xref: /openbmc/linux/block/blk-iolatency.c (revision b60a5b8d)
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.
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
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>
72 #include <linux/sched/loadavg.h>
73 #include <linux/sched/signal.h>
74 #include <trace/events/block.h>
75 #include <linux/blk-mq.h>
76 #include "blk-rq-qos.h"
77 #include "blk-stat.h"
78 
79 #define DEFAULT_SCALE_COOKIE 1000000U
80 
81 static struct blkcg_policy blkcg_policy_iolatency;
82 struct iolatency_grp;
83 
84 struct blk_iolatency {
85 	struct rq_qos rqos;
86 	struct timer_list timer;
87 	atomic_t enabled;
88 };
89 
90 static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
91 {
92 	return container_of(rqos, struct blk_iolatency, rqos);
93 }
94 
95 static inline bool blk_iolatency_enabled(struct blk_iolatency *blkiolat)
96 {
97 	return atomic_read(&blkiolat->enabled) > 0;
98 }
99 
100 struct child_latency_info {
101 	spinlock_t lock;
102 
103 	/* Last time we adjusted the scale of everybody. */
104 	u64 last_scale_event;
105 
106 	/* The latency that we missed. */
107 	u64 scale_lat;
108 
109 	/* Total io's from all of our children for the last summation. */
110 	u64 nr_samples;
111 
112 	/* The guy who actually changed the latency numbers. */
113 	struct iolatency_grp *scale_grp;
114 
115 	/* Cookie to tell if we need to scale up or down. */
116 	atomic_t scale_cookie;
117 };
118 
119 struct percentile_stats {
120 	u64 total;
121 	u64 missed;
122 };
123 
124 struct latency_stat {
125 	union {
126 		struct percentile_stats ps;
127 		struct blk_rq_stat rqs;
128 	};
129 };
130 
131 struct iolatency_grp {
132 	struct blkg_policy_data pd;
133 	struct latency_stat __percpu *stats;
134 	struct latency_stat cur_stat;
135 	struct blk_iolatency *blkiolat;
136 	struct rq_depth rq_depth;
137 	struct rq_wait rq_wait;
138 	atomic64_t window_start;
139 	atomic_t scale_cookie;
140 	u64 min_lat_nsec;
141 	u64 cur_win_nsec;
142 
143 	/* total running average of our io latency. */
144 	u64 lat_avg;
145 
146 	/* Our current number of IO's for the last summation. */
147 	u64 nr_samples;
148 
149 	bool ssd;
150 	struct child_latency_info child_lat;
151 };
152 
153 #define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
154 #define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
155 /*
156  * These are the constants used to fake the fixed-point moving average
157  * calculation just like load average.  The call to calc_load() folds
158  * (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg.  The sampling
159  * window size is bucketed to try to approximately calculate average
160  * latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
161  * elapse immediately.  Note, windows only elapse with IO activity.  Idle
162  * periods extend the most recent window.
163  */
164 #define BLKIOLATENCY_NR_EXP_FACTORS 5
165 #define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
166 				      (BLKIOLATENCY_NR_EXP_FACTORS - 1))
167 static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
168 	2045, // exp(1/600) - 600 samples
169 	2039, // exp(1/240) - 240 samples
170 	2031, // exp(1/120) - 120 samples
171 	2023, // exp(1/80)  - 80 samples
172 	2014, // exp(1/60)  - 60 samples
173 };
174 
175 static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
176 {
177 	return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
178 }
179 
180 static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
181 {
182 	return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
183 }
184 
185 static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
186 {
187 	return pd_to_blkg(&iolat->pd);
188 }
189 
190 static inline void latency_stat_init(struct iolatency_grp *iolat,
191 				     struct latency_stat *stat)
192 {
193 	if (iolat->ssd) {
194 		stat->ps.total = 0;
195 		stat->ps.missed = 0;
196 	} else
197 		blk_rq_stat_init(&stat->rqs);
198 }
199 
200 static inline void latency_stat_sum(struct iolatency_grp *iolat,
201 				    struct latency_stat *sum,
202 				    struct latency_stat *stat)
203 {
204 	if (iolat->ssd) {
205 		sum->ps.total += stat->ps.total;
206 		sum->ps.missed += stat->ps.missed;
207 	} else
208 		blk_rq_stat_sum(&sum->rqs, &stat->rqs);
209 }
210 
211 static inline void latency_stat_record_time(struct iolatency_grp *iolat,
212 					    u64 req_time)
213 {
214 	struct latency_stat *stat = get_cpu_ptr(iolat->stats);
215 	if (iolat->ssd) {
216 		if (req_time >= iolat->min_lat_nsec)
217 			stat->ps.missed++;
218 		stat->ps.total++;
219 	} else
220 		blk_rq_stat_add(&stat->rqs, req_time);
221 	put_cpu_ptr(stat);
222 }
223 
224 static inline bool latency_sum_ok(struct iolatency_grp *iolat,
225 				  struct latency_stat *stat)
226 {
227 	if (iolat->ssd) {
228 		u64 thresh = div64_u64(stat->ps.total, 10);
229 		thresh = max(thresh, 1ULL);
230 		return stat->ps.missed < thresh;
231 	}
232 	return stat->rqs.mean <= iolat->min_lat_nsec;
233 }
234 
235 static inline u64 latency_stat_samples(struct iolatency_grp *iolat,
236 				       struct latency_stat *stat)
237 {
238 	if (iolat->ssd)
239 		return stat->ps.total;
240 	return stat->rqs.nr_samples;
241 }
242 
243 static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat,
244 					      struct latency_stat *stat)
245 {
246 	int exp_idx;
247 
248 	if (iolat->ssd)
249 		return;
250 
251 	/*
252 	 * calc_load() takes in a number stored in fixed point representation.
253 	 * Because we are using this for IO time in ns, the values stored
254 	 * are significantly larger than the FIXED_1 denominator (2048).
255 	 * Therefore, rounding errors in the calculation are negligible and
256 	 * can be ignored.
257 	 */
258 	exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
259 			div64_u64(iolat->cur_win_nsec,
260 				  BLKIOLATENCY_EXP_BUCKET_SIZE));
261 	iolat->lat_avg = calc_load(iolat->lat_avg,
262 				   iolatency_exp_factors[exp_idx],
263 				   stat->rqs.mean);
264 }
265 
266 static void iolat_cleanup_cb(struct rq_wait *rqw, void *private_data)
267 {
268 	atomic_dec(&rqw->inflight);
269 	wake_up(&rqw->wait);
270 }
271 
272 static bool iolat_acquire_inflight(struct rq_wait *rqw, void *private_data)
273 {
274 	struct iolatency_grp *iolat = private_data;
275 	return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
276 }
277 
278 static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
279 				       struct iolatency_grp *iolat,
280 				       bool issue_as_root,
281 				       bool use_memdelay)
282 {
283 	struct rq_wait *rqw = &iolat->rq_wait;
284 	unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
285 
286 	if (use_delay)
287 		blkcg_schedule_throttle(rqos->q, use_memdelay);
288 
289 	/*
290 	 * To avoid priority inversions we want to just take a slot if we are
291 	 * issuing as root.  If we're being killed off there's no point in
292 	 * delaying things, we may have been killed by OOM so throttling may
293 	 * make recovery take even longer, so just let the IO's through so the
294 	 * task can go away.
295 	 */
296 	if (issue_as_root || fatal_signal_pending(current)) {
297 		atomic_inc(&rqw->inflight);
298 		return;
299 	}
300 
301 	rq_qos_wait(rqw, iolat, iolat_acquire_inflight, iolat_cleanup_cb);
302 }
303 
304 #define SCALE_DOWN_FACTOR 2
305 #define SCALE_UP_FACTOR 4
306 
307 static inline unsigned long scale_amount(unsigned long qd, bool up)
308 {
309 	return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
310 }
311 
312 /*
313  * We scale the qd down faster than we scale up, so we need to use this helper
314  * to adjust the scale_cookie accordingly so we don't prematurely get
315  * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
316  *
317  * Each group has their own local copy of the last scale cookie they saw, so if
318  * the global scale cookie goes up or down they know which way they need to go
319  * based on their last knowledge of it.
320  */
321 static void scale_cookie_change(struct blk_iolatency *blkiolat,
322 				struct child_latency_info *lat_info,
323 				bool up)
324 {
325 	unsigned long qd = blkiolat->rqos.q->nr_requests;
326 	unsigned long scale = scale_amount(qd, up);
327 	unsigned long old = atomic_read(&lat_info->scale_cookie);
328 	unsigned long max_scale = qd << 1;
329 	unsigned long diff = 0;
330 
331 	if (old < DEFAULT_SCALE_COOKIE)
332 		diff = DEFAULT_SCALE_COOKIE - old;
333 
334 	if (up) {
335 		if (scale + old > DEFAULT_SCALE_COOKIE)
336 			atomic_set(&lat_info->scale_cookie,
337 				   DEFAULT_SCALE_COOKIE);
338 		else if (diff > qd)
339 			atomic_inc(&lat_info->scale_cookie);
340 		else
341 			atomic_add(scale, &lat_info->scale_cookie);
342 	} else {
343 		/*
344 		 * We don't want to dig a hole so deep that it takes us hours to
345 		 * dig out of it.  Just enough that we don't throttle/unthrottle
346 		 * with jagged workloads but can still unthrottle once pressure
347 		 * has sufficiently dissipated.
348 		 */
349 		if (diff > qd) {
350 			if (diff < max_scale)
351 				atomic_dec(&lat_info->scale_cookie);
352 		} else {
353 			atomic_sub(scale, &lat_info->scale_cookie);
354 		}
355 	}
356 }
357 
358 /*
359  * Change the queue depth of the iolatency_grp.  We add/subtract 1/16th of the
360  * queue depth at a time so we don't get wild swings and hopefully dial in to
361  * fairer distribution of the overall queue depth.
362  */
363 static void scale_change(struct iolatency_grp *iolat, bool up)
364 {
365 	unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
366 	unsigned long scale = scale_amount(qd, up);
367 	unsigned long old = iolat->rq_depth.max_depth;
368 
369 	if (old > qd)
370 		old = qd;
371 
372 	if (up) {
373 		if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
374 			return;
375 
376 		if (old < qd) {
377 			old += scale;
378 			old = min(old, qd);
379 			iolat->rq_depth.max_depth = old;
380 			wake_up_all(&iolat->rq_wait.wait);
381 		}
382 	} else {
383 		old >>= 1;
384 		iolat->rq_depth.max_depth = max(old, 1UL);
385 	}
386 }
387 
388 /* Check our parent and see if the scale cookie has changed. */
389 static void check_scale_change(struct iolatency_grp *iolat)
390 {
391 	struct iolatency_grp *parent;
392 	struct child_latency_info *lat_info;
393 	unsigned int cur_cookie;
394 	unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
395 	u64 scale_lat;
396 	unsigned int old;
397 	int direction = 0;
398 
399 	if (lat_to_blkg(iolat)->parent == NULL)
400 		return;
401 
402 	parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
403 	if (!parent)
404 		return;
405 
406 	lat_info = &parent->child_lat;
407 	cur_cookie = atomic_read(&lat_info->scale_cookie);
408 	scale_lat = READ_ONCE(lat_info->scale_lat);
409 
410 	if (cur_cookie < our_cookie)
411 		direction = -1;
412 	else if (cur_cookie > our_cookie)
413 		direction = 1;
414 	else
415 		return;
416 
417 	old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
418 
419 	/* Somebody beat us to the punch, just bail. */
420 	if (old != our_cookie)
421 		return;
422 
423 	if (direction < 0 && iolat->min_lat_nsec) {
424 		u64 samples_thresh;
425 
426 		if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
427 			return;
428 
429 		/*
430 		 * Sometimes high priority groups are their own worst enemy, so
431 		 * instead of taking it out on some poor other group that did 5%
432 		 * or less of the IO's for the last summation just skip this
433 		 * scale down event.
434 		 */
435 		samples_thresh = lat_info->nr_samples * 5;
436 		samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
437 		if (iolat->nr_samples <= samples_thresh)
438 			return;
439 	}
440 
441 	/* We're as low as we can go. */
442 	if (iolat->rq_depth.max_depth == 1 && direction < 0) {
443 		blkcg_use_delay(lat_to_blkg(iolat));
444 		return;
445 	}
446 
447 	/* We're back to the default cookie, unthrottle all the things. */
448 	if (cur_cookie == DEFAULT_SCALE_COOKIE) {
449 		blkcg_clear_delay(lat_to_blkg(iolat));
450 		iolat->rq_depth.max_depth = UINT_MAX;
451 		wake_up_all(&iolat->rq_wait.wait);
452 		return;
453 	}
454 
455 	scale_change(iolat, direction > 0);
456 }
457 
458 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
459 {
460 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
461 	struct blkcg_gq *blkg = bio->bi_blkg;
462 	bool issue_as_root = bio_issue_as_root_blkg(bio);
463 
464 	if (!blk_iolatency_enabled(blkiolat))
465 		return;
466 
467 	while (blkg && blkg->parent) {
468 		struct iolatency_grp *iolat = blkg_to_lat(blkg);
469 		if (!iolat) {
470 			blkg = blkg->parent;
471 			continue;
472 		}
473 
474 		check_scale_change(iolat);
475 		__blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
476 				     (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
477 		blkg = blkg->parent;
478 	}
479 	if (!timer_pending(&blkiolat->timer))
480 		mod_timer(&blkiolat->timer, jiffies + HZ);
481 }
482 
483 static void iolatency_record_time(struct iolatency_grp *iolat,
484 				  struct bio_issue *issue, u64 now,
485 				  bool issue_as_root)
486 {
487 	u64 start = bio_issue_time(issue);
488 	u64 req_time;
489 
490 	/*
491 	 * Have to do this so we are truncated to the correct time that our
492 	 * issue is truncated to.
493 	 */
494 	now = __bio_issue_time(now);
495 
496 	if (now <= start)
497 		return;
498 
499 	req_time = now - start;
500 
501 	/*
502 	 * We don't want to count issue_as_root bio's in the cgroups latency
503 	 * statistics as it could skew the numbers downwards.
504 	 */
505 	if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
506 		u64 sub = iolat->min_lat_nsec;
507 		if (req_time < sub)
508 			blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
509 		return;
510 	}
511 
512 	latency_stat_record_time(iolat, req_time);
513 }
514 
515 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
516 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
517 
518 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
519 {
520 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
521 	struct iolatency_grp *parent;
522 	struct child_latency_info *lat_info;
523 	struct latency_stat stat;
524 	unsigned long flags;
525 	int cpu;
526 
527 	latency_stat_init(iolat, &stat);
528 	preempt_disable();
529 	for_each_online_cpu(cpu) {
530 		struct latency_stat *s;
531 		s = per_cpu_ptr(iolat->stats, cpu);
532 		latency_stat_sum(iolat, &stat, s);
533 		latency_stat_init(iolat, s);
534 	}
535 	preempt_enable();
536 
537 	parent = blkg_to_lat(blkg->parent);
538 	if (!parent)
539 		return;
540 
541 	lat_info = &parent->child_lat;
542 
543 	iolat_update_total_lat_avg(iolat, &stat);
544 
545 	/* Everything is ok and we don't need to adjust the scale. */
546 	if (latency_sum_ok(iolat, &stat) &&
547 	    atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
548 		return;
549 
550 	/* Somebody beat us to the punch, just bail. */
551 	spin_lock_irqsave(&lat_info->lock, flags);
552 
553 	latency_stat_sum(iolat, &iolat->cur_stat, &stat);
554 	lat_info->nr_samples -= iolat->nr_samples;
555 	lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
556 	iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
557 
558 	if ((lat_info->last_scale_event >= now ||
559 	    now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
560 		goto out;
561 
562 	if (latency_sum_ok(iolat, &iolat->cur_stat) &&
563 	    latency_sum_ok(iolat, &stat)) {
564 		if (latency_stat_samples(iolat, &iolat->cur_stat) <
565 		    BLKIOLATENCY_MIN_GOOD_SAMPLES)
566 			goto out;
567 		if (lat_info->scale_grp == iolat) {
568 			lat_info->last_scale_event = now;
569 			scale_cookie_change(iolat->blkiolat, lat_info, true);
570 		}
571 	} else if (lat_info->scale_lat == 0 ||
572 		   lat_info->scale_lat >= iolat->min_lat_nsec) {
573 		lat_info->last_scale_event = now;
574 		if (!lat_info->scale_grp ||
575 		    lat_info->scale_lat > iolat->min_lat_nsec) {
576 			WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
577 			lat_info->scale_grp = iolat;
578 		}
579 		scale_cookie_change(iolat->blkiolat, lat_info, false);
580 	}
581 	latency_stat_init(iolat, &iolat->cur_stat);
582 out:
583 	spin_unlock_irqrestore(&lat_info->lock, flags);
584 }
585 
586 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
587 {
588 	struct blkcg_gq *blkg;
589 	struct rq_wait *rqw;
590 	struct iolatency_grp *iolat;
591 	u64 window_start;
592 	u64 now = ktime_to_ns(ktime_get());
593 	bool issue_as_root = bio_issue_as_root_blkg(bio);
594 	bool enabled = false;
595 	int inflight = 0;
596 
597 	blkg = bio->bi_blkg;
598 	if (!blkg || !bio_flagged(bio, BIO_TRACKED))
599 		return;
600 
601 	iolat = blkg_to_lat(bio->bi_blkg);
602 	if (!iolat)
603 		return;
604 
605 	enabled = blk_iolatency_enabled(iolat->blkiolat);
606 	if (!enabled)
607 		return;
608 
609 	while (blkg && blkg->parent) {
610 		iolat = blkg_to_lat(blkg);
611 		if (!iolat) {
612 			blkg = blkg->parent;
613 			continue;
614 		}
615 		rqw = &iolat->rq_wait;
616 
617 		inflight = atomic_dec_return(&rqw->inflight);
618 		WARN_ON_ONCE(inflight < 0);
619 		if (iolat->min_lat_nsec == 0)
620 			goto next;
621 		iolatency_record_time(iolat, &bio->bi_issue, now,
622 				      issue_as_root);
623 		window_start = atomic64_read(&iolat->window_start);
624 		if (now > window_start &&
625 		    (now - window_start) >= iolat->cur_win_nsec) {
626 			if (atomic64_cmpxchg(&iolat->window_start,
627 					window_start, now) == window_start)
628 				iolatency_check_latencies(iolat, now);
629 		}
630 next:
631 		wake_up(&rqw->wait);
632 		blkg = blkg->parent;
633 	}
634 }
635 
636 static void blkcg_iolatency_cleanup(struct rq_qos *rqos, struct bio *bio)
637 {
638 	struct blkcg_gq *blkg;
639 
640 	blkg = bio->bi_blkg;
641 	while (blkg && blkg->parent) {
642 		struct rq_wait *rqw;
643 		struct iolatency_grp *iolat;
644 
645 		iolat = blkg_to_lat(blkg);
646 		if (!iolat)
647 			goto next;
648 
649 		rqw = &iolat->rq_wait;
650 		atomic_dec(&rqw->inflight);
651 		wake_up(&rqw->wait);
652 next:
653 		blkg = blkg->parent;
654 	}
655 }
656 
657 static void blkcg_iolatency_exit(struct rq_qos *rqos)
658 {
659 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
660 
661 	del_timer_sync(&blkiolat->timer);
662 	blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
663 	kfree(blkiolat);
664 }
665 
666 static struct rq_qos_ops blkcg_iolatency_ops = {
667 	.throttle = blkcg_iolatency_throttle,
668 	.cleanup = blkcg_iolatency_cleanup,
669 	.done_bio = blkcg_iolatency_done_bio,
670 	.exit = blkcg_iolatency_exit,
671 };
672 
673 static void blkiolatency_timer_fn(struct timer_list *t)
674 {
675 	struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
676 	struct blkcg_gq *blkg;
677 	struct cgroup_subsys_state *pos_css;
678 	u64 now = ktime_to_ns(ktime_get());
679 
680 	rcu_read_lock();
681 	blkg_for_each_descendant_pre(blkg, pos_css,
682 				     blkiolat->rqos.q->root_blkg) {
683 		struct iolatency_grp *iolat;
684 		struct child_latency_info *lat_info;
685 		unsigned long flags;
686 		u64 cookie;
687 
688 		/*
689 		 * We could be exiting, don't access the pd unless we have a
690 		 * ref on the blkg.
691 		 */
692 		if (!blkg_tryget(blkg))
693 			continue;
694 
695 		iolat = blkg_to_lat(blkg);
696 		if (!iolat)
697 			goto next;
698 
699 		lat_info = &iolat->child_lat;
700 		cookie = atomic_read(&lat_info->scale_cookie);
701 
702 		if (cookie >= DEFAULT_SCALE_COOKIE)
703 			goto next;
704 
705 		spin_lock_irqsave(&lat_info->lock, flags);
706 		if (lat_info->last_scale_event >= now)
707 			goto next_lock;
708 
709 		/*
710 		 * We scaled down but don't have a scale_grp, scale up and carry
711 		 * on.
712 		 */
713 		if (lat_info->scale_grp == NULL) {
714 			scale_cookie_change(iolat->blkiolat, lat_info, true);
715 			goto next_lock;
716 		}
717 
718 		/*
719 		 * It's been 5 seconds since our last scale event, clear the
720 		 * scale grp in case the group that needed the scale down isn't
721 		 * doing any IO currently.
722 		 */
723 		if (now - lat_info->last_scale_event >=
724 		    ((u64)NSEC_PER_SEC * 5))
725 			lat_info->scale_grp = NULL;
726 next_lock:
727 		spin_unlock_irqrestore(&lat_info->lock, flags);
728 next:
729 		blkg_put(blkg);
730 	}
731 	rcu_read_unlock();
732 }
733 
734 int blk_iolatency_init(struct request_queue *q)
735 {
736 	struct blk_iolatency *blkiolat;
737 	struct rq_qos *rqos;
738 	int ret;
739 
740 	blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
741 	if (!blkiolat)
742 		return -ENOMEM;
743 
744 	rqos = &blkiolat->rqos;
745 	rqos->id = RQ_QOS_CGROUP;
746 	rqos->ops = &blkcg_iolatency_ops;
747 	rqos->q = q;
748 
749 	rq_qos_add(q, rqos);
750 
751 	ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
752 	if (ret) {
753 		rq_qos_del(q, rqos);
754 		kfree(blkiolat);
755 		return ret;
756 	}
757 
758 	timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
759 
760 	return 0;
761 }
762 
763 /*
764  * return 1 for enabling iolatency, return -1 for disabling iolatency, otherwise
765  * return 0.
766  */
767 static int iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
768 {
769 	struct iolatency_grp *iolat = blkg_to_lat(blkg);
770 	u64 oldval = iolat->min_lat_nsec;
771 
772 	iolat->min_lat_nsec = val;
773 	iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
774 	iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
775 				    BLKIOLATENCY_MAX_WIN_SIZE);
776 
777 	if (!oldval && val)
778 		return 1;
779 	if (oldval && !val)
780 		return -1;
781 	return 0;
782 }
783 
784 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
785 {
786 	if (blkg->parent) {
787 		struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
788 		struct child_latency_info *lat_info;
789 		if (!iolat)
790 			return;
791 
792 		lat_info = &iolat->child_lat;
793 		spin_lock(&lat_info->lock);
794 		atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
795 		lat_info->last_scale_event = 0;
796 		lat_info->scale_grp = NULL;
797 		lat_info->scale_lat = 0;
798 		spin_unlock(&lat_info->lock);
799 	}
800 }
801 
802 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
803 			     size_t nbytes, loff_t off)
804 {
805 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
806 	struct blkcg_gq *blkg;
807 	struct blkg_conf_ctx ctx;
808 	struct iolatency_grp *iolat;
809 	char *p, *tok;
810 	u64 lat_val = 0;
811 	u64 oldval;
812 	int ret;
813 	int enable = 0;
814 
815 	ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
816 	if (ret)
817 		return ret;
818 
819 	iolat = blkg_to_lat(ctx.blkg);
820 	p = ctx.body;
821 
822 	ret = -EINVAL;
823 	while ((tok = strsep(&p, " "))) {
824 		char key[16];
825 		char val[21];	/* 18446744073709551616 */
826 
827 		if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
828 			goto out;
829 
830 		if (!strcmp(key, "target")) {
831 			u64 v;
832 
833 			if (!strcmp(val, "max"))
834 				lat_val = 0;
835 			else if (sscanf(val, "%llu", &v) == 1)
836 				lat_val = v * NSEC_PER_USEC;
837 			else
838 				goto out;
839 		} else {
840 			goto out;
841 		}
842 	}
843 
844 	/* Walk up the tree to see if our new val is lower than it should be. */
845 	blkg = ctx.blkg;
846 	oldval = iolat->min_lat_nsec;
847 
848 	enable = iolatency_set_min_lat_nsec(blkg, lat_val);
849 	if (enable) {
850 		WARN_ON_ONCE(!blk_get_queue(blkg->q));
851 		blkg_get(blkg);
852 	}
853 
854 	if (oldval != iolat->min_lat_nsec) {
855 		iolatency_clear_scaling(blkg);
856 	}
857 
858 	ret = 0;
859 out:
860 	blkg_conf_finish(&ctx);
861 	if (ret == 0 && enable) {
862 		struct iolatency_grp *tmp = blkg_to_lat(blkg);
863 		struct blk_iolatency *blkiolat = tmp->blkiolat;
864 
865 		blk_mq_freeze_queue(blkg->q);
866 
867 		if (enable == 1)
868 			atomic_inc(&blkiolat->enabled);
869 		else if (enable == -1)
870 			atomic_dec(&blkiolat->enabled);
871 		else
872 			WARN_ON_ONCE(1);
873 
874 		blk_mq_unfreeze_queue(blkg->q);
875 
876 		blkg_put(blkg);
877 		blk_put_queue(blkg->q);
878 	}
879 	return ret ?: nbytes;
880 }
881 
882 static u64 iolatency_prfill_limit(struct seq_file *sf,
883 				  struct blkg_policy_data *pd, int off)
884 {
885 	struct iolatency_grp *iolat = pd_to_lat(pd);
886 	const char *dname = blkg_dev_name(pd->blkg);
887 
888 	if (!dname || !iolat->min_lat_nsec)
889 		return 0;
890 	seq_printf(sf, "%s target=%llu\n",
891 		   dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
892 	return 0;
893 }
894 
895 static int iolatency_print_limit(struct seq_file *sf, void *v)
896 {
897 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
898 			  iolatency_prfill_limit,
899 			  &blkcg_policy_iolatency, seq_cft(sf)->private, false);
900 	return 0;
901 }
902 
903 static size_t iolatency_ssd_stat(struct iolatency_grp *iolat, char *buf,
904 				 size_t size)
905 {
906 	struct latency_stat stat;
907 	int cpu;
908 
909 	latency_stat_init(iolat, &stat);
910 	preempt_disable();
911 	for_each_online_cpu(cpu) {
912 		struct latency_stat *s;
913 		s = per_cpu_ptr(iolat->stats, cpu);
914 		latency_stat_sum(iolat, &stat, s);
915 	}
916 	preempt_enable();
917 
918 	if (iolat->rq_depth.max_depth == UINT_MAX)
919 		return scnprintf(buf, size, " missed=%llu total=%llu depth=max",
920 				 (unsigned long long)stat.ps.missed,
921 				 (unsigned long long)stat.ps.total);
922 	return scnprintf(buf, size, " missed=%llu total=%llu depth=%u",
923 			 (unsigned long long)stat.ps.missed,
924 			 (unsigned long long)stat.ps.total,
925 			 iolat->rq_depth.max_depth);
926 }
927 
928 static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf,
929 				size_t size)
930 {
931 	struct iolatency_grp *iolat = pd_to_lat(pd);
932 	unsigned long long avg_lat;
933 	unsigned long long cur_win;
934 
935 	if (iolat->ssd)
936 		return iolatency_ssd_stat(iolat, buf, size);
937 
938 	avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
939 	cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
940 	if (iolat->rq_depth.max_depth == UINT_MAX)
941 		return scnprintf(buf, size, " depth=max avg_lat=%llu win=%llu",
942 				 avg_lat, cur_win);
943 
944 	return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu",
945 			 iolat->rq_depth.max_depth, avg_lat, cur_win);
946 }
947 
948 
949 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp, int node)
950 {
951 	struct iolatency_grp *iolat;
952 
953 	iolat = kzalloc_node(sizeof(*iolat), gfp, node);
954 	if (!iolat)
955 		return NULL;
956 	iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
957 				       __alignof__(struct latency_stat), gfp);
958 	if (!iolat->stats) {
959 		kfree(iolat);
960 		return NULL;
961 	}
962 	return &iolat->pd;
963 }
964 
965 static void iolatency_pd_init(struct blkg_policy_data *pd)
966 {
967 	struct iolatency_grp *iolat = pd_to_lat(pd);
968 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
969 	struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
970 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
971 	u64 now = ktime_to_ns(ktime_get());
972 	int cpu;
973 
974 	if (blk_queue_nonrot(blkg->q))
975 		iolat->ssd = true;
976 	else
977 		iolat->ssd = false;
978 
979 	for_each_possible_cpu(cpu) {
980 		struct latency_stat *stat;
981 		stat = per_cpu_ptr(iolat->stats, cpu);
982 		latency_stat_init(iolat, stat);
983 	}
984 
985 	latency_stat_init(iolat, &iolat->cur_stat);
986 	rq_wait_init(&iolat->rq_wait);
987 	spin_lock_init(&iolat->child_lat.lock);
988 	iolat->rq_depth.queue_depth = blkg->q->nr_requests;
989 	iolat->rq_depth.max_depth = UINT_MAX;
990 	iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
991 	iolat->blkiolat = blkiolat;
992 	iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
993 	atomic64_set(&iolat->window_start, now);
994 
995 	/*
996 	 * We init things in list order, so the pd for the parent may not be
997 	 * init'ed yet for whatever reason.
998 	 */
999 	if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
1000 		struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
1001 		atomic_set(&iolat->scale_cookie,
1002 			   atomic_read(&parent->child_lat.scale_cookie));
1003 	} else {
1004 		atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
1005 	}
1006 
1007 	atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
1008 }
1009 
1010 static void iolatency_pd_offline(struct blkg_policy_data *pd)
1011 {
1012 	struct iolatency_grp *iolat = pd_to_lat(pd);
1013 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
1014 	struct blk_iolatency *blkiolat = iolat->blkiolat;
1015 	int ret;
1016 
1017 	ret = iolatency_set_min_lat_nsec(blkg, 0);
1018 	if (ret == 1)
1019 		atomic_inc(&blkiolat->enabled);
1020 	if (ret == -1)
1021 		atomic_dec(&blkiolat->enabled);
1022 	iolatency_clear_scaling(blkg);
1023 }
1024 
1025 static void iolatency_pd_free(struct blkg_policy_data *pd)
1026 {
1027 	struct iolatency_grp *iolat = pd_to_lat(pd);
1028 	free_percpu(iolat->stats);
1029 	kfree(iolat);
1030 }
1031 
1032 static struct cftype iolatency_files[] = {
1033 	{
1034 		.name = "latency",
1035 		.flags = CFTYPE_NOT_ON_ROOT,
1036 		.seq_show = iolatency_print_limit,
1037 		.write = iolatency_set_limit,
1038 	},
1039 	{}
1040 };
1041 
1042 static struct blkcg_policy blkcg_policy_iolatency = {
1043 	.dfl_cftypes	= iolatency_files,
1044 	.pd_alloc_fn	= iolatency_pd_alloc,
1045 	.pd_init_fn	= iolatency_pd_init,
1046 	.pd_offline_fn	= iolatency_pd_offline,
1047 	.pd_free_fn	= iolatency_pd_free,
1048 	.pd_stat_fn	= iolatency_pd_stat,
1049 };
1050 
1051 static int __init iolatency_init(void)
1052 {
1053 	return blkcg_policy_register(&blkcg_policy_iolatency);
1054 }
1055 
1056 static void __exit iolatency_exit(void)
1057 {
1058 	return blkcg_policy_unregister(&blkcg_policy_iolatency);
1059 }
1060 
1061 module_init(iolatency_init);
1062 module_exit(iolatency_exit);
1063