598 | * A feedback loop based on Proportional-Integral-Derivative (PID) controller. | |
3599 | * | |
3600 | * The P term is refaulted/(evicted+protected) from a tier in the generation | |
3601 | * currently being evicted; the I term is the exponential moving average of the | |
3602 | * P term over the generations previously evicted, using the smoothing factor | |
3603 | * 1/2; the D term isn't supported. | |
3604 | * | |
3605 | * The setpoint (SP) is always the first tier of one type; the process variable | |
3606 | * (PV) is either any tier of the other type or any other tier of the same | |
3607 | * type. | |
3608 | * | |
3609 | * The error is the difference between the SP and the PV; the correction is to | |
3610 | * turn off protection when SP>PV or turn on protection when SP<PV. | |
3611 | * | |
3612 | * For future optimizations: | |
3613 | * 1. The D term may discount the other two terms over time so that long-lived | |
3614 | * generations can resist stale information. | |
3615 | */ | |
3616 | struct ctrl_pos { | |
3617 | unsigned long refaulted; | |
3618 | unsigned long total; | |
3619 | int gain; | |
3620 | }; | |
3621 | ||
3622 | static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain, | |
3623 | struct ctrl_pos *pos) | |
3624 | { | |
391655fe | 3625 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
ac35a490 YZ |
3626 | int hist = lru_hist_from_seq(lrugen->min_seq[type]); |
3627 | ||
3628 | pos->refaulted = lrugen->avg_refaulted[type][tier] + | |
3629 | atomic_long_read(&lrugen->refaulted[hist][type][tier]); | |
3630 | pos->total = lrugen->avg_total[type][tier] + | |
3631 | atomic_long_read(&lrugen->evicted[hist][type][tier]); | |
3632 | if (tier) | |
3633 | pos->total += lrugen->protected[hist][type][tier - 1]; | |
3634 | pos->gain = gain; | |
3635 | } | |
3636 | ||
3637 | static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover) | |
3638 | { | |
3639 | int hist, tier; | |
391655fe | 3640 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
ac35a490 YZ |
3641 | bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1; |
3642 | unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1; | |
3643 | ||
3644 | lockdep_assert_held(&lruvec->lru_lock); | |
3645 | ||
3646 | if (!carryover && !clear) | |
3647 | return; | |
3648 | ||
3649 | hist = lru_hist_from_seq(seq); | |
3650 | ||
3651 | for (tier = 0; tier < MAX_NR_TIERS; tier++) { | |
3652 | if (carryover) { | |
3653 | unsigned long sum; | |
3654 | ||
3655 | sum = lrugen->avg_refaulted[type][tier] + | |
3656 | atomic_long_read(&lrugen->refaulted[hist][type][tier]); | |
3657 | WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2); | |
3658 | ||
3659 | sum = lrugen->avg_total[type][tier] + | |
3660 | atomic_long_read(&lrugen->evicted[hist][type][tier]); | |
3661 | if (tier) | |
3662 | sum += lrugen->protected[hist][type][tier - 1]; | |
3663 | WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2); | |
3664 | } | |
3665 | ||
3666 | if (clear) { | |
3667 | atomic_long_set(&lrugen->refaulted[hist][type][tier], 0); | |
3668 | atomic_long_set(&lrugen->evicted[hist][type][tier], 0); | |
3669 | if (tier) | |
3670 | WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0); | |
3671 | } | |
3672 | } | |
3673 | } | |
3674 | ||
3675 | static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv) | |
3676 | { | |
3677 | /* | |
3678 | * Return true if the PV has a limited number of refaults or a lower | |
3679 | * refaulted/total than the SP. | |
3680 | */ | |
3681 | return pv->refaulted < MIN_LRU_BATCH || | |
3682 | pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <= | |
3683 | (sp->refaulted + 1) * pv->total * pv->gain; | |
3684 | } | |
3685 | ||
3686 | /****************************************************************************** | |
3687 | * the aging | |
3688 | ******************************************************************************/ | |
3689 | ||
018ee47f YZ |
3690 | /* promote pages accessed through page tables */ |
3691 | static int folio_update_gen(struct folio *folio, int gen) | |
3692 | { | |
3693 | unsigned long new_flags, old_flags = READ_ONCE(folio->flags); | |
3694 | ||
3695 | VM_WARN_ON_ONCE(gen >= MAX_NR_GENS); | |
3696 | VM_WARN_ON_ONCE(!rcu_read_lock_held()); | |
3697 | ||
3698 | do { | |
3699 | /* lru_gen_del_folio() has isolated this page? */ | |
3700 | if (!(old_flags & LRU_GEN_MASK)) { | |
49fd9b6d | 3701 | /* for shrink_folio_list() */ |
018ee47f YZ |
3702 | new_flags = old_flags | BIT(PG_referenced); |
3703 | continue; | |
3704 | } | |
3705 | ||
3706 | new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS); | |
3707 | new_flags |= (gen + 1UL) << LRU_GEN_PGOFF; | |
3708 | } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags)); | |
3709 | ||
3710 | return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1; | |
3711 | } | |
3712 | ||
ac35a490 YZ |
3713 | /* protect pages accessed multiple times through file descriptors */ |
3714 | static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming) | |
3715 | { | |
3716 | int type = folio_is_file_lru(folio); | |
391655fe | 3717 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
ac35a490 YZ |
3718 | int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]); |
3719 | unsigned long new_flags, old_flags = READ_ONCE(folio->flags); | |
3720 | ||
3721 | VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio); | |
3722 | ||
3723 | do { | |
018ee47f YZ |
3724 | new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1; |
3725 | /* folio_update_gen() has promoted this page? */ | |
3726 | if (new_gen >= 0 && new_gen != old_gen) | |
3727 | return new_gen; | |
3728 | ||
ac35a490 YZ |
3729 | new_gen = (old_gen + 1) % MAX_NR_GENS; |
3730 | ||
3731 | new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS); | |
3732 | new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF; | |
3733 | /* for folio_end_writeback() */ | |
3734 | if (reclaiming) | |
3735 | new_flags |= BIT(PG_reclaim); | |
3736 | } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags)); | |
3737 | ||
3738 | lru_gen_update_size(lruvec, folio, old_gen, new_gen); | |
3739 | ||
3740 | return new_gen; | |
3741 | } | |
3742 | ||
bd74fdae YZ |
3743 | static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio, |
3744 | int old_gen, int new_gen) | |
3745 | { | |
3746 | int type = folio_is_file_lru(folio); | |
3747 | int zone = folio_zonenum(folio); | |
3748 | int delta = folio_nr_pages(folio); | |
3749 | ||
3750 | VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS); | |
3751 | VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS); | |
3752 | ||
3753 | walk->batched++; | |
3754 | ||
3755 | walk->nr_pages[old_gen][type][zone] -= delta; | |
3756 | walk->nr_pages[new_gen][type][zone] += delta; | |
3757 | } | |
3758 | ||
3759 | static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk) | |
3760 | { | |
3761 | int gen, type, zone; | |
391655fe | 3762 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
bd74fdae YZ |
3763 | |
3764 | walk->batched = 0; | |
3765 | ||
3766 | for_each_gen_type_zone(gen, type, zone) { | |
3767 | enum lru_list lru = type * LRU_INACTIVE_FILE; | |
3768 | int delta = walk->nr_pages[gen][type][zone]; | |
3769 | ||
3770 | if (!delta) | |
3771 | continue; | |
3772 | ||
3773 | walk->nr_pages[gen][type][zone] = 0; | |
3774 | WRITE_ONCE(lrugen->nr_pages[gen][type][zone], | |
3775 | lrugen->nr_pages[gen][type][zone] + delta); | |
3776 | ||
3777 | if (lru_gen_is_active(lruvec, gen)) | |
3778 | lru += LRU_ACTIVE; | |
3779 | __update_lru_size(lruvec, lru, zone, delta); | |
3780 | } | |
3781 | } | |
3782 | ||
3783 | static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args) | |
3784 | { | |
3785 | struct address_space *mapping; | |
3786 | struct vm_area_struct *vma = args->vma; | |
3787 | struct lru_gen_mm_walk *walk = args->private; | |
3788 | ||
3789 | if (!vma_is_accessible(vma)) | |
3790 | return true; | |
3791 | ||
3792 | if (is_vm_hugetlb_page(vma)) | |
3793 | return true; | |
3794 | ||
3795 | if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL | VM_SEQ_READ | VM_RAND_READ)) | |
3796 | return true; | |
3797 | ||
3798 | if (vma == get_gate_vma(vma->vm_mm)) | |
3799 | return true; | |
3800 | ||
3801 | if (vma_is_anonymous(vma)) | |
3802 | return !walk->can_swap; | |
3803 | ||
3804 | if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping)) | |
3805 | return true; | |
3806 | ||
3807 | mapping = vma->vm_file->f_mapping; | |
3808 | if (mapping_unevictable(mapping)) | |
3809 | return true; | |
3810 | ||
3811 | if (shmem_mapping(mapping)) | |
3812 | return !walk->can_swap; | |
3813 | ||
3814 | /* to exclude special mappings like dax, etc. */ | |
3815 | return !mapping->a_ops->read_folio; | |
3816 | } | |
3817 | ||
3818 | /* | |
3819 | * Some userspace memory allocators map many single-page VMAs. Instead of | |
3820 | * returning back to the PGD table for each of such VMAs, finish an entire PMD | |
3821 | * table to reduce zigzags and improve cache performance. | |
3822 | */ | |
3823 | static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args, | |
3824 | unsigned long *vm_start, unsigned long *vm_end) | |
3825 | { | |
3826 | unsigned long start = round_up(*vm_end, size); | |
3827 | unsigned long end = (start | ~mask) + 1; | |
78ba531f | 3828 | VMA_ITERATOR(vmi, args->mm, start); |
bd74fdae YZ |
3829 | |
3830 | VM_WARN_ON_ONCE(mask & size); | |
3831 | VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask)); | |
3832 | ||
78ba531f | 3833 | for_each_vma(vmi, args->vma) { |
bd74fdae YZ |
3834 | if (end && end <= args->vma->vm_start) |
3835 | return false; | |
3836 | ||
78ba531f | 3837 | if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args)) |
bd74fdae | 3838 | continue; |
bd74fdae YZ |
3839 | |
3840 | *vm_start = max(start, args->vma->vm_start); | |
3841 | *vm_end = min(end - 1, args->vma->vm_end - 1) + 1; | |
3842 | ||
3843 | return true; | |
3844 | } | |
3845 | ||
3846 | return false; | |
3847 | } | |
3848 | ||
018ee47f YZ |
3849 | static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr) |
3850 | { | |
3851 | unsigned long pfn = pte_pfn(pte); | |
3852 | ||
3853 | VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end); | |
3854 | ||
3855 | if (!pte_present(pte) || is_zero_pfn(pfn)) | |
3856 | return -1; | |
3857 | ||
3858 | if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte))) | |
3859 | return -1; | |
3860 | ||
3861 | if (WARN_ON_ONCE(!pfn_valid(pfn))) | |
3862 | return -1; | |
3863 | ||
3864 | return pfn; | |
3865 | } | |
3866 | ||
bd74fdae YZ |
3867 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG) |
3868 | static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr) | |
3869 | { | |
3870 | unsigned long pfn = pmd_pfn(pmd); | |
3871 | ||
3872 | VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end); | |
3873 | ||
3874 | if (!pmd_present(pmd) || is_huge_zero_pmd(pmd)) | |
3875 | return -1; | |
3876 | ||
3877 | if (WARN_ON_ONCE(pmd_devmap(pmd))) | |
3878 | return -1; | |
3879 | ||
3880 | if (WARN_ON_ONCE(!pfn_valid(pfn))) | |
3881 | return -1; | |
3882 | ||
3883 | return pfn; | |
3884 | } | |
3885 | #endif | |
3886 | ||
018ee47f | 3887 | static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg, |
bd74fdae | 3888 | struct pglist_data *pgdat, bool can_swap) |
018ee47f YZ |
3889 | { |
3890 | struct folio *folio; | |
3891 | ||
3892 | /* try to avoid unnecessary memory loads */ | |
3893 | if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat)) | |
3894 | return NULL; | |
3895 | ||
3896 | folio = pfn_folio(pfn); | |
3897 | if (folio_nid(folio) != pgdat->node_id) | |
3898 | return NULL; | |
3899 | ||
3900 | if (folio_memcg_rcu(folio) != memcg) | |
3901 | return NULL; | |
3902 | ||
bd74fdae YZ |
3903 | /* file VMAs can contain anon pages from COW */ |
3904 | if (!folio_is_file_lru(folio) && !can_swap) | |
3905 | return NULL; | |
3906 | ||
018ee47f YZ |
3907 | return folio; |
3908 | } | |
3909 | ||
bd74fdae YZ |
3910 | static bool suitable_to_scan(int total, int young) |
3911 | { | |
3912 | int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8); | |
3913 | ||
3914 | /* suitable if the average number of young PTEs per cacheline is >=1 */ | |
3915 | return young * n >= total; | |
3916 | } | |
3917 | ||
3918 | static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end, | |
3919 | struct mm_walk *args) | |
3920 | { | |
3921 | int i; | |
3922 | pte_t *pte; | |
3923 | spinlock_t *ptl; | |
3924 | unsigned long addr; | |
3925 | int total = 0; | |
3926 | int young = 0; | |
3927 | struct lru_gen_mm_walk *walk = args->private; | |
3928 | struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec); | |
3929 | struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec); | |
3930 | int old_gen, new_gen = lru_gen_from_seq(walk->max_seq); | |
3931 | ||
3932 | VM_WARN_ON_ONCE(pmd_leaf(*pmd)); | |
3933 | ||
3934 | ptl = pte_lockptr(args->mm, pmd); | |
3935 | if (!spin_trylock(ptl)) | |
3936 | return false; | |
3937 | ||
3938 | arch_enter_lazy_mmu_mode(); | |
3939 | ||
3940 | pte = pte_offset_map(pmd, start & PMD_MASK); | |
3941 | restart: | |
3942 | for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) { | |
3943 | unsigned long pfn; | |
3944 | struct folio *folio; | |
3945 | ||
3946 | total++; | |
3947 | walk->mm_stats[MM_LEAF_TOTAL]++; | |
3948 | ||
3949 | pfn = get_pte_pfn(pte[i], args->vma, addr); | |
3950 | if (pfn == -1) | |
3951 | continue; | |
3952 | ||
3953 | if (!pte_young(pte[i])) { | |
3954 | walk->mm_stats[MM_LEAF_OLD]++; | |
3955 | continue; | |
3956 | } | |
3957 | ||
3958 | folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap); | |
3959 | if (!folio) | |
3960 | continue; | |
3961 | ||
3962 | if (!ptep_test_and_clear_young(args->vma, addr, pte + i)) | |
3963 | VM_WARN_ON_ONCE(true); | |
3964 | ||
3965 | young++; | |
3966 | walk->mm_stats[MM_LEAF_YOUNG]++; | |
3967 | ||
3968 | if (pte_dirty(pte[i]) && !folio_test_dirty(folio) && | |
3969 | !(folio_test_anon(folio) && folio_test_swapbacked(folio) && | |
3970 | !folio_test_swapcache(folio))) | |
3971 | folio_mark_dirty(folio); | |
3972 | ||
3973 | old_gen = folio_update_gen(folio, new_gen); | |
3974 | if (old_gen >= 0 && old_gen != new_gen) | |
3975 | update_batch_size(walk, folio, old_gen, new_gen); | |
3976 | } | |
3977 | ||
3978 | if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end)) | |
3979 | goto restart; | |
3980 | ||
3981 | pte_unmap(pte); | |
3982 | ||
3983 | arch_leave_lazy_mmu_mode(); | |
3984 | spin_unlock(ptl); | |
3985 | ||
3986 | return suitable_to_scan(total, young); | |
3987 | } | |
3988 | ||
3989 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG) | |
3990 | static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma, | |
3991 | struct mm_walk *args, unsigned long *bitmap, unsigned long *start) | |
3992 | { | |
3993 | int i; | |
3994 | pmd_t *pmd; | |
3995 | spinlock_t *ptl; | |
3996 | struct lru_gen_mm_walk *walk = args->private; | |
3997 | struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec); | |
3998 | struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec); | |
3999 | int old_gen, new_gen = lru_gen_from_seq(walk->max_seq); | |
4000 | ||
4001 | VM_WARN_ON_ONCE(pud_leaf(*pud)); | |
4002 | ||
4003 | /* try to batch at most 1+MIN_LRU_BATCH+1 entries */ | |
4004 | if (*start == -1) { | |
4005 | *start = next; | |
4006 | return; | |
4007 | } | |
4008 | ||
4009 | i = next == -1 ? 0 : pmd_index(next) - pmd_index(*start); | |
4010 | if (i && i <= MIN_LRU_BATCH) { | |
4011 | __set_bit(i - 1, bitmap); | |
4012 | return; | |
4013 | } | |
4014 | ||
4015 | pmd = pmd_offset(pud, *start); | |
4016 | ||
4017 | ptl = pmd_lockptr(args->mm, pmd); | |
4018 | if (!spin_trylock(ptl)) | |
4019 | goto done; | |
4020 | ||
4021 | arch_enter_lazy_mmu_mode(); | |
4022 | ||
4023 | do { | |
4024 | unsigned long pfn; | |
4025 | struct folio *folio; | |
4026 | unsigned long addr = i ? (*start & PMD_MASK) + i * PMD_SIZE : *start; | |
4027 | ||
4028 | pfn = get_pmd_pfn(pmd[i], vma, addr); | |
4029 | if (pfn == -1) | |
4030 | goto next; | |
4031 | ||
4032 | if (!pmd_trans_huge(pmd[i])) { | |
4aaf269c | 4033 | if (arch_has_hw_nonleaf_pmd_young() && |
354ed597 | 4034 | get_cap(LRU_GEN_NONLEAF_YOUNG)) |
bd74fdae YZ |
4035 | pmdp_test_and_clear_young(vma, addr, pmd + i); |
4036 | goto next; | |
4037 | } | |
4038 | ||
4039 | folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap); | |
4040 | if (!folio) | |
4041 | goto next; | |
4042 | ||
4043 | if (!pmdp_test_and_clear_young(vma, addr, pmd + i)) | |
4044 | goto next; | |
4045 | ||
4046 | walk->mm_stats[MM_LEAF_YOUNG]++; | |
4047 | ||
4048 | if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) && | |
4049 | !(folio_test_anon(folio) && folio_test_swapbacked(folio) && | |
4050 | !folio_test_swapcache(folio))) | |
4051 | folio_mark_dirty(folio); | |
4052 | ||
4053 | old_gen = folio_update_gen(folio, new_gen); | |
4054 | if (old_gen >= 0 && old_gen != new_gen) | |
4055 | update_batch_size(walk, folio, old_gen, new_gen); | |
4056 | next: | |
4057 | i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1; | |
4058 | } while (i <= MIN_LRU_BATCH); | |
4059 | ||
4060 | arch_leave_lazy_mmu_mode(); | |
4061 | spin_unlock(ptl); | |
4062 | done: | |
4063 | *start = -1; | |
4064 | bitmap_zero(bitmap, MIN_LRU_BATCH); | |
4065 | } | |
4066 | #else | |
4067 | static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma, | |
4068 | struct mm_walk *args, unsigned long *bitmap, unsigned long *start) | |
4069 | { | |
4070 | } | |
4071 | #endif | |
4072 | ||
4073 | static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end, | |
4074 | struct mm_walk *args) | |
4075 | { | |
4076 | int i; | |
4077 | pmd_t *pmd; | |
4078 | unsigned long next; | |
4079 | unsigned long addr; | |
4080 | struct vm_area_struct *vma; | |
4081 | unsigned long pos = -1; | |
4082 | struct lru_gen_mm_walk *walk = args->private; | |
4083 | unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {}; | |
4084 | ||
4085 | VM_WARN_ON_ONCE(pud_leaf(*pud)); | |
4086 | ||
4087 | /* | |
4088 | * Finish an entire PMD in two passes: the first only reaches to PTE | |
4089 | * tables to avoid taking the PMD lock; the second, if necessary, takes | |
4090 | * the PMD lock to clear the accessed bit in PMD entries. | |
4091 | */ | |
4092 | pmd = pmd_offset(pud, start & PUD_MASK); | |
4093 | restart: | |
4094 | /* walk_pte_range() may call get_next_vma() */ | |
4095 | vma = args->vma; | |
4096 | for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) { | |
dab6e717 | 4097 | pmd_t val = pmdp_get_lockless(pmd + i); |
bd74fdae YZ |
4098 | |
4099 | next = pmd_addr_end(addr, end); | |
4100 | ||
4101 | if (!pmd_present(val) || is_huge_zero_pmd(val)) { | |
4102 | walk->mm_stats[MM_LEAF_TOTAL]++; | |
4103 | continue; | |
4104 | } | |
4105 | ||
4106 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
4107 | if (pmd_trans_huge(val)) { | |
4108 | unsigned long pfn = pmd_pfn(val); | |
4109 | struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec); | |
4110 | ||
4111 | walk->mm_stats[MM_LEAF_TOTAL]++; | |
4112 | ||
4113 | if (!pmd_young(val)) { | |
4114 | walk->mm_stats[MM_LEAF_OLD]++; | |
4115 | continue; | |
4116 | } | |
4117 | ||
4118 | /* try to avoid unnecessary memory loads */ | |
4119 | if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat)) | |
4120 | continue; | |
4121 | ||
4122 | walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos); | |
4123 | continue; | |
4124 | } | |
4125 | #endif | |
4126 | walk->mm_stats[MM_NONLEAF_TOTAL]++; | |
4127 | ||
4aaf269c JG |
4128 | if (arch_has_hw_nonleaf_pmd_young() && |
4129 | get_cap(LRU_GEN_NONLEAF_YOUNG)) { | |
354ed597 YZ |
4130 | if (!pmd_young(val)) |
4131 | continue; | |
bd74fdae | 4132 | |
354ed597 YZ |
4133 | walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos); |
4134 | } | |
4aaf269c | 4135 | |
bd74fdae YZ |
4136 | if (!walk->force_scan && !test_bloom_filter(walk->lruvec, walk->max_seq, pmd + i)) |
4137 | continue; | |
4138 | ||
4139 | walk->mm_stats[MM_NONLEAF_FOUND]++; | |
4140 | ||
4141 | if (!walk_pte_range(&val, addr, next, args)) | |
4142 | continue; | |
4143 | ||
4144 | walk->mm_stats[MM_NONLEAF_ADDED]++; | |
4145 | ||
4146 | /* carry over to the next generation */ | |
4147 | update_bloom_filter(walk->lruvec, walk->max_seq + 1, pmd + i); | |
4148 | } | |
4149 | ||
4150 | walk_pmd_range_locked(pud, -1, vma, args, bitmap, &pos); | |
4151 | ||
4152 | if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end)) | |
4153 | goto restart; | |
4154 | } | |
4155 | ||
4156 | static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end, | |
4157 | struct mm_walk *args) | |
4158 | { | |
4159 | int i; | |
4160 | pud_t *pud; | |
4161 | unsigned long addr; | |
4162 | unsigned long next; | |
4163 | struct lru_gen_mm_walk *walk = args->private; | |
4164 | ||
4165 | VM_WARN_ON_ONCE(p4d_leaf(*p4d)); | |
4166 | ||
4167 | pud = pud_offset(p4d, start & P4D_MASK); | |
4168 | restart: | |
4169 | for (i = pud_index(start), addr = start; addr != end; i++, addr = next) { | |
4170 | pud_t val = READ_ONCE(pud[i]); | |
4171 | ||
4172 | next = pud_addr_end(addr, end); | |
4173 | ||
4174 | if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val))) | |
4175 | continue; | |
4176 | ||
4177 | walk_pmd_range(&val, addr, next, args); | |
4178 | ||
4179 | /* a racy check to curtail the waiting time */ | |
4180 | if (wq_has_sleeper(&walk->lruvec->mm_state.wait)) | |
4181 | return 1; | |
4182 | ||
4183 | if (need_resched() || walk->batched >= MAX_LRU_BATCH) { | |
4184 | end = (addr | ~PUD_MASK) + 1; | |
4185 | goto done; | |
4186 | } | |
4187 | } | |
4188 | ||
4189 | if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end)) | |
4190 | goto restart; | |
4191 | ||
4192 | end = round_up(end, P4D_SIZE); | |
4193 | done: | |
4194 | if (!end || !args->vma) | |
4195 | return 1; | |
4196 | ||
4197 | walk->next_addr = max(end, args->vma->vm_start); | |
4198 | ||
4199 | return -EAGAIN; | |
4200 | } | |
4201 | ||
4202 | static void walk_mm(struct lruvec *lruvec, struct mm_struct *mm, struct lru_gen_mm_walk *walk) | |
4203 | { | |
4204 | static const struct mm_walk_ops mm_walk_ops = { | |
4205 | .test_walk = should_skip_vma, | |
4206 | .p4d_entry = walk_pud_range, | |
4207 | }; | |
4208 | ||
4209 | int err; | |
4210 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
4211 | ||
4212 | walk->next_addr = FIRST_USER_ADDRESS; | |
4213 | ||
4214 | do { | |
4215 | err = -EBUSY; | |
4216 | ||
4217 | /* folio_update_gen() requires stable folio_memcg() */ | |
4218 | if (!mem_cgroup_trylock_pages(memcg)) | |
4219 | break; | |
4220 | ||
4221 | /* the caller might be holding the lock for write */ | |
4222 | if (mmap_read_trylock(mm)) { | |
4223 | err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk); | |
4224 | ||
4225 | mmap_read_unlock(mm); | |
4226 | } | |
4227 | ||
4228 | mem_cgroup_unlock_pages(); | |
4229 | ||
4230 | if (walk->batched) { | |
4231 | spin_lock_irq(&lruvec->lru_lock); | |
4232 | reset_batch_size(lruvec, walk); | |
4233 | spin_unlock_irq(&lruvec->lru_lock); | |
4234 | } | |
4235 | ||
4236 | cond_resched(); | |
4237 | } while (err == -EAGAIN); | |
4238 | } | |
4239 | ||
4240 | static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat) | |
4241 | { | |
4242 | struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk; | |
4243 | ||
4244 | if (pgdat && current_is_kswapd()) { | |
4245 | VM_WARN_ON_ONCE(walk); | |
4246 | ||
4247 | walk = &pgdat->mm_walk; | |
4248 | } else if (!pgdat && !walk) { | |
4249 | VM_WARN_ON_ONCE(current_is_kswapd()); | |
4250 | ||
4251 | walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN); | |
4252 | } | |
4253 | ||
4254 | current->reclaim_state->mm_walk = walk; | |
4255 | ||
4256 | return walk; | |
4257 | } | |
4258 | ||
4259 | static void clear_mm_walk(void) | |
4260 | { | |
4261 | struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk; | |
4262 | ||
4263 | VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages))); | |
4264 | VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats))); | |
4265 | ||
4266 | current->reclaim_state->mm_walk = NULL; | |
4267 | ||
4268 | if (!current_is_kswapd()) | |
4269 | kfree(walk); | |
4270 | } | |
4271 | ||
d6c3af7d | 4272 | static bool inc_min_seq(struct lruvec *lruvec, int type, bool can_swap) |
ac35a490 | 4273 | { |
d6c3af7d YZ |
4274 | int zone; |
4275 | int remaining = MAX_LRU_BATCH; | |
391655fe | 4276 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
d6c3af7d YZ |
4277 | int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]); |
4278 | ||
4279 | if (type == LRU_GEN_ANON && !can_swap) | |
4280 | goto done; | |
4281 | ||
4282 | /* prevent cold/hot inversion if force_scan is true */ | |
4283 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
6df1b221 | 4284 | struct list_head *head = &lrugen->folios[old_gen][type][zone]; |
d6c3af7d YZ |
4285 | |
4286 | while (!list_empty(head)) { | |
4287 | struct folio *folio = lru_to_folio(head); | |
4288 | ||
4289 | VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio); | |
4290 | VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio); | |
4291 | VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio); | |
4292 | VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio); | |
ac35a490 | 4293 | |
d6c3af7d | 4294 | new_gen = folio_inc_gen(lruvec, folio, false); |
6df1b221 | 4295 | list_move_tail(&folio->lru, &lrugen->folios[new_gen][type][zone]); |
d6c3af7d YZ |
4296 | |
4297 | if (!--remaining) | |
4298 | return false; | |
4299 | } | |
4300 | } | |
4301 | done: | |
ac35a490 YZ |
4302 | reset_ctrl_pos(lruvec, type, true); |
4303 | WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1); | |
d6c3af7d YZ |
4304 | |
4305 | return true; | |
ac35a490 YZ |
4306 | } |
4307 | ||
4308 | static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap) | |
4309 | { | |
4310 | int gen, type, zone; | |
4311 | bool success = false; | |
391655fe | 4312 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
ac35a490 YZ |
4313 | DEFINE_MIN_SEQ(lruvec); |
4314 | ||
4315 | VM_WARN_ON_ONCE(!seq_is_valid(lruvec)); | |
4316 | ||
4317 | /* find the oldest populated generation */ | |
4318 | for (type = !can_swap; type < ANON_AND_FILE; type++) { | |
4319 | while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) { | |
4320 | gen = lru_gen_from_seq(min_seq[type]); | |
4321 | ||
4322 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
6df1b221 | 4323 | if (!list_empty(&lrugen->folios[gen][type][zone])) |
ac35a490 YZ |
4324 | goto next; |
4325 | } | |
4326 | ||
4327 | min_seq[type]++; | |
4328 | } | |
4329 | next: | |
4330 | ; | |
4331 | } | |
4332 | ||
391655fe | 4333 | /* see the comment on lru_gen_folio */ |
ac35a490 YZ |
4334 | if (can_swap) { |
4335 | min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]); | |
4336 | min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]); | |
4337 | } | |
4338 | ||
4339 | for (type = !can_swap; type < ANON_AND_FILE; type++) { | |
4340 | if (min_seq[type] == lrugen->min_seq[type]) | |
4341 | continue; | |
4342 | ||
4343 | reset_ctrl_pos(lruvec, type, true); | |
4344 | WRITE_ONCE(lrugen->min_seq[type], min_seq[type]); | |
4345 | success = true; | |
4346 | } | |
4347 | ||
4348 | return success; | |
4349 | } | |
4350 | ||
d6c3af7d | 4351 | static void inc_max_seq(struct lruvec *lruvec, bool can_swap, bool force_scan) |
ac35a490 YZ |
4352 | { |
4353 | int prev, next; | |
4354 | int type, zone; | |
391655fe | 4355 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
ac35a490 YZ |
4356 | |
4357 | spin_lock_irq(&lruvec->lru_lock); | |
4358 | ||
4359 | VM_WARN_ON_ONCE(!seq_is_valid(lruvec)); | |
4360 | ||
ac35a490 YZ |
4361 | for (type = ANON_AND_FILE - 1; type >= 0; type--) { |
4362 | if (get_nr_gens(lruvec, type) != MAX_NR_GENS) | |
4363 | continue; | |
4364 | ||
d6c3af7d | 4365 | VM_WARN_ON_ONCE(!force_scan && (type == LRU_GEN_FILE || can_swap)); |
ac35a490 | 4366 | |
d6c3af7d YZ |
4367 | while (!inc_min_seq(lruvec, type, can_swap)) { |
4368 | spin_unlock_irq(&lruvec->lru_lock); | |
4369 | cond_resched(); | |
4370 | spin_lock_irq(&lruvec->lru_lock); | |
4371 | } | |
ac35a490 YZ |
4372 | } |
4373 | ||
4374 | /* | |
4375 | * Update the active/inactive LRU sizes for compatibility. Both sides of | |
4376 | * the current max_seq need to be covered, since max_seq+1 can overlap | |
4377 | * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do | |
4378 | * overlap, cold/hot inversion happens. | |
4379 | */ | |
4380 | prev = lru_gen_from_seq(lrugen->max_seq - 1); | |
4381 | next = lru_gen_from_seq(lrugen->max_seq + 1); | |
4382 | ||
4383 | for (type = 0; type < ANON_AND_FILE; type++) { | |
4384 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
4385 | enum lru_list lru = type * LRU_INACTIVE_FILE; | |
4386 | long delta = lrugen->nr_pages[prev][type][zone] - | |
4387 | lrugen->nr_pages[next][type][zone]; | |
4388 | ||
4389 | if (!delta) | |
4390 | continue; | |
4391 | ||
4392 | __update_lru_size(lruvec, lru, zone, delta); | |
4393 | __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta); | |
4394 | } | |
4395 | } | |
4396 | ||
4397 | for (type = 0; type < ANON_AND_FILE; type++) | |
4398 | reset_ctrl_pos(lruvec, type, false); | |
4399 | ||
1332a809 | 4400 | WRITE_ONCE(lrugen->timestamps[next], jiffies); |
ac35a490 YZ |
4401 | /* make sure preceding modifications appear */ |
4402 | smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1); | |
bd74fdae | 4403 | |
ac35a490 YZ |
4404 | spin_unlock_irq(&lruvec->lru_lock); |
4405 | } | |
4406 | ||
bd74fdae | 4407 | static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq, |
d6c3af7d | 4408 | struct scan_control *sc, bool can_swap, bool force_scan) |
bd74fdae YZ |
4409 | { |
4410 | bool success; | |
4411 | struct lru_gen_mm_walk *walk; | |
4412 | struct mm_struct *mm = NULL; | |
391655fe | 4413 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
bd74fdae YZ |
4414 | |
4415 | VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq)); | |
4416 | ||
4417 | /* see the comment in iterate_mm_list() */ | |
4418 | if (max_seq <= READ_ONCE(lruvec->mm_state.seq)) { | |
4419 | success = false; | |
4420 | goto done; | |
4421 | } | |
4422 | ||
4423 | /* | |
4424 | * If the hardware doesn't automatically set the accessed bit, fallback | |
4425 | * to lru_gen_look_around(), which only clears the accessed bit in a | |
4426 | * handful of PTEs. Spreading the work out over a period of time usually | |
4427 | * is less efficient, but it avoids bursty page faults. | |
4428 | */ | |
d6c3af7d | 4429 | if (!force_scan && !(arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))) { |
bd74fdae YZ |
4430 | success = iterate_mm_list_nowalk(lruvec, max_seq); |
4431 | goto done; | |
4432 | } | |
4433 | ||
4434 | walk = set_mm_walk(NULL); | |
4435 | if (!walk) { | |
4436 | success = iterate_mm_list_nowalk(lruvec, max_seq); | |
4437 | goto done; | |
4438 | } | |
4439 | ||
4440 | walk->lruvec = lruvec; | |
4441 | walk->max_seq = max_seq; | |
4442 | walk->can_swap = can_swap; | |
d6c3af7d | 4443 | walk->force_scan = force_scan; |
bd74fdae YZ |
4444 | |
4445 | do { | |
4446 | success = iterate_mm_list(lruvec, walk, &mm); | |
4447 | if (mm) | |
4448 | walk_mm(lruvec, mm, walk); | |
4449 | ||
4450 | cond_resched(); | |
4451 | } while (mm); | |
4452 | done: | |
4453 | if (!success) { | |
4454 | if (sc->priority <= DEF_PRIORITY - 2) | |
4455 | wait_event_killable(lruvec->mm_state.wait, | |
4456 | max_seq < READ_ONCE(lrugen->max_seq)); | |
4457 | ||
4458 | return max_seq < READ_ONCE(lrugen->max_seq); | |
4459 | } | |
4460 | ||
4461 | VM_WARN_ON_ONCE(max_seq != READ_ONCE(lrugen->max_seq)); | |
4462 | ||
d6c3af7d | 4463 | inc_max_seq(lruvec, can_swap, force_scan); |
bd74fdae YZ |
4464 | /* either this sees any waiters or they will see updated max_seq */ |
4465 | if (wq_has_sleeper(&lruvec->mm_state.wait)) | |
4466 | wake_up_all(&lruvec->mm_state.wait); | |
4467 | ||
bd74fdae YZ |
4468 | return true; |
4469 | } | |
4470 | ||
ac35a490 YZ |
4471 | static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq, unsigned long *min_seq, |
4472 | struct scan_control *sc, bool can_swap, unsigned long *nr_to_scan) | |
4473 | { | |
4474 | int gen, type, zone; | |
4475 | unsigned long old = 0; | |
4476 | unsigned long young = 0; | |
4477 | unsigned long total = 0; | |
391655fe | 4478 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
ac35a490 YZ |
4479 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); |
4480 | ||
4481 | for (type = !can_swap; type < ANON_AND_FILE; type++) { | |
4482 | unsigned long seq; | |
4483 | ||
4484 | for (seq = min_seq[type]; seq <= max_seq; seq++) { | |
4485 | unsigned long size = 0; | |
4486 | ||
4487 | gen = lru_gen_from_seq(seq); | |
4488 | ||
4489 | for (zone = 0; zone < MAX_NR_ZONES; zone++) | |
4490 | size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L); | |
4491 | ||
4492 | total += size; | |
4493 | if (seq == max_seq) | |
4494 | young += size; | |
4495 | else if (seq + MIN_NR_GENS == max_seq) | |
4496 | old += size; | |
4497 | } | |
4498 | } | |
4499 | ||
4500 | /* try to scrape all its memory if this memcg was deleted */ | |
4501 | *nr_to_scan = mem_cgroup_online(memcg) ? (total >> sc->priority) : total; | |
4502 | ||
4503 | /* | |
4504 | * The aging tries to be lazy to reduce the overhead, while the eviction | |
4505 | * stalls when the number of generations reaches MIN_NR_GENS. Hence, the | |
4506 | * ideal number of generations is MIN_NR_GENS+1. | |
4507 | */ | |
4508 | if (min_seq[!can_swap] + MIN_NR_GENS > max_seq) | |
4509 | return true; | |
4510 | if (min_seq[!can_swap] + MIN_NR_GENS < max_seq) | |
4511 | return false; | |
4512 | ||
4513 | /* | |
4514 | * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1) | |
4515 | * of the total number of pages for each generation. A reasonable range | |
4516 | * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The | |
4517 | * aging cares about the upper bound of hot pages, while the eviction | |
4518 | * cares about the lower bound of cold pages. | |
4519 | */ | |
4520 | if (young * MIN_NR_GENS > total) | |
4521 | return true; | |
4522 | if (old * (MIN_NR_GENS + 2) < total) | |
4523 | return true; | |
4524 | ||
4525 | return false; | |
4526 | } | |
4527 | ||
1332a809 | 4528 | static bool age_lruvec(struct lruvec *lruvec, struct scan_control *sc, unsigned long min_ttl) |
ac35a490 YZ |
4529 | { |
4530 | bool need_aging; | |
4531 | unsigned long nr_to_scan; | |
4532 | int swappiness = get_swappiness(lruvec, sc); | |
4533 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
4534 | DEFINE_MAX_SEQ(lruvec); | |
4535 | DEFINE_MIN_SEQ(lruvec); | |
4536 | ||
4537 | VM_WARN_ON_ONCE(sc->memcg_low_reclaim); | |
4538 | ||
4539 | mem_cgroup_calculate_protection(NULL, memcg); | |
4540 | ||
adb82130 | 4541 | if (mem_cgroup_below_min(NULL, memcg)) |
1332a809 | 4542 | return false; |
ac35a490 YZ |
4543 | |
4544 | need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, swappiness, &nr_to_scan); | |
1332a809 YZ |
4545 | |
4546 | if (min_ttl) { | |
4547 | int gen = lru_gen_from_seq(min_seq[LRU_GEN_FILE]); | |
4548 | unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]); | |
4549 | ||
4550 | if (time_is_after_jiffies(birth + min_ttl)) | |
4551 | return false; | |
4552 | ||
4553 | /* the size is likely too small to be helpful */ | |
4554 | if (!nr_to_scan && sc->priority != DEF_PRIORITY) | |
4555 | return false; | |
4556 | } | |
4557 | ||
ac35a490 | 4558 | if (need_aging) |
d6c3af7d | 4559 | try_to_inc_max_seq(lruvec, max_seq, sc, swappiness, false); |
1332a809 YZ |
4560 | |
4561 | return true; | |
ac35a490 YZ |
4562 | } |
4563 | ||
1332a809 YZ |
4564 | /* to protect the working set of the last N jiffies */ |
4565 | static unsigned long lru_gen_min_ttl __read_mostly; | |
4566 | ||
ac35a490 YZ |
4567 | static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc) |
4568 | { | |
4569 | struct mem_cgroup *memcg; | |
1332a809 YZ |
4570 | bool success = false; |
4571 | unsigned long min_ttl = READ_ONCE(lru_gen_min_ttl); | |
ac35a490 YZ |
4572 | |
4573 | VM_WARN_ON_ONCE(!current_is_kswapd()); | |
4574 | ||
f76c8337 YZ |
4575 | sc->last_reclaimed = sc->nr_reclaimed; |
4576 | ||
4577 | /* | |
4578 | * To reduce the chance of going into the aging path, which can be | |
4579 | * costly, optimistically skip it if the flag below was cleared in the | |
4580 | * eviction path. This improves the overall performance when multiple | |
4581 | * memcgs are available. | |
4582 | */ | |
4583 | if (!sc->memcgs_need_aging) { | |
4584 | sc->memcgs_need_aging = true; | |
4585 | return; | |
4586 | } | |
4587 | ||
bd74fdae YZ |
4588 | set_mm_walk(pgdat); |
4589 | ||
ac35a490 YZ |
4590 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
4591 | do { | |
4592 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); | |
4593 | ||
1332a809 YZ |
4594 | if (age_lruvec(lruvec, sc, min_ttl)) |
4595 | success = true; | |
ac35a490 YZ |
4596 | |
4597 | cond_resched(); | |
4598 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL))); | |
bd74fdae YZ |
4599 | |
4600 | clear_mm_walk(); | |
1332a809 YZ |
4601 | |
4602 | /* check the order to exclude compaction-induced reclaim */ | |
4603 | if (success || !min_ttl || sc->order) | |
4604 | return; | |
4605 | ||
4606 | /* | |
4607 | * The main goal is to OOM kill if every generation from all memcgs is | |
4608 | * younger than min_ttl. However, another possibility is all memcgs are | |
4609 | * either below min or empty. | |
4610 | */ | |
4611 | if (mutex_trylock(&oom_lock)) { | |
4612 | struct oom_control oc = { | |
4613 | .gfp_mask = sc->gfp_mask, | |
4614 | }; | |
4615 | ||
4616 | out_of_memory(&oc); | |
4617 | ||
4618 | mutex_unlock(&oom_lock); | |
4619 | } | |
ac35a490 YZ |
4620 | } |
4621 | ||
018ee47f | 4622 | /* |
49fd9b6d | 4623 | * This function exploits spatial locality when shrink_folio_list() walks the |
bd74fdae YZ |
4624 | * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If |
4625 | * the scan was done cacheline efficiently, it adds the PMD entry pointing to | |
4626 | * the PTE table to the Bloom filter. This forms a feedback loop between the | |
4627 | * eviction and the aging. | |
018ee47f YZ |
4628 | */ |
4629 | void lru_gen_look_around(struct page_vma_mapped_walk *pvmw) | |
4630 | { | |
4631 | int i; | |
4632 | pte_t *pte; | |
4633 | unsigned long start; | |
4634 | unsigned long end; | |
4635 | unsigned long addr; | |
bd74fdae YZ |
4636 | struct lru_gen_mm_walk *walk; |
4637 | int young = 0; | |
018ee47f YZ |
4638 | unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {}; |
4639 | struct folio *folio = pfn_folio(pvmw->pfn); | |
4640 | struct mem_cgroup *memcg = folio_memcg(folio); | |
4641 | struct pglist_data *pgdat = folio_pgdat(folio); | |
4642 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); | |
4643 | DEFINE_MAX_SEQ(lruvec); | |
4644 | int old_gen, new_gen = lru_gen_from_seq(max_seq); | |
4645 | ||
4646 | lockdep_assert_held(pvmw->ptl); | |
4647 | VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio); | |
4648 | ||
4649 | if (spin_is_contended(pvmw->ptl)) | |
4650 | return; | |
4651 | ||
bd74fdae YZ |
4652 | /* avoid taking the LRU lock under the PTL when possible */ |
4653 | walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL; | |
4654 | ||
018ee47f YZ |
4655 | start = max(pvmw->address & PMD_MASK, pvmw->vma->vm_start); |
4656 | end = min(pvmw->address | ~PMD_MASK, pvmw->vma->vm_end - 1) + 1; | |
4657 | ||
4658 | if (end - start > MIN_LRU_BATCH * PAGE_SIZE) { | |
4659 | if (pvmw->address - start < MIN_LRU_BATCH * PAGE_SIZE / 2) | |
4660 | end = start + MIN_LRU_BATCH * PAGE_SIZE; | |
4661 | else if (end - pvmw->address < MIN_LRU_BATCH * PAGE_SIZE / 2) | |
4662 | start = end - MIN_LRU_BATCH * PAGE_SIZE; | |
4663 | else { | |
4664 | start = pvmw->address - MIN_LRU_BATCH * PAGE_SIZE / 2; | |
4665 | end = pvmw->address + MIN_LRU_BATCH * PAGE_SIZE / 2; | |
4666 | } | |
4667 | } | |
4668 | ||
4669 | pte = pvmw->pte - (pvmw->address - start) / PAGE_SIZE; | |
4670 | ||
4671 | rcu_read_lock(); | |
4672 | arch_enter_lazy_mmu_mode(); | |
4673 | ||
4674 | for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) { | |
4675 | unsigned long pfn; | |
4676 | ||
4677 | pfn = get_pte_pfn(pte[i], pvmw->vma, addr); | |
4678 | if (pfn == -1) | |
4679 | continue; | |
4680 | ||
4681 | if (!pte_young(pte[i])) | |
4682 | continue; | |
4683 | ||
bd74fdae | 4684 | folio = get_pfn_folio(pfn, memcg, pgdat, !walk || walk->can_swap); |
018ee47f YZ |
4685 | if (!folio) |
4686 | continue; | |
4687 | ||
4688 | if (!ptep_test_and_clear_young(pvmw->vma, addr, pte + i)) | |
4689 | VM_WARN_ON_ONCE(true); | |
4690 | ||
bd74fdae YZ |
4691 | young++; |
4692 | ||
018ee47f YZ |
4693 | if (pte_dirty(pte[i]) && !folio_test_dirty(folio) && |
4694 | !(folio_test_anon(folio) && folio_test_swapbacked(folio) && | |
4695 | !folio_test_swapcache(folio))) | |
4696 | folio_mark_dirty(folio); | |
4697 | ||
4698 | old_gen = folio_lru_gen(folio); | |
4699 | if (old_gen < 0) | |
4700 | folio_set_referenced(folio); | |
4701 | else if (old_gen != new_gen) | |
4702 | __set_bit(i, bitmap); | |
4703 | } | |
4704 | ||
4705 | arch_leave_lazy_mmu_mode(); | |
4706 | rcu_read_unlock(); | |
4707 | ||
bd74fdae YZ |
4708 | /* feedback from rmap walkers to page table walkers */ |
4709 | if (suitable_to_scan(i, young)) | |
4710 | update_bloom_filter(lruvec, max_seq, pvmw->pmd); | |
4711 | ||
4712 | if (!walk && bitmap_weight(bitmap, MIN_LRU_BATCH) < PAGEVEC_SIZE) { | |
018ee47f YZ |
4713 | for_each_set_bit(i, bitmap, MIN_LRU_BATCH) { |
4714 | folio = pfn_folio(pte_pfn(pte[i])); | |
4715 | folio_activate(folio); | |
4716 | } | |
4717 | return; | |
4718 | } | |
4719 | ||
4720 | /* folio_update_gen() requires stable folio_memcg() */ | |
4721 | if (!mem_cgroup_trylock_pages(memcg)) | |
4722 | return; | |
4723 | ||
bd74fdae YZ |
4724 | if (!walk) { |
4725 | spin_lock_irq(&lruvec->lru_lock); | |
4726 | new_gen = lru_gen_from_seq(lruvec->lrugen.max_seq); | |
4727 | } | |
018ee47f YZ |
4728 | |
4729 | for_each_set_bit(i, bitmap, MIN_LRU_BATCH) { | |
4730 | folio = pfn_folio(pte_pfn(pte[i])); | |
4731 | if (folio_memcg_rcu(folio) != memcg) | |
4732 | continue; | |
4733 | ||
4734 | old_gen = folio_update_gen(folio, new_gen); | |
4735 | if (old_gen < 0 || old_gen == new_gen) | |
4736 | continue; | |
4737 | ||
bd74fdae YZ |
4738 | if (walk) |
4739 | update_batch_size(walk, folio, old_gen, new_gen); | |
4740 | else | |
4741 | lru_gen_update_size(lruvec, folio, old_gen, new_gen); | |
018ee47f YZ |
4742 | } |
4743 | ||
bd74fdae YZ |
4744 | if (!walk) |
4745 | spin_unlock_irq(&lruvec->lru_lock); | |
018ee47f YZ |
4746 | |
4747 | mem_cgroup_unlock_pages(); | |
4748 | } | |
4749 | ||
ac35a490 YZ |
4750 | /****************************************************************************** |
4751 | * the eviction | |
4752 | ******************************************************************************/ | |
4753 | ||
4754 | static bool sort_folio(struct lruvec *lruvec, struct folio *folio, int tier_idx) | |
4755 | { | |
4756 | bool success; | |
4757 | int gen = folio_lru_gen(folio); | |
4758 | int type = folio_is_file_lru(folio); | |
4759 | int zone = folio_zonenum(folio); | |
4760 | int delta = folio_nr_pages(folio); | |
4761 | int refs = folio_lru_refs(folio); | |
4762 | int tier = lru_tier_from_refs(refs); | |
391655fe | 4763 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
ac35a490 YZ |
4764 | |
4765 | VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio); | |
4766 | ||
4767 | /* unevictable */ | |
4768 | if (!folio_evictable(folio)) { | |
4769 | success = lru_gen_del_folio(lruvec, folio, true); | |
4770 | VM_WARN_ON_ONCE_FOLIO(!success, folio); | |
4771 | folio_set_unevictable(folio); | |
4772 | lruvec_add_folio(lruvec, folio); | |
4773 | __count_vm_events(UNEVICTABLE_PGCULLED, delta); | |
4774 | return true; | |
4775 | } | |
4776 | ||
4777 | /* dirty lazyfree */ | |
4778 | if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) { | |
4779 | success = lru_gen_del_folio(lruvec, folio, true); | |
4780 | VM_WARN_ON_ONCE_FOLIO(!success, folio); | |
4781 | folio_set_swapbacked(folio); | |
4782 | lruvec_add_folio_tail(lruvec, folio); | |
4783 | return true; | |
4784 | } | |
4785 | ||
018ee47f YZ |
4786 | /* promoted */ |
4787 | if (gen != lru_gen_from_seq(lrugen->min_seq[type])) { | |
6df1b221 | 4788 | list_move(&folio->lru, &lrugen->folios[gen][type][zone]); |
018ee47f YZ |
4789 | return true; |
4790 | } | |
4791 | ||
ac35a490 YZ |
4792 | /* protected */ |
4793 | if (tier > tier_idx) { | |
4794 | int hist = lru_hist_from_seq(lrugen->min_seq[type]); | |
4795 | ||
4796 | gen = folio_inc_gen(lruvec, folio, false); | |
6df1b221 | 4797 | list_move_tail(&folio->lru, &lrugen->folios[gen][type][zone]); |
ac35a490 YZ |
4798 | |
4799 | WRITE_ONCE(lrugen->protected[hist][type][tier - 1], | |
4800 | lrugen->protected[hist][type][tier - 1] + delta); | |
4801 | __mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta); | |
4802 | return true; | |
4803 | } | |
4804 | ||
4805 | /* waiting for writeback */ | |
4806 | if (folio_test_locked(folio) || folio_test_writeback(folio) || | |
4807 | (type == LRU_GEN_FILE && folio_test_dirty(folio))) { | |
4808 | gen = folio_inc_gen(lruvec, folio, true); | |
6df1b221 | 4809 | list_move(&folio->lru, &lrugen->folios[gen][type][zone]); |
ac35a490 YZ |
4810 | return true; |
4811 | } | |
4812 | ||
4813 | return false; | |
4814 | } | |
4815 | ||
4816 | static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc) | |
4817 | { | |
4818 | bool success; | |
4819 | ||
4820 | /* unmapping inhibited */ | |
4821 | if (!sc->may_unmap && folio_mapped(folio)) | |
4822 | return false; | |
4823 | ||
4824 | /* swapping inhibited */ | |
4825 | if (!(sc->may_writepage && (sc->gfp_mask & __GFP_IO)) && | |
4826 | (folio_test_dirty(folio) || | |
4827 | (folio_test_anon(folio) && !folio_test_swapcache(folio)))) | |
4828 | return false; | |
4829 | ||
4830 | /* raced with release_pages() */ | |
4831 | if (!folio_try_get(folio)) | |
4832 | return false; | |
4833 | ||
4834 | /* raced with another isolation */ | |
4835 | if (!folio_test_clear_lru(folio)) { | |
4836 | folio_put(folio); | |
4837 | return false; | |
4838 | } | |
4839 | ||
4840 | /* see the comment on MAX_NR_TIERS */ | |
4841 | if (!folio_test_referenced(folio)) | |
4842 | set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0); | |
4843 | ||
49fd9b6d | 4844 | /* for shrink_folio_list() */ |
ac35a490 YZ |
4845 | folio_clear_reclaim(folio); |
4846 | folio_clear_referenced(folio); | |
4847 | ||
4848 | success = lru_gen_del_folio(lruvec, folio, true); | |
4849 | VM_WARN_ON_ONCE_FOLIO(!success, folio); | |
4850 | ||
4851 | return true; | |
4852 | } | |
4853 | ||
4854 | static int scan_folios(struct lruvec *lruvec, struct scan_control *sc, | |
4855 | int type, int tier, struct list_head *list) | |
4856 | { | |
4857 | int gen, zone; | |
4858 | enum vm_event_item item; | |
4859 | int sorted = 0; | |
4860 | int scanned = 0; | |
4861 | int isolated = 0; | |
4862 | int remaining = MAX_LRU_BATCH; | |
391655fe | 4863 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
ac35a490 YZ |
4864 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); |
4865 | ||
4866 | VM_WARN_ON_ONCE(!list_empty(list)); | |
4867 | ||
4868 | if (get_nr_gens(lruvec, type) == MIN_NR_GENS) | |
4869 | return 0; | |
4870 | ||
4871 | gen = lru_gen_from_seq(lrugen->min_seq[type]); | |
4872 | ||
4873 | for (zone = sc->reclaim_idx; zone >= 0; zone--) { | |
4874 | LIST_HEAD(moved); | |
4875 | int skipped = 0; | |
6df1b221 | 4876 | struct list_head *head = &lrugen->folios[gen][type][zone]; |
ac35a490 YZ |
4877 | |
4878 | while (!list_empty(head)) { | |
4879 | struct folio *folio = lru_to_folio(head); | |
4880 | int delta = folio_nr_pages(folio); | |
4881 | ||
4882 | VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio); | |
4883 | VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio); | |
4884 | VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio); | |
4885 | VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio); | |
4886 | ||
4887 | scanned += delta; | |
4888 | ||
4889 | if (sort_folio(lruvec, folio, tier)) | |
4890 | sorted += delta; | |
4891 | else if (isolate_folio(lruvec, folio, sc)) { | |
4892 | list_add(&folio->lru, list); | |
4893 | isolated += delta; | |
4894 | } else { | |
4895 | list_move(&folio->lru, &moved); | |
4896 | skipped += delta; | |
4897 | } | |
4898 | ||
4899 | if (!--remaining || max(isolated, skipped) >= MIN_LRU_BATCH) | |
4900 | break; | |
4901 | } | |
4902 | ||
4903 | if (skipped) { | |
4904 | list_splice(&moved, head); | |
4905 | __count_zid_vm_events(PGSCAN_SKIP, zone, skipped); | |
4906 | } | |
4907 | ||
4908 | if (!remaining || isolated >= MIN_LRU_BATCH) | |
4909 | break; | |
4910 | } | |
4911 | ||
57e9cc50 | 4912 | item = PGSCAN_KSWAPD + reclaimer_offset(); |
ac35a490 YZ |
4913 | if (!cgroup_reclaim(sc)) { |
4914 | __count_vm_events(item, isolated); | |
4915 | __count_vm_events(PGREFILL, sorted); | |
4916 | } | |
4917 | __count_memcg_events(memcg, item, isolated); | |
4918 | __count_memcg_events(memcg, PGREFILL, sorted); | |
4919 | __count_vm_events(PGSCAN_ANON + type, isolated); | |
4920 | ||
4921 | /* | |
4922 | * There might not be eligible pages due to reclaim_idx, may_unmap and | |
4923 | * may_writepage. Check the remaining to prevent livelock if it's not | |
4924 | * making progress. | |
4925 | */ | |
4926 | return isolated || !remaining ? scanned : 0; | |
4927 | } | |
4928 | ||
4929 | static int get_tier_idx(struct lruvec *lruvec, int type) | |
4930 | { | |
4931 | int tier; | |
4932 | struct ctrl_pos sp, pv; | |
4933 | ||
4934 | /* | |
4935 | * To leave a margin for fluctuations, use a larger gain factor (1:2). | |
4936 | * This value is chosen because any other tier would have at least twice | |
4937 | * as many refaults as the first tier. | |
4938 | */ | |
4939 | read_ctrl_pos(lruvec, type, 0, 1, &sp); | |
4940 | for (tier = 1; tier < MAX_NR_TIERS; tier++) { | |
4941 | read_ctrl_pos(lruvec, type, tier, 2, &pv); | |
4942 | if (!positive_ctrl_err(&sp, &pv)) | |
4943 | break; | |
4944 | } | |
4945 | ||
4946 | return tier - 1; | |
4947 | } | |
4948 | ||
4949 | static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx) | |
4950 | { | |
4951 | int type, tier; | |
4952 | struct ctrl_pos sp, pv; | |
4953 | int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness }; | |
4954 | ||
4955 | /* | |
4956 | * Compare the first tier of anon with that of file to determine which | |
4957 | * type to scan. Also need to compare other tiers of the selected type | |
4958 | * with the first tier of the other type to determine the last tier (of | |
4959 | * the selected type) to evict. | |
4960 | */ | |
4961 | read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp); | |
4962 | read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv); | |
4963 | type = positive_ctrl_err(&sp, &pv); | |
4964 | ||
4965 | read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp); | |
4966 | for (tier = 1; tier < MAX_NR_TIERS; tier++) { | |
4967 | read_ctrl_pos(lruvec, type, tier, gain[type], &pv); | |
4968 | if (!positive_ctrl_err(&sp, &pv)) | |
4969 | break; | |
4970 | } | |
4971 | ||
4972 | *tier_idx = tier - 1; | |
4973 | ||
4974 | return type; | |
4975 | } | |
4976 | ||
4977 | static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness, | |
4978 | int *type_scanned, struct list_head *list) | |
4979 | { | |
4980 | int i; | |
4981 | int type; | |
4982 | int scanned; | |
4983 | int tier = -1; | |
4984 | DEFINE_MIN_SEQ(lruvec); | |
4985 | ||
4986 | /* | |
4987 | * Try to make the obvious choice first. When anon and file are both | |
4988 | * available from the same generation, interpret swappiness 1 as file | |
4989 | * first and 200 as anon first. | |
4990 | */ | |
4991 | if (!swappiness) | |
4992 | type = LRU_GEN_FILE; | |
4993 | else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE]) | |
4994 | type = LRU_GEN_ANON; | |
4995 | else if (swappiness == 1) | |
4996 | type = LRU_GEN_FILE; | |
4997 | else if (swappiness == 200) | |
4998 | type = LRU_GEN_ANON; | |
4999 | else | |
5000 | type = get_type_to_scan(lruvec, swappiness, &tier); | |
5001 | ||
5002 | for (i = !swappiness; i < ANON_AND_FILE; i++) { | |
5003 | if (tier < 0) | |
5004 | tier = get_tier_idx(lruvec, type); | |
5005 | ||
5006 | scanned = scan_folios(lruvec, sc, type, tier, list); | |
5007 | if (scanned) | |
5008 | break; | |
5009 | ||
5010 | type = !type; | |
5011 | tier = -1; | |
5012 | } | |
5013 | ||
5014 | *type_scanned = type; | |
5015 | ||
5016 | return scanned; | |
5017 | } | |
5018 | ||
a579086c | 5019 | static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness) |
ac35a490 YZ |
5020 | { |
5021 | int type; | |
5022 | int scanned; | |
5023 | int reclaimed; | |
5024 | LIST_HEAD(list); | |
359a5e14 | 5025 | LIST_HEAD(clean); |
ac35a490 | 5026 | struct folio *folio; |
359a5e14 | 5027 | struct folio *next; |
ac35a490 YZ |
5028 | enum vm_event_item item; |
5029 | struct reclaim_stat stat; | |
bd74fdae | 5030 | struct lru_gen_mm_walk *walk; |
359a5e14 | 5031 | bool skip_retry = false; |
ac35a490 YZ |
5032 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); |
5033 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); | |
5034 | ||
5035 | spin_lock_irq(&lruvec->lru_lock); | |
5036 | ||
5037 | scanned = isolate_folios(lruvec, sc, swappiness, &type, &list); | |
5038 | ||
5039 | scanned += try_to_inc_min_seq(lruvec, swappiness); | |
5040 | ||
5041 | if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS) | |
5042 | scanned = 0; | |
5043 | ||
5044 | spin_unlock_irq(&lruvec->lru_lock); | |
5045 | ||
5046 | if (list_empty(&list)) | |
5047 | return scanned; | |
359a5e14 | 5048 | retry: |
49fd9b6d | 5049 | reclaimed = shrink_folio_list(&list, pgdat, sc, &stat, false); |
359a5e14 | 5050 | sc->nr_reclaimed += reclaimed; |
ac35a490 | 5051 | |
359a5e14 YZ |
5052 | list_for_each_entry_safe_reverse(folio, next, &list, lru) { |
5053 | if (!folio_evictable(folio)) { | |
5054 | list_del(&folio->lru); | |
5055 | folio_putback_lru(folio); | |
5056 | continue; | |
5057 | } | |
ac35a490 | 5058 | |
ac35a490 | 5059 | if (folio_test_reclaim(folio) && |
359a5e14 YZ |
5060 | (folio_test_dirty(folio) || folio_test_writeback(folio))) { |
5061 | /* restore LRU_REFS_FLAGS cleared by isolate_folio() */ | |
5062 | if (folio_test_workingset(folio)) | |
5063 | folio_set_referenced(folio); | |
5064 | continue; | |
5065 | } | |
5066 | ||
5067 | if (skip_retry || folio_test_active(folio) || folio_test_referenced(folio) || | |
5068 | folio_mapped(folio) || folio_test_locked(folio) || | |
5069 | folio_test_dirty(folio) || folio_test_writeback(folio)) { | |
5070 | /* don't add rejected folios to the oldest generation */ | |
5071 | set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, | |
5072 | BIT(PG_active)); | |
5073 | continue; | |
5074 | } | |
5075 | ||
5076 | /* retry folios that may have missed folio_rotate_reclaimable() */ | |
5077 | list_move(&folio->lru, &clean); | |
5078 | sc->nr_scanned -= folio_nr_pages(folio); | |
ac35a490 YZ |
5079 | } |
5080 | ||
5081 | spin_lock_irq(&lruvec->lru_lock); | |
5082 | ||
49fd9b6d | 5083 | move_folios_to_lru(lruvec, &list); |
ac35a490 | 5084 | |
bd74fdae YZ |
5085 | walk = current->reclaim_state->mm_walk; |
5086 | if (walk && walk->batched) | |
5087 | reset_batch_size(lruvec, walk); | |
5088 | ||
57e9cc50 | 5089 | item = PGSTEAL_KSWAPD + reclaimer_offset(); |
ac35a490 YZ |
5090 | if (!cgroup_reclaim(sc)) |
5091 | __count_vm_events(item, reclaimed); | |
5092 | __count_memcg_events(memcg, item, reclaimed); | |
5093 | __count_vm_events(PGSTEAL_ANON + type, reclaimed); | |
5094 | ||
5095 | spin_unlock_irq(&lruvec->lru_lock); | |
5096 | ||
5097 | mem_cgroup_uncharge_list(&list); | |
5098 | free_unref_page_list(&list); | |
5099 | ||
359a5e14 YZ |
5100 | INIT_LIST_HEAD(&list); |
5101 | list_splice_init(&clean, &list); | |
5102 | ||
5103 | if (!list_empty(&list)) { | |
5104 | skip_retry = true; | |
5105 | goto retry; | |
5106 | } | |
ac35a490 YZ |
5107 | |
5108 | return scanned; | |
5109 | } | |
5110 | ||
bd74fdae YZ |
5111 | /* |
5112 | * For future optimizations: | |
5113 | * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg | |
5114 | * reclaim. | |
5115 | */ | |
ac35a490 | 5116 | static unsigned long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, |
f76c8337 | 5117 | bool can_swap, bool *need_aging) |
ac35a490 | 5118 | { |
ac35a490 YZ |
5119 | unsigned long nr_to_scan; |
5120 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
5121 | DEFINE_MAX_SEQ(lruvec); | |
5122 | DEFINE_MIN_SEQ(lruvec); | |
5123 | ||
adb82130 YA |
5124 | if (mem_cgroup_below_min(sc->target_mem_cgroup, memcg) || |
5125 | (mem_cgroup_below_low(sc->target_mem_cgroup, memcg) && | |
5126 | !sc->memcg_low_reclaim)) | |
ac35a490 YZ |
5127 | return 0; |
5128 | ||
f76c8337 YZ |
5129 | *need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, can_swap, &nr_to_scan); |
5130 | if (!*need_aging) | |
ac35a490 YZ |
5131 | return nr_to_scan; |
5132 | ||
5133 | /* skip the aging path at the default priority */ | |
5134 | if (sc->priority == DEF_PRIORITY) | |
5135 | goto done; | |
5136 | ||
5137 | /* leave the work to lru_gen_age_node() */ | |
5138 | if (current_is_kswapd()) | |
5139 | return 0; | |
5140 | ||
d6c3af7d | 5141 | if (try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, false)) |
bd74fdae | 5142 | return nr_to_scan; |
ac35a490 YZ |
5143 | done: |
5144 | return min_seq[!can_swap] + MIN_NR_GENS <= max_seq ? nr_to_scan : 0; | |
5145 | } | |
5146 | ||
a579086c | 5147 | static unsigned long get_nr_to_reclaim(struct scan_control *sc) |
f76c8337 | 5148 | { |
a579086c YZ |
5149 | /* don't abort memcg reclaim to ensure fairness */ |
5150 | if (!global_reclaim(sc)) | |
5151 | return -1; | |
f76c8337 | 5152 | |
a579086c YZ |
5153 | /* discount the previous progress for kswapd */ |
5154 | if (current_is_kswapd()) | |
5155 | return sc->nr_to_reclaim + sc->last_reclaimed; | |
f76c8337 | 5156 | |
a579086c | 5157 | return max(sc->nr_to_reclaim, compact_gap(sc->order)); |
f76c8337 YZ |
5158 | } |
5159 | ||
ac35a490 YZ |
5160 | static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) |
5161 | { | |
5162 | struct blk_plug plug; | |
f76c8337 | 5163 | bool need_aging = false; |
ac35a490 | 5164 | unsigned long scanned = 0; |
f76c8337 | 5165 | unsigned long reclaimed = sc->nr_reclaimed; |
a579086c | 5166 | unsigned long nr_to_reclaim = get_nr_to_reclaim(sc); |
ac35a490 YZ |
5167 | |
5168 | lru_add_drain(); | |
5169 | ||
5170 | blk_start_plug(&plug); | |
5171 | ||
bd74fdae YZ |
5172 | set_mm_walk(lruvec_pgdat(lruvec)); |
5173 | ||
ac35a490 YZ |
5174 | while (true) { |
5175 | int delta; | |
5176 | int swappiness; | |
5177 | unsigned long nr_to_scan; | |
5178 | ||
5179 | if (sc->may_swap) | |
5180 | swappiness = get_swappiness(lruvec, sc); | |
5181 | else if (!cgroup_reclaim(sc) && get_swappiness(lruvec, sc)) | |
5182 | swappiness = 1; | |
5183 | else | |
5184 | swappiness = 0; | |
5185 | ||
f76c8337 | 5186 | nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness, &need_aging); |
ac35a490 | 5187 | if (!nr_to_scan) |
f76c8337 | 5188 | goto done; |
ac35a490 | 5189 | |
a579086c | 5190 | delta = evict_folios(lruvec, sc, swappiness); |
ac35a490 | 5191 | if (!delta) |
f76c8337 | 5192 | goto done; |
ac35a490 YZ |
5193 | |
5194 | scanned += delta; | |
5195 | if (scanned >= nr_to_scan) | |
5196 | break; | |
5197 | ||
a579086c | 5198 | if (sc->nr_reclaimed >= nr_to_reclaim) |
f76c8337 YZ |
5199 | break; |
5200 | ||
ac35a490 YZ |
5201 | cond_resched(); |
5202 | } | |
5203 | ||
f76c8337 YZ |
5204 | /* see the comment in lru_gen_age_node() */ |
5205 | if (sc->nr_reclaimed - reclaimed >= MIN_LRU_BATCH && !need_aging) | |
5206 | sc->memcgs_need_aging = false; | |
5207 | done: | |
bd74fdae YZ |
5208 | clear_mm_walk(); |
5209 | ||
ac35a490 YZ |
5210 | blk_finish_plug(&plug); |
5211 | } | |
5212 | ||
354ed597 YZ |
5213 | /****************************************************************************** |
5214 | * state change | |
5215 | ******************************************************************************/ | |
5216 | ||
5217 | static bool __maybe_unused state_is_valid(struct lruvec *lruvec) | |
5218 | { | |
391655fe | 5219 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
354ed597 YZ |
5220 | |
5221 | if (lrugen->enabled) { | |
5222 | enum lru_list lru; | |
5223 | ||
5224 | for_each_evictable_lru(lru) { | |
5225 | if (!list_empty(&lruvec->lists[lru])) | |
5226 | return false; | |
5227 | } | |
5228 | } else { | |
5229 | int gen, type, zone; | |
5230 | ||
5231 | for_each_gen_type_zone(gen, type, zone) { | |
6df1b221 | 5232 | if (!list_empty(&lrugen->folios[gen][type][zone])) |
354ed597 YZ |
5233 | return false; |
5234 | } | |
5235 | } | |
5236 | ||
5237 | return true; | |
5238 | } | |
5239 | ||
5240 | static bool fill_evictable(struct lruvec *lruvec) | |
5241 | { | |
5242 | enum lru_list lru; | |
5243 | int remaining = MAX_LRU_BATCH; | |
5244 | ||
5245 | for_each_evictable_lru(lru) { | |
5246 | int type = is_file_lru(lru); | |
5247 | bool active = is_active_lru(lru); | |
5248 | struct list_head *head = &lruvec->lists[lru]; | |
5249 | ||
5250 | while (!list_empty(head)) { | |
5251 | bool success; | |
5252 | struct folio *folio = lru_to_folio(head); | |
5253 | ||
5254 | VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio); | |
5255 | VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio); | |
5256 | VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio); | |
5257 | VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio); | |
5258 | ||
5259 | lruvec_del_folio(lruvec, folio); | |
5260 | success = lru_gen_add_folio(lruvec, folio, false); | |
5261 | VM_WARN_ON_ONCE(!success); | |
5262 | ||
5263 | if (!--remaining) | |
5264 | return false; | |
5265 | } | |
5266 | } | |
5267 | ||
5268 | return true; | |
5269 | } | |
5270 | ||
5271 | static bool drain_evictable(struct lruvec *lruvec) | |
5272 | { | |
5273 | int gen, type, zone; | |
5274 | int remaining = MAX_LRU_BATCH; | |
5275 | ||
5276 | for_each_gen_type_zone(gen, type, zone) { | |
6df1b221 | 5277 | struct list_head *head = &lruvec->lrugen.folios[gen][type][zone]; |
354ed597 YZ |
5278 | |
5279 | while (!list_empty(head)) { | |
5280 | bool success; | |
5281 | struct folio *folio = lru_to_folio(head); | |
5282 | ||
5283 | VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio); | |
5284 | VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio); | |
5285 | VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio); | |
5286 | VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio); | |
5287 | ||
5288 | success = lru_gen_del_folio(lruvec, folio, false); | |
5289 | VM_WARN_ON_ONCE(!success); | |
5290 | lruvec_add_folio(lruvec, folio); | |
5291 | ||
5292 | if (!--remaining) | |
5293 | return false; | |
5294 | } | |
5295 | } | |
5296 | ||
5297 | return true; | |
5298 | } | |
5299 | ||
5300 | static void lru_gen_change_state(bool enabled) | |
5301 | { | |
5302 | static DEFINE_MUTEX(state_mutex); | |
5303 | ||
5304 | struct mem_cgroup *memcg; | |
5305 | ||
5306 | cgroup_lock(); | |
5307 | cpus_read_lock(); | |
5308 | get_online_mems(); | |
5309 | mutex_lock(&state_mutex); | |
5310 | ||
5311 | if (enabled == lru_gen_enabled()) | |
5312 | goto unlock; | |
5313 | ||
5314 | if (enabled) | |
5315 | static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]); | |
5316 | else | |
5317 | static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]); | |
5318 | ||
5319 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
5320 | do { | |
5321 | int nid; | |
5322 | ||
5323 | for_each_node(nid) { | |
5324 | struct lruvec *lruvec = get_lruvec(memcg, nid); | |
5325 | ||
354ed597 YZ |
5326 | spin_lock_irq(&lruvec->lru_lock); |
5327 | ||
5328 | VM_WARN_ON_ONCE(!seq_is_valid(lruvec)); | |
5329 | VM_WARN_ON_ONCE(!state_is_valid(lruvec)); | |
5330 | ||
5331 | lruvec->lrugen.enabled = enabled; | |
5332 | ||
5333 | while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) { | |
5334 | spin_unlock_irq(&lruvec->lru_lock); | |
5335 | cond_resched(); | |
5336 | spin_lock_irq(&lruvec->lru_lock); | |
5337 | } | |
5338 | ||
5339 | spin_unlock_irq(&lruvec->lru_lock); | |
5340 | } | |
5341 | ||
5342 | cond_resched(); | |
5343 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL))); | |
5344 | unlock: | |
5345 | mutex_unlock(&state_mutex); | |
5346 | put_online_mems(); | |
5347 | cpus_read_unlock(); | |
5348 | cgroup_unlock(); | |
5349 | } | |
5350 | ||
5351 | /****************************************************************************** | |
5352 | * sysfs interface | |
5353 | ******************************************************************************/ | |
5354 | ||
1332a809 YZ |
5355 | static ssize_t show_min_ttl(struct kobject *kobj, struct kobj_attribute *attr, char *buf) |
5356 | { | |
5357 | return sprintf(buf, "%u\n", jiffies_to_msecs(READ_ONCE(lru_gen_min_ttl))); | |
5358 | } | |
5359 | ||
07017acb | 5360 | /* see Documentation/admin-guide/mm/multigen_lru.rst for details */ |
1332a809 YZ |
5361 | static ssize_t store_min_ttl(struct kobject *kobj, struct kobj_attribute *attr, |
5362 | const char *buf, size_t len) | |
5363 | { | |
5364 | unsigned int msecs; | |
5365 | ||
5366 | if (kstrtouint(buf, 0, &msecs)) | |
5367 | return -EINVAL; | |
5368 | ||
5369 | WRITE_ONCE(lru_gen_min_ttl, msecs_to_jiffies(msecs)); | |
5370 | ||
5371 | return len; | |
5372 | } | |
5373 | ||
5374 | static struct kobj_attribute lru_gen_min_ttl_attr = __ATTR( | |
5375 | min_ttl_ms, 0644, show_min_ttl, store_min_ttl | |
5376 | ); | |
5377 | ||
354ed597 YZ |
5378 | static ssize_t show_enabled(struct kobject *kobj, struct kobj_attribute *attr, char *buf) |
5379 | { | |
5380 | unsigned int caps = 0; | |
5381 | ||
5382 | if (get_cap(LRU_GEN_CORE)) | |
5383 | caps |= BIT(LRU_GEN_CORE); | |
5384 | ||
5385 | if (arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK)) | |
5386 | caps |= BIT(LRU_GEN_MM_WALK); | |
5387 | ||
4aaf269c | 5388 | if (arch_has_hw_nonleaf_pmd_young() && get_cap(LRU_GEN_NONLEAF_YOUNG)) |
354ed597 YZ |
5389 | caps |= BIT(LRU_GEN_NONLEAF_YOUNG); |
5390 | ||
8ef9c32a | 5391 | return sysfs_emit(buf, "0x%04x\n", caps); |
354ed597 YZ |
5392 | } |
5393 | ||
07017acb | 5394 | /* see Documentation/admin-guide/mm/multigen_lru.rst for details */ |
354ed597 YZ |
5395 | static ssize_t store_enabled(struct kobject *kobj, struct kobj_attribute *attr, |
5396 | const char *buf, size_t len) | |
5397 | { | |
5398 | int i; | |
5399 | unsigned int caps; | |
5400 | ||
5401 | if (tolower(*buf) == 'n') | |
5402 | caps = 0; | |
5403 | else if (tolower(*buf) == 'y') | |
5404 | caps = -1; | |
5405 | else if (kstrtouint(buf, 0, &caps)) | |
5406 | return -EINVAL; | |
5407 | ||
5408 | for (i = 0; i < NR_LRU_GEN_CAPS; i++) { | |
5409 | bool enabled = caps & BIT(i); | |
5410 | ||
5411 | if (i == LRU_GEN_CORE) | |
5412 | lru_gen_change_state(enabled); | |
5413 | else if (enabled) | |
5414 | static_branch_enable(&lru_gen_caps[i]); | |
5415 | else | |
5416 | static_branch_disable(&lru_gen_caps[i]); | |
5417 | } | |
5418 | ||
5419 | return len; | |
5420 | } | |
5421 | ||
5422 | static struct kobj_attribute lru_gen_enabled_attr = __ATTR( | |
5423 | enabled, 0644, show_enabled, store_enabled | |
5424 | ); | |
5425 | ||
5426 | static struct attribute *lru_gen_attrs[] = { | |
1332a809 | 5427 | &lru_gen_min_ttl_attr.attr, |
354ed597 YZ |
5428 | &lru_gen_enabled_attr.attr, |
5429 | NULL | |
5430 | }; | |
5431 | ||
5432 | static struct attribute_group lru_gen_attr_group = { | |
5433 | .name = "lru_gen", | |
5434 | .attrs = lru_gen_attrs, | |
5435 | }; | |
5436 | ||
d6c3af7d YZ |
5437 | /****************************************************************************** |
5438 | * debugfs interface | |
5439 | ******************************************************************************/ | |
5440 | ||
5441 | static void *lru_gen_seq_start(struct seq_file *m, loff_t *pos) | |
5442 | { | |
5443 | struct mem_cgroup *memcg; | |
5444 | loff_t nr_to_skip = *pos; | |
5445 | ||
5446 | m->private = kvmalloc(PATH_MAX, GFP_KERNEL); | |
5447 | if (!m->private) | |
5448 | return ERR_PTR(-ENOMEM); | |
5449 | ||
5450 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
5451 | do { | |
5452 | int nid; | |
5453 | ||
5454 | for_each_node_state(nid, N_MEMORY) { | |
5455 | if (!nr_to_skip--) | |
5456 | return get_lruvec(memcg, nid); | |
5457 | } | |
5458 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL))); | |
5459 | ||
5460 | return NULL; | |
5461 | } | |
5462 | ||
5463 | static void lru_gen_seq_stop(struct seq_file *m, void *v) | |
5464 | { | |
5465 | if (!IS_ERR_OR_NULL(v)) | |
5466 | mem_cgroup_iter_break(NULL, lruvec_memcg(v)); | |
5467 | ||
5468 | kvfree(m->private); | |
5469 | m->private = NULL; | |
5470 | } | |
5471 | ||
5472 | static void *lru_gen_seq_next(struct seq_file *m, void *v, loff_t *pos) | |
5473 | { | |
5474 | int nid = lruvec_pgdat(v)->node_id; | |
5475 | struct mem_cgroup *memcg = lruvec_memcg(v); | |
5476 | ||
5477 | ++*pos; | |
5478 | ||
5479 | nid = next_memory_node(nid); | |
5480 | if (nid == MAX_NUMNODES) { | |
5481 | memcg = mem_cgroup_iter(NULL, memcg, NULL); | |
5482 | if (!memcg) | |
5483 | return NULL; | |
5484 | ||
5485 | nid = first_memory_node; | |
5486 | } | |
5487 | ||
5488 | return get_lruvec(memcg, nid); | |
5489 | } | |
5490 | ||
5491 | static void lru_gen_seq_show_full(struct seq_file *m, struct lruvec *lruvec, | |
5492 | unsigned long max_seq, unsigned long *min_seq, | |
5493 | unsigned long seq) | |
5494 | { | |
5495 | int i; | |
5496 | int type, tier; | |
5497 | int hist = lru_hist_from_seq(seq); | |
391655fe | 5498 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
d6c3af7d YZ |
5499 | |
5500 | for (tier = 0; tier < MAX_NR_TIERS; tier++) { | |
5501 | seq_printf(m, " %10d", tier); | |
5502 | for (type = 0; type < ANON_AND_FILE; type++) { | |
5503 | const char *s = " "; | |
5504 | unsigned long n[3] = {}; | |
5505 | ||
5506 | if (seq == max_seq) { | |
5507 | s = "RT "; | |
5508 | n[0] = READ_ONCE(lrugen->avg_refaulted[type][tier]); | |
5509 | n[1] = READ_ONCE(lrugen->avg_total[type][tier]); | |
5510 | } else if (seq == min_seq[type] || NR_HIST_GENS > 1) { | |
5511 | s = "rep"; | |
5512 | n[0] = atomic_long_read(&lrugen->refaulted[hist][type][tier]); | |
5513 | n[1] = atomic_long_read(&lrugen->evicted[hist][type][tier]); | |
5514 | if (tier) | |
5515 | n[2] = READ_ONCE(lrugen->protected[hist][type][tier - 1]); | |
5516 | } | |
5517 | ||
5518 | for (i = 0; i < 3; i++) | |
5519 | seq_printf(m, " %10lu%c", n[i], s[i]); | |
5520 | } | |
5521 | seq_putc(m, '\n'); | |
5522 | } | |
5523 | ||
5524 | seq_puts(m, " "); | |
5525 | for (i = 0; i < NR_MM_STATS; i++) { | |
5526 | const char *s = " "; | |
5527 | unsigned long n = 0; | |
5528 | ||
5529 | if (seq == max_seq && NR_HIST_GENS == 1) { | |
5530 | s = "LOYNFA"; | |
5531 | n = READ_ONCE(lruvec->mm_state.stats[hist][i]); | |
5532 | } else if (seq != max_seq && NR_HIST_GENS > 1) { | |
5533 | s = "loynfa"; | |
5534 | n = READ_ONCE(lruvec->mm_state.stats[hist][i]); | |
5535 | } | |
5536 | ||
5537 | seq_printf(m, " %10lu%c", n, s[i]); | |
5538 | } | |
5539 | seq_putc(m, '\n'); | |
5540 | } | |
5541 | ||
07017acb | 5542 | /* see Documentation/admin-guide/mm/multigen_lru.rst for details */ |
d6c3af7d YZ |
5543 | static int lru_gen_seq_show(struct seq_file *m, void *v) |
5544 | { | |
5545 | unsigned long seq; | |
5546 | bool full = !debugfs_real_fops(m->file)->write; | |
5547 | struct lruvec *lruvec = v; | |
391655fe | 5548 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
d6c3af7d YZ |
5549 | int nid = lruvec_pgdat(lruvec)->node_id; |
5550 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
5551 | DEFINE_MAX_SEQ(lruvec); | |
5552 | DEFINE_MIN_SEQ(lruvec); | |
5553 | ||
5554 | if (nid == first_memory_node) { | |
5555 | const char *path = memcg ? m->private : ""; | |
5556 | ||
5557 | #ifdef CONFIG_MEMCG | |
5558 | if (memcg) | |
5559 | cgroup_path(memcg->css.cgroup, m->private, PATH_MAX); | |
5560 | #endif | |
5561 | seq_printf(m, "memcg %5hu %s\n", mem_cgroup_id(memcg), path); | |
5562 | } | |
5563 | ||
5564 | seq_printf(m, " node %5d\n", nid); | |
5565 | ||
5566 | if (!full) | |
5567 | seq = min_seq[LRU_GEN_ANON]; | |
5568 | else if (max_seq >= MAX_NR_GENS) | |
5569 | seq = max_seq - MAX_NR_GENS + 1; | |
5570 | else | |
5571 | seq = 0; | |
5572 | ||
5573 | for (; seq <= max_seq; seq++) { | |
5574 | int type, zone; | |
5575 | int gen = lru_gen_from_seq(seq); | |
5576 | unsigned long birth = READ_ONCE(lruvec->lrugen.timestamps[gen]); | |
5577 | ||
5578 | seq_printf(m, " %10lu %10u", seq, jiffies_to_msecs(jiffies - birth)); | |
5579 | ||
5580 | for (type = 0; type < ANON_AND_FILE; type++) { | |
5581 | unsigned long size = 0; | |
5582 | char mark = full && seq < min_seq[type] ? 'x' : ' '; | |
5583 | ||
5584 | for (zone = 0; zone < MAX_NR_ZONES; zone++) | |
5585 | size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L); | |
5586 | ||
5587 | seq_printf(m, " %10lu%c", size, mark); | |
5588 | } | |
5589 | ||
5590 | seq_putc(m, '\n'); | |
5591 | ||
5592 | if (full) | |
5593 | lru_gen_seq_show_full(m, lruvec, max_seq, min_seq, seq); | |
5594 | } | |
5595 | ||
5596 | return 0; | |
5597 | } | |
5598 | ||
5599 | static const struct seq_operations lru_gen_seq_ops = { | |
5600 | .start = lru_gen_seq_start, | |
5601 | .stop = lru_gen_seq_stop, | |
5602 | .next = lru_gen_seq_next, | |
5603 | .show = lru_gen_seq_show, | |
5604 | }; | |
5605 | ||
5606 | static int run_aging(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc, | |
5607 | bool can_swap, bool force_scan) | |
5608 | { | |
5609 | DEFINE_MAX_SEQ(lruvec); | |
5610 | DEFINE_MIN_SEQ(lruvec); | |
5611 | ||
5612 | if (seq < max_seq) | |
5613 | return 0; | |
5614 | ||
5615 | if (seq > max_seq) | |
5616 | return -EINVAL; | |
5617 | ||
5618 | if (!force_scan && min_seq[!can_swap] + MAX_NR_GENS - 1 <= max_seq) | |
5619 | return -ERANGE; | |
5620 | ||
5621 | try_to_inc_max_seq(lruvec, max_seq, sc, can_swap, force_scan); | |
5622 | ||
5623 | return 0; | |
5624 | } | |
5625 | ||
5626 | static int run_eviction(struct lruvec *lruvec, unsigned long seq, struct scan_control *sc, | |
5627 | int swappiness, unsigned long nr_to_reclaim) | |
5628 | { | |
5629 | DEFINE_MAX_SEQ(lruvec); | |
5630 | ||
5631 | if (seq + MIN_NR_GENS > max_seq) | |
5632 | return -EINVAL; | |
5633 | ||
5634 | sc->nr_reclaimed = 0; | |
5635 | ||
5636 | while (!signal_pending(current)) { | |
5637 | DEFINE_MIN_SEQ(lruvec); | |
5638 | ||
5639 | if (seq < min_seq[!swappiness]) | |
5640 | return 0; | |
5641 | ||
5642 | if (sc->nr_reclaimed >= nr_to_reclaim) | |
5643 | return 0; | |
5644 | ||
a579086c | 5645 | if (!evict_folios(lruvec, sc, swappiness)) |
d6c3af7d YZ |
5646 | return 0; |
5647 | ||
5648 | cond_resched(); | |
5649 | } | |
5650 | ||
5651 | return -EINTR; | |
5652 | } | |
5653 | ||
5654 | static int run_cmd(char cmd, int memcg_id, int nid, unsigned long seq, | |
5655 | struct scan_control *sc, int swappiness, unsigned long opt) | |
5656 | { | |
5657 | struct lruvec *lruvec; | |
5658 | int err = -EINVAL; | |
5659 | struct mem_cgroup *memcg = NULL; | |
5660 | ||
5661 | if (nid < 0 || nid >= MAX_NUMNODES || !node_state(nid, N_MEMORY)) | |
5662 | return -EINVAL; | |
5663 | ||
5664 | if (!mem_cgroup_disabled()) { | |
5665 | rcu_read_lock(); | |
5666 | memcg = mem_cgroup_from_id(memcg_id); | |
5667 | #ifdef CONFIG_MEMCG | |
5668 | if (memcg && !css_tryget(&memcg->css)) | |
5669 | memcg = NULL; | |
5670 | #endif | |
5671 | rcu_read_unlock(); | |
5672 | ||
5673 | if (!memcg) | |
5674 | return -EINVAL; | |
5675 | } | |
5676 | ||
5677 | if (memcg_id != mem_cgroup_id(memcg)) | |
5678 | goto done; | |
5679 | ||
5680 | lruvec = get_lruvec(memcg, nid); | |
5681 | ||
5682 | if (swappiness < 0) | |
5683 | swappiness = get_swappiness(lruvec, sc); | |
5684 | else if (swappiness > 200) | |
5685 | goto done; | |
5686 | ||
5687 | switch (cmd) { | |
5688 | case '+': | |
5689 | err = run_aging(lruvec, seq, sc, swappiness, opt); | |
5690 | break; | |
5691 | case '-': | |
5692 | err = run_eviction(lruvec, seq, sc, swappiness, opt); | |
5693 | break; | |
5694 | } | |
5695 | done: | |
5696 | mem_cgroup_put(memcg); | |
5697 | ||
5698 | return err; | |
5699 | } | |
5700 | ||
07017acb | 5701 | /* see Documentation/admin-guide/mm/multigen_lru.rst for details */ |
d6c3af7d YZ |
5702 | static ssize_t lru_gen_seq_write(struct file *file, const char __user *src, |
5703 | size_t len, loff_t *pos) | |
5704 | { | |
5705 | void *buf; | |
5706 | char *cur, *next; | |
5707 | unsigned int flags; | |
5708 | struct blk_plug plug; | |
5709 | int err = -EINVAL; | |
5710 | struct scan_control sc = { | |
5711 | .may_writepage = true, | |
5712 | .may_unmap = true, | |
5713 | .may_swap = true, | |
5714 | .reclaim_idx = MAX_NR_ZONES - 1, | |
5715 | .gfp_mask = GFP_KERNEL, | |
5716 | }; | |
5717 | ||
5718 | buf = kvmalloc(len + 1, GFP_KERNEL); | |
5719 | if (!buf) | |
5720 | return -ENOMEM; | |
5721 | ||
5722 | if (copy_from_user(buf, src, len)) { | |
5723 | kvfree(buf); | |
5724 | return -EFAULT; | |
5725 | } | |
5726 | ||
5727 | set_task_reclaim_state(current, &sc.reclaim_state); | |
5728 | flags = memalloc_noreclaim_save(); | |
5729 | blk_start_plug(&plug); | |
5730 | if (!set_mm_walk(NULL)) { | |
5731 | err = -ENOMEM; | |
5732 | goto done; | |
5733 | } | |
5734 | ||
5735 | next = buf; | |
5736 | next[len] = '\0'; | |
5737 | ||
5738 | while ((cur = strsep(&next, ",;\n"))) { | |
5739 | int n; | |
5740 | int end; | |
5741 | char cmd; | |
5742 | unsigned int memcg_id; | |
5743 | unsigned int nid; | |
5744 | unsigned long seq; | |
5745 | unsigned int swappiness = -1; | |
5746 | unsigned long opt = -1; | |
5747 | ||
5748 | cur = skip_spaces(cur); | |
5749 | if (!*cur) | |
5750 | continue; | |
5751 | ||
5752 | n = sscanf(cur, "%c %u %u %lu %n %u %n %lu %n", &cmd, &memcg_id, &nid, | |
5753 | &seq, &end, &swappiness, &end, &opt, &end); | |
5754 | if (n < 4 || cur[end]) { | |
5755 | err = -EINVAL; | |
5756 | break; | |
5757 | } | |
5758 | ||
5759 | err = run_cmd(cmd, memcg_id, nid, seq, &sc, swappiness, opt); | |
5760 | if (err) | |
5761 | break; | |
5762 | } | |
5763 | done: | |
5764 | clear_mm_walk(); | |
5765 | blk_finish_plug(&plug); | |
5766 | memalloc_noreclaim_restore(flags); | |
5767 | set_task_reclaim_state(current, NULL); | |
5768 | ||
5769 | kvfree(buf); | |
5770 | ||
5771 | return err ? : len; | |
5772 | } | |
5773 | ||
5774 | static int lru_gen_seq_open(struct inode *inode, struct file *file) | |
5775 | { | |
5776 | return seq_open(file, &lru_gen_seq_ops); | |
5777 | } | |
5778 | ||
5779 | static const struct file_operations lru_gen_rw_fops = { | |
5780 | .open = lru_gen_seq_open, | |
5781 | .read = seq_read, | |
5782 | .write = lru_gen_seq_write, | |
5783 | .llseek = seq_lseek, | |
5784 | .release = seq_release, | |
5785 | }; | |
5786 | ||
5787 | static const struct file_operations lru_gen_ro_fops = { | |
5788 | .open = lru_gen_seq_open, | |
5789 | .read = seq_read, | |
5790 | .llseek = seq_lseek, | |
5791 | .release = seq_release, | |
5792 | }; | |
5793 | ||
ec1c86b2 YZ |
5794 | /****************************************************************************** |
5795 | * initialization | |
5796 | ******************************************************************************/ | |
5797 | ||
5798 | void lru_gen_init_lruvec(struct lruvec *lruvec) | |
5799 | { | |
1332a809 | 5800 | int i; |
ec1c86b2 | 5801 | int gen, type, zone; |
391655fe | 5802 | struct lru_gen_folio *lrugen = &lruvec->lrugen; |
ec1c86b2 YZ |
5803 | |
5804 | lrugen->max_seq = MIN_NR_GENS + 1; | |
354ed597 | 5805 | lrugen->enabled = lru_gen_enabled(); |
ec1c86b2 | 5806 | |
1332a809 YZ |
5807 | for (i = 0; i <= MIN_NR_GENS + 1; i++) |
5808 | lrugen->timestamps[i] = jiffies; | |
5809 | ||
ec1c86b2 | 5810 | for_each_gen_type_zone(gen, type, zone) |
6df1b221 | 5811 | INIT_LIST_HEAD(&lrugen->folios[gen][type][zone]); |
bd74fdae YZ |
5812 | |
5813 | lruvec->mm_state.seq = MIN_NR_GENS; | |
5814 | init_waitqueue_head(&lruvec->mm_state.wait); | |
ec1c86b2 YZ |
5815 | } |
5816 | ||
5817 | #ifdef CONFIG_MEMCG | |
5818 | void lru_gen_init_memcg(struct mem_cgroup *memcg) | |
5819 | { | |
bd74fdae YZ |
5820 | INIT_LIST_HEAD(&memcg->mm_list.fifo); |
5821 | spin_lock_init(&memcg->mm_list.lock); | |
ec1c86b2 YZ |
5822 | } |
5823 | ||
5824 | void lru_gen_exit_memcg(struct mem_cgroup *memcg) | |
5825 | { | |
bd74fdae | 5826 | int i; |
ec1c86b2 YZ |
5827 | int nid; |
5828 | ||
5829 | for_each_node(nid) { | |
5830 | struct lruvec *lruvec = get_lruvec(memcg, nid); | |
5831 | ||
5832 | VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0, | |
5833 | sizeof(lruvec->lrugen.nr_pages))); | |
bd74fdae YZ |
5834 | |
5835 | for (i = 0; i < NR_BLOOM_FILTERS; i++) { | |
5836 | bitmap_free(lruvec->mm_state.filters[i]); | |
5837 | lruvec->mm_state.filters[i] = NULL; | |
5838 | } | |
ec1c86b2 YZ |
5839 | } |
5840 | } | |
5841 | #endif | |
5842 | ||
5843 | static int __init init_lru_gen(void) | |
5844 | { | |
5845 | BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS); | |
5846 | BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS); | |
5847 | ||
354ed597 YZ |
5848 | if (sysfs_create_group(mm_kobj, &lru_gen_attr_group)) |
5849 | pr_err("lru_gen: failed to create sysfs group\n"); | |
5850 | ||
d6c3af7d YZ |
5851 | debugfs_create_file("lru_gen", 0644, NULL, NULL, &lru_gen_rw_fops); |
5852 | debugfs_create_file("lru_gen_full", 0444, NULL, NULL, &lru_gen_ro_fops); | |
5853 | ||
ec1c86b2 YZ |
5854 | return 0; |
5855 | }; | |
5856 | late_initcall(init_lru_gen); | |
5857 | ||
ac35a490 YZ |
5858 | #else /* !CONFIG_LRU_GEN */ |
5859 | ||
5860 | static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc) | |
5861 | { | |
5862 | } | |
5863 | ||
5864 | static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) | |
5865 | { | |
5866 | } | |
5867 | ||
ec1c86b2 YZ |
5868 | #endif /* CONFIG_LRU_GEN */ |
5869 | ||
afaf07a6 | 5870 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) |
9b4f98cd JW |
5871 | { |
5872 | unsigned long nr[NR_LRU_LISTS]; | |
e82e0561 | 5873 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
5874 | unsigned long nr_to_scan; |
5875 | enum lru_list lru; | |
5876 | unsigned long nr_reclaimed = 0; | |
5877 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
f53af428 | 5878 | bool proportional_reclaim; |
9b4f98cd JW |
5879 | struct blk_plug plug; |
5880 | ||
ac35a490 YZ |
5881 | if (lru_gen_enabled()) { |
5882 | lru_gen_shrink_lruvec(lruvec, sc); | |
5883 | return; | |
5884 | } | |
5885 | ||
afaf07a6 | 5886 | get_scan_count(lruvec, sc, nr); |
9b4f98cd | 5887 | |
e82e0561 MG |
5888 | /* Record the original scan target for proportional adjustments later */ |
5889 | memcpy(targets, nr, sizeof(nr)); | |
5890 | ||
1a501907 MG |
5891 | /* |
5892 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
5893 | * event that can occur when there is little memory pressure e.g. | |
5894 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
5895 | * when the requested number of pages are reclaimed when scanning at | |
5896 | * DEF_PRIORITY on the assumption that the fact we are direct | |
5897 | * reclaiming implies that kswapd is not keeping up and it is best to | |
5898 | * do a batch of work at once. For memcg reclaim one check is made to | |
5899 | * abort proportional reclaim if either the file or anon lru has already | |
5900 | * dropped to zero at the first pass. | |
5901 | */ | |
f53af428 JW |
5902 | proportional_reclaim = (!cgroup_reclaim(sc) && !current_is_kswapd() && |
5903 | sc->priority == DEF_PRIORITY); | |
1a501907 | 5904 | |
9b4f98cd JW |
5905 | blk_start_plug(&plug); |
5906 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
5907 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
5908 | unsigned long nr_anon, nr_file, percentage; |
5909 | unsigned long nr_scanned; | |
5910 | ||
9b4f98cd JW |
5911 | for_each_evictable_lru(lru) { |
5912 | if (nr[lru]) { | |
5913 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
5914 | nr[lru] -= nr_to_scan; | |
5915 | ||
5916 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
3b991208 | 5917 | lruvec, sc); |
9b4f98cd JW |
5918 | } |
5919 | } | |
e82e0561 | 5920 | |
bd041733 MH |
5921 | cond_resched(); |
5922 | ||
f53af428 | 5923 | if (nr_reclaimed < nr_to_reclaim || proportional_reclaim) |
e82e0561 MG |
5924 | continue; |
5925 | ||
e82e0561 MG |
5926 | /* |
5927 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 5928 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
5929 | * proportionally what was requested by get_scan_count(). We |
5930 | * stop reclaiming one LRU and reduce the amount scanning | |
5931 | * proportional to the original scan target. | |
5932 | */ | |
5933 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
5934 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
5935 | ||
1a501907 MG |
5936 | /* |
5937 | * It's just vindictive to attack the larger once the smaller | |
5938 | * has gone to zero. And given the way we stop scanning the | |
5939 | * smaller below, this makes sure that we only make one nudge | |
5940 | * towards proportionality once we've got nr_to_reclaim. | |
5941 | */ | |
5942 | if (!nr_file || !nr_anon) | |
5943 | break; | |
5944 | ||
e82e0561 MG |
5945 | if (nr_file > nr_anon) { |
5946 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
5947 | targets[LRU_ACTIVE_ANON] + 1; | |
5948 | lru = LRU_BASE; | |
5949 | percentage = nr_anon * 100 / scan_target; | |
5950 | } else { | |
5951 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
5952 | targets[LRU_ACTIVE_FILE] + 1; | |
5953 | lru = LRU_FILE; | |
5954 | percentage = nr_file * 100 / scan_target; | |
5955 | } | |
5956 | ||
5957 | /* Stop scanning the smaller of the LRU */ | |
5958 | nr[lru] = 0; | |
5959 | nr[lru + LRU_ACTIVE] = 0; | |
5960 | ||
5961 | /* | |
5962 | * Recalculate the other LRU scan count based on its original | |
5963 | * scan target and the percentage scanning already complete | |
5964 | */ | |
5965 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
5966 | nr_scanned = targets[lru] - nr[lru]; | |
5967 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
5968 | nr[lru] -= min(nr[lru], nr_scanned); | |
5969 | ||
5970 | lru += LRU_ACTIVE; | |
5971 | nr_scanned = targets[lru] - nr[lru]; | |
5972 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
5973 | nr[lru] -= min(nr[lru], nr_scanned); | |
9b4f98cd JW |
5974 | } |
5975 | blk_finish_plug(&plug); | |
5976 | sc->nr_reclaimed += nr_reclaimed; | |
5977 | ||
5978 | /* | |
5979 | * Even if we did not try to evict anon pages at all, we want to | |
5980 | * rebalance the anon lru active/inactive ratio. | |
5981 | */ | |
2f368a9f DH |
5982 | if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) && |
5983 | inactive_is_low(lruvec, LRU_INACTIVE_ANON)) | |
9b4f98cd JW |
5984 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
5985 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
5986 | } |
5987 | ||
23b9da55 | 5988 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 5989 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 5990 | { |
d84da3f9 | 5991 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 5992 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 5993 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
5994 | return true; |
5995 | ||
5996 | return false; | |
5997 | } | |
5998 | ||
3e7d3449 | 5999 | /* |
23b9da55 MG |
6000 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
6001 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
6002 | * true if more pages should be reclaimed such that when the page allocator | |
df3a45f9 | 6003 | * calls try_to_compact_pages() that it will have enough free pages to succeed. |
23b9da55 | 6004 | * It will give up earlier than that if there is difficulty reclaiming pages. |
3e7d3449 | 6005 | */ |
a9dd0a83 | 6006 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 | 6007 | unsigned long nr_reclaimed, |
3e7d3449 MG |
6008 | struct scan_control *sc) |
6009 | { | |
6010 | unsigned long pages_for_compaction; | |
6011 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 6012 | int z; |
3e7d3449 MG |
6013 | |
6014 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 6015 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
6016 | return false; |
6017 | ||
5ee04716 VB |
6018 | /* |
6019 | * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX | |
6020 | * number of pages that were scanned. This will return to the caller | |
6021 | * with the risk reclaim/compaction and the resulting allocation attempt | |
6022 | * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL | |
6023 | * allocations through requiring that the full LRU list has been scanned | |
6024 | * first, by assuming that zero delta of sc->nr_scanned means full LRU | |
6025 | * scan, but that approximation was wrong, and there were corner cases | |
6026 | * where always a non-zero amount of pages were scanned. | |
6027 | */ | |
6028 | if (!nr_reclaimed) | |
6029 | return false; | |
3e7d3449 | 6030 | |
3e7d3449 | 6031 | /* If compaction would go ahead or the allocation would succeed, stop */ |
a9dd0a83 MG |
6032 | for (z = 0; z <= sc->reclaim_idx; z++) { |
6033 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 6034 | if (!managed_zone(zone)) |
a9dd0a83 MG |
6035 | continue; |
6036 | ||
6037 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 6038 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
6039 | case COMPACT_CONTINUE: |
6040 | return false; | |
6041 | default: | |
6042 | /* check next zone */ | |
6043 | ; | |
6044 | } | |
3e7d3449 | 6045 | } |
1c6c1597 HD |
6046 | |
6047 | /* | |
6048 | * If we have not reclaimed enough pages for compaction and the | |
6049 | * inactive lists are large enough, continue reclaiming | |
6050 | */ | |
6051 | pages_for_compaction = compact_gap(sc->order); | |
6052 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); | |
a2a36488 | 6053 | if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc)) |
1c6c1597 HD |
6054 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
6055 | ||
5ee04716 | 6056 | return inactive_lru_pages > pages_for_compaction; |
3e7d3449 MG |
6057 | } |
6058 | ||
0f6a5cff | 6059 | static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 6060 | { |
0f6a5cff | 6061 | struct mem_cgroup *target_memcg = sc->target_mem_cgroup; |
d2af3397 | 6062 | struct mem_cgroup *memcg; |
1da177e4 | 6063 | |
0f6a5cff | 6064 | memcg = mem_cgroup_iter(target_memcg, NULL, NULL); |
d2af3397 | 6065 | do { |
afaf07a6 | 6066 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
d2af3397 JW |
6067 | unsigned long reclaimed; |
6068 | unsigned long scanned; | |
5660048c | 6069 | |
e3336cab XP |
6070 | /* |
6071 | * This loop can become CPU-bound when target memcgs | |
6072 | * aren't eligible for reclaim - either because they | |
6073 | * don't have any reclaimable pages, or because their | |
6074 | * memory is explicitly protected. Avoid soft lockups. | |
6075 | */ | |
6076 | cond_resched(); | |
6077 | ||
45c7f7e1 CD |
6078 | mem_cgroup_calculate_protection(target_memcg, memcg); |
6079 | ||
adb82130 | 6080 | if (mem_cgroup_below_min(target_memcg, memcg)) { |
d2af3397 JW |
6081 | /* |
6082 | * Hard protection. | |
6083 | * If there is no reclaimable memory, OOM. | |
6084 | */ | |
6085 | continue; | |
adb82130 | 6086 | } else if (mem_cgroup_below_low(target_memcg, memcg)) { |
d2af3397 JW |
6087 | /* |
6088 | * Soft protection. | |
6089 | * Respect the protection only as long as | |
6090 | * there is an unprotected supply | |
6091 | * of reclaimable memory from other cgroups. | |
6092 | */ | |
6093 | if (!sc->memcg_low_reclaim) { | |
6094 | sc->memcg_low_skipped = 1; | |
bf8d5d52 | 6095 | continue; |
241994ed | 6096 | } |
d2af3397 | 6097 | memcg_memory_event(memcg, MEMCG_LOW); |
d2af3397 | 6098 | } |
241994ed | 6099 | |
d2af3397 JW |
6100 | reclaimed = sc->nr_reclaimed; |
6101 | scanned = sc->nr_scanned; | |
afaf07a6 JW |
6102 | |
6103 | shrink_lruvec(lruvec, sc); | |
70ddf637 | 6104 | |
d2af3397 JW |
6105 | shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, |
6106 | sc->priority); | |
6b4f7799 | 6107 | |
d2af3397 | 6108 | /* Record the group's reclaim efficiency */ |
73b73bac YA |
6109 | if (!sc->proactive) |
6110 | vmpressure(sc->gfp_mask, memcg, false, | |
6111 | sc->nr_scanned - scanned, | |
6112 | sc->nr_reclaimed - reclaimed); | |
70ddf637 | 6113 | |
0f6a5cff JW |
6114 | } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL))); |
6115 | } | |
6116 | ||
6c9e0907 | 6117 | static void shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
0f6a5cff JW |
6118 | { |
6119 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
0f6a5cff | 6120 | unsigned long nr_reclaimed, nr_scanned; |
1b05117d | 6121 | struct lruvec *target_lruvec; |
0f6a5cff JW |
6122 | bool reclaimable = false; |
6123 | ||
1b05117d JW |
6124 | target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat); |
6125 | ||
0f6a5cff JW |
6126 | again: |
6127 | memset(&sc->nr, 0, sizeof(sc->nr)); | |
6128 | ||
6129 | nr_reclaimed = sc->nr_reclaimed; | |
6130 | nr_scanned = sc->nr_scanned; | |
6131 | ||
f1e1a7be | 6132 | prepare_scan_count(pgdat, sc); |
53138cea | 6133 | |
0f6a5cff | 6134 | shrink_node_memcgs(pgdat, sc); |
2344d7e4 | 6135 | |
d2af3397 JW |
6136 | if (reclaim_state) { |
6137 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
6138 | reclaim_state->reclaimed_slab = 0; | |
6139 | } | |
d108c772 | 6140 | |
d2af3397 | 6141 | /* Record the subtree's reclaim efficiency */ |
73b73bac YA |
6142 | if (!sc->proactive) |
6143 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
6144 | sc->nr_scanned - nr_scanned, | |
6145 | sc->nr_reclaimed - nr_reclaimed); | |
d108c772 | 6146 | |
d2af3397 JW |
6147 | if (sc->nr_reclaimed - nr_reclaimed) |
6148 | reclaimable = true; | |
d108c772 | 6149 | |
d2af3397 JW |
6150 | if (current_is_kswapd()) { |
6151 | /* | |
6152 | * If reclaim is isolating dirty pages under writeback, | |
6153 | * it implies that the long-lived page allocation rate | |
6154 | * is exceeding the page laundering rate. Either the | |
6155 | * global limits are not being effective at throttling | |
6156 | * processes due to the page distribution throughout | |
6157 | * zones or there is heavy usage of a slow backing | |
6158 | * device. The only option is to throttle from reclaim | |
6159 | * context which is not ideal as there is no guarantee | |
6160 | * the dirtying process is throttled in the same way | |
6161 | * balance_dirty_pages() manages. | |
6162 | * | |
6163 | * Once a node is flagged PGDAT_WRITEBACK, kswapd will | |
6164 | * count the number of pages under pages flagged for | |
6165 | * immediate reclaim and stall if any are encountered | |
6166 | * in the nr_immediate check below. | |
6167 | */ | |
6168 | if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken) | |
6169 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
d108c772 | 6170 | |
d2af3397 JW |
6171 | /* Allow kswapd to start writing pages during reclaim.*/ |
6172 | if (sc->nr.unqueued_dirty == sc->nr.file_taken) | |
6173 | set_bit(PGDAT_DIRTY, &pgdat->flags); | |
e3c1ac58 | 6174 | |
d108c772 | 6175 | /* |
1eba09c1 | 6176 | * If kswapd scans pages marked for immediate |
d2af3397 JW |
6177 | * reclaim and under writeback (nr_immediate), it |
6178 | * implies that pages are cycling through the LRU | |
8cd7c588 MG |
6179 | * faster than they are written so forcibly stall |
6180 | * until some pages complete writeback. | |
d108c772 | 6181 | */ |
d2af3397 | 6182 | if (sc->nr.immediate) |
c3f4a9a2 | 6183 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK); |
d2af3397 JW |
6184 | } |
6185 | ||
6186 | /* | |
8cd7c588 MG |
6187 | * Tag a node/memcg as congested if all the dirty pages were marked |
6188 | * for writeback and immediate reclaim (counted in nr.congested). | |
1b05117d | 6189 | * |
d2af3397 | 6190 | * Legacy memcg will stall in page writeback so avoid forcibly |
8cd7c588 | 6191 | * stalling in reclaim_throttle(). |
d2af3397 | 6192 | */ |
1b05117d JW |
6193 | if ((current_is_kswapd() || |
6194 | (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) && | |
d2af3397 | 6195 | sc->nr.dirty && sc->nr.dirty == sc->nr.congested) |
1b05117d | 6196 | set_bit(LRUVEC_CONGESTED, &target_lruvec->flags); |
d2af3397 JW |
6197 | |
6198 | /* | |
8cd7c588 MG |
6199 | * Stall direct reclaim for IO completions if the lruvec is |
6200 | * node is congested. Allow kswapd to continue until it | |
d2af3397 JW |
6201 | * starts encountering unqueued dirty pages or cycling through |
6202 | * the LRU too quickly. | |
6203 | */ | |
1b05117d JW |
6204 | if (!current_is_kswapd() && current_may_throttle() && |
6205 | !sc->hibernation_mode && | |
6206 | test_bit(LRUVEC_CONGESTED, &target_lruvec->flags)) | |
1b4e3f26 | 6207 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED); |
d108c772 | 6208 | |
d2af3397 JW |
6209 | if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
6210 | sc)) | |
6211 | goto again; | |
2344d7e4 | 6212 | |
c73322d0 JW |
6213 | /* |
6214 | * Kswapd gives up on balancing particular nodes after too | |
6215 | * many failures to reclaim anything from them and goes to | |
6216 | * sleep. On reclaim progress, reset the failure counter. A | |
6217 | * successful direct reclaim run will revive a dormant kswapd. | |
6218 | */ | |
6219 | if (reclaimable) | |
6220 | pgdat->kswapd_failures = 0; | |
f16015fb JW |
6221 | } |
6222 | ||
53853e2d | 6223 | /* |
fdd4c614 VB |
6224 | * Returns true if compaction should go ahead for a costly-order request, or |
6225 | * the allocation would already succeed without compaction. Return false if we | |
6226 | * should reclaim first. | |
53853e2d | 6227 | */ |
4f588331 | 6228 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 6229 | { |
31483b6a | 6230 | unsigned long watermark; |
fdd4c614 | 6231 | enum compact_result suitable; |
fe4b1b24 | 6232 | |
fdd4c614 VB |
6233 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
6234 | if (suitable == COMPACT_SUCCESS) | |
6235 | /* Allocation should succeed already. Don't reclaim. */ | |
6236 | return true; | |
6237 | if (suitable == COMPACT_SKIPPED) | |
6238 | /* Compaction cannot yet proceed. Do reclaim. */ | |
6239 | return false; | |
fe4b1b24 | 6240 | |
53853e2d | 6241 | /* |
fdd4c614 VB |
6242 | * Compaction is already possible, but it takes time to run and there |
6243 | * are potentially other callers using the pages just freed. So proceed | |
6244 | * with reclaim to make a buffer of free pages available to give | |
6245 | * compaction a reasonable chance of completing and allocating the page. | |
6246 | * Note that we won't actually reclaim the whole buffer in one attempt | |
6247 | * as the target watermark in should_continue_reclaim() is lower. But if | |
6248 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 6249 | */ |
fdd4c614 | 6250 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 6251 | |
fdd4c614 | 6252 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
6253 | } |
6254 | ||
69392a40 MG |
6255 | static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc) |
6256 | { | |
66ce520b MG |
6257 | /* |
6258 | * If reclaim is making progress greater than 12% efficiency then | |
6259 | * wake all the NOPROGRESS throttled tasks. | |
6260 | */ | |
6261 | if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) { | |
69392a40 MG |
6262 | wait_queue_head_t *wqh; |
6263 | ||
6264 | wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS]; | |
6265 | if (waitqueue_active(wqh)) | |
6266 | wake_up(wqh); | |
6267 | ||
6268 | return; | |
6269 | } | |
6270 | ||
6271 | /* | |
1b4e3f26 MG |
6272 | * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will |
6273 | * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages | |
6274 | * under writeback and marked for immediate reclaim at the tail of the | |
6275 | * LRU. | |
69392a40 | 6276 | */ |
1b4e3f26 | 6277 | if (current_is_kswapd() || cgroup_reclaim(sc)) |
69392a40 MG |
6278 | return; |
6279 | ||
6280 | /* Throttle if making no progress at high prioities. */ | |
1b4e3f26 | 6281 | if (sc->priority == 1 && !sc->nr_reclaimed) |
c3f4a9a2 | 6282 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS); |
69392a40 MG |
6283 | } |
6284 | ||
1da177e4 LT |
6285 | /* |
6286 | * This is the direct reclaim path, for page-allocating processes. We only | |
6287 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
6288 | * request. | |
6289 | * | |
1da177e4 LT |
6290 | * If a zone is deemed to be full of pinned pages then just give it a light |
6291 | * scan then give up on it. | |
6292 | */ | |
0a0337e0 | 6293 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 6294 | { |
dd1a239f | 6295 | struct zoneref *z; |
54a6eb5c | 6296 | struct zone *zone; |
0608f43d AM |
6297 | unsigned long nr_soft_reclaimed; |
6298 | unsigned long nr_soft_scanned; | |
619d0d76 | 6299 | gfp_t orig_mask; |
79dafcdc | 6300 | pg_data_t *last_pgdat = NULL; |
1b4e3f26 | 6301 | pg_data_t *first_pgdat = NULL; |
1cfb419b | 6302 | |
cc715d99 MG |
6303 | /* |
6304 | * If the number of buffer_heads in the machine exceeds the maximum | |
6305 | * allowed level, force direct reclaim to scan the highmem zone as | |
6306 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
6307 | */ | |
619d0d76 | 6308 | orig_mask = sc->gfp_mask; |
b2e18757 | 6309 | if (buffer_heads_over_limit) { |
cc715d99 | 6310 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 6311 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 6312 | } |
cc715d99 | 6313 | |
d4debc66 | 6314 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 6315 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
6316 | /* |
6317 | * Take care memory controller reclaiming has small influence | |
6318 | * to global LRU. | |
6319 | */ | |
b5ead35e | 6320 | if (!cgroup_reclaim(sc)) { |
344736f2 VD |
6321 | if (!cpuset_zone_allowed(zone, |
6322 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 6323 | continue; |
65ec02cb | 6324 | |
0b06496a JW |
6325 | /* |
6326 | * If we already have plenty of memory free for | |
6327 | * compaction in this zone, don't free any more. | |
6328 | * Even though compaction is invoked for any | |
6329 | * non-zero order, only frequent costly order | |
6330 | * reclamation is disruptive enough to become a | |
6331 | * noticeable problem, like transparent huge | |
6332 | * page allocations. | |
6333 | */ | |
6334 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
6335 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 6336 | compaction_ready(zone, sc)) { |
0b06496a JW |
6337 | sc->compaction_ready = true; |
6338 | continue; | |
e0887c19 | 6339 | } |
0b06496a | 6340 | |
79dafcdc MG |
6341 | /* |
6342 | * Shrink each node in the zonelist once. If the | |
6343 | * zonelist is ordered by zone (not the default) then a | |
6344 | * node may be shrunk multiple times but in that case | |
6345 | * the user prefers lower zones being preserved. | |
6346 | */ | |
6347 | if (zone->zone_pgdat == last_pgdat) | |
6348 | continue; | |
6349 | ||
0608f43d AM |
6350 | /* |
6351 | * This steals pages from memory cgroups over softlimit | |
6352 | * and returns the number of reclaimed pages and | |
6353 | * scanned pages. This works for global memory pressure | |
6354 | * and balancing, not for a memcg's limit. | |
6355 | */ | |
6356 | nr_soft_scanned = 0; | |
ef8f2327 | 6357 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
6358 | sc->order, sc->gfp_mask, |
6359 | &nr_soft_scanned); | |
6360 | sc->nr_reclaimed += nr_soft_reclaimed; | |
6361 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 6362 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 6363 | } |
408d8544 | 6364 | |
1b4e3f26 MG |
6365 | if (!first_pgdat) |
6366 | first_pgdat = zone->zone_pgdat; | |
6367 | ||
79dafcdc MG |
6368 | /* See comment about same check for global reclaim above */ |
6369 | if (zone->zone_pgdat == last_pgdat) | |
6370 | continue; | |
6371 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 6372 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 6373 | } |
e0c23279 | 6374 | |
80082938 MG |
6375 | if (first_pgdat) |
6376 | consider_reclaim_throttle(first_pgdat, sc); | |
1b4e3f26 | 6377 | |
619d0d76 WY |
6378 | /* |
6379 | * Restore to original mask to avoid the impact on the caller if we | |
6380 | * promoted it to __GFP_HIGHMEM. | |
6381 | */ | |
6382 | sc->gfp_mask = orig_mask; | |
1da177e4 | 6383 | } |
4f98a2fe | 6384 | |
b910718a | 6385 | static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat) |
2a2e4885 | 6386 | { |
b910718a JW |
6387 | struct lruvec *target_lruvec; |
6388 | unsigned long refaults; | |
2a2e4885 | 6389 | |
ac35a490 YZ |
6390 | if (lru_gen_enabled()) |
6391 | return; | |
6392 | ||
b910718a | 6393 | target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat); |
170b04b7 | 6394 | refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON); |
e9c2dbc8 | 6395 | target_lruvec->refaults[WORKINGSET_ANON] = refaults; |
170b04b7 | 6396 | refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE); |
e9c2dbc8 | 6397 | target_lruvec->refaults[WORKINGSET_FILE] = refaults; |
2a2e4885 JW |
6398 | } |
6399 | ||
1da177e4 LT |
6400 | /* |
6401 | * This is the main entry point to direct page reclaim. | |
6402 | * | |
6403 | * If a full scan of the inactive list fails to free enough memory then we | |
6404 | * are "out of memory" and something needs to be killed. | |
6405 | * | |
6406 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
6407 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
6408 | * caller can't do much about. We kick the writeback threads and take explicit |
6409 | * naps in the hope that some of these pages can be written. But if the | |
6410 | * allocating task holds filesystem locks which prevent writeout this might not | |
6411 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
6412 | * |
6413 | * returns: 0, if no pages reclaimed | |
6414 | * else, the number of pages reclaimed | |
1da177e4 | 6415 | */ |
dac1d27b | 6416 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 6417 | struct scan_control *sc) |
1da177e4 | 6418 | { |
241994ed | 6419 | int initial_priority = sc->priority; |
2a2e4885 JW |
6420 | pg_data_t *last_pgdat; |
6421 | struct zoneref *z; | |
6422 | struct zone *zone; | |
241994ed | 6423 | retry: |
873b4771 KK |
6424 | delayacct_freepages_start(); |
6425 | ||
b5ead35e | 6426 | if (!cgroup_reclaim(sc)) |
7cc30fcf | 6427 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 6428 | |
9e3b2f8c | 6429 | do { |
73b73bac YA |
6430 | if (!sc->proactive) |
6431 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, | |
6432 | sc->priority); | |
66e1707b | 6433 | sc->nr_scanned = 0; |
0a0337e0 | 6434 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 6435 | |
bb21c7ce | 6436 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
6437 | break; |
6438 | ||
6439 | if (sc->compaction_ready) | |
6440 | break; | |
1da177e4 | 6441 | |
0e50ce3b MK |
6442 | /* |
6443 | * If we're getting trouble reclaiming, start doing | |
6444 | * writepage even in laptop mode. | |
6445 | */ | |
6446 | if (sc->priority < DEF_PRIORITY - 2) | |
6447 | sc->may_writepage = 1; | |
0b06496a | 6448 | } while (--sc->priority >= 0); |
bb21c7ce | 6449 | |
2a2e4885 JW |
6450 | last_pgdat = NULL; |
6451 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
6452 | sc->nodemask) { | |
6453 | if (zone->zone_pgdat == last_pgdat) | |
6454 | continue; | |
6455 | last_pgdat = zone->zone_pgdat; | |
1b05117d | 6456 | |
2a2e4885 | 6457 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); |
1b05117d JW |
6458 | |
6459 | if (cgroup_reclaim(sc)) { | |
6460 | struct lruvec *lruvec; | |
6461 | ||
6462 | lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, | |
6463 | zone->zone_pgdat); | |
6464 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
6465 | } | |
2a2e4885 JW |
6466 | } |
6467 | ||
873b4771 KK |
6468 | delayacct_freepages_end(); |
6469 | ||
bb21c7ce KM |
6470 | if (sc->nr_reclaimed) |
6471 | return sc->nr_reclaimed; | |
6472 | ||
0cee34fd | 6473 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 6474 | if (sc->compaction_ready) |
7335084d MG |
6475 | return 1; |
6476 | ||
b91ac374 JW |
6477 | /* |
6478 | * We make inactive:active ratio decisions based on the node's | |
6479 | * composition of memory, but a restrictive reclaim_idx or a | |
6480 | * memory.low cgroup setting can exempt large amounts of | |
6481 | * memory from reclaim. Neither of which are very common, so | |
6482 | * instead of doing costly eligibility calculations of the | |
6483 | * entire cgroup subtree up front, we assume the estimates are | |
6484 | * good, and retry with forcible deactivation if that fails. | |
6485 | */ | |
6486 | if (sc->skipped_deactivate) { | |
6487 | sc->priority = initial_priority; | |
6488 | sc->force_deactivate = 1; | |
6489 | sc->skipped_deactivate = 0; | |
6490 | goto retry; | |
6491 | } | |
6492 | ||
241994ed | 6493 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
d6622f63 | 6494 | if (sc->memcg_low_skipped) { |
241994ed | 6495 | sc->priority = initial_priority; |
b91ac374 | 6496 | sc->force_deactivate = 0; |
d6622f63 YX |
6497 | sc->memcg_low_reclaim = 1; |
6498 | sc->memcg_low_skipped = 0; | |
241994ed JW |
6499 | goto retry; |
6500 | } | |
6501 | ||
bb21c7ce | 6502 | return 0; |
1da177e4 LT |
6503 | } |
6504 | ||
c73322d0 | 6505 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
5515061d MG |
6506 | { |
6507 | struct zone *zone; | |
6508 | unsigned long pfmemalloc_reserve = 0; | |
6509 | unsigned long free_pages = 0; | |
6510 | int i; | |
6511 | bool wmark_ok; | |
6512 | ||
c73322d0 JW |
6513 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
6514 | return true; | |
6515 | ||
5515061d MG |
6516 | for (i = 0; i <= ZONE_NORMAL; i++) { |
6517 | zone = &pgdat->node_zones[i]; | |
d450abd8 JW |
6518 | if (!managed_zone(zone)) |
6519 | continue; | |
6520 | ||
6521 | if (!zone_reclaimable_pages(zone)) | |
675becce MG |
6522 | continue; |
6523 | ||
5515061d MG |
6524 | pfmemalloc_reserve += min_wmark_pages(zone); |
6525 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
6526 | } | |
6527 | ||
675becce MG |
6528 | /* If there are no reserves (unexpected config) then do not throttle */ |
6529 | if (!pfmemalloc_reserve) | |
6530 | return true; | |
6531 | ||
5515061d MG |
6532 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
6533 | ||
6534 | /* kswapd must be awake if processes are being throttled */ | |
6535 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
97a225e6 JK |
6536 | if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL) |
6537 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL); | |
5644e1fb | 6538 | |
5515061d MG |
6539 | wake_up_interruptible(&pgdat->kswapd_wait); |
6540 | } | |
6541 | ||
6542 | return wmark_ok; | |
6543 | } | |
6544 | ||
6545 | /* | |
6546 | * Throttle direct reclaimers if backing storage is backed by the network | |
6547 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
6548 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
6549 | * when the low watermark is reached. |
6550 | * | |
6551 | * Returns true if a fatal signal was delivered during throttling. If this | |
6552 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 6553 | */ |
50694c28 | 6554 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
6555 | nodemask_t *nodemask) |
6556 | { | |
675becce | 6557 | struct zoneref *z; |
5515061d | 6558 | struct zone *zone; |
675becce | 6559 | pg_data_t *pgdat = NULL; |
5515061d MG |
6560 | |
6561 | /* | |
6562 | * Kernel threads should not be throttled as they may be indirectly | |
6563 | * responsible for cleaning pages necessary for reclaim to make forward | |
6564 | * progress. kjournald for example may enter direct reclaim while | |
6565 | * committing a transaction where throttling it could forcing other | |
6566 | * processes to block on log_wait_commit(). | |
6567 | */ | |
6568 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
6569 | goto out; |
6570 | ||
6571 | /* | |
6572 | * If a fatal signal is pending, this process should not throttle. | |
6573 | * It should return quickly so it can exit and free its memory | |
6574 | */ | |
6575 | if (fatal_signal_pending(current)) | |
6576 | goto out; | |
5515061d | 6577 | |
675becce MG |
6578 | /* |
6579 | * Check if the pfmemalloc reserves are ok by finding the first node | |
6580 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
6581 | * GFP_KERNEL will be required for allocating network buffers when | |
6582 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
6583 | * | |
6584 | * Throttling is based on the first usable node and throttled processes | |
6585 | * wait on a queue until kswapd makes progress and wakes them. There | |
6586 | * is an affinity then between processes waking up and where reclaim | |
6587 | * progress has been made assuming the process wakes on the same node. | |
6588 | * More importantly, processes running on remote nodes will not compete | |
6589 | * for remote pfmemalloc reserves and processes on different nodes | |
6590 | * should make reasonable progress. | |
6591 | */ | |
6592 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 6593 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
6594 | if (zone_idx(zone) > ZONE_NORMAL) |
6595 | continue; | |
6596 | ||
6597 | /* Throttle based on the first usable node */ | |
6598 | pgdat = zone->zone_pgdat; | |
c73322d0 | 6599 | if (allow_direct_reclaim(pgdat)) |
675becce MG |
6600 | goto out; |
6601 | break; | |
6602 | } | |
6603 | ||
6604 | /* If no zone was usable by the allocation flags then do not throttle */ | |
6605 | if (!pgdat) | |
50694c28 | 6606 | goto out; |
5515061d | 6607 | |
68243e76 MG |
6608 | /* Account for the throttling */ |
6609 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
6610 | ||
5515061d MG |
6611 | /* |
6612 | * If the caller cannot enter the filesystem, it's possible that it | |
6613 | * is due to the caller holding an FS lock or performing a journal | |
6614 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
6615 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
6616 | * blocked waiting on the same lock. Instead, throttle for up to a | |
6617 | * second before continuing. | |
6618 | */ | |
2e786d9e | 6619 | if (!(gfp_mask & __GFP_FS)) |
5515061d | 6620 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, |
c73322d0 | 6621 | allow_direct_reclaim(pgdat), HZ); |
2e786d9e ML |
6622 | else |
6623 | /* Throttle until kswapd wakes the process */ | |
6624 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
6625 | allow_direct_reclaim(pgdat)); | |
50694c28 | 6626 | |
50694c28 MG |
6627 | if (fatal_signal_pending(current)) |
6628 | return true; | |
6629 | ||
6630 | out: | |
6631 | return false; | |
5515061d MG |
6632 | } |
6633 | ||
dac1d27b | 6634 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 6635 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 6636 | { |
33906bc5 | 6637 | unsigned long nr_reclaimed; |
66e1707b | 6638 | struct scan_control sc = { |
ee814fe2 | 6639 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
f2f43e56 | 6640 | .gfp_mask = current_gfp_context(gfp_mask), |
b2e18757 | 6641 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
6642 | .order = order, |
6643 | .nodemask = nodemask, | |
6644 | .priority = DEF_PRIORITY, | |
66e1707b | 6645 | .may_writepage = !laptop_mode, |
a6dc60f8 | 6646 | .may_unmap = 1, |
2e2e4259 | 6647 | .may_swap = 1, |
66e1707b BS |
6648 | }; |
6649 | ||
bb451fdf GT |
6650 | /* |
6651 | * scan_control uses s8 fields for order, priority, and reclaim_idx. | |
6652 | * Confirm they are large enough for max values. | |
6653 | */ | |
6654 | BUILD_BUG_ON(MAX_ORDER > S8_MAX); | |
6655 | BUILD_BUG_ON(DEF_PRIORITY > S8_MAX); | |
6656 | BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX); | |
6657 | ||
5515061d | 6658 | /* |
50694c28 MG |
6659 | * Do not enter reclaim if fatal signal was delivered while throttled. |
6660 | * 1 is returned so that the page allocator does not OOM kill at this | |
6661 | * point. | |
5515061d | 6662 | */ |
f2f43e56 | 6663 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
5515061d MG |
6664 | return 1; |
6665 | ||
1732d2b0 | 6666 | set_task_reclaim_state(current, &sc.reclaim_state); |
3481c37f | 6667 | trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask); |
33906bc5 | 6668 | |
3115cd91 | 6669 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
6670 | |
6671 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
1732d2b0 | 6672 | set_task_reclaim_state(current, NULL); |
33906bc5 MG |
6673 | |
6674 | return nr_reclaimed; | |
66e1707b BS |
6675 | } |
6676 | ||
c255a458 | 6677 | #ifdef CONFIG_MEMCG |
66e1707b | 6678 | |
d2e5fb92 | 6679 | /* Only used by soft limit reclaim. Do not reuse for anything else. */ |
a9dd0a83 | 6680 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 6681 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 6682 | pg_data_t *pgdat, |
0ae5e89c | 6683 | unsigned long *nr_scanned) |
4e416953 | 6684 | { |
afaf07a6 | 6685 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
4e416953 | 6686 | struct scan_control sc = { |
b8f5c566 | 6687 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 6688 | .target_mem_cgroup = memcg, |
4e416953 BS |
6689 | .may_writepage = !laptop_mode, |
6690 | .may_unmap = 1, | |
b2e18757 | 6691 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 6692 | .may_swap = !noswap, |
4e416953 | 6693 | }; |
0ae5e89c | 6694 | |
d2e5fb92 MH |
6695 | WARN_ON_ONCE(!current->reclaim_state); |
6696 | ||
4e416953 BS |
6697 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
6698 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 6699 | |
9e3b2f8c | 6700 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
3481c37f | 6701 | sc.gfp_mask); |
bdce6d9e | 6702 | |
4e416953 BS |
6703 | /* |
6704 | * NOTE: Although we can get the priority field, using it | |
6705 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 6706 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
6707 | * will pick up pages from other mem cgroup's as well. We hack |
6708 | * the priority and make it zero. | |
6709 | */ | |
afaf07a6 | 6710 | shrink_lruvec(lruvec, &sc); |
bdce6d9e KM |
6711 | |
6712 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
6713 | ||
0ae5e89c | 6714 | *nr_scanned = sc.nr_scanned; |
0308f7cf | 6715 | |
4e416953 BS |
6716 | return sc.nr_reclaimed; |
6717 | } | |
6718 | ||
72835c86 | 6719 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 6720 | unsigned long nr_pages, |
a7885eb8 | 6721 | gfp_t gfp_mask, |
12a5d395 MA |
6722 | unsigned int reclaim_options, |
6723 | nodemask_t *nodemask) | |
66e1707b | 6724 | { |
bdce6d9e | 6725 | unsigned long nr_reclaimed; |
499118e9 | 6726 | unsigned int noreclaim_flag; |
66e1707b | 6727 | struct scan_control sc = { |
b70a2a21 | 6728 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 6729 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
a09ed5e0 | 6730 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
b2e18757 | 6731 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
6732 | .target_mem_cgroup = memcg, |
6733 | .priority = DEF_PRIORITY, | |
6734 | .may_writepage = !laptop_mode, | |
6735 | .may_unmap = 1, | |
73b73bac YA |
6736 | .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP), |
6737 | .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE), | |
12a5d395 | 6738 | .nodemask = nodemask, |
a09ed5e0 | 6739 | }; |
889976db | 6740 | /* |
fa40d1ee SB |
6741 | * Traverse the ZONELIST_FALLBACK zonelist of the current node to put |
6742 | * equal pressure on all the nodes. This is based on the assumption that | |
6743 | * the reclaim does not bail out early. | |
889976db | 6744 | */ |
fa40d1ee | 6745 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
889976db | 6746 | |
fa40d1ee | 6747 | set_task_reclaim_state(current, &sc.reclaim_state); |
3481c37f | 6748 | trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask); |
499118e9 | 6749 | noreclaim_flag = memalloc_noreclaim_save(); |
eb414681 | 6750 | |
3115cd91 | 6751 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
eb414681 | 6752 | |
499118e9 | 6753 | memalloc_noreclaim_restore(noreclaim_flag); |
bdce6d9e | 6754 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); |
1732d2b0 | 6755 | set_task_reclaim_state(current, NULL); |
bdce6d9e KM |
6756 | |
6757 | return nr_reclaimed; | |
66e1707b BS |
6758 | } |
6759 | #endif | |
6760 | ||
ac35a490 | 6761 | static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc) |
f16015fb | 6762 | { |
b95a2f2d | 6763 | struct mem_cgroup *memcg; |
b91ac374 | 6764 | struct lruvec *lruvec; |
f16015fb | 6765 | |
ac35a490 YZ |
6766 | if (lru_gen_enabled()) { |
6767 | lru_gen_age_node(pgdat, sc); | |
6768 | return; | |
6769 | } | |
6770 | ||
2f368a9f | 6771 | if (!can_age_anon_pages(pgdat, sc)) |
b95a2f2d JW |
6772 | return; |
6773 | ||
b91ac374 JW |
6774 | lruvec = mem_cgroup_lruvec(NULL, pgdat); |
6775 | if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON)) | |
6776 | return; | |
6777 | ||
b95a2f2d JW |
6778 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
6779 | do { | |
b91ac374 JW |
6780 | lruvec = mem_cgroup_lruvec(memcg, pgdat); |
6781 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
6782 | sc, LRU_ACTIVE_ANON); | |
b95a2f2d JW |
6783 | memcg = mem_cgroup_iter(NULL, memcg, NULL); |
6784 | } while (memcg); | |
f16015fb JW |
6785 | } |
6786 | ||
97a225e6 | 6787 | static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx) |
1c30844d MG |
6788 | { |
6789 | int i; | |
6790 | struct zone *zone; | |
6791 | ||
6792 | /* | |
6793 | * Check for watermark boosts top-down as the higher zones | |
6794 | * are more likely to be boosted. Both watermarks and boosts | |
1eba09c1 | 6795 | * should not be checked at the same time as reclaim would |
1c30844d MG |
6796 | * start prematurely when there is no boosting and a lower |
6797 | * zone is balanced. | |
6798 | */ | |
97a225e6 | 6799 | for (i = highest_zoneidx; i >= 0; i--) { |
1c30844d MG |
6800 | zone = pgdat->node_zones + i; |
6801 | if (!managed_zone(zone)) | |
6802 | continue; | |
6803 | ||
6804 | if (zone->watermark_boost) | |
6805 | return true; | |
6806 | } | |
6807 | ||
6808 | return false; | |
6809 | } | |
6810 | ||
e716f2eb MG |
6811 | /* |
6812 | * Returns true if there is an eligible zone balanced for the request order | |
97a225e6 | 6813 | * and highest_zoneidx |
e716f2eb | 6814 | */ |
97a225e6 | 6815 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx) |
60cefed4 | 6816 | { |
e716f2eb MG |
6817 | int i; |
6818 | unsigned long mark = -1; | |
6819 | struct zone *zone; | |
60cefed4 | 6820 | |
1c30844d MG |
6821 | /* |
6822 | * Check watermarks bottom-up as lower zones are more likely to | |
6823 | * meet watermarks. | |
6824 | */ | |
97a225e6 | 6825 | for (i = 0; i <= highest_zoneidx; i++) { |
e716f2eb | 6826 | zone = pgdat->node_zones + i; |
6256c6b4 | 6827 | |
e716f2eb MG |
6828 | if (!managed_zone(zone)) |
6829 | continue; | |
6830 | ||
c574bbe9 HY |
6831 | if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) |
6832 | mark = wmark_pages(zone, WMARK_PROMO); | |
6833 | else | |
6834 | mark = high_wmark_pages(zone); | |
97a225e6 | 6835 | if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx)) |
e716f2eb MG |
6836 | return true; |
6837 | } | |
6838 | ||
6839 | /* | |
36c26128 | 6840 | * If a node has no managed zone within highest_zoneidx, it does not |
e716f2eb MG |
6841 | * need balancing by definition. This can happen if a zone-restricted |
6842 | * allocation tries to wake a remote kswapd. | |
6843 | */ | |
6844 | if (mark == -1) | |
6845 | return true; | |
6846 | ||
6847 | return false; | |
60cefed4 JW |
6848 | } |
6849 | ||
631b6e08 MG |
6850 | /* Clear pgdat state for congested, dirty or under writeback. */ |
6851 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
6852 | { | |
1b05117d JW |
6853 | struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat); |
6854 | ||
6855 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
631b6e08 MG |
6856 | clear_bit(PGDAT_DIRTY, &pgdat->flags); |
6857 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
6858 | } | |
6859 | ||
5515061d MG |
6860 | /* |
6861 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
6862 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
6863 | * | |
6864 | * Returns true if kswapd is ready to sleep | |
6865 | */ | |
97a225e6 JK |
6866 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, |
6867 | int highest_zoneidx) | |
f50de2d3 | 6868 | { |
5515061d | 6869 | /* |
9e5e3661 | 6870 | * The throttled processes are normally woken up in balance_pgdat() as |
c73322d0 | 6871 | * soon as allow_direct_reclaim() is true. But there is a potential |
9e5e3661 VB |
6872 | * race between when kswapd checks the watermarks and a process gets |
6873 | * throttled. There is also a potential race if processes get | |
6874 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
6875 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
6876 | * the wake up checks. If kswapd is going to sleep, no process should | |
6877 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
6878 | * the wake up is premature, processes will wake kswapd and get | |
6879 | * throttled again. The difference from wake ups in balance_pgdat() is | |
6880 | * that here we are under prepare_to_wait(). | |
5515061d | 6881 | */ |
9e5e3661 VB |
6882 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
6883 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 6884 | |
c73322d0 JW |
6885 | /* Hopeless node, leave it to direct reclaim */ |
6886 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
6887 | return true; | |
6888 | ||
97a225e6 | 6889 | if (pgdat_balanced(pgdat, order, highest_zoneidx)) { |
e716f2eb MG |
6890 | clear_pgdat_congested(pgdat); |
6891 | return true; | |
1d82de61 MG |
6892 | } |
6893 | ||
333b0a45 | 6894 | return false; |
f50de2d3 MG |
6895 | } |
6896 | ||
75485363 | 6897 | /* |
1d82de61 MG |
6898 | * kswapd shrinks a node of pages that are at or below the highest usable |
6899 | * zone that is currently unbalanced. | |
b8e83b94 MG |
6900 | * |
6901 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
6902 | * reclaim or if the lack of progress was due to pages under writeback. |
6903 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 6904 | */ |
1d82de61 | 6905 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 6906 | struct scan_control *sc) |
75485363 | 6907 | { |
1d82de61 MG |
6908 | struct zone *zone; |
6909 | int z; | |
75485363 | 6910 | |
1d82de61 MG |
6911 | /* Reclaim a number of pages proportional to the number of zones */ |
6912 | sc->nr_to_reclaim = 0; | |
970a39a3 | 6913 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 6914 | zone = pgdat->node_zones + z; |
6aa303de | 6915 | if (!managed_zone(zone)) |
1d82de61 | 6916 | continue; |
7c954f6d | 6917 | |
1d82de61 MG |
6918 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
6919 | } | |
7c954f6d MG |
6920 | |
6921 | /* | |
1d82de61 MG |
6922 | * Historically care was taken to put equal pressure on all zones but |
6923 | * now pressure is applied based on node LRU order. | |
7c954f6d | 6924 | */ |
970a39a3 | 6925 | shrink_node(pgdat, sc); |
283aba9f | 6926 | |
7c954f6d | 6927 | /* |
1d82de61 MG |
6928 | * Fragmentation may mean that the system cannot be rebalanced for |
6929 | * high-order allocations. If twice the allocation size has been | |
6930 | * reclaimed then recheck watermarks only at order-0 to prevent | |
6931 | * excessive reclaim. Assume that a process requested a high-order | |
6932 | * can direct reclaim/compact. | |
7c954f6d | 6933 | */ |
9861a62c | 6934 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 6935 | sc->order = 0; |
7c954f6d | 6936 | |
b8e83b94 | 6937 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
6938 | } |
6939 | ||
c49c2c47 MG |
6940 | /* Page allocator PCP high watermark is lowered if reclaim is active. */ |
6941 | static inline void | |
6942 | update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active) | |
6943 | { | |
6944 | int i; | |
6945 | struct zone *zone; | |
6946 | ||
6947 | for (i = 0; i <= highest_zoneidx; i++) { | |
6948 | zone = pgdat->node_zones + i; | |
6949 | ||
6950 | if (!managed_zone(zone)) | |
6951 | continue; | |
6952 | ||
6953 | if (active) | |
6954 | set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags); | |
6955 | else | |
6956 | clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags); | |
6957 | } | |
6958 | } | |
6959 | ||
6960 | static inline void | |
6961 | set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx) | |
6962 | { | |
6963 | update_reclaim_active(pgdat, highest_zoneidx, true); | |
6964 | } | |
6965 | ||
6966 | static inline void | |
6967 | clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx) | |
6968 | { | |
6969 | update_reclaim_active(pgdat, highest_zoneidx, false); | |
6970 | } | |
6971 | ||
1da177e4 | 6972 | /* |
1d82de61 MG |
6973 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
6974 | * that are eligible for use by the caller until at least one zone is | |
6975 | * balanced. | |
1da177e4 | 6976 | * |
1d82de61 | 6977 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
6978 | * |
6979 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 6980 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
8bb4e7a2 | 6981 | * found to have free_pages <= high_wmark_pages(zone), any page in that zone |
1d82de61 MG |
6982 | * or lower is eligible for reclaim until at least one usable zone is |
6983 | * balanced. | |
1da177e4 | 6984 | */ |
97a225e6 | 6985 | static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx) |
1da177e4 | 6986 | { |
1da177e4 | 6987 | int i; |
0608f43d AM |
6988 | unsigned long nr_soft_reclaimed; |
6989 | unsigned long nr_soft_scanned; | |
eb414681 | 6990 | unsigned long pflags; |
1c30844d MG |
6991 | unsigned long nr_boost_reclaim; |
6992 | unsigned long zone_boosts[MAX_NR_ZONES] = { 0, }; | |
6993 | bool boosted; | |
1d82de61 | 6994 | struct zone *zone; |
179e9639 AM |
6995 | struct scan_control sc = { |
6996 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 6997 | .order = order, |
a6dc60f8 | 6998 | .may_unmap = 1, |
179e9639 | 6999 | }; |
93781325 | 7000 | |
1732d2b0 | 7001 | set_task_reclaim_state(current, &sc.reclaim_state); |
eb414681 | 7002 | psi_memstall_enter(&pflags); |
4f3eaf45 | 7003 | __fs_reclaim_acquire(_THIS_IP_); |
93781325 | 7004 | |
f8891e5e | 7005 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 7006 | |
1c30844d MG |
7007 | /* |
7008 | * Account for the reclaim boost. Note that the zone boost is left in | |
7009 | * place so that parallel allocations that are near the watermark will | |
7010 | * stall or direct reclaim until kswapd is finished. | |
7011 | */ | |
7012 | nr_boost_reclaim = 0; | |
97a225e6 | 7013 | for (i = 0; i <= highest_zoneidx; i++) { |
1c30844d MG |
7014 | zone = pgdat->node_zones + i; |
7015 | if (!managed_zone(zone)) | |
7016 | continue; | |
7017 | ||
7018 | nr_boost_reclaim += zone->watermark_boost; | |
7019 | zone_boosts[i] = zone->watermark_boost; | |
7020 | } | |
7021 | boosted = nr_boost_reclaim; | |
7022 | ||
7023 | restart: | |
c49c2c47 | 7024 | set_reclaim_active(pgdat, highest_zoneidx); |
1c30844d | 7025 | sc.priority = DEF_PRIORITY; |
9e3b2f8c | 7026 | do { |
c73322d0 | 7027 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
b8e83b94 | 7028 | bool raise_priority = true; |
1c30844d | 7029 | bool balanced; |
93781325 | 7030 | bool ret; |
b8e83b94 | 7031 | |
97a225e6 | 7032 | sc.reclaim_idx = highest_zoneidx; |
1da177e4 | 7033 | |
86c79f6b | 7034 | /* |
84c7a777 MG |
7035 | * If the number of buffer_heads exceeds the maximum allowed |
7036 | * then consider reclaiming from all zones. This has a dual | |
7037 | * purpose -- on 64-bit systems it is expected that | |
7038 | * buffer_heads are stripped during active rotation. On 32-bit | |
7039 | * systems, highmem pages can pin lowmem memory and shrinking | |
7040 | * buffers can relieve lowmem pressure. Reclaim may still not | |
7041 | * go ahead if all eligible zones for the original allocation | |
7042 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
7043 | */ |
7044 | if (buffer_heads_over_limit) { | |
7045 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
7046 | zone = pgdat->node_zones + i; | |
6aa303de | 7047 | if (!managed_zone(zone)) |
86c79f6b | 7048 | continue; |
cc715d99 | 7049 | |
970a39a3 | 7050 | sc.reclaim_idx = i; |
e1dbeda6 | 7051 | break; |
1da177e4 | 7052 | } |
1da177e4 | 7053 | } |
dafcb73e | 7054 | |
86c79f6b | 7055 | /* |
1c30844d MG |
7056 | * If the pgdat is imbalanced then ignore boosting and preserve |
7057 | * the watermarks for a later time and restart. Note that the | |
7058 | * zone watermarks will be still reset at the end of balancing | |
7059 | * on the grounds that the normal reclaim should be enough to | |
7060 | * re-evaluate if boosting is required when kswapd next wakes. | |
7061 | */ | |
97a225e6 | 7062 | balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx); |
1c30844d MG |
7063 | if (!balanced && nr_boost_reclaim) { |
7064 | nr_boost_reclaim = 0; | |
7065 | goto restart; | |
7066 | } | |
7067 | ||
7068 | /* | |
7069 | * If boosting is not active then only reclaim if there are no | |
7070 | * eligible zones. Note that sc.reclaim_idx is not used as | |
7071 | * buffer_heads_over_limit may have adjusted it. | |
86c79f6b | 7072 | */ |
1c30844d | 7073 | if (!nr_boost_reclaim && balanced) |
e716f2eb | 7074 | goto out; |
e1dbeda6 | 7075 | |
1c30844d MG |
7076 | /* Limit the priority of boosting to avoid reclaim writeback */ |
7077 | if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2) | |
7078 | raise_priority = false; | |
7079 | ||
7080 | /* | |
7081 | * Do not writeback or swap pages for boosted reclaim. The | |
7082 | * intent is to relieve pressure not issue sub-optimal IO | |
7083 | * from reclaim context. If no pages are reclaimed, the | |
7084 | * reclaim will be aborted. | |
7085 | */ | |
7086 | sc.may_writepage = !laptop_mode && !nr_boost_reclaim; | |
7087 | sc.may_swap = !nr_boost_reclaim; | |
1c30844d | 7088 | |
1d82de61 | 7089 | /* |
ac35a490 YZ |
7090 | * Do some background aging, to give pages a chance to be |
7091 | * referenced before reclaiming. All pages are rotated | |
7092 | * regardless of classzone as this is about consistent aging. | |
1d82de61 | 7093 | */ |
ac35a490 | 7094 | kswapd_age_node(pgdat, &sc); |
1d82de61 | 7095 | |
b7ea3c41 MG |
7096 | /* |
7097 | * If we're getting trouble reclaiming, start doing writepage | |
7098 | * even in laptop mode. | |
7099 | */ | |
047d72c3 | 7100 | if (sc.priority < DEF_PRIORITY - 2) |
b7ea3c41 MG |
7101 | sc.may_writepage = 1; |
7102 | ||
1d82de61 MG |
7103 | /* Call soft limit reclaim before calling shrink_node. */ |
7104 | sc.nr_scanned = 0; | |
7105 | nr_soft_scanned = 0; | |
ef8f2327 | 7106 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
7107 | sc.gfp_mask, &nr_soft_scanned); |
7108 | sc.nr_reclaimed += nr_soft_reclaimed; | |
7109 | ||
1da177e4 | 7110 | /* |
1d82de61 MG |
7111 | * There should be no need to raise the scanning priority if |
7112 | * enough pages are already being scanned that that high | |
7113 | * watermark would be met at 100% efficiency. | |
1da177e4 | 7114 | */ |
970a39a3 | 7115 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 7116 | raise_priority = false; |
5515061d MG |
7117 | |
7118 | /* | |
7119 | * If the low watermark is met there is no need for processes | |
7120 | * to be throttled on pfmemalloc_wait as they should not be | |
7121 | * able to safely make forward progress. Wake them | |
7122 | */ | |
7123 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
c73322d0 | 7124 | allow_direct_reclaim(pgdat)) |
cfc51155 | 7125 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 7126 | |
b8e83b94 | 7127 | /* Check if kswapd should be suspending */ |
4f3eaf45 | 7128 | __fs_reclaim_release(_THIS_IP_); |
93781325 | 7129 | ret = try_to_freeze(); |
4f3eaf45 | 7130 | __fs_reclaim_acquire(_THIS_IP_); |
93781325 | 7131 | if (ret || kthread_should_stop()) |
b8e83b94 | 7132 | break; |
8357376d | 7133 | |
73ce02e9 | 7134 | /* |
b8e83b94 MG |
7135 | * Raise priority if scanning rate is too low or there was no |
7136 | * progress in reclaiming pages | |
73ce02e9 | 7137 | */ |
c73322d0 | 7138 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
1c30844d MG |
7139 | nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed); |
7140 | ||
7141 | /* | |
7142 | * If reclaim made no progress for a boost, stop reclaim as | |
7143 | * IO cannot be queued and it could be an infinite loop in | |
7144 | * extreme circumstances. | |
7145 | */ | |
7146 | if (nr_boost_reclaim && !nr_reclaimed) | |
7147 | break; | |
7148 | ||
c73322d0 | 7149 | if (raise_priority || !nr_reclaimed) |
b8e83b94 | 7150 | sc.priority--; |
1d82de61 | 7151 | } while (sc.priority >= 1); |
1da177e4 | 7152 | |
c73322d0 JW |
7153 | if (!sc.nr_reclaimed) |
7154 | pgdat->kswapd_failures++; | |
7155 | ||
b8e83b94 | 7156 | out: |
c49c2c47 MG |
7157 | clear_reclaim_active(pgdat, highest_zoneidx); |
7158 | ||
1c30844d MG |
7159 | /* If reclaim was boosted, account for the reclaim done in this pass */ |
7160 | if (boosted) { | |
7161 | unsigned long flags; | |
7162 | ||
97a225e6 | 7163 | for (i = 0; i <= highest_zoneidx; i++) { |
1c30844d MG |
7164 | if (!zone_boosts[i]) |
7165 | continue; | |
7166 | ||
7167 | /* Increments are under the zone lock */ | |
7168 | zone = pgdat->node_zones + i; | |
7169 | spin_lock_irqsave(&zone->lock, flags); | |
7170 | zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]); | |
7171 | spin_unlock_irqrestore(&zone->lock, flags); | |
7172 | } | |
7173 | ||
7174 | /* | |
7175 | * As there is now likely space, wakeup kcompact to defragment | |
7176 | * pageblocks. | |
7177 | */ | |
97a225e6 | 7178 | wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx); |
1c30844d MG |
7179 | } |
7180 | ||
2a2e4885 | 7181 | snapshot_refaults(NULL, pgdat); |
4f3eaf45 | 7182 | __fs_reclaim_release(_THIS_IP_); |
eb414681 | 7183 | psi_memstall_leave(&pflags); |
1732d2b0 | 7184 | set_task_reclaim_state(current, NULL); |
e5ca8071 | 7185 | |
0abdee2b | 7186 | /* |
1d82de61 MG |
7187 | * Return the order kswapd stopped reclaiming at as |
7188 | * prepare_kswapd_sleep() takes it into account. If another caller | |
7189 | * entered the allocator slow path while kswapd was awake, order will | |
7190 | * remain at the higher level. | |
0abdee2b | 7191 | */ |
1d82de61 | 7192 | return sc.order; |
1da177e4 LT |
7193 | } |
7194 | ||
e716f2eb | 7195 | /* |
97a225e6 JK |
7196 | * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to |
7197 | * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is | |
7198 | * not a valid index then either kswapd runs for first time or kswapd couldn't | |
7199 | * sleep after previous reclaim attempt (node is still unbalanced). In that | |
7200 | * case return the zone index of the previous kswapd reclaim cycle. | |
e716f2eb | 7201 | */ |
97a225e6 JK |
7202 | static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat, |
7203 | enum zone_type prev_highest_zoneidx) | |
e716f2eb | 7204 | { |
97a225e6 | 7205 | enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx); |
5644e1fb | 7206 | |
97a225e6 | 7207 | return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx; |
e716f2eb MG |
7208 | } |
7209 | ||
38087d9b | 7210 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
97a225e6 | 7211 | unsigned int highest_zoneidx) |
f0bc0a60 KM |
7212 | { |
7213 | long remaining = 0; | |
7214 | DEFINE_WAIT(wait); | |
7215 | ||
7216 | if (freezing(current) || kthread_should_stop()) | |
7217 | return; | |
7218 | ||
7219 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
7220 | ||
333b0a45 SG |
7221 | /* |
7222 | * Try to sleep for a short interval. Note that kcompactd will only be | |
7223 | * woken if it is possible to sleep for a short interval. This is | |
7224 | * deliberate on the assumption that if reclaim cannot keep an | |
7225 | * eligible zone balanced that it's also unlikely that compaction will | |
7226 | * succeed. | |
7227 | */ | |
97a225e6 | 7228 | if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) { |
fd901c95 VB |
7229 | /* |
7230 | * Compaction records what page blocks it recently failed to | |
7231 | * isolate pages from and skips them in the future scanning. | |
7232 | * When kswapd is going to sleep, it is reasonable to assume | |
7233 | * that pages and compaction may succeed so reset the cache. | |
7234 | */ | |
7235 | reset_isolation_suitable(pgdat); | |
7236 | ||
7237 | /* | |
7238 | * We have freed the memory, now we should compact it to make | |
7239 | * allocation of the requested order possible. | |
7240 | */ | |
97a225e6 | 7241 | wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx); |
fd901c95 | 7242 | |
f0bc0a60 | 7243 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
7244 | |
7245 | /* | |
97a225e6 | 7246 | * If woken prematurely then reset kswapd_highest_zoneidx and |
38087d9b MG |
7247 | * order. The values will either be from a wakeup request or |
7248 | * the previous request that slept prematurely. | |
7249 | */ | |
7250 | if (remaining) { | |
97a225e6 JK |
7251 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, |
7252 | kswapd_highest_zoneidx(pgdat, | |
7253 | highest_zoneidx)); | |
5644e1fb QC |
7254 | |
7255 | if (READ_ONCE(pgdat->kswapd_order) < reclaim_order) | |
7256 | WRITE_ONCE(pgdat->kswapd_order, reclaim_order); | |
38087d9b MG |
7257 | } |
7258 | ||
f0bc0a60 KM |
7259 | finish_wait(&pgdat->kswapd_wait, &wait); |
7260 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
7261 | } | |
7262 | ||
7263 | /* | |
7264 | * After a short sleep, check if it was a premature sleep. If not, then | |
7265 | * go fully to sleep until explicitly woken up. | |
7266 | */ | |
d9f21d42 | 7267 | if (!remaining && |
97a225e6 | 7268 | prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) { |
f0bc0a60 KM |
7269 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
7270 | ||
7271 | /* | |
7272 | * vmstat counters are not perfectly accurate and the estimated | |
7273 | * value for counters such as NR_FREE_PAGES can deviate from the | |
7274 | * true value by nr_online_cpus * threshold. To avoid the zone | |
7275 | * watermarks being breached while under pressure, we reduce the | |
7276 | * per-cpu vmstat threshold while kswapd is awake and restore | |
7277 | * them before going back to sleep. | |
7278 | */ | |
7279 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
7280 | |
7281 | if (!kthread_should_stop()) | |
7282 | schedule(); | |
7283 | ||
f0bc0a60 KM |
7284 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
7285 | } else { | |
7286 | if (remaining) | |
7287 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
7288 | else | |
7289 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
7290 | } | |
7291 | finish_wait(&pgdat->kswapd_wait, &wait); | |
7292 | } | |
7293 | ||
1da177e4 LT |
7294 | /* |
7295 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 7296 | * from the init process. |
1da177e4 LT |
7297 | * |
7298 | * This basically trickles out pages so that we have _some_ | |
7299 | * free memory available even if there is no other activity | |
7300 | * that frees anything up. This is needed for things like routing | |
7301 | * etc, where we otherwise might have all activity going on in | |
7302 | * asynchronous contexts that cannot page things out. | |
7303 | * | |
7304 | * If there are applications that are active memory-allocators | |
7305 | * (most normal use), this basically shouldn't matter. | |
7306 | */ | |
7307 | static int kswapd(void *p) | |
7308 | { | |
e716f2eb | 7309 | unsigned int alloc_order, reclaim_order; |
97a225e6 | 7310 | unsigned int highest_zoneidx = MAX_NR_ZONES - 1; |
68d68ff6 | 7311 | pg_data_t *pgdat = (pg_data_t *)p; |
1da177e4 | 7312 | struct task_struct *tsk = current; |
a70f7302 | 7313 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 7314 | |
174596a0 | 7315 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 7316 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
7317 | |
7318 | /* | |
7319 | * Tell the memory management that we're a "memory allocator", | |
7320 | * and that if we need more memory we should get access to it | |
7321 | * regardless (see "__alloc_pages()"). "kswapd" should | |
7322 | * never get caught in the normal page freeing logic. | |
7323 | * | |
7324 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
7325 | * you need a small amount of memory in order to be able to | |
7326 | * page out something else, and this flag essentially protects | |
7327 | * us from recursively trying to free more memory as we're | |
7328 | * trying to free the first piece of memory in the first place). | |
7329 | */ | |
b698f0a1 | 7330 | tsk->flags |= PF_MEMALLOC | PF_KSWAPD; |
83144186 | 7331 | set_freezable(); |
1da177e4 | 7332 | |
5644e1fb | 7333 | WRITE_ONCE(pgdat->kswapd_order, 0); |
97a225e6 | 7334 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES); |
8cd7c588 | 7335 | atomic_set(&pgdat->nr_writeback_throttled, 0); |
1da177e4 | 7336 | for ( ; ; ) { |
6f6313d4 | 7337 | bool ret; |
3e1d1d28 | 7338 | |
5644e1fb | 7339 | alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order); |
97a225e6 JK |
7340 | highest_zoneidx = kswapd_highest_zoneidx(pgdat, |
7341 | highest_zoneidx); | |
e716f2eb | 7342 | |
38087d9b MG |
7343 | kswapd_try_sleep: |
7344 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
97a225e6 | 7345 | highest_zoneidx); |
215ddd66 | 7346 | |
97a225e6 | 7347 | /* Read the new order and highest_zoneidx */ |
2b47a24c | 7348 | alloc_order = READ_ONCE(pgdat->kswapd_order); |
97a225e6 JK |
7349 | highest_zoneidx = kswapd_highest_zoneidx(pgdat, |
7350 | highest_zoneidx); | |
5644e1fb | 7351 | WRITE_ONCE(pgdat->kswapd_order, 0); |
97a225e6 | 7352 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES); |
1da177e4 | 7353 | |
8fe23e05 DR |
7354 | ret = try_to_freeze(); |
7355 | if (kthread_should_stop()) | |
7356 | break; | |
7357 | ||
7358 | /* | |
7359 | * We can speed up thawing tasks if we don't call balance_pgdat | |
7360 | * after returning from the refrigerator | |
7361 | */ | |
38087d9b MG |
7362 | if (ret) |
7363 | continue; | |
7364 | ||
7365 | /* | |
7366 | * Reclaim begins at the requested order but if a high-order | |
7367 | * reclaim fails then kswapd falls back to reclaiming for | |
7368 | * order-0. If that happens, kswapd will consider sleeping | |
7369 | * for the order it finished reclaiming at (reclaim_order) | |
7370 | * but kcompactd is woken to compact for the original | |
7371 | * request (alloc_order). | |
7372 | */ | |
97a225e6 | 7373 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx, |
e5146b12 | 7374 | alloc_order); |
97a225e6 JK |
7375 | reclaim_order = balance_pgdat(pgdat, alloc_order, |
7376 | highest_zoneidx); | |
38087d9b MG |
7377 | if (reclaim_order < alloc_order) |
7378 | goto kswapd_try_sleep; | |
1da177e4 | 7379 | } |
b0a8cc58 | 7380 | |
b698f0a1 | 7381 | tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD); |
71abdc15 | 7382 | |
1da177e4 LT |
7383 | return 0; |
7384 | } | |
7385 | ||
7386 | /* | |
5ecd9d40 DR |
7387 | * A zone is low on free memory or too fragmented for high-order memory. If |
7388 | * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's | |
7389 | * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim | |
7390 | * has failed or is not needed, still wake up kcompactd if only compaction is | |
7391 | * needed. | |
1da177e4 | 7392 | */ |
5ecd9d40 | 7393 | void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, |
97a225e6 | 7394 | enum zone_type highest_zoneidx) |
1da177e4 LT |
7395 | { |
7396 | pg_data_t *pgdat; | |
5644e1fb | 7397 | enum zone_type curr_idx; |
1da177e4 | 7398 | |
6aa303de | 7399 | if (!managed_zone(zone)) |
1da177e4 LT |
7400 | return; |
7401 | ||
5ecd9d40 | 7402 | if (!cpuset_zone_allowed(zone, gfp_flags)) |
1da177e4 | 7403 | return; |
5644e1fb | 7404 | |
88f5acf8 | 7405 | pgdat = zone->zone_pgdat; |
97a225e6 | 7406 | curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx); |
5644e1fb | 7407 | |
97a225e6 JK |
7408 | if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx) |
7409 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx); | |
5644e1fb QC |
7410 | |
7411 | if (READ_ONCE(pgdat->kswapd_order) < order) | |
7412 | WRITE_ONCE(pgdat->kswapd_order, order); | |
dffcac2c | 7413 | |
8d0986e2 | 7414 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 7415 | return; |
e1a55637 | 7416 | |
5ecd9d40 DR |
7417 | /* Hopeless node, leave it to direct reclaim if possible */ |
7418 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES || | |
97a225e6 JK |
7419 | (pgdat_balanced(pgdat, order, highest_zoneidx) && |
7420 | !pgdat_watermark_boosted(pgdat, highest_zoneidx))) { | |
5ecd9d40 DR |
7421 | /* |
7422 | * There may be plenty of free memory available, but it's too | |
7423 | * fragmented for high-order allocations. Wake up kcompactd | |
7424 | * and rely on compaction_suitable() to determine if it's | |
7425 | * needed. If it fails, it will defer subsequent attempts to | |
7426 | * ratelimit its work. | |
7427 | */ | |
7428 | if (!(gfp_flags & __GFP_DIRECT_RECLAIM)) | |
97a225e6 | 7429 | wakeup_kcompactd(pgdat, order, highest_zoneidx); |
e716f2eb | 7430 | return; |
5ecd9d40 | 7431 | } |
88f5acf8 | 7432 | |
97a225e6 | 7433 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order, |
5ecd9d40 | 7434 | gfp_flags); |
8d0986e2 | 7435 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
7436 | } |
7437 | ||
c6f37f12 | 7438 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 7439 | /* |
7b51755c | 7440 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
7441 | * freed pages. |
7442 | * | |
7443 | * Rather than trying to age LRUs the aim is to preserve the overall | |
7444 | * LRU order by reclaiming preferentially | |
7445 | * inactive > active > active referenced > active mapped | |
1da177e4 | 7446 | */ |
7b51755c | 7447 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 7448 | { |
d6277db4 | 7449 | struct scan_control sc = { |
ee814fe2 | 7450 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 7451 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 7452 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 7453 | .priority = DEF_PRIORITY, |
d6277db4 | 7454 | .may_writepage = 1, |
ee814fe2 JW |
7455 | .may_unmap = 1, |
7456 | .may_swap = 1, | |
7b51755c | 7457 | .hibernation_mode = 1, |
1da177e4 | 7458 | }; |
a09ed5e0 | 7459 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c | 7460 | unsigned long nr_reclaimed; |
499118e9 | 7461 | unsigned int noreclaim_flag; |
1da177e4 | 7462 | |
d92a8cfc | 7463 | fs_reclaim_acquire(sc.gfp_mask); |
93781325 | 7464 | noreclaim_flag = memalloc_noreclaim_save(); |
1732d2b0 | 7465 | set_task_reclaim_state(current, &sc.reclaim_state); |
d6277db4 | 7466 | |
3115cd91 | 7467 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 7468 | |
1732d2b0 | 7469 | set_task_reclaim_state(current, NULL); |
499118e9 | 7470 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 7471 | fs_reclaim_release(sc.gfp_mask); |
d6277db4 | 7472 | |
7b51755c | 7473 | return nr_reclaimed; |
1da177e4 | 7474 | } |
c6f37f12 | 7475 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 7476 | |
3218ae14 YG |
7477 | /* |
7478 | * This kswapd start function will be called by init and node-hot-add. | |
3218ae14 | 7479 | */ |
b87c517a | 7480 | void kswapd_run(int nid) |
3218ae14 YG |
7481 | { |
7482 | pg_data_t *pgdat = NODE_DATA(nid); | |
3218ae14 | 7483 | |
b4a0215e KW |
7484 | pgdat_kswapd_lock(pgdat); |
7485 | if (!pgdat->kswapd) { | |
7486 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
7487 | if (IS_ERR(pgdat->kswapd)) { | |
7488 | /* failure at boot is fatal */ | |
7489 | BUG_ON(system_state < SYSTEM_RUNNING); | |
7490 | pr_err("Failed to start kswapd on node %d\n", nid); | |
7491 | pgdat->kswapd = NULL; | |
7492 | } | |
3218ae14 | 7493 | } |
b4a0215e | 7494 | pgdat_kswapd_unlock(pgdat); |
3218ae14 YG |
7495 | } |
7496 | ||
8fe23e05 | 7497 | /* |
d8adde17 | 7498 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
e8da368a | 7499 | * be holding mem_hotplug_begin/done(). |
8fe23e05 DR |
7500 | */ |
7501 | void kswapd_stop(int nid) | |
7502 | { | |
b4a0215e KW |
7503 | pg_data_t *pgdat = NODE_DATA(nid); |
7504 | struct task_struct *kswapd; | |
8fe23e05 | 7505 | |
b4a0215e KW |
7506 | pgdat_kswapd_lock(pgdat); |
7507 | kswapd = pgdat->kswapd; | |
d8adde17 | 7508 | if (kswapd) { |
8fe23e05 | 7509 | kthread_stop(kswapd); |
b4a0215e | 7510 | pgdat->kswapd = NULL; |
d8adde17 | 7511 | } |
b4a0215e | 7512 | pgdat_kswapd_unlock(pgdat); |
8fe23e05 DR |
7513 | } |
7514 | ||
1da177e4 LT |
7515 | static int __init kswapd_init(void) |
7516 | { | |
6b700b5b | 7517 | int nid; |
69e05944 | 7518 | |
1da177e4 | 7519 | swap_setup(); |
48fb2e24 | 7520 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 7521 | kswapd_run(nid); |
1da177e4 LT |
7522 | return 0; |
7523 | } | |
7524 | ||
7525 | module_init(kswapd_init) | |
9eeff239 CL |
7526 | |
7527 | #ifdef CONFIG_NUMA | |
7528 | /* | |
a5f5f91d | 7529 | * Node reclaim mode |
9eeff239 | 7530 | * |
a5f5f91d | 7531 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 7532 | * the watermarks. |
9eeff239 | 7533 | */ |
a5f5f91d | 7534 | int node_reclaim_mode __read_mostly; |
9eeff239 | 7535 | |
a92f7126 | 7536 | /* |
a5f5f91d | 7537 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
7538 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
7539 | * a zone. | |
7540 | */ | |
a5f5f91d | 7541 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 7542 | |
9614634f | 7543 | /* |
a5f5f91d | 7544 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
7545 | * occur. |
7546 | */ | |
7547 | int sysctl_min_unmapped_ratio = 1; | |
7548 | ||
0ff38490 CL |
7549 | /* |
7550 | * If the number of slab pages in a zone grows beyond this percentage then | |
7551 | * slab reclaim needs to occur. | |
7552 | */ | |
7553 | int sysctl_min_slab_ratio = 5; | |
7554 | ||
11fb9989 | 7555 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 7556 | { |
11fb9989 MG |
7557 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
7558 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
7559 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
7560 | |
7561 | /* | |
7562 | * It's possible for there to be more file mapped pages than | |
7563 | * accounted for by the pages on the file LRU lists because | |
7564 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
7565 | */ | |
7566 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
7567 | } | |
7568 | ||
7569 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 7570 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 7571 | { |
d031a157 AM |
7572 | unsigned long nr_pagecache_reclaimable; |
7573 | unsigned long delta = 0; | |
90afa5de MG |
7574 | |
7575 | /* | |
95bbc0c7 | 7576 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 7577 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 7578 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
7579 | * a better estimate |
7580 | */ | |
a5f5f91d MG |
7581 | if (node_reclaim_mode & RECLAIM_UNMAP) |
7582 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 7583 | else |
a5f5f91d | 7584 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
7585 | |
7586 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
7587 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
7588 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
7589 | |
7590 | /* Watch for any possible underflows due to delta */ | |
7591 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
7592 | delta = nr_pagecache_reclaimable; | |
7593 | ||
7594 | return nr_pagecache_reclaimable - delta; | |
7595 | } | |
7596 | ||
9eeff239 | 7597 | /* |
a5f5f91d | 7598 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 7599 | */ |
a5f5f91d | 7600 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 7601 | { |
7fb2d46d | 7602 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 7603 | const unsigned long nr_pages = 1 << order; |
9eeff239 | 7604 | struct task_struct *p = current; |
499118e9 | 7605 | unsigned int noreclaim_flag; |
179e9639 | 7606 | struct scan_control sc = { |
62b726c1 | 7607 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
f2f43e56 | 7608 | .gfp_mask = current_gfp_context(gfp_mask), |
bd2f6199 | 7609 | .order = order, |
a5f5f91d MG |
7610 | .priority = NODE_RECLAIM_PRIORITY, |
7611 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
7612 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 7613 | .may_swap = 1, |
f2f43e56 | 7614 | .reclaim_idx = gfp_zone(gfp_mask), |
179e9639 | 7615 | }; |
57f29762 | 7616 | unsigned long pflags; |
9eeff239 | 7617 | |
132bb8cf YS |
7618 | trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order, |
7619 | sc.gfp_mask); | |
7620 | ||
9eeff239 | 7621 | cond_resched(); |
57f29762 | 7622 | psi_memstall_enter(&pflags); |
93781325 | 7623 | fs_reclaim_acquire(sc.gfp_mask); |
d4f7796e | 7624 | /* |
95bbc0c7 | 7625 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 7626 | */ |
499118e9 | 7627 | noreclaim_flag = memalloc_noreclaim_save(); |
1732d2b0 | 7628 | set_task_reclaim_state(p, &sc.reclaim_state); |
c84db23c | 7629 | |
d8ff6fde ML |
7630 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages || |
7631 | node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) { | |
0ff38490 | 7632 | /* |
894befec | 7633 | * Free memory by calling shrink node with increasing |
0ff38490 CL |
7634 | * priorities until we have enough memory freed. |
7635 | */ | |
0ff38490 | 7636 | do { |
970a39a3 | 7637 | shrink_node(pgdat, &sc); |
9e3b2f8c | 7638 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 7639 | } |
c84db23c | 7640 | |
1732d2b0 | 7641 | set_task_reclaim_state(p, NULL); |
499118e9 | 7642 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 7643 | fs_reclaim_release(sc.gfp_mask); |
57f29762 | 7644 | psi_memstall_leave(&pflags); |
132bb8cf YS |
7645 | |
7646 | trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed); | |
7647 | ||
a79311c1 | 7648 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 7649 | } |
179e9639 | 7650 | |
a5f5f91d | 7651 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 7652 | { |
d773ed6b | 7653 | int ret; |
179e9639 AM |
7654 | |
7655 | /* | |
a5f5f91d | 7656 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 7657 | * slab pages if we are over the defined limits. |
34aa1330 | 7658 | * |
9614634f CL |
7659 | * A small portion of unmapped file backed pages is needed for |
7660 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
7661 | * thrown out if the node is overallocated. So we do not reclaim |
7662 | * if less than a specified percentage of the node is used by | |
9614634f | 7663 | * unmapped file backed pages. |
179e9639 | 7664 | */ |
a5f5f91d | 7665 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
d42f3245 RG |
7666 | node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <= |
7667 | pgdat->min_slab_pages) | |
a5f5f91d | 7668 | return NODE_RECLAIM_FULL; |
179e9639 AM |
7669 | |
7670 | /* | |
d773ed6b | 7671 | * Do not scan if the allocation should not be delayed. |
179e9639 | 7672 | */ |
d0164adc | 7673 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 7674 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
7675 | |
7676 | /* | |
a5f5f91d | 7677 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
7678 | * have associated processors. This will favor the local processor |
7679 | * over remote processors and spread off node memory allocations | |
7680 | * as wide as possible. | |
7681 | */ | |
a5f5f91d MG |
7682 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
7683 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 7684 | |
a5f5f91d MG |
7685 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
7686 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 7687 | |
a5f5f91d MG |
7688 | ret = __node_reclaim(pgdat, gfp_mask, order); |
7689 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 7690 | |
24cf7251 MG |
7691 | if (!ret) |
7692 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
7693 | ||
d773ed6b | 7694 | return ret; |
179e9639 | 7695 | } |
9eeff239 | 7696 | #endif |
894bc310 | 7697 | |
77414d19 MWO |
7698 | void check_move_unevictable_pages(struct pagevec *pvec) |
7699 | { | |
7700 | struct folio_batch fbatch; | |
7701 | unsigned i; | |
7702 | ||
7703 | folio_batch_init(&fbatch); | |
7704 | for (i = 0; i < pvec->nr; i++) { | |
7705 | struct page *page = pvec->pages[i]; | |
7706 | ||
7707 | if (PageTransTail(page)) | |
7708 | continue; | |
7709 | folio_batch_add(&fbatch, page_folio(page)); | |
7710 | } | |
7711 | check_move_unevictable_folios(&fbatch); | |
7712 | } | |
7713 | EXPORT_SYMBOL_GPL(check_move_unevictable_pages); | |
7714 | ||
89e004ea | 7715 | /** |
77414d19 MWO |
7716 | * check_move_unevictable_folios - Move evictable folios to appropriate zone |
7717 | * lru list | |
7718 | * @fbatch: Batch of lru folios to check. | |
89e004ea | 7719 | * |
77414d19 | 7720 | * Checks folios for evictability, if an evictable folio is in the unevictable |
64e3d12f | 7721 | * lru list, moves it to the appropriate evictable lru list. This function |
77414d19 | 7722 | * should be only used for lru folios. |
89e004ea | 7723 | */ |
77414d19 | 7724 | void check_move_unevictable_folios(struct folio_batch *fbatch) |
89e004ea | 7725 | { |
6168d0da | 7726 | struct lruvec *lruvec = NULL; |
24513264 HD |
7727 | int pgscanned = 0; |
7728 | int pgrescued = 0; | |
7729 | int i; | |
89e004ea | 7730 | |
77414d19 MWO |
7731 | for (i = 0; i < fbatch->nr; i++) { |
7732 | struct folio *folio = fbatch->folios[i]; | |
7733 | int nr_pages = folio_nr_pages(folio); | |
8d8869ca | 7734 | |
8d8869ca | 7735 | pgscanned += nr_pages; |
89e004ea | 7736 | |
77414d19 MWO |
7737 | /* block memcg migration while the folio moves between lrus */ |
7738 | if (!folio_test_clear_lru(folio)) | |
d25b5bd8 AS |
7739 | continue; |
7740 | ||
0de340cb | 7741 | lruvec = folio_lruvec_relock_irq(folio, lruvec); |
77414d19 MWO |
7742 | if (folio_evictable(folio) && folio_test_unevictable(folio)) { |
7743 | lruvec_del_folio(lruvec, folio); | |
7744 | folio_clear_unevictable(folio); | |
7745 | lruvec_add_folio(lruvec, folio); | |
8d8869ca | 7746 | pgrescued += nr_pages; |
89e004ea | 7747 | } |
77414d19 | 7748 | folio_set_lru(folio); |
24513264 | 7749 | } |
89e004ea | 7750 | |
6168d0da | 7751 | if (lruvec) { |
24513264 HD |
7752 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
7753 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
6168d0da | 7754 | unlock_page_lruvec_irq(lruvec); |
d25b5bd8 AS |
7755 | } else if (pgscanned) { |
7756 | count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
89e004ea | 7757 | } |
89e004ea | 7758 | } |
77414d19 | 7759 | EXPORT_SYMBOL_GPL(check_move_unevictable_folios); |