Commit | Line | Data |
---|---|---|
d7067512 JB |
1 | /* |
2 | * Block rq-qos base io controller | |
3 | * | |
4 | * This works similar to wbt with a few exceptions | |
5 | * | |
6 | * - It's bio based, so the latency covers the whole block layer in addition to | |
7 | * the actual io. | |
8 | * - We will throttle all IO that comes in here if we need to. | |
9 | * - We use the mean latency over the 100ms window. This is because writes can | |
10 | * be particularly fast, which could give us a false sense of the impact of | |
11 | * other workloads on our protected workload. | |
a284390b JB |
12 | * - By default there's no throttling, we set the queue_depth to UINT_MAX so |
13 | * that we can have as many outstanding bio's as we're allowed to. Only at | |
d7067512 JB |
14 | * throttle time do we pay attention to the actual queue depth. |
15 | * | |
16 | * The hierarchy works like the cpu controller does, we track the latency at | |
17 | * every configured node, and each configured node has it's own independent | |
18 | * queue depth. This means that we only care about our latency targets at the | |
19 | * peer level. Some group at the bottom of the hierarchy isn't going to affect | |
20 | * a group at the end of some other path if we're only configred at leaf level. | |
21 | * | |
22 | * Consider the following | |
23 | * | |
24 | * root blkg | |
25 | * / \ | |
26 | * fast (target=5ms) slow (target=10ms) | |
27 | * / \ / \ | |
28 | * a b normal(15ms) unloved | |
29 | * | |
30 | * "a" and "b" have no target, but their combined io under "fast" cannot exceed | |
31 | * an average latency of 5ms. If it does then we will throttle the "slow" | |
32 | * group. In the case of "normal", if it exceeds its 15ms target, we will | |
33 | * throttle "unloved", but nobody else. | |
34 | * | |
35 | * In this example "fast", "slow", and "normal" will be the only groups actually | |
36 | * accounting their io latencies. We have to walk up the heirarchy to the root | |
37 | * on every submit and complete so we can do the appropriate stat recording and | |
38 | * adjust the queue depth of ourselves if needed. | |
39 | * | |
40 | * There are 2 ways we throttle IO. | |
41 | * | |
42 | * 1) Queue depth throttling. As we throttle down we will adjust the maximum | |
43 | * number of IO's we're allowed to have in flight. This starts at (u64)-1 down | |
44 | * to 1. If the group is only ever submitting IO for itself then this is the | |
45 | * only way we throttle. | |
46 | * | |
47 | * 2) Induced delay throttling. This is for the case that a group is generating | |
48 | * IO that has to be issued by the root cg to avoid priority inversion. So think | |
49 | * REQ_META or REQ_SWAP. If we are already at qd == 1 and we're getting a lot | |
50 | * of work done for us on behalf of the root cg and are being asked to scale | |
51 | * down more then we induce a latency at userspace return. We accumulate the | |
52 | * total amount of time we need to be punished by doing | |
53 | * | |
54 | * total_time += min_lat_nsec - actual_io_completion | |
55 | * | |
56 | * and then at throttle time will do | |
57 | * | |
58 | * throttle_time = min(total_time, NSEC_PER_SEC) | |
59 | * | |
60 | * This induced delay will throttle back the activity that is generating the | |
61 | * root cg issued io's, wethere that's some metadata intensive operation or the | |
62 | * group is using so much memory that it is pushing us into swap. | |
63 | * | |
64 | * Copyright (C) 2018 Josef Bacik | |
65 | */ | |
66 | #include <linux/kernel.h> | |
67 | #include <linux/blk_types.h> | |
68 | #include <linux/backing-dev.h> | |
69 | #include <linux/module.h> | |
70 | #include <linux/timer.h> | |
71 | #include <linux/memcontrol.h> | |
c480bcf9 | 72 | #include <linux/sched/loadavg.h> |
d7067512 JB |
73 | #include <linux/sched/signal.h> |
74 | #include <trace/events/block.h> | |
75 | #include "blk-rq-qos.h" | |
76 | #include "blk-stat.h" | |
77 | ||
78 | #define DEFAULT_SCALE_COOKIE 1000000U | |
79 | ||
80 | static struct blkcg_policy blkcg_policy_iolatency; | |
81 | struct iolatency_grp; | |
82 | ||
83 | struct blk_iolatency { | |
84 | struct rq_qos rqos; | |
85 | struct timer_list timer; | |
86 | atomic_t enabled; | |
87 | }; | |
88 | ||
89 | static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos) | |
90 | { | |
91 | return container_of(rqos, struct blk_iolatency, rqos); | |
92 | } | |
93 | ||
94 | static inline bool blk_iolatency_enabled(struct blk_iolatency *blkiolat) | |
95 | { | |
96 | return atomic_read(&blkiolat->enabled) > 0; | |
97 | } | |
98 | ||
99 | struct child_latency_info { | |
100 | spinlock_t lock; | |
101 | ||
102 | /* Last time we adjusted the scale of everybody. */ | |
103 | u64 last_scale_event; | |
104 | ||
105 | /* The latency that we missed. */ | |
106 | u64 scale_lat; | |
107 | ||
108 | /* Total io's from all of our children for the last summation. */ | |
109 | u64 nr_samples; | |
110 | ||
111 | /* The guy who actually changed the latency numbers. */ | |
112 | struct iolatency_grp *scale_grp; | |
113 | ||
114 | /* Cookie to tell if we need to scale up or down. */ | |
115 | atomic_t scale_cookie; | |
116 | }; | |
117 | ||
1fa2840e JB |
118 | struct percentile_stats { |
119 | u64 total; | |
120 | u64 missed; | |
121 | }; | |
122 | ||
123 | struct latency_stat { | |
124 | union { | |
125 | struct percentile_stats ps; | |
126 | struct blk_rq_stat rqs; | |
127 | }; | |
128 | }; | |
129 | ||
d7067512 JB |
130 | struct iolatency_grp { |
131 | struct blkg_policy_data pd; | |
1fa2840e | 132 | struct latency_stat __percpu *stats; |
451bb7c3 | 133 | struct latency_stat cur_stat; |
d7067512 JB |
134 | struct blk_iolatency *blkiolat; |
135 | struct rq_depth rq_depth; | |
136 | struct rq_wait rq_wait; | |
137 | atomic64_t window_start; | |
138 | atomic_t scale_cookie; | |
139 | u64 min_lat_nsec; | |
140 | u64 cur_win_nsec; | |
141 | ||
142 | /* total running average of our io latency. */ | |
c480bcf9 | 143 | u64 lat_avg; |
d7067512 JB |
144 | |
145 | /* Our current number of IO's for the last summation. */ | |
146 | u64 nr_samples; | |
147 | ||
1fa2840e | 148 | bool ssd; |
d7067512 JB |
149 | struct child_latency_info child_lat; |
150 | }; | |
151 | ||
c480bcf9 DZF |
152 | #define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC) |
153 | #define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC | |
154 | /* | |
155 | * These are the constants used to fake the fixed-point moving average | |
8508cf3f | 156 | * calculation just like load average. The call to calc_load() folds |
c480bcf9 DZF |
157 | * (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg. The sampling |
158 | * window size is bucketed to try to approximately calculate average | |
159 | * latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows | |
160 | * elapse immediately. Note, windows only elapse with IO activity. Idle | |
161 | * periods extend the most recent window. | |
162 | */ | |
163 | #define BLKIOLATENCY_NR_EXP_FACTORS 5 | |
164 | #define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \ | |
165 | (BLKIOLATENCY_NR_EXP_FACTORS - 1)) | |
166 | static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = { | |
167 | 2045, // exp(1/600) - 600 samples | |
168 | 2039, // exp(1/240) - 240 samples | |
169 | 2031, // exp(1/120) - 120 samples | |
170 | 2023, // exp(1/80) - 80 samples | |
171 | 2014, // exp(1/60) - 60 samples | |
172 | }; | |
173 | ||
d7067512 JB |
174 | static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd) |
175 | { | |
176 | return pd ? container_of(pd, struct iolatency_grp, pd) : NULL; | |
177 | } | |
178 | ||
179 | static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg) | |
180 | { | |
181 | return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency)); | |
182 | } | |
183 | ||
184 | static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat) | |
185 | { | |
186 | return pd_to_blkg(&iolat->pd); | |
187 | } | |
188 | ||
1fa2840e JB |
189 | static inline void latency_stat_init(struct iolatency_grp *iolat, |
190 | struct latency_stat *stat) | |
191 | { | |
192 | if (iolat->ssd) { | |
193 | stat->ps.total = 0; | |
194 | stat->ps.missed = 0; | |
195 | } else | |
196 | blk_rq_stat_init(&stat->rqs); | |
197 | } | |
198 | ||
199 | static inline void latency_stat_sum(struct iolatency_grp *iolat, | |
200 | struct latency_stat *sum, | |
201 | struct latency_stat *stat) | |
202 | { | |
203 | if (iolat->ssd) { | |
204 | sum->ps.total += stat->ps.total; | |
205 | sum->ps.missed += stat->ps.missed; | |
206 | } else | |
207 | blk_rq_stat_sum(&sum->rqs, &stat->rqs); | |
208 | } | |
209 | ||
210 | static inline void latency_stat_record_time(struct iolatency_grp *iolat, | |
211 | u64 req_time) | |
212 | { | |
213 | struct latency_stat *stat = get_cpu_ptr(iolat->stats); | |
214 | if (iolat->ssd) { | |
215 | if (req_time >= iolat->min_lat_nsec) | |
216 | stat->ps.missed++; | |
217 | stat->ps.total++; | |
218 | } else | |
219 | blk_rq_stat_add(&stat->rqs, req_time); | |
220 | put_cpu_ptr(stat); | |
221 | } | |
222 | ||
223 | static inline bool latency_sum_ok(struct iolatency_grp *iolat, | |
224 | struct latency_stat *stat) | |
225 | { | |
226 | if (iolat->ssd) { | |
227 | u64 thresh = div64_u64(stat->ps.total, 10); | |
228 | thresh = max(thresh, 1ULL); | |
229 | return stat->ps.missed < thresh; | |
230 | } | |
231 | return stat->rqs.mean <= iolat->min_lat_nsec; | |
232 | } | |
233 | ||
234 | static inline u64 latency_stat_samples(struct iolatency_grp *iolat, | |
235 | struct latency_stat *stat) | |
236 | { | |
237 | if (iolat->ssd) | |
238 | return stat->ps.total; | |
239 | return stat->rqs.nr_samples; | |
240 | } | |
241 | ||
242 | static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat, | |
243 | struct latency_stat *stat) | |
244 | { | |
245 | int exp_idx; | |
246 | ||
247 | if (iolat->ssd) | |
248 | return; | |
249 | ||
250 | /* | |
8508cf3f | 251 | * calc_load() takes in a number stored in fixed point representation. |
1fa2840e JB |
252 | * Because we are using this for IO time in ns, the values stored |
253 | * are significantly larger than the FIXED_1 denominator (2048). | |
254 | * Therefore, rounding errors in the calculation are negligible and | |
255 | * can be ignored. | |
256 | */ | |
257 | exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1, | |
258 | div64_u64(iolat->cur_win_nsec, | |
259 | BLKIOLATENCY_EXP_BUCKET_SIZE)); | |
8508cf3f JW |
260 | iolat->lat_avg = calc_load(iolat->lat_avg, |
261 | iolatency_exp_factors[exp_idx], | |
262 | stat->rqs.mean); | |
1fa2840e JB |
263 | } |
264 | ||
d7067512 JB |
265 | static inline bool iolatency_may_queue(struct iolatency_grp *iolat, |
266 | wait_queue_entry_t *wait, | |
267 | bool first_block) | |
268 | { | |
269 | struct rq_wait *rqw = &iolat->rq_wait; | |
270 | ||
271 | if (first_block && waitqueue_active(&rqw->wait) && | |
272 | rqw->wait.head.next != &wait->entry) | |
273 | return false; | |
274 | return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth); | |
275 | } | |
276 | ||
277 | static void __blkcg_iolatency_throttle(struct rq_qos *rqos, | |
278 | struct iolatency_grp *iolat, | |
279 | spinlock_t *lock, bool issue_as_root, | |
280 | bool use_memdelay) | |
281 | __releases(lock) | |
282 | __acquires(lock) | |
283 | { | |
284 | struct rq_wait *rqw = &iolat->rq_wait; | |
285 | unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay); | |
286 | DEFINE_WAIT(wait); | |
287 | bool first_block = true; | |
288 | ||
289 | if (use_delay) | |
290 | blkcg_schedule_throttle(rqos->q, use_memdelay); | |
291 | ||
292 | /* | |
293 | * To avoid priority inversions we want to just take a slot if we are | |
294 | * issuing as root. If we're being killed off there's no point in | |
295 | * delaying things, we may have been killed by OOM so throttling may | |
296 | * make recovery take even longer, so just let the IO's through so the | |
297 | * task can go away. | |
298 | */ | |
299 | if (issue_as_root || fatal_signal_pending(current)) { | |
300 | atomic_inc(&rqw->inflight); | |
301 | return; | |
302 | } | |
303 | ||
304 | if (iolatency_may_queue(iolat, &wait, first_block)) | |
305 | return; | |
306 | ||
307 | do { | |
308 | prepare_to_wait_exclusive(&rqw->wait, &wait, | |
309 | TASK_UNINTERRUPTIBLE); | |
310 | ||
311 | if (iolatency_may_queue(iolat, &wait, first_block)) | |
312 | break; | |
313 | first_block = false; | |
314 | ||
315 | if (lock) { | |
316 | spin_unlock_irq(lock); | |
317 | io_schedule(); | |
318 | spin_lock_irq(lock); | |
319 | } else { | |
320 | io_schedule(); | |
321 | } | |
322 | } while (1); | |
323 | ||
324 | finish_wait(&rqw->wait, &wait); | |
325 | } | |
326 | ||
327 | #define SCALE_DOWN_FACTOR 2 | |
328 | #define SCALE_UP_FACTOR 4 | |
329 | ||
330 | static inline unsigned long scale_amount(unsigned long qd, bool up) | |
331 | { | |
332 | return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL); | |
333 | } | |
334 | ||
335 | /* | |
336 | * We scale the qd down faster than we scale up, so we need to use this helper | |
337 | * to adjust the scale_cookie accordingly so we don't prematurely get | |
338 | * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much. | |
339 | * | |
340 | * Each group has their own local copy of the last scale cookie they saw, so if | |
341 | * the global scale cookie goes up or down they know which way they need to go | |
342 | * based on their last knowledge of it. | |
343 | */ | |
344 | static void scale_cookie_change(struct blk_iolatency *blkiolat, | |
345 | struct child_latency_info *lat_info, | |
346 | bool up) | |
347 | { | |
ff4cee08 | 348 | unsigned long qd = blkiolat->rqos.q->nr_requests; |
d7067512 JB |
349 | unsigned long scale = scale_amount(qd, up); |
350 | unsigned long old = atomic_read(&lat_info->scale_cookie); | |
351 | unsigned long max_scale = qd << 1; | |
352 | unsigned long diff = 0; | |
353 | ||
354 | if (old < DEFAULT_SCALE_COOKIE) | |
355 | diff = DEFAULT_SCALE_COOKIE - old; | |
356 | ||
357 | if (up) { | |
358 | if (scale + old > DEFAULT_SCALE_COOKIE) | |
359 | atomic_set(&lat_info->scale_cookie, | |
360 | DEFAULT_SCALE_COOKIE); | |
361 | else if (diff > qd) | |
362 | atomic_inc(&lat_info->scale_cookie); | |
363 | else | |
364 | atomic_add(scale, &lat_info->scale_cookie); | |
365 | } else { | |
366 | /* | |
367 | * We don't want to dig a hole so deep that it takes us hours to | |
368 | * dig out of it. Just enough that we don't throttle/unthrottle | |
369 | * with jagged workloads but can still unthrottle once pressure | |
370 | * has sufficiently dissipated. | |
371 | */ | |
372 | if (diff > qd) { | |
373 | if (diff < max_scale) | |
374 | atomic_dec(&lat_info->scale_cookie); | |
375 | } else { | |
376 | atomic_sub(scale, &lat_info->scale_cookie); | |
377 | } | |
378 | } | |
379 | } | |
380 | ||
381 | /* | |
382 | * Change the queue depth of the iolatency_grp. We add/subtract 1/16th of the | |
383 | * queue depth at a time so we don't get wild swings and hopefully dial in to | |
384 | * fairer distribution of the overall queue depth. | |
385 | */ | |
386 | static void scale_change(struct iolatency_grp *iolat, bool up) | |
387 | { | |
ff4cee08 | 388 | unsigned long qd = iolat->blkiolat->rqos.q->nr_requests; |
d7067512 JB |
389 | unsigned long scale = scale_amount(qd, up); |
390 | unsigned long old = iolat->rq_depth.max_depth; | |
d7067512 JB |
391 | |
392 | if (old > qd) | |
393 | old = qd; | |
394 | ||
395 | if (up) { | |
396 | if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat))) | |
397 | return; | |
398 | ||
399 | if (old < qd) { | |
d7067512 JB |
400 | old += scale; |
401 | old = min(old, qd); | |
402 | iolat->rq_depth.max_depth = old; | |
403 | wake_up_all(&iolat->rq_wait.wait); | |
404 | } | |
9f60511a | 405 | } else { |
d7067512 | 406 | old >>= 1; |
d7067512 JB |
407 | iolat->rq_depth.max_depth = max(old, 1UL); |
408 | } | |
409 | } | |
410 | ||
411 | /* Check our parent and see if the scale cookie has changed. */ | |
412 | static void check_scale_change(struct iolatency_grp *iolat) | |
413 | { | |
414 | struct iolatency_grp *parent; | |
415 | struct child_latency_info *lat_info; | |
416 | unsigned int cur_cookie; | |
417 | unsigned int our_cookie = atomic_read(&iolat->scale_cookie); | |
418 | u64 scale_lat; | |
419 | unsigned int old; | |
420 | int direction = 0; | |
421 | ||
422 | if (lat_to_blkg(iolat)->parent == NULL) | |
423 | return; | |
424 | ||
425 | parent = blkg_to_lat(lat_to_blkg(iolat)->parent); | |
426 | if (!parent) | |
427 | return; | |
428 | ||
429 | lat_info = &parent->child_lat; | |
430 | cur_cookie = atomic_read(&lat_info->scale_cookie); | |
431 | scale_lat = READ_ONCE(lat_info->scale_lat); | |
432 | ||
433 | if (cur_cookie < our_cookie) | |
434 | direction = -1; | |
435 | else if (cur_cookie > our_cookie) | |
436 | direction = 1; | |
437 | else | |
438 | return; | |
439 | ||
440 | old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie); | |
441 | ||
442 | /* Somebody beat us to the punch, just bail. */ | |
443 | if (old != our_cookie) | |
444 | return; | |
445 | ||
446 | if (direction < 0 && iolat->min_lat_nsec) { | |
447 | u64 samples_thresh; | |
448 | ||
449 | if (!scale_lat || iolat->min_lat_nsec <= scale_lat) | |
450 | return; | |
451 | ||
452 | /* | |
453 | * Sometimes high priority groups are their own worst enemy, so | |
454 | * instead of taking it out on some poor other group that did 5% | |
455 | * or less of the IO's for the last summation just skip this | |
456 | * scale down event. | |
457 | */ | |
458 | samples_thresh = lat_info->nr_samples * 5; | |
22ed8a93 | 459 | samples_thresh = max(1ULL, div64_u64(samples_thresh, 100)); |
d7067512 JB |
460 | if (iolat->nr_samples <= samples_thresh) |
461 | return; | |
462 | } | |
463 | ||
464 | /* We're as low as we can go. */ | |
465 | if (iolat->rq_depth.max_depth == 1 && direction < 0) { | |
466 | blkcg_use_delay(lat_to_blkg(iolat)); | |
467 | return; | |
468 | } | |
469 | ||
470 | /* We're back to the default cookie, unthrottle all the things. */ | |
471 | if (cur_cookie == DEFAULT_SCALE_COOKIE) { | |
472 | blkcg_clear_delay(lat_to_blkg(iolat)); | |
a284390b | 473 | iolat->rq_depth.max_depth = UINT_MAX; |
d7067512 JB |
474 | wake_up_all(&iolat->rq_wait.wait); |
475 | return; | |
476 | } | |
477 | ||
478 | scale_change(iolat, direction > 0); | |
479 | } | |
480 | ||
481 | static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio, | |
482 | spinlock_t *lock) | |
483 | { | |
484 | struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos); | |
b5f2954d DZ |
485 | struct blkcg *blkcg; |
486 | struct blkcg_gq *blkg; | |
487 | struct request_queue *q = rqos->q; | |
d7067512 JB |
488 | bool issue_as_root = bio_issue_as_root_blkg(bio); |
489 | ||
490 | if (!blk_iolatency_enabled(blkiolat)) | |
491 | return; | |
492 | ||
b5f2954d DZ |
493 | rcu_read_lock(); |
494 | blkcg = bio_blkcg(bio); | |
495 | bio_associate_blkcg(bio, &blkcg->css); | |
496 | blkg = blkg_lookup(blkcg, q); | |
497 | if (unlikely(!blkg)) { | |
498 | if (!lock) | |
499 | spin_lock_irq(q->queue_lock); | |
500 | blkg = blkg_lookup_create(blkcg, q); | |
501 | if (IS_ERR(blkg)) | |
502 | blkg = NULL; | |
503 | if (!lock) | |
504 | spin_unlock_irq(q->queue_lock); | |
505 | } | |
506 | if (!blkg) | |
507 | goto out; | |
508 | ||
509 | bio_issue_init(&bio->bi_issue, bio_sectors(bio)); | |
510 | bio_associate_blkg(bio, blkg); | |
511 | out: | |
512 | rcu_read_unlock(); | |
d7067512 JB |
513 | while (blkg && blkg->parent) { |
514 | struct iolatency_grp *iolat = blkg_to_lat(blkg); | |
515 | if (!iolat) { | |
516 | blkg = blkg->parent; | |
517 | continue; | |
518 | } | |
519 | ||
520 | check_scale_change(iolat); | |
521 | __blkcg_iolatency_throttle(rqos, iolat, lock, issue_as_root, | |
522 | (bio->bi_opf & REQ_SWAP) == REQ_SWAP); | |
523 | blkg = blkg->parent; | |
524 | } | |
525 | if (!timer_pending(&blkiolat->timer)) | |
526 | mod_timer(&blkiolat->timer, jiffies + HZ); | |
527 | } | |
528 | ||
529 | static void iolatency_record_time(struct iolatency_grp *iolat, | |
530 | struct bio_issue *issue, u64 now, | |
531 | bool issue_as_root) | |
532 | { | |
d7067512 JB |
533 | u64 start = bio_issue_time(issue); |
534 | u64 req_time; | |
535 | ||
71e9690b JB |
536 | /* |
537 | * Have to do this so we are truncated to the correct time that our | |
538 | * issue is truncated to. | |
539 | */ | |
540 | now = __bio_issue_time(now); | |
541 | ||
d7067512 JB |
542 | if (now <= start) |
543 | return; | |
544 | ||
545 | req_time = now - start; | |
546 | ||
547 | /* | |
548 | * We don't want to count issue_as_root bio's in the cgroups latency | |
549 | * statistics as it could skew the numbers downwards. | |
550 | */ | |
a284390b | 551 | if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) { |
d7067512 JB |
552 | u64 sub = iolat->min_lat_nsec; |
553 | if (req_time < sub) | |
554 | blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time); | |
555 | return; | |
556 | } | |
557 | ||
1fa2840e | 558 | latency_stat_record_time(iolat, req_time); |
d7067512 JB |
559 | } |
560 | ||
561 | #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC) | |
562 | #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5 | |
563 | ||
564 | static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now) | |
565 | { | |
566 | struct blkcg_gq *blkg = lat_to_blkg(iolat); | |
567 | struct iolatency_grp *parent; | |
568 | struct child_latency_info *lat_info; | |
1fa2840e | 569 | struct latency_stat stat; |
d7067512 | 570 | unsigned long flags; |
1fa2840e | 571 | int cpu; |
d7067512 | 572 | |
1fa2840e | 573 | latency_stat_init(iolat, &stat); |
d7067512 JB |
574 | preempt_disable(); |
575 | for_each_online_cpu(cpu) { | |
1fa2840e | 576 | struct latency_stat *s; |
d7067512 | 577 | s = per_cpu_ptr(iolat->stats, cpu); |
1fa2840e JB |
578 | latency_stat_sum(iolat, &stat, s); |
579 | latency_stat_init(iolat, s); | |
d7067512 JB |
580 | } |
581 | preempt_enable(); | |
582 | ||
d7067512 JB |
583 | parent = blkg_to_lat(blkg->parent); |
584 | if (!parent) | |
585 | return; | |
586 | ||
587 | lat_info = &parent->child_lat; | |
588 | ||
1fa2840e | 589 | iolat_update_total_lat_avg(iolat, &stat); |
d7067512 JB |
590 | |
591 | /* Everything is ok and we don't need to adjust the scale. */ | |
1fa2840e | 592 | if (latency_sum_ok(iolat, &stat) && |
d7067512 JB |
593 | atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE) |
594 | return; | |
595 | ||
596 | /* Somebody beat us to the punch, just bail. */ | |
597 | spin_lock_irqsave(&lat_info->lock, flags); | |
451bb7c3 JB |
598 | |
599 | latency_stat_sum(iolat, &iolat->cur_stat, &stat); | |
d7067512 | 600 | lat_info->nr_samples -= iolat->nr_samples; |
451bb7c3 JB |
601 | lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat); |
602 | iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat); | |
d7067512 JB |
603 | |
604 | if ((lat_info->last_scale_event >= now || | |
451bb7c3 | 605 | now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME)) |
d7067512 JB |
606 | goto out; |
607 | ||
451bb7c3 JB |
608 | if (latency_sum_ok(iolat, &iolat->cur_stat) && |
609 | latency_sum_ok(iolat, &stat)) { | |
610 | if (latency_stat_samples(iolat, &iolat->cur_stat) < | |
1fa2840e JB |
611 | BLKIOLATENCY_MIN_GOOD_SAMPLES) |
612 | goto out; | |
d7067512 JB |
613 | if (lat_info->scale_grp == iolat) { |
614 | lat_info->last_scale_event = now; | |
615 | scale_cookie_change(iolat->blkiolat, lat_info, true); | |
616 | } | |
451bb7c3 JB |
617 | } else if (lat_info->scale_lat == 0 || |
618 | lat_info->scale_lat >= iolat->min_lat_nsec) { | |
d7067512 JB |
619 | lat_info->last_scale_event = now; |
620 | if (!lat_info->scale_grp || | |
621 | lat_info->scale_lat > iolat->min_lat_nsec) { | |
622 | WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec); | |
623 | lat_info->scale_grp = iolat; | |
624 | } | |
625 | scale_cookie_change(iolat->blkiolat, lat_info, false); | |
626 | } | |
451bb7c3 | 627 | latency_stat_init(iolat, &iolat->cur_stat); |
d7067512 JB |
628 | out: |
629 | spin_unlock_irqrestore(&lat_info->lock, flags); | |
630 | } | |
631 | ||
632 | static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio) | |
633 | { | |
634 | struct blkcg_gq *blkg; | |
635 | struct rq_wait *rqw; | |
636 | struct iolatency_grp *iolat; | |
637 | u64 window_start; | |
638 | u64 now = ktime_to_ns(ktime_get()); | |
639 | bool issue_as_root = bio_issue_as_root_blkg(bio); | |
640 | bool enabled = false; | |
641 | ||
642 | blkg = bio->bi_blkg; | |
643 | if (!blkg) | |
644 | return; | |
645 | ||
646 | iolat = blkg_to_lat(bio->bi_blkg); | |
647 | if (!iolat) | |
648 | return; | |
649 | ||
650 | enabled = blk_iolatency_enabled(iolat->blkiolat); | |
651 | while (blkg && blkg->parent) { | |
652 | iolat = blkg_to_lat(blkg); | |
653 | if (!iolat) { | |
654 | blkg = blkg->parent; | |
655 | continue; | |
656 | } | |
657 | rqw = &iolat->rq_wait; | |
658 | ||
659 | atomic_dec(&rqw->inflight); | |
660 | if (!enabled || iolat->min_lat_nsec == 0) | |
661 | goto next; | |
662 | iolatency_record_time(iolat, &bio->bi_issue, now, | |
663 | issue_as_root); | |
664 | window_start = atomic64_read(&iolat->window_start); | |
665 | if (now > window_start && | |
666 | (now - window_start) >= iolat->cur_win_nsec) { | |
667 | if (atomic64_cmpxchg(&iolat->window_start, | |
668 | window_start, now) == window_start) | |
669 | iolatency_check_latencies(iolat, now); | |
670 | } | |
671 | next: | |
672 | wake_up(&rqw->wait); | |
673 | blkg = blkg->parent; | |
674 | } | |
675 | } | |
676 | ||
677 | static void blkcg_iolatency_cleanup(struct rq_qos *rqos, struct bio *bio) | |
678 | { | |
679 | struct blkcg_gq *blkg; | |
680 | ||
681 | blkg = bio->bi_blkg; | |
682 | while (blkg && blkg->parent) { | |
683 | struct rq_wait *rqw; | |
684 | struct iolatency_grp *iolat; | |
685 | ||
686 | iolat = blkg_to_lat(blkg); | |
687 | if (!iolat) | |
688 | goto next; | |
689 | ||
690 | rqw = &iolat->rq_wait; | |
691 | atomic_dec(&rqw->inflight); | |
692 | wake_up(&rqw->wait); | |
693 | next: | |
694 | blkg = blkg->parent; | |
695 | } | |
696 | } | |
697 | ||
698 | static void blkcg_iolatency_exit(struct rq_qos *rqos) | |
699 | { | |
700 | struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos); | |
701 | ||
702 | del_timer_sync(&blkiolat->timer); | |
703 | blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency); | |
704 | kfree(blkiolat); | |
705 | } | |
706 | ||
707 | static struct rq_qos_ops blkcg_iolatency_ops = { | |
708 | .throttle = blkcg_iolatency_throttle, | |
709 | .cleanup = blkcg_iolatency_cleanup, | |
710 | .done_bio = blkcg_iolatency_done_bio, | |
711 | .exit = blkcg_iolatency_exit, | |
712 | }; | |
713 | ||
714 | static void blkiolatency_timer_fn(struct timer_list *t) | |
715 | { | |
716 | struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer); | |
717 | struct blkcg_gq *blkg; | |
718 | struct cgroup_subsys_state *pos_css; | |
719 | u64 now = ktime_to_ns(ktime_get()); | |
720 | ||
721 | rcu_read_lock(); | |
722 | blkg_for_each_descendant_pre(blkg, pos_css, | |
723 | blkiolat->rqos.q->root_blkg) { | |
724 | struct iolatency_grp *iolat; | |
725 | struct child_latency_info *lat_info; | |
726 | unsigned long flags; | |
727 | u64 cookie; | |
728 | ||
729 | /* | |
730 | * We could be exiting, don't access the pd unless we have a | |
731 | * ref on the blkg. | |
732 | */ | |
b5f2954d | 733 | if (!blkg_try_get(blkg)) |
d7067512 JB |
734 | continue; |
735 | ||
736 | iolat = blkg_to_lat(blkg); | |
737 | if (!iolat) | |
52a1199c | 738 | goto next; |
d7067512 JB |
739 | |
740 | lat_info = &iolat->child_lat; | |
741 | cookie = atomic_read(&lat_info->scale_cookie); | |
742 | ||
743 | if (cookie >= DEFAULT_SCALE_COOKIE) | |
744 | goto next; | |
745 | ||
746 | spin_lock_irqsave(&lat_info->lock, flags); | |
747 | if (lat_info->last_scale_event >= now) | |
748 | goto next_lock; | |
749 | ||
750 | /* | |
751 | * We scaled down but don't have a scale_grp, scale up and carry | |
752 | * on. | |
753 | */ | |
754 | if (lat_info->scale_grp == NULL) { | |
755 | scale_cookie_change(iolat->blkiolat, lat_info, true); | |
756 | goto next_lock; | |
757 | } | |
758 | ||
759 | /* | |
760 | * It's been 5 seconds since our last scale event, clear the | |
761 | * scale grp in case the group that needed the scale down isn't | |
762 | * doing any IO currently. | |
763 | */ | |
764 | if (now - lat_info->last_scale_event >= | |
765 | ((u64)NSEC_PER_SEC * 5)) | |
766 | lat_info->scale_grp = NULL; | |
767 | next_lock: | |
768 | spin_unlock_irqrestore(&lat_info->lock, flags); | |
769 | next: | |
770 | blkg_put(blkg); | |
771 | } | |
772 | rcu_read_unlock(); | |
773 | } | |
774 | ||
775 | int blk_iolatency_init(struct request_queue *q) | |
776 | { | |
777 | struct blk_iolatency *blkiolat; | |
778 | struct rq_qos *rqos; | |
779 | int ret; | |
780 | ||
781 | blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL); | |
782 | if (!blkiolat) | |
783 | return -ENOMEM; | |
784 | ||
785 | rqos = &blkiolat->rqos; | |
786 | rqos->id = RQ_QOS_CGROUP; | |
787 | rqos->ops = &blkcg_iolatency_ops; | |
788 | rqos->q = q; | |
789 | ||
790 | rq_qos_add(q, rqos); | |
791 | ||
792 | ret = blkcg_activate_policy(q, &blkcg_policy_iolatency); | |
793 | if (ret) { | |
794 | rq_qos_del(q, rqos); | |
795 | kfree(blkiolat); | |
796 | return ret; | |
797 | } | |
798 | ||
799 | timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0); | |
800 | ||
801 | return 0; | |
802 | } | |
803 | ||
804 | static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val) | |
805 | { | |
806 | struct iolatency_grp *iolat = blkg_to_lat(blkg); | |
807 | struct blk_iolatency *blkiolat = iolat->blkiolat; | |
808 | u64 oldval = iolat->min_lat_nsec; | |
809 | ||
810 | iolat->min_lat_nsec = val; | |
c480bcf9 DZF |
811 | iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE); |
812 | iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec, | |
813 | BLKIOLATENCY_MAX_WIN_SIZE); | |
d7067512 JB |
814 | |
815 | if (!oldval && val) | |
816 | atomic_inc(&blkiolat->enabled); | |
817 | if (oldval && !val) | |
818 | atomic_dec(&blkiolat->enabled); | |
819 | } | |
820 | ||
821 | static void iolatency_clear_scaling(struct blkcg_gq *blkg) | |
822 | { | |
823 | if (blkg->parent) { | |
824 | struct iolatency_grp *iolat = blkg_to_lat(blkg->parent); | |
825 | struct child_latency_info *lat_info; | |
826 | if (!iolat) | |
827 | return; | |
828 | ||
829 | lat_info = &iolat->child_lat; | |
830 | spin_lock(&lat_info->lock); | |
831 | atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE); | |
832 | lat_info->last_scale_event = 0; | |
833 | lat_info->scale_grp = NULL; | |
834 | lat_info->scale_lat = 0; | |
835 | spin_unlock(&lat_info->lock); | |
836 | } | |
837 | } | |
838 | ||
839 | static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf, | |
840 | size_t nbytes, loff_t off) | |
841 | { | |
842 | struct blkcg *blkcg = css_to_blkcg(of_css(of)); | |
843 | struct blkcg_gq *blkg; | |
d7067512 JB |
844 | struct blkg_conf_ctx ctx; |
845 | struct iolatency_grp *iolat; | |
846 | char *p, *tok; | |
847 | u64 lat_val = 0; | |
848 | u64 oldval; | |
849 | int ret; | |
850 | ||
851 | ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx); | |
852 | if (ret) | |
853 | return ret; | |
854 | ||
855 | iolat = blkg_to_lat(ctx.blkg); | |
d7067512 JB |
856 | p = ctx.body; |
857 | ||
858 | ret = -EINVAL; | |
859 | while ((tok = strsep(&p, " "))) { | |
860 | char key[16]; | |
861 | char val[21]; /* 18446744073709551616 */ | |
862 | ||
863 | if (sscanf(tok, "%15[^=]=%20s", key, val) != 2) | |
864 | goto out; | |
865 | ||
866 | if (!strcmp(key, "target")) { | |
867 | u64 v; | |
868 | ||
869 | if (!strcmp(val, "max")) | |
870 | lat_val = 0; | |
871 | else if (sscanf(val, "%llu", &v) == 1) | |
872 | lat_val = v * NSEC_PER_USEC; | |
873 | else | |
874 | goto out; | |
875 | } else { | |
876 | goto out; | |
877 | } | |
878 | } | |
879 | ||
880 | /* Walk up the tree to see if our new val is lower than it should be. */ | |
881 | blkg = ctx.blkg; | |
882 | oldval = iolat->min_lat_nsec; | |
883 | ||
884 | iolatency_set_min_lat_nsec(blkg, lat_val); | |
885 | if (oldval != iolat->min_lat_nsec) { | |
886 | iolatency_clear_scaling(blkg); | |
887 | } | |
888 | ||
889 | ret = 0; | |
890 | out: | |
891 | blkg_conf_finish(&ctx); | |
892 | return ret ?: nbytes; | |
893 | } | |
894 | ||
895 | static u64 iolatency_prfill_limit(struct seq_file *sf, | |
896 | struct blkg_policy_data *pd, int off) | |
897 | { | |
898 | struct iolatency_grp *iolat = pd_to_lat(pd); | |
899 | const char *dname = blkg_dev_name(pd->blkg); | |
900 | ||
901 | if (!dname || !iolat->min_lat_nsec) | |
902 | return 0; | |
903 | seq_printf(sf, "%s target=%llu\n", | |
88b7210c | 904 | dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC)); |
d7067512 JB |
905 | return 0; |
906 | } | |
907 | ||
908 | static int iolatency_print_limit(struct seq_file *sf, void *v) | |
909 | { | |
910 | blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)), | |
911 | iolatency_prfill_limit, | |
912 | &blkcg_policy_iolatency, seq_cft(sf)->private, false); | |
913 | return 0; | |
914 | } | |
915 | ||
1fa2840e JB |
916 | static size_t iolatency_ssd_stat(struct iolatency_grp *iolat, char *buf, |
917 | size_t size) | |
918 | { | |
919 | struct latency_stat stat; | |
920 | int cpu; | |
921 | ||
922 | latency_stat_init(iolat, &stat); | |
923 | preempt_disable(); | |
924 | for_each_online_cpu(cpu) { | |
925 | struct latency_stat *s; | |
926 | s = per_cpu_ptr(iolat->stats, cpu); | |
927 | latency_stat_sum(iolat, &stat, s); | |
928 | } | |
929 | preempt_enable(); | |
930 | ||
931 | if (iolat->rq_depth.max_depth == UINT_MAX) | |
932 | return scnprintf(buf, size, " missed=%llu total=%llu depth=max", | |
933 | (unsigned long long)stat.ps.missed, | |
934 | (unsigned long long)stat.ps.total); | |
935 | return scnprintf(buf, size, " missed=%llu total=%llu depth=%u", | |
936 | (unsigned long long)stat.ps.missed, | |
937 | (unsigned long long)stat.ps.total, | |
938 | iolat->rq_depth.max_depth); | |
939 | } | |
940 | ||
d7067512 JB |
941 | static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf, |
942 | size_t size) | |
943 | { | |
944 | struct iolatency_grp *iolat = pd_to_lat(pd); | |
1fa2840e JB |
945 | unsigned long long avg_lat; |
946 | unsigned long long cur_win; | |
947 | ||
948 | if (iolat->ssd) | |
949 | return iolatency_ssd_stat(iolat, buf, size); | |
d7067512 | 950 | |
1fa2840e JB |
951 | avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC); |
952 | cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC); | |
a284390b | 953 | if (iolat->rq_depth.max_depth == UINT_MAX) |
c480bcf9 DZF |
954 | return scnprintf(buf, size, " depth=max avg_lat=%llu win=%llu", |
955 | avg_lat, cur_win); | |
d7067512 | 956 | |
c480bcf9 DZF |
957 | return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu", |
958 | iolat->rq_depth.max_depth, avg_lat, cur_win); | |
d7067512 JB |
959 | } |
960 | ||
961 | ||
962 | static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp, int node) | |
963 | { | |
964 | struct iolatency_grp *iolat; | |
965 | ||
966 | iolat = kzalloc_node(sizeof(*iolat), gfp, node); | |
967 | if (!iolat) | |
968 | return NULL; | |
1fa2840e JB |
969 | iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat), |
970 | __alignof__(struct latency_stat), gfp); | |
d7067512 JB |
971 | if (!iolat->stats) { |
972 | kfree(iolat); | |
973 | return NULL; | |
974 | } | |
975 | return &iolat->pd; | |
976 | } | |
977 | ||
978 | static void iolatency_pd_init(struct blkg_policy_data *pd) | |
979 | { | |
980 | struct iolatency_grp *iolat = pd_to_lat(pd); | |
981 | struct blkcg_gq *blkg = lat_to_blkg(iolat); | |
982 | struct rq_qos *rqos = blkcg_rq_qos(blkg->q); | |
983 | struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos); | |
984 | u64 now = ktime_to_ns(ktime_get()); | |
985 | int cpu; | |
986 | ||
1fa2840e JB |
987 | if (blk_queue_nonrot(blkg->q)) |
988 | iolat->ssd = true; | |
989 | else | |
990 | iolat->ssd = false; | |
991 | ||
d7067512 | 992 | for_each_possible_cpu(cpu) { |
1fa2840e | 993 | struct latency_stat *stat; |
d7067512 | 994 | stat = per_cpu_ptr(iolat->stats, cpu); |
1fa2840e | 995 | latency_stat_init(iolat, stat); |
d7067512 JB |
996 | } |
997 | ||
451bb7c3 | 998 | latency_stat_init(iolat, &iolat->cur_stat); |
d7067512 JB |
999 | rq_wait_init(&iolat->rq_wait); |
1000 | spin_lock_init(&iolat->child_lat.lock); | |
ff4cee08 | 1001 | iolat->rq_depth.queue_depth = blkg->q->nr_requests; |
a284390b | 1002 | iolat->rq_depth.max_depth = UINT_MAX; |
d7067512 JB |
1003 | iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth; |
1004 | iolat->blkiolat = blkiolat; | |
1005 | iolat->cur_win_nsec = 100 * NSEC_PER_MSEC; | |
1006 | atomic64_set(&iolat->window_start, now); | |
1007 | ||
1008 | /* | |
1009 | * We init things in list order, so the pd for the parent may not be | |
1010 | * init'ed yet for whatever reason. | |
1011 | */ | |
1012 | if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) { | |
1013 | struct iolatency_grp *parent = blkg_to_lat(blkg->parent); | |
1014 | atomic_set(&iolat->scale_cookie, | |
1015 | atomic_read(&parent->child_lat.scale_cookie)); | |
1016 | } else { | |
1017 | atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE); | |
1018 | } | |
1019 | ||
1020 | atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE); | |
1021 | } | |
1022 | ||
1023 | static void iolatency_pd_offline(struct blkg_policy_data *pd) | |
1024 | { | |
1025 | struct iolatency_grp *iolat = pd_to_lat(pd); | |
1026 | struct blkcg_gq *blkg = lat_to_blkg(iolat); | |
1027 | ||
1028 | iolatency_set_min_lat_nsec(blkg, 0); | |
1029 | iolatency_clear_scaling(blkg); | |
1030 | } | |
1031 | ||
1032 | static void iolatency_pd_free(struct blkg_policy_data *pd) | |
1033 | { | |
1034 | struct iolatency_grp *iolat = pd_to_lat(pd); | |
1035 | free_percpu(iolat->stats); | |
1036 | kfree(iolat); | |
1037 | } | |
1038 | ||
1039 | static struct cftype iolatency_files[] = { | |
1040 | { | |
1041 | .name = "latency", | |
1042 | .flags = CFTYPE_NOT_ON_ROOT, | |
1043 | .seq_show = iolatency_print_limit, | |
1044 | .write = iolatency_set_limit, | |
1045 | }, | |
1046 | {} | |
1047 | }; | |
1048 | ||
1049 | static struct blkcg_policy blkcg_policy_iolatency = { | |
1050 | .dfl_cftypes = iolatency_files, | |
1051 | .pd_alloc_fn = iolatency_pd_alloc, | |
1052 | .pd_init_fn = iolatency_pd_init, | |
1053 | .pd_offline_fn = iolatency_pd_offline, | |
1054 | .pd_free_fn = iolatency_pd_free, | |
1055 | .pd_stat_fn = iolatency_pd_stat, | |
1056 | }; | |
1057 | ||
1058 | static int __init iolatency_init(void) | |
1059 | { | |
1060 | return blkcg_policy_register(&blkcg_policy_iolatency); | |
1061 | } | |
1062 | ||
1063 | static void __exit iolatency_exit(void) | |
1064 | { | |
1065 | return blkcg_policy_unregister(&blkcg_policy_iolatency); | |
1066 | } | |
1067 | ||
1068 | module_init(iolatency_init); | |
1069 | module_exit(iolatency_exit); |