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> |
929bea7c | 46 | #include <linux/oom.h> |
64e3d12f | 47 | #include <linux/pagevec.h> |
268bb0ce | 48 | #include <linux/prefetch.h> |
b1de0d13 | 49 | #include <linux/printk.h> |
f9fe48be | 50 | #include <linux/dax.h> |
eb414681 | 51 | #include <linux/psi.h> |
bd74fdae YZ |
52 | #include <linux/pagewalk.h> |
53 | #include <linux/shmem_fs.h> | |
354ed597 | 54 | #include <linux/ctype.h> |
1da177e4 LT |
55 | |
56 | #include <asm/tlbflush.h> | |
57 | #include <asm/div64.h> | |
58 | ||
59 | #include <linux/swapops.h> | |
117aad1e | 60 | #include <linux/balloon_compaction.h> |
c574bbe9 | 61 | #include <linux/sched/sysctl.h> |
1da177e4 | 62 | |
0f8053a5 | 63 | #include "internal.h" |
014bb1de | 64 | #include "swap.h" |
0f8053a5 | 65 | |
33906bc5 MG |
66 | #define CREATE_TRACE_POINTS |
67 | #include <trace/events/vmscan.h> | |
68 | ||
1da177e4 | 69 | struct scan_control { |
22fba335 KM |
70 | /* How many pages shrink_list() should reclaim */ |
71 | unsigned long nr_to_reclaim; | |
72 | ||
ee814fe2 JW |
73 | /* |
74 | * Nodemask of nodes allowed by the caller. If NULL, all nodes | |
75 | * are scanned. | |
76 | */ | |
77 | nodemask_t *nodemask; | |
9e3b2f8c | 78 | |
f16015fb JW |
79 | /* |
80 | * The memory cgroup that hit its limit and as a result is the | |
81 | * primary target of this reclaim invocation. | |
82 | */ | |
83 | struct mem_cgroup *target_mem_cgroup; | |
66e1707b | 84 | |
7cf111bc JW |
85 | /* |
86 | * Scan pressure balancing between anon and file LRUs | |
87 | */ | |
88 | unsigned long anon_cost; | |
89 | unsigned long file_cost; | |
90 | ||
b91ac374 JW |
91 | /* Can active pages be deactivated as part of reclaim? */ |
92 | #define DEACTIVATE_ANON 1 | |
93 | #define DEACTIVATE_FILE 2 | |
94 | unsigned int may_deactivate:2; | |
95 | unsigned int force_deactivate:1; | |
96 | unsigned int skipped_deactivate:1; | |
97 | ||
1276ad68 | 98 | /* Writepage batching in laptop mode; RECLAIM_WRITE */ |
ee814fe2 JW |
99 | unsigned int may_writepage:1; |
100 | ||
101 | /* Can mapped pages be reclaimed? */ | |
102 | unsigned int may_unmap:1; | |
103 | ||
104 | /* Can pages be swapped as part of reclaim? */ | |
105 | unsigned int may_swap:1; | |
106 | ||
73b73bac YA |
107 | /* Proactive reclaim invoked by userspace through memory.reclaim */ |
108 | unsigned int proactive:1; | |
109 | ||
d6622f63 | 110 | /* |
f56ce412 JW |
111 | * Cgroup memory below memory.low is protected as long as we |
112 | * don't threaten to OOM. If any cgroup is reclaimed at | |
113 | * reduced force or passed over entirely due to its memory.low | |
114 | * setting (memcg_low_skipped), and nothing is reclaimed as a | |
115 | * result, then go back for one more cycle that reclaims the protected | |
116 | * memory (memcg_low_reclaim) to avert OOM. | |
d6622f63 YX |
117 | */ |
118 | unsigned int memcg_low_reclaim:1; | |
119 | unsigned int memcg_low_skipped:1; | |
241994ed | 120 | |
ee814fe2 JW |
121 | unsigned int hibernation_mode:1; |
122 | ||
123 | /* One of the zones is ready for compaction */ | |
124 | unsigned int compaction_ready:1; | |
125 | ||
b91ac374 JW |
126 | /* There is easily reclaimable cold cache in the current node */ |
127 | unsigned int cache_trim_mode:1; | |
128 | ||
53138cea JW |
129 | /* The file pages on the current node are dangerously low */ |
130 | unsigned int file_is_tiny:1; | |
131 | ||
26aa2d19 DH |
132 | /* Always discard instead of demoting to lower tier memory */ |
133 | unsigned int no_demotion:1; | |
134 | ||
f76c8337 YZ |
135 | #ifdef CONFIG_LRU_GEN |
136 | /* help kswapd make better choices among multiple memcgs */ | |
137 | unsigned int memcgs_need_aging:1; | |
138 | unsigned long last_reclaimed; | |
139 | #endif | |
140 | ||
bb451fdf GT |
141 | /* Allocation order */ |
142 | s8 order; | |
143 | ||
144 | /* Scan (total_size >> priority) pages at once */ | |
145 | s8 priority; | |
146 | ||
147 | /* The highest zone to isolate pages for reclaim from */ | |
148 | s8 reclaim_idx; | |
149 | ||
150 | /* This context's GFP mask */ | |
151 | gfp_t gfp_mask; | |
152 | ||
ee814fe2 JW |
153 | /* Incremented by the number of inactive pages that were scanned */ |
154 | unsigned long nr_scanned; | |
155 | ||
156 | /* Number of pages freed so far during a call to shrink_zones() */ | |
157 | unsigned long nr_reclaimed; | |
d108c772 AR |
158 | |
159 | struct { | |
160 | unsigned int dirty; | |
161 | unsigned int unqueued_dirty; | |
162 | unsigned int congested; | |
163 | unsigned int writeback; | |
164 | unsigned int immediate; | |
165 | unsigned int file_taken; | |
166 | unsigned int taken; | |
167 | } nr; | |
e5ca8071 YS |
168 | |
169 | /* for recording the reclaimed slab by now */ | |
170 | struct reclaim_state reclaim_state; | |
1da177e4 LT |
171 | }; |
172 | ||
1da177e4 | 173 | #ifdef ARCH_HAS_PREFETCHW |
166e3d32 | 174 | #define prefetchw_prev_lru_folio(_folio, _base, _field) \ |
1da177e4 | 175 | do { \ |
166e3d32 MWO |
176 | if ((_folio)->lru.prev != _base) { \ |
177 | struct folio *prev; \ | |
1da177e4 | 178 | \ |
166e3d32 | 179 | prev = lru_to_folio(&(_folio->lru)); \ |
1da177e4 LT |
180 | prefetchw(&prev->_field); \ |
181 | } \ | |
182 | } while (0) | |
183 | #else | |
166e3d32 | 184 | #define prefetchw_prev_lru_folio(_folio, _base, _field) do { } while (0) |
1da177e4 LT |
185 | #endif |
186 | ||
187 | /* | |
c843966c | 188 | * From 0 .. 200. Higher means more swappy. |
1da177e4 LT |
189 | */ |
190 | int vm_swappiness = 60; | |
1da177e4 | 191 | |
0a432dcb YS |
192 | static void set_task_reclaim_state(struct task_struct *task, |
193 | struct reclaim_state *rs) | |
194 | { | |
195 | /* Check for an overwrite */ | |
196 | WARN_ON_ONCE(rs && task->reclaim_state); | |
197 | ||
198 | /* Check for the nulling of an already-nulled member */ | |
199 | WARN_ON_ONCE(!rs && !task->reclaim_state); | |
200 | ||
201 | task->reclaim_state = rs; | |
202 | } | |
203 | ||
5035ebc6 RG |
204 | LIST_HEAD(shrinker_list); |
205 | DECLARE_RWSEM(shrinker_rwsem); | |
1da177e4 | 206 | |
0a432dcb | 207 | #ifdef CONFIG_MEMCG |
a2fb1261 | 208 | static int shrinker_nr_max; |
2bfd3637 | 209 | |
3c6f17e6 | 210 | /* The shrinker_info is expanded in a batch of BITS_PER_LONG */ |
a2fb1261 YS |
211 | static inline int shrinker_map_size(int nr_items) |
212 | { | |
213 | return (DIV_ROUND_UP(nr_items, BITS_PER_LONG) * sizeof(unsigned long)); | |
214 | } | |
2bfd3637 | 215 | |
3c6f17e6 YS |
216 | static inline int shrinker_defer_size(int nr_items) |
217 | { | |
218 | return (round_up(nr_items, BITS_PER_LONG) * sizeof(atomic_long_t)); | |
219 | } | |
220 | ||
468ab843 YS |
221 | static struct shrinker_info *shrinker_info_protected(struct mem_cgroup *memcg, |
222 | int nid) | |
223 | { | |
224 | return rcu_dereference_protected(memcg->nodeinfo[nid]->shrinker_info, | |
225 | lockdep_is_held(&shrinker_rwsem)); | |
226 | } | |
227 | ||
e4262c4f | 228 | static int expand_one_shrinker_info(struct mem_cgroup *memcg, |
3c6f17e6 YS |
229 | int map_size, int defer_size, |
230 | int old_map_size, int old_defer_size) | |
2bfd3637 | 231 | { |
e4262c4f | 232 | struct shrinker_info *new, *old; |
2bfd3637 YS |
233 | struct mem_cgroup_per_node *pn; |
234 | int nid; | |
3c6f17e6 | 235 | int size = map_size + defer_size; |
2bfd3637 | 236 | |
2bfd3637 YS |
237 | for_each_node(nid) { |
238 | pn = memcg->nodeinfo[nid]; | |
468ab843 | 239 | old = shrinker_info_protected(memcg, nid); |
2bfd3637 YS |
240 | /* Not yet online memcg */ |
241 | if (!old) | |
242 | return 0; | |
243 | ||
244 | new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid); | |
245 | if (!new) | |
246 | return -ENOMEM; | |
247 | ||
3c6f17e6 YS |
248 | new->nr_deferred = (atomic_long_t *)(new + 1); |
249 | new->map = (void *)new->nr_deferred + defer_size; | |
250 | ||
251 | /* map: set all old bits, clear all new bits */ | |
252 | memset(new->map, (int)0xff, old_map_size); | |
253 | memset((void *)new->map + old_map_size, 0, map_size - old_map_size); | |
254 | /* nr_deferred: copy old values, clear all new values */ | |
255 | memcpy(new->nr_deferred, old->nr_deferred, old_defer_size); | |
256 | memset((void *)new->nr_deferred + old_defer_size, 0, | |
257 | defer_size - old_defer_size); | |
2bfd3637 | 258 | |
e4262c4f | 259 | rcu_assign_pointer(pn->shrinker_info, new); |
72673e86 | 260 | kvfree_rcu(old, rcu); |
2bfd3637 YS |
261 | } |
262 | ||
263 | return 0; | |
264 | } | |
265 | ||
e4262c4f | 266 | void free_shrinker_info(struct mem_cgroup *memcg) |
2bfd3637 YS |
267 | { |
268 | struct mem_cgroup_per_node *pn; | |
e4262c4f | 269 | struct shrinker_info *info; |
2bfd3637 YS |
270 | int nid; |
271 | ||
2bfd3637 YS |
272 | for_each_node(nid) { |
273 | pn = memcg->nodeinfo[nid]; | |
e4262c4f YS |
274 | info = rcu_dereference_protected(pn->shrinker_info, true); |
275 | kvfree(info); | |
276 | rcu_assign_pointer(pn->shrinker_info, NULL); | |
2bfd3637 YS |
277 | } |
278 | } | |
279 | ||
e4262c4f | 280 | int alloc_shrinker_info(struct mem_cgroup *memcg) |
2bfd3637 | 281 | { |
e4262c4f | 282 | struct shrinker_info *info; |
2bfd3637 | 283 | int nid, size, ret = 0; |
3c6f17e6 | 284 | int map_size, defer_size = 0; |
2bfd3637 | 285 | |
d27cf2aa | 286 | down_write(&shrinker_rwsem); |
3c6f17e6 YS |
287 | map_size = shrinker_map_size(shrinker_nr_max); |
288 | defer_size = shrinker_defer_size(shrinker_nr_max); | |
289 | size = map_size + defer_size; | |
2bfd3637 | 290 | for_each_node(nid) { |
e4262c4f YS |
291 | info = kvzalloc_node(sizeof(*info) + size, GFP_KERNEL, nid); |
292 | if (!info) { | |
293 | free_shrinker_info(memcg); | |
2bfd3637 YS |
294 | ret = -ENOMEM; |
295 | break; | |
296 | } | |
3c6f17e6 YS |
297 | info->nr_deferred = (atomic_long_t *)(info + 1); |
298 | info->map = (void *)info->nr_deferred + defer_size; | |
e4262c4f | 299 | rcu_assign_pointer(memcg->nodeinfo[nid]->shrinker_info, info); |
2bfd3637 | 300 | } |
d27cf2aa | 301 | up_write(&shrinker_rwsem); |
2bfd3637 YS |
302 | |
303 | return ret; | |
304 | } | |
305 | ||
3c6f17e6 YS |
306 | static inline bool need_expand(int nr_max) |
307 | { | |
308 | return round_up(nr_max, BITS_PER_LONG) > | |
309 | round_up(shrinker_nr_max, BITS_PER_LONG); | |
310 | } | |
311 | ||
e4262c4f | 312 | static int expand_shrinker_info(int new_id) |
2bfd3637 | 313 | { |
3c6f17e6 | 314 | int ret = 0; |
a2fb1261 | 315 | int new_nr_max = new_id + 1; |
3c6f17e6 YS |
316 | int map_size, defer_size = 0; |
317 | int old_map_size, old_defer_size = 0; | |
2bfd3637 YS |
318 | struct mem_cgroup *memcg; |
319 | ||
3c6f17e6 | 320 | if (!need_expand(new_nr_max)) |
a2fb1261 | 321 | goto out; |
2bfd3637 | 322 | |
2bfd3637 | 323 | if (!root_mem_cgroup) |
d27cf2aa YS |
324 | goto out; |
325 | ||
326 | lockdep_assert_held(&shrinker_rwsem); | |
2bfd3637 | 327 | |
3c6f17e6 YS |
328 | map_size = shrinker_map_size(new_nr_max); |
329 | defer_size = shrinker_defer_size(new_nr_max); | |
330 | old_map_size = shrinker_map_size(shrinker_nr_max); | |
331 | old_defer_size = shrinker_defer_size(shrinker_nr_max); | |
332 | ||
2bfd3637 YS |
333 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
334 | do { | |
3c6f17e6 YS |
335 | ret = expand_one_shrinker_info(memcg, map_size, defer_size, |
336 | old_map_size, old_defer_size); | |
2bfd3637 YS |
337 | if (ret) { |
338 | mem_cgroup_iter_break(NULL, memcg); | |
d27cf2aa | 339 | goto out; |
2bfd3637 YS |
340 | } |
341 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); | |
d27cf2aa | 342 | out: |
2bfd3637 | 343 | if (!ret) |
a2fb1261 | 344 | shrinker_nr_max = new_nr_max; |
d27cf2aa | 345 | |
2bfd3637 YS |
346 | return ret; |
347 | } | |
348 | ||
349 | void set_shrinker_bit(struct mem_cgroup *memcg, int nid, int shrinker_id) | |
350 | { | |
351 | if (shrinker_id >= 0 && memcg && !mem_cgroup_is_root(memcg)) { | |
e4262c4f | 352 | struct shrinker_info *info; |
2bfd3637 YS |
353 | |
354 | rcu_read_lock(); | |
e4262c4f | 355 | info = rcu_dereference(memcg->nodeinfo[nid]->shrinker_info); |
2bfd3637 YS |
356 | /* Pairs with smp mb in shrink_slab() */ |
357 | smp_mb__before_atomic(); | |
e4262c4f | 358 | set_bit(shrinker_id, info->map); |
2bfd3637 YS |
359 | rcu_read_unlock(); |
360 | } | |
361 | } | |
362 | ||
b4c2b231 | 363 | static DEFINE_IDR(shrinker_idr); |
b4c2b231 KT |
364 | |
365 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) | |
366 | { | |
367 | int id, ret = -ENOMEM; | |
368 | ||
476b30a0 YS |
369 | if (mem_cgroup_disabled()) |
370 | return -ENOSYS; | |
371 | ||
b4c2b231 KT |
372 | down_write(&shrinker_rwsem); |
373 | /* This may call shrinker, so it must use down_read_trylock() */ | |
41ca668a | 374 | id = idr_alloc(&shrinker_idr, shrinker, 0, 0, GFP_KERNEL); |
b4c2b231 KT |
375 | if (id < 0) |
376 | goto unlock; | |
377 | ||
0a4465d3 | 378 | if (id >= shrinker_nr_max) { |
e4262c4f | 379 | if (expand_shrinker_info(id)) { |
0a4465d3 KT |
380 | idr_remove(&shrinker_idr, id); |
381 | goto unlock; | |
382 | } | |
0a4465d3 | 383 | } |
b4c2b231 KT |
384 | shrinker->id = id; |
385 | ret = 0; | |
386 | unlock: | |
387 | up_write(&shrinker_rwsem); | |
388 | return ret; | |
389 | } | |
390 | ||
391 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
392 | { | |
393 | int id = shrinker->id; | |
394 | ||
395 | BUG_ON(id < 0); | |
396 | ||
41ca668a YS |
397 | lockdep_assert_held(&shrinker_rwsem); |
398 | ||
b4c2b231 | 399 | idr_remove(&shrinker_idr, id); |
b4c2b231 | 400 | } |
b4c2b231 | 401 | |
86750830 YS |
402 | static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker, |
403 | struct mem_cgroup *memcg) | |
404 | { | |
405 | struct shrinker_info *info; | |
406 | ||
407 | info = shrinker_info_protected(memcg, nid); | |
408 | return atomic_long_xchg(&info->nr_deferred[shrinker->id], 0); | |
409 | } | |
410 | ||
411 | static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker, | |
412 | struct mem_cgroup *memcg) | |
413 | { | |
414 | struct shrinker_info *info; | |
415 | ||
416 | info = shrinker_info_protected(memcg, nid); | |
417 | return atomic_long_add_return(nr, &info->nr_deferred[shrinker->id]); | |
418 | } | |
419 | ||
a178015c YS |
420 | void reparent_shrinker_deferred(struct mem_cgroup *memcg) |
421 | { | |
422 | int i, nid; | |
423 | long nr; | |
424 | struct mem_cgroup *parent; | |
425 | struct shrinker_info *child_info, *parent_info; | |
426 | ||
427 | parent = parent_mem_cgroup(memcg); | |
428 | if (!parent) | |
429 | parent = root_mem_cgroup; | |
430 | ||
431 | /* Prevent from concurrent shrinker_info expand */ | |
432 | down_read(&shrinker_rwsem); | |
433 | for_each_node(nid) { | |
434 | child_info = shrinker_info_protected(memcg, nid); | |
435 | parent_info = shrinker_info_protected(parent, nid); | |
436 | for (i = 0; i < shrinker_nr_max; i++) { | |
437 | nr = atomic_long_read(&child_info->nr_deferred[i]); | |
438 | atomic_long_add(nr, &parent_info->nr_deferred[i]); | |
439 | } | |
440 | } | |
441 | up_read(&shrinker_rwsem); | |
442 | } | |
443 | ||
b5ead35e | 444 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 445 | { |
b5ead35e | 446 | return sc->target_mem_cgroup; |
89b5fae5 | 447 | } |
97c9341f TH |
448 | |
449 | /** | |
b5ead35e | 450 | * writeback_throttling_sane - is the usual dirty throttling mechanism available? |
97c9341f TH |
451 | * @sc: scan_control in question |
452 | * | |
453 | * The normal page dirty throttling mechanism in balance_dirty_pages() is | |
454 | * completely broken with the legacy memcg and direct stalling in | |
455 | * shrink_page_list() is used for throttling instead, which lacks all the | |
456 | * niceties such as fairness, adaptive pausing, bandwidth proportional | |
457 | * allocation and configurability. | |
458 | * | |
459 | * This function tests whether the vmscan currently in progress can assume | |
460 | * that the normal dirty throttling mechanism is operational. | |
461 | */ | |
b5ead35e | 462 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f | 463 | { |
b5ead35e | 464 | if (!cgroup_reclaim(sc)) |
97c9341f TH |
465 | return true; |
466 | #ifdef CONFIG_CGROUP_WRITEBACK | |
69234ace | 467 | if (cgroup_subsys_on_dfl(memory_cgrp_subsys)) |
97c9341f TH |
468 | return true; |
469 | #endif | |
470 | return false; | |
471 | } | |
91a45470 | 472 | #else |
0a432dcb YS |
473 | static int prealloc_memcg_shrinker(struct shrinker *shrinker) |
474 | { | |
476b30a0 | 475 | return -ENOSYS; |
0a432dcb YS |
476 | } |
477 | ||
478 | static void unregister_memcg_shrinker(struct shrinker *shrinker) | |
479 | { | |
480 | } | |
481 | ||
86750830 YS |
482 | static long xchg_nr_deferred_memcg(int nid, struct shrinker *shrinker, |
483 | struct mem_cgroup *memcg) | |
484 | { | |
485 | return 0; | |
486 | } | |
487 | ||
488 | static long add_nr_deferred_memcg(long nr, int nid, struct shrinker *shrinker, | |
489 | struct mem_cgroup *memcg) | |
490 | { | |
491 | return 0; | |
492 | } | |
493 | ||
b5ead35e | 494 | static bool cgroup_reclaim(struct scan_control *sc) |
89b5fae5 | 495 | { |
b5ead35e | 496 | return false; |
89b5fae5 | 497 | } |
97c9341f | 498 | |
b5ead35e | 499 | static bool writeback_throttling_sane(struct scan_control *sc) |
97c9341f TH |
500 | { |
501 | return true; | |
502 | } | |
91a45470 KH |
503 | #endif |
504 | ||
86750830 YS |
505 | static long xchg_nr_deferred(struct shrinker *shrinker, |
506 | struct shrink_control *sc) | |
507 | { | |
508 | int nid = sc->nid; | |
509 | ||
510 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) | |
511 | nid = 0; | |
512 | ||
513 | if (sc->memcg && | |
514 | (shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
515 | return xchg_nr_deferred_memcg(nid, shrinker, | |
516 | sc->memcg); | |
517 | ||
518 | return atomic_long_xchg(&shrinker->nr_deferred[nid], 0); | |
519 | } | |
520 | ||
521 | ||
522 | static long add_nr_deferred(long nr, struct shrinker *shrinker, | |
523 | struct shrink_control *sc) | |
524 | { | |
525 | int nid = sc->nid; | |
526 | ||
527 | if (!(shrinker->flags & SHRINKER_NUMA_AWARE)) | |
528 | nid = 0; | |
529 | ||
530 | if (sc->memcg && | |
531 | (shrinker->flags & SHRINKER_MEMCG_AWARE)) | |
532 | return add_nr_deferred_memcg(nr, nid, shrinker, | |
533 | sc->memcg); | |
534 | ||
535 | return atomic_long_add_return(nr, &shrinker->nr_deferred[nid]); | |
536 | } | |
537 | ||
26aa2d19 DH |
538 | static bool can_demote(int nid, struct scan_control *sc) |
539 | { | |
20b51af1 HY |
540 | if (!numa_demotion_enabled) |
541 | return false; | |
3f1509c5 JW |
542 | if (sc && sc->no_demotion) |
543 | return false; | |
26aa2d19 DH |
544 | if (next_demotion_node(nid) == NUMA_NO_NODE) |
545 | return false; | |
546 | ||
20b51af1 | 547 | return true; |
26aa2d19 DH |
548 | } |
549 | ||
a2a36488 KB |
550 | static inline bool can_reclaim_anon_pages(struct mem_cgroup *memcg, |
551 | int nid, | |
552 | struct scan_control *sc) | |
553 | { | |
554 | if (memcg == NULL) { | |
555 | /* | |
556 | * For non-memcg reclaim, is there | |
557 | * space in any swap device? | |
558 | */ | |
559 | if (get_nr_swap_pages() > 0) | |
560 | return true; | |
561 | } else { | |
562 | /* Is the memcg below its swap limit? */ | |
563 | if (mem_cgroup_get_nr_swap_pages(memcg) > 0) | |
564 | return true; | |
565 | } | |
566 | ||
567 | /* | |
568 | * The page can not be swapped. | |
569 | * | |
570 | * Can it be reclaimed from this node via demotion? | |
571 | */ | |
572 | return can_demote(nid, sc); | |
573 | } | |
574 | ||
5a1c84b4 MG |
575 | /* |
576 | * This misses isolated pages which are not accounted for to save counters. | |
577 | * As the data only determines if reclaim or compaction continues, it is | |
578 | * not expected that isolated pages will be a dominating factor. | |
579 | */ | |
580 | unsigned long zone_reclaimable_pages(struct zone *zone) | |
581 | { | |
582 | unsigned long nr; | |
583 | ||
584 | nr = zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_FILE) + | |
585 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_FILE); | |
a2a36488 | 586 | if (can_reclaim_anon_pages(NULL, zone_to_nid(zone), NULL)) |
5a1c84b4 MG |
587 | nr += zone_page_state_snapshot(zone, NR_ZONE_INACTIVE_ANON) + |
588 | zone_page_state_snapshot(zone, NR_ZONE_ACTIVE_ANON); | |
589 | ||
590 | return nr; | |
591 | } | |
592 | ||
fd538803 MH |
593 | /** |
594 | * lruvec_lru_size - Returns the number of pages on the given LRU list. | |
595 | * @lruvec: lru vector | |
596 | * @lru: lru to use | |
8b3a899a | 597 | * @zone_idx: zones to consider (use MAX_NR_ZONES - 1 for the whole LRU list) |
fd538803 | 598 | */ |
2091339d YZ |
599 | static unsigned long lruvec_lru_size(struct lruvec *lruvec, enum lru_list lru, |
600 | int zone_idx) | |
c9f299d9 | 601 | { |
de3b0150 | 602 | unsigned long size = 0; |
fd538803 MH |
603 | int zid; |
604 | ||
8b3a899a | 605 | for (zid = 0; zid <= zone_idx; zid++) { |
fd538803 | 606 | struct zone *zone = &lruvec_pgdat(lruvec)->node_zones[zid]; |
c9f299d9 | 607 | |
fd538803 MH |
608 | if (!managed_zone(zone)) |
609 | continue; | |
610 | ||
611 | if (!mem_cgroup_disabled()) | |
de3b0150 | 612 | size += mem_cgroup_get_zone_lru_size(lruvec, lru, zid); |
fd538803 | 613 | else |
de3b0150 | 614 | size += zone_page_state(zone, NR_ZONE_LRU_BASE + lru); |
fd538803 | 615 | } |
de3b0150 | 616 | return size; |
b4536f0c MH |
617 | } |
618 | ||
1da177e4 | 619 | /* |
1d3d4437 | 620 | * Add a shrinker callback to be called from the vm. |
1da177e4 | 621 | */ |
e33c267a | 622 | static int __prealloc_shrinker(struct shrinker *shrinker) |
1da177e4 | 623 | { |
476b30a0 YS |
624 | unsigned int size; |
625 | int err; | |
626 | ||
627 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) { | |
628 | err = prealloc_memcg_shrinker(shrinker); | |
629 | if (err != -ENOSYS) | |
630 | return err; | |
1d3d4437 | 631 | |
476b30a0 YS |
632 | shrinker->flags &= ~SHRINKER_MEMCG_AWARE; |
633 | } | |
634 | ||
635 | size = sizeof(*shrinker->nr_deferred); | |
1d3d4437 GC |
636 | if (shrinker->flags & SHRINKER_NUMA_AWARE) |
637 | size *= nr_node_ids; | |
638 | ||
639 | shrinker->nr_deferred = kzalloc(size, GFP_KERNEL); | |
640 | if (!shrinker->nr_deferred) | |
641 | return -ENOMEM; | |
b4c2b231 | 642 | |
8e04944f TH |
643 | return 0; |
644 | } | |
645 | ||
e33c267a RG |
646 | #ifdef CONFIG_SHRINKER_DEBUG |
647 | int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...) | |
648 | { | |
649 | va_list ap; | |
650 | int err; | |
651 | ||
652 | va_start(ap, fmt); | |
653 | shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap); | |
654 | va_end(ap); | |
655 | if (!shrinker->name) | |
656 | return -ENOMEM; | |
657 | ||
658 | err = __prealloc_shrinker(shrinker); | |
14773bfa | 659 | if (err) { |
e33c267a | 660 | kfree_const(shrinker->name); |
14773bfa TH |
661 | shrinker->name = NULL; |
662 | } | |
e33c267a RG |
663 | |
664 | return err; | |
665 | } | |
666 | #else | |
667 | int prealloc_shrinker(struct shrinker *shrinker, const char *fmt, ...) | |
668 | { | |
669 | return __prealloc_shrinker(shrinker); | |
670 | } | |
671 | #endif | |
672 | ||
8e04944f TH |
673 | void free_prealloced_shrinker(struct shrinker *shrinker) |
674 | { | |
e33c267a RG |
675 | #ifdef CONFIG_SHRINKER_DEBUG |
676 | kfree_const(shrinker->name); | |
14773bfa | 677 | shrinker->name = NULL; |
e33c267a | 678 | #endif |
41ca668a YS |
679 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) { |
680 | down_write(&shrinker_rwsem); | |
b4c2b231 | 681 | unregister_memcg_shrinker(shrinker); |
41ca668a | 682 | up_write(&shrinker_rwsem); |
476b30a0 | 683 | return; |
41ca668a | 684 | } |
b4c2b231 | 685 | |
8e04944f TH |
686 | kfree(shrinker->nr_deferred); |
687 | shrinker->nr_deferred = NULL; | |
688 | } | |
1d3d4437 | 689 | |
8e04944f TH |
690 | void register_shrinker_prepared(struct shrinker *shrinker) |
691 | { | |
8e1f936b RR |
692 | down_write(&shrinker_rwsem); |
693 | list_add_tail(&shrinker->list, &shrinker_list); | |
41ca668a | 694 | shrinker->flags |= SHRINKER_REGISTERED; |
5035ebc6 | 695 | shrinker_debugfs_add(shrinker); |
8e1f936b | 696 | up_write(&shrinker_rwsem); |
8e04944f TH |
697 | } |
698 | ||
e33c267a | 699 | static int __register_shrinker(struct shrinker *shrinker) |
8e04944f | 700 | { |
e33c267a | 701 | int err = __prealloc_shrinker(shrinker); |
8e04944f TH |
702 | |
703 | if (err) | |
704 | return err; | |
705 | register_shrinker_prepared(shrinker); | |
1d3d4437 | 706 | return 0; |
1da177e4 | 707 | } |
e33c267a RG |
708 | |
709 | #ifdef CONFIG_SHRINKER_DEBUG | |
710 | int register_shrinker(struct shrinker *shrinker, const char *fmt, ...) | |
711 | { | |
712 | va_list ap; | |
713 | int err; | |
714 | ||
715 | va_start(ap, fmt); | |
716 | shrinker->name = kvasprintf_const(GFP_KERNEL, fmt, ap); | |
717 | va_end(ap); | |
718 | if (!shrinker->name) | |
719 | return -ENOMEM; | |
720 | ||
721 | err = __register_shrinker(shrinker); | |
14773bfa | 722 | if (err) { |
e33c267a | 723 | kfree_const(shrinker->name); |
14773bfa TH |
724 | shrinker->name = NULL; |
725 | } | |
e33c267a RG |
726 | return err; |
727 | } | |
728 | #else | |
729 | int register_shrinker(struct shrinker *shrinker, const char *fmt, ...) | |
730 | { | |
731 | return __register_shrinker(shrinker); | |
732 | } | |
733 | #endif | |
8e1f936b | 734 | EXPORT_SYMBOL(register_shrinker); |
1da177e4 LT |
735 | |
736 | /* | |
737 | * Remove one | |
738 | */ | |
8e1f936b | 739 | void unregister_shrinker(struct shrinker *shrinker) |
1da177e4 | 740 | { |
41ca668a | 741 | if (!(shrinker->flags & SHRINKER_REGISTERED)) |
bb422a73 | 742 | return; |
41ca668a | 743 | |
1da177e4 LT |
744 | down_write(&shrinker_rwsem); |
745 | list_del(&shrinker->list); | |
41ca668a YS |
746 | shrinker->flags &= ~SHRINKER_REGISTERED; |
747 | if (shrinker->flags & SHRINKER_MEMCG_AWARE) | |
748 | unregister_memcg_shrinker(shrinker); | |
5035ebc6 | 749 | shrinker_debugfs_remove(shrinker); |
1da177e4 | 750 | up_write(&shrinker_rwsem); |
41ca668a | 751 | |
ae393321 | 752 | kfree(shrinker->nr_deferred); |
bb422a73 | 753 | shrinker->nr_deferred = NULL; |
1da177e4 | 754 | } |
8e1f936b | 755 | EXPORT_SYMBOL(unregister_shrinker); |
1da177e4 | 756 | |
880121be CK |
757 | /** |
758 | * synchronize_shrinkers - Wait for all running shrinkers to complete. | |
759 | * | |
760 | * This is equivalent to calling unregister_shrink() and register_shrinker(), | |
761 | * but atomically and with less overhead. This is useful to guarantee that all | |
762 | * shrinker invocations have seen an update, before freeing memory, similar to | |
763 | * rcu. | |
764 | */ | |
765 | void synchronize_shrinkers(void) | |
766 | { | |
767 | down_write(&shrinker_rwsem); | |
768 | up_write(&shrinker_rwsem); | |
769 | } | |
770 | EXPORT_SYMBOL(synchronize_shrinkers); | |
771 | ||
1da177e4 | 772 | #define SHRINK_BATCH 128 |
1d3d4437 | 773 | |
cb731d6c | 774 | static unsigned long do_shrink_slab(struct shrink_control *shrinkctl, |
9092c71b | 775 | struct shrinker *shrinker, int priority) |
1d3d4437 GC |
776 | { |
777 | unsigned long freed = 0; | |
778 | unsigned long long delta; | |
779 | long total_scan; | |
d5bc5fd3 | 780 | long freeable; |
1d3d4437 GC |
781 | long nr; |
782 | long new_nr; | |
1d3d4437 GC |
783 | long batch_size = shrinker->batch ? shrinker->batch |
784 | : SHRINK_BATCH; | |
5f33a080 | 785 | long scanned = 0, next_deferred; |
1d3d4437 | 786 | |
d5bc5fd3 | 787 | freeable = shrinker->count_objects(shrinker, shrinkctl); |
9b996468 KT |
788 | if (freeable == 0 || freeable == SHRINK_EMPTY) |
789 | return freeable; | |
1d3d4437 GC |
790 | |
791 | /* | |
792 | * copy the current shrinker scan count into a local variable | |
793 | * and zero it so that other concurrent shrinker invocations | |
794 | * don't also do this scanning work. | |
795 | */ | |
86750830 | 796 | nr = xchg_nr_deferred(shrinker, shrinkctl); |
1d3d4437 | 797 | |
4b85afbd JW |
798 | if (shrinker->seeks) { |
799 | delta = freeable >> priority; | |
800 | delta *= 4; | |
801 | do_div(delta, shrinker->seeks); | |
802 | } else { | |
803 | /* | |
804 | * These objects don't require any IO to create. Trim | |
805 | * them aggressively under memory pressure to keep | |
806 | * them from causing refetches in the IO caches. | |
807 | */ | |
808 | delta = freeable / 2; | |
809 | } | |
172b06c3 | 810 | |
18bb473e | 811 | total_scan = nr >> priority; |
1d3d4437 | 812 | total_scan += delta; |
18bb473e | 813 | total_scan = min(total_scan, (2 * freeable)); |
1d3d4437 GC |
814 | |
815 | trace_mm_shrink_slab_start(shrinker, shrinkctl, nr, | |
9092c71b | 816 | freeable, delta, total_scan, priority); |
1d3d4437 | 817 | |
0b1fb40a VD |
818 | /* |
819 | * Normally, we should not scan less than batch_size objects in one | |
820 | * pass to avoid too frequent shrinker calls, but if the slab has less | |
821 | * than batch_size objects in total and we are really tight on memory, | |
822 | * we will try to reclaim all available objects, otherwise we can end | |
823 | * up failing allocations although there are plenty of reclaimable | |
824 | * objects spread over several slabs with usage less than the | |
825 | * batch_size. | |
826 | * | |
827 | * We detect the "tight on memory" situations by looking at the total | |
828 | * number of objects we want to scan (total_scan). If it is greater | |
d5bc5fd3 | 829 | * than the total number of objects on slab (freeable), we must be |
0b1fb40a VD |
830 | * scanning at high prio and therefore should try to reclaim as much as |
831 | * possible. | |
832 | */ | |
833 | while (total_scan >= batch_size || | |
d5bc5fd3 | 834 | total_scan >= freeable) { |
a0b02131 | 835 | unsigned long ret; |
0b1fb40a | 836 | unsigned long nr_to_scan = min(batch_size, total_scan); |
1d3d4437 | 837 | |
0b1fb40a | 838 | shrinkctl->nr_to_scan = nr_to_scan; |
d460acb5 | 839 | shrinkctl->nr_scanned = nr_to_scan; |
a0b02131 DC |
840 | ret = shrinker->scan_objects(shrinker, shrinkctl); |
841 | if (ret == SHRINK_STOP) | |
842 | break; | |
843 | freed += ret; | |
1d3d4437 | 844 | |
d460acb5 CW |
845 | count_vm_events(SLABS_SCANNED, shrinkctl->nr_scanned); |
846 | total_scan -= shrinkctl->nr_scanned; | |
847 | scanned += shrinkctl->nr_scanned; | |
1d3d4437 GC |
848 | |
849 | cond_resched(); | |
850 | } | |
851 | ||
18bb473e YS |
852 | /* |
853 | * The deferred work is increased by any new work (delta) that wasn't | |
854 | * done, decreased by old deferred work that was done now. | |
855 | * | |
856 | * And it is capped to two times of the freeable items. | |
857 | */ | |
858 | next_deferred = max_t(long, (nr + delta - scanned), 0); | |
859 | next_deferred = min(next_deferred, (2 * freeable)); | |
860 | ||
1d3d4437 GC |
861 | /* |
862 | * move the unused scan count back into the shrinker in a | |
86750830 | 863 | * manner that handles concurrent updates. |
1d3d4437 | 864 | */ |
86750830 | 865 | new_nr = add_nr_deferred(next_deferred, shrinker, shrinkctl); |
1d3d4437 | 866 | |
8efb4b59 | 867 | trace_mm_shrink_slab_end(shrinker, shrinkctl->nid, freed, nr, new_nr, total_scan); |
1d3d4437 | 868 | return freed; |
1495f230 YH |
869 | } |
870 | ||
0a432dcb | 871 | #ifdef CONFIG_MEMCG |
b0dedc49 KT |
872 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
873 | struct mem_cgroup *memcg, int priority) | |
874 | { | |
e4262c4f | 875 | struct shrinker_info *info; |
b8e57efa KT |
876 | unsigned long ret, freed = 0; |
877 | int i; | |
b0dedc49 | 878 | |
0a432dcb | 879 | if (!mem_cgroup_online(memcg)) |
b0dedc49 KT |
880 | return 0; |
881 | ||
882 | if (!down_read_trylock(&shrinker_rwsem)) | |
883 | return 0; | |
884 | ||
468ab843 | 885 | info = shrinker_info_protected(memcg, nid); |
e4262c4f | 886 | if (unlikely(!info)) |
b0dedc49 KT |
887 | goto unlock; |
888 | ||
e4262c4f | 889 | for_each_set_bit(i, info->map, shrinker_nr_max) { |
b0dedc49 KT |
890 | struct shrink_control sc = { |
891 | .gfp_mask = gfp_mask, | |
892 | .nid = nid, | |
893 | .memcg = memcg, | |
894 | }; | |
895 | struct shrinker *shrinker; | |
896 | ||
897 | shrinker = idr_find(&shrinker_idr, i); | |
41ca668a | 898 | if (unlikely(!shrinker || !(shrinker->flags & SHRINKER_REGISTERED))) { |
7e010df5 | 899 | if (!shrinker) |
e4262c4f | 900 | clear_bit(i, info->map); |
b0dedc49 KT |
901 | continue; |
902 | } | |
903 | ||
0a432dcb YS |
904 | /* Call non-slab shrinkers even though kmem is disabled */ |
905 | if (!memcg_kmem_enabled() && | |
906 | !(shrinker->flags & SHRINKER_NONSLAB)) | |
907 | continue; | |
908 | ||
b0dedc49 | 909 | ret = do_shrink_slab(&sc, shrinker, priority); |
f90280d6 | 910 | if (ret == SHRINK_EMPTY) { |
e4262c4f | 911 | clear_bit(i, info->map); |
f90280d6 KT |
912 | /* |
913 | * After the shrinker reported that it had no objects to | |
914 | * free, but before we cleared the corresponding bit in | |
915 | * the memcg shrinker map, a new object might have been | |
916 | * added. To make sure, we have the bit set in this | |
917 | * case, we invoke the shrinker one more time and reset | |
918 | * the bit if it reports that it is not empty anymore. | |
919 | * The memory barrier here pairs with the barrier in | |
2bfd3637 | 920 | * set_shrinker_bit(): |
f90280d6 KT |
921 | * |
922 | * list_lru_add() shrink_slab_memcg() | |
923 | * list_add_tail() clear_bit() | |
924 | * <MB> <MB> | |
925 | * set_bit() do_shrink_slab() | |
926 | */ | |
927 | smp_mb__after_atomic(); | |
928 | ret = do_shrink_slab(&sc, shrinker, priority); | |
929 | if (ret == SHRINK_EMPTY) | |
930 | ret = 0; | |
931 | else | |
2bfd3637 | 932 | set_shrinker_bit(memcg, nid, i); |
f90280d6 | 933 | } |
b0dedc49 KT |
934 | freed += ret; |
935 | ||
936 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
937 | freed = freed ? : 1; | |
938 | break; | |
939 | } | |
940 | } | |
941 | unlock: | |
942 | up_read(&shrinker_rwsem); | |
943 | return freed; | |
944 | } | |
0a432dcb | 945 | #else /* CONFIG_MEMCG */ |
b0dedc49 KT |
946 | static unsigned long shrink_slab_memcg(gfp_t gfp_mask, int nid, |
947 | struct mem_cgroup *memcg, int priority) | |
948 | { | |
949 | return 0; | |
950 | } | |
0a432dcb | 951 | #endif /* CONFIG_MEMCG */ |
b0dedc49 | 952 | |
6b4f7799 | 953 | /** |
cb731d6c | 954 | * shrink_slab - shrink slab caches |
6b4f7799 JW |
955 | * @gfp_mask: allocation context |
956 | * @nid: node whose slab caches to target | |
cb731d6c | 957 | * @memcg: memory cgroup whose slab caches to target |
9092c71b | 958 | * @priority: the reclaim priority |
1da177e4 | 959 | * |
6b4f7799 | 960 | * Call the shrink functions to age shrinkable caches. |
1da177e4 | 961 | * |
6b4f7799 JW |
962 | * @nid is passed along to shrinkers with SHRINKER_NUMA_AWARE set, |
963 | * unaware shrinkers will receive a node id of 0 instead. | |
1da177e4 | 964 | * |
aeed1d32 VD |
965 | * @memcg specifies the memory cgroup to target. Unaware shrinkers |
966 | * are called only if it is the root cgroup. | |
cb731d6c | 967 | * |
9092c71b JB |
968 | * @priority is sc->priority, we take the number of objects and >> by priority |
969 | * in order to get the scan target. | |
b15e0905 | 970 | * |
6b4f7799 | 971 | * Returns the number of reclaimed slab objects. |
1da177e4 | 972 | */ |
cb731d6c VD |
973 | static unsigned long shrink_slab(gfp_t gfp_mask, int nid, |
974 | struct mem_cgroup *memcg, | |
9092c71b | 975 | int priority) |
1da177e4 | 976 | { |
b8e57efa | 977 | unsigned long ret, freed = 0; |
1da177e4 LT |
978 | struct shrinker *shrinker; |
979 | ||
fa1e512f YS |
980 | /* |
981 | * The root memcg might be allocated even though memcg is disabled | |
982 | * via "cgroup_disable=memory" boot parameter. This could make | |
983 | * mem_cgroup_is_root() return false, then just run memcg slab | |
984 | * shrink, but skip global shrink. This may result in premature | |
985 | * oom. | |
986 | */ | |
987 | if (!mem_cgroup_disabled() && !mem_cgroup_is_root(memcg)) | |
b0dedc49 | 988 | return shrink_slab_memcg(gfp_mask, nid, memcg, priority); |
cb731d6c | 989 | |
e830c63a | 990 | if (!down_read_trylock(&shrinker_rwsem)) |
f06590bd | 991 | goto out; |
1da177e4 LT |
992 | |
993 | list_for_each_entry(shrinker, &shrinker_list, list) { | |
6b4f7799 JW |
994 | struct shrink_control sc = { |
995 | .gfp_mask = gfp_mask, | |
996 | .nid = nid, | |
cb731d6c | 997 | .memcg = memcg, |
6b4f7799 | 998 | }; |
ec97097b | 999 | |
9b996468 KT |
1000 | ret = do_shrink_slab(&sc, shrinker, priority); |
1001 | if (ret == SHRINK_EMPTY) | |
1002 | ret = 0; | |
1003 | freed += ret; | |
e496612c MK |
1004 | /* |
1005 | * Bail out if someone want to register a new shrinker to | |
55b65a57 | 1006 | * prevent the registration from being stalled for long periods |
e496612c MK |
1007 | * by parallel ongoing shrinking. |
1008 | */ | |
1009 | if (rwsem_is_contended(&shrinker_rwsem)) { | |
1010 | freed = freed ? : 1; | |
1011 | break; | |
1012 | } | |
1da177e4 | 1013 | } |
6b4f7799 | 1014 | |
1da177e4 | 1015 | up_read(&shrinker_rwsem); |
f06590bd MK |
1016 | out: |
1017 | cond_resched(); | |
24f7c6b9 | 1018 | return freed; |
1da177e4 LT |
1019 | } |
1020 | ||
e4b424b7 | 1021 | static void drop_slab_node(int nid) |
cb731d6c VD |
1022 | { |
1023 | unsigned long freed; | |
1399af7e | 1024 | int shift = 0; |
cb731d6c VD |
1025 | |
1026 | do { | |
1027 | struct mem_cgroup *memcg = NULL; | |
1028 | ||
069c411d CZ |
1029 | if (fatal_signal_pending(current)) |
1030 | return; | |
1031 | ||
cb731d6c | 1032 | freed = 0; |
aeed1d32 | 1033 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
cb731d6c | 1034 | do { |
9092c71b | 1035 | freed += shrink_slab(GFP_KERNEL, nid, memcg, 0); |
cb731d6c | 1036 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL)) != NULL); |
1399af7e | 1037 | } while ((freed >> shift++) > 1); |
cb731d6c VD |
1038 | } |
1039 | ||
1040 | void drop_slab(void) | |
1041 | { | |
1042 | int nid; | |
1043 | ||
1044 | for_each_online_node(nid) | |
1045 | drop_slab_node(nid); | |
1046 | } | |
1047 | ||
e0cd5e7f | 1048 | static inline int is_page_cache_freeable(struct folio *folio) |
1da177e4 | 1049 | { |
ceddc3a5 JW |
1050 | /* |
1051 | * A freeable page cache page is referenced only by the caller | |
67891fff MW |
1052 | * that isolated the page, the page cache and optional buffer |
1053 | * heads at page->private. | |
ceddc3a5 | 1054 | */ |
e0cd5e7f MWO |
1055 | return folio_ref_count(folio) - folio_test_private(folio) == |
1056 | 1 + folio_nr_pages(folio); | |
1da177e4 LT |
1057 | } |
1058 | ||
1da177e4 | 1059 | /* |
e0cd5e7f | 1060 | * We detected a synchronous write error writing a folio out. Probably |
1da177e4 LT |
1061 | * -ENOSPC. We need to propagate that into the address_space for a subsequent |
1062 | * fsync(), msync() or close(). | |
1063 | * | |
1064 | * The tricky part is that after writepage we cannot touch the mapping: nothing | |
e0cd5e7f MWO |
1065 | * prevents it from being freed up. But we have a ref on the folio and once |
1066 | * that folio is locked, the mapping is pinned. | |
1da177e4 | 1067 | * |
e0cd5e7f | 1068 | * We're allowed to run sleeping folio_lock() here because we know the caller has |
1da177e4 LT |
1069 | * __GFP_FS. |
1070 | */ | |
1071 | static void handle_write_error(struct address_space *mapping, | |
e0cd5e7f | 1072 | struct folio *folio, int error) |
1da177e4 | 1073 | { |
e0cd5e7f MWO |
1074 | folio_lock(folio); |
1075 | if (folio_mapping(folio) == mapping) | |
3e9f45bd | 1076 | mapping_set_error(mapping, error); |
e0cd5e7f | 1077 | folio_unlock(folio); |
1da177e4 LT |
1078 | } |
1079 | ||
1b4e3f26 MG |
1080 | static bool skip_throttle_noprogress(pg_data_t *pgdat) |
1081 | { | |
1082 | int reclaimable = 0, write_pending = 0; | |
1083 | int i; | |
1084 | ||
1085 | /* | |
1086 | * If kswapd is disabled, reschedule if necessary but do not | |
1087 | * throttle as the system is likely near OOM. | |
1088 | */ | |
1089 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
1090 | return true; | |
1091 | ||
1092 | /* | |
1093 | * If there are a lot of dirty/writeback pages then do not | |
1094 | * throttle as throttling will occur when the pages cycle | |
1095 | * towards the end of the LRU if still under writeback. | |
1096 | */ | |
1097 | for (i = 0; i < MAX_NR_ZONES; i++) { | |
1098 | struct zone *zone = pgdat->node_zones + i; | |
1099 | ||
36c26128 | 1100 | if (!managed_zone(zone)) |
1b4e3f26 MG |
1101 | continue; |
1102 | ||
1103 | reclaimable += zone_reclaimable_pages(zone); | |
1104 | write_pending += zone_page_state_snapshot(zone, | |
1105 | NR_ZONE_WRITE_PENDING); | |
1106 | } | |
1107 | if (2 * write_pending <= reclaimable) | |
1108 | return true; | |
1109 | ||
1110 | return false; | |
1111 | } | |
1112 | ||
c3f4a9a2 | 1113 | void reclaim_throttle(pg_data_t *pgdat, enum vmscan_throttle_state reason) |
8cd7c588 MG |
1114 | { |
1115 | wait_queue_head_t *wqh = &pgdat->reclaim_wait[reason]; | |
c3f4a9a2 | 1116 | long timeout, ret; |
8cd7c588 MG |
1117 | DEFINE_WAIT(wait); |
1118 | ||
1119 | /* | |
1120 | * Do not throttle IO workers, kthreads other than kswapd or | |
1121 | * workqueues. They may be required for reclaim to make | |
1122 | * forward progress (e.g. journalling workqueues or kthreads). | |
1123 | */ | |
1124 | if (!current_is_kswapd() && | |
b485c6f1 MG |
1125 | current->flags & (PF_IO_WORKER|PF_KTHREAD)) { |
1126 | cond_resched(); | |
8cd7c588 | 1127 | return; |
b485c6f1 | 1128 | } |
8cd7c588 | 1129 | |
c3f4a9a2 MG |
1130 | /* |
1131 | * These figures are pulled out of thin air. | |
1132 | * VMSCAN_THROTTLE_ISOLATED is a transient condition based on too many | |
1133 | * parallel reclaimers which is a short-lived event so the timeout is | |
1134 | * short. Failing to make progress or waiting on writeback are | |
1135 | * potentially long-lived events so use a longer timeout. This is shaky | |
1136 | * logic as a failure to make progress could be due to anything from | |
1137 | * writeback to a slow device to excessive references pages at the tail | |
1138 | * of the inactive LRU. | |
1139 | */ | |
1140 | switch(reason) { | |
1141 | case VMSCAN_THROTTLE_WRITEBACK: | |
1142 | timeout = HZ/10; | |
1143 | ||
1144 | if (atomic_inc_return(&pgdat->nr_writeback_throttled) == 1) { | |
1145 | WRITE_ONCE(pgdat->nr_reclaim_start, | |
1146 | node_page_state(pgdat, NR_THROTTLED_WRITTEN)); | |
1147 | } | |
1148 | ||
1149 | break; | |
1b4e3f26 MG |
1150 | case VMSCAN_THROTTLE_CONGESTED: |
1151 | fallthrough; | |
c3f4a9a2 | 1152 | case VMSCAN_THROTTLE_NOPROGRESS: |
1b4e3f26 MG |
1153 | if (skip_throttle_noprogress(pgdat)) { |
1154 | cond_resched(); | |
1155 | return; | |
1156 | } | |
1157 | ||
1158 | timeout = 1; | |
1159 | ||
c3f4a9a2 MG |
1160 | break; |
1161 | case VMSCAN_THROTTLE_ISOLATED: | |
1162 | timeout = HZ/50; | |
1163 | break; | |
1164 | default: | |
1165 | WARN_ON_ONCE(1); | |
1166 | timeout = HZ; | |
1167 | break; | |
8cd7c588 MG |
1168 | } |
1169 | ||
1170 | prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); | |
1171 | ret = schedule_timeout(timeout); | |
1172 | finish_wait(wqh, &wait); | |
d818fca1 | 1173 | |
c3f4a9a2 | 1174 | if (reason == VMSCAN_THROTTLE_WRITEBACK) |
d818fca1 | 1175 | atomic_dec(&pgdat->nr_writeback_throttled); |
8cd7c588 MG |
1176 | |
1177 | trace_mm_vmscan_throttled(pgdat->node_id, jiffies_to_usecs(timeout), | |
1178 | jiffies_to_usecs(timeout - ret), | |
1179 | reason); | |
1180 | } | |
1181 | ||
1182 | /* | |
1183 | * Account for pages written if tasks are throttled waiting on dirty | |
1184 | * pages to clean. If enough pages have been cleaned since throttling | |
1185 | * started then wakeup the throttled tasks. | |
1186 | */ | |
512b7931 | 1187 | void __acct_reclaim_writeback(pg_data_t *pgdat, struct folio *folio, |
8cd7c588 MG |
1188 | int nr_throttled) |
1189 | { | |
1190 | unsigned long nr_written; | |
1191 | ||
512b7931 | 1192 | node_stat_add_folio(folio, NR_THROTTLED_WRITTEN); |
8cd7c588 MG |
1193 | |
1194 | /* | |
1195 | * This is an inaccurate read as the per-cpu deltas may not | |
1196 | * be synchronised. However, given that the system is | |
1197 | * writeback throttled, it is not worth taking the penalty | |
1198 | * of getting an accurate count. At worst, the throttle | |
1199 | * timeout guarantees forward progress. | |
1200 | */ | |
1201 | nr_written = node_page_state(pgdat, NR_THROTTLED_WRITTEN) - | |
1202 | READ_ONCE(pgdat->nr_reclaim_start); | |
1203 | ||
1204 | if (nr_written > SWAP_CLUSTER_MAX * nr_throttled) | |
1205 | wake_up(&pgdat->reclaim_wait[VMSCAN_THROTTLE_WRITEBACK]); | |
1206 | } | |
1207 | ||
04e62a29 CL |
1208 | /* possible outcome of pageout() */ |
1209 | typedef enum { | |
1210 | /* failed to write page out, page is locked */ | |
1211 | PAGE_KEEP, | |
1212 | /* move page to the active list, page is locked */ | |
1213 | PAGE_ACTIVATE, | |
1214 | /* page has been sent to the disk successfully, page is unlocked */ | |
1215 | PAGE_SUCCESS, | |
1216 | /* page is clean and locked */ | |
1217 | PAGE_CLEAN, | |
1218 | } pageout_t; | |
1219 | ||
1da177e4 | 1220 | /* |
1742f19f AM |
1221 | * pageout is called by shrink_page_list() for each dirty page. |
1222 | * Calls ->writepage(). | |
1da177e4 | 1223 | */ |
2282679f N |
1224 | static pageout_t pageout(struct folio *folio, struct address_space *mapping, |
1225 | struct swap_iocb **plug) | |
1da177e4 LT |
1226 | { |
1227 | /* | |
e0cd5e7f | 1228 | * If the folio is dirty, only perform writeback if that write |
1da177e4 LT |
1229 | * will be non-blocking. To prevent this allocation from being |
1230 | * stalled by pagecache activity. But note that there may be | |
1231 | * stalls if we need to run get_block(). We could test | |
1232 | * PagePrivate for that. | |
1233 | * | |
8174202b | 1234 | * If this process is currently in __generic_file_write_iter() against |
e0cd5e7f | 1235 | * this folio's queue, we can perform writeback even if that |
1da177e4 LT |
1236 | * will block. |
1237 | * | |
e0cd5e7f | 1238 | * If the folio is swapcache, write it back even if that would |
1da177e4 LT |
1239 | * block, for some throttling. This happens by accident, because |
1240 | * swap_backing_dev_info is bust: it doesn't reflect the | |
1241 | * congestion state of the swapdevs. Easy to fix, if needed. | |
1da177e4 | 1242 | */ |
e0cd5e7f | 1243 | if (!is_page_cache_freeable(folio)) |
1da177e4 LT |
1244 | return PAGE_KEEP; |
1245 | if (!mapping) { | |
1246 | /* | |
e0cd5e7f MWO |
1247 | * Some data journaling orphaned folios can have |
1248 | * folio->mapping == NULL while being dirty with clean buffers. | |
1da177e4 | 1249 | */ |
e0cd5e7f | 1250 | if (folio_test_private(folio)) { |
68189fef | 1251 | if (try_to_free_buffers(folio)) { |
e0cd5e7f MWO |
1252 | folio_clear_dirty(folio); |
1253 | pr_info("%s: orphaned folio\n", __func__); | |
1da177e4 LT |
1254 | return PAGE_CLEAN; |
1255 | } | |
1256 | } | |
1257 | return PAGE_KEEP; | |
1258 | } | |
1259 | if (mapping->a_ops->writepage == NULL) | |
1260 | return PAGE_ACTIVATE; | |
1da177e4 | 1261 | |
e0cd5e7f | 1262 | if (folio_clear_dirty_for_io(folio)) { |
1da177e4 LT |
1263 | int res; |
1264 | struct writeback_control wbc = { | |
1265 | .sync_mode = WB_SYNC_NONE, | |
1266 | .nr_to_write = SWAP_CLUSTER_MAX, | |
111ebb6e OH |
1267 | .range_start = 0, |
1268 | .range_end = LLONG_MAX, | |
1da177e4 | 1269 | .for_reclaim = 1, |
2282679f | 1270 | .swap_plug = plug, |
1da177e4 LT |
1271 | }; |
1272 | ||
e0cd5e7f MWO |
1273 | folio_set_reclaim(folio); |
1274 | res = mapping->a_ops->writepage(&folio->page, &wbc); | |
1da177e4 | 1275 | if (res < 0) |
e0cd5e7f | 1276 | handle_write_error(mapping, folio, res); |
994fc28c | 1277 | if (res == AOP_WRITEPAGE_ACTIVATE) { |
e0cd5e7f | 1278 | folio_clear_reclaim(folio); |
1da177e4 LT |
1279 | return PAGE_ACTIVATE; |
1280 | } | |
c661b078 | 1281 | |
e0cd5e7f | 1282 | if (!folio_test_writeback(folio)) { |
1da177e4 | 1283 | /* synchronous write or broken a_ops? */ |
e0cd5e7f | 1284 | folio_clear_reclaim(folio); |
1da177e4 | 1285 | } |
e0cd5e7f MWO |
1286 | trace_mm_vmscan_write_folio(folio); |
1287 | node_stat_add_folio(folio, NR_VMSCAN_WRITE); | |
1da177e4 LT |
1288 | return PAGE_SUCCESS; |
1289 | } | |
1290 | ||
1291 | return PAGE_CLEAN; | |
1292 | } | |
1293 | ||
a649fd92 | 1294 | /* |
e286781d NP |
1295 | * Same as remove_mapping, but if the page is removed from the mapping, it |
1296 | * gets returned with a refcount of 0. | |
a649fd92 | 1297 | */ |
be7c07d6 | 1298 | static int __remove_mapping(struct address_space *mapping, struct folio *folio, |
b910718a | 1299 | bool reclaimed, struct mem_cgroup *target_memcg) |
49d2e9cc | 1300 | { |
bd4c82c2 | 1301 | int refcount; |
aae466b0 | 1302 | void *shadow = NULL; |
c4843a75 | 1303 | |
be7c07d6 MWO |
1304 | BUG_ON(!folio_test_locked(folio)); |
1305 | BUG_ON(mapping != folio_mapping(folio)); | |
49d2e9cc | 1306 | |
be7c07d6 | 1307 | if (!folio_test_swapcache(folio)) |
51b8c1fe | 1308 | spin_lock(&mapping->host->i_lock); |
30472509 | 1309 | xa_lock_irq(&mapping->i_pages); |
49d2e9cc | 1310 | /* |
0fd0e6b0 NP |
1311 | * The non racy check for a busy page. |
1312 | * | |
1313 | * Must be careful with the order of the tests. When someone has | |
1314 | * a ref to the page, it may be possible that they dirty it then | |
1315 | * drop the reference. So if PageDirty is tested before page_count | |
1316 | * here, then the following race may occur: | |
1317 | * | |
1318 | * get_user_pages(&page); | |
1319 | * [user mapping goes away] | |
1320 | * write_to(page); | |
1321 | * !PageDirty(page) [good] | |
1322 | * SetPageDirty(page); | |
1323 | * put_page(page); | |
1324 | * !page_count(page) [good, discard it] | |
1325 | * | |
1326 | * [oops, our write_to data is lost] | |
1327 | * | |
1328 | * Reversing the order of the tests ensures such a situation cannot | |
1329 | * escape unnoticed. The smp_rmb is needed to ensure the page->flags | |
0139aa7b | 1330 | * load is not satisfied before that of page->_refcount. |
0fd0e6b0 NP |
1331 | * |
1332 | * Note that if SetPageDirty is always performed via set_page_dirty, | |
b93b0163 | 1333 | * and thus under the i_pages lock, then this ordering is not required. |
49d2e9cc | 1334 | */ |
be7c07d6 MWO |
1335 | refcount = 1 + folio_nr_pages(folio); |
1336 | if (!folio_ref_freeze(folio, refcount)) | |
49d2e9cc | 1337 | goto cannot_free; |
1c4c3b99 | 1338 | /* note: atomic_cmpxchg in page_ref_freeze provides the smp_rmb */ |
be7c07d6 MWO |
1339 | if (unlikely(folio_test_dirty(folio))) { |
1340 | folio_ref_unfreeze(folio, refcount); | |
49d2e9cc | 1341 | goto cannot_free; |
e286781d | 1342 | } |
49d2e9cc | 1343 | |
be7c07d6 MWO |
1344 | if (folio_test_swapcache(folio)) { |
1345 | swp_entry_t swap = folio_swap_entry(folio); | |
ac35a490 YZ |
1346 | |
1347 | /* get a shadow entry before mem_cgroup_swapout() clears folio_memcg() */ | |
aae466b0 | 1348 | if (reclaimed && !mapping_exiting(mapping)) |
8927f647 | 1349 | shadow = workingset_eviction(folio, target_memcg); |
ac35a490 | 1350 | mem_cgroup_swapout(folio, swap); |
ceff9d33 | 1351 | __delete_from_swap_cache(folio, swap, shadow); |
30472509 | 1352 | xa_unlock_irq(&mapping->i_pages); |
be7c07d6 | 1353 | put_swap_page(&folio->page, swap); |
e286781d | 1354 | } else { |
d2329aa0 | 1355 | void (*free_folio)(struct folio *); |
6072d13c | 1356 | |
d2329aa0 | 1357 | free_folio = mapping->a_ops->free_folio; |
a528910e JW |
1358 | /* |
1359 | * Remember a shadow entry for reclaimed file cache in | |
1360 | * order to detect refaults, thus thrashing, later on. | |
1361 | * | |
1362 | * But don't store shadows in an address space that is | |
238c3046 | 1363 | * already exiting. This is not just an optimization, |
a528910e JW |
1364 | * inode reclaim needs to empty out the radix tree or |
1365 | * the nodes are lost. Don't plant shadows behind its | |
1366 | * back. | |
f9fe48be RZ |
1367 | * |
1368 | * We also don't store shadows for DAX mappings because the | |
1369 | * only page cache pages found in these are zero pages | |
1370 | * covering holes, and because we don't want to mix DAX | |
1371 | * exceptional entries and shadow exceptional entries in the | |
b93b0163 | 1372 | * same address_space. |
a528910e | 1373 | */ |
be7c07d6 | 1374 | if (reclaimed && folio_is_file_lru(folio) && |
f9fe48be | 1375 | !mapping_exiting(mapping) && !dax_mapping(mapping)) |
8927f647 MWO |
1376 | shadow = workingset_eviction(folio, target_memcg); |
1377 | __filemap_remove_folio(folio, shadow); | |
30472509 | 1378 | xa_unlock_irq(&mapping->i_pages); |
51b8c1fe JW |
1379 | if (mapping_shrinkable(mapping)) |
1380 | inode_add_lru(mapping->host); | |
1381 | spin_unlock(&mapping->host->i_lock); | |
6072d13c | 1382 | |
d2329aa0 MWO |
1383 | if (free_folio) |
1384 | free_folio(folio); | |
49d2e9cc CL |
1385 | } |
1386 | ||
49d2e9cc CL |
1387 | return 1; |
1388 | ||
1389 | cannot_free: | |
30472509 | 1390 | xa_unlock_irq(&mapping->i_pages); |
be7c07d6 | 1391 | if (!folio_test_swapcache(folio)) |
51b8c1fe | 1392 | spin_unlock(&mapping->host->i_lock); |
49d2e9cc CL |
1393 | return 0; |
1394 | } | |
1395 | ||
5100da38 MWO |
1396 | /** |
1397 | * remove_mapping() - Attempt to remove a folio from its mapping. | |
1398 | * @mapping: The address space. | |
1399 | * @folio: The folio to remove. | |
1400 | * | |
1401 | * If the folio is dirty, under writeback or if someone else has a ref | |
1402 | * on it, removal will fail. | |
1403 | * Return: The number of pages removed from the mapping. 0 if the folio | |
1404 | * could not be removed. | |
1405 | * Context: The caller should have a single refcount on the folio and | |
1406 | * hold its lock. | |
e286781d | 1407 | */ |
5100da38 | 1408 | long remove_mapping(struct address_space *mapping, struct folio *folio) |
e286781d | 1409 | { |
be7c07d6 | 1410 | if (__remove_mapping(mapping, folio, false, NULL)) { |
e286781d | 1411 | /* |
5100da38 | 1412 | * Unfreezing the refcount with 1 effectively |
e286781d NP |
1413 | * drops the pagecache ref for us without requiring another |
1414 | * atomic operation. | |
1415 | */ | |
be7c07d6 | 1416 | folio_ref_unfreeze(folio, 1); |
5100da38 | 1417 | return folio_nr_pages(folio); |
e286781d NP |
1418 | } |
1419 | return 0; | |
1420 | } | |
1421 | ||
894bc310 | 1422 | /** |
ca6d60f3 MWO |
1423 | * folio_putback_lru - Put previously isolated folio onto appropriate LRU list. |
1424 | * @folio: Folio to be returned to an LRU list. | |
894bc310 | 1425 | * |
ca6d60f3 MWO |
1426 | * Add previously isolated @folio to appropriate LRU list. |
1427 | * The folio may still be unevictable for other reasons. | |
894bc310 | 1428 | * |
ca6d60f3 | 1429 | * Context: lru_lock must not be held, interrupts must be enabled. |
894bc310 | 1430 | */ |
ca6d60f3 | 1431 | void folio_putback_lru(struct folio *folio) |
894bc310 | 1432 | { |
ca6d60f3 MWO |
1433 | folio_add_lru(folio); |
1434 | folio_put(folio); /* drop ref from isolate */ | |
894bc310 LS |
1435 | } |
1436 | ||
dfc8d636 JW |
1437 | enum page_references { |
1438 | PAGEREF_RECLAIM, | |
1439 | PAGEREF_RECLAIM_CLEAN, | |
64574746 | 1440 | PAGEREF_KEEP, |
dfc8d636 JW |
1441 | PAGEREF_ACTIVATE, |
1442 | }; | |
1443 | ||
d92013d1 | 1444 | static enum page_references folio_check_references(struct folio *folio, |
dfc8d636 JW |
1445 | struct scan_control *sc) |
1446 | { | |
d92013d1 | 1447 | int referenced_ptes, referenced_folio; |
dfc8d636 | 1448 | unsigned long vm_flags; |
dfc8d636 | 1449 | |
b3ac0413 MWO |
1450 | referenced_ptes = folio_referenced(folio, 1, sc->target_mem_cgroup, |
1451 | &vm_flags); | |
d92013d1 | 1452 | referenced_folio = folio_test_clear_referenced(folio); |
dfc8d636 | 1453 | |
dfc8d636 | 1454 | /* |
d92013d1 MWO |
1455 | * The supposedly reclaimable folio was found to be in a VM_LOCKED vma. |
1456 | * Let the folio, now marked Mlocked, be moved to the unevictable list. | |
dfc8d636 JW |
1457 | */ |
1458 | if (vm_flags & VM_LOCKED) | |
47d4f3ee | 1459 | return PAGEREF_ACTIVATE; |
dfc8d636 | 1460 | |
6d4675e6 MK |
1461 | /* rmap lock contention: rotate */ |
1462 | if (referenced_ptes == -1) | |
1463 | return PAGEREF_KEEP; | |
1464 | ||
64574746 | 1465 | if (referenced_ptes) { |
64574746 | 1466 | /* |
d92013d1 | 1467 | * All mapped folios start out with page table |
64574746 | 1468 | * references from the instantiating fault, so we need |
9030fb0b | 1469 | * to look twice if a mapped file/anon folio is used more |
64574746 JW |
1470 | * than once. |
1471 | * | |
1472 | * Mark it and spare it for another trip around the | |
1473 | * inactive list. Another page table reference will | |
1474 | * lead to its activation. | |
1475 | * | |
d92013d1 MWO |
1476 | * Note: the mark is set for activated folios as well |
1477 | * so that recently deactivated but used folios are | |
64574746 JW |
1478 | * quickly recovered. |
1479 | */ | |
d92013d1 | 1480 | folio_set_referenced(folio); |
64574746 | 1481 | |
d92013d1 | 1482 | if (referenced_folio || referenced_ptes > 1) |
64574746 JW |
1483 | return PAGEREF_ACTIVATE; |
1484 | ||
c909e993 | 1485 | /* |
d92013d1 | 1486 | * Activate file-backed executable folios after first usage. |
c909e993 | 1487 | */ |
f19a27e3 | 1488 | if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) |
c909e993 KK |
1489 | return PAGEREF_ACTIVATE; |
1490 | ||
64574746 JW |
1491 | return PAGEREF_KEEP; |
1492 | } | |
dfc8d636 | 1493 | |
d92013d1 | 1494 | /* Reclaim if clean, defer dirty folios to writeback */ |
f19a27e3 | 1495 | if (referenced_folio && folio_is_file_lru(folio)) |
64574746 JW |
1496 | return PAGEREF_RECLAIM_CLEAN; |
1497 | ||
1498 | return PAGEREF_RECLAIM; | |
dfc8d636 JW |
1499 | } |
1500 | ||
e2be15f6 | 1501 | /* Check if a page is dirty or under writeback */ |
e20c41b1 | 1502 | static void folio_check_dirty_writeback(struct folio *folio, |
e2be15f6 MG |
1503 | bool *dirty, bool *writeback) |
1504 | { | |
b4597226 MG |
1505 | struct address_space *mapping; |
1506 | ||
e2be15f6 MG |
1507 | /* |
1508 | * Anonymous pages are not handled by flushers and must be written | |
32a331a7 ML |
1509 | * from reclaim context. Do not stall reclaim based on them. |
1510 | * MADV_FREE anonymous pages are put into inactive file list too. | |
1511 | * They could be mistakenly treated as file lru. So further anon | |
1512 | * test is needed. | |
e2be15f6 | 1513 | */ |
e20c41b1 MWO |
1514 | if (!folio_is_file_lru(folio) || |
1515 | (folio_test_anon(folio) && !folio_test_swapbacked(folio))) { | |
e2be15f6 MG |
1516 | *dirty = false; |
1517 | *writeback = false; | |
1518 | return; | |
1519 | } | |
1520 | ||
e20c41b1 MWO |
1521 | /* By default assume that the folio flags are accurate */ |
1522 | *dirty = folio_test_dirty(folio); | |
1523 | *writeback = folio_test_writeback(folio); | |
b4597226 MG |
1524 | |
1525 | /* Verify dirty/writeback state if the filesystem supports it */ | |
e20c41b1 | 1526 | if (!folio_test_private(folio)) |
b4597226 MG |
1527 | return; |
1528 | ||
e20c41b1 | 1529 | mapping = folio_mapping(folio); |
b4597226 | 1530 | if (mapping && mapping->a_ops->is_dirty_writeback) |
520f301c | 1531 | mapping->a_ops->is_dirty_writeback(folio, dirty, writeback); |
e2be15f6 MG |
1532 | } |
1533 | ||
26aa2d19 DH |
1534 | static struct page *alloc_demote_page(struct page *page, unsigned long node) |
1535 | { | |
1536 | struct migration_target_control mtc = { | |
1537 | /* | |
1538 | * Allocate from 'node', or fail quickly and quietly. | |
1539 | * When this happens, 'page' will likely just be discarded | |
1540 | * instead of migrated. | |
1541 | */ | |
1542 | .gfp_mask = (GFP_HIGHUSER_MOVABLE & ~__GFP_RECLAIM) | | |
1543 | __GFP_THISNODE | __GFP_NOWARN | | |
1544 | __GFP_NOMEMALLOC | GFP_NOWAIT, | |
1545 | .nid = node | |
1546 | }; | |
1547 | ||
1548 | return alloc_migration_target(page, (unsigned long)&mtc); | |
1549 | } | |
1550 | ||
1551 | /* | |
1552 | * Take pages on @demote_list and attempt to demote them to | |
1553 | * another node. Pages which are not demoted are left on | |
1554 | * @demote_pages. | |
1555 | */ | |
1556 | static unsigned int demote_page_list(struct list_head *demote_pages, | |
1557 | struct pglist_data *pgdat) | |
1558 | { | |
1559 | int target_nid = next_demotion_node(pgdat->node_id); | |
1560 | unsigned int nr_succeeded; | |
26aa2d19 DH |
1561 | |
1562 | if (list_empty(demote_pages)) | |
1563 | return 0; | |
1564 | ||
1565 | if (target_nid == NUMA_NO_NODE) | |
1566 | return 0; | |
1567 | ||
1568 | /* Demotion ignores all cpuset and mempolicy settings */ | |
cb75463c | 1569 | migrate_pages(demote_pages, alloc_demote_page, NULL, |
26aa2d19 DH |
1570 | target_nid, MIGRATE_ASYNC, MR_DEMOTION, |
1571 | &nr_succeeded); | |
1572 | ||
668e4147 YS |
1573 | if (current_is_kswapd()) |
1574 | __count_vm_events(PGDEMOTE_KSWAPD, nr_succeeded); | |
1575 | else | |
1576 | __count_vm_events(PGDEMOTE_DIRECT, nr_succeeded); | |
1577 | ||
26aa2d19 DH |
1578 | return nr_succeeded; |
1579 | } | |
1580 | ||
c28a0e96 | 1581 | static bool may_enter_fs(struct folio *folio, gfp_t gfp_mask) |
d791ea67 N |
1582 | { |
1583 | if (gfp_mask & __GFP_FS) | |
1584 | return true; | |
c28a0e96 | 1585 | if (!folio_test_swapcache(folio) || !(gfp_mask & __GFP_IO)) |
d791ea67 N |
1586 | return false; |
1587 | /* | |
1588 | * We can "enter_fs" for swap-cache with only __GFP_IO | |
1589 | * providing this isn't SWP_FS_OPS. | |
1590 | * ->flags can be updated non-atomicially (scan_swap_map_slots), | |
1591 | * but that will never affect SWP_FS_OPS, so the data_race | |
1592 | * is safe. | |
1593 | */ | |
b98c359f | 1594 | return !data_race(folio_swap_flags(folio) & SWP_FS_OPS); |
d791ea67 N |
1595 | } |
1596 | ||
1da177e4 | 1597 | /* |
1742f19f | 1598 | * shrink_page_list() returns the number of reclaimed pages |
1da177e4 | 1599 | */ |
730ec8c0 MS |
1600 | static unsigned int shrink_page_list(struct list_head *page_list, |
1601 | struct pglist_data *pgdat, | |
1602 | struct scan_control *sc, | |
730ec8c0 MS |
1603 | struct reclaim_stat *stat, |
1604 | bool ignore_references) | |
1da177e4 LT |
1605 | { |
1606 | LIST_HEAD(ret_pages); | |
abe4c3b5 | 1607 | LIST_HEAD(free_pages); |
26aa2d19 | 1608 | LIST_HEAD(demote_pages); |
730ec8c0 MS |
1609 | unsigned int nr_reclaimed = 0; |
1610 | unsigned int pgactivate = 0; | |
26aa2d19 | 1611 | bool do_demote_pass; |
2282679f | 1612 | struct swap_iocb *plug = NULL; |
1da177e4 | 1613 | |
060f005f | 1614 | memset(stat, 0, sizeof(*stat)); |
1da177e4 | 1615 | cond_resched(); |
26aa2d19 | 1616 | do_demote_pass = can_demote(pgdat->node_id, sc); |
1da177e4 | 1617 | |
26aa2d19 | 1618 | retry: |
1da177e4 LT |
1619 | while (!list_empty(page_list)) { |
1620 | struct address_space *mapping; | |
be7c07d6 | 1621 | struct folio *folio; |
8940b34a | 1622 | enum page_references references = PAGEREF_RECLAIM; |
d791ea67 | 1623 | bool dirty, writeback; |
98879b3b | 1624 | unsigned int nr_pages; |
1da177e4 LT |
1625 | |
1626 | cond_resched(); | |
1627 | ||
be7c07d6 MWO |
1628 | folio = lru_to_folio(page_list); |
1629 | list_del(&folio->lru); | |
1da177e4 | 1630 | |
c28a0e96 | 1631 | if (!folio_trylock(folio)) |
1da177e4 LT |
1632 | goto keep; |
1633 | ||
c28a0e96 | 1634 | VM_BUG_ON_FOLIO(folio_test_active(folio), folio); |
1da177e4 | 1635 | |
c28a0e96 | 1636 | nr_pages = folio_nr_pages(folio); |
98879b3b | 1637 | |
c28a0e96 | 1638 | /* Account the number of base pages */ |
98879b3b | 1639 | sc->nr_scanned += nr_pages; |
80e43426 | 1640 | |
c28a0e96 | 1641 | if (unlikely(!folio_evictable(folio))) |
ad6b6704 | 1642 | goto activate_locked; |
894bc310 | 1643 | |
1bee2c16 | 1644 | if (!sc->may_unmap && folio_mapped(folio)) |
80e43426 CL |
1645 | goto keep_locked; |
1646 | ||
018ee47f YZ |
1647 | /* folio_update_gen() tried to promote this page? */ |
1648 | if (lru_gen_enabled() && !ignore_references && | |
1649 | folio_mapped(folio) && folio_test_referenced(folio)) | |
1650 | goto keep_locked; | |
1651 | ||
e2be15f6 | 1652 | /* |
894befec | 1653 | * The number of dirty pages determines if a node is marked |
8cd7c588 | 1654 | * reclaim_congested. kswapd will stall and start writing |
c28a0e96 | 1655 | * folios if the tail of the LRU is all dirty unqueued folios. |
e2be15f6 | 1656 | */ |
e20c41b1 | 1657 | folio_check_dirty_writeback(folio, &dirty, &writeback); |
e2be15f6 | 1658 | if (dirty || writeback) |
c79b7b96 | 1659 | stat->nr_dirty += nr_pages; |
e2be15f6 MG |
1660 | |
1661 | if (dirty && !writeback) | |
c79b7b96 | 1662 | stat->nr_unqueued_dirty += nr_pages; |
e2be15f6 | 1663 | |
d04e8acd | 1664 | /* |
c28a0e96 MWO |
1665 | * Treat this folio as congested if folios are cycling |
1666 | * through the LRU so quickly that the folios marked | |
1667 | * for immediate reclaim are making it to the end of | |
1668 | * the LRU a second time. | |
d04e8acd | 1669 | */ |
c28a0e96 | 1670 | if (writeback && folio_test_reclaim(folio)) |
c79b7b96 | 1671 | stat->nr_congested += nr_pages; |
e2be15f6 | 1672 | |
283aba9f | 1673 | /* |
d33e4e14 | 1674 | * If a folio at the tail of the LRU is under writeback, there |
283aba9f MG |
1675 | * are three cases to consider. |
1676 | * | |
c28a0e96 MWO |
1677 | * 1) If reclaim is encountering an excessive number |
1678 | * of folios under writeback and this folio has both | |
1679 | * the writeback and reclaim flags set, then it | |
d33e4e14 MWO |
1680 | * indicates that folios are being queued for I/O but |
1681 | * are being recycled through the LRU before the I/O | |
1682 | * can complete. Waiting on the folio itself risks an | |
1683 | * indefinite stall if it is impossible to writeback | |
1684 | * the folio due to I/O error or disconnected storage | |
1685 | * so instead note that the LRU is being scanned too | |
1686 | * quickly and the caller can stall after the folio | |
1687 | * list has been processed. | |
283aba9f | 1688 | * |
d33e4e14 | 1689 | * 2) Global or new memcg reclaim encounters a folio that is |
ecf5fc6e MH |
1690 | * not marked for immediate reclaim, or the caller does not |
1691 | * have __GFP_FS (or __GFP_IO if it's simply going to swap, | |
d33e4e14 | 1692 | * not to fs). In this case mark the folio for immediate |
97c9341f | 1693 | * reclaim and continue scanning. |
283aba9f | 1694 | * |
d791ea67 | 1695 | * Require may_enter_fs() because we would wait on fs, which |
d33e4e14 MWO |
1696 | * may not have submitted I/O yet. And the loop driver might |
1697 | * enter reclaim, and deadlock if it waits on a folio for | |
283aba9f MG |
1698 | * which it is needed to do the write (loop masks off |
1699 | * __GFP_IO|__GFP_FS for this reason); but more thought | |
1700 | * would probably show more reasons. | |
1701 | * | |
d33e4e14 MWO |
1702 | * 3) Legacy memcg encounters a folio that already has the |
1703 | * reclaim flag set. memcg does not have any dirty folio | |
283aba9f | 1704 | * throttling so we could easily OOM just because too many |
d33e4e14 | 1705 | * folios are in writeback and there is nothing else to |
283aba9f | 1706 | * reclaim. Wait for the writeback to complete. |
c55e8d03 | 1707 | * |
d33e4e14 MWO |
1708 | * In cases 1) and 2) we activate the folios to get them out of |
1709 | * the way while we continue scanning for clean folios on the | |
c55e8d03 JW |
1710 | * inactive list and refilling from the active list. The |
1711 | * observation here is that waiting for disk writes is more | |
1712 | * expensive than potentially causing reloads down the line. | |
1713 | * Since they're marked for immediate reclaim, they won't put | |
1714 | * memory pressure on the cache working set any longer than it | |
1715 | * takes to write them to disk. | |
283aba9f | 1716 | */ |
d33e4e14 | 1717 | if (folio_test_writeback(folio)) { |
283aba9f MG |
1718 | /* Case 1 above */ |
1719 | if (current_is_kswapd() && | |
d33e4e14 | 1720 | folio_test_reclaim(folio) && |
599d0c95 | 1721 | test_bit(PGDAT_WRITEBACK, &pgdat->flags)) { |
c79b7b96 | 1722 | stat->nr_immediate += nr_pages; |
c55e8d03 | 1723 | goto activate_locked; |
283aba9f MG |
1724 | |
1725 | /* Case 2 above */ | |
b5ead35e | 1726 | } else if (writeback_throttling_sane(sc) || |
d33e4e14 | 1727 | !folio_test_reclaim(folio) || |
c28a0e96 | 1728 | !may_enter_fs(folio, sc->gfp_mask)) { |
c3b94f44 | 1729 | /* |
d33e4e14 | 1730 | * This is slightly racy - |
c28a0e96 MWO |
1731 | * folio_end_writeback() might have |
1732 | * just cleared the reclaim flag, then | |
1733 | * setting the reclaim flag here ends up | |
1734 | * interpreted as the readahead flag - but | |
1735 | * that does not matter enough to care. | |
1736 | * What we do want is for this folio to | |
1737 | * have the reclaim flag set next time | |
1738 | * memcg reclaim reaches the tests above, | |
1739 | * so it will then wait for writeback to | |
1740 | * avoid OOM; and it's also appropriate | |
d33e4e14 | 1741 | * in global reclaim. |
c3b94f44 | 1742 | */ |
d33e4e14 | 1743 | folio_set_reclaim(folio); |
c79b7b96 | 1744 | stat->nr_writeback += nr_pages; |
c55e8d03 | 1745 | goto activate_locked; |
283aba9f MG |
1746 | |
1747 | /* Case 3 above */ | |
1748 | } else { | |
d33e4e14 MWO |
1749 | folio_unlock(folio); |
1750 | folio_wait_writeback(folio); | |
1751 | /* then go back and try same folio again */ | |
1752 | list_add_tail(&folio->lru, page_list); | |
7fadc820 | 1753 | continue; |
e62e384e | 1754 | } |
c661b078 | 1755 | } |
1da177e4 | 1756 | |
8940b34a | 1757 | if (!ignore_references) |
d92013d1 | 1758 | references = folio_check_references(folio, sc); |
02c6de8d | 1759 | |
dfc8d636 JW |
1760 | switch (references) { |
1761 | case PAGEREF_ACTIVATE: | |
1da177e4 | 1762 | goto activate_locked; |
64574746 | 1763 | case PAGEREF_KEEP: |
98879b3b | 1764 | stat->nr_ref_keep += nr_pages; |
64574746 | 1765 | goto keep_locked; |
dfc8d636 JW |
1766 | case PAGEREF_RECLAIM: |
1767 | case PAGEREF_RECLAIM_CLEAN: | |
c28a0e96 | 1768 | ; /* try to reclaim the folio below */ |
dfc8d636 | 1769 | } |
1da177e4 | 1770 | |
26aa2d19 | 1771 | /* |
c28a0e96 | 1772 | * Before reclaiming the folio, try to relocate |
26aa2d19 DH |
1773 | * its contents to another node. |
1774 | */ | |
1775 | if (do_demote_pass && | |
c28a0e96 MWO |
1776 | (thp_migration_supported() || !folio_test_large(folio))) { |
1777 | list_add(&folio->lru, &demote_pages); | |
1778 | folio_unlock(folio); | |
26aa2d19 DH |
1779 | continue; |
1780 | } | |
1781 | ||
1da177e4 LT |
1782 | /* |
1783 | * Anonymous process memory has backing store? | |
1784 | * Try to allocate it some swap space here. | |
c28a0e96 | 1785 | * Lazyfree folio could be freed directly |
1da177e4 | 1786 | */ |
c28a0e96 MWO |
1787 | if (folio_test_anon(folio) && folio_test_swapbacked(folio)) { |
1788 | if (!folio_test_swapcache(folio)) { | |
bd4c82c2 HY |
1789 | if (!(sc->gfp_mask & __GFP_IO)) |
1790 | goto keep_locked; | |
d4b4084a | 1791 | if (folio_maybe_dma_pinned(folio)) |
feb889fb | 1792 | goto keep_locked; |
c28a0e96 MWO |
1793 | if (folio_test_large(folio)) { |
1794 | /* cannot split folio, skip it */ | |
d4b4084a | 1795 | if (!can_split_folio(folio, NULL)) |
bd4c82c2 HY |
1796 | goto activate_locked; |
1797 | /* | |
c28a0e96 | 1798 | * Split folios without a PMD map right |
bd4c82c2 HY |
1799 | * away. Chances are some or all of the |
1800 | * tail pages can be freed without IO. | |
1801 | */ | |
d4b4084a | 1802 | if (!folio_entire_mapcount(folio) && |
346cf613 MWO |
1803 | split_folio_to_list(folio, |
1804 | page_list)) | |
bd4c82c2 HY |
1805 | goto activate_locked; |
1806 | } | |
09c02e56 MWO |
1807 | if (!add_to_swap(folio)) { |
1808 | if (!folio_test_large(folio)) | |
98879b3b | 1809 | goto activate_locked_split; |
bd4c82c2 | 1810 | /* Fallback to swap normal pages */ |
346cf613 MWO |
1811 | if (split_folio_to_list(folio, |
1812 | page_list)) | |
bd4c82c2 | 1813 | goto activate_locked; |
fe490cc0 HY |
1814 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE |
1815 | count_vm_event(THP_SWPOUT_FALLBACK); | |
1816 | #endif | |
09c02e56 | 1817 | if (!add_to_swap(folio)) |
98879b3b | 1818 | goto activate_locked_split; |
bd4c82c2 | 1819 | } |
bd4c82c2 | 1820 | } |
c28a0e96 MWO |
1821 | } else if (folio_test_swapbacked(folio) && |
1822 | folio_test_large(folio)) { | |
1823 | /* Split shmem folio */ | |
346cf613 | 1824 | if (split_folio_to_list(folio, page_list)) |
7751b2da | 1825 | goto keep_locked; |
e2be15f6 | 1826 | } |
1da177e4 | 1827 | |
98879b3b | 1828 | /* |
c28a0e96 MWO |
1829 | * If the folio was split above, the tail pages will make |
1830 | * their own pass through this function and be accounted | |
1831 | * then. | |
98879b3b | 1832 | */ |
c28a0e96 | 1833 | if ((nr_pages > 1) && !folio_test_large(folio)) { |
98879b3b YS |
1834 | sc->nr_scanned -= (nr_pages - 1); |
1835 | nr_pages = 1; | |
1836 | } | |
1837 | ||
1da177e4 | 1838 | /* |
1bee2c16 | 1839 | * The folio is mapped into the page tables of one or more |
1da177e4 LT |
1840 | * processes. Try to unmap it here. |
1841 | */ | |
1bee2c16 | 1842 | if (folio_mapped(folio)) { |
013339df | 1843 | enum ttu_flags flags = TTU_BATCH_FLUSH; |
1bee2c16 | 1844 | bool was_swapbacked = folio_test_swapbacked(folio); |
bd4c82c2 | 1845 | |
1bee2c16 | 1846 | if (folio_test_pmd_mappable(folio)) |
bd4c82c2 | 1847 | flags |= TTU_SPLIT_HUGE_PMD; |
1f318a9b | 1848 | |
869f7ee6 | 1849 | try_to_unmap(folio, flags); |
1bee2c16 | 1850 | if (folio_mapped(folio)) { |
98879b3b | 1851 | stat->nr_unmap_fail += nr_pages; |
1bee2c16 MWO |
1852 | if (!was_swapbacked && |
1853 | folio_test_swapbacked(folio)) | |
1f318a9b | 1854 | stat->nr_lazyfree_fail += nr_pages; |
1da177e4 | 1855 | goto activate_locked; |
1da177e4 LT |
1856 | } |
1857 | } | |
1858 | ||
5441d490 | 1859 | mapping = folio_mapping(folio); |
49bd2bf9 | 1860 | if (folio_test_dirty(folio)) { |
ee72886d | 1861 | /* |
49bd2bf9 | 1862 | * Only kswapd can writeback filesystem folios |
4eda4823 | 1863 | * to avoid risk of stack overflow. But avoid |
49bd2bf9 | 1864 | * injecting inefficient single-folio I/O into |
4eda4823 | 1865 | * flusher writeback as much as possible: only |
49bd2bf9 MWO |
1866 | * write folios when we've encountered many |
1867 | * dirty folios, and when we've already scanned | |
1868 | * the rest of the LRU for clean folios and see | |
1869 | * the same dirty folios again (with the reclaim | |
1870 | * flag set). | |
ee72886d | 1871 | */ |
49bd2bf9 MWO |
1872 | if (folio_is_file_lru(folio) && |
1873 | (!current_is_kswapd() || | |
1874 | !folio_test_reclaim(folio) || | |
4eda4823 | 1875 | !test_bit(PGDAT_DIRTY, &pgdat->flags))) { |
49ea7eb6 MG |
1876 | /* |
1877 | * Immediately reclaim when written back. | |
49bd2bf9 MWO |
1878 | * Similar in principle to deactivate_page() |
1879 | * except we already have the folio isolated | |
49ea7eb6 MG |
1880 | * and know it's dirty |
1881 | */ | |
49bd2bf9 MWO |
1882 | node_stat_mod_folio(folio, NR_VMSCAN_IMMEDIATE, |
1883 | nr_pages); | |
1884 | folio_set_reclaim(folio); | |
49ea7eb6 | 1885 | |
c55e8d03 | 1886 | goto activate_locked; |
ee72886d MG |
1887 | } |
1888 | ||
dfc8d636 | 1889 | if (references == PAGEREF_RECLAIM_CLEAN) |
1da177e4 | 1890 | goto keep_locked; |
c28a0e96 | 1891 | if (!may_enter_fs(folio, sc->gfp_mask)) |
1da177e4 | 1892 | goto keep_locked; |
52a8363e | 1893 | if (!sc->may_writepage) |
1da177e4 LT |
1894 | goto keep_locked; |
1895 | ||
d950c947 | 1896 | /* |
49bd2bf9 MWO |
1897 | * Folio is dirty. Flush the TLB if a writable entry |
1898 | * potentially exists to avoid CPU writes after I/O | |
d950c947 MG |
1899 | * starts and then write it out here. |
1900 | */ | |
1901 | try_to_unmap_flush_dirty(); | |
2282679f | 1902 | switch (pageout(folio, mapping, &plug)) { |
1da177e4 LT |
1903 | case PAGE_KEEP: |
1904 | goto keep_locked; | |
1905 | case PAGE_ACTIVATE: | |
1906 | goto activate_locked; | |
1907 | case PAGE_SUCCESS: | |
c79b7b96 | 1908 | stat->nr_pageout += nr_pages; |
96f8bf4f | 1909 | |
49bd2bf9 | 1910 | if (folio_test_writeback(folio)) |
41ac1999 | 1911 | goto keep; |
49bd2bf9 | 1912 | if (folio_test_dirty(folio)) |
1da177e4 | 1913 | goto keep; |
7d3579e8 | 1914 | |
1da177e4 LT |
1915 | /* |
1916 | * A synchronous write - probably a ramdisk. Go | |
49bd2bf9 | 1917 | * ahead and try to reclaim the folio. |
1da177e4 | 1918 | */ |
49bd2bf9 | 1919 | if (!folio_trylock(folio)) |
1da177e4 | 1920 | goto keep; |
49bd2bf9 MWO |
1921 | if (folio_test_dirty(folio) || |
1922 | folio_test_writeback(folio)) | |
1da177e4 | 1923 | goto keep_locked; |
49bd2bf9 | 1924 | mapping = folio_mapping(folio); |
01359eb2 | 1925 | fallthrough; |
1da177e4 | 1926 | case PAGE_CLEAN: |
49bd2bf9 | 1927 | ; /* try to free the folio below */ |
1da177e4 LT |
1928 | } |
1929 | } | |
1930 | ||
1931 | /* | |
0a36111c MWO |
1932 | * If the folio has buffers, try to free the buffer |
1933 | * mappings associated with this folio. If we succeed | |
1934 | * we try to free the folio as well. | |
1da177e4 | 1935 | * |
0a36111c MWO |
1936 | * We do this even if the folio is dirty. |
1937 | * filemap_release_folio() does not perform I/O, but it | |
1938 | * is possible for a folio to have the dirty flag set, | |
1939 | * but it is actually clean (all its buffers are clean). | |
1940 | * This happens if the buffers were written out directly, | |
1941 | * with submit_bh(). ext3 will do this, as well as | |
1942 | * the blockdev mapping. filemap_release_folio() will | |
1943 | * discover that cleanness and will drop the buffers | |
1944 | * and mark the folio clean - it can be freed. | |
1da177e4 | 1945 | * |
0a36111c MWO |
1946 | * Rarely, folios can have buffers and no ->mapping. |
1947 | * These are the folios which were not successfully | |
1948 | * invalidated in truncate_cleanup_folio(). We try to | |
1949 | * drop those buffers here and if that worked, and the | |
1950 | * folio is no longer mapped into process address space | |
1951 | * (refcount == 1) it can be freed. Otherwise, leave | |
1952 | * the folio on the LRU so it is swappable. | |
1da177e4 | 1953 | */ |
0a36111c MWO |
1954 | if (folio_has_private(folio)) { |
1955 | if (!filemap_release_folio(folio, sc->gfp_mask)) | |
1da177e4 | 1956 | goto activate_locked; |
0a36111c MWO |
1957 | if (!mapping && folio_ref_count(folio) == 1) { |
1958 | folio_unlock(folio); | |
1959 | if (folio_put_testzero(folio)) | |
e286781d NP |
1960 | goto free_it; |
1961 | else { | |
1962 | /* | |
1963 | * rare race with speculative reference. | |
1964 | * the speculative reference will free | |
0a36111c | 1965 | * this folio shortly, so we may |
e286781d NP |
1966 | * increment nr_reclaimed here (and |
1967 | * leave it off the LRU). | |
1968 | */ | |
9aafcffc | 1969 | nr_reclaimed += nr_pages; |
e286781d NP |
1970 | continue; |
1971 | } | |
1972 | } | |
1da177e4 LT |
1973 | } |
1974 | ||
64daa5d8 | 1975 | if (folio_test_anon(folio) && !folio_test_swapbacked(folio)) { |
802a3a92 | 1976 | /* follow __remove_mapping for reference */ |
64daa5d8 | 1977 | if (!folio_ref_freeze(folio, 1)) |
802a3a92 | 1978 | goto keep_locked; |
d17be2d9 | 1979 | /* |
64daa5d8 | 1980 | * The folio has only one reference left, which is |
d17be2d9 | 1981 | * from the isolation. After the caller puts the |
64daa5d8 MWO |
1982 | * folio back on the lru and drops the reference, the |
1983 | * folio will be freed anyway. It doesn't matter | |
1984 | * which lru it goes on. So we don't bother checking | |
1985 | * the dirty flag here. | |
d17be2d9 | 1986 | */ |
64daa5d8 MWO |
1987 | count_vm_events(PGLAZYFREED, nr_pages); |
1988 | count_memcg_folio_events(folio, PGLAZYFREED, nr_pages); | |
be7c07d6 | 1989 | } else if (!mapping || !__remove_mapping(mapping, folio, true, |
b910718a | 1990 | sc->target_mem_cgroup)) |
802a3a92 | 1991 | goto keep_locked; |
9a1ea439 | 1992 | |
c28a0e96 | 1993 | folio_unlock(folio); |
e286781d | 1994 | free_it: |
98879b3b | 1995 | /* |
c28a0e96 MWO |
1996 | * Folio may get swapped out as a whole, need to account |
1997 | * all pages in it. | |
98879b3b YS |
1998 | */ |
1999 | nr_reclaimed += nr_pages; | |
abe4c3b5 MG |
2000 | |
2001 | /* | |
2002 | * Is there need to periodically free_page_list? It would | |
2003 | * appear not as the counts should be low | |
2004 | */ | |
c28a0e96 | 2005 | if (unlikely(folio_test_large(folio))) |
5375336c | 2006 | destroy_large_folio(folio); |
7ae88534 | 2007 | else |
c28a0e96 | 2008 | list_add(&folio->lru, &free_pages); |
1da177e4 LT |
2009 | continue; |
2010 | ||
98879b3b YS |
2011 | activate_locked_split: |
2012 | /* | |
2013 | * The tail pages that are failed to add into swap cache | |
2014 | * reach here. Fixup nr_scanned and nr_pages. | |
2015 | */ | |
2016 | if (nr_pages > 1) { | |
2017 | sc->nr_scanned -= (nr_pages - 1); | |
2018 | nr_pages = 1; | |
2019 | } | |
1da177e4 | 2020 | activate_locked: |
68a22394 | 2021 | /* Not a candidate for swapping, so reclaim swap space. */ |
246b6480 MWO |
2022 | if (folio_test_swapcache(folio) && |
2023 | (mem_cgroup_swap_full(&folio->page) || | |
2024 | folio_test_mlocked(folio))) | |
2025 | try_to_free_swap(&folio->page); | |
2026 | VM_BUG_ON_FOLIO(folio_test_active(folio), folio); | |
2027 | if (!folio_test_mlocked(folio)) { | |
2028 | int type = folio_is_file_lru(folio); | |
2029 | folio_set_active(folio); | |
98879b3b | 2030 | stat->nr_activate[type] += nr_pages; |
246b6480 | 2031 | count_memcg_folio_events(folio, PGACTIVATE, nr_pages); |
ad6b6704 | 2032 | } |
1da177e4 | 2033 | keep_locked: |
c28a0e96 | 2034 | folio_unlock(folio); |
1da177e4 | 2035 | keep: |
c28a0e96 MWO |
2036 | list_add(&folio->lru, &ret_pages); |
2037 | VM_BUG_ON_FOLIO(folio_test_lru(folio) || | |
2038 | folio_test_unevictable(folio), folio); | |
1da177e4 | 2039 | } |
26aa2d19 DH |
2040 | /* 'page_list' is always empty here */ |
2041 | ||
c28a0e96 | 2042 | /* Migrate folios selected for demotion */ |
26aa2d19 | 2043 | nr_reclaimed += demote_page_list(&demote_pages, pgdat); |
c28a0e96 | 2044 | /* Folios that could not be demoted are still in @demote_pages */ |
26aa2d19 | 2045 | if (!list_empty(&demote_pages)) { |
c28a0e96 | 2046 | /* Folios which weren't demoted go back on @page_list for retry: */ |
26aa2d19 DH |
2047 | list_splice_init(&demote_pages, page_list); |
2048 | do_demote_pass = false; | |
2049 | goto retry; | |
2050 | } | |
abe4c3b5 | 2051 | |
98879b3b YS |
2052 | pgactivate = stat->nr_activate[0] + stat->nr_activate[1]; |
2053 | ||
747db954 | 2054 | mem_cgroup_uncharge_list(&free_pages); |
72b252ae | 2055 | try_to_unmap_flush(); |
2d4894b5 | 2056 | free_unref_page_list(&free_pages); |
abe4c3b5 | 2057 | |
1da177e4 | 2058 | list_splice(&ret_pages, page_list); |
886cf190 | 2059 | count_vm_events(PGACTIVATE, pgactivate); |
060f005f | 2060 | |
2282679f N |
2061 | if (plug) |
2062 | swap_write_unplug(plug); | |
05ff5137 | 2063 | return nr_reclaimed; |
1da177e4 LT |
2064 | } |
2065 | ||
730ec8c0 | 2066 | unsigned int reclaim_clean_pages_from_list(struct zone *zone, |
b8cecb93 | 2067 | struct list_head *folio_list) |
02c6de8d MK |
2068 | { |
2069 | struct scan_control sc = { | |
2070 | .gfp_mask = GFP_KERNEL, | |
02c6de8d MK |
2071 | .may_unmap = 1, |
2072 | }; | |
1f318a9b | 2073 | struct reclaim_stat stat; |
730ec8c0 | 2074 | unsigned int nr_reclaimed; |
b8cecb93 MWO |
2075 | struct folio *folio, *next; |
2076 | LIST_HEAD(clean_folios); | |
2d2b8d2b | 2077 | unsigned int noreclaim_flag; |
02c6de8d | 2078 | |
b8cecb93 MWO |
2079 | list_for_each_entry_safe(folio, next, folio_list, lru) { |
2080 | if (!folio_test_hugetlb(folio) && folio_is_file_lru(folio) && | |
2081 | !folio_test_dirty(folio) && !__folio_test_movable(folio) && | |
2082 | !folio_test_unevictable(folio)) { | |
2083 | folio_clear_active(folio); | |
2084 | list_move(&folio->lru, &clean_folios); | |
02c6de8d MK |
2085 | } |
2086 | } | |
2087 | ||
2d2b8d2b YZ |
2088 | /* |
2089 | * We should be safe here since we are only dealing with file pages and | |
2090 | * we are not kswapd and therefore cannot write dirty file pages. But | |
2091 | * call memalloc_noreclaim_save() anyway, just in case these conditions | |
2092 | * change in the future. | |
2093 | */ | |
2094 | noreclaim_flag = memalloc_noreclaim_save(); | |
b8cecb93 | 2095 | nr_reclaimed = shrink_page_list(&clean_folios, zone->zone_pgdat, &sc, |
013339df | 2096 | &stat, true); |
2d2b8d2b YZ |
2097 | memalloc_noreclaim_restore(noreclaim_flag); |
2098 | ||
b8cecb93 | 2099 | list_splice(&clean_folios, folio_list); |
2da9f630 NP |
2100 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, |
2101 | -(long)nr_reclaimed); | |
1f318a9b JK |
2102 | /* |
2103 | * Since lazyfree pages are isolated from file LRU from the beginning, | |
2104 | * they will rotate back to anonymous LRU in the end if it failed to | |
2105 | * discard so isolated count will be mismatched. | |
2106 | * Compensate the isolated count for both LRU lists. | |
2107 | */ | |
2108 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_ANON, | |
2109 | stat.nr_lazyfree_fail); | |
2110 | mod_node_page_state(zone->zone_pgdat, NR_ISOLATED_FILE, | |
2da9f630 | 2111 | -(long)stat.nr_lazyfree_fail); |
1f318a9b | 2112 | return nr_reclaimed; |
02c6de8d MK |
2113 | } |
2114 | ||
7ee36a14 MG |
2115 | /* |
2116 | * Update LRU sizes after isolating pages. The LRU size updates must | |
55b65a57 | 2117 | * be complete before mem_cgroup_update_lru_size due to a sanity check. |
7ee36a14 MG |
2118 | */ |
2119 | static __always_inline void update_lru_sizes(struct lruvec *lruvec, | |
b4536f0c | 2120 | enum lru_list lru, unsigned long *nr_zone_taken) |
7ee36a14 | 2121 | { |
7ee36a14 MG |
2122 | int zid; |
2123 | ||
7ee36a14 MG |
2124 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
2125 | if (!nr_zone_taken[zid]) | |
2126 | continue; | |
2127 | ||
a892cb6b | 2128 | update_lru_size(lruvec, lru, zid, -nr_zone_taken[zid]); |
b4536f0c MH |
2129 | } |
2130 | ||
7ee36a14 MG |
2131 | } |
2132 | ||
f611fab7 | 2133 | /* |
15b44736 HD |
2134 | * Isolating page from the lruvec to fill in @dst list by nr_to_scan times. |
2135 | * | |
2136 | * lruvec->lru_lock is heavily contended. Some of the functions that | |
1da177e4 LT |
2137 | * shrink the lists perform better by taking out a batch of pages |
2138 | * and working on them outside the LRU lock. | |
2139 | * | |
2140 | * For pagecache intensive workloads, this function is the hottest | |
2141 | * spot in the kernel (apart from copy_*_user functions). | |
2142 | * | |
15b44736 | 2143 | * Lru_lock must be held before calling this function. |
1da177e4 | 2144 | * |
791b48b6 | 2145 | * @nr_to_scan: The number of eligible pages to look through on the list. |
5dc35979 | 2146 | * @lruvec: The LRU vector to pull pages from. |
1da177e4 | 2147 | * @dst: The temp list to put pages on to. |
f626012d | 2148 | * @nr_scanned: The number of pages that were scanned. |
fe2c2a10 | 2149 | * @sc: The scan_control struct for this reclaim session |
3cb99451 | 2150 | * @lru: LRU list id for isolating |
1da177e4 LT |
2151 | * |
2152 | * returns how many pages were moved onto *@dst. | |
2153 | */ | |
69e05944 | 2154 | static unsigned long isolate_lru_pages(unsigned long nr_to_scan, |
5dc35979 | 2155 | struct lruvec *lruvec, struct list_head *dst, |
fe2c2a10 | 2156 | unsigned long *nr_scanned, struct scan_control *sc, |
a9e7c39f | 2157 | enum lru_list lru) |
1da177e4 | 2158 | { |
75b00af7 | 2159 | struct list_head *src = &lruvec->lists[lru]; |
69e05944 | 2160 | unsigned long nr_taken = 0; |
599d0c95 | 2161 | unsigned long nr_zone_taken[MAX_NR_ZONES] = { 0 }; |
7cc30fcf | 2162 | unsigned long nr_skipped[MAX_NR_ZONES] = { 0, }; |
3db65812 | 2163 | unsigned long skipped = 0; |
791b48b6 | 2164 | unsigned long scan, total_scan, nr_pages; |
166e3d32 | 2165 | LIST_HEAD(folios_skipped); |
1da177e4 | 2166 | |
98879b3b | 2167 | total_scan = 0; |
791b48b6 | 2168 | scan = 0; |
98879b3b | 2169 | while (scan < nr_to_scan && !list_empty(src)) { |
89f6c88a | 2170 | struct list_head *move_to = src; |
166e3d32 | 2171 | struct folio *folio; |
5ad333eb | 2172 | |
166e3d32 MWO |
2173 | folio = lru_to_folio(src); |
2174 | prefetchw_prev_lru_folio(folio, src, flags); | |
1da177e4 | 2175 | |
166e3d32 | 2176 | nr_pages = folio_nr_pages(folio); |
98879b3b YS |
2177 | total_scan += nr_pages; |
2178 | ||
166e3d32 MWO |
2179 | if (folio_zonenum(folio) > sc->reclaim_idx) { |
2180 | nr_skipped[folio_zonenum(folio)] += nr_pages; | |
2181 | move_to = &folios_skipped; | |
89f6c88a | 2182 | goto move; |
b2e18757 MG |
2183 | } |
2184 | ||
791b48b6 | 2185 | /* |
166e3d32 MWO |
2186 | * Do not count skipped folios because that makes the function |
2187 | * return with no isolated folios if the LRU mostly contains | |
2188 | * ineligible folios. This causes the VM to not reclaim any | |
2189 | * folios, triggering a premature OOM. | |
2190 | * Account all pages in a folio. | |
791b48b6 | 2191 | */ |
98879b3b | 2192 | scan += nr_pages; |
89f6c88a | 2193 | |
166e3d32 | 2194 | if (!folio_test_lru(folio)) |
89f6c88a | 2195 | goto move; |
166e3d32 | 2196 | if (!sc->may_unmap && folio_mapped(folio)) |
89f6c88a HD |
2197 | goto move; |
2198 | ||
c2135f7c | 2199 | /* |
166e3d32 MWO |
2200 | * Be careful not to clear the lru flag until after we're |
2201 | * sure the folio is not being freed elsewhere -- the | |
2202 | * folio release code relies on it. | |
c2135f7c | 2203 | */ |
166e3d32 | 2204 | if (unlikely(!folio_try_get(folio))) |
89f6c88a | 2205 | goto move; |
5ad333eb | 2206 | |
166e3d32 MWO |
2207 | if (!folio_test_clear_lru(folio)) { |
2208 | /* Another thread is already isolating this folio */ | |
2209 | folio_put(folio); | |
89f6c88a | 2210 | goto move; |
5ad333eb | 2211 | } |
c2135f7c AS |
2212 | |
2213 | nr_taken += nr_pages; | |
166e3d32 | 2214 | nr_zone_taken[folio_zonenum(folio)] += nr_pages; |
89f6c88a HD |
2215 | move_to = dst; |
2216 | move: | |
166e3d32 | 2217 | list_move(&folio->lru, move_to); |
1da177e4 LT |
2218 | } |
2219 | ||
b2e18757 | 2220 | /* |
166e3d32 | 2221 | * Splice any skipped folios to the start of the LRU list. Note that |
b2e18757 MG |
2222 | * this disrupts the LRU order when reclaiming for lower zones but |
2223 | * we cannot splice to the tail. If we did then the SWAP_CLUSTER_MAX | |
166e3d32 | 2224 | * scanning would soon rescan the same folios to skip and waste lots |
b2cb6826 | 2225 | * of cpu cycles. |
b2e18757 | 2226 | */ |
166e3d32 | 2227 | if (!list_empty(&folios_skipped)) { |
7cc30fcf MG |
2228 | int zid; |
2229 | ||
166e3d32 | 2230 | list_splice(&folios_skipped, src); |
7cc30fcf MG |
2231 | for (zid = 0; zid < MAX_NR_ZONES; zid++) { |
2232 | if (!nr_skipped[zid]) | |
2233 | continue; | |
2234 | ||
2235 | __count_zid_vm_events(PGSCAN_SKIP, zid, nr_skipped[zid]); | |
1265e3a6 | 2236 | skipped += nr_skipped[zid]; |
7cc30fcf MG |
2237 | } |
2238 | } | |
791b48b6 | 2239 | *nr_scanned = total_scan; |
1265e3a6 | 2240 | trace_mm_vmscan_lru_isolate(sc->reclaim_idx, sc->order, nr_to_scan, |
89f6c88a HD |
2241 | total_scan, skipped, nr_taken, |
2242 | sc->may_unmap ? 0 : ISOLATE_UNMAPPED, lru); | |
b4536f0c | 2243 | update_lru_sizes(lruvec, lru, nr_zone_taken); |
1da177e4 LT |
2244 | return nr_taken; |
2245 | } | |
2246 | ||
62695a84 | 2247 | /** |
d1d8a3b4 MWO |
2248 | * folio_isolate_lru() - Try to isolate a folio from its LRU list. |
2249 | * @folio: Folio to isolate from its LRU list. | |
62695a84 | 2250 | * |
d1d8a3b4 MWO |
2251 | * Isolate a @folio from an LRU list and adjust the vmstat statistic |
2252 | * corresponding to whatever LRU list the folio was on. | |
62695a84 | 2253 | * |
d1d8a3b4 MWO |
2254 | * The folio will have its LRU flag cleared. If it was found on the |
2255 | * active list, it will have the Active flag set. If it was found on the | |
2256 | * unevictable list, it will have the Unevictable flag set. These flags | |
894bc310 | 2257 | * may need to be cleared by the caller before letting the page go. |
62695a84 | 2258 | * |
d1d8a3b4 | 2259 | * Context: |
a5d09bed | 2260 | * |
62695a84 | 2261 | * (1) Must be called with an elevated refcount on the page. This is a |
d1d8a3b4 | 2262 | * fundamental difference from isolate_lru_pages() (which is called |
62695a84 | 2263 | * without a stable reference). |
d1d8a3b4 MWO |
2264 | * (2) The lru_lock must not be held. |
2265 | * (3) Interrupts must be enabled. | |
2266 | * | |
2267 | * Return: 0 if the folio was removed from an LRU list. | |
2268 | * -EBUSY if the folio was not on an LRU list. | |
62695a84 | 2269 | */ |
d1d8a3b4 | 2270 | int folio_isolate_lru(struct folio *folio) |
62695a84 NP |
2271 | { |
2272 | int ret = -EBUSY; | |
2273 | ||
d1d8a3b4 | 2274 | VM_BUG_ON_FOLIO(!folio_ref_count(folio), folio); |
0c917313 | 2275 | |
d1d8a3b4 | 2276 | if (folio_test_clear_lru(folio)) { |
fa9add64 | 2277 | struct lruvec *lruvec; |
62695a84 | 2278 | |
d1d8a3b4 | 2279 | folio_get(folio); |
e809c3fe | 2280 | lruvec = folio_lruvec_lock_irq(folio); |
d1d8a3b4 | 2281 | lruvec_del_folio(lruvec, folio); |
6168d0da | 2282 | unlock_page_lruvec_irq(lruvec); |
d25b5bd8 | 2283 | ret = 0; |
62695a84 | 2284 | } |
d25b5bd8 | 2285 | |
62695a84 NP |
2286 | return ret; |
2287 | } | |
2288 | ||
35cd7815 | 2289 | /* |
d37dd5dc | 2290 | * A direct reclaimer may isolate SWAP_CLUSTER_MAX pages from the LRU list and |
178821b8 | 2291 | * then get rescheduled. When there are massive number of tasks doing page |
d37dd5dc FW |
2292 | * allocation, such sleeping direct reclaimers may keep piling up on each CPU, |
2293 | * the LRU list will go small and be scanned faster than necessary, leading to | |
2294 | * unnecessary swapping, thrashing and OOM. | |
35cd7815 | 2295 | */ |
599d0c95 | 2296 | static int too_many_isolated(struct pglist_data *pgdat, int file, |
35cd7815 RR |
2297 | struct scan_control *sc) |
2298 | { | |
2299 | unsigned long inactive, isolated; | |
d818fca1 | 2300 | bool too_many; |
35cd7815 RR |
2301 | |
2302 | if (current_is_kswapd()) | |
2303 | return 0; | |
2304 | ||
b5ead35e | 2305 | if (!writeback_throttling_sane(sc)) |
35cd7815 RR |
2306 | return 0; |
2307 | ||
2308 | if (file) { | |
599d0c95 MG |
2309 | inactive = node_page_state(pgdat, NR_INACTIVE_FILE); |
2310 | isolated = node_page_state(pgdat, NR_ISOLATED_FILE); | |
35cd7815 | 2311 | } else { |
599d0c95 MG |
2312 | inactive = node_page_state(pgdat, NR_INACTIVE_ANON); |
2313 | isolated = node_page_state(pgdat, NR_ISOLATED_ANON); | |
35cd7815 RR |
2314 | } |
2315 | ||
3cf23841 FW |
2316 | /* |
2317 | * GFP_NOIO/GFP_NOFS callers are allowed to isolate more pages, so they | |
2318 | * won't get blocked by normal direct-reclaimers, forming a circular | |
2319 | * deadlock. | |
2320 | */ | |
d0164adc | 2321 | if ((sc->gfp_mask & (__GFP_IO | __GFP_FS)) == (__GFP_IO | __GFP_FS)) |
3cf23841 FW |
2322 | inactive >>= 3; |
2323 | ||
d818fca1 MG |
2324 | too_many = isolated > inactive; |
2325 | ||
2326 | /* Wake up tasks throttled due to too_many_isolated. */ | |
2327 | if (!too_many) | |
2328 | wake_throttle_isolated(pgdat); | |
2329 | ||
2330 | return too_many; | |
35cd7815 RR |
2331 | } |
2332 | ||
a222f341 | 2333 | /* |
ff00a170 MWO |
2334 | * move_pages_to_lru() moves folios from private @list to appropriate LRU list. |
2335 | * On return, @list is reused as a list of folios to be freed by the caller. | |
a222f341 KT |
2336 | * |
2337 | * Returns the number of pages moved to the given lruvec. | |
2338 | */ | |
9ef56b78 MS |
2339 | static unsigned int move_pages_to_lru(struct lruvec *lruvec, |
2340 | struct list_head *list) | |
66635629 | 2341 | { |
a222f341 | 2342 | int nr_pages, nr_moved = 0; |
ff00a170 | 2343 | LIST_HEAD(folios_to_free); |
66635629 | 2344 | |
a222f341 | 2345 | while (!list_empty(list)) { |
ff00a170 MWO |
2346 | struct folio *folio = lru_to_folio(list); |
2347 | ||
2348 | VM_BUG_ON_FOLIO(folio_test_lru(folio), folio); | |
2349 | list_del(&folio->lru); | |
2350 | if (unlikely(!folio_evictable(folio))) { | |
6168d0da | 2351 | spin_unlock_irq(&lruvec->lru_lock); |
ff00a170 | 2352 | folio_putback_lru(folio); |
6168d0da | 2353 | spin_lock_irq(&lruvec->lru_lock); |
66635629 MG |
2354 | continue; |
2355 | } | |
fa9add64 | 2356 | |
3d06afab | 2357 | /* |
ff00a170 | 2358 | * The folio_set_lru needs to be kept here for list integrity. |
3d06afab AS |
2359 | * Otherwise: |
2360 | * #0 move_pages_to_lru #1 release_pages | |
ff00a170 MWO |
2361 | * if (!folio_put_testzero()) |
2362 | * if (folio_put_testzero()) | |
2363 | * !lru //skip lru_lock | |
2364 | * folio_set_lru() | |
2365 | * list_add(&folio->lru,) | |
2366 | * list_add(&folio->lru,) | |
3d06afab | 2367 | */ |
ff00a170 | 2368 | folio_set_lru(folio); |
a222f341 | 2369 | |
ff00a170 MWO |
2370 | if (unlikely(folio_put_testzero(folio))) { |
2371 | __folio_clear_lru_flags(folio); | |
2bcf8879 | 2372 | |
ff00a170 | 2373 | if (unlikely(folio_test_large(folio))) { |
6168d0da | 2374 | spin_unlock_irq(&lruvec->lru_lock); |
5375336c | 2375 | destroy_large_folio(folio); |
6168d0da | 2376 | spin_lock_irq(&lruvec->lru_lock); |
2bcf8879 | 2377 | } else |
ff00a170 | 2378 | list_add(&folio->lru, &folios_to_free); |
3d06afab AS |
2379 | |
2380 | continue; | |
66635629 | 2381 | } |
3d06afab | 2382 | |
afca9157 AS |
2383 | /* |
2384 | * All pages were isolated from the same lruvec (and isolation | |
2385 | * inhibits memcg migration). | |
2386 | */ | |
ff00a170 MWO |
2387 | VM_BUG_ON_FOLIO(!folio_matches_lruvec(folio, lruvec), folio); |
2388 | lruvec_add_folio(lruvec, folio); | |
2389 | nr_pages = folio_nr_pages(folio); | |
3d06afab | 2390 | nr_moved += nr_pages; |
ff00a170 | 2391 | if (folio_test_active(folio)) |
3d06afab | 2392 | workingset_age_nonresident(lruvec, nr_pages); |
66635629 | 2393 | } |
66635629 | 2394 | |
3f79768f HD |
2395 | /* |
2396 | * To save our caller's stack, now use input list for pages to free. | |
2397 | */ | |
ff00a170 | 2398 | list_splice(&folios_to_free, list); |
a222f341 KT |
2399 | |
2400 | return nr_moved; | |
66635629 MG |
2401 | } |
2402 | ||
399ba0b9 | 2403 | /* |
5829f7db ML |
2404 | * If a kernel thread (such as nfsd for loop-back mounts) services a backing |
2405 | * device by writing to the page cache it sets PF_LOCAL_THROTTLE. In this case | |
2406 | * we should not throttle. Otherwise it is safe to do so. | |
399ba0b9 N |
2407 | */ |
2408 | static int current_may_throttle(void) | |
2409 | { | |
b9b1335e | 2410 | return !(current->flags & PF_LOCAL_THROTTLE); |
399ba0b9 N |
2411 | } |
2412 | ||
1da177e4 | 2413 | /* |
b2e18757 | 2414 | * shrink_inactive_list() is a helper for shrink_node(). It returns the number |
1742f19f | 2415 | * of reclaimed pages |
1da177e4 | 2416 | */ |
9ef56b78 | 2417 | static unsigned long |
1a93be0e | 2418 | shrink_inactive_list(unsigned long nr_to_scan, struct lruvec *lruvec, |
9e3b2f8c | 2419 | struct scan_control *sc, enum lru_list lru) |
1da177e4 LT |
2420 | { |
2421 | LIST_HEAD(page_list); | |
e247dbce | 2422 | unsigned long nr_scanned; |
730ec8c0 | 2423 | unsigned int nr_reclaimed = 0; |
e247dbce | 2424 | unsigned long nr_taken; |
060f005f | 2425 | struct reclaim_stat stat; |
497a6c1b | 2426 | bool file = is_file_lru(lru); |
f46b7912 | 2427 | enum vm_event_item item; |
599d0c95 | 2428 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
db73ee0d | 2429 | bool stalled = false; |
78dc583d | 2430 | |
599d0c95 | 2431 | while (unlikely(too_many_isolated(pgdat, file, sc))) { |
db73ee0d MH |
2432 | if (stalled) |
2433 | return 0; | |
2434 | ||
2435 | /* wait a bit for the reclaimer. */ | |
db73ee0d | 2436 | stalled = true; |
c3f4a9a2 | 2437 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_ISOLATED); |
35cd7815 RR |
2438 | |
2439 | /* We are about to die and free our memory. Return now. */ | |
2440 | if (fatal_signal_pending(current)) | |
2441 | return SWAP_CLUSTER_MAX; | |
2442 | } | |
2443 | ||
1da177e4 | 2444 | lru_add_drain(); |
f80c0673 | 2445 | |
6168d0da | 2446 | spin_lock_irq(&lruvec->lru_lock); |
b35ea17b | 2447 | |
5dc35979 | 2448 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &page_list, |
a9e7c39f | 2449 | &nr_scanned, sc, lru); |
95d918fc | 2450 | |
599d0c95 | 2451 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
f46b7912 | 2452 | item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT; |
b5ead35e | 2453 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
2454 | __count_vm_events(item, nr_scanned); |
2455 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_scanned); | |
497a6c1b JW |
2456 | __count_vm_events(PGSCAN_ANON + file, nr_scanned); |
2457 | ||
6168d0da | 2458 | spin_unlock_irq(&lruvec->lru_lock); |
b35ea17b | 2459 | |
d563c050 | 2460 | if (nr_taken == 0) |
66635629 | 2461 | return 0; |
5ad333eb | 2462 | |
013339df | 2463 | nr_reclaimed = shrink_page_list(&page_list, pgdat, sc, &stat, false); |
c661b078 | 2464 | |
6168d0da | 2465 | spin_lock_irq(&lruvec->lru_lock); |
497a6c1b JW |
2466 | move_pages_to_lru(lruvec, &page_list); |
2467 | ||
2468 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); | |
f46b7912 | 2469 | item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT; |
b5ead35e | 2470 | if (!cgroup_reclaim(sc)) |
f46b7912 KT |
2471 | __count_vm_events(item, nr_reclaimed); |
2472 | __count_memcg_events(lruvec_memcg(lruvec), item, nr_reclaimed); | |
497a6c1b | 2473 | __count_vm_events(PGSTEAL_ANON + file, nr_reclaimed); |
6168d0da | 2474 | spin_unlock_irq(&lruvec->lru_lock); |
3f79768f | 2475 | |
75cc3c91 | 2476 | lru_note_cost(lruvec, file, stat.nr_pageout); |
747db954 | 2477 | mem_cgroup_uncharge_list(&page_list); |
2d4894b5 | 2478 | free_unref_page_list(&page_list); |
e11da5b4 | 2479 | |
1c610d5f AR |
2480 | /* |
2481 | * If dirty pages are scanned that are not queued for IO, it | |
2482 | * implies that flushers are not doing their job. This can | |
2483 | * happen when memory pressure pushes dirty pages to the end of | |
2484 | * the LRU before the dirty limits are breached and the dirty | |
2485 | * data has expired. It can also happen when the proportion of | |
2486 | * dirty pages grows not through writes but through memory | |
2487 | * pressure reclaiming all the clean cache. And in some cases, | |
2488 | * the flushers simply cannot keep up with the allocation | |
2489 | * rate. Nudge the flusher threads in case they are asleep. | |
2490 | */ | |
2491 | if (stat.nr_unqueued_dirty == nr_taken) | |
2492 | wakeup_flusher_threads(WB_REASON_VMSCAN); | |
2493 | ||
d108c772 AR |
2494 | sc->nr.dirty += stat.nr_dirty; |
2495 | sc->nr.congested += stat.nr_congested; | |
2496 | sc->nr.unqueued_dirty += stat.nr_unqueued_dirty; | |
2497 | sc->nr.writeback += stat.nr_writeback; | |
2498 | sc->nr.immediate += stat.nr_immediate; | |
2499 | sc->nr.taken += nr_taken; | |
2500 | if (file) | |
2501 | sc->nr.file_taken += nr_taken; | |
8e950282 | 2502 | |
599d0c95 | 2503 | trace_mm_vmscan_lru_shrink_inactive(pgdat->node_id, |
d51d1e64 | 2504 | nr_scanned, nr_reclaimed, &stat, sc->priority, file); |
05ff5137 | 2505 | return nr_reclaimed; |
1da177e4 LT |
2506 | } |
2507 | ||
15b44736 | 2508 | /* |
07f67a8d | 2509 | * shrink_active_list() moves folios from the active LRU to the inactive LRU. |
15b44736 | 2510 | * |
07f67a8d | 2511 | * We move them the other way if the folio is referenced by one or more |
15b44736 HD |
2512 | * processes. |
2513 | * | |
07f67a8d | 2514 | * If the folios are mostly unmapped, the processing is fast and it is |
15b44736 | 2515 | * appropriate to hold lru_lock across the whole operation. But if |
07f67a8d MWO |
2516 | * the folios are mapped, the processing is slow (folio_referenced()), so |
2517 | * we should drop lru_lock around each folio. It's impossible to balance | |
2518 | * this, so instead we remove the folios from the LRU while processing them. | |
2519 | * It is safe to rely on the active flag against the non-LRU folios in here | |
2520 | * because nobody will play with that bit on a non-LRU folio. | |
15b44736 | 2521 | * |
07f67a8d MWO |
2522 | * The downside is that we have to touch folio->_refcount against each folio. |
2523 | * But we had to alter folio->flags anyway. | |
15b44736 | 2524 | */ |
f626012d | 2525 | static void shrink_active_list(unsigned long nr_to_scan, |
1a93be0e | 2526 | struct lruvec *lruvec, |
f16015fb | 2527 | struct scan_control *sc, |
9e3b2f8c | 2528 | enum lru_list lru) |
1da177e4 | 2529 | { |
44c241f1 | 2530 | unsigned long nr_taken; |
f626012d | 2531 | unsigned long nr_scanned; |
6fe6b7e3 | 2532 | unsigned long vm_flags; |
07f67a8d | 2533 | LIST_HEAD(l_hold); /* The folios which were snipped off */ |
8cab4754 | 2534 | LIST_HEAD(l_active); |
b69408e8 | 2535 | LIST_HEAD(l_inactive); |
9d998b4f MH |
2536 | unsigned nr_deactivate, nr_activate; |
2537 | unsigned nr_rotated = 0; | |
3cb99451 | 2538 | int file = is_file_lru(lru); |
599d0c95 | 2539 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
1da177e4 LT |
2540 | |
2541 | lru_add_drain(); | |
f80c0673 | 2542 | |
6168d0da | 2543 | spin_lock_irq(&lruvec->lru_lock); |
925b7673 | 2544 | |
5dc35979 | 2545 | nr_taken = isolate_lru_pages(nr_to_scan, lruvec, &l_hold, |
a9e7c39f | 2546 | &nr_scanned, sc, lru); |
89b5fae5 | 2547 | |
599d0c95 | 2548 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, nr_taken); |
1cfb419b | 2549 | |
912c0572 SB |
2550 | if (!cgroup_reclaim(sc)) |
2551 | __count_vm_events(PGREFILL, nr_scanned); | |
2fa2690c | 2552 | __count_memcg_events(lruvec_memcg(lruvec), PGREFILL, nr_scanned); |
9d5e6a9f | 2553 | |
6168d0da | 2554 | spin_unlock_irq(&lruvec->lru_lock); |
1da177e4 | 2555 | |
1da177e4 | 2556 | while (!list_empty(&l_hold)) { |
b3ac0413 | 2557 | struct folio *folio; |
b3ac0413 | 2558 | |
1da177e4 | 2559 | cond_resched(); |
b3ac0413 MWO |
2560 | folio = lru_to_folio(&l_hold); |
2561 | list_del(&folio->lru); | |
7e9cd484 | 2562 | |
07f67a8d MWO |
2563 | if (unlikely(!folio_evictable(folio))) { |
2564 | folio_putback_lru(folio); | |
894bc310 LS |
2565 | continue; |
2566 | } | |
2567 | ||
cc715d99 | 2568 | if (unlikely(buffer_heads_over_limit)) { |
36a3b14b MWO |
2569 | if (folio_test_private(folio) && folio_trylock(folio)) { |
2570 | if (folio_test_private(folio)) | |
07f67a8d MWO |
2571 | filemap_release_folio(folio, 0); |
2572 | folio_unlock(folio); | |
cc715d99 MG |
2573 | } |
2574 | } | |
2575 | ||
6d4675e6 | 2576 | /* Referenced or rmap lock contention: rotate */ |
b3ac0413 | 2577 | if (folio_referenced(folio, 0, sc->target_mem_cgroup, |
6d4675e6 | 2578 | &vm_flags) != 0) { |
8cab4754 | 2579 | /* |
07f67a8d | 2580 | * Identify referenced, file-backed active folios and |
8cab4754 WF |
2581 | * give them one more trip around the active list. So |
2582 | * that executable code get better chances to stay in | |
07f67a8d | 2583 | * memory under moderate memory pressure. Anon folios |
8cab4754 | 2584 | * are not likely to be evicted by use-once streaming |
07f67a8d | 2585 | * IO, plus JVM can create lots of anon VM_EXEC folios, |
8cab4754 WF |
2586 | * so we ignore them here. |
2587 | */ | |
07f67a8d MWO |
2588 | if ((vm_flags & VM_EXEC) && folio_is_file_lru(folio)) { |
2589 | nr_rotated += folio_nr_pages(folio); | |
2590 | list_add(&folio->lru, &l_active); | |
8cab4754 WF |
2591 | continue; |
2592 | } | |
2593 | } | |
7e9cd484 | 2594 | |
07f67a8d MWO |
2595 | folio_clear_active(folio); /* we are de-activating */ |
2596 | folio_set_workingset(folio); | |
2597 | list_add(&folio->lru, &l_inactive); | |
1da177e4 LT |
2598 | } |
2599 | ||
b555749a | 2600 | /* |
07f67a8d | 2601 | * Move folios back to the lru list. |
b555749a | 2602 | */ |
6168d0da | 2603 | spin_lock_irq(&lruvec->lru_lock); |
556adecb | 2604 | |
a222f341 KT |
2605 | nr_activate = move_pages_to_lru(lruvec, &l_active); |
2606 | nr_deactivate = move_pages_to_lru(lruvec, &l_inactive); | |
07f67a8d | 2607 | /* Keep all free folios in l_active list */ |
f372d89e | 2608 | list_splice(&l_inactive, &l_active); |
9851ac13 KT |
2609 | |
2610 | __count_vm_events(PGDEACTIVATE, nr_deactivate); | |
2611 | __count_memcg_events(lruvec_memcg(lruvec), PGDEACTIVATE, nr_deactivate); | |
2612 | ||
599d0c95 | 2613 | __mod_node_page_state(pgdat, NR_ISOLATED_ANON + file, -nr_taken); |
6168d0da | 2614 | spin_unlock_irq(&lruvec->lru_lock); |
2bcf8879 | 2615 | |
f372d89e KT |
2616 | mem_cgroup_uncharge_list(&l_active); |
2617 | free_unref_page_list(&l_active); | |
9d998b4f MH |
2618 | trace_mm_vmscan_lru_shrink_active(pgdat->node_id, nr_taken, nr_activate, |
2619 | nr_deactivate, nr_rotated, sc->priority, file); | |
1da177e4 LT |
2620 | } |
2621 | ||
1fe47c0b ML |
2622 | static unsigned int reclaim_page_list(struct list_head *page_list, |
2623 | struct pglist_data *pgdat) | |
1a4e58cc | 2624 | { |
1a4e58cc | 2625 | struct reclaim_stat dummy_stat; |
1fe47c0b ML |
2626 | unsigned int nr_reclaimed; |
2627 | struct folio *folio; | |
1a4e58cc MK |
2628 | struct scan_control sc = { |
2629 | .gfp_mask = GFP_KERNEL, | |
1a4e58cc MK |
2630 | .may_writepage = 1, |
2631 | .may_unmap = 1, | |
2632 | .may_swap = 1, | |
26aa2d19 | 2633 | .no_demotion = 1, |
1a4e58cc MK |
2634 | }; |
2635 | ||
1fe47c0b ML |
2636 | nr_reclaimed = shrink_page_list(page_list, pgdat, &sc, &dummy_stat, false); |
2637 | while (!list_empty(page_list)) { | |
2638 | folio = lru_to_folio(page_list); | |
2639 | list_del(&folio->lru); | |
2640 | folio_putback_lru(folio); | |
2641 | } | |
2642 | ||
2643 | return nr_reclaimed; | |
2644 | } | |
2645 | ||
a83f0551 | 2646 | unsigned long reclaim_pages(struct list_head *folio_list) |
1fe47c0b | 2647 | { |
ed657e55 | 2648 | int nid; |
1fe47c0b | 2649 | unsigned int nr_reclaimed = 0; |
a83f0551 | 2650 | LIST_HEAD(node_folio_list); |
1fe47c0b ML |
2651 | unsigned int noreclaim_flag; |
2652 | ||
a83f0551 | 2653 | if (list_empty(folio_list)) |
1ae65e27 WY |
2654 | return nr_reclaimed; |
2655 | ||
2d2b8d2b YZ |
2656 | noreclaim_flag = memalloc_noreclaim_save(); |
2657 | ||
a83f0551 | 2658 | nid = folio_nid(lru_to_folio(folio_list)); |
1ae65e27 | 2659 | do { |
a83f0551 | 2660 | struct folio *folio = lru_to_folio(folio_list); |
1a4e58cc | 2661 | |
a83f0551 MWO |
2662 | if (nid == folio_nid(folio)) { |
2663 | folio_clear_active(folio); | |
2664 | list_move(&folio->lru, &node_folio_list); | |
1a4e58cc MK |
2665 | continue; |
2666 | } | |
2667 | ||
a83f0551 MWO |
2668 | nr_reclaimed += reclaim_page_list(&node_folio_list, NODE_DATA(nid)); |
2669 | nid = folio_nid(lru_to_folio(folio_list)); | |
2670 | } while (!list_empty(folio_list)); | |
1a4e58cc | 2671 | |
a83f0551 | 2672 | nr_reclaimed += reclaim_page_list(&node_folio_list, NODE_DATA(nid)); |
1a4e58cc | 2673 | |
2d2b8d2b YZ |
2674 | memalloc_noreclaim_restore(noreclaim_flag); |
2675 | ||
1a4e58cc MK |
2676 | return nr_reclaimed; |
2677 | } | |
2678 | ||
b91ac374 JW |
2679 | static unsigned long shrink_list(enum lru_list lru, unsigned long nr_to_scan, |
2680 | struct lruvec *lruvec, struct scan_control *sc) | |
2681 | { | |
2682 | if (is_active_lru(lru)) { | |
2683 | if (sc->may_deactivate & (1 << is_file_lru(lru))) | |
2684 | shrink_active_list(nr_to_scan, lruvec, sc, lru); | |
2685 | else | |
2686 | sc->skipped_deactivate = 1; | |
2687 | return 0; | |
2688 | } | |
2689 | ||
2690 | return shrink_inactive_list(nr_to_scan, lruvec, sc, lru); | |
2691 | } | |
2692 | ||
59dc76b0 RR |
2693 | /* |
2694 | * The inactive anon list should be small enough that the VM never has | |
2695 | * to do too much work. | |
14797e23 | 2696 | * |
59dc76b0 RR |
2697 | * The inactive file list should be small enough to leave most memory |
2698 | * to the established workingset on the scan-resistant active list, | |
2699 | * but large enough to avoid thrashing the aggregate readahead window. | |
56e49d21 | 2700 | * |
59dc76b0 RR |
2701 | * Both inactive lists should also be large enough that each inactive |
2702 | * page has a chance to be referenced again before it is reclaimed. | |
56e49d21 | 2703 | * |
2a2e4885 JW |
2704 | * If that fails and refaulting is observed, the inactive list grows. |
2705 | * | |
59dc76b0 | 2706 | * The inactive_ratio is the target ratio of ACTIVE to INACTIVE pages |
3a50d14d | 2707 | * on this LRU, maintained by the pageout code. An inactive_ratio |
59dc76b0 | 2708 | * of 3 means 3:1 or 25% of the pages are kept on the inactive list. |
56e49d21 | 2709 | * |
59dc76b0 RR |
2710 | * total target max |
2711 | * memory ratio inactive | |
2712 | * ------------------------------------- | |
2713 | * 10MB 1 5MB | |
2714 | * 100MB 1 50MB | |
2715 | * 1GB 3 250MB | |
2716 | * 10GB 10 0.9GB | |
2717 | * 100GB 31 3GB | |
2718 | * 1TB 101 10GB | |
2719 | * 10TB 320 32GB | |
56e49d21 | 2720 | */ |
b91ac374 | 2721 | static bool inactive_is_low(struct lruvec *lruvec, enum lru_list inactive_lru) |
56e49d21 | 2722 | { |
b91ac374 | 2723 | enum lru_list active_lru = inactive_lru + LRU_ACTIVE; |
2a2e4885 JW |
2724 | unsigned long inactive, active; |
2725 | unsigned long inactive_ratio; | |
59dc76b0 | 2726 | unsigned long gb; |
e3790144 | 2727 | |
b91ac374 JW |
2728 | inactive = lruvec_page_state(lruvec, NR_LRU_BASE + inactive_lru); |
2729 | active = lruvec_page_state(lruvec, NR_LRU_BASE + active_lru); | |
f8d1a311 | 2730 | |
b91ac374 | 2731 | gb = (inactive + active) >> (30 - PAGE_SHIFT); |
4002570c | 2732 | if (gb) |
b91ac374 JW |
2733 | inactive_ratio = int_sqrt(10 * gb); |
2734 | else | |
2735 | inactive_ratio = 1; | |
fd538803 | 2736 | |
59dc76b0 | 2737 | return inactive * inactive_ratio < active; |
b39415b2 RR |
2738 | } |
2739 | ||
9a265114 JW |
2740 | enum scan_balance { |
2741 | SCAN_EQUAL, | |
2742 | SCAN_FRACT, | |
2743 | SCAN_ANON, | |
2744 | SCAN_FILE, | |
2745 | }; | |
2746 | ||
f1e1a7be YZ |
2747 | static void prepare_scan_count(pg_data_t *pgdat, struct scan_control *sc) |
2748 | { | |
2749 | unsigned long file; | |
2750 | struct lruvec *target_lruvec; | |
2751 | ||
ac35a490 YZ |
2752 | if (lru_gen_enabled()) |
2753 | return; | |
2754 | ||
f1e1a7be YZ |
2755 | target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat); |
2756 | ||
2757 | /* | |
2758 | * Flush the memory cgroup stats, so that we read accurate per-memcg | |
2759 | * lruvec stats for heuristics. | |
2760 | */ | |
2761 | mem_cgroup_flush_stats(); | |
2762 | ||
2763 | /* | |
2764 | * Determine the scan balance between anon and file LRUs. | |
2765 | */ | |
2766 | spin_lock_irq(&target_lruvec->lru_lock); | |
2767 | sc->anon_cost = target_lruvec->anon_cost; | |
2768 | sc->file_cost = target_lruvec->file_cost; | |
2769 | spin_unlock_irq(&target_lruvec->lru_lock); | |
2770 | ||
2771 | /* | |
2772 | * Target desirable inactive:active list ratios for the anon | |
2773 | * and file LRU lists. | |
2774 | */ | |
2775 | if (!sc->force_deactivate) { | |
2776 | unsigned long refaults; | |
2777 | ||
2778 | /* | |
2779 | * When refaults are being observed, it means a new | |
2780 | * workingset is being established. Deactivate to get | |
2781 | * rid of any stale active pages quickly. | |
2782 | */ | |
2783 | refaults = lruvec_page_state(target_lruvec, | |
2784 | WORKINGSET_ACTIVATE_ANON); | |
2785 | if (refaults != target_lruvec->refaults[WORKINGSET_ANON] || | |
2786 | inactive_is_low(target_lruvec, LRU_INACTIVE_ANON)) | |
2787 | sc->may_deactivate |= DEACTIVATE_ANON; | |
2788 | else | |
2789 | sc->may_deactivate &= ~DEACTIVATE_ANON; | |
2790 | ||
2791 | refaults = lruvec_page_state(target_lruvec, | |
2792 | WORKINGSET_ACTIVATE_FILE); | |
2793 | if (refaults != target_lruvec->refaults[WORKINGSET_FILE] || | |
2794 | inactive_is_low(target_lruvec, LRU_INACTIVE_FILE)) | |
2795 | sc->may_deactivate |= DEACTIVATE_FILE; | |
2796 | else | |
2797 | sc->may_deactivate &= ~DEACTIVATE_FILE; | |
2798 | } else | |
2799 | sc->may_deactivate = DEACTIVATE_ANON | DEACTIVATE_FILE; | |
2800 | ||
2801 | /* | |
2802 | * If we have plenty of inactive file pages that aren't | |
2803 | * thrashing, try to reclaim those first before touching | |
2804 | * anonymous pages. | |
2805 | */ | |
2806 | file = lruvec_page_state(target_lruvec, NR_INACTIVE_FILE); | |
2807 | if (file >> sc->priority && !(sc->may_deactivate & DEACTIVATE_FILE)) | |
2808 | sc->cache_trim_mode = 1; | |
2809 | else | |
2810 | sc->cache_trim_mode = 0; | |
2811 | ||
2812 | /* | |
2813 | * Prevent the reclaimer from falling into the cache trap: as | |
2814 | * cache pages start out inactive, every cache fault will tip | |
2815 | * the scan balance towards the file LRU. And as the file LRU | |
2816 | * shrinks, so does the window for rotation from references. | |
2817 | * This means we have a runaway feedback loop where a tiny | |
2818 | * thrashing file LRU becomes infinitely more attractive than | |
2819 | * anon pages. Try to detect this based on file LRU size. | |
2820 | */ | |
2821 | if (!cgroup_reclaim(sc)) { | |
2822 | unsigned long total_high_wmark = 0; | |
2823 | unsigned long free, anon; | |
2824 | int z; | |
2825 | ||
2826 | free = sum_zone_node_page_state(pgdat->node_id, NR_FREE_PAGES); | |
2827 | file = node_page_state(pgdat, NR_ACTIVE_FILE) + | |
2828 | node_page_state(pgdat, NR_INACTIVE_FILE); | |
2829 | ||
2830 | for (z = 0; z < MAX_NR_ZONES; z++) { | |
2831 | struct zone *zone = &pgdat->node_zones[z]; | |
2832 | ||
2833 | if (!managed_zone(zone)) | |
2834 | continue; | |
2835 | ||
2836 | total_high_wmark += high_wmark_pages(zone); | |
2837 | } | |
2838 | ||
2839 | /* | |
2840 | * Consider anon: if that's low too, this isn't a | |
2841 | * runaway file reclaim problem, but rather just | |
2842 | * extreme pressure. Reclaim as per usual then. | |
2843 | */ | |
2844 | anon = node_page_state(pgdat, NR_INACTIVE_ANON); | |
2845 | ||
2846 | sc->file_is_tiny = | |
2847 | file + free <= total_high_wmark && | |
2848 | !(sc->may_deactivate & DEACTIVATE_ANON) && | |
2849 | anon >> sc->priority; | |
2850 | } | |
2851 | } | |
2852 | ||
4f98a2fe RR |
2853 | /* |
2854 | * Determine how aggressively the anon and file LRU lists should be | |
02e458d8 | 2855 | * scanned. |
4f98a2fe | 2856 | * |
be7bd59d WL |
2857 | * nr[0] = anon inactive pages to scan; nr[1] = anon active pages to scan |
2858 | * nr[2] = file inactive pages to scan; nr[3] = file active pages to scan | |
4f98a2fe | 2859 | */ |
afaf07a6 JW |
2860 | static void get_scan_count(struct lruvec *lruvec, struct scan_control *sc, |
2861 | unsigned long *nr) | |
4f98a2fe | 2862 | { |
a2a36488 | 2863 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); |
afaf07a6 | 2864 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); |
d483a5dd | 2865 | unsigned long anon_cost, file_cost, total_cost; |
33377678 | 2866 | int swappiness = mem_cgroup_swappiness(memcg); |
ed017373 | 2867 | u64 fraction[ANON_AND_FILE]; |
9a265114 | 2868 | u64 denominator = 0; /* gcc */ |
9a265114 | 2869 | enum scan_balance scan_balance; |
4f98a2fe | 2870 | unsigned long ap, fp; |
4111304d | 2871 | enum lru_list lru; |
76a33fc3 SL |
2872 | |
2873 | /* If we have no swap space, do not bother scanning anon pages. */ | |
a2a36488 | 2874 | if (!sc->may_swap || !can_reclaim_anon_pages(memcg, pgdat->node_id, sc)) { |
9a265114 | 2875 | scan_balance = SCAN_FILE; |
76a33fc3 SL |
2876 | goto out; |
2877 | } | |
4f98a2fe | 2878 | |
10316b31 JW |
2879 | /* |
2880 | * Global reclaim will swap to prevent OOM even with no | |
2881 | * swappiness, but memcg users want to use this knob to | |
2882 | * disable swapping for individual groups completely when | |
2883 | * using the memory controller's swap limit feature would be | |
2884 | * too expensive. | |
2885 | */ | |
b5ead35e | 2886 | if (cgroup_reclaim(sc) && !swappiness) { |
9a265114 | 2887 | scan_balance = SCAN_FILE; |
10316b31 JW |
2888 | goto out; |
2889 | } | |
2890 | ||
2891 | /* | |
2892 | * Do not apply any pressure balancing cleverness when the | |
2893 | * system is close to OOM, scan both anon and file equally | |
2894 | * (unless the swappiness setting disagrees with swapping). | |
2895 | */ | |
02695175 | 2896 | if (!sc->priority && swappiness) { |
9a265114 | 2897 | scan_balance = SCAN_EQUAL; |
10316b31 JW |
2898 | goto out; |
2899 | } | |
2900 | ||
62376251 | 2901 | /* |
53138cea | 2902 | * If the system is almost out of file pages, force-scan anon. |
62376251 | 2903 | */ |
b91ac374 | 2904 | if (sc->file_is_tiny) { |
53138cea JW |
2905 | scan_balance = SCAN_ANON; |
2906 | goto out; | |
62376251 JW |
2907 | } |
2908 | ||
7c5bd705 | 2909 | /* |
b91ac374 JW |
2910 | * If there is enough inactive page cache, we do not reclaim |
2911 | * anything from the anonymous working right now. | |
7c5bd705 | 2912 | */ |
b91ac374 | 2913 | if (sc->cache_trim_mode) { |
9a265114 | 2914 | scan_balance = SCAN_FILE; |
7c5bd705 JW |
2915 | goto out; |
2916 | } | |
2917 | ||
9a265114 | 2918 | scan_balance = SCAN_FRACT; |
58c37f6e | 2919 | /* |
314b57fb JW |
2920 | * Calculate the pressure balance between anon and file pages. |
2921 | * | |
2922 | * The amount of pressure we put on each LRU is inversely | |
2923 | * proportional to the cost of reclaiming each list, as | |
2924 | * determined by the share of pages that are refaulting, times | |
2925 | * the relative IO cost of bringing back a swapped out | |
2926 | * anonymous page vs reloading a filesystem page (swappiness). | |
2927 | * | |
d483a5dd JW |
2928 | * Although we limit that influence to ensure no list gets |
2929 | * left behind completely: at least a third of the pressure is | |
2930 | * applied, before swappiness. | |
2931 | * | |
314b57fb | 2932 | * With swappiness at 100, anon and file have equal IO cost. |
58c37f6e | 2933 | */ |
d483a5dd JW |
2934 | total_cost = sc->anon_cost + sc->file_cost; |
2935 | anon_cost = total_cost + sc->anon_cost; | |
2936 | file_cost = total_cost + sc->file_cost; | |
2937 | total_cost = anon_cost + file_cost; | |
58c37f6e | 2938 | |
d483a5dd JW |
2939 | ap = swappiness * (total_cost + 1); |
2940 | ap /= anon_cost + 1; | |
4f98a2fe | 2941 | |
d483a5dd JW |
2942 | fp = (200 - swappiness) * (total_cost + 1); |
2943 | fp /= file_cost + 1; | |
4f98a2fe | 2944 | |
76a33fc3 SL |
2945 | fraction[0] = ap; |
2946 | fraction[1] = fp; | |
a4fe1631 | 2947 | denominator = ap + fp; |
76a33fc3 | 2948 | out: |
688035f7 JW |
2949 | for_each_evictable_lru(lru) { |
2950 | int file = is_file_lru(lru); | |
9783aa99 | 2951 | unsigned long lruvec_size; |
f56ce412 | 2952 | unsigned long low, min; |
688035f7 | 2953 | unsigned long scan; |
9783aa99 CD |
2954 | |
2955 | lruvec_size = lruvec_lru_size(lruvec, lru, sc->reclaim_idx); | |
f56ce412 JW |
2956 | mem_cgroup_protection(sc->target_mem_cgroup, memcg, |
2957 | &min, &low); | |
9783aa99 | 2958 | |
f56ce412 | 2959 | if (min || low) { |
9783aa99 CD |
2960 | /* |
2961 | * Scale a cgroup's reclaim pressure by proportioning | |
2962 | * its current usage to its memory.low or memory.min | |
2963 | * setting. | |
2964 | * | |
2965 | * This is important, as otherwise scanning aggression | |
2966 | * becomes extremely binary -- from nothing as we | |
2967 | * approach the memory protection threshold, to totally | |
2968 | * nominal as we exceed it. This results in requiring | |
2969 | * setting extremely liberal protection thresholds. It | |
2970 | * also means we simply get no protection at all if we | |
2971 | * set it too low, which is not ideal. | |
1bc63fb1 CD |
2972 | * |
2973 | * If there is any protection in place, we reduce scan | |
2974 | * pressure by how much of the total memory used is | |
2975 | * within protection thresholds. | |
9783aa99 | 2976 | * |
9de7ca46 CD |
2977 | * There is one special case: in the first reclaim pass, |
2978 | * we skip over all groups that are within their low | |
2979 | * protection. If that fails to reclaim enough pages to | |
2980 | * satisfy the reclaim goal, we come back and override | |
2981 | * the best-effort low protection. However, we still | |
2982 | * ideally want to honor how well-behaved groups are in | |
2983 | * that case instead of simply punishing them all | |
2984 | * equally. As such, we reclaim them based on how much | |
1bc63fb1 CD |
2985 | * memory they are using, reducing the scan pressure |
2986 | * again by how much of the total memory used is under | |
2987 | * hard protection. | |
9783aa99 | 2988 | */ |
1bc63fb1 | 2989 | unsigned long cgroup_size = mem_cgroup_size(memcg); |
f56ce412 JW |
2990 | unsigned long protection; |
2991 | ||
2992 | /* memory.low scaling, make sure we retry before OOM */ | |
2993 | if (!sc->memcg_low_reclaim && low > min) { | |
2994 | protection = low; | |
2995 | sc->memcg_low_skipped = 1; | |
2996 | } else { | |
2997 | protection = min; | |
2998 | } | |
1bc63fb1 CD |
2999 | |
3000 | /* Avoid TOCTOU with earlier protection check */ | |
3001 | cgroup_size = max(cgroup_size, protection); | |
3002 | ||
3003 | scan = lruvec_size - lruvec_size * protection / | |
32d4f4b7 | 3004 | (cgroup_size + 1); |
9783aa99 CD |
3005 | |
3006 | /* | |
1bc63fb1 | 3007 | * Minimally target SWAP_CLUSTER_MAX pages to keep |
55b65a57 | 3008 | * reclaim moving forwards, avoiding decrementing |
9de7ca46 | 3009 | * sc->priority further than desirable. |
9783aa99 | 3010 | */ |
1bc63fb1 | 3011 | scan = max(scan, SWAP_CLUSTER_MAX); |
9783aa99 CD |
3012 | } else { |
3013 | scan = lruvec_size; | |
3014 | } | |
3015 | ||
3016 | scan >>= sc->priority; | |
6b4f7799 | 3017 | |
688035f7 JW |
3018 | /* |
3019 | * If the cgroup's already been deleted, make sure to | |
3020 | * scrape out the remaining cache. | |
3021 | */ | |
3022 | if (!scan && !mem_cgroup_online(memcg)) | |
9783aa99 | 3023 | scan = min(lruvec_size, SWAP_CLUSTER_MAX); |
6b4f7799 | 3024 | |
688035f7 JW |
3025 | switch (scan_balance) { |
3026 | case SCAN_EQUAL: | |
3027 | /* Scan lists relative to size */ | |
3028 | break; | |
3029 | case SCAN_FRACT: | |
9a265114 | 3030 | /* |
688035f7 JW |
3031 | * Scan types proportional to swappiness and |
3032 | * their relative recent reclaim efficiency. | |
76073c64 GS |
3033 | * Make sure we don't miss the last page on |
3034 | * the offlined memory cgroups because of a | |
3035 | * round-off error. | |
9a265114 | 3036 | */ |
76073c64 GS |
3037 | scan = mem_cgroup_online(memcg) ? |
3038 | div64_u64(scan * fraction[file], denominator) : | |
3039 | DIV64_U64_ROUND_UP(scan * fraction[file], | |
68600f62 | 3040 | denominator); |
688035f7 JW |
3041 | break; |
3042 | case SCAN_FILE: | |
3043 | case SCAN_ANON: | |
3044 | /* Scan one type exclusively */ | |
e072bff6 | 3045 | if ((scan_balance == SCAN_FILE) != file) |
688035f7 | 3046 | scan = 0; |
688035f7 JW |
3047 | break; |
3048 | default: | |
3049 | /* Look ma, no brain */ | |
3050 | BUG(); | |
9a265114 | 3051 | } |
688035f7 | 3052 | |
688035f7 | 3053 | nr[lru] = scan; |
76a33fc3 | 3054 | } |
6e08a369 | 3055 | } |
4f98a2fe | 3056 | |
2f368a9f DH |
3057 | /* |
3058 | * Anonymous LRU management is a waste if there is | |
3059 | * ultimately no way to reclaim the memory. | |
3060 | */ | |
3061 | static bool can_age_anon_pages(struct pglist_data *pgdat, | |
3062 | struct scan_control *sc) | |
3063 | { | |
3064 | /* Aging the anon LRU is valuable if swap is present: */ | |
3065 | if (total_swap_pages > 0) | |
3066 | return true; | |
3067 | ||
3068 | /* Also valuable if anon pages can be demoted: */ | |
3069 | return can_demote(pgdat->node_id, sc); | |
3070 | } | |
3071 | ||
ec1c86b2 YZ |
3072 | #ifdef CONFIG_LRU_GEN |
3073 | ||
354ed597 YZ |
3074 | #ifdef CONFIG_LRU_GEN_ENABLED |
3075 | DEFINE_STATIC_KEY_ARRAY_TRUE(lru_gen_caps, NR_LRU_GEN_CAPS); | |
3076 | #define get_cap(cap) static_branch_likely(&lru_gen_caps[cap]) | |
3077 | #else | |
3078 | DEFINE_STATIC_KEY_ARRAY_FALSE(lru_gen_caps, NR_LRU_GEN_CAPS); | |
3079 | #define get_cap(cap) static_branch_unlikely(&lru_gen_caps[cap]) | |
3080 | #endif | |
3081 | ||
ec1c86b2 YZ |
3082 | /****************************************************************************** |
3083 | * shorthand helpers | |
3084 | ******************************************************************************/ | |
3085 | ||
ac35a490 YZ |
3086 | #define LRU_REFS_FLAGS (BIT(PG_referenced) | BIT(PG_workingset)) |
3087 | ||
3088 | #define DEFINE_MAX_SEQ(lruvec) \ | |
3089 | unsigned long max_seq = READ_ONCE((lruvec)->lrugen.max_seq) | |
3090 | ||
3091 | #define DEFINE_MIN_SEQ(lruvec) \ | |
3092 | unsigned long min_seq[ANON_AND_FILE] = { \ | |
3093 | READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_ANON]), \ | |
3094 | READ_ONCE((lruvec)->lrugen.min_seq[LRU_GEN_FILE]), \ | |
3095 | } | |
3096 | ||
ec1c86b2 YZ |
3097 | #define for_each_gen_type_zone(gen, type, zone) \ |
3098 | for ((gen) = 0; (gen) < MAX_NR_GENS; (gen)++) \ | |
3099 | for ((type) = 0; (type) < ANON_AND_FILE; (type)++) \ | |
3100 | for ((zone) = 0; (zone) < MAX_NR_ZONES; (zone)++) | |
3101 | ||
bd74fdae | 3102 | static struct lruvec *get_lruvec(struct mem_cgroup *memcg, int nid) |
ec1c86b2 YZ |
3103 | { |
3104 | struct pglist_data *pgdat = NODE_DATA(nid); | |
3105 | ||
3106 | #ifdef CONFIG_MEMCG | |
3107 | if (memcg) { | |
3108 | struct lruvec *lruvec = &memcg->nodeinfo[nid]->lruvec; | |
3109 | ||
3110 | /* for hotadd_new_pgdat() */ | |
3111 | if (!lruvec->pgdat) | |
3112 | lruvec->pgdat = pgdat; | |
3113 | ||
3114 | return lruvec; | |
3115 | } | |
3116 | #endif | |
3117 | VM_WARN_ON_ONCE(!mem_cgroup_disabled()); | |
3118 | ||
3119 | return pgdat ? &pgdat->__lruvec : NULL; | |
3120 | } | |
3121 | ||
ac35a490 YZ |
3122 | static int get_swappiness(struct lruvec *lruvec, struct scan_control *sc) |
3123 | { | |
3124 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
3125 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); | |
3126 | ||
3127 | if (!can_demote(pgdat->node_id, sc) && | |
3128 | mem_cgroup_get_nr_swap_pages(memcg) < MIN_LRU_BATCH) | |
3129 | return 0; | |
3130 | ||
3131 | return mem_cgroup_swappiness(memcg); | |
3132 | } | |
3133 | ||
3134 | static int get_nr_gens(struct lruvec *lruvec, int type) | |
3135 | { | |
3136 | return lruvec->lrugen.max_seq - lruvec->lrugen.min_seq[type] + 1; | |
3137 | } | |
3138 | ||
3139 | static bool __maybe_unused seq_is_valid(struct lruvec *lruvec) | |
3140 | { | |
3141 | /* see the comment on lru_gen_struct */ | |
3142 | return get_nr_gens(lruvec, LRU_GEN_FILE) >= MIN_NR_GENS && | |
3143 | get_nr_gens(lruvec, LRU_GEN_FILE) <= get_nr_gens(lruvec, LRU_GEN_ANON) && | |
3144 | get_nr_gens(lruvec, LRU_GEN_ANON) <= MAX_NR_GENS; | |
3145 | } | |
3146 | ||
bd74fdae YZ |
3147 | /****************************************************************************** |
3148 | * mm_struct list | |
3149 | ******************************************************************************/ | |
3150 | ||
3151 | static struct lru_gen_mm_list *get_mm_list(struct mem_cgroup *memcg) | |
3152 | { | |
3153 | static struct lru_gen_mm_list mm_list = { | |
3154 | .fifo = LIST_HEAD_INIT(mm_list.fifo), | |
3155 | .lock = __SPIN_LOCK_UNLOCKED(mm_list.lock), | |
3156 | }; | |
3157 | ||
3158 | #ifdef CONFIG_MEMCG | |
3159 | if (memcg) | |
3160 | return &memcg->mm_list; | |
3161 | #endif | |
3162 | VM_WARN_ON_ONCE(!mem_cgroup_disabled()); | |
3163 | ||
3164 | return &mm_list; | |
3165 | } | |
3166 | ||
3167 | void lru_gen_add_mm(struct mm_struct *mm) | |
3168 | { | |
3169 | int nid; | |
3170 | struct mem_cgroup *memcg = get_mem_cgroup_from_mm(mm); | |
3171 | struct lru_gen_mm_list *mm_list = get_mm_list(memcg); | |
3172 | ||
3173 | VM_WARN_ON_ONCE(!list_empty(&mm->lru_gen.list)); | |
3174 | #ifdef CONFIG_MEMCG | |
3175 | VM_WARN_ON_ONCE(mm->lru_gen.memcg); | |
3176 | mm->lru_gen.memcg = memcg; | |
3177 | #endif | |
3178 | spin_lock(&mm_list->lock); | |
3179 | ||
3180 | for_each_node_state(nid, N_MEMORY) { | |
3181 | struct lruvec *lruvec = get_lruvec(memcg, nid); | |
3182 | ||
3183 | if (!lruvec) | |
3184 | continue; | |
3185 | ||
3186 | /* the first addition since the last iteration */ | |
3187 | if (lruvec->mm_state.tail == &mm_list->fifo) | |
3188 | lruvec->mm_state.tail = &mm->lru_gen.list; | |
3189 | } | |
3190 | ||
3191 | list_add_tail(&mm->lru_gen.list, &mm_list->fifo); | |
3192 | ||
3193 | spin_unlock(&mm_list->lock); | |
3194 | } | |
3195 | ||
3196 | void lru_gen_del_mm(struct mm_struct *mm) | |
3197 | { | |
3198 | int nid; | |
3199 | struct lru_gen_mm_list *mm_list; | |
3200 | struct mem_cgroup *memcg = NULL; | |
3201 | ||
3202 | if (list_empty(&mm->lru_gen.list)) | |
3203 | return; | |
3204 | ||
3205 | #ifdef CONFIG_MEMCG | |
3206 | memcg = mm->lru_gen.memcg; | |
3207 | #endif | |
3208 | mm_list = get_mm_list(memcg); | |
3209 | ||
3210 | spin_lock(&mm_list->lock); | |
3211 | ||
3212 | for_each_node(nid) { | |
3213 | struct lruvec *lruvec = get_lruvec(memcg, nid); | |
3214 | ||
3215 | if (!lruvec) | |
3216 | continue; | |
3217 | ||
3218 | /* where the last iteration ended (exclusive) */ | |
3219 | if (lruvec->mm_state.tail == &mm->lru_gen.list) | |
3220 | lruvec->mm_state.tail = lruvec->mm_state.tail->next; | |
3221 | ||
3222 | /* where the current iteration continues (inclusive) */ | |
3223 | if (lruvec->mm_state.head != &mm->lru_gen.list) | |
3224 | continue; | |
3225 | ||
3226 | lruvec->mm_state.head = lruvec->mm_state.head->next; | |
3227 | /* the deletion ends the current iteration */ | |
3228 | if (lruvec->mm_state.head == &mm_list->fifo) | |
3229 | WRITE_ONCE(lruvec->mm_state.seq, lruvec->mm_state.seq + 1); | |
3230 | } | |
3231 | ||
3232 | list_del_init(&mm->lru_gen.list); | |
3233 | ||
3234 | spin_unlock(&mm_list->lock); | |
3235 | ||
3236 | #ifdef CONFIG_MEMCG | |
3237 | mem_cgroup_put(mm->lru_gen.memcg); | |
3238 | mm->lru_gen.memcg = NULL; | |
3239 | #endif | |
3240 | } | |
3241 | ||
3242 | #ifdef CONFIG_MEMCG | |
3243 | void lru_gen_migrate_mm(struct mm_struct *mm) | |
3244 | { | |
3245 | struct mem_cgroup *memcg; | |
3246 | struct task_struct *task = rcu_dereference_protected(mm->owner, true); | |
3247 | ||
3248 | VM_WARN_ON_ONCE(task->mm != mm); | |
3249 | lockdep_assert_held(&task->alloc_lock); | |
3250 | ||
3251 | /* for mm_update_next_owner() */ | |
3252 | if (mem_cgroup_disabled()) | |
3253 | return; | |
3254 | ||
3255 | rcu_read_lock(); | |
3256 | memcg = mem_cgroup_from_task(task); | |
3257 | rcu_read_unlock(); | |
3258 | if (memcg == mm->lru_gen.memcg) | |
3259 | return; | |
3260 | ||
3261 | VM_WARN_ON_ONCE(!mm->lru_gen.memcg); | |
3262 | VM_WARN_ON_ONCE(list_empty(&mm->lru_gen.list)); | |
3263 | ||
3264 | lru_gen_del_mm(mm); | |
3265 | lru_gen_add_mm(mm); | |
3266 | } | |
3267 | #endif | |
3268 | ||
3269 | /* | |
3270 | * Bloom filters with m=1<<15, k=2 and the false positive rates of ~1/5 when | |
3271 | * n=10,000 and ~1/2 when n=20,000, where, conventionally, m is the number of | |
3272 | * bits in a bitmap, k is the number of hash functions and n is the number of | |
3273 | * inserted items. | |
3274 | * | |
3275 | * Page table walkers use one of the two filters to reduce their search space. | |
3276 | * To get rid of non-leaf entries that no longer have enough leaf entries, the | |
3277 | * aging uses the double-buffering technique to flip to the other filter each | |
3278 | * time it produces a new generation. For non-leaf entries that have enough | |
3279 | * leaf entries, the aging carries them over to the next generation in | |
3280 | * walk_pmd_range(); the eviction also report them when walking the rmap | |
3281 | * in lru_gen_look_around(). | |
3282 | * | |
3283 | * For future optimizations: | |
3284 | * 1. It's not necessary to keep both filters all the time. The spare one can be | |
3285 | * freed after the RCU grace period and reallocated if needed again. | |
3286 | * 2. And when reallocating, it's worth scaling its size according to the number | |
3287 | * of inserted entries in the other filter, to reduce the memory overhead on | |
3288 | * small systems and false positives on large systems. | |
3289 | * 3. Jenkins' hash function is an alternative to Knuth's. | |
3290 | */ | |
3291 | #define BLOOM_FILTER_SHIFT 15 | |
3292 | ||
3293 | static inline int filter_gen_from_seq(unsigned long seq) | |
3294 | { | |
3295 | return seq % NR_BLOOM_FILTERS; | |
3296 | } | |
3297 | ||
3298 | static void get_item_key(void *item, int *key) | |
3299 | { | |
3300 | u32 hash = hash_ptr(item, BLOOM_FILTER_SHIFT * 2); | |
3301 | ||
3302 | BUILD_BUG_ON(BLOOM_FILTER_SHIFT * 2 > BITS_PER_TYPE(u32)); | |
3303 | ||
3304 | key[0] = hash & (BIT(BLOOM_FILTER_SHIFT) - 1); | |
3305 | key[1] = hash >> BLOOM_FILTER_SHIFT; | |
3306 | } | |
3307 | ||
3308 | static void reset_bloom_filter(struct lruvec *lruvec, unsigned long seq) | |
3309 | { | |
3310 | unsigned long *filter; | |
3311 | int gen = filter_gen_from_seq(seq); | |
3312 | ||
3313 | filter = lruvec->mm_state.filters[gen]; | |
3314 | if (filter) { | |
3315 | bitmap_clear(filter, 0, BIT(BLOOM_FILTER_SHIFT)); | |
3316 | return; | |
3317 | } | |
3318 | ||
3319 | filter = bitmap_zalloc(BIT(BLOOM_FILTER_SHIFT), | |
3320 | __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN); | |
3321 | WRITE_ONCE(lruvec->mm_state.filters[gen], filter); | |
3322 | } | |
3323 | ||
3324 | static void update_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item) | |
3325 | { | |
3326 | int key[2]; | |
3327 | unsigned long *filter; | |
3328 | int gen = filter_gen_from_seq(seq); | |
3329 | ||
3330 | filter = READ_ONCE(lruvec->mm_state.filters[gen]); | |
3331 | if (!filter) | |
3332 | return; | |
3333 | ||
3334 | get_item_key(item, key); | |
3335 | ||
3336 | if (!test_bit(key[0], filter)) | |
3337 | set_bit(key[0], filter); | |
3338 | if (!test_bit(key[1], filter)) | |
3339 | set_bit(key[1], filter); | |
3340 | } | |
3341 | ||
3342 | static bool test_bloom_filter(struct lruvec *lruvec, unsigned long seq, void *item) | |
3343 | { | |
3344 | int key[2]; | |
3345 | unsigned long *filter; | |
3346 | int gen = filter_gen_from_seq(seq); | |
3347 | ||
3348 | filter = READ_ONCE(lruvec->mm_state.filters[gen]); | |
3349 | if (!filter) | |
3350 | return true; | |
3351 | ||
3352 | get_item_key(item, key); | |
3353 | ||
3354 | return test_bit(key[0], filter) && test_bit(key[1], filter); | |
3355 | } | |
3356 | ||
3357 | static void reset_mm_stats(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, bool last) | |
3358 | { | |
3359 | int i; | |
3360 | int hist; | |
3361 | ||
3362 | lockdep_assert_held(&get_mm_list(lruvec_memcg(lruvec))->lock); | |
3363 | ||
3364 | if (walk) { | |
3365 | hist = lru_hist_from_seq(walk->max_seq); | |
3366 | ||
3367 | for (i = 0; i < NR_MM_STATS; i++) { | |
3368 | WRITE_ONCE(lruvec->mm_state.stats[hist][i], | |
3369 | lruvec->mm_state.stats[hist][i] + walk->mm_stats[i]); | |
3370 | walk->mm_stats[i] = 0; | |
3371 | } | |
3372 | } | |
3373 | ||
3374 | if (NR_HIST_GENS > 1 && last) { | |
3375 | hist = lru_hist_from_seq(lruvec->mm_state.seq + 1); | |
3376 | ||
3377 | for (i = 0; i < NR_MM_STATS; i++) | |
3378 | WRITE_ONCE(lruvec->mm_state.stats[hist][i], 0); | |
3379 | } | |
3380 | } | |
3381 | ||
3382 | static bool should_skip_mm(struct mm_struct *mm, struct lru_gen_mm_walk *walk) | |
3383 | { | |
3384 | int type; | |
3385 | unsigned long size = 0; | |
3386 | struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec); | |
3387 | int key = pgdat->node_id % BITS_PER_TYPE(mm->lru_gen.bitmap); | |
3388 | ||
3389 | if (!walk->force_scan && !test_bit(key, &mm->lru_gen.bitmap)) | |
3390 | return true; | |
3391 | ||
3392 | clear_bit(key, &mm->lru_gen.bitmap); | |
3393 | ||
3394 | for (type = !walk->can_swap; type < ANON_AND_FILE; type++) { | |
3395 | size += type ? get_mm_counter(mm, MM_FILEPAGES) : | |
3396 | get_mm_counter(mm, MM_ANONPAGES) + | |
3397 | get_mm_counter(mm, MM_SHMEMPAGES); | |
3398 | } | |
3399 | ||
3400 | if (size < MIN_LRU_BATCH) | |
3401 | return true; | |
3402 | ||
3403 | return !mmget_not_zero(mm); | |
3404 | } | |
3405 | ||
3406 | static bool iterate_mm_list(struct lruvec *lruvec, struct lru_gen_mm_walk *walk, | |
3407 | struct mm_struct **iter) | |
3408 | { | |
3409 | bool first = false; | |
3410 | bool last = true; | |
3411 | struct mm_struct *mm = NULL; | |
3412 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
3413 | struct lru_gen_mm_list *mm_list = get_mm_list(memcg); | |
3414 | struct lru_gen_mm_state *mm_state = &lruvec->mm_state; | |
3415 | ||
3416 | /* | |
3417 | * There are four interesting cases for this page table walker: | |
3418 | * 1. It tries to start a new iteration of mm_list with a stale max_seq; | |
3419 | * there is nothing left to do. | |
3420 | * 2. It's the first of the current generation, and it needs to reset | |
3421 | * the Bloom filter for the next generation. | |
3422 | * 3. It reaches the end of mm_list, and it needs to increment | |
3423 | * mm_state->seq; the iteration is done. | |
3424 | * 4. It's the last of the current generation, and it needs to reset the | |
3425 | * mm stats counters for the next generation. | |
3426 | */ | |
3427 | spin_lock(&mm_list->lock); | |
3428 | ||
3429 | VM_WARN_ON_ONCE(mm_state->seq + 1 < walk->max_seq); | |
3430 | VM_WARN_ON_ONCE(*iter && mm_state->seq > walk->max_seq); | |
3431 | VM_WARN_ON_ONCE(*iter && !mm_state->nr_walkers); | |
3432 | ||
3433 | if (walk->max_seq <= mm_state->seq) { | |
3434 | if (!*iter) | |
3435 | last = false; | |
3436 | goto done; | |
3437 | } | |
3438 | ||
3439 | if (!mm_state->nr_walkers) { | |
3440 | VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo); | |
3441 | ||
3442 | mm_state->head = mm_list->fifo.next; | |
3443 | first = true; | |
3444 | } | |
3445 | ||
3446 | while (!mm && mm_state->head != &mm_list->fifo) { | |
3447 | mm = list_entry(mm_state->head, struct mm_struct, lru_gen.list); | |
3448 | ||
3449 | mm_state->head = mm_state->head->next; | |
3450 | ||
3451 | /* force scan for those added after the last iteration */ | |
3452 | if (!mm_state->tail || mm_state->tail == &mm->lru_gen.list) { | |
3453 | mm_state->tail = mm_state->head; | |
3454 | walk->force_scan = true; | |
3455 | } | |
3456 | ||
3457 | if (should_skip_mm(mm, walk)) | |
3458 | mm = NULL; | |
3459 | } | |
3460 | ||
3461 | if (mm_state->head == &mm_list->fifo) | |
3462 | WRITE_ONCE(mm_state->seq, mm_state->seq + 1); | |
3463 | done: | |
3464 | if (*iter && !mm) | |
3465 | mm_state->nr_walkers--; | |
3466 | if (!*iter && mm) | |
3467 | mm_state->nr_walkers++; | |
3468 | ||
3469 | if (mm_state->nr_walkers) | |
3470 | last = false; | |
3471 | ||
3472 | if (*iter || last) | |
3473 | reset_mm_stats(lruvec, walk, last); | |
3474 | ||
3475 | spin_unlock(&mm_list->lock); | |
3476 | ||
3477 | if (mm && first) | |
3478 | reset_bloom_filter(lruvec, walk->max_seq + 1); | |
3479 | ||
3480 | if (*iter) | |
3481 | mmput_async(*iter); | |
3482 | ||
3483 | *iter = mm; | |
3484 | ||
3485 | return last; | |
3486 | } | |
3487 | ||
3488 | static bool iterate_mm_list_nowalk(struct lruvec *lruvec, unsigned long max_seq) | |
3489 | { | |
3490 | bool success = false; | |
3491 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
3492 | struct lru_gen_mm_list *mm_list = get_mm_list(memcg); | |
3493 | struct lru_gen_mm_state *mm_state = &lruvec->mm_state; | |
3494 | ||
3495 | spin_lock(&mm_list->lock); | |
3496 | ||
3497 | VM_WARN_ON_ONCE(mm_state->seq + 1 < max_seq); | |
3498 | ||
3499 | if (max_seq > mm_state->seq && !mm_state->nr_walkers) { | |
3500 | VM_WARN_ON_ONCE(mm_state->head && mm_state->head != &mm_list->fifo); | |
3501 | ||
3502 | WRITE_ONCE(mm_state->seq, mm_state->seq + 1); | |
3503 | reset_mm_stats(lruvec, NULL, true); | |
3504 | success = true; | |
3505 | } | |
3506 | ||
3507 | spin_unlock(&mm_list->lock); | |
3508 | ||
3509 | return success; | |
3510 | } | |
3511 | ||
ac35a490 YZ |
3512 | /****************************************************************************** |
3513 | * refault feedback loop | |
3514 | ******************************************************************************/ | |
3515 | ||
3516 | /* | |
3517 | * A feedback loop based on Proportional-Integral-Derivative (PID) controller. | |
3518 | * | |
3519 | * The P term is refaulted/(evicted+protected) from a tier in the generation | |
3520 | * currently being evicted; the I term is the exponential moving average of the | |
3521 | * P term over the generations previously evicted, using the smoothing factor | |
3522 | * 1/2; the D term isn't supported. | |
3523 | * | |
3524 | * The setpoint (SP) is always the first tier of one type; the process variable | |
3525 | * (PV) is either any tier of the other type or any other tier of the same | |
3526 | * type. | |
3527 | * | |
3528 | * The error is the difference between the SP and the PV; the correction is to | |
3529 | * turn off protection when SP>PV or turn on protection when SP<PV. | |
3530 | * | |
3531 | * For future optimizations: | |
3532 | * 1. The D term may discount the other two terms over time so that long-lived | |
3533 | * generations can resist stale information. | |
3534 | */ | |
3535 | struct ctrl_pos { | |
3536 | unsigned long refaulted; | |
3537 | unsigned long total; | |
3538 | int gain; | |
3539 | }; | |
3540 | ||
3541 | static void read_ctrl_pos(struct lruvec *lruvec, int type, int tier, int gain, | |
3542 | struct ctrl_pos *pos) | |
3543 | { | |
3544 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
3545 | int hist = lru_hist_from_seq(lrugen->min_seq[type]); | |
3546 | ||
3547 | pos->refaulted = lrugen->avg_refaulted[type][tier] + | |
3548 | atomic_long_read(&lrugen->refaulted[hist][type][tier]); | |
3549 | pos->total = lrugen->avg_total[type][tier] + | |
3550 | atomic_long_read(&lrugen->evicted[hist][type][tier]); | |
3551 | if (tier) | |
3552 | pos->total += lrugen->protected[hist][type][tier - 1]; | |
3553 | pos->gain = gain; | |
3554 | } | |
3555 | ||
3556 | static void reset_ctrl_pos(struct lruvec *lruvec, int type, bool carryover) | |
3557 | { | |
3558 | int hist, tier; | |
3559 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
3560 | bool clear = carryover ? NR_HIST_GENS == 1 : NR_HIST_GENS > 1; | |
3561 | unsigned long seq = carryover ? lrugen->min_seq[type] : lrugen->max_seq + 1; | |
3562 | ||
3563 | lockdep_assert_held(&lruvec->lru_lock); | |
3564 | ||
3565 | if (!carryover && !clear) | |
3566 | return; | |
3567 | ||
3568 | hist = lru_hist_from_seq(seq); | |
3569 | ||
3570 | for (tier = 0; tier < MAX_NR_TIERS; tier++) { | |
3571 | if (carryover) { | |
3572 | unsigned long sum; | |
3573 | ||
3574 | sum = lrugen->avg_refaulted[type][tier] + | |
3575 | atomic_long_read(&lrugen->refaulted[hist][type][tier]); | |
3576 | WRITE_ONCE(lrugen->avg_refaulted[type][tier], sum / 2); | |
3577 | ||
3578 | sum = lrugen->avg_total[type][tier] + | |
3579 | atomic_long_read(&lrugen->evicted[hist][type][tier]); | |
3580 | if (tier) | |
3581 | sum += lrugen->protected[hist][type][tier - 1]; | |
3582 | WRITE_ONCE(lrugen->avg_total[type][tier], sum / 2); | |
3583 | } | |
3584 | ||
3585 | if (clear) { | |
3586 | atomic_long_set(&lrugen->refaulted[hist][type][tier], 0); | |
3587 | atomic_long_set(&lrugen->evicted[hist][type][tier], 0); | |
3588 | if (tier) | |
3589 | WRITE_ONCE(lrugen->protected[hist][type][tier - 1], 0); | |
3590 | } | |
3591 | } | |
3592 | } | |
3593 | ||
3594 | static bool positive_ctrl_err(struct ctrl_pos *sp, struct ctrl_pos *pv) | |
3595 | { | |
3596 | /* | |
3597 | * Return true if the PV has a limited number of refaults or a lower | |
3598 | * refaulted/total than the SP. | |
3599 | */ | |
3600 | return pv->refaulted < MIN_LRU_BATCH || | |
3601 | pv->refaulted * (sp->total + MIN_LRU_BATCH) * sp->gain <= | |
3602 | (sp->refaulted + 1) * pv->total * pv->gain; | |
3603 | } | |
3604 | ||
3605 | /****************************************************************************** | |
3606 | * the aging | |
3607 | ******************************************************************************/ | |
3608 | ||
018ee47f YZ |
3609 | /* promote pages accessed through page tables */ |
3610 | static int folio_update_gen(struct folio *folio, int gen) | |
3611 | { | |
3612 | unsigned long new_flags, old_flags = READ_ONCE(folio->flags); | |
3613 | ||
3614 | VM_WARN_ON_ONCE(gen >= MAX_NR_GENS); | |
3615 | VM_WARN_ON_ONCE(!rcu_read_lock_held()); | |
3616 | ||
3617 | do { | |
3618 | /* lru_gen_del_folio() has isolated this page? */ | |
3619 | if (!(old_flags & LRU_GEN_MASK)) { | |
3620 | /* for shrink_page_list() */ | |
3621 | new_flags = old_flags | BIT(PG_referenced); | |
3622 | continue; | |
3623 | } | |
3624 | ||
3625 | new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS); | |
3626 | new_flags |= (gen + 1UL) << LRU_GEN_PGOFF; | |
3627 | } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags)); | |
3628 | ||
3629 | return ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1; | |
3630 | } | |
3631 | ||
ac35a490 YZ |
3632 | /* protect pages accessed multiple times through file descriptors */ |
3633 | static int folio_inc_gen(struct lruvec *lruvec, struct folio *folio, bool reclaiming) | |
3634 | { | |
3635 | int type = folio_is_file_lru(folio); | |
3636 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
3637 | int new_gen, old_gen = lru_gen_from_seq(lrugen->min_seq[type]); | |
3638 | unsigned long new_flags, old_flags = READ_ONCE(folio->flags); | |
3639 | ||
3640 | VM_WARN_ON_ONCE_FOLIO(!(old_flags & LRU_GEN_MASK), folio); | |
3641 | ||
3642 | do { | |
018ee47f YZ |
3643 | new_gen = ((old_flags & LRU_GEN_MASK) >> LRU_GEN_PGOFF) - 1; |
3644 | /* folio_update_gen() has promoted this page? */ | |
3645 | if (new_gen >= 0 && new_gen != old_gen) | |
3646 | return new_gen; | |
3647 | ||
ac35a490 YZ |
3648 | new_gen = (old_gen + 1) % MAX_NR_GENS; |
3649 | ||
3650 | new_flags = old_flags & ~(LRU_GEN_MASK | LRU_REFS_MASK | LRU_REFS_FLAGS); | |
3651 | new_flags |= (new_gen + 1UL) << LRU_GEN_PGOFF; | |
3652 | /* for folio_end_writeback() */ | |
3653 | if (reclaiming) | |
3654 | new_flags |= BIT(PG_reclaim); | |
3655 | } while (!try_cmpxchg(&folio->flags, &old_flags, new_flags)); | |
3656 | ||
3657 | lru_gen_update_size(lruvec, folio, old_gen, new_gen); | |
3658 | ||
3659 | return new_gen; | |
3660 | } | |
3661 | ||
bd74fdae YZ |
3662 | static void update_batch_size(struct lru_gen_mm_walk *walk, struct folio *folio, |
3663 | int old_gen, int new_gen) | |
3664 | { | |
3665 | int type = folio_is_file_lru(folio); | |
3666 | int zone = folio_zonenum(folio); | |
3667 | int delta = folio_nr_pages(folio); | |
3668 | ||
3669 | VM_WARN_ON_ONCE(old_gen >= MAX_NR_GENS); | |
3670 | VM_WARN_ON_ONCE(new_gen >= MAX_NR_GENS); | |
3671 | ||
3672 | walk->batched++; | |
3673 | ||
3674 | walk->nr_pages[old_gen][type][zone] -= delta; | |
3675 | walk->nr_pages[new_gen][type][zone] += delta; | |
3676 | } | |
3677 | ||
3678 | static void reset_batch_size(struct lruvec *lruvec, struct lru_gen_mm_walk *walk) | |
3679 | { | |
3680 | int gen, type, zone; | |
3681 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
3682 | ||
3683 | walk->batched = 0; | |
3684 | ||
3685 | for_each_gen_type_zone(gen, type, zone) { | |
3686 | enum lru_list lru = type * LRU_INACTIVE_FILE; | |
3687 | int delta = walk->nr_pages[gen][type][zone]; | |
3688 | ||
3689 | if (!delta) | |
3690 | continue; | |
3691 | ||
3692 | walk->nr_pages[gen][type][zone] = 0; | |
3693 | WRITE_ONCE(lrugen->nr_pages[gen][type][zone], | |
3694 | lrugen->nr_pages[gen][type][zone] + delta); | |
3695 | ||
3696 | if (lru_gen_is_active(lruvec, gen)) | |
3697 | lru += LRU_ACTIVE; | |
3698 | __update_lru_size(lruvec, lru, zone, delta); | |
3699 | } | |
3700 | } | |
3701 | ||
3702 | static int should_skip_vma(unsigned long start, unsigned long end, struct mm_walk *args) | |
3703 | { | |
3704 | struct address_space *mapping; | |
3705 | struct vm_area_struct *vma = args->vma; | |
3706 | struct lru_gen_mm_walk *walk = args->private; | |
3707 | ||
3708 | if (!vma_is_accessible(vma)) | |
3709 | return true; | |
3710 | ||
3711 | if (is_vm_hugetlb_page(vma)) | |
3712 | return true; | |
3713 | ||
3714 | if (vma->vm_flags & (VM_LOCKED | VM_SPECIAL | VM_SEQ_READ | VM_RAND_READ)) | |
3715 | return true; | |
3716 | ||
3717 | if (vma == get_gate_vma(vma->vm_mm)) | |
3718 | return true; | |
3719 | ||
3720 | if (vma_is_anonymous(vma)) | |
3721 | return !walk->can_swap; | |
3722 | ||
3723 | if (WARN_ON_ONCE(!vma->vm_file || !vma->vm_file->f_mapping)) | |
3724 | return true; | |
3725 | ||
3726 | mapping = vma->vm_file->f_mapping; | |
3727 | if (mapping_unevictable(mapping)) | |
3728 | return true; | |
3729 | ||
3730 | if (shmem_mapping(mapping)) | |
3731 | return !walk->can_swap; | |
3732 | ||
3733 | /* to exclude special mappings like dax, etc. */ | |
3734 | return !mapping->a_ops->read_folio; | |
3735 | } | |
3736 | ||
3737 | /* | |
3738 | * Some userspace memory allocators map many single-page VMAs. Instead of | |
3739 | * returning back to the PGD table for each of such VMAs, finish an entire PMD | |
3740 | * table to reduce zigzags and improve cache performance. | |
3741 | */ | |
3742 | static bool get_next_vma(unsigned long mask, unsigned long size, struct mm_walk *args, | |
3743 | unsigned long *vm_start, unsigned long *vm_end) | |
3744 | { | |
3745 | unsigned long start = round_up(*vm_end, size); | |
3746 | unsigned long end = (start | ~mask) + 1; | |
3747 | ||
3748 | VM_WARN_ON_ONCE(mask & size); | |
3749 | VM_WARN_ON_ONCE((start & mask) != (*vm_start & mask)); | |
3750 | ||
3751 | while (args->vma) { | |
3752 | if (start >= args->vma->vm_end) { | |
3753 | args->vma = args->vma->vm_next; | |
3754 | continue; | |
3755 | } | |
3756 | ||
3757 | if (end && end <= args->vma->vm_start) | |
3758 | return false; | |
3759 | ||
3760 | if (should_skip_vma(args->vma->vm_start, args->vma->vm_end, args)) { | |
3761 | args->vma = args->vma->vm_next; | |
3762 | continue; | |
3763 | } | |
3764 | ||
3765 | *vm_start = max(start, args->vma->vm_start); | |
3766 | *vm_end = min(end - 1, args->vma->vm_end - 1) + 1; | |
3767 | ||
3768 | return true; | |
3769 | } | |
3770 | ||
3771 | return false; | |
3772 | } | |
3773 | ||
018ee47f YZ |
3774 | static unsigned long get_pte_pfn(pte_t pte, struct vm_area_struct *vma, unsigned long addr) |
3775 | { | |
3776 | unsigned long pfn = pte_pfn(pte); | |
3777 | ||
3778 | VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end); | |
3779 | ||
3780 | if (!pte_present(pte) || is_zero_pfn(pfn)) | |
3781 | return -1; | |
3782 | ||
3783 | if (WARN_ON_ONCE(pte_devmap(pte) || pte_special(pte))) | |
3784 | return -1; | |
3785 | ||
3786 | if (WARN_ON_ONCE(!pfn_valid(pfn))) | |
3787 | return -1; | |
3788 | ||
3789 | return pfn; | |
3790 | } | |
3791 | ||
bd74fdae YZ |
3792 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG) |
3793 | static unsigned long get_pmd_pfn(pmd_t pmd, struct vm_area_struct *vma, unsigned long addr) | |
3794 | { | |
3795 | unsigned long pfn = pmd_pfn(pmd); | |
3796 | ||
3797 | VM_WARN_ON_ONCE(addr < vma->vm_start || addr >= vma->vm_end); | |
3798 | ||
3799 | if (!pmd_present(pmd) || is_huge_zero_pmd(pmd)) | |
3800 | return -1; | |
3801 | ||
3802 | if (WARN_ON_ONCE(pmd_devmap(pmd))) | |
3803 | return -1; | |
3804 | ||
3805 | if (WARN_ON_ONCE(!pfn_valid(pfn))) | |
3806 | return -1; | |
3807 | ||
3808 | return pfn; | |
3809 | } | |
3810 | #endif | |
3811 | ||
018ee47f | 3812 | static struct folio *get_pfn_folio(unsigned long pfn, struct mem_cgroup *memcg, |
bd74fdae | 3813 | struct pglist_data *pgdat, bool can_swap) |
018ee47f YZ |
3814 | { |
3815 | struct folio *folio; | |
3816 | ||
3817 | /* try to avoid unnecessary memory loads */ | |
3818 | if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat)) | |
3819 | return NULL; | |
3820 | ||
3821 | folio = pfn_folio(pfn); | |
3822 | if (folio_nid(folio) != pgdat->node_id) | |
3823 | return NULL; | |
3824 | ||
3825 | if (folio_memcg_rcu(folio) != memcg) | |
3826 | return NULL; | |
3827 | ||
bd74fdae YZ |
3828 | /* file VMAs can contain anon pages from COW */ |
3829 | if (!folio_is_file_lru(folio) && !can_swap) | |
3830 | return NULL; | |
3831 | ||
018ee47f YZ |
3832 | return folio; |
3833 | } | |
3834 | ||
bd74fdae YZ |
3835 | static bool suitable_to_scan(int total, int young) |
3836 | { | |
3837 | int n = clamp_t(int, cache_line_size() / sizeof(pte_t), 2, 8); | |
3838 | ||
3839 | /* suitable if the average number of young PTEs per cacheline is >=1 */ | |
3840 | return young * n >= total; | |
3841 | } | |
3842 | ||
3843 | static bool walk_pte_range(pmd_t *pmd, unsigned long start, unsigned long end, | |
3844 | struct mm_walk *args) | |
3845 | { | |
3846 | int i; | |
3847 | pte_t *pte; | |
3848 | spinlock_t *ptl; | |
3849 | unsigned long addr; | |
3850 | int total = 0; | |
3851 | int young = 0; | |
3852 | struct lru_gen_mm_walk *walk = args->private; | |
3853 | struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec); | |
3854 | struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec); | |
3855 | int old_gen, new_gen = lru_gen_from_seq(walk->max_seq); | |
3856 | ||
3857 | VM_WARN_ON_ONCE(pmd_leaf(*pmd)); | |
3858 | ||
3859 | ptl = pte_lockptr(args->mm, pmd); | |
3860 | if (!spin_trylock(ptl)) | |
3861 | return false; | |
3862 | ||
3863 | arch_enter_lazy_mmu_mode(); | |
3864 | ||
3865 | pte = pte_offset_map(pmd, start & PMD_MASK); | |
3866 | restart: | |
3867 | for (i = pte_index(start), addr = start; addr != end; i++, addr += PAGE_SIZE) { | |
3868 | unsigned long pfn; | |
3869 | struct folio *folio; | |
3870 | ||
3871 | total++; | |
3872 | walk->mm_stats[MM_LEAF_TOTAL]++; | |
3873 | ||
3874 | pfn = get_pte_pfn(pte[i], args->vma, addr); | |
3875 | if (pfn == -1) | |
3876 | continue; | |
3877 | ||
3878 | if (!pte_young(pte[i])) { | |
3879 | walk->mm_stats[MM_LEAF_OLD]++; | |
3880 | continue; | |
3881 | } | |
3882 | ||
3883 | folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap); | |
3884 | if (!folio) | |
3885 | continue; | |
3886 | ||
3887 | if (!ptep_test_and_clear_young(args->vma, addr, pte + i)) | |
3888 | VM_WARN_ON_ONCE(true); | |
3889 | ||
3890 | young++; | |
3891 | walk->mm_stats[MM_LEAF_YOUNG]++; | |
3892 | ||
3893 | if (pte_dirty(pte[i]) && !folio_test_dirty(folio) && | |
3894 | !(folio_test_anon(folio) && folio_test_swapbacked(folio) && | |
3895 | !folio_test_swapcache(folio))) | |
3896 | folio_mark_dirty(folio); | |
3897 | ||
3898 | old_gen = folio_update_gen(folio, new_gen); | |
3899 | if (old_gen >= 0 && old_gen != new_gen) | |
3900 | update_batch_size(walk, folio, old_gen, new_gen); | |
3901 | } | |
3902 | ||
3903 | if (i < PTRS_PER_PTE && get_next_vma(PMD_MASK, PAGE_SIZE, args, &start, &end)) | |
3904 | goto restart; | |
3905 | ||
3906 | pte_unmap(pte); | |
3907 | ||
3908 | arch_leave_lazy_mmu_mode(); | |
3909 | spin_unlock(ptl); | |
3910 | ||
3911 | return suitable_to_scan(total, young); | |
3912 | } | |
3913 | ||
3914 | #if defined(CONFIG_TRANSPARENT_HUGEPAGE) || defined(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG) | |
3915 | static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma, | |
3916 | struct mm_walk *args, unsigned long *bitmap, unsigned long *start) | |
3917 | { | |
3918 | int i; | |
3919 | pmd_t *pmd; | |
3920 | spinlock_t *ptl; | |
3921 | struct lru_gen_mm_walk *walk = args->private; | |
3922 | struct mem_cgroup *memcg = lruvec_memcg(walk->lruvec); | |
3923 | struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec); | |
3924 | int old_gen, new_gen = lru_gen_from_seq(walk->max_seq); | |
3925 | ||
3926 | VM_WARN_ON_ONCE(pud_leaf(*pud)); | |
3927 | ||
3928 | /* try to batch at most 1+MIN_LRU_BATCH+1 entries */ | |
3929 | if (*start == -1) { | |
3930 | *start = next; | |
3931 | return; | |
3932 | } | |
3933 | ||
3934 | i = next == -1 ? 0 : pmd_index(next) - pmd_index(*start); | |
3935 | if (i && i <= MIN_LRU_BATCH) { | |
3936 | __set_bit(i - 1, bitmap); | |
3937 | return; | |
3938 | } | |
3939 | ||
3940 | pmd = pmd_offset(pud, *start); | |
3941 | ||
3942 | ptl = pmd_lockptr(args->mm, pmd); | |
3943 | if (!spin_trylock(ptl)) | |
3944 | goto done; | |
3945 | ||
3946 | arch_enter_lazy_mmu_mode(); | |
3947 | ||
3948 | do { | |
3949 | unsigned long pfn; | |
3950 | struct folio *folio; | |
3951 | unsigned long addr = i ? (*start & PMD_MASK) + i * PMD_SIZE : *start; | |
3952 | ||
3953 | pfn = get_pmd_pfn(pmd[i], vma, addr); | |
3954 | if (pfn == -1) | |
3955 | goto next; | |
3956 | ||
3957 | if (!pmd_trans_huge(pmd[i])) { | |
354ed597 YZ |
3958 | if (IS_ENABLED(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG) && |
3959 | get_cap(LRU_GEN_NONLEAF_YOUNG)) | |
bd74fdae YZ |
3960 | pmdp_test_and_clear_young(vma, addr, pmd + i); |
3961 | goto next; | |
3962 | } | |
3963 | ||
3964 | folio = get_pfn_folio(pfn, memcg, pgdat, walk->can_swap); | |
3965 | if (!folio) | |
3966 | goto next; | |
3967 | ||
3968 | if (!pmdp_test_and_clear_young(vma, addr, pmd + i)) | |
3969 | goto next; | |
3970 | ||
3971 | walk->mm_stats[MM_LEAF_YOUNG]++; | |
3972 | ||
3973 | if (pmd_dirty(pmd[i]) && !folio_test_dirty(folio) && | |
3974 | !(folio_test_anon(folio) && folio_test_swapbacked(folio) && | |
3975 | !folio_test_swapcache(folio))) | |
3976 | folio_mark_dirty(folio); | |
3977 | ||
3978 | old_gen = folio_update_gen(folio, new_gen); | |
3979 | if (old_gen >= 0 && old_gen != new_gen) | |
3980 | update_batch_size(walk, folio, old_gen, new_gen); | |
3981 | next: | |
3982 | i = i > MIN_LRU_BATCH ? 0 : find_next_bit(bitmap, MIN_LRU_BATCH, i) + 1; | |
3983 | } while (i <= MIN_LRU_BATCH); | |
3984 | ||
3985 | arch_leave_lazy_mmu_mode(); | |
3986 | spin_unlock(ptl); | |
3987 | done: | |
3988 | *start = -1; | |
3989 | bitmap_zero(bitmap, MIN_LRU_BATCH); | |
3990 | } | |
3991 | #else | |
3992 | static void walk_pmd_range_locked(pud_t *pud, unsigned long next, struct vm_area_struct *vma, | |
3993 | struct mm_walk *args, unsigned long *bitmap, unsigned long *start) | |
3994 | { | |
3995 | } | |
3996 | #endif | |
3997 | ||
3998 | static void walk_pmd_range(pud_t *pud, unsigned long start, unsigned long end, | |
3999 | struct mm_walk *args) | |
4000 | { | |
4001 | int i; | |
4002 | pmd_t *pmd; | |
4003 | unsigned long next; | |
4004 | unsigned long addr; | |
4005 | struct vm_area_struct *vma; | |
4006 | unsigned long pos = -1; | |
4007 | struct lru_gen_mm_walk *walk = args->private; | |
4008 | unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {}; | |
4009 | ||
4010 | VM_WARN_ON_ONCE(pud_leaf(*pud)); | |
4011 | ||
4012 | /* | |
4013 | * Finish an entire PMD in two passes: the first only reaches to PTE | |
4014 | * tables to avoid taking the PMD lock; the second, if necessary, takes | |
4015 | * the PMD lock to clear the accessed bit in PMD entries. | |
4016 | */ | |
4017 | pmd = pmd_offset(pud, start & PUD_MASK); | |
4018 | restart: | |
4019 | /* walk_pte_range() may call get_next_vma() */ | |
4020 | vma = args->vma; | |
4021 | for (i = pmd_index(start), addr = start; addr != end; i++, addr = next) { | |
4022 | pmd_t val = pmd_read_atomic(pmd + i); | |
4023 | ||
4024 | /* for pmd_read_atomic() */ | |
4025 | barrier(); | |
4026 | ||
4027 | next = pmd_addr_end(addr, end); | |
4028 | ||
4029 | if (!pmd_present(val) || is_huge_zero_pmd(val)) { | |
4030 | walk->mm_stats[MM_LEAF_TOTAL]++; | |
4031 | continue; | |
4032 | } | |
4033 | ||
4034 | #ifdef CONFIG_TRANSPARENT_HUGEPAGE | |
4035 | if (pmd_trans_huge(val)) { | |
4036 | unsigned long pfn = pmd_pfn(val); | |
4037 | struct pglist_data *pgdat = lruvec_pgdat(walk->lruvec); | |
4038 | ||
4039 | walk->mm_stats[MM_LEAF_TOTAL]++; | |
4040 | ||
4041 | if (!pmd_young(val)) { | |
4042 | walk->mm_stats[MM_LEAF_OLD]++; | |
4043 | continue; | |
4044 | } | |
4045 | ||
4046 | /* try to avoid unnecessary memory loads */ | |
4047 | if (pfn < pgdat->node_start_pfn || pfn >= pgdat_end_pfn(pgdat)) | |
4048 | continue; | |
4049 | ||
4050 | walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos); | |
4051 | continue; | |
4052 | } | |
4053 | #endif | |
4054 | walk->mm_stats[MM_NONLEAF_TOTAL]++; | |
4055 | ||
4056 | #ifdef CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG | |
354ed597 YZ |
4057 | if (get_cap(LRU_GEN_NONLEAF_YOUNG)) { |
4058 | if (!pmd_young(val)) | |
4059 | continue; | |
bd74fdae | 4060 | |
354ed597 YZ |
4061 | walk_pmd_range_locked(pud, addr, vma, args, bitmap, &pos); |
4062 | } | |
bd74fdae YZ |
4063 | #endif |
4064 | if (!walk->force_scan && !test_bloom_filter(walk->lruvec, walk->max_seq, pmd + i)) | |
4065 | continue; | |
4066 | ||
4067 | walk->mm_stats[MM_NONLEAF_FOUND]++; | |
4068 | ||
4069 | if (!walk_pte_range(&val, addr, next, args)) | |
4070 | continue; | |
4071 | ||
4072 | walk->mm_stats[MM_NONLEAF_ADDED]++; | |
4073 | ||
4074 | /* carry over to the next generation */ | |
4075 | update_bloom_filter(walk->lruvec, walk->max_seq + 1, pmd + i); | |
4076 | } | |
4077 | ||
4078 | walk_pmd_range_locked(pud, -1, vma, args, bitmap, &pos); | |
4079 | ||
4080 | if (i < PTRS_PER_PMD && get_next_vma(PUD_MASK, PMD_SIZE, args, &start, &end)) | |
4081 | goto restart; | |
4082 | } | |
4083 | ||
4084 | static int walk_pud_range(p4d_t *p4d, unsigned long start, unsigned long end, | |
4085 | struct mm_walk *args) | |
4086 | { | |
4087 | int i; | |
4088 | pud_t *pud; | |
4089 | unsigned long addr; | |
4090 | unsigned long next; | |
4091 | struct lru_gen_mm_walk *walk = args->private; | |
4092 | ||
4093 | VM_WARN_ON_ONCE(p4d_leaf(*p4d)); | |
4094 | ||
4095 | pud = pud_offset(p4d, start & P4D_MASK); | |
4096 | restart: | |
4097 | for (i = pud_index(start), addr = start; addr != end; i++, addr = next) { | |
4098 | pud_t val = READ_ONCE(pud[i]); | |
4099 | ||
4100 | next = pud_addr_end(addr, end); | |
4101 | ||
4102 | if (!pud_present(val) || WARN_ON_ONCE(pud_leaf(val))) | |
4103 | continue; | |
4104 | ||
4105 | walk_pmd_range(&val, addr, next, args); | |
4106 | ||
4107 | /* a racy check to curtail the waiting time */ | |
4108 | if (wq_has_sleeper(&walk->lruvec->mm_state.wait)) | |
4109 | return 1; | |
4110 | ||
4111 | if (need_resched() || walk->batched >= MAX_LRU_BATCH) { | |
4112 | end = (addr | ~PUD_MASK) + 1; | |
4113 | goto done; | |
4114 | } | |
4115 | } | |
4116 | ||
4117 | if (i < PTRS_PER_PUD && get_next_vma(P4D_MASK, PUD_SIZE, args, &start, &end)) | |
4118 | goto restart; | |
4119 | ||
4120 | end = round_up(end, P4D_SIZE); | |
4121 | done: | |
4122 | if (!end || !args->vma) | |
4123 | return 1; | |
4124 | ||
4125 | walk->next_addr = max(end, args->vma->vm_start); | |
4126 | ||
4127 | return -EAGAIN; | |
4128 | } | |
4129 | ||
4130 | static void walk_mm(struct lruvec *lruvec, struct mm_struct *mm, struct lru_gen_mm_walk *walk) | |
4131 | { | |
4132 | static const struct mm_walk_ops mm_walk_ops = { | |
4133 | .test_walk = should_skip_vma, | |
4134 | .p4d_entry = walk_pud_range, | |
4135 | }; | |
4136 | ||
4137 | int err; | |
4138 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
4139 | ||
4140 | walk->next_addr = FIRST_USER_ADDRESS; | |
4141 | ||
4142 | do { | |
4143 | err = -EBUSY; | |
4144 | ||
4145 | /* folio_update_gen() requires stable folio_memcg() */ | |
4146 | if (!mem_cgroup_trylock_pages(memcg)) | |
4147 | break; | |
4148 | ||
4149 | /* the caller might be holding the lock for write */ | |
4150 | if (mmap_read_trylock(mm)) { | |
4151 | err = walk_page_range(mm, walk->next_addr, ULONG_MAX, &mm_walk_ops, walk); | |
4152 | ||
4153 | mmap_read_unlock(mm); | |
4154 | } | |
4155 | ||
4156 | mem_cgroup_unlock_pages(); | |
4157 | ||
4158 | if (walk->batched) { | |
4159 | spin_lock_irq(&lruvec->lru_lock); | |
4160 | reset_batch_size(lruvec, walk); | |
4161 | spin_unlock_irq(&lruvec->lru_lock); | |
4162 | } | |
4163 | ||
4164 | cond_resched(); | |
4165 | } while (err == -EAGAIN); | |
4166 | } | |
4167 | ||
4168 | static struct lru_gen_mm_walk *set_mm_walk(struct pglist_data *pgdat) | |
4169 | { | |
4170 | struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk; | |
4171 | ||
4172 | if (pgdat && current_is_kswapd()) { | |
4173 | VM_WARN_ON_ONCE(walk); | |
4174 | ||
4175 | walk = &pgdat->mm_walk; | |
4176 | } else if (!pgdat && !walk) { | |
4177 | VM_WARN_ON_ONCE(current_is_kswapd()); | |
4178 | ||
4179 | walk = kzalloc(sizeof(*walk), __GFP_HIGH | __GFP_NOMEMALLOC | __GFP_NOWARN); | |
4180 | } | |
4181 | ||
4182 | current->reclaim_state->mm_walk = walk; | |
4183 | ||
4184 | return walk; | |
4185 | } | |
4186 | ||
4187 | static void clear_mm_walk(void) | |
4188 | { | |
4189 | struct lru_gen_mm_walk *walk = current->reclaim_state->mm_walk; | |
4190 | ||
4191 | VM_WARN_ON_ONCE(walk && memchr_inv(walk->nr_pages, 0, sizeof(walk->nr_pages))); | |
4192 | VM_WARN_ON_ONCE(walk && memchr_inv(walk->mm_stats, 0, sizeof(walk->mm_stats))); | |
4193 | ||
4194 | current->reclaim_state->mm_walk = NULL; | |
4195 | ||
4196 | if (!current_is_kswapd()) | |
4197 | kfree(walk); | |
4198 | } | |
4199 | ||
ac35a490 YZ |
4200 | static void inc_min_seq(struct lruvec *lruvec, int type) |
4201 | { | |
4202 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
4203 | ||
4204 | reset_ctrl_pos(lruvec, type, true); | |
4205 | WRITE_ONCE(lrugen->min_seq[type], lrugen->min_seq[type] + 1); | |
4206 | } | |
4207 | ||
4208 | static bool try_to_inc_min_seq(struct lruvec *lruvec, bool can_swap) | |
4209 | { | |
4210 | int gen, type, zone; | |
4211 | bool success = false; | |
4212 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
4213 | DEFINE_MIN_SEQ(lruvec); | |
4214 | ||
4215 | VM_WARN_ON_ONCE(!seq_is_valid(lruvec)); | |
4216 | ||
4217 | /* find the oldest populated generation */ | |
4218 | for (type = !can_swap; type < ANON_AND_FILE; type++) { | |
4219 | while (min_seq[type] + MIN_NR_GENS <= lrugen->max_seq) { | |
4220 | gen = lru_gen_from_seq(min_seq[type]); | |
4221 | ||
4222 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
4223 | if (!list_empty(&lrugen->lists[gen][type][zone])) | |
4224 | goto next; | |
4225 | } | |
4226 | ||
4227 | min_seq[type]++; | |
4228 | } | |
4229 | next: | |
4230 | ; | |
4231 | } | |
4232 | ||
4233 | /* see the comment on lru_gen_struct */ | |
4234 | if (can_swap) { | |
4235 | min_seq[LRU_GEN_ANON] = min(min_seq[LRU_GEN_ANON], min_seq[LRU_GEN_FILE]); | |
4236 | min_seq[LRU_GEN_FILE] = max(min_seq[LRU_GEN_ANON], lrugen->min_seq[LRU_GEN_FILE]); | |
4237 | } | |
4238 | ||
4239 | for (type = !can_swap; type < ANON_AND_FILE; type++) { | |
4240 | if (min_seq[type] == lrugen->min_seq[type]) | |
4241 | continue; | |
4242 | ||
4243 | reset_ctrl_pos(lruvec, type, true); | |
4244 | WRITE_ONCE(lrugen->min_seq[type], min_seq[type]); | |
4245 | success = true; | |
4246 | } | |
4247 | ||
4248 | return success; | |
4249 | } | |
4250 | ||
bd74fdae | 4251 | static void inc_max_seq(struct lruvec *lruvec, bool can_swap) |
ac35a490 YZ |
4252 | { |
4253 | int prev, next; | |
4254 | int type, zone; | |
4255 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
4256 | ||
4257 | spin_lock_irq(&lruvec->lru_lock); | |
4258 | ||
4259 | VM_WARN_ON_ONCE(!seq_is_valid(lruvec)); | |
4260 | ||
ac35a490 YZ |
4261 | for (type = ANON_AND_FILE - 1; type >= 0; type--) { |
4262 | if (get_nr_gens(lruvec, type) != MAX_NR_GENS) | |
4263 | continue; | |
4264 | ||
4265 | VM_WARN_ON_ONCE(type == LRU_GEN_FILE || can_swap); | |
4266 | ||
4267 | inc_min_seq(lruvec, type); | |
4268 | } | |
4269 | ||
4270 | /* | |
4271 | * Update the active/inactive LRU sizes for compatibility. Both sides of | |
4272 | * the current max_seq need to be covered, since max_seq+1 can overlap | |
4273 | * with min_seq[LRU_GEN_ANON] if swapping is constrained. And if they do | |
4274 | * overlap, cold/hot inversion happens. | |
4275 | */ | |
4276 | prev = lru_gen_from_seq(lrugen->max_seq - 1); | |
4277 | next = lru_gen_from_seq(lrugen->max_seq + 1); | |
4278 | ||
4279 | for (type = 0; type < ANON_AND_FILE; type++) { | |
4280 | for (zone = 0; zone < MAX_NR_ZONES; zone++) { | |
4281 | enum lru_list lru = type * LRU_INACTIVE_FILE; | |
4282 | long delta = lrugen->nr_pages[prev][type][zone] - | |
4283 | lrugen->nr_pages[next][type][zone]; | |
4284 | ||
4285 | if (!delta) | |
4286 | continue; | |
4287 | ||
4288 | __update_lru_size(lruvec, lru, zone, delta); | |
4289 | __update_lru_size(lruvec, lru + LRU_ACTIVE, zone, -delta); | |
4290 | } | |
4291 | } | |
4292 | ||
4293 | for (type = 0; type < ANON_AND_FILE; type++) | |
4294 | reset_ctrl_pos(lruvec, type, false); | |
4295 | ||
4296 | /* make sure preceding modifications appear */ | |
4297 | smp_store_release(&lrugen->max_seq, lrugen->max_seq + 1); | |
bd74fdae | 4298 | |
ac35a490 YZ |
4299 | spin_unlock_irq(&lruvec->lru_lock); |
4300 | } | |
4301 | ||
bd74fdae YZ |
4302 | static bool try_to_inc_max_seq(struct lruvec *lruvec, unsigned long max_seq, |
4303 | struct scan_control *sc, bool can_swap) | |
4304 | { | |
4305 | bool success; | |
4306 | struct lru_gen_mm_walk *walk; | |
4307 | struct mm_struct *mm = NULL; | |
4308 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
4309 | ||
4310 | VM_WARN_ON_ONCE(max_seq > READ_ONCE(lrugen->max_seq)); | |
4311 | ||
4312 | /* see the comment in iterate_mm_list() */ | |
4313 | if (max_seq <= READ_ONCE(lruvec->mm_state.seq)) { | |
4314 | success = false; | |
4315 | goto done; | |
4316 | } | |
4317 | ||
4318 | /* | |
4319 | * If the hardware doesn't automatically set the accessed bit, fallback | |
4320 | * to lru_gen_look_around(), which only clears the accessed bit in a | |
4321 | * handful of PTEs. Spreading the work out over a period of time usually | |
4322 | * is less efficient, but it avoids bursty page faults. | |
4323 | */ | |
354ed597 | 4324 | if (!(arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK))) { |
bd74fdae YZ |
4325 | success = iterate_mm_list_nowalk(lruvec, max_seq); |
4326 | goto done; | |
4327 | } | |
4328 | ||
4329 | walk = set_mm_walk(NULL); | |
4330 | if (!walk) { | |
4331 | success = iterate_mm_list_nowalk(lruvec, max_seq); | |
4332 | goto done; | |
4333 | } | |
4334 | ||
4335 | walk->lruvec = lruvec; | |
4336 | walk->max_seq = max_seq; | |
4337 | walk->can_swap = can_swap; | |
4338 | walk->force_scan = false; | |
4339 | ||
4340 | do { | |
4341 | success = iterate_mm_list(lruvec, walk, &mm); | |
4342 | if (mm) | |
4343 | walk_mm(lruvec, mm, walk); | |
4344 | ||
4345 | cond_resched(); | |
4346 | } while (mm); | |
4347 | done: | |
4348 | if (!success) { | |
4349 | if (sc->priority <= DEF_PRIORITY - 2) | |
4350 | wait_event_killable(lruvec->mm_state.wait, | |
4351 | max_seq < READ_ONCE(lrugen->max_seq)); | |
4352 | ||
4353 | return max_seq < READ_ONCE(lrugen->max_seq); | |
4354 | } | |
4355 | ||
4356 | VM_WARN_ON_ONCE(max_seq != READ_ONCE(lrugen->max_seq)); | |
4357 | ||
4358 | inc_max_seq(lruvec, can_swap); | |
4359 | /* either this sees any waiters or they will see updated max_seq */ | |
4360 | if (wq_has_sleeper(&lruvec->mm_state.wait)) | |
4361 | wake_up_all(&lruvec->mm_state.wait); | |
4362 | ||
4363 | wakeup_flusher_threads(WB_REASON_VMSCAN); | |
4364 | ||
4365 | return true; | |
4366 | } | |
4367 | ||
ac35a490 YZ |
4368 | static bool should_run_aging(struct lruvec *lruvec, unsigned long max_seq, unsigned long *min_seq, |
4369 | struct scan_control *sc, bool can_swap, unsigned long *nr_to_scan) | |
4370 | { | |
4371 | int gen, type, zone; | |
4372 | unsigned long old = 0; | |
4373 | unsigned long young = 0; | |
4374 | unsigned long total = 0; | |
4375 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
4376 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
4377 | ||
4378 | for (type = !can_swap; type < ANON_AND_FILE; type++) { | |
4379 | unsigned long seq; | |
4380 | ||
4381 | for (seq = min_seq[type]; seq <= max_seq; seq++) { | |
4382 | unsigned long size = 0; | |
4383 | ||
4384 | gen = lru_gen_from_seq(seq); | |
4385 | ||
4386 | for (zone = 0; zone < MAX_NR_ZONES; zone++) | |
4387 | size += max(READ_ONCE(lrugen->nr_pages[gen][type][zone]), 0L); | |
4388 | ||
4389 | total += size; | |
4390 | if (seq == max_seq) | |
4391 | young += size; | |
4392 | else if (seq + MIN_NR_GENS == max_seq) | |
4393 | old += size; | |
4394 | } | |
4395 | } | |
4396 | ||
4397 | /* try to scrape all its memory if this memcg was deleted */ | |
4398 | *nr_to_scan = mem_cgroup_online(memcg) ? (total >> sc->priority) : total; | |
4399 | ||
4400 | /* | |
4401 | * The aging tries to be lazy to reduce the overhead, while the eviction | |
4402 | * stalls when the number of generations reaches MIN_NR_GENS. Hence, the | |
4403 | * ideal number of generations is MIN_NR_GENS+1. | |
4404 | */ | |
4405 | if (min_seq[!can_swap] + MIN_NR_GENS > max_seq) | |
4406 | return true; | |
4407 | if (min_seq[!can_swap] + MIN_NR_GENS < max_seq) | |
4408 | return false; | |
4409 | ||
4410 | /* | |
4411 | * It's also ideal to spread pages out evenly, i.e., 1/(MIN_NR_GENS+1) | |
4412 | * of the total number of pages for each generation. A reasonable range | |
4413 | * for this average portion is [1/MIN_NR_GENS, 1/(MIN_NR_GENS+2)]. The | |
4414 | * aging cares about the upper bound of hot pages, while the eviction | |
4415 | * cares about the lower bound of cold pages. | |
4416 | */ | |
4417 | if (young * MIN_NR_GENS > total) | |
4418 | return true; | |
4419 | if (old * (MIN_NR_GENS + 2) < total) | |
4420 | return true; | |
4421 | ||
4422 | return false; | |
4423 | } | |
4424 | ||
4425 | static void age_lruvec(struct lruvec *lruvec, struct scan_control *sc) | |
4426 | { | |
4427 | bool need_aging; | |
4428 | unsigned long nr_to_scan; | |
4429 | int swappiness = get_swappiness(lruvec, sc); | |
4430 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
4431 | DEFINE_MAX_SEQ(lruvec); | |
4432 | DEFINE_MIN_SEQ(lruvec); | |
4433 | ||
4434 | VM_WARN_ON_ONCE(sc->memcg_low_reclaim); | |
4435 | ||
4436 | mem_cgroup_calculate_protection(NULL, memcg); | |
4437 | ||
4438 | if (mem_cgroup_below_min(memcg)) | |
4439 | return; | |
4440 | ||
4441 | need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, swappiness, &nr_to_scan); | |
4442 | if (need_aging) | |
bd74fdae | 4443 | try_to_inc_max_seq(lruvec, max_seq, sc, swappiness); |
ac35a490 YZ |
4444 | } |
4445 | ||
4446 | static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc) | |
4447 | { | |
4448 | struct mem_cgroup *memcg; | |
4449 | ||
4450 | VM_WARN_ON_ONCE(!current_is_kswapd()); | |
4451 | ||
f76c8337 YZ |
4452 | sc->last_reclaimed = sc->nr_reclaimed; |
4453 | ||
4454 | /* | |
4455 | * To reduce the chance of going into the aging path, which can be | |
4456 | * costly, optimistically skip it if the flag below was cleared in the | |
4457 | * eviction path. This improves the overall performance when multiple | |
4458 | * memcgs are available. | |
4459 | */ | |
4460 | if (!sc->memcgs_need_aging) { | |
4461 | sc->memcgs_need_aging = true; | |
4462 | return; | |
4463 | } | |
4464 | ||
bd74fdae YZ |
4465 | set_mm_walk(pgdat); |
4466 | ||
ac35a490 YZ |
4467 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
4468 | do { | |
4469 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); | |
4470 | ||
4471 | age_lruvec(lruvec, sc); | |
4472 | ||
4473 | cond_resched(); | |
4474 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL))); | |
bd74fdae YZ |
4475 | |
4476 | clear_mm_walk(); | |
ac35a490 YZ |
4477 | } |
4478 | ||
018ee47f YZ |
4479 | /* |
4480 | * This function exploits spatial locality when shrink_page_list() walks the | |
bd74fdae YZ |
4481 | * rmap. It scans the adjacent PTEs of a young PTE and promotes hot pages. If |
4482 | * the scan was done cacheline efficiently, it adds the PMD entry pointing to | |
4483 | * the PTE table to the Bloom filter. This forms a feedback loop between the | |
4484 | * eviction and the aging. | |
018ee47f YZ |
4485 | */ |
4486 | void lru_gen_look_around(struct page_vma_mapped_walk *pvmw) | |
4487 | { | |
4488 | int i; | |
4489 | pte_t *pte; | |
4490 | unsigned long start; | |
4491 | unsigned long end; | |
4492 | unsigned long addr; | |
bd74fdae YZ |
4493 | struct lru_gen_mm_walk *walk; |
4494 | int young = 0; | |
018ee47f YZ |
4495 | unsigned long bitmap[BITS_TO_LONGS(MIN_LRU_BATCH)] = {}; |
4496 | struct folio *folio = pfn_folio(pvmw->pfn); | |
4497 | struct mem_cgroup *memcg = folio_memcg(folio); | |
4498 | struct pglist_data *pgdat = folio_pgdat(folio); | |
4499 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); | |
4500 | DEFINE_MAX_SEQ(lruvec); | |
4501 | int old_gen, new_gen = lru_gen_from_seq(max_seq); | |
4502 | ||
4503 | lockdep_assert_held(pvmw->ptl); | |
4504 | VM_WARN_ON_ONCE_FOLIO(folio_test_lru(folio), folio); | |
4505 | ||
4506 | if (spin_is_contended(pvmw->ptl)) | |
4507 | return; | |
4508 | ||
bd74fdae YZ |
4509 | /* avoid taking the LRU lock under the PTL when possible */ |
4510 | walk = current->reclaim_state ? current->reclaim_state->mm_walk : NULL; | |
4511 | ||
018ee47f YZ |
4512 | start = max(pvmw->address & PMD_MASK, pvmw->vma->vm_start); |
4513 | end = min(pvmw->address | ~PMD_MASK, pvmw->vma->vm_end - 1) + 1; | |
4514 | ||
4515 | if (end - start > MIN_LRU_BATCH * PAGE_SIZE) { | |
4516 | if (pvmw->address - start < MIN_LRU_BATCH * PAGE_SIZE / 2) | |
4517 | end = start + MIN_LRU_BATCH * PAGE_SIZE; | |
4518 | else if (end - pvmw->address < MIN_LRU_BATCH * PAGE_SIZE / 2) | |
4519 | start = end - MIN_LRU_BATCH * PAGE_SIZE; | |
4520 | else { | |
4521 | start = pvmw->address - MIN_LRU_BATCH * PAGE_SIZE / 2; | |
4522 | end = pvmw->address + MIN_LRU_BATCH * PAGE_SIZE / 2; | |
4523 | } | |
4524 | } | |
4525 | ||
4526 | pte = pvmw->pte - (pvmw->address - start) / PAGE_SIZE; | |
4527 | ||
4528 | rcu_read_lock(); | |
4529 | arch_enter_lazy_mmu_mode(); | |
4530 | ||
4531 | for (i = 0, addr = start; addr != end; i++, addr += PAGE_SIZE) { | |
4532 | unsigned long pfn; | |
4533 | ||
4534 | pfn = get_pte_pfn(pte[i], pvmw->vma, addr); | |
4535 | if (pfn == -1) | |
4536 | continue; | |
4537 | ||
4538 | if (!pte_young(pte[i])) | |
4539 | continue; | |
4540 | ||
bd74fdae | 4541 | folio = get_pfn_folio(pfn, memcg, pgdat, !walk || walk->can_swap); |
018ee47f YZ |
4542 | if (!folio) |
4543 | continue; | |
4544 | ||
4545 | if (!ptep_test_and_clear_young(pvmw->vma, addr, pte + i)) | |
4546 | VM_WARN_ON_ONCE(true); | |
4547 | ||
bd74fdae YZ |
4548 | young++; |
4549 | ||
018ee47f YZ |
4550 | if (pte_dirty(pte[i]) && !folio_test_dirty(folio) && |
4551 | !(folio_test_anon(folio) && folio_test_swapbacked(folio) && | |
4552 | !folio_test_swapcache(folio))) | |
4553 | folio_mark_dirty(folio); | |
4554 | ||
4555 | old_gen = folio_lru_gen(folio); | |
4556 | if (old_gen < 0) | |
4557 | folio_set_referenced(folio); | |
4558 | else if (old_gen != new_gen) | |
4559 | __set_bit(i, bitmap); | |
4560 | } | |
4561 | ||
4562 | arch_leave_lazy_mmu_mode(); | |
4563 | rcu_read_unlock(); | |
4564 | ||
bd74fdae YZ |
4565 | /* feedback from rmap walkers to page table walkers */ |
4566 | if (suitable_to_scan(i, young)) | |
4567 | update_bloom_filter(lruvec, max_seq, pvmw->pmd); | |
4568 | ||
4569 | if (!walk && bitmap_weight(bitmap, MIN_LRU_BATCH) < PAGEVEC_SIZE) { | |
018ee47f YZ |
4570 | for_each_set_bit(i, bitmap, MIN_LRU_BATCH) { |
4571 | folio = pfn_folio(pte_pfn(pte[i])); | |
4572 | folio_activate(folio); | |
4573 | } | |
4574 | return; | |
4575 | } | |
4576 | ||
4577 | /* folio_update_gen() requires stable folio_memcg() */ | |
4578 | if (!mem_cgroup_trylock_pages(memcg)) | |
4579 | return; | |
4580 | ||
bd74fdae YZ |
4581 | if (!walk) { |
4582 | spin_lock_irq(&lruvec->lru_lock); | |
4583 | new_gen = lru_gen_from_seq(lruvec->lrugen.max_seq); | |
4584 | } | |
018ee47f YZ |
4585 | |
4586 | for_each_set_bit(i, bitmap, MIN_LRU_BATCH) { | |
4587 | folio = pfn_folio(pte_pfn(pte[i])); | |
4588 | if (folio_memcg_rcu(folio) != memcg) | |
4589 | continue; | |
4590 | ||
4591 | old_gen = folio_update_gen(folio, new_gen); | |
4592 | if (old_gen < 0 || old_gen == new_gen) | |
4593 | continue; | |
4594 | ||
bd74fdae YZ |
4595 | if (walk) |
4596 | update_batch_size(walk, folio, old_gen, new_gen); | |
4597 | else | |
4598 | lru_gen_update_size(lruvec, folio, old_gen, new_gen); | |
018ee47f YZ |
4599 | } |
4600 | ||
bd74fdae YZ |
4601 | if (!walk) |
4602 | spin_unlock_irq(&lruvec->lru_lock); | |
018ee47f YZ |
4603 | |
4604 | mem_cgroup_unlock_pages(); | |
4605 | } | |
4606 | ||
ac35a490 YZ |
4607 | /****************************************************************************** |
4608 | * the eviction | |
4609 | ******************************************************************************/ | |
4610 | ||
4611 | static bool sort_folio(struct lruvec *lruvec, struct folio *folio, int tier_idx) | |
4612 | { | |
4613 | bool success; | |
4614 | int gen = folio_lru_gen(folio); | |
4615 | int type = folio_is_file_lru(folio); | |
4616 | int zone = folio_zonenum(folio); | |
4617 | int delta = folio_nr_pages(folio); | |
4618 | int refs = folio_lru_refs(folio); | |
4619 | int tier = lru_tier_from_refs(refs); | |
4620 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
4621 | ||
4622 | VM_WARN_ON_ONCE_FOLIO(gen >= MAX_NR_GENS, folio); | |
4623 | ||
4624 | /* unevictable */ | |
4625 | if (!folio_evictable(folio)) { | |
4626 | success = lru_gen_del_folio(lruvec, folio, true); | |
4627 | VM_WARN_ON_ONCE_FOLIO(!success, folio); | |
4628 | folio_set_unevictable(folio); | |
4629 | lruvec_add_folio(lruvec, folio); | |
4630 | __count_vm_events(UNEVICTABLE_PGCULLED, delta); | |
4631 | return true; | |
4632 | } | |
4633 | ||
4634 | /* dirty lazyfree */ | |
4635 | if (type == LRU_GEN_FILE && folio_test_anon(folio) && folio_test_dirty(folio)) { | |
4636 | success = lru_gen_del_folio(lruvec, folio, true); | |
4637 | VM_WARN_ON_ONCE_FOLIO(!success, folio); | |
4638 | folio_set_swapbacked(folio); | |
4639 | lruvec_add_folio_tail(lruvec, folio); | |
4640 | return true; | |
4641 | } | |
4642 | ||
018ee47f YZ |
4643 | /* promoted */ |
4644 | if (gen != lru_gen_from_seq(lrugen->min_seq[type])) { | |
4645 | list_move(&folio->lru, &lrugen->lists[gen][type][zone]); | |
4646 | return true; | |
4647 | } | |
4648 | ||
ac35a490 YZ |
4649 | /* protected */ |
4650 | if (tier > tier_idx) { | |
4651 | int hist = lru_hist_from_seq(lrugen->min_seq[type]); | |
4652 | ||
4653 | gen = folio_inc_gen(lruvec, folio, false); | |
4654 | list_move_tail(&folio->lru, &lrugen->lists[gen][type][zone]); | |
4655 | ||
4656 | WRITE_ONCE(lrugen->protected[hist][type][tier - 1], | |
4657 | lrugen->protected[hist][type][tier - 1] + delta); | |
4658 | __mod_lruvec_state(lruvec, WORKINGSET_ACTIVATE_BASE + type, delta); | |
4659 | return true; | |
4660 | } | |
4661 | ||
4662 | /* waiting for writeback */ | |
4663 | if (folio_test_locked(folio) || folio_test_writeback(folio) || | |
4664 | (type == LRU_GEN_FILE && folio_test_dirty(folio))) { | |
4665 | gen = folio_inc_gen(lruvec, folio, true); | |
4666 | list_move(&folio->lru, &lrugen->lists[gen][type][zone]); | |
4667 | return true; | |
4668 | } | |
4669 | ||
4670 | return false; | |
4671 | } | |
4672 | ||
4673 | static bool isolate_folio(struct lruvec *lruvec, struct folio *folio, struct scan_control *sc) | |
4674 | { | |
4675 | bool success; | |
4676 | ||
4677 | /* unmapping inhibited */ | |
4678 | if (!sc->may_unmap && folio_mapped(folio)) | |
4679 | return false; | |
4680 | ||
4681 | /* swapping inhibited */ | |
4682 | if (!(sc->may_writepage && (sc->gfp_mask & __GFP_IO)) && | |
4683 | (folio_test_dirty(folio) || | |
4684 | (folio_test_anon(folio) && !folio_test_swapcache(folio)))) | |
4685 | return false; | |
4686 | ||
4687 | /* raced with release_pages() */ | |
4688 | if (!folio_try_get(folio)) | |
4689 | return false; | |
4690 | ||
4691 | /* raced with another isolation */ | |
4692 | if (!folio_test_clear_lru(folio)) { | |
4693 | folio_put(folio); | |
4694 | return false; | |
4695 | } | |
4696 | ||
4697 | /* see the comment on MAX_NR_TIERS */ | |
4698 | if (!folio_test_referenced(folio)) | |
4699 | set_mask_bits(&folio->flags, LRU_REFS_MASK | LRU_REFS_FLAGS, 0); | |
4700 | ||
4701 | /* for shrink_page_list() */ | |
4702 | folio_clear_reclaim(folio); | |
4703 | folio_clear_referenced(folio); | |
4704 | ||
4705 | success = lru_gen_del_folio(lruvec, folio, true); | |
4706 | VM_WARN_ON_ONCE_FOLIO(!success, folio); | |
4707 | ||
4708 | return true; | |
4709 | } | |
4710 | ||
4711 | static int scan_folios(struct lruvec *lruvec, struct scan_control *sc, | |
4712 | int type, int tier, struct list_head *list) | |
4713 | { | |
4714 | int gen, zone; | |
4715 | enum vm_event_item item; | |
4716 | int sorted = 0; | |
4717 | int scanned = 0; | |
4718 | int isolated = 0; | |
4719 | int remaining = MAX_LRU_BATCH; | |
4720 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
4721 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
4722 | ||
4723 | VM_WARN_ON_ONCE(!list_empty(list)); | |
4724 | ||
4725 | if (get_nr_gens(lruvec, type) == MIN_NR_GENS) | |
4726 | return 0; | |
4727 | ||
4728 | gen = lru_gen_from_seq(lrugen->min_seq[type]); | |
4729 | ||
4730 | for (zone = sc->reclaim_idx; zone >= 0; zone--) { | |
4731 | LIST_HEAD(moved); | |
4732 | int skipped = 0; | |
4733 | struct list_head *head = &lrugen->lists[gen][type][zone]; | |
4734 | ||
4735 | while (!list_empty(head)) { | |
4736 | struct folio *folio = lru_to_folio(head); | |
4737 | int delta = folio_nr_pages(folio); | |
4738 | ||
4739 | VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio); | |
4740 | VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio); | |
4741 | VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio); | |
4742 | VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio); | |
4743 | ||
4744 | scanned += delta; | |
4745 | ||
4746 | if (sort_folio(lruvec, folio, tier)) | |
4747 | sorted += delta; | |
4748 | else if (isolate_folio(lruvec, folio, sc)) { | |
4749 | list_add(&folio->lru, list); | |
4750 | isolated += delta; | |
4751 | } else { | |
4752 | list_move(&folio->lru, &moved); | |
4753 | skipped += delta; | |
4754 | } | |
4755 | ||
4756 | if (!--remaining || max(isolated, skipped) >= MIN_LRU_BATCH) | |
4757 | break; | |
4758 | } | |
4759 | ||
4760 | if (skipped) { | |
4761 | list_splice(&moved, head); | |
4762 | __count_zid_vm_events(PGSCAN_SKIP, zone, skipped); | |
4763 | } | |
4764 | ||
4765 | if (!remaining || isolated >= MIN_LRU_BATCH) | |
4766 | break; | |
4767 | } | |
4768 | ||
4769 | item = current_is_kswapd() ? PGSCAN_KSWAPD : PGSCAN_DIRECT; | |
4770 | if (!cgroup_reclaim(sc)) { | |
4771 | __count_vm_events(item, isolated); | |
4772 | __count_vm_events(PGREFILL, sorted); | |
4773 | } | |
4774 | __count_memcg_events(memcg, item, isolated); | |
4775 | __count_memcg_events(memcg, PGREFILL, sorted); | |
4776 | __count_vm_events(PGSCAN_ANON + type, isolated); | |
4777 | ||
4778 | /* | |
4779 | * There might not be eligible pages due to reclaim_idx, may_unmap and | |
4780 | * may_writepage. Check the remaining to prevent livelock if it's not | |
4781 | * making progress. | |
4782 | */ | |
4783 | return isolated || !remaining ? scanned : 0; | |
4784 | } | |
4785 | ||
4786 | static int get_tier_idx(struct lruvec *lruvec, int type) | |
4787 | { | |
4788 | int tier; | |
4789 | struct ctrl_pos sp, pv; | |
4790 | ||
4791 | /* | |
4792 | * To leave a margin for fluctuations, use a larger gain factor (1:2). | |
4793 | * This value is chosen because any other tier would have at least twice | |
4794 | * as many refaults as the first tier. | |
4795 | */ | |
4796 | read_ctrl_pos(lruvec, type, 0, 1, &sp); | |
4797 | for (tier = 1; tier < MAX_NR_TIERS; tier++) { | |
4798 | read_ctrl_pos(lruvec, type, tier, 2, &pv); | |
4799 | if (!positive_ctrl_err(&sp, &pv)) | |
4800 | break; | |
4801 | } | |
4802 | ||
4803 | return tier - 1; | |
4804 | } | |
4805 | ||
4806 | static int get_type_to_scan(struct lruvec *lruvec, int swappiness, int *tier_idx) | |
4807 | { | |
4808 | int type, tier; | |
4809 | struct ctrl_pos sp, pv; | |
4810 | int gain[ANON_AND_FILE] = { swappiness, 200 - swappiness }; | |
4811 | ||
4812 | /* | |
4813 | * Compare the first tier of anon with that of file to determine which | |
4814 | * type to scan. Also need to compare other tiers of the selected type | |
4815 | * with the first tier of the other type to determine the last tier (of | |
4816 | * the selected type) to evict. | |
4817 | */ | |
4818 | read_ctrl_pos(lruvec, LRU_GEN_ANON, 0, gain[LRU_GEN_ANON], &sp); | |
4819 | read_ctrl_pos(lruvec, LRU_GEN_FILE, 0, gain[LRU_GEN_FILE], &pv); | |
4820 | type = positive_ctrl_err(&sp, &pv); | |
4821 | ||
4822 | read_ctrl_pos(lruvec, !type, 0, gain[!type], &sp); | |
4823 | for (tier = 1; tier < MAX_NR_TIERS; tier++) { | |
4824 | read_ctrl_pos(lruvec, type, tier, gain[type], &pv); | |
4825 | if (!positive_ctrl_err(&sp, &pv)) | |
4826 | break; | |
4827 | } | |
4828 | ||
4829 | *tier_idx = tier - 1; | |
4830 | ||
4831 | return type; | |
4832 | } | |
4833 | ||
4834 | static int isolate_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness, | |
4835 | int *type_scanned, struct list_head *list) | |
4836 | { | |
4837 | int i; | |
4838 | int type; | |
4839 | int scanned; | |
4840 | int tier = -1; | |
4841 | DEFINE_MIN_SEQ(lruvec); | |
4842 | ||
4843 | /* | |
4844 | * Try to make the obvious choice first. When anon and file are both | |
4845 | * available from the same generation, interpret swappiness 1 as file | |
4846 | * first and 200 as anon first. | |
4847 | */ | |
4848 | if (!swappiness) | |
4849 | type = LRU_GEN_FILE; | |
4850 | else if (min_seq[LRU_GEN_ANON] < min_seq[LRU_GEN_FILE]) | |
4851 | type = LRU_GEN_ANON; | |
4852 | else if (swappiness == 1) | |
4853 | type = LRU_GEN_FILE; | |
4854 | else if (swappiness == 200) | |
4855 | type = LRU_GEN_ANON; | |
4856 | else | |
4857 | type = get_type_to_scan(lruvec, swappiness, &tier); | |
4858 | ||
4859 | for (i = !swappiness; i < ANON_AND_FILE; i++) { | |
4860 | if (tier < 0) | |
4861 | tier = get_tier_idx(lruvec, type); | |
4862 | ||
4863 | scanned = scan_folios(lruvec, sc, type, tier, list); | |
4864 | if (scanned) | |
4865 | break; | |
4866 | ||
4867 | type = !type; | |
4868 | tier = -1; | |
4869 | } | |
4870 | ||
4871 | *type_scanned = type; | |
4872 | ||
4873 | return scanned; | |
4874 | } | |
4875 | ||
f76c8337 YZ |
4876 | static int evict_folios(struct lruvec *lruvec, struct scan_control *sc, int swappiness, |
4877 | bool *need_swapping) | |
ac35a490 YZ |
4878 | { |
4879 | int type; | |
4880 | int scanned; | |
4881 | int reclaimed; | |
4882 | LIST_HEAD(list); | |
4883 | struct folio *folio; | |
4884 | enum vm_event_item item; | |
4885 | struct reclaim_stat stat; | |
bd74fdae | 4886 | struct lru_gen_mm_walk *walk; |
ac35a490 YZ |
4887 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); |
4888 | struct pglist_data *pgdat = lruvec_pgdat(lruvec); | |
4889 | ||
4890 | spin_lock_irq(&lruvec->lru_lock); | |
4891 | ||
4892 | scanned = isolate_folios(lruvec, sc, swappiness, &type, &list); | |
4893 | ||
4894 | scanned += try_to_inc_min_seq(lruvec, swappiness); | |
4895 | ||
4896 | if (get_nr_gens(lruvec, !swappiness) == MIN_NR_GENS) | |
4897 | scanned = 0; | |
4898 | ||
4899 | spin_unlock_irq(&lruvec->lru_lock); | |
4900 | ||
4901 | if (list_empty(&list)) | |
4902 | return scanned; | |
4903 | ||
4904 | reclaimed = shrink_page_list(&list, pgdat, sc, &stat, false); | |
4905 | ||
4906 | list_for_each_entry(folio, &list, lru) { | |
4907 | /* restore LRU_REFS_FLAGS cleared by isolate_folio() */ | |
4908 | if (folio_test_workingset(folio)) | |
4909 | folio_set_referenced(folio); | |
4910 | ||
4911 | /* don't add rejected pages to the oldest generation */ | |
4912 | if (folio_test_reclaim(folio) && | |
4913 | (folio_test_dirty(folio) || folio_test_writeback(folio))) | |
4914 | folio_clear_active(folio); | |
4915 | else | |
4916 | folio_set_active(folio); | |
4917 | } | |
4918 | ||
4919 | spin_lock_irq(&lruvec->lru_lock); | |
4920 | ||
4921 | move_pages_to_lru(lruvec, &list); | |
4922 | ||
bd74fdae YZ |
4923 | walk = current->reclaim_state->mm_walk; |
4924 | if (walk && walk->batched) | |
4925 | reset_batch_size(lruvec, walk); | |
4926 | ||
ac35a490 YZ |
4927 | item = current_is_kswapd() ? PGSTEAL_KSWAPD : PGSTEAL_DIRECT; |
4928 | if (!cgroup_reclaim(sc)) | |
4929 | __count_vm_events(item, reclaimed); | |
4930 | __count_memcg_events(memcg, item, reclaimed); | |
4931 | __count_vm_events(PGSTEAL_ANON + type, reclaimed); | |
4932 | ||
4933 | spin_unlock_irq(&lruvec->lru_lock); | |
4934 | ||
4935 | mem_cgroup_uncharge_list(&list); | |
4936 | free_unref_page_list(&list); | |
4937 | ||
4938 | sc->nr_reclaimed += reclaimed; | |
4939 | ||
f76c8337 YZ |
4940 | if (need_swapping && type == LRU_GEN_ANON) |
4941 | *need_swapping = true; | |
4942 | ||
ac35a490 YZ |
4943 | return scanned; |
4944 | } | |
4945 | ||
bd74fdae YZ |
4946 | /* |
4947 | * For future optimizations: | |
4948 | * 1. Defer try_to_inc_max_seq() to workqueues to reduce latency for memcg | |
4949 | * reclaim. | |
4950 | */ | |
ac35a490 | 4951 | static unsigned long get_nr_to_scan(struct lruvec *lruvec, struct scan_control *sc, |
f76c8337 | 4952 | bool can_swap, bool *need_aging) |
ac35a490 | 4953 | { |
ac35a490 YZ |
4954 | unsigned long nr_to_scan; |
4955 | struct mem_cgroup *memcg = lruvec_memcg(lruvec); | |
4956 | DEFINE_MAX_SEQ(lruvec); | |
4957 | DEFINE_MIN_SEQ(lruvec); | |
4958 | ||
4959 | if (mem_cgroup_below_min(memcg) || | |
4960 | (mem_cgroup_below_low(memcg) && !sc->memcg_low_reclaim)) | |
4961 | return 0; | |
4962 | ||
f76c8337 YZ |
4963 | *need_aging = should_run_aging(lruvec, max_seq, min_seq, sc, can_swap, &nr_to_scan); |
4964 | if (!*need_aging) | |
ac35a490 YZ |
4965 | return nr_to_scan; |
4966 | ||
4967 | /* skip the aging path at the default priority */ | |
4968 | if (sc->priority == DEF_PRIORITY) | |
4969 | goto done; | |
4970 | ||
4971 | /* leave the work to lru_gen_age_node() */ | |
4972 | if (current_is_kswapd()) | |
4973 | return 0; | |
4974 | ||
bd74fdae YZ |
4975 | if (try_to_inc_max_seq(lruvec, max_seq, sc, can_swap)) |
4976 | return nr_to_scan; | |
ac35a490 YZ |
4977 | done: |
4978 | return min_seq[!can_swap] + MIN_NR_GENS <= max_seq ? nr_to_scan : 0; | |
4979 | } | |
4980 | ||
f76c8337 YZ |
4981 | static bool should_abort_scan(struct lruvec *lruvec, unsigned long seq, |
4982 | struct scan_control *sc, bool need_swapping) | |
4983 | { | |
4984 | int i; | |
4985 | DEFINE_MAX_SEQ(lruvec); | |
4986 | ||
4987 | if (!current_is_kswapd()) { | |
4988 | /* age each memcg once to ensure fairness */ | |
4989 | if (max_seq - seq > 1) | |
4990 | return true; | |
4991 | ||
4992 | /* over-swapping can increase allocation latency */ | |
4993 | if (sc->nr_reclaimed >= sc->nr_to_reclaim && need_swapping) | |
4994 | return true; | |
4995 | ||
4996 | /* give this thread a chance to exit and free its memory */ | |
4997 | if (fatal_signal_pending(current)) { | |
4998 | sc->nr_reclaimed += MIN_LRU_BATCH; | |
4999 | return true; | |
5000 | } | |
5001 | ||
5002 | if (cgroup_reclaim(sc)) | |
5003 | return false; | |
5004 | } else if (sc->nr_reclaimed - sc->last_reclaimed < sc->nr_to_reclaim) | |
5005 | return false; | |
5006 | ||
5007 | /* keep scanning at low priorities to ensure fairness */ | |
5008 | if (sc->priority > DEF_PRIORITY - 2) | |
5009 | return false; | |
5010 | ||
5011 | /* | |
5012 | * A minimum amount of work was done under global memory pressure. For | |
5013 | * kswapd, it may be overshooting. For direct reclaim, the target isn't | |
5014 | * met, and yet the allocation may still succeed, since kswapd may have | |
5015 | * caught up. In either case, it's better to stop now, and restart if | |
5016 | * necessary. | |
5017 | */ | |
5018 | for (i = 0; i <= sc->reclaim_idx; i++) { | |
5019 | unsigned long wmark; | |
5020 | struct zone *zone = lruvec_pgdat(lruvec)->node_zones + i; | |
5021 | ||
5022 | if (!managed_zone(zone)) | |
5023 | continue; | |
5024 | ||
5025 | wmark = current_is_kswapd() ? high_wmark_pages(zone) : low_wmark_pages(zone); | |
5026 | if (wmark > zone_page_state(zone, NR_FREE_PAGES)) | |
5027 | return false; | |
5028 | } | |
5029 | ||
5030 | sc->nr_reclaimed += MIN_LRU_BATCH; | |
5031 | ||
5032 | return true; | |
5033 | } | |
5034 | ||
ac35a490 YZ |
5035 | static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) |
5036 | { | |
5037 | struct blk_plug plug; | |
f76c8337 YZ |
5038 | bool need_aging = false; |
5039 | bool need_swapping = false; | |
ac35a490 | 5040 | unsigned long scanned = 0; |
f76c8337 YZ |
5041 | unsigned long reclaimed = sc->nr_reclaimed; |
5042 | DEFINE_MAX_SEQ(lruvec); | |
ac35a490 YZ |
5043 | |
5044 | lru_add_drain(); | |
5045 | ||
5046 | blk_start_plug(&plug); | |
5047 | ||
bd74fdae YZ |
5048 | set_mm_walk(lruvec_pgdat(lruvec)); |
5049 | ||
ac35a490 YZ |
5050 | while (true) { |
5051 | int delta; | |
5052 | int swappiness; | |
5053 | unsigned long nr_to_scan; | |
5054 | ||
5055 | if (sc->may_swap) | |
5056 | swappiness = get_swappiness(lruvec, sc); | |
5057 | else if (!cgroup_reclaim(sc) && get_swappiness(lruvec, sc)) | |
5058 | swappiness = 1; | |
5059 | else | |
5060 | swappiness = 0; | |
5061 | ||
f76c8337 | 5062 | nr_to_scan = get_nr_to_scan(lruvec, sc, swappiness, &need_aging); |
ac35a490 | 5063 | if (!nr_to_scan) |
f76c8337 | 5064 | goto done; |
ac35a490 | 5065 | |
f76c8337 | 5066 | delta = evict_folios(lruvec, sc, swappiness, &need_swapping); |
ac35a490 | 5067 | if (!delta) |
f76c8337 | 5068 | goto done; |
ac35a490 YZ |
5069 | |
5070 | scanned += delta; | |
5071 | if (scanned >= nr_to_scan) | |
5072 | break; | |
5073 | ||
f76c8337 YZ |
5074 | if (should_abort_scan(lruvec, max_seq, sc, need_swapping)) |
5075 | break; | |
5076 | ||
ac35a490 YZ |
5077 | cond_resched(); |
5078 | } | |
5079 | ||
f76c8337 YZ |
5080 | /* see the comment in lru_gen_age_node() */ |
5081 | if (sc->nr_reclaimed - reclaimed >= MIN_LRU_BATCH && !need_aging) | |
5082 | sc->memcgs_need_aging = false; | |
5083 | done: | |
bd74fdae YZ |
5084 | clear_mm_walk(); |
5085 | ||
ac35a490 YZ |
5086 | blk_finish_plug(&plug); |
5087 | } | |
5088 | ||
354ed597 YZ |
5089 | /****************************************************************************** |
5090 | * state change | |
5091 | ******************************************************************************/ | |
5092 | ||
5093 | static bool __maybe_unused state_is_valid(struct lruvec *lruvec) | |
5094 | { | |
5095 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
5096 | ||
5097 | if (lrugen->enabled) { | |
5098 | enum lru_list lru; | |
5099 | ||
5100 | for_each_evictable_lru(lru) { | |
5101 | if (!list_empty(&lruvec->lists[lru])) | |
5102 | return false; | |
5103 | } | |
5104 | } else { | |
5105 | int gen, type, zone; | |
5106 | ||
5107 | for_each_gen_type_zone(gen, type, zone) { | |
5108 | if (!list_empty(&lrugen->lists[gen][type][zone])) | |
5109 | return false; | |
5110 | } | |
5111 | } | |
5112 | ||
5113 | return true; | |
5114 | } | |
5115 | ||
5116 | static bool fill_evictable(struct lruvec *lruvec) | |
5117 | { | |
5118 | enum lru_list lru; | |
5119 | int remaining = MAX_LRU_BATCH; | |
5120 | ||
5121 | for_each_evictable_lru(lru) { | |
5122 | int type = is_file_lru(lru); | |
5123 | bool active = is_active_lru(lru); | |
5124 | struct list_head *head = &lruvec->lists[lru]; | |
5125 | ||
5126 | while (!list_empty(head)) { | |
5127 | bool success; | |
5128 | struct folio *folio = lru_to_folio(head); | |
5129 | ||
5130 | VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio); | |
5131 | VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio) != active, folio); | |
5132 | VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio); | |
5133 | VM_WARN_ON_ONCE_FOLIO(folio_lru_gen(folio) != -1, folio); | |
5134 | ||
5135 | lruvec_del_folio(lruvec, folio); | |
5136 | success = lru_gen_add_folio(lruvec, folio, false); | |
5137 | VM_WARN_ON_ONCE(!success); | |
5138 | ||
5139 | if (!--remaining) | |
5140 | return false; | |
5141 | } | |
5142 | } | |
5143 | ||
5144 | return true; | |
5145 | } | |
5146 | ||
5147 | static bool drain_evictable(struct lruvec *lruvec) | |
5148 | { | |
5149 | int gen, type, zone; | |
5150 | int remaining = MAX_LRU_BATCH; | |
5151 | ||
5152 | for_each_gen_type_zone(gen, type, zone) { | |
5153 | struct list_head *head = &lruvec->lrugen.lists[gen][type][zone]; | |
5154 | ||
5155 | while (!list_empty(head)) { | |
5156 | bool success; | |
5157 | struct folio *folio = lru_to_folio(head); | |
5158 | ||
5159 | VM_WARN_ON_ONCE_FOLIO(folio_test_unevictable(folio), folio); | |
5160 | VM_WARN_ON_ONCE_FOLIO(folio_test_active(folio), folio); | |
5161 | VM_WARN_ON_ONCE_FOLIO(folio_is_file_lru(folio) != type, folio); | |
5162 | VM_WARN_ON_ONCE_FOLIO(folio_zonenum(folio) != zone, folio); | |
5163 | ||
5164 | success = lru_gen_del_folio(lruvec, folio, false); | |
5165 | VM_WARN_ON_ONCE(!success); | |
5166 | lruvec_add_folio(lruvec, folio); | |
5167 | ||
5168 | if (!--remaining) | |
5169 | return false; | |
5170 | } | |
5171 | } | |
5172 | ||
5173 | return true; | |
5174 | } | |
5175 | ||
5176 | static void lru_gen_change_state(bool enabled) | |
5177 | { | |
5178 | static DEFINE_MUTEX(state_mutex); | |
5179 | ||
5180 | struct mem_cgroup *memcg; | |
5181 | ||
5182 | cgroup_lock(); | |
5183 | cpus_read_lock(); | |
5184 | get_online_mems(); | |
5185 | mutex_lock(&state_mutex); | |
5186 | ||
5187 | if (enabled == lru_gen_enabled()) | |
5188 | goto unlock; | |
5189 | ||
5190 | if (enabled) | |
5191 | static_branch_enable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]); | |
5192 | else | |
5193 | static_branch_disable_cpuslocked(&lru_gen_caps[LRU_GEN_CORE]); | |
5194 | ||
5195 | memcg = mem_cgroup_iter(NULL, NULL, NULL); | |
5196 | do { | |
5197 | int nid; | |
5198 | ||
5199 | for_each_node(nid) { | |
5200 | struct lruvec *lruvec = get_lruvec(memcg, nid); | |
5201 | ||
5202 | if (!lruvec) | |
5203 | continue; | |
5204 | ||
5205 | spin_lock_irq(&lruvec->lru_lock); | |
5206 | ||
5207 | VM_WARN_ON_ONCE(!seq_is_valid(lruvec)); | |
5208 | VM_WARN_ON_ONCE(!state_is_valid(lruvec)); | |
5209 | ||
5210 | lruvec->lrugen.enabled = enabled; | |
5211 | ||
5212 | while (!(enabled ? fill_evictable(lruvec) : drain_evictable(lruvec))) { | |
5213 | spin_unlock_irq(&lruvec->lru_lock); | |
5214 | cond_resched(); | |
5215 | spin_lock_irq(&lruvec->lru_lock); | |
5216 | } | |
5217 | ||
5218 | spin_unlock_irq(&lruvec->lru_lock); | |
5219 | } | |
5220 | ||
5221 | cond_resched(); | |
5222 | } while ((memcg = mem_cgroup_iter(NULL, memcg, NULL))); | |
5223 | unlock: | |
5224 | mutex_unlock(&state_mutex); | |
5225 | put_online_mems(); | |
5226 | cpus_read_unlock(); | |
5227 | cgroup_unlock(); | |
5228 | } | |
5229 | ||
5230 | /****************************************************************************** | |
5231 | * sysfs interface | |
5232 | ******************************************************************************/ | |
5233 | ||
5234 | static ssize_t show_enabled(struct kobject *kobj, struct kobj_attribute *attr, char *buf) | |
5235 | { | |
5236 | unsigned int caps = 0; | |
5237 | ||
5238 | if (get_cap(LRU_GEN_CORE)) | |
5239 | caps |= BIT(LRU_GEN_CORE); | |
5240 | ||
5241 | if (arch_has_hw_pte_young() && get_cap(LRU_GEN_MM_WALK)) | |
5242 | caps |= BIT(LRU_GEN_MM_WALK); | |
5243 | ||
5244 | if (IS_ENABLED(CONFIG_ARCH_HAS_NONLEAF_PMD_YOUNG) && get_cap(LRU_GEN_NONLEAF_YOUNG)) | |
5245 | caps |= BIT(LRU_GEN_NONLEAF_YOUNG); | |
5246 | ||
5247 | return snprintf(buf, PAGE_SIZE, "0x%04x\n", caps); | |
5248 | } | |
5249 | ||
5250 | static ssize_t store_enabled(struct kobject *kobj, struct kobj_attribute *attr, | |
5251 | const char *buf, size_t len) | |
5252 | { | |
5253 | int i; | |
5254 | unsigned int caps; | |
5255 | ||
5256 | if (tolower(*buf) == 'n') | |
5257 | caps = 0; | |
5258 | else if (tolower(*buf) == 'y') | |
5259 | caps = -1; | |
5260 | else if (kstrtouint(buf, 0, &caps)) | |
5261 | return -EINVAL; | |
5262 | ||
5263 | for (i = 0; i < NR_LRU_GEN_CAPS; i++) { | |
5264 | bool enabled = caps & BIT(i); | |
5265 | ||
5266 | if (i == LRU_GEN_CORE) | |
5267 | lru_gen_change_state(enabled); | |
5268 | else if (enabled) | |
5269 | static_branch_enable(&lru_gen_caps[i]); | |
5270 | else | |
5271 | static_branch_disable(&lru_gen_caps[i]); | |
5272 | } | |
5273 | ||
5274 | return len; | |
5275 | } | |
5276 | ||
5277 | static struct kobj_attribute lru_gen_enabled_attr = __ATTR( | |
5278 | enabled, 0644, show_enabled, store_enabled | |
5279 | ); | |
5280 | ||
5281 | static struct attribute *lru_gen_attrs[] = { | |
5282 | &lru_gen_enabled_attr.attr, | |
5283 | NULL | |
5284 | }; | |
5285 | ||
5286 | static struct attribute_group lru_gen_attr_group = { | |
5287 | .name = "lru_gen", | |
5288 | .attrs = lru_gen_attrs, | |
5289 | }; | |
5290 | ||
ec1c86b2 YZ |
5291 | /****************************************************************************** |
5292 | * initialization | |
5293 | ******************************************************************************/ | |
5294 | ||
5295 | void lru_gen_init_lruvec(struct lruvec *lruvec) | |
5296 | { | |
5297 | int gen, type, zone; | |
5298 | struct lru_gen_struct *lrugen = &lruvec->lrugen; | |
5299 | ||
5300 | lrugen->max_seq = MIN_NR_GENS + 1; | |
354ed597 | 5301 | lrugen->enabled = lru_gen_enabled(); |
ec1c86b2 YZ |
5302 | |
5303 | for_each_gen_type_zone(gen, type, zone) | |
5304 | INIT_LIST_HEAD(&lrugen->lists[gen][type][zone]); | |
bd74fdae YZ |
5305 | |
5306 | lruvec->mm_state.seq = MIN_NR_GENS; | |
5307 | init_waitqueue_head(&lruvec->mm_state.wait); | |
ec1c86b2 YZ |
5308 | } |
5309 | ||
5310 | #ifdef CONFIG_MEMCG | |
5311 | void lru_gen_init_memcg(struct mem_cgroup *memcg) | |
5312 | { | |
bd74fdae YZ |
5313 | INIT_LIST_HEAD(&memcg->mm_list.fifo); |
5314 | spin_lock_init(&memcg->mm_list.lock); | |
ec1c86b2 YZ |
5315 | } |
5316 | ||
5317 | void lru_gen_exit_memcg(struct mem_cgroup *memcg) | |
5318 | { | |
bd74fdae | 5319 | int i; |
ec1c86b2 YZ |
5320 | int nid; |
5321 | ||
5322 | for_each_node(nid) { | |
5323 | struct lruvec *lruvec = get_lruvec(memcg, nid); | |
5324 | ||
5325 | VM_WARN_ON_ONCE(memchr_inv(lruvec->lrugen.nr_pages, 0, | |
5326 | sizeof(lruvec->lrugen.nr_pages))); | |
bd74fdae YZ |
5327 | |
5328 | for (i = 0; i < NR_BLOOM_FILTERS; i++) { | |
5329 | bitmap_free(lruvec->mm_state.filters[i]); | |
5330 | lruvec->mm_state.filters[i] = NULL; | |
5331 | } | |
ec1c86b2 YZ |
5332 | } |
5333 | } | |
5334 | #endif | |
5335 | ||
5336 | static int __init init_lru_gen(void) | |
5337 | { | |
5338 | BUILD_BUG_ON(MIN_NR_GENS + 1 >= MAX_NR_GENS); | |
5339 | BUILD_BUG_ON(BIT(LRU_GEN_WIDTH) <= MAX_NR_GENS); | |
5340 | ||
354ed597 YZ |
5341 | if (sysfs_create_group(mm_kobj, &lru_gen_attr_group)) |
5342 | pr_err("lru_gen: failed to create sysfs group\n"); | |
5343 | ||
ec1c86b2 YZ |
5344 | return 0; |
5345 | }; | |
5346 | late_initcall(init_lru_gen); | |
5347 | ||
ac35a490 YZ |
5348 | #else /* !CONFIG_LRU_GEN */ |
5349 | ||
5350 | static void lru_gen_age_node(struct pglist_data *pgdat, struct scan_control *sc) | |
5351 | { | |
5352 | } | |
5353 | ||
5354 | static void lru_gen_shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) | |
5355 | { | |
5356 | } | |
5357 | ||
ec1c86b2 YZ |
5358 | #endif /* CONFIG_LRU_GEN */ |
5359 | ||
afaf07a6 | 5360 | static void shrink_lruvec(struct lruvec *lruvec, struct scan_control *sc) |
9b4f98cd JW |
5361 | { |
5362 | unsigned long nr[NR_LRU_LISTS]; | |
e82e0561 | 5363 | unsigned long targets[NR_LRU_LISTS]; |
9b4f98cd JW |
5364 | unsigned long nr_to_scan; |
5365 | enum lru_list lru; | |
5366 | unsigned long nr_reclaimed = 0; | |
5367 | unsigned long nr_to_reclaim = sc->nr_to_reclaim; | |
5368 | struct blk_plug plug; | |
1a501907 | 5369 | bool scan_adjusted; |
9b4f98cd | 5370 | |
ac35a490 YZ |
5371 | if (lru_gen_enabled()) { |
5372 | lru_gen_shrink_lruvec(lruvec, sc); | |
5373 | return; | |
5374 | } | |
5375 | ||
afaf07a6 | 5376 | get_scan_count(lruvec, sc, nr); |
9b4f98cd | 5377 | |
e82e0561 MG |
5378 | /* Record the original scan target for proportional adjustments later */ |
5379 | memcpy(targets, nr, sizeof(nr)); | |
5380 | ||
1a501907 MG |
5381 | /* |
5382 | * Global reclaiming within direct reclaim at DEF_PRIORITY is a normal | |
5383 | * event that can occur when there is little memory pressure e.g. | |
5384 | * multiple streaming readers/writers. Hence, we do not abort scanning | |
5385 | * when the requested number of pages are reclaimed when scanning at | |
5386 | * DEF_PRIORITY on the assumption that the fact we are direct | |
5387 | * reclaiming implies that kswapd is not keeping up and it is best to | |
5388 | * do a batch of work at once. For memcg reclaim one check is made to | |
5389 | * abort proportional reclaim if either the file or anon lru has already | |
5390 | * dropped to zero at the first pass. | |
5391 | */ | |
b5ead35e | 5392 | scan_adjusted = (!cgroup_reclaim(sc) && !current_is_kswapd() && |
1a501907 MG |
5393 | sc->priority == DEF_PRIORITY); |
5394 | ||
9b4f98cd JW |
5395 | blk_start_plug(&plug); |
5396 | while (nr[LRU_INACTIVE_ANON] || nr[LRU_ACTIVE_FILE] || | |
5397 | nr[LRU_INACTIVE_FILE]) { | |
e82e0561 MG |
5398 | unsigned long nr_anon, nr_file, percentage; |
5399 | unsigned long nr_scanned; | |
5400 | ||
9b4f98cd JW |
5401 | for_each_evictable_lru(lru) { |
5402 | if (nr[lru]) { | |
5403 | nr_to_scan = min(nr[lru], SWAP_CLUSTER_MAX); | |
5404 | nr[lru] -= nr_to_scan; | |
5405 | ||
5406 | nr_reclaimed += shrink_list(lru, nr_to_scan, | |
3b991208 | 5407 | lruvec, sc); |
9b4f98cd JW |
5408 | } |
5409 | } | |
e82e0561 | 5410 | |
bd041733 MH |
5411 | cond_resched(); |
5412 | ||
e82e0561 MG |
5413 | if (nr_reclaimed < nr_to_reclaim || scan_adjusted) |
5414 | continue; | |
5415 | ||
e82e0561 MG |
5416 | /* |
5417 | * For kswapd and memcg, reclaim at least the number of pages | |
1a501907 | 5418 | * requested. Ensure that the anon and file LRUs are scanned |
e82e0561 MG |
5419 | * proportionally what was requested by get_scan_count(). We |
5420 | * stop reclaiming one LRU and reduce the amount scanning | |
5421 | * proportional to the original scan target. | |
5422 | */ | |
5423 | nr_file = nr[LRU_INACTIVE_FILE] + nr[LRU_ACTIVE_FILE]; | |
5424 | nr_anon = nr[LRU_INACTIVE_ANON] + nr[LRU_ACTIVE_ANON]; | |
5425 | ||
1a501907 MG |
5426 | /* |
5427 | * It's just vindictive to attack the larger once the smaller | |
5428 | * has gone to zero. And given the way we stop scanning the | |
5429 | * smaller below, this makes sure that we only make one nudge | |
5430 | * towards proportionality once we've got nr_to_reclaim. | |
5431 | */ | |
5432 | if (!nr_file || !nr_anon) | |
5433 | break; | |
5434 | ||
e82e0561 MG |
5435 | if (nr_file > nr_anon) { |
5436 | unsigned long scan_target = targets[LRU_INACTIVE_ANON] + | |
5437 | targets[LRU_ACTIVE_ANON] + 1; | |
5438 | lru = LRU_BASE; | |
5439 | percentage = nr_anon * 100 / scan_target; | |
5440 | } else { | |
5441 | unsigned long scan_target = targets[LRU_INACTIVE_FILE] + | |
5442 | targets[LRU_ACTIVE_FILE] + 1; | |
5443 | lru = LRU_FILE; | |
5444 | percentage = nr_file * 100 / scan_target; | |
5445 | } | |
5446 | ||
5447 | /* Stop scanning the smaller of the LRU */ | |
5448 | nr[lru] = 0; | |
5449 | nr[lru + LRU_ACTIVE] = 0; | |
5450 | ||
5451 | /* | |
5452 | * Recalculate the other LRU scan count based on its original | |
5453 | * scan target and the percentage scanning already complete | |
5454 | */ | |
5455 | lru = (lru == LRU_FILE) ? LRU_BASE : LRU_FILE; | |
5456 | nr_scanned = targets[lru] - nr[lru]; | |
5457 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
5458 | nr[lru] -= min(nr[lru], nr_scanned); | |
5459 | ||
5460 | lru += LRU_ACTIVE; | |
5461 | nr_scanned = targets[lru] - nr[lru]; | |
5462 | nr[lru] = targets[lru] * (100 - percentage) / 100; | |
5463 | nr[lru] -= min(nr[lru], nr_scanned); | |
5464 | ||
5465 | scan_adjusted = true; | |
9b4f98cd JW |
5466 | } |
5467 | blk_finish_plug(&plug); | |
5468 | sc->nr_reclaimed += nr_reclaimed; | |
5469 | ||
5470 | /* | |
5471 | * Even if we did not try to evict anon pages at all, we want to | |
5472 | * rebalance the anon lru active/inactive ratio. | |
5473 | */ | |
2f368a9f DH |
5474 | if (can_age_anon_pages(lruvec_pgdat(lruvec), sc) && |
5475 | inactive_is_low(lruvec, LRU_INACTIVE_ANON)) | |
9b4f98cd JW |
5476 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, |
5477 | sc, LRU_ACTIVE_ANON); | |
9b4f98cd JW |
5478 | } |
5479 | ||
23b9da55 | 5480 | /* Use reclaim/compaction for costly allocs or under memory pressure */ |
9e3b2f8c | 5481 | static bool in_reclaim_compaction(struct scan_control *sc) |
23b9da55 | 5482 | { |
d84da3f9 | 5483 | if (IS_ENABLED(CONFIG_COMPACTION) && sc->order && |
23b9da55 | 5484 | (sc->order > PAGE_ALLOC_COSTLY_ORDER || |
9e3b2f8c | 5485 | sc->priority < DEF_PRIORITY - 2)) |
23b9da55 MG |
5486 | return true; |
5487 | ||
5488 | return false; | |
5489 | } | |
5490 | ||
3e7d3449 | 5491 | /* |
23b9da55 MG |
5492 | * Reclaim/compaction is used for high-order allocation requests. It reclaims |
5493 | * order-0 pages before compacting the zone. should_continue_reclaim() returns | |
5494 | * true if more pages should be reclaimed such that when the page allocator | |
df3a45f9 | 5495 | * calls try_to_compact_pages() that it will have enough free pages to succeed. |
23b9da55 | 5496 | * It will give up earlier than that if there is difficulty reclaiming pages. |
3e7d3449 | 5497 | */ |
a9dd0a83 | 5498 | static inline bool should_continue_reclaim(struct pglist_data *pgdat, |
3e7d3449 | 5499 | unsigned long nr_reclaimed, |
3e7d3449 MG |
5500 | struct scan_control *sc) |
5501 | { | |
5502 | unsigned long pages_for_compaction; | |
5503 | unsigned long inactive_lru_pages; | |
a9dd0a83 | 5504 | int z; |
3e7d3449 MG |
5505 | |
5506 | /* If not in reclaim/compaction mode, stop */ | |
9e3b2f8c | 5507 | if (!in_reclaim_compaction(sc)) |
3e7d3449 MG |
5508 | return false; |
5509 | ||
5ee04716 VB |
5510 | /* |
5511 | * Stop if we failed to reclaim any pages from the last SWAP_CLUSTER_MAX | |
5512 | * number of pages that were scanned. This will return to the caller | |
5513 | * with the risk reclaim/compaction and the resulting allocation attempt | |
5514 | * fails. In the past we have tried harder for __GFP_RETRY_MAYFAIL | |
5515 | * allocations through requiring that the full LRU list has been scanned | |
5516 | * first, by assuming that zero delta of sc->nr_scanned means full LRU | |
5517 | * scan, but that approximation was wrong, and there were corner cases | |
5518 | * where always a non-zero amount of pages were scanned. | |
5519 | */ | |
5520 | if (!nr_reclaimed) | |
5521 | return false; | |
3e7d3449 | 5522 | |
3e7d3449 | 5523 | /* If compaction would go ahead or the allocation would succeed, stop */ |
a9dd0a83 MG |
5524 | for (z = 0; z <= sc->reclaim_idx; z++) { |
5525 | struct zone *zone = &pgdat->node_zones[z]; | |
6aa303de | 5526 | if (!managed_zone(zone)) |
a9dd0a83 MG |
5527 | continue; |
5528 | ||
5529 | switch (compaction_suitable(zone, sc->order, 0, sc->reclaim_idx)) { | |
cf378319 | 5530 | case COMPACT_SUCCESS: |
a9dd0a83 MG |
5531 | case COMPACT_CONTINUE: |
5532 | return false; | |
5533 | default: | |
5534 | /* check next zone */ | |
5535 | ; | |
5536 | } | |
3e7d3449 | 5537 | } |
1c6c1597 HD |
5538 | |
5539 | /* | |
5540 | * If we have not reclaimed enough pages for compaction and the | |
5541 | * inactive lists are large enough, continue reclaiming | |
5542 | */ | |
5543 | pages_for_compaction = compact_gap(sc->order); | |
5544 | inactive_lru_pages = node_page_state(pgdat, NR_INACTIVE_FILE); | |
a2a36488 | 5545 | if (can_reclaim_anon_pages(NULL, pgdat->node_id, sc)) |
1c6c1597 HD |
5546 | inactive_lru_pages += node_page_state(pgdat, NR_INACTIVE_ANON); |
5547 | ||
5ee04716 | 5548 | return inactive_lru_pages > pages_for_compaction; |
3e7d3449 MG |
5549 | } |
5550 | ||
0f6a5cff | 5551 | static void shrink_node_memcgs(pg_data_t *pgdat, struct scan_control *sc) |
1da177e4 | 5552 | { |
0f6a5cff | 5553 | struct mem_cgroup *target_memcg = sc->target_mem_cgroup; |
d2af3397 | 5554 | struct mem_cgroup *memcg; |
1da177e4 | 5555 | |
0f6a5cff | 5556 | memcg = mem_cgroup_iter(target_memcg, NULL, NULL); |
d2af3397 | 5557 | do { |
afaf07a6 | 5558 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
d2af3397 JW |
5559 | unsigned long reclaimed; |
5560 | unsigned long scanned; | |
5660048c | 5561 | |
e3336cab XP |
5562 | /* |
5563 | * This loop can become CPU-bound when target memcgs | |
5564 | * aren't eligible for reclaim - either because they | |
5565 | * don't have any reclaimable pages, or because their | |
5566 | * memory is explicitly protected. Avoid soft lockups. | |
5567 | */ | |
5568 | cond_resched(); | |
5569 | ||
45c7f7e1 CD |
5570 | mem_cgroup_calculate_protection(target_memcg, memcg); |
5571 | ||
5572 | if (mem_cgroup_below_min(memcg)) { | |
d2af3397 JW |
5573 | /* |
5574 | * Hard protection. | |
5575 | * If there is no reclaimable memory, OOM. | |
5576 | */ | |
5577 | continue; | |
45c7f7e1 | 5578 | } else if (mem_cgroup_below_low(memcg)) { |
d2af3397 JW |
5579 | /* |
5580 | * Soft protection. | |
5581 | * Respect the protection only as long as | |
5582 | * there is an unprotected supply | |
5583 | * of reclaimable memory from other cgroups. | |
5584 | */ | |
5585 | if (!sc->memcg_low_reclaim) { | |
5586 | sc->memcg_low_skipped = 1; | |
bf8d5d52 | 5587 | continue; |
241994ed | 5588 | } |
d2af3397 | 5589 | memcg_memory_event(memcg, MEMCG_LOW); |
d2af3397 | 5590 | } |
241994ed | 5591 | |
d2af3397 JW |
5592 | reclaimed = sc->nr_reclaimed; |
5593 | scanned = sc->nr_scanned; | |
afaf07a6 JW |
5594 | |
5595 | shrink_lruvec(lruvec, sc); | |
70ddf637 | 5596 | |
d2af3397 JW |
5597 | shrink_slab(sc->gfp_mask, pgdat->node_id, memcg, |
5598 | sc->priority); | |
6b4f7799 | 5599 | |
d2af3397 | 5600 | /* Record the group's reclaim efficiency */ |
73b73bac YA |
5601 | if (!sc->proactive) |
5602 | vmpressure(sc->gfp_mask, memcg, false, | |
5603 | sc->nr_scanned - scanned, | |
5604 | sc->nr_reclaimed - reclaimed); | |
70ddf637 | 5605 | |
0f6a5cff JW |
5606 | } while ((memcg = mem_cgroup_iter(target_memcg, memcg, NULL))); |
5607 | } | |
5608 | ||
6c9e0907 | 5609 | static void shrink_node(pg_data_t *pgdat, struct scan_control *sc) |
0f6a5cff JW |
5610 | { |
5611 | struct reclaim_state *reclaim_state = current->reclaim_state; | |
0f6a5cff | 5612 | unsigned long nr_reclaimed, nr_scanned; |
1b05117d | 5613 | struct lruvec *target_lruvec; |
0f6a5cff JW |
5614 | bool reclaimable = false; |
5615 | ||
1b05117d JW |
5616 | target_lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, pgdat); |
5617 | ||
0f6a5cff JW |
5618 | again: |
5619 | memset(&sc->nr, 0, sizeof(sc->nr)); | |
5620 | ||
5621 | nr_reclaimed = sc->nr_reclaimed; | |
5622 | nr_scanned = sc->nr_scanned; | |
5623 | ||
f1e1a7be | 5624 | prepare_scan_count(pgdat, sc); |
53138cea | 5625 | |
0f6a5cff | 5626 | shrink_node_memcgs(pgdat, sc); |
2344d7e4 | 5627 | |
d2af3397 JW |
5628 | if (reclaim_state) { |
5629 | sc->nr_reclaimed += reclaim_state->reclaimed_slab; | |
5630 | reclaim_state->reclaimed_slab = 0; | |
5631 | } | |
d108c772 | 5632 | |
d2af3397 | 5633 | /* Record the subtree's reclaim efficiency */ |
73b73bac YA |
5634 | if (!sc->proactive) |
5635 | vmpressure(sc->gfp_mask, sc->target_mem_cgroup, true, | |
5636 | sc->nr_scanned - nr_scanned, | |
5637 | sc->nr_reclaimed - nr_reclaimed); | |
d108c772 | 5638 | |
d2af3397 JW |
5639 | if (sc->nr_reclaimed - nr_reclaimed) |
5640 | reclaimable = true; | |
d108c772 | 5641 | |
d2af3397 JW |
5642 | if (current_is_kswapd()) { |
5643 | /* | |
5644 | * If reclaim is isolating dirty pages under writeback, | |
5645 | * it implies that the long-lived page allocation rate | |
5646 | * is exceeding the page laundering rate. Either the | |
5647 | * global limits are not being effective at throttling | |
5648 | * processes due to the page distribution throughout | |
5649 | * zones or there is heavy usage of a slow backing | |
5650 | * device. The only option is to throttle from reclaim | |
5651 | * context which is not ideal as there is no guarantee | |
5652 | * the dirtying process is throttled in the same way | |
5653 | * balance_dirty_pages() manages. | |
5654 | * | |
5655 | * Once a node is flagged PGDAT_WRITEBACK, kswapd will | |
5656 | * count the number of pages under pages flagged for | |
5657 | * immediate reclaim and stall if any are encountered | |
5658 | * in the nr_immediate check below. | |
5659 | */ | |
5660 | if (sc->nr.writeback && sc->nr.writeback == sc->nr.taken) | |
5661 | set_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
d108c772 | 5662 | |
d2af3397 JW |
5663 | /* Allow kswapd to start writing pages during reclaim.*/ |
5664 | if (sc->nr.unqueued_dirty == sc->nr.file_taken) | |
5665 | set_bit(PGDAT_DIRTY, &pgdat->flags); | |
e3c1ac58 | 5666 | |
d108c772 | 5667 | /* |
1eba09c1 | 5668 | * If kswapd scans pages marked for immediate |
d2af3397 JW |
5669 | * reclaim and under writeback (nr_immediate), it |
5670 | * implies that pages are cycling through the LRU | |
8cd7c588 MG |
5671 | * faster than they are written so forcibly stall |
5672 | * until some pages complete writeback. | |
d108c772 | 5673 | */ |
d2af3397 | 5674 | if (sc->nr.immediate) |
c3f4a9a2 | 5675 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_WRITEBACK); |
d2af3397 JW |
5676 | } |
5677 | ||
5678 | /* | |
8cd7c588 MG |
5679 | * Tag a node/memcg as congested if all the dirty pages were marked |
5680 | * for writeback and immediate reclaim (counted in nr.congested). | |
1b05117d | 5681 | * |
d2af3397 | 5682 | * Legacy memcg will stall in page writeback so avoid forcibly |
8cd7c588 | 5683 | * stalling in reclaim_throttle(). |
d2af3397 | 5684 | */ |
1b05117d JW |
5685 | if ((current_is_kswapd() || |
5686 | (cgroup_reclaim(sc) && writeback_throttling_sane(sc))) && | |
d2af3397 | 5687 | sc->nr.dirty && sc->nr.dirty == sc->nr.congested) |
1b05117d | 5688 | set_bit(LRUVEC_CONGESTED, &target_lruvec->flags); |
d2af3397 JW |
5689 | |
5690 | /* | |
8cd7c588 MG |
5691 | * Stall direct reclaim for IO completions if the lruvec is |
5692 | * node is congested. Allow kswapd to continue until it | |
d2af3397 JW |
5693 | * starts encountering unqueued dirty pages or cycling through |
5694 | * the LRU too quickly. | |
5695 | */ | |
1b05117d JW |
5696 | if (!current_is_kswapd() && current_may_throttle() && |
5697 | !sc->hibernation_mode && | |
5698 | test_bit(LRUVEC_CONGESTED, &target_lruvec->flags)) | |
1b4e3f26 | 5699 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_CONGESTED); |
d108c772 | 5700 | |
d2af3397 JW |
5701 | if (should_continue_reclaim(pgdat, sc->nr_reclaimed - nr_reclaimed, |
5702 | sc)) | |
5703 | goto again; | |
2344d7e4 | 5704 | |
c73322d0 JW |
5705 | /* |
5706 | * Kswapd gives up on balancing particular nodes after too | |
5707 | * many failures to reclaim anything from them and goes to | |
5708 | * sleep. On reclaim progress, reset the failure counter. A | |
5709 | * successful direct reclaim run will revive a dormant kswapd. | |
5710 | */ | |
5711 | if (reclaimable) | |
5712 | pgdat->kswapd_failures = 0; | |
f16015fb JW |
5713 | } |
5714 | ||
53853e2d | 5715 | /* |
fdd4c614 VB |
5716 | * Returns true if compaction should go ahead for a costly-order request, or |
5717 | * the allocation would already succeed without compaction. Return false if we | |
5718 | * should reclaim first. | |
53853e2d | 5719 | */ |
4f588331 | 5720 | static inline bool compaction_ready(struct zone *zone, struct scan_control *sc) |
fe4b1b24 | 5721 | { |
31483b6a | 5722 | unsigned long watermark; |
fdd4c614 | 5723 | enum compact_result suitable; |
fe4b1b24 | 5724 | |
fdd4c614 VB |
5725 | suitable = compaction_suitable(zone, sc->order, 0, sc->reclaim_idx); |
5726 | if (suitable == COMPACT_SUCCESS) | |
5727 | /* Allocation should succeed already. Don't reclaim. */ | |
5728 | return true; | |
5729 | if (suitable == COMPACT_SKIPPED) | |
5730 | /* Compaction cannot yet proceed. Do reclaim. */ | |
5731 | return false; | |
fe4b1b24 | 5732 | |
53853e2d | 5733 | /* |
fdd4c614 VB |
5734 | * Compaction is already possible, but it takes time to run and there |
5735 | * are potentially other callers using the pages just freed. So proceed | |
5736 | * with reclaim to make a buffer of free pages available to give | |
5737 | * compaction a reasonable chance of completing and allocating the page. | |
5738 | * Note that we won't actually reclaim the whole buffer in one attempt | |
5739 | * as the target watermark in should_continue_reclaim() is lower. But if | |
5740 | * we are already above the high+gap watermark, don't reclaim at all. | |
53853e2d | 5741 | */ |
fdd4c614 | 5742 | watermark = high_wmark_pages(zone) + compact_gap(sc->order); |
fe4b1b24 | 5743 | |
fdd4c614 | 5744 | return zone_watermark_ok_safe(zone, 0, watermark, sc->reclaim_idx); |
fe4b1b24 MG |
5745 | } |
5746 | ||
69392a40 MG |
5747 | static void consider_reclaim_throttle(pg_data_t *pgdat, struct scan_control *sc) |
5748 | { | |
66ce520b MG |
5749 | /* |
5750 | * If reclaim is making progress greater than 12% efficiency then | |
5751 | * wake all the NOPROGRESS throttled tasks. | |
5752 | */ | |
5753 | if (sc->nr_reclaimed > (sc->nr_scanned >> 3)) { | |
69392a40 MG |
5754 | wait_queue_head_t *wqh; |
5755 | ||
5756 | wqh = &pgdat->reclaim_wait[VMSCAN_THROTTLE_NOPROGRESS]; | |
5757 | if (waitqueue_active(wqh)) | |
5758 | wake_up(wqh); | |
5759 | ||
5760 | return; | |
5761 | } | |
5762 | ||
5763 | /* | |
1b4e3f26 MG |
5764 | * Do not throttle kswapd or cgroup reclaim on NOPROGRESS as it will |
5765 | * throttle on VMSCAN_THROTTLE_WRITEBACK if there are too many pages | |
5766 | * under writeback and marked for immediate reclaim at the tail of the | |
5767 | * LRU. | |
69392a40 | 5768 | */ |
1b4e3f26 | 5769 | if (current_is_kswapd() || cgroup_reclaim(sc)) |
69392a40 MG |
5770 | return; |
5771 | ||
5772 | /* Throttle if making no progress at high prioities. */ | |
1b4e3f26 | 5773 | if (sc->priority == 1 && !sc->nr_reclaimed) |
c3f4a9a2 | 5774 | reclaim_throttle(pgdat, VMSCAN_THROTTLE_NOPROGRESS); |
69392a40 MG |
5775 | } |
5776 | ||
1da177e4 LT |
5777 | /* |
5778 | * This is the direct reclaim path, for page-allocating processes. We only | |
5779 | * try to reclaim pages from zones which will satisfy the caller's allocation | |
5780 | * request. | |
5781 | * | |
1da177e4 LT |
5782 | * If a zone is deemed to be full of pinned pages then just give it a light |
5783 | * scan then give up on it. | |
5784 | */ | |
0a0337e0 | 5785 | static void shrink_zones(struct zonelist *zonelist, struct scan_control *sc) |
1da177e4 | 5786 | { |
dd1a239f | 5787 | struct zoneref *z; |
54a6eb5c | 5788 | struct zone *zone; |
0608f43d AM |
5789 | unsigned long nr_soft_reclaimed; |
5790 | unsigned long nr_soft_scanned; | |
619d0d76 | 5791 | gfp_t orig_mask; |
79dafcdc | 5792 | pg_data_t *last_pgdat = NULL; |
1b4e3f26 | 5793 | pg_data_t *first_pgdat = NULL; |
1cfb419b | 5794 | |
cc715d99 MG |
5795 | /* |
5796 | * If the number of buffer_heads in the machine exceeds the maximum | |
5797 | * allowed level, force direct reclaim to scan the highmem zone as | |
5798 | * highmem pages could be pinning lowmem pages storing buffer_heads | |
5799 | */ | |
619d0d76 | 5800 | orig_mask = sc->gfp_mask; |
b2e18757 | 5801 | if (buffer_heads_over_limit) { |
cc715d99 | 5802 | sc->gfp_mask |= __GFP_HIGHMEM; |
4f588331 | 5803 | sc->reclaim_idx = gfp_zone(sc->gfp_mask); |
b2e18757 | 5804 | } |
cc715d99 | 5805 | |
d4debc66 | 5806 | for_each_zone_zonelist_nodemask(zone, z, zonelist, |
b2e18757 | 5807 | sc->reclaim_idx, sc->nodemask) { |
1cfb419b KH |
5808 | /* |
5809 | * Take care memory controller reclaiming has small influence | |
5810 | * to global LRU. | |
5811 | */ | |
b5ead35e | 5812 | if (!cgroup_reclaim(sc)) { |
344736f2 VD |
5813 | if (!cpuset_zone_allowed(zone, |
5814 | GFP_KERNEL | __GFP_HARDWALL)) | |
1cfb419b | 5815 | continue; |
65ec02cb | 5816 | |
0b06496a JW |
5817 | /* |
5818 | * If we already have plenty of memory free for | |
5819 | * compaction in this zone, don't free any more. | |
5820 | * Even though compaction is invoked for any | |
5821 | * non-zero order, only frequent costly order | |
5822 | * reclamation is disruptive enough to become a | |
5823 | * noticeable problem, like transparent huge | |
5824 | * page allocations. | |
5825 | */ | |
5826 | if (IS_ENABLED(CONFIG_COMPACTION) && | |
5827 | sc->order > PAGE_ALLOC_COSTLY_ORDER && | |
4f588331 | 5828 | compaction_ready(zone, sc)) { |
0b06496a JW |
5829 | sc->compaction_ready = true; |
5830 | continue; | |
e0887c19 | 5831 | } |
0b06496a | 5832 | |
79dafcdc MG |
5833 | /* |
5834 | * Shrink each node in the zonelist once. If the | |
5835 | * zonelist is ordered by zone (not the default) then a | |
5836 | * node may be shrunk multiple times but in that case | |
5837 | * the user prefers lower zones being preserved. | |
5838 | */ | |
5839 | if (zone->zone_pgdat == last_pgdat) | |
5840 | continue; | |
5841 | ||
0608f43d AM |
5842 | /* |
5843 | * This steals pages from memory cgroups over softlimit | |
5844 | * and returns the number of reclaimed pages and | |
5845 | * scanned pages. This works for global memory pressure | |
5846 | * and balancing, not for a memcg's limit. | |
5847 | */ | |
5848 | nr_soft_scanned = 0; | |
ef8f2327 | 5849 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(zone->zone_pgdat, |
0608f43d AM |
5850 | sc->order, sc->gfp_mask, |
5851 | &nr_soft_scanned); | |
5852 | sc->nr_reclaimed += nr_soft_reclaimed; | |
5853 | sc->nr_scanned += nr_soft_scanned; | |
ac34a1a3 | 5854 | /* need some check for avoid more shrink_zone() */ |
1cfb419b | 5855 | } |
408d8544 | 5856 | |
1b4e3f26 MG |
5857 | if (!first_pgdat) |
5858 | first_pgdat = zone->zone_pgdat; | |
5859 | ||
79dafcdc MG |
5860 | /* See comment about same check for global reclaim above */ |
5861 | if (zone->zone_pgdat == last_pgdat) | |
5862 | continue; | |
5863 | last_pgdat = zone->zone_pgdat; | |
970a39a3 | 5864 | shrink_node(zone->zone_pgdat, sc); |
1da177e4 | 5865 | } |
e0c23279 | 5866 | |
80082938 MG |
5867 | if (first_pgdat) |
5868 | consider_reclaim_throttle(first_pgdat, sc); | |
1b4e3f26 | 5869 | |
619d0d76 WY |
5870 | /* |
5871 | * Restore to original mask to avoid the impact on the caller if we | |
5872 | * promoted it to __GFP_HIGHMEM. | |
5873 | */ | |
5874 | sc->gfp_mask = orig_mask; | |
1da177e4 | 5875 | } |
4f98a2fe | 5876 | |
b910718a | 5877 | static void snapshot_refaults(struct mem_cgroup *target_memcg, pg_data_t *pgdat) |
2a2e4885 | 5878 | { |
b910718a JW |
5879 | struct lruvec *target_lruvec; |
5880 | unsigned long refaults; | |
2a2e4885 | 5881 | |
ac35a490 YZ |
5882 | if (lru_gen_enabled()) |
5883 | return; | |
5884 | ||
b910718a | 5885 | target_lruvec = mem_cgroup_lruvec(target_memcg, pgdat); |
170b04b7 | 5886 | refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_ANON); |
e9c2dbc8 | 5887 | target_lruvec->refaults[WORKINGSET_ANON] = refaults; |
170b04b7 | 5888 | refaults = lruvec_page_state(target_lruvec, WORKINGSET_ACTIVATE_FILE); |
e9c2dbc8 | 5889 | target_lruvec->refaults[WORKINGSET_FILE] = refaults; |
2a2e4885 JW |
5890 | } |
5891 | ||
1da177e4 LT |
5892 | /* |
5893 | * This is the main entry point to direct page reclaim. | |
5894 | * | |
5895 | * If a full scan of the inactive list fails to free enough memory then we | |
5896 | * are "out of memory" and something needs to be killed. | |
5897 | * | |
5898 | * If the caller is !__GFP_FS then the probability of a failure is reasonably | |
5899 | * high - the zone may be full of dirty or under-writeback pages, which this | |
5b0830cb JA |
5900 | * caller can't do much about. We kick the writeback threads and take explicit |
5901 | * naps in the hope that some of these pages can be written. But if the | |
5902 | * allocating task holds filesystem locks which prevent writeout this might not | |
5903 | * work, and the allocation attempt will fail. | |
a41f24ea NA |
5904 | * |
5905 | * returns: 0, if no pages reclaimed | |
5906 | * else, the number of pages reclaimed | |
1da177e4 | 5907 | */ |
dac1d27b | 5908 | static unsigned long do_try_to_free_pages(struct zonelist *zonelist, |
3115cd91 | 5909 | struct scan_control *sc) |
1da177e4 | 5910 | { |
241994ed | 5911 | int initial_priority = sc->priority; |
2a2e4885 JW |
5912 | pg_data_t *last_pgdat; |
5913 | struct zoneref *z; | |
5914 | struct zone *zone; | |
241994ed | 5915 | retry: |
873b4771 KK |
5916 | delayacct_freepages_start(); |
5917 | ||
b5ead35e | 5918 | if (!cgroup_reclaim(sc)) |
7cc30fcf | 5919 | __count_zid_vm_events(ALLOCSTALL, sc->reclaim_idx, 1); |
1da177e4 | 5920 | |
9e3b2f8c | 5921 | do { |
73b73bac YA |
5922 | if (!sc->proactive) |
5923 | vmpressure_prio(sc->gfp_mask, sc->target_mem_cgroup, | |
5924 | sc->priority); | |
66e1707b | 5925 | sc->nr_scanned = 0; |
0a0337e0 | 5926 | shrink_zones(zonelist, sc); |
c6a8a8c5 | 5927 | |
bb21c7ce | 5928 | if (sc->nr_reclaimed >= sc->nr_to_reclaim) |
0b06496a JW |
5929 | break; |
5930 | ||
5931 | if (sc->compaction_ready) | |
5932 | break; | |
1da177e4 | 5933 | |
0e50ce3b MK |
5934 | /* |
5935 | * If we're getting trouble reclaiming, start doing | |
5936 | * writepage even in laptop mode. | |
5937 | */ | |
5938 | if (sc->priority < DEF_PRIORITY - 2) | |
5939 | sc->may_writepage = 1; | |
0b06496a | 5940 | } while (--sc->priority >= 0); |
bb21c7ce | 5941 | |
2a2e4885 JW |
5942 | last_pgdat = NULL; |
5943 | for_each_zone_zonelist_nodemask(zone, z, zonelist, sc->reclaim_idx, | |
5944 | sc->nodemask) { | |
5945 | if (zone->zone_pgdat == last_pgdat) | |
5946 | continue; | |
5947 | last_pgdat = zone->zone_pgdat; | |
1b05117d | 5948 | |
2a2e4885 | 5949 | snapshot_refaults(sc->target_mem_cgroup, zone->zone_pgdat); |
1b05117d JW |
5950 | |
5951 | if (cgroup_reclaim(sc)) { | |
5952 | struct lruvec *lruvec; | |
5953 | ||
5954 | lruvec = mem_cgroup_lruvec(sc->target_mem_cgroup, | |
5955 | zone->zone_pgdat); | |
5956 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
5957 | } | |
2a2e4885 JW |
5958 | } |
5959 | ||
873b4771 KK |
5960 | delayacct_freepages_end(); |
5961 | ||
bb21c7ce KM |
5962 | if (sc->nr_reclaimed) |
5963 | return sc->nr_reclaimed; | |
5964 | ||
0cee34fd | 5965 | /* Aborted reclaim to try compaction? don't OOM, then */ |
0b06496a | 5966 | if (sc->compaction_ready) |
7335084d MG |
5967 | return 1; |
5968 | ||
b91ac374 JW |
5969 | /* |
5970 | * We make inactive:active ratio decisions based on the node's | |
5971 | * composition of memory, but a restrictive reclaim_idx or a | |
5972 | * memory.low cgroup setting can exempt large amounts of | |
5973 | * memory from reclaim. Neither of which are very common, so | |
5974 | * instead of doing costly eligibility calculations of the | |
5975 | * entire cgroup subtree up front, we assume the estimates are | |
5976 | * good, and retry with forcible deactivation if that fails. | |
5977 | */ | |
5978 | if (sc->skipped_deactivate) { | |
5979 | sc->priority = initial_priority; | |
5980 | sc->force_deactivate = 1; | |
5981 | sc->skipped_deactivate = 0; | |
5982 | goto retry; | |
5983 | } | |
5984 | ||
241994ed | 5985 | /* Untapped cgroup reserves? Don't OOM, retry. */ |
d6622f63 | 5986 | if (sc->memcg_low_skipped) { |
241994ed | 5987 | sc->priority = initial_priority; |
b91ac374 | 5988 | sc->force_deactivate = 0; |
d6622f63 YX |
5989 | sc->memcg_low_reclaim = 1; |
5990 | sc->memcg_low_skipped = 0; | |
241994ed JW |
5991 | goto retry; |
5992 | } | |
5993 | ||
bb21c7ce | 5994 | return 0; |
1da177e4 LT |
5995 | } |
5996 | ||
c73322d0 | 5997 | static bool allow_direct_reclaim(pg_data_t *pgdat) |
5515061d MG |
5998 | { |
5999 | struct zone *zone; | |
6000 | unsigned long pfmemalloc_reserve = 0; | |
6001 | unsigned long free_pages = 0; | |
6002 | int i; | |
6003 | bool wmark_ok; | |
6004 | ||
c73322d0 JW |
6005 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) |
6006 | return true; | |
6007 | ||
5515061d MG |
6008 | for (i = 0; i <= ZONE_NORMAL; i++) { |
6009 | zone = &pgdat->node_zones[i]; | |
d450abd8 JW |
6010 | if (!managed_zone(zone)) |
6011 | continue; | |
6012 | ||
6013 | if (!zone_reclaimable_pages(zone)) | |
675becce MG |
6014 | continue; |
6015 | ||
5515061d MG |
6016 | pfmemalloc_reserve += min_wmark_pages(zone); |
6017 | free_pages += zone_page_state(zone, NR_FREE_PAGES); | |
6018 | } | |
6019 | ||
675becce MG |
6020 | /* If there are no reserves (unexpected config) then do not throttle */ |
6021 | if (!pfmemalloc_reserve) | |
6022 | return true; | |
6023 | ||
5515061d MG |
6024 | wmark_ok = free_pages > pfmemalloc_reserve / 2; |
6025 | ||
6026 | /* kswapd must be awake if processes are being throttled */ | |
6027 | if (!wmark_ok && waitqueue_active(&pgdat->kswapd_wait)) { | |
97a225e6 JK |
6028 | if (READ_ONCE(pgdat->kswapd_highest_zoneidx) > ZONE_NORMAL) |
6029 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, ZONE_NORMAL); | |
5644e1fb | 6030 | |
5515061d MG |
6031 | wake_up_interruptible(&pgdat->kswapd_wait); |
6032 | } | |
6033 | ||
6034 | return wmark_ok; | |
6035 | } | |
6036 | ||
6037 | /* | |
6038 | * Throttle direct reclaimers if backing storage is backed by the network | |
6039 | * and the PFMEMALLOC reserve for the preferred node is getting dangerously | |
6040 | * depleted. kswapd will continue to make progress and wake the processes | |
50694c28 MG |
6041 | * when the low watermark is reached. |
6042 | * | |
6043 | * Returns true if a fatal signal was delivered during throttling. If this | |
6044 | * happens, the page allocator should not consider triggering the OOM killer. | |
5515061d | 6045 | */ |
50694c28 | 6046 | static bool throttle_direct_reclaim(gfp_t gfp_mask, struct zonelist *zonelist, |
5515061d MG |
6047 | nodemask_t *nodemask) |
6048 | { | |
675becce | 6049 | struct zoneref *z; |
5515061d | 6050 | struct zone *zone; |
675becce | 6051 | pg_data_t *pgdat = NULL; |
5515061d MG |
6052 | |
6053 | /* | |
6054 | * Kernel threads should not be throttled as they may be indirectly | |
6055 | * responsible for cleaning pages necessary for reclaim to make forward | |
6056 | * progress. kjournald for example may enter direct reclaim while | |
6057 | * committing a transaction where throttling it could forcing other | |
6058 | * processes to block on log_wait_commit(). | |
6059 | */ | |
6060 | if (current->flags & PF_KTHREAD) | |
50694c28 MG |
6061 | goto out; |
6062 | ||
6063 | /* | |
6064 | * If a fatal signal is pending, this process should not throttle. | |
6065 | * It should return quickly so it can exit and free its memory | |
6066 | */ | |
6067 | if (fatal_signal_pending(current)) | |
6068 | goto out; | |
5515061d | 6069 | |
675becce MG |
6070 | /* |
6071 | * Check if the pfmemalloc reserves are ok by finding the first node | |
6072 | * with a usable ZONE_NORMAL or lower zone. The expectation is that | |
6073 | * GFP_KERNEL will be required for allocating network buffers when | |
6074 | * swapping over the network so ZONE_HIGHMEM is unusable. | |
6075 | * | |
6076 | * Throttling is based on the first usable node and throttled processes | |
6077 | * wait on a queue until kswapd makes progress and wakes them. There | |
6078 | * is an affinity then between processes waking up and where reclaim | |
6079 | * progress has been made assuming the process wakes on the same node. | |
6080 | * More importantly, processes running on remote nodes will not compete | |
6081 | * for remote pfmemalloc reserves and processes on different nodes | |
6082 | * should make reasonable progress. | |
6083 | */ | |
6084 | for_each_zone_zonelist_nodemask(zone, z, zonelist, | |
17636faa | 6085 | gfp_zone(gfp_mask), nodemask) { |
675becce MG |
6086 | if (zone_idx(zone) > ZONE_NORMAL) |
6087 | continue; | |
6088 | ||
6089 | /* Throttle based on the first usable node */ | |
6090 | pgdat = zone->zone_pgdat; | |
c73322d0 | 6091 | if (allow_direct_reclaim(pgdat)) |
675becce MG |
6092 | goto out; |
6093 | break; | |
6094 | } | |
6095 | ||
6096 | /* If no zone was usable by the allocation flags then do not throttle */ | |
6097 | if (!pgdat) | |
50694c28 | 6098 | goto out; |
5515061d | 6099 | |
68243e76 MG |
6100 | /* Account for the throttling */ |
6101 | count_vm_event(PGSCAN_DIRECT_THROTTLE); | |
6102 | ||
5515061d MG |
6103 | /* |
6104 | * If the caller cannot enter the filesystem, it's possible that it | |
6105 | * is due to the caller holding an FS lock or performing a journal | |
6106 | * transaction in the case of a filesystem like ext[3|4]. In this case, | |
6107 | * it is not safe to block on pfmemalloc_wait as kswapd could be | |
6108 | * blocked waiting on the same lock. Instead, throttle for up to a | |
6109 | * second before continuing. | |
6110 | */ | |
2e786d9e | 6111 | if (!(gfp_mask & __GFP_FS)) |
5515061d | 6112 | wait_event_interruptible_timeout(pgdat->pfmemalloc_wait, |
c73322d0 | 6113 | allow_direct_reclaim(pgdat), HZ); |
2e786d9e ML |
6114 | else |
6115 | /* Throttle until kswapd wakes the process */ | |
6116 | wait_event_killable(zone->zone_pgdat->pfmemalloc_wait, | |
6117 | allow_direct_reclaim(pgdat)); | |
50694c28 | 6118 | |
50694c28 MG |
6119 | if (fatal_signal_pending(current)) |
6120 | return true; | |
6121 | ||
6122 | out: | |
6123 | return false; | |
5515061d MG |
6124 | } |
6125 | ||
dac1d27b | 6126 | unsigned long try_to_free_pages(struct zonelist *zonelist, int order, |
327c0e96 | 6127 | gfp_t gfp_mask, nodemask_t *nodemask) |
66e1707b | 6128 | { |
33906bc5 | 6129 | unsigned long nr_reclaimed; |
66e1707b | 6130 | struct scan_control sc = { |
ee814fe2 | 6131 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
f2f43e56 | 6132 | .gfp_mask = current_gfp_context(gfp_mask), |
b2e18757 | 6133 | .reclaim_idx = gfp_zone(gfp_mask), |
ee814fe2 JW |
6134 | .order = order, |
6135 | .nodemask = nodemask, | |
6136 | .priority = DEF_PRIORITY, | |
66e1707b | 6137 | .may_writepage = !laptop_mode, |
a6dc60f8 | 6138 | .may_unmap = 1, |
2e2e4259 | 6139 | .may_swap = 1, |
66e1707b BS |
6140 | }; |
6141 | ||
bb451fdf GT |
6142 | /* |
6143 | * scan_control uses s8 fields for order, priority, and reclaim_idx. | |
6144 | * Confirm they are large enough for max values. | |
6145 | */ | |
6146 | BUILD_BUG_ON(MAX_ORDER > S8_MAX); | |
6147 | BUILD_BUG_ON(DEF_PRIORITY > S8_MAX); | |
6148 | BUILD_BUG_ON(MAX_NR_ZONES > S8_MAX); | |
6149 | ||
5515061d | 6150 | /* |
50694c28 MG |
6151 | * Do not enter reclaim if fatal signal was delivered while throttled. |
6152 | * 1 is returned so that the page allocator does not OOM kill at this | |
6153 | * point. | |
5515061d | 6154 | */ |
f2f43e56 | 6155 | if (throttle_direct_reclaim(sc.gfp_mask, zonelist, nodemask)) |
5515061d MG |
6156 | return 1; |
6157 | ||
1732d2b0 | 6158 | set_task_reclaim_state(current, &sc.reclaim_state); |
3481c37f | 6159 | trace_mm_vmscan_direct_reclaim_begin(order, sc.gfp_mask); |
33906bc5 | 6160 | |
3115cd91 | 6161 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
33906bc5 MG |
6162 | |
6163 | trace_mm_vmscan_direct_reclaim_end(nr_reclaimed); | |
1732d2b0 | 6164 | set_task_reclaim_state(current, NULL); |
33906bc5 MG |
6165 | |
6166 | return nr_reclaimed; | |
66e1707b BS |
6167 | } |
6168 | ||
c255a458 | 6169 | #ifdef CONFIG_MEMCG |
66e1707b | 6170 | |
d2e5fb92 | 6171 | /* Only used by soft limit reclaim. Do not reuse for anything else. */ |
a9dd0a83 | 6172 | unsigned long mem_cgroup_shrink_node(struct mem_cgroup *memcg, |
4e416953 | 6173 | gfp_t gfp_mask, bool noswap, |
ef8f2327 | 6174 | pg_data_t *pgdat, |
0ae5e89c | 6175 | unsigned long *nr_scanned) |
4e416953 | 6176 | { |
afaf07a6 | 6177 | struct lruvec *lruvec = mem_cgroup_lruvec(memcg, pgdat); |
4e416953 | 6178 | struct scan_control sc = { |
b8f5c566 | 6179 | .nr_to_reclaim = SWAP_CLUSTER_MAX, |
ee814fe2 | 6180 | .target_mem_cgroup = memcg, |
4e416953 BS |
6181 | .may_writepage = !laptop_mode, |
6182 | .may_unmap = 1, | |
b2e18757 | 6183 | .reclaim_idx = MAX_NR_ZONES - 1, |
4e416953 | 6184 | .may_swap = !noswap, |
4e416953 | 6185 | }; |
0ae5e89c | 6186 | |
d2e5fb92 MH |
6187 | WARN_ON_ONCE(!current->reclaim_state); |
6188 | ||
4e416953 BS |
6189 | sc.gfp_mask = (gfp_mask & GFP_RECLAIM_MASK) | |
6190 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK); | |
bdce6d9e | 6191 | |
9e3b2f8c | 6192 | trace_mm_vmscan_memcg_softlimit_reclaim_begin(sc.order, |
3481c37f | 6193 | sc.gfp_mask); |
bdce6d9e | 6194 | |
4e416953 BS |
6195 | /* |
6196 | * NOTE: Although we can get the priority field, using it | |
6197 | * here is not a good idea, since it limits the pages we can scan. | |
a9dd0a83 | 6198 | * if we don't reclaim here, the shrink_node from balance_pgdat |
4e416953 BS |
6199 | * will pick up pages from other mem cgroup's as well. We hack |
6200 | * the priority and make it zero. | |
6201 | */ | |
afaf07a6 | 6202 | shrink_lruvec(lruvec, &sc); |
bdce6d9e KM |
6203 | |
6204 | trace_mm_vmscan_memcg_softlimit_reclaim_end(sc.nr_reclaimed); | |
6205 | ||
0ae5e89c | 6206 | *nr_scanned = sc.nr_scanned; |
0308f7cf | 6207 | |
4e416953 BS |
6208 | return sc.nr_reclaimed; |
6209 | } | |
6210 | ||
72835c86 | 6211 | unsigned long try_to_free_mem_cgroup_pages(struct mem_cgroup *memcg, |
b70a2a21 | 6212 | unsigned long nr_pages, |
a7885eb8 | 6213 | gfp_t gfp_mask, |
73b73bac | 6214 | unsigned int reclaim_options) |
66e1707b | 6215 | { |
bdce6d9e | 6216 | unsigned long nr_reclaimed; |
499118e9 | 6217 | unsigned int noreclaim_flag; |
66e1707b | 6218 | struct scan_control sc = { |
b70a2a21 | 6219 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
7dea19f9 | 6220 | .gfp_mask = (current_gfp_context(gfp_mask) & GFP_RECLAIM_MASK) | |
a09ed5e0 | 6221 | (GFP_HIGHUSER_MOVABLE & ~GFP_RECLAIM_MASK), |
b2e18757 | 6222 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 JW |
6223 | .target_mem_cgroup = memcg, |
6224 | .priority = DEF_PRIORITY, | |
6225 | .may_writepage = !laptop_mode, | |
6226 | .may_unmap = 1, | |
73b73bac YA |
6227 | .may_swap = !!(reclaim_options & MEMCG_RECLAIM_MAY_SWAP), |
6228 | .proactive = !!(reclaim_options & MEMCG_RECLAIM_PROACTIVE), | |
a09ed5e0 | 6229 | }; |
889976db | 6230 | /* |
fa40d1ee SB |
6231 | * Traverse the ZONELIST_FALLBACK zonelist of the current node to put |
6232 | * equal pressure on all the nodes. This is based on the assumption that | |
6233 | * the reclaim does not bail out early. | |
889976db | 6234 | */ |
fa40d1ee | 6235 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
889976db | 6236 | |
fa40d1ee | 6237 | set_task_reclaim_state(current, &sc.reclaim_state); |
3481c37f | 6238 | trace_mm_vmscan_memcg_reclaim_begin(0, sc.gfp_mask); |
499118e9 | 6239 | noreclaim_flag = memalloc_noreclaim_save(); |
eb414681 | 6240 | |
3115cd91 | 6241 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
eb414681 | 6242 | |
499118e9 | 6243 | memalloc_noreclaim_restore(noreclaim_flag); |
bdce6d9e | 6244 | trace_mm_vmscan_memcg_reclaim_end(nr_reclaimed); |
1732d2b0 | 6245 | set_task_reclaim_state(current, NULL); |
bdce6d9e KM |
6246 | |
6247 | return nr_reclaimed; | |
66e1707b BS |
6248 | } |
6249 | #endif | |
6250 | ||
ac35a490 | 6251 | static void kswapd_age_node(struct pglist_data *pgdat, struct scan_control *sc) |
f16015fb | 6252 | { |
b95a2f2d | 6253 | struct mem_cgroup *memcg; |
b91ac374 | 6254 | struct lruvec *lruvec; |
f16015fb | 6255 | |
ac35a490 YZ |
6256 | if (lru_gen_enabled()) { |
6257 | lru_gen_age_node(pgdat, sc); | |
6258 | return; | |
6259 | } | |
6260 | ||
2f368a9f | 6261 | if (!can_age_anon_pages(pgdat, sc)) |
b95a2f2d JW |
6262 | return; |
6263 | ||
b91ac374 JW |
6264 | lruvec = mem_cgroup_lruvec(NULL, pgdat); |
6265 | if (!inactive_is_low(lruvec, LRU_INACTIVE_ANON)) | |
6266 | return; | |
6267 | ||
b95a2f2d JW |
6268 | memcg = mem_cgroup_iter(NULL, NULL, NULL); |
6269 | do { | |
b91ac374 JW |
6270 | lruvec = mem_cgroup_lruvec(memcg, pgdat); |
6271 | shrink_active_list(SWAP_CLUSTER_MAX, lruvec, | |
6272 | sc, LRU_ACTIVE_ANON); | |
b95a2f2d JW |
6273 | memcg = mem_cgroup_iter(NULL, memcg, NULL); |
6274 | } while (memcg); | |
f16015fb JW |
6275 | } |
6276 | ||
97a225e6 | 6277 | static bool pgdat_watermark_boosted(pg_data_t *pgdat, int highest_zoneidx) |
1c30844d MG |
6278 | { |
6279 | int i; | |
6280 | struct zone *zone; | |
6281 | ||
6282 | /* | |
6283 | * Check for watermark boosts top-down as the higher zones | |
6284 | * are more likely to be boosted. Both watermarks and boosts | |
1eba09c1 | 6285 | * should not be checked at the same time as reclaim would |
1c30844d MG |
6286 | * start prematurely when there is no boosting and a lower |
6287 | * zone is balanced. | |
6288 | */ | |
97a225e6 | 6289 | for (i = highest_zoneidx; i >= 0; i--) { |
1c30844d MG |
6290 | zone = pgdat->node_zones + i; |
6291 | if (!managed_zone(zone)) | |
6292 | continue; | |
6293 | ||
6294 | if (zone->watermark_boost) | |
6295 | return true; | |
6296 | } | |
6297 | ||
6298 | return false; | |
6299 | } | |
6300 | ||
e716f2eb MG |
6301 | /* |
6302 | * Returns true if there is an eligible zone balanced for the request order | |
97a225e6 | 6303 | * and highest_zoneidx |
e716f2eb | 6304 | */ |
97a225e6 | 6305 | static bool pgdat_balanced(pg_data_t *pgdat, int order, int highest_zoneidx) |
60cefed4 | 6306 | { |
e716f2eb MG |
6307 | int i; |
6308 | unsigned long mark = -1; | |
6309 | struct zone *zone; | |
60cefed4 | 6310 | |
1c30844d MG |
6311 | /* |
6312 | * Check watermarks bottom-up as lower zones are more likely to | |
6313 | * meet watermarks. | |
6314 | */ | |
97a225e6 | 6315 | for (i = 0; i <= highest_zoneidx; i++) { |
e716f2eb | 6316 | zone = pgdat->node_zones + i; |
6256c6b4 | 6317 | |
e716f2eb MG |
6318 | if (!managed_zone(zone)) |
6319 | continue; | |
6320 | ||
c574bbe9 HY |
6321 | if (sysctl_numa_balancing_mode & NUMA_BALANCING_MEMORY_TIERING) |
6322 | mark = wmark_pages(zone, WMARK_PROMO); | |
6323 | else | |
6324 | mark = high_wmark_pages(zone); | |
97a225e6 | 6325 | if (zone_watermark_ok_safe(zone, order, mark, highest_zoneidx)) |
e716f2eb MG |
6326 | return true; |
6327 | } | |
6328 | ||
6329 | /* | |
36c26128 | 6330 | * If a node has no managed zone within highest_zoneidx, it does not |
e716f2eb MG |
6331 | * need balancing by definition. This can happen if a zone-restricted |
6332 | * allocation tries to wake a remote kswapd. | |
6333 | */ | |
6334 | if (mark == -1) | |
6335 | return true; | |
6336 | ||
6337 | return false; | |
60cefed4 JW |
6338 | } |
6339 | ||
631b6e08 MG |
6340 | /* Clear pgdat state for congested, dirty or under writeback. */ |
6341 | static void clear_pgdat_congested(pg_data_t *pgdat) | |
6342 | { | |
1b05117d JW |
6343 | struct lruvec *lruvec = mem_cgroup_lruvec(NULL, pgdat); |
6344 | ||
6345 | clear_bit(LRUVEC_CONGESTED, &lruvec->flags); | |
631b6e08 MG |
6346 | clear_bit(PGDAT_DIRTY, &pgdat->flags); |
6347 | clear_bit(PGDAT_WRITEBACK, &pgdat->flags); | |
6348 | } | |
6349 | ||
5515061d MG |
6350 | /* |
6351 | * Prepare kswapd for sleeping. This verifies that there are no processes | |
6352 | * waiting in throttle_direct_reclaim() and that watermarks have been met. | |
6353 | * | |
6354 | * Returns true if kswapd is ready to sleep | |
6355 | */ | |
97a225e6 JK |
6356 | static bool prepare_kswapd_sleep(pg_data_t *pgdat, int order, |
6357 | int highest_zoneidx) | |
f50de2d3 | 6358 | { |
5515061d | 6359 | /* |
9e5e3661 | 6360 | * The throttled processes are normally woken up in balance_pgdat() as |
c73322d0 | 6361 | * soon as allow_direct_reclaim() is true. But there is a potential |
9e5e3661 VB |
6362 | * race between when kswapd checks the watermarks and a process gets |
6363 | * throttled. There is also a potential race if processes get | |
6364 | * throttled, kswapd wakes, a large process exits thereby balancing the | |
6365 | * zones, which causes kswapd to exit balance_pgdat() before reaching | |
6366 | * the wake up checks. If kswapd is going to sleep, no process should | |
6367 | * be sleeping on pfmemalloc_wait, so wake them now if necessary. If | |
6368 | * the wake up is premature, processes will wake kswapd and get | |
6369 | * throttled again. The difference from wake ups in balance_pgdat() is | |
6370 | * that here we are under prepare_to_wait(). | |
5515061d | 6371 | */ |
9e5e3661 VB |
6372 | if (waitqueue_active(&pgdat->pfmemalloc_wait)) |
6373 | wake_up_all(&pgdat->pfmemalloc_wait); | |
f50de2d3 | 6374 | |
c73322d0 JW |
6375 | /* Hopeless node, leave it to direct reclaim */ |
6376 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES) | |
6377 | return true; | |
6378 | ||
97a225e6 | 6379 | if (pgdat_balanced(pgdat, order, highest_zoneidx)) { |
e716f2eb MG |
6380 | clear_pgdat_congested(pgdat); |
6381 | return true; | |
1d82de61 MG |
6382 | } |
6383 | ||
333b0a45 | 6384 | return false; |
f50de2d3 MG |
6385 | } |
6386 | ||
75485363 | 6387 | /* |
1d82de61 MG |
6388 | * kswapd shrinks a node of pages that are at or below the highest usable |
6389 | * zone that is currently unbalanced. | |
b8e83b94 MG |
6390 | * |
6391 | * Returns true if kswapd scanned at least the requested number of pages to | |
283aba9f MG |
6392 | * reclaim or if the lack of progress was due to pages under writeback. |
6393 | * This is used to determine if the scanning priority needs to be raised. | |
75485363 | 6394 | */ |
1d82de61 | 6395 | static bool kswapd_shrink_node(pg_data_t *pgdat, |
accf6242 | 6396 | struct scan_control *sc) |
75485363 | 6397 | { |
1d82de61 MG |
6398 | struct zone *zone; |
6399 | int z; | |
75485363 | 6400 | |
1d82de61 MG |
6401 | /* Reclaim a number of pages proportional to the number of zones */ |
6402 | sc->nr_to_reclaim = 0; | |
970a39a3 | 6403 | for (z = 0; z <= sc->reclaim_idx; z++) { |
1d82de61 | 6404 | zone = pgdat->node_zones + z; |
6aa303de | 6405 | if (!managed_zone(zone)) |
1d82de61 | 6406 | continue; |
7c954f6d | 6407 | |
1d82de61 MG |
6408 | sc->nr_to_reclaim += max(high_wmark_pages(zone), SWAP_CLUSTER_MAX); |
6409 | } | |
7c954f6d MG |
6410 | |
6411 | /* | |
1d82de61 MG |
6412 | * Historically care was taken to put equal pressure on all zones but |
6413 | * now pressure is applied based on node LRU order. | |
7c954f6d | 6414 | */ |
970a39a3 | 6415 | shrink_node(pgdat, sc); |
283aba9f | 6416 | |
7c954f6d | 6417 | /* |
1d82de61 MG |
6418 | * Fragmentation may mean that the system cannot be rebalanced for |
6419 | * high-order allocations. If twice the allocation size has been | |
6420 | * reclaimed then recheck watermarks only at order-0 to prevent | |
6421 | * excessive reclaim. Assume that a process requested a high-order | |
6422 | * can direct reclaim/compact. | |
7c954f6d | 6423 | */ |
9861a62c | 6424 | if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order)) |
1d82de61 | 6425 | sc->order = 0; |
7c954f6d | 6426 | |
b8e83b94 | 6427 | return sc->nr_scanned >= sc->nr_to_reclaim; |
75485363 MG |
6428 | } |
6429 | ||
c49c2c47 MG |
6430 | /* Page allocator PCP high watermark is lowered if reclaim is active. */ |
6431 | static inline void | |
6432 | update_reclaim_active(pg_data_t *pgdat, int highest_zoneidx, bool active) | |
6433 | { | |
6434 | int i; | |
6435 | struct zone *zone; | |
6436 | ||
6437 | for (i = 0; i <= highest_zoneidx; i++) { | |
6438 | zone = pgdat->node_zones + i; | |
6439 | ||
6440 | if (!managed_zone(zone)) | |
6441 | continue; | |
6442 | ||
6443 | if (active) | |
6444 | set_bit(ZONE_RECLAIM_ACTIVE, &zone->flags); | |
6445 | else | |
6446 | clear_bit(ZONE_RECLAIM_ACTIVE, &zone->flags); | |
6447 | } | |
6448 | } | |
6449 | ||
6450 | static inline void | |
6451 | set_reclaim_active(pg_data_t *pgdat, int highest_zoneidx) | |
6452 | { | |
6453 | update_reclaim_active(pgdat, highest_zoneidx, true); | |
6454 | } | |
6455 | ||
6456 | static inline void | |
6457 | clear_reclaim_active(pg_data_t *pgdat, int highest_zoneidx) | |
6458 | { | |
6459 | update_reclaim_active(pgdat, highest_zoneidx, false); | |
6460 | } | |
6461 | ||
1da177e4 | 6462 | /* |
1d82de61 MG |
6463 | * For kswapd, balance_pgdat() will reclaim pages across a node from zones |
6464 | * that are eligible for use by the caller until at least one zone is | |
6465 | * balanced. | |
1da177e4 | 6466 | * |
1d82de61 | 6467 | * Returns the order kswapd finished reclaiming at. |
1da177e4 LT |
6468 | * |
6469 | * kswapd scans the zones in the highmem->normal->dma direction. It skips | |
41858966 | 6470 | * zones which have free_pages > high_wmark_pages(zone), but once a zone is |
8bb4e7a2 | 6471 | * found to have free_pages <= high_wmark_pages(zone), any page in that zone |
1d82de61 MG |
6472 | * or lower is eligible for reclaim until at least one usable zone is |
6473 | * balanced. | |
1da177e4 | 6474 | */ |
97a225e6 | 6475 | static int balance_pgdat(pg_data_t *pgdat, int order, int highest_zoneidx) |
1da177e4 | 6476 | { |
1da177e4 | 6477 | int i; |
0608f43d AM |
6478 | unsigned long nr_soft_reclaimed; |
6479 | unsigned long nr_soft_scanned; | |
eb414681 | 6480 | unsigned long pflags; |
1c30844d MG |
6481 | unsigned long nr_boost_reclaim; |
6482 | unsigned long zone_boosts[MAX_NR_ZONES] = { 0, }; | |
6483 | bool boosted; | |
1d82de61 | 6484 | struct zone *zone; |
179e9639 AM |
6485 | struct scan_control sc = { |
6486 | .gfp_mask = GFP_KERNEL, | |
ee814fe2 | 6487 | .order = order, |
a6dc60f8 | 6488 | .may_unmap = 1, |
179e9639 | 6489 | }; |
93781325 | 6490 | |
1732d2b0 | 6491 | set_task_reclaim_state(current, &sc.reclaim_state); |
eb414681 | 6492 | psi_memstall_enter(&pflags); |
4f3eaf45 | 6493 | __fs_reclaim_acquire(_THIS_IP_); |
93781325 | 6494 | |
f8891e5e | 6495 | count_vm_event(PAGEOUTRUN); |
1da177e4 | 6496 | |
1c30844d MG |
6497 | /* |
6498 | * Account for the reclaim boost. Note that the zone boost is left in | |
6499 | * place so that parallel allocations that are near the watermark will | |
6500 | * stall or direct reclaim until kswapd is finished. | |
6501 | */ | |
6502 | nr_boost_reclaim = 0; | |
97a225e6 | 6503 | for (i = 0; i <= highest_zoneidx; i++) { |
1c30844d MG |
6504 | zone = pgdat->node_zones + i; |
6505 | if (!managed_zone(zone)) | |
6506 | continue; | |
6507 | ||
6508 | nr_boost_reclaim += zone->watermark_boost; | |
6509 | zone_boosts[i] = zone->watermark_boost; | |
6510 | } | |
6511 | boosted = nr_boost_reclaim; | |
6512 | ||
6513 | restart: | |
c49c2c47 | 6514 | set_reclaim_active(pgdat, highest_zoneidx); |
1c30844d | 6515 | sc.priority = DEF_PRIORITY; |
9e3b2f8c | 6516 | do { |
c73322d0 | 6517 | unsigned long nr_reclaimed = sc.nr_reclaimed; |
b8e83b94 | 6518 | bool raise_priority = true; |
1c30844d | 6519 | bool balanced; |
93781325 | 6520 | bool ret; |
b8e83b94 | 6521 | |
97a225e6 | 6522 | sc.reclaim_idx = highest_zoneidx; |
1da177e4 | 6523 | |
86c79f6b | 6524 | /* |
84c7a777 MG |
6525 | * If the number of buffer_heads exceeds the maximum allowed |
6526 | * then consider reclaiming from all zones. This has a dual | |
6527 | * purpose -- on 64-bit systems it is expected that | |
6528 | * buffer_heads are stripped during active rotation. On 32-bit | |
6529 | * systems, highmem pages can pin lowmem memory and shrinking | |
6530 | * buffers can relieve lowmem pressure. Reclaim may still not | |
6531 | * go ahead if all eligible zones for the original allocation | |
6532 | * request are balanced to avoid excessive reclaim from kswapd. | |
86c79f6b MG |
6533 | */ |
6534 | if (buffer_heads_over_limit) { | |
6535 | for (i = MAX_NR_ZONES - 1; i >= 0; i--) { | |
6536 | zone = pgdat->node_zones + i; | |
6aa303de | 6537 | if (!managed_zone(zone)) |
86c79f6b | 6538 | continue; |
cc715d99 | 6539 | |
970a39a3 | 6540 | sc.reclaim_idx = i; |
e1dbeda6 | 6541 | break; |
1da177e4 | 6542 | } |
1da177e4 | 6543 | } |
dafcb73e | 6544 | |
86c79f6b | 6545 | /* |
1c30844d MG |
6546 | * If the pgdat is imbalanced then ignore boosting and preserve |
6547 | * the watermarks for a later time and restart. Note that the | |
6548 | * zone watermarks will be still reset at the end of balancing | |
6549 | * on the grounds that the normal reclaim should be enough to | |
6550 | * re-evaluate if boosting is required when kswapd next wakes. | |
6551 | */ | |
97a225e6 | 6552 | balanced = pgdat_balanced(pgdat, sc.order, highest_zoneidx); |
1c30844d MG |
6553 | if (!balanced && nr_boost_reclaim) { |
6554 | nr_boost_reclaim = 0; | |
6555 | goto restart; | |
6556 | } | |
6557 | ||
6558 | /* | |
6559 | * If boosting is not active then only reclaim if there are no | |
6560 | * eligible zones. Note that sc.reclaim_idx is not used as | |
6561 | * buffer_heads_over_limit may have adjusted it. | |
86c79f6b | 6562 | */ |
1c30844d | 6563 | if (!nr_boost_reclaim && balanced) |
e716f2eb | 6564 | goto out; |
e1dbeda6 | 6565 | |
1c30844d MG |
6566 | /* Limit the priority of boosting to avoid reclaim writeback */ |
6567 | if (nr_boost_reclaim && sc.priority == DEF_PRIORITY - 2) | |
6568 | raise_priority = false; | |
6569 | ||
6570 | /* | |
6571 | * Do not writeback or swap pages for boosted reclaim. The | |
6572 | * intent is to relieve pressure not issue sub-optimal IO | |
6573 | * from reclaim context. If no pages are reclaimed, the | |
6574 | * reclaim will be aborted. | |
6575 | */ | |
6576 | sc.may_writepage = !laptop_mode && !nr_boost_reclaim; | |
6577 | sc.may_swap = !nr_boost_reclaim; | |
1c30844d | 6578 | |
1d82de61 | 6579 | /* |
ac35a490 YZ |
6580 | * Do some background aging, to give pages a chance to be |
6581 | * referenced before reclaiming. All pages are rotated | |
6582 | * regardless of classzone as this is about consistent aging. | |
1d82de61 | 6583 | */ |
ac35a490 | 6584 | kswapd_age_node(pgdat, &sc); |
1d82de61 | 6585 | |
b7ea3c41 MG |
6586 | /* |
6587 | * If we're getting trouble reclaiming, start doing writepage | |
6588 | * even in laptop mode. | |
6589 | */ | |
047d72c3 | 6590 | if (sc.priority < DEF_PRIORITY - 2) |
b7ea3c41 MG |
6591 | sc.may_writepage = 1; |
6592 | ||
1d82de61 MG |
6593 | /* Call soft limit reclaim before calling shrink_node. */ |
6594 | sc.nr_scanned = 0; | |
6595 | nr_soft_scanned = 0; | |
ef8f2327 | 6596 | nr_soft_reclaimed = mem_cgroup_soft_limit_reclaim(pgdat, sc.order, |
1d82de61 MG |
6597 | sc.gfp_mask, &nr_soft_scanned); |
6598 | sc.nr_reclaimed += nr_soft_reclaimed; | |
6599 | ||
1da177e4 | 6600 | /* |
1d82de61 MG |
6601 | * There should be no need to raise the scanning priority if |
6602 | * enough pages are already being scanned that that high | |
6603 | * watermark would be met at 100% efficiency. | |
1da177e4 | 6604 | */ |
970a39a3 | 6605 | if (kswapd_shrink_node(pgdat, &sc)) |
1d82de61 | 6606 | raise_priority = false; |
5515061d MG |
6607 | |
6608 | /* | |
6609 | * If the low watermark is met there is no need for processes | |
6610 | * to be throttled on pfmemalloc_wait as they should not be | |
6611 | * able to safely make forward progress. Wake them | |
6612 | */ | |
6613 | if (waitqueue_active(&pgdat->pfmemalloc_wait) && | |
c73322d0 | 6614 | allow_direct_reclaim(pgdat)) |
cfc51155 | 6615 | wake_up_all(&pgdat->pfmemalloc_wait); |
5515061d | 6616 | |
b8e83b94 | 6617 | /* Check if kswapd should be suspending */ |
4f3eaf45 | 6618 | __fs_reclaim_release(_THIS_IP_); |
93781325 | 6619 | ret = try_to_freeze(); |
4f3eaf45 | 6620 | __fs_reclaim_acquire(_THIS_IP_); |
93781325 | 6621 | if (ret || kthread_should_stop()) |
b8e83b94 | 6622 | break; |
8357376d | 6623 | |
73ce02e9 | 6624 | /* |
b8e83b94 MG |
6625 | * Raise priority if scanning rate is too low or there was no |
6626 | * progress in reclaiming pages | |
73ce02e9 | 6627 | */ |
c73322d0 | 6628 | nr_reclaimed = sc.nr_reclaimed - nr_reclaimed; |
1c30844d MG |
6629 | nr_boost_reclaim -= min(nr_boost_reclaim, nr_reclaimed); |
6630 | ||
6631 | /* | |
6632 | * If reclaim made no progress for a boost, stop reclaim as | |
6633 | * IO cannot be queued and it could be an infinite loop in | |
6634 | * extreme circumstances. | |
6635 | */ | |
6636 | if (nr_boost_reclaim && !nr_reclaimed) | |
6637 | break; | |
6638 | ||
c73322d0 | 6639 | if (raise_priority || !nr_reclaimed) |
b8e83b94 | 6640 | sc.priority--; |
1d82de61 | 6641 | } while (sc.priority >= 1); |
1da177e4 | 6642 | |
c73322d0 JW |
6643 | if (!sc.nr_reclaimed) |
6644 | pgdat->kswapd_failures++; | |
6645 | ||
b8e83b94 | 6646 | out: |
c49c2c47 MG |
6647 | clear_reclaim_active(pgdat, highest_zoneidx); |
6648 | ||
1c30844d MG |
6649 | /* If reclaim was boosted, account for the reclaim done in this pass */ |
6650 | if (boosted) { | |
6651 | unsigned long flags; | |
6652 | ||
97a225e6 | 6653 | for (i = 0; i <= highest_zoneidx; i++) { |
1c30844d MG |
6654 | if (!zone_boosts[i]) |
6655 | continue; | |
6656 | ||
6657 | /* Increments are under the zone lock */ | |
6658 | zone = pgdat->node_zones + i; | |
6659 | spin_lock_irqsave(&zone->lock, flags); | |
6660 | zone->watermark_boost -= min(zone->watermark_boost, zone_boosts[i]); | |
6661 | spin_unlock_irqrestore(&zone->lock, flags); | |
6662 | } | |
6663 | ||
6664 | /* | |
6665 | * As there is now likely space, wakeup kcompact to defragment | |
6666 | * pageblocks. | |
6667 | */ | |
97a225e6 | 6668 | wakeup_kcompactd(pgdat, pageblock_order, highest_zoneidx); |
1c30844d MG |
6669 | } |
6670 | ||
2a2e4885 | 6671 | snapshot_refaults(NULL, pgdat); |
4f3eaf45 | 6672 | __fs_reclaim_release(_THIS_IP_); |
eb414681 | 6673 | psi_memstall_leave(&pflags); |
1732d2b0 | 6674 | set_task_reclaim_state(current, NULL); |
e5ca8071 | 6675 | |
0abdee2b | 6676 | /* |
1d82de61 MG |
6677 | * Return the order kswapd stopped reclaiming at as |
6678 | * prepare_kswapd_sleep() takes it into account. If another caller | |
6679 | * entered the allocator slow path while kswapd was awake, order will | |
6680 | * remain at the higher level. | |
0abdee2b | 6681 | */ |
1d82de61 | 6682 | return sc.order; |
1da177e4 LT |
6683 | } |
6684 | ||
e716f2eb | 6685 | /* |
97a225e6 JK |
6686 | * The pgdat->kswapd_highest_zoneidx is used to pass the highest zone index to |
6687 | * be reclaimed by kswapd from the waker. If the value is MAX_NR_ZONES which is | |
6688 | * not a valid index then either kswapd runs for first time or kswapd couldn't | |
6689 | * sleep after previous reclaim attempt (node is still unbalanced). In that | |
6690 | * case return the zone index of the previous kswapd reclaim cycle. | |
e716f2eb | 6691 | */ |
97a225e6 JK |
6692 | static enum zone_type kswapd_highest_zoneidx(pg_data_t *pgdat, |
6693 | enum zone_type prev_highest_zoneidx) | |
e716f2eb | 6694 | { |
97a225e6 | 6695 | enum zone_type curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx); |
5644e1fb | 6696 | |
97a225e6 | 6697 | return curr_idx == MAX_NR_ZONES ? prev_highest_zoneidx : curr_idx; |
e716f2eb MG |
6698 | } |
6699 | ||
38087d9b | 6700 | static void kswapd_try_to_sleep(pg_data_t *pgdat, int alloc_order, int reclaim_order, |
97a225e6 | 6701 | unsigned int highest_zoneidx) |
f0bc0a60 KM |
6702 | { |
6703 | long remaining = 0; | |
6704 | DEFINE_WAIT(wait); | |
6705 | ||
6706 | if (freezing(current) || kthread_should_stop()) | |
6707 | return; | |
6708 | ||
6709 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
6710 | ||
333b0a45 SG |
6711 | /* |
6712 | * Try to sleep for a short interval. Note that kcompactd will only be | |
6713 | * woken if it is possible to sleep for a short interval. This is | |
6714 | * deliberate on the assumption that if reclaim cannot keep an | |
6715 | * eligible zone balanced that it's also unlikely that compaction will | |
6716 | * succeed. | |
6717 | */ | |
97a225e6 | 6718 | if (prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) { |
fd901c95 VB |
6719 | /* |
6720 | * Compaction records what page blocks it recently failed to | |
6721 | * isolate pages from and skips them in the future scanning. | |
6722 | * When kswapd is going to sleep, it is reasonable to assume | |
6723 | * that pages and compaction may succeed so reset the cache. | |
6724 | */ | |
6725 | reset_isolation_suitable(pgdat); | |
6726 | ||
6727 | /* | |
6728 | * We have freed the memory, now we should compact it to make | |
6729 | * allocation of the requested order possible. | |
6730 | */ | |
97a225e6 | 6731 | wakeup_kcompactd(pgdat, alloc_order, highest_zoneidx); |
fd901c95 | 6732 | |
f0bc0a60 | 6733 | remaining = schedule_timeout(HZ/10); |
38087d9b MG |
6734 | |
6735 | /* | |
97a225e6 | 6736 | * If woken prematurely then reset kswapd_highest_zoneidx and |
38087d9b MG |
6737 | * order. The values will either be from a wakeup request or |
6738 | * the previous request that slept prematurely. | |
6739 | */ | |
6740 | if (remaining) { | |
97a225e6 JK |
6741 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, |
6742 | kswapd_highest_zoneidx(pgdat, | |
6743 | highest_zoneidx)); | |
5644e1fb QC |
6744 | |
6745 | if (READ_ONCE(pgdat->kswapd_order) < reclaim_order) | |
6746 | WRITE_ONCE(pgdat->kswapd_order, reclaim_order); | |
38087d9b MG |
6747 | } |
6748 | ||
f0bc0a60 KM |
6749 | finish_wait(&pgdat->kswapd_wait, &wait); |
6750 | prepare_to_wait(&pgdat->kswapd_wait, &wait, TASK_INTERRUPTIBLE); | |
6751 | } | |
6752 | ||
6753 | /* | |
6754 | * After a short sleep, check if it was a premature sleep. If not, then | |
6755 | * go fully to sleep until explicitly woken up. | |
6756 | */ | |
d9f21d42 | 6757 | if (!remaining && |
97a225e6 | 6758 | prepare_kswapd_sleep(pgdat, reclaim_order, highest_zoneidx)) { |
f0bc0a60 KM |
6759 | trace_mm_vmscan_kswapd_sleep(pgdat->node_id); |
6760 | ||
6761 | /* | |
6762 | * vmstat counters are not perfectly accurate and the estimated | |
6763 | * value for counters such as NR_FREE_PAGES can deviate from the | |
6764 | * true value by nr_online_cpus * threshold. To avoid the zone | |
6765 | * watermarks being breached while under pressure, we reduce the | |
6766 | * per-cpu vmstat threshold while kswapd is awake and restore | |
6767 | * them before going back to sleep. | |
6768 | */ | |
6769 | set_pgdat_percpu_threshold(pgdat, calculate_normal_threshold); | |
1c7e7f6c AK |
6770 | |
6771 | if (!kthread_should_stop()) | |
6772 | schedule(); | |
6773 | ||
f0bc0a60 KM |
6774 | set_pgdat_percpu_threshold(pgdat, calculate_pressure_threshold); |
6775 | } else { | |
6776 | if (remaining) | |
6777 | count_vm_event(KSWAPD_LOW_WMARK_HIT_QUICKLY); | |
6778 | else | |
6779 | count_vm_event(KSWAPD_HIGH_WMARK_HIT_QUICKLY); | |
6780 | } | |
6781 | finish_wait(&pgdat->kswapd_wait, &wait); | |
6782 | } | |
6783 | ||
1da177e4 LT |
6784 | /* |
6785 | * The background pageout daemon, started as a kernel thread | |
4f98a2fe | 6786 | * from the init process. |
1da177e4 LT |
6787 | * |
6788 | * This basically trickles out pages so that we have _some_ | |
6789 | * free memory available even if there is no other activity | |
6790 | * that frees anything up. This is needed for things like routing | |
6791 | * etc, where we otherwise might have all activity going on in | |
6792 | * asynchronous contexts that cannot page things out. | |
6793 | * | |
6794 | * If there are applications that are active memory-allocators | |
6795 | * (most normal use), this basically shouldn't matter. | |
6796 | */ | |
6797 | static int kswapd(void *p) | |
6798 | { | |
e716f2eb | 6799 | unsigned int alloc_order, reclaim_order; |
97a225e6 | 6800 | unsigned int highest_zoneidx = MAX_NR_ZONES - 1; |
68d68ff6 | 6801 | pg_data_t *pgdat = (pg_data_t *)p; |
1da177e4 | 6802 | struct task_struct *tsk = current; |
a70f7302 | 6803 | const struct cpumask *cpumask = cpumask_of_node(pgdat->node_id); |
1da177e4 | 6804 | |
174596a0 | 6805 | if (!cpumask_empty(cpumask)) |
c5f59f08 | 6806 | set_cpus_allowed_ptr(tsk, cpumask); |
1da177e4 LT |
6807 | |
6808 | /* | |
6809 | * Tell the memory management that we're a "memory allocator", | |
6810 | * and that if we need more memory we should get access to it | |
6811 | * regardless (see "__alloc_pages()"). "kswapd" should | |
6812 | * never get caught in the normal page freeing logic. | |
6813 | * | |
6814 | * (Kswapd normally doesn't need memory anyway, but sometimes | |
6815 | * you need a small amount of memory in order to be able to | |
6816 | * page out something else, and this flag essentially protects | |
6817 | * us from recursively trying to free more memory as we're | |
6818 | * trying to free the first piece of memory in the first place). | |
6819 | */ | |
b698f0a1 | 6820 | tsk->flags |= PF_MEMALLOC | PF_KSWAPD; |
83144186 | 6821 | set_freezable(); |
1da177e4 | 6822 | |
5644e1fb | 6823 | WRITE_ONCE(pgdat->kswapd_order, 0); |
97a225e6 | 6824 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES); |
8cd7c588 | 6825 | atomic_set(&pgdat->nr_writeback_throttled, 0); |
1da177e4 | 6826 | for ( ; ; ) { |
6f6313d4 | 6827 | bool ret; |
3e1d1d28 | 6828 | |
5644e1fb | 6829 | alloc_order = reclaim_order = READ_ONCE(pgdat->kswapd_order); |
97a225e6 JK |
6830 | highest_zoneidx = kswapd_highest_zoneidx(pgdat, |
6831 | highest_zoneidx); | |
e716f2eb | 6832 | |
38087d9b MG |
6833 | kswapd_try_sleep: |
6834 | kswapd_try_to_sleep(pgdat, alloc_order, reclaim_order, | |
97a225e6 | 6835 | highest_zoneidx); |
215ddd66 | 6836 | |
97a225e6 | 6837 | /* Read the new order and highest_zoneidx */ |
2b47a24c | 6838 | alloc_order = READ_ONCE(pgdat->kswapd_order); |
97a225e6 JK |
6839 | highest_zoneidx = kswapd_highest_zoneidx(pgdat, |
6840 | highest_zoneidx); | |
5644e1fb | 6841 | WRITE_ONCE(pgdat->kswapd_order, 0); |
97a225e6 | 6842 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, MAX_NR_ZONES); |
1da177e4 | 6843 | |
8fe23e05 DR |
6844 | ret = try_to_freeze(); |
6845 | if (kthread_should_stop()) | |
6846 | break; | |
6847 | ||
6848 | /* | |
6849 | * We can speed up thawing tasks if we don't call balance_pgdat | |
6850 | * after returning from the refrigerator | |
6851 | */ | |
38087d9b MG |
6852 | if (ret) |
6853 | continue; | |
6854 | ||
6855 | /* | |
6856 | * Reclaim begins at the requested order but if a high-order | |
6857 | * reclaim fails then kswapd falls back to reclaiming for | |
6858 | * order-0. If that happens, kswapd will consider sleeping | |
6859 | * for the order it finished reclaiming at (reclaim_order) | |
6860 | * but kcompactd is woken to compact for the original | |
6861 | * request (alloc_order). | |
6862 | */ | |
97a225e6 | 6863 | trace_mm_vmscan_kswapd_wake(pgdat->node_id, highest_zoneidx, |
e5146b12 | 6864 | alloc_order); |
97a225e6 JK |
6865 | reclaim_order = balance_pgdat(pgdat, alloc_order, |
6866 | highest_zoneidx); | |
38087d9b MG |
6867 | if (reclaim_order < alloc_order) |
6868 | goto kswapd_try_sleep; | |
1da177e4 | 6869 | } |
b0a8cc58 | 6870 | |
b698f0a1 | 6871 | tsk->flags &= ~(PF_MEMALLOC | PF_KSWAPD); |
71abdc15 | 6872 | |
1da177e4 LT |
6873 | return 0; |
6874 | } | |
6875 | ||
6876 | /* | |
5ecd9d40 DR |
6877 | * A zone is low on free memory or too fragmented for high-order memory. If |
6878 | * kswapd should reclaim (direct reclaim is deferred), wake it up for the zone's | |
6879 | * pgdat. It will wake up kcompactd after reclaiming memory. If kswapd reclaim | |
6880 | * has failed or is not needed, still wake up kcompactd if only compaction is | |
6881 | * needed. | |
1da177e4 | 6882 | */ |
5ecd9d40 | 6883 | void wakeup_kswapd(struct zone *zone, gfp_t gfp_flags, int order, |
97a225e6 | 6884 | enum zone_type highest_zoneidx) |
1da177e4 LT |
6885 | { |
6886 | pg_data_t *pgdat; | |
5644e1fb | 6887 | enum zone_type curr_idx; |
1da177e4 | 6888 | |
6aa303de | 6889 | if (!managed_zone(zone)) |
1da177e4 LT |
6890 | return; |
6891 | ||
5ecd9d40 | 6892 | if (!cpuset_zone_allowed(zone, gfp_flags)) |
1da177e4 | 6893 | return; |
5644e1fb | 6894 | |
88f5acf8 | 6895 | pgdat = zone->zone_pgdat; |
97a225e6 | 6896 | curr_idx = READ_ONCE(pgdat->kswapd_highest_zoneidx); |
5644e1fb | 6897 | |
97a225e6 JK |
6898 | if (curr_idx == MAX_NR_ZONES || curr_idx < highest_zoneidx) |
6899 | WRITE_ONCE(pgdat->kswapd_highest_zoneidx, highest_zoneidx); | |
5644e1fb QC |
6900 | |
6901 | if (READ_ONCE(pgdat->kswapd_order) < order) | |
6902 | WRITE_ONCE(pgdat->kswapd_order, order); | |
dffcac2c | 6903 | |
8d0986e2 | 6904 | if (!waitqueue_active(&pgdat->kswapd_wait)) |
1da177e4 | 6905 | return; |
e1a55637 | 6906 | |
5ecd9d40 DR |
6907 | /* Hopeless node, leave it to direct reclaim if possible */ |
6908 | if (pgdat->kswapd_failures >= MAX_RECLAIM_RETRIES || | |
97a225e6 JK |
6909 | (pgdat_balanced(pgdat, order, highest_zoneidx) && |
6910 | !pgdat_watermark_boosted(pgdat, highest_zoneidx))) { | |
5ecd9d40 DR |
6911 | /* |
6912 | * There may be plenty of free memory available, but it's too | |
6913 | * fragmented for high-order allocations. Wake up kcompactd | |
6914 | * and rely on compaction_suitable() to determine if it's | |
6915 | * needed. If it fails, it will defer subsequent attempts to | |
6916 | * ratelimit its work. | |
6917 | */ | |
6918 | if (!(gfp_flags & __GFP_DIRECT_RECLAIM)) | |
97a225e6 | 6919 | wakeup_kcompactd(pgdat, order, highest_zoneidx); |
e716f2eb | 6920 | return; |
5ecd9d40 | 6921 | } |
88f5acf8 | 6922 | |
97a225e6 | 6923 | trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, highest_zoneidx, order, |
5ecd9d40 | 6924 | gfp_flags); |
8d0986e2 | 6925 | wake_up_interruptible(&pgdat->kswapd_wait); |
1da177e4 LT |
6926 | } |
6927 | ||
c6f37f12 | 6928 | #ifdef CONFIG_HIBERNATION |
1da177e4 | 6929 | /* |
7b51755c | 6930 | * Try to free `nr_to_reclaim' of memory, system-wide, and return the number of |
d6277db4 RW |
6931 | * freed pages. |
6932 | * | |
6933 | * Rather than trying to age LRUs the aim is to preserve the overall | |
6934 | * LRU order by reclaiming preferentially | |
6935 | * inactive > active > active referenced > active mapped | |
1da177e4 | 6936 | */ |
7b51755c | 6937 | unsigned long shrink_all_memory(unsigned long nr_to_reclaim) |
1da177e4 | 6938 | { |
d6277db4 | 6939 | struct scan_control sc = { |
ee814fe2 | 6940 | .nr_to_reclaim = nr_to_reclaim, |
7b51755c | 6941 | .gfp_mask = GFP_HIGHUSER_MOVABLE, |
b2e18757 | 6942 | .reclaim_idx = MAX_NR_ZONES - 1, |
ee814fe2 | 6943 | .priority = DEF_PRIORITY, |
d6277db4 | 6944 | .may_writepage = 1, |
ee814fe2 JW |
6945 | .may_unmap = 1, |
6946 | .may_swap = 1, | |
7b51755c | 6947 | .hibernation_mode = 1, |
1da177e4 | 6948 | }; |
a09ed5e0 | 6949 | struct zonelist *zonelist = node_zonelist(numa_node_id(), sc.gfp_mask); |
7b51755c | 6950 | unsigned long nr_reclaimed; |
499118e9 | 6951 | unsigned int noreclaim_flag; |
1da177e4 | 6952 | |
d92a8cfc | 6953 | fs_reclaim_acquire(sc.gfp_mask); |
93781325 | 6954 | noreclaim_flag = memalloc_noreclaim_save(); |
1732d2b0 | 6955 | set_task_reclaim_state(current, &sc.reclaim_state); |
d6277db4 | 6956 | |
3115cd91 | 6957 | nr_reclaimed = do_try_to_free_pages(zonelist, &sc); |
d979677c | 6958 | |
1732d2b0 | 6959 | set_task_reclaim_state(current, NULL); |
499118e9 | 6960 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 6961 | fs_reclaim_release(sc.gfp_mask); |
d6277db4 | 6962 | |
7b51755c | 6963 | return nr_reclaimed; |
1da177e4 | 6964 | } |
c6f37f12 | 6965 | #endif /* CONFIG_HIBERNATION */ |
1da177e4 | 6966 | |
3218ae14 YG |
6967 | /* |
6968 | * This kswapd start function will be called by init and node-hot-add. | |
3218ae14 | 6969 | */ |
b87c517a | 6970 | void kswapd_run(int nid) |
3218ae14 YG |
6971 | { |
6972 | pg_data_t *pgdat = NODE_DATA(nid); | |
3218ae14 | 6973 | |
b4a0215e KW |
6974 | pgdat_kswapd_lock(pgdat); |
6975 | if (!pgdat->kswapd) { | |
6976 | pgdat->kswapd = kthread_run(kswapd, pgdat, "kswapd%d", nid); | |
6977 | if (IS_ERR(pgdat->kswapd)) { | |
6978 | /* failure at boot is fatal */ | |
6979 | BUG_ON(system_state < SYSTEM_RUNNING); | |
6980 | pr_err("Failed to start kswapd on node %d\n", nid); | |
6981 | pgdat->kswapd = NULL; | |
6982 | } | |
3218ae14 | 6983 | } |
b4a0215e | 6984 | pgdat_kswapd_unlock(pgdat); |
3218ae14 YG |
6985 | } |
6986 | ||
8fe23e05 | 6987 | /* |
d8adde17 | 6988 | * Called by memory hotplug when all memory in a node is offlined. Caller must |
e8da368a | 6989 | * be holding mem_hotplug_begin/done(). |
8fe23e05 DR |
6990 | */ |
6991 | void kswapd_stop(int nid) | |
6992 | { | |
b4a0215e KW |
6993 | pg_data_t *pgdat = NODE_DATA(nid); |
6994 | struct task_struct *kswapd; | |
8fe23e05 | 6995 | |
b4a0215e KW |
6996 | pgdat_kswapd_lock(pgdat); |
6997 | kswapd = pgdat->kswapd; | |
d8adde17 | 6998 | if (kswapd) { |
8fe23e05 | 6999 | kthread_stop(kswapd); |
b4a0215e | 7000 | pgdat->kswapd = NULL; |
d8adde17 | 7001 | } |
b4a0215e | 7002 | pgdat_kswapd_unlock(pgdat); |
8fe23e05 DR |
7003 | } |
7004 | ||
1da177e4 LT |
7005 | static int __init kswapd_init(void) |
7006 | { | |
6b700b5b | 7007 | int nid; |
69e05944 | 7008 | |
1da177e4 | 7009 | swap_setup(); |
48fb2e24 | 7010 | for_each_node_state(nid, N_MEMORY) |
3218ae14 | 7011 | kswapd_run(nid); |
1da177e4 LT |
7012 | return 0; |
7013 | } | |
7014 | ||
7015 | module_init(kswapd_init) | |
9eeff239 CL |
7016 | |
7017 | #ifdef CONFIG_NUMA | |
7018 | /* | |
a5f5f91d | 7019 | * Node reclaim mode |
9eeff239 | 7020 | * |
a5f5f91d | 7021 | * If non-zero call node_reclaim when the number of free pages falls below |
9eeff239 | 7022 | * the watermarks. |
9eeff239 | 7023 | */ |
a5f5f91d | 7024 | int node_reclaim_mode __read_mostly; |
9eeff239 | 7025 | |
a92f7126 | 7026 | /* |
a5f5f91d | 7027 | * Priority for NODE_RECLAIM. This determines the fraction of pages |
a92f7126 CL |
7028 | * of a node considered for each zone_reclaim. 4 scans 1/16th of |
7029 | * a zone. | |
7030 | */ | |
a5f5f91d | 7031 | #define NODE_RECLAIM_PRIORITY 4 |
a92f7126 | 7032 | |
9614634f | 7033 | /* |
a5f5f91d | 7034 | * Percentage of pages in a zone that must be unmapped for node_reclaim to |
9614634f CL |
7035 | * occur. |
7036 | */ | |
7037 | int sysctl_min_unmapped_ratio = 1; | |
7038 | ||
0ff38490 CL |
7039 | /* |
7040 | * If the number of slab pages in a zone grows beyond this percentage then | |
7041 | * slab reclaim needs to occur. | |
7042 | */ | |
7043 | int sysctl_min_slab_ratio = 5; | |
7044 | ||
11fb9989 | 7045 | static inline unsigned long node_unmapped_file_pages(struct pglist_data *pgdat) |
90afa5de | 7046 | { |
11fb9989 MG |
7047 | unsigned long file_mapped = node_page_state(pgdat, NR_FILE_MAPPED); |
7048 | unsigned long file_lru = node_page_state(pgdat, NR_INACTIVE_FILE) + | |
7049 | node_page_state(pgdat, NR_ACTIVE_FILE); | |
90afa5de MG |
7050 | |
7051 | /* | |
7052 | * It's possible for there to be more file mapped pages than | |
7053 | * accounted for by the pages on the file LRU lists because | |
7054 | * tmpfs pages accounted for as ANON can also be FILE_MAPPED | |
7055 | */ | |
7056 | return (file_lru > file_mapped) ? (file_lru - file_mapped) : 0; | |
7057 | } | |
7058 | ||
7059 | /* Work out how many page cache pages we can reclaim in this reclaim_mode */ | |
a5f5f91d | 7060 | static unsigned long node_pagecache_reclaimable(struct pglist_data *pgdat) |
90afa5de | 7061 | { |
d031a157 AM |
7062 | unsigned long nr_pagecache_reclaimable; |
7063 | unsigned long delta = 0; | |
90afa5de MG |
7064 | |
7065 | /* | |
95bbc0c7 | 7066 | * If RECLAIM_UNMAP is set, then all file pages are considered |
90afa5de | 7067 | * potentially reclaimable. Otherwise, we have to worry about |
11fb9989 | 7068 | * pages like swapcache and node_unmapped_file_pages() provides |
90afa5de MG |
7069 | * a better estimate |
7070 | */ | |
a5f5f91d MG |
7071 | if (node_reclaim_mode & RECLAIM_UNMAP) |
7072 | nr_pagecache_reclaimable = node_page_state(pgdat, NR_FILE_PAGES); | |
90afa5de | 7073 | else |
a5f5f91d | 7074 | nr_pagecache_reclaimable = node_unmapped_file_pages(pgdat); |
90afa5de MG |
7075 | |
7076 | /* If we can't clean pages, remove dirty pages from consideration */ | |
a5f5f91d MG |
7077 | if (!(node_reclaim_mode & RECLAIM_WRITE)) |
7078 | delta += node_page_state(pgdat, NR_FILE_DIRTY); | |
90afa5de MG |
7079 | |
7080 | /* Watch for any possible underflows due to delta */ | |
7081 | if (unlikely(delta > nr_pagecache_reclaimable)) | |
7082 | delta = nr_pagecache_reclaimable; | |
7083 | ||
7084 | return nr_pagecache_reclaimable - delta; | |
7085 | } | |
7086 | ||
9eeff239 | 7087 | /* |
a5f5f91d | 7088 | * Try to free up some pages from this node through reclaim. |
9eeff239 | 7089 | */ |
a5f5f91d | 7090 | static int __node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
9eeff239 | 7091 | { |
7fb2d46d | 7092 | /* Minimum pages needed in order to stay on node */ |
69e05944 | 7093 | const unsigned long nr_pages = 1 << order; |
9eeff239 | 7094 | struct task_struct *p = current; |
499118e9 | 7095 | unsigned int noreclaim_flag; |
179e9639 | 7096 | struct scan_control sc = { |
62b726c1 | 7097 | .nr_to_reclaim = max(nr_pages, SWAP_CLUSTER_MAX), |
f2f43e56 | 7098 | .gfp_mask = current_gfp_context(gfp_mask), |
bd2f6199 | 7099 | .order = order, |
a5f5f91d MG |
7100 | .priority = NODE_RECLAIM_PRIORITY, |
7101 | .may_writepage = !!(node_reclaim_mode & RECLAIM_WRITE), | |
7102 | .may_unmap = !!(node_reclaim_mode & RECLAIM_UNMAP), | |
ee814fe2 | 7103 | .may_swap = 1, |
f2f43e56 | 7104 | .reclaim_idx = gfp_zone(gfp_mask), |
179e9639 | 7105 | }; |
57f29762 | 7106 | unsigned long pflags; |
9eeff239 | 7107 | |
132bb8cf YS |
7108 | trace_mm_vmscan_node_reclaim_begin(pgdat->node_id, order, |
7109 | sc.gfp_mask); | |
7110 | ||
9eeff239 | 7111 | cond_resched(); |
57f29762 | 7112 | psi_memstall_enter(&pflags); |
93781325 | 7113 | fs_reclaim_acquire(sc.gfp_mask); |
d4f7796e | 7114 | /* |
95bbc0c7 | 7115 | * We need to be able to allocate from the reserves for RECLAIM_UNMAP |
d4f7796e | 7116 | */ |
499118e9 | 7117 | noreclaim_flag = memalloc_noreclaim_save(); |
1732d2b0 | 7118 | set_task_reclaim_state(p, &sc.reclaim_state); |
c84db23c | 7119 | |
d8ff6fde ML |
7120 | if (node_pagecache_reclaimable(pgdat) > pgdat->min_unmapped_pages || |
7121 | node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) > pgdat->min_slab_pages) { | |
0ff38490 | 7122 | /* |
894befec | 7123 | * Free memory by calling shrink node with increasing |
0ff38490 CL |
7124 | * priorities until we have enough memory freed. |
7125 | */ | |
0ff38490 | 7126 | do { |
970a39a3 | 7127 | shrink_node(pgdat, &sc); |
9e3b2f8c | 7128 | } while (sc.nr_reclaimed < nr_pages && --sc.priority >= 0); |
0ff38490 | 7129 | } |
c84db23c | 7130 | |
1732d2b0 | 7131 | set_task_reclaim_state(p, NULL); |
499118e9 | 7132 | memalloc_noreclaim_restore(noreclaim_flag); |
93781325 | 7133 | fs_reclaim_release(sc.gfp_mask); |
57f29762 | 7134 | psi_memstall_leave(&pflags); |
132bb8cf YS |
7135 | |
7136 | trace_mm_vmscan_node_reclaim_end(sc.nr_reclaimed); | |
7137 | ||
a79311c1 | 7138 | return sc.nr_reclaimed >= nr_pages; |
9eeff239 | 7139 | } |
179e9639 | 7140 | |
a5f5f91d | 7141 | int node_reclaim(struct pglist_data *pgdat, gfp_t gfp_mask, unsigned int order) |
179e9639 | 7142 | { |
d773ed6b | 7143 | int ret; |
179e9639 AM |
7144 | |
7145 | /* | |
a5f5f91d | 7146 | * Node reclaim reclaims unmapped file backed pages and |
0ff38490 | 7147 | * slab pages if we are over the defined limits. |
34aa1330 | 7148 | * |
9614634f CL |
7149 | * A small portion of unmapped file backed pages is needed for |
7150 | * file I/O otherwise pages read by file I/O will be immediately | |
a5f5f91d MG |
7151 | * thrown out if the node is overallocated. So we do not reclaim |
7152 | * if less than a specified percentage of the node is used by | |
9614634f | 7153 | * unmapped file backed pages. |
179e9639 | 7154 | */ |
a5f5f91d | 7155 | if (node_pagecache_reclaimable(pgdat) <= pgdat->min_unmapped_pages && |
d42f3245 RG |
7156 | node_page_state_pages(pgdat, NR_SLAB_RECLAIMABLE_B) <= |
7157 | pgdat->min_slab_pages) | |
a5f5f91d | 7158 | return NODE_RECLAIM_FULL; |
179e9639 AM |
7159 | |
7160 | /* | |
d773ed6b | 7161 | * Do not scan if the allocation should not be delayed. |
179e9639 | 7162 | */ |
d0164adc | 7163 | if (!gfpflags_allow_blocking(gfp_mask) || (current->flags & PF_MEMALLOC)) |
a5f5f91d | 7164 | return NODE_RECLAIM_NOSCAN; |
179e9639 AM |
7165 | |
7166 | /* | |
a5f5f91d | 7167 | * Only run node reclaim on the local node or on nodes that do not |
179e9639 AM |
7168 | * have associated processors. This will favor the local processor |
7169 | * over remote processors and spread off node memory allocations | |
7170 | * as wide as possible. | |
7171 | */ | |
a5f5f91d MG |
7172 | if (node_state(pgdat->node_id, N_CPU) && pgdat->node_id != numa_node_id()) |
7173 | return NODE_RECLAIM_NOSCAN; | |
d773ed6b | 7174 | |
a5f5f91d MG |
7175 | if (test_and_set_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags)) |
7176 | return NODE_RECLAIM_NOSCAN; | |
fa5e084e | 7177 | |
a5f5f91d MG |
7178 | ret = __node_reclaim(pgdat, gfp_mask, order); |
7179 | clear_bit(PGDAT_RECLAIM_LOCKED, &pgdat->flags); | |
d773ed6b | 7180 | |
24cf7251 MG |
7181 | if (!ret) |
7182 | count_vm_event(PGSCAN_ZONE_RECLAIM_FAILED); | |
7183 | ||
d773ed6b | 7184 | return ret; |
179e9639 | 7185 | } |
9eeff239 | 7186 | #endif |
894bc310 | 7187 | |
77414d19 MWO |
7188 | void check_move_unevictable_pages(struct pagevec *pvec) |
7189 | { | |
7190 | struct folio_batch fbatch; | |
7191 | unsigned i; | |
7192 | ||
7193 | folio_batch_init(&fbatch); | |
7194 | for (i = 0; i < pvec->nr; i++) { | |
7195 | struct page *page = pvec->pages[i]; | |
7196 | ||
7197 | if (PageTransTail(page)) | |
7198 | continue; | |
7199 | folio_batch_add(&fbatch, page_folio(page)); | |
7200 | } | |
7201 | check_move_unevictable_folios(&fbatch); | |
7202 | } | |
7203 | EXPORT_SYMBOL_GPL(check_move_unevictable_pages); | |
7204 | ||
89e004ea | 7205 | /** |
77414d19 MWO |
7206 | * check_move_unevictable_folios - Move evictable folios to appropriate zone |
7207 | * lru list | |
7208 | * @fbatch: Batch of lru folios to check. | |
89e004ea | 7209 | * |
77414d19 | 7210 | * Checks folios for evictability, if an evictable folio is in the unevictable |
64e3d12f | 7211 | * lru list, moves it to the appropriate evictable lru list. This function |
77414d19 | 7212 | * should be only used for lru folios. |
89e004ea | 7213 | */ |
77414d19 | 7214 | void check_move_unevictable_folios(struct folio_batch *fbatch) |
89e004ea | 7215 | { |
6168d0da | 7216 | struct lruvec *lruvec = NULL; |
24513264 HD |
7217 | int pgscanned = 0; |
7218 | int pgrescued = 0; | |
7219 | int i; | |
89e004ea | 7220 | |
77414d19 MWO |
7221 | for (i = 0; i < fbatch->nr; i++) { |
7222 | struct folio *folio = fbatch->folios[i]; | |
7223 | int nr_pages = folio_nr_pages(folio); | |
8d8869ca | 7224 | |
8d8869ca | 7225 | pgscanned += nr_pages; |
89e004ea | 7226 | |
77414d19 MWO |
7227 | /* block memcg migration while the folio moves between lrus */ |
7228 | if (!folio_test_clear_lru(folio)) | |
d25b5bd8 AS |
7229 | continue; |
7230 | ||
0de340cb | 7231 | lruvec = folio_lruvec_relock_irq(folio, lruvec); |
77414d19 MWO |
7232 | if (folio_evictable(folio) && folio_test_unevictable(folio)) { |
7233 | lruvec_del_folio(lruvec, folio); | |
7234 | folio_clear_unevictable(folio); | |
7235 | lruvec_add_folio(lruvec, folio); | |
8d8869ca | 7236 | pgrescued += nr_pages; |
89e004ea | 7237 | } |
77414d19 | 7238 | folio_set_lru(folio); |
24513264 | 7239 | } |
89e004ea | 7240 | |
6168d0da | 7241 | if (lruvec) { |
24513264 HD |
7242 | __count_vm_events(UNEVICTABLE_PGRESCUED, pgrescued); |
7243 | __count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
6168d0da | 7244 | unlock_page_lruvec_irq(lruvec); |
d25b5bd8 AS |
7245 | } else if (pgscanned) { |
7246 | count_vm_events(UNEVICTABLE_PGSCANNED, pgscanned); | |
89e004ea | 7247 | } |
89e004ea | 7248 | } |
77414d19 | 7249 | EXPORT_SYMBOL_GPL(check_move_unevictable_folios); |