xref: /linux/block/blk-iolatency.c (revision b1d29ba82cf2bc784f4c963ddd6a2cf29e229b33)
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 "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 
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 
130 struct iolatency_grp {
131 	struct blkg_policy_data pd;
132 	struct latency_stat __percpu *stats;
133 	struct latency_stat cur_stat;
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. */
143 	u64 lat_avg;
144 
145 	/* Our current number of IO's for the last summation. */
146 	u64 nr_samples;
147 
148 	bool ssd;
149 	struct child_latency_info child_lat;
150 };
151 
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
156  * calculation just like load average.  The call to CALC_LOAD folds
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 
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 
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 	/*
251 	 * CALC_LOAD takes in a number stored in fixed point representation.
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));
260 	CALC_LOAD(iolat->lat_avg, iolatency_exp_factors[exp_idx], stat->rqs.mean);
261 }
262 
263 static inline bool iolatency_may_queue(struct iolatency_grp *iolat,
264 				       wait_queue_entry_t *wait,
265 				       bool first_block)
266 {
267 	struct rq_wait *rqw = &iolat->rq_wait;
268 
269 	if (first_block && waitqueue_active(&rqw->wait) &&
270 	    rqw->wait.head.next != &wait->entry)
271 		return false;
272 	return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
273 }
274 
275 static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
276 				       struct iolatency_grp *iolat,
277 				       spinlock_t *lock, bool issue_as_root,
278 				       bool use_memdelay)
279 	__releases(lock)
280 	__acquires(lock)
281 {
282 	struct rq_wait *rqw = &iolat->rq_wait;
283 	unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
284 	DEFINE_WAIT(wait);
285 	bool first_block = true;
286 
287 	if (use_delay)
288 		blkcg_schedule_throttle(rqos->q, use_memdelay);
289 
290 	/*
291 	 * To avoid priority inversions we want to just take a slot if we are
292 	 * issuing as root.  If we're being killed off there's no point in
293 	 * delaying things, we may have been killed by OOM so throttling may
294 	 * make recovery take even longer, so just let the IO's through so the
295 	 * task can go away.
296 	 */
297 	if (issue_as_root || fatal_signal_pending(current)) {
298 		atomic_inc(&rqw->inflight);
299 		return;
300 	}
301 
302 	if (iolatency_may_queue(iolat, &wait, first_block))
303 		return;
304 
305 	do {
306 		prepare_to_wait_exclusive(&rqw->wait, &wait,
307 					  TASK_UNINTERRUPTIBLE);
308 
309 		if (iolatency_may_queue(iolat, &wait, first_block))
310 			break;
311 		first_block = false;
312 
313 		if (lock) {
314 			spin_unlock_irq(lock);
315 			io_schedule();
316 			spin_lock_irq(lock);
317 		} else {
318 			io_schedule();
319 		}
320 	} while (1);
321 
322 	finish_wait(&rqw->wait, &wait);
323 }
324 
325 #define SCALE_DOWN_FACTOR 2
326 #define SCALE_UP_FACTOR 4
327 
328 static inline unsigned long scale_amount(unsigned long qd, bool up)
329 {
330 	return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
331 }
332 
333 /*
334  * We scale the qd down faster than we scale up, so we need to use this helper
335  * to adjust the scale_cookie accordingly so we don't prematurely get
336  * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
337  *
338  * Each group has their own local copy of the last scale cookie they saw, so if
339  * the global scale cookie goes up or down they know which way they need to go
340  * based on their last knowledge of it.
341  */
342 static void scale_cookie_change(struct blk_iolatency *blkiolat,
343 				struct child_latency_info *lat_info,
344 				bool up)
345 {
346 	unsigned long qd = blkiolat->rqos.q->nr_requests;
347 	unsigned long scale = scale_amount(qd, up);
348 	unsigned long old = atomic_read(&lat_info->scale_cookie);
349 	unsigned long max_scale = qd << 1;
350 	unsigned long diff = 0;
351 
352 	if (old < DEFAULT_SCALE_COOKIE)
353 		diff = DEFAULT_SCALE_COOKIE - old;
354 
355 	if (up) {
356 		if (scale + old > DEFAULT_SCALE_COOKIE)
357 			atomic_set(&lat_info->scale_cookie,
358 				   DEFAULT_SCALE_COOKIE);
359 		else if (diff > qd)
360 			atomic_inc(&lat_info->scale_cookie);
361 		else
362 			atomic_add(scale, &lat_info->scale_cookie);
363 	} else {
364 		/*
365 		 * We don't want to dig a hole so deep that it takes us hours to
366 		 * dig out of it.  Just enough that we don't throttle/unthrottle
367 		 * with jagged workloads but can still unthrottle once pressure
368 		 * has sufficiently dissipated.
369 		 */
370 		if (diff > qd) {
371 			if (diff < max_scale)
372 				atomic_dec(&lat_info->scale_cookie);
373 		} else {
374 			atomic_sub(scale, &lat_info->scale_cookie);
375 		}
376 	}
377 }
378 
379 /*
380  * Change the queue depth of the iolatency_grp.  We add/subtract 1/16th of the
381  * queue depth at a time so we don't get wild swings and hopefully dial in to
382  * fairer distribution of the overall queue depth.
383  */
384 static void scale_change(struct iolatency_grp *iolat, bool up)
385 {
386 	unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
387 	unsigned long scale = scale_amount(qd, up);
388 	unsigned long old = iolat->rq_depth.max_depth;
389 
390 	if (old > qd)
391 		old = qd;
392 
393 	if (up) {
394 		if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
395 			return;
396 
397 		if (old < qd) {
398 			old += scale;
399 			old = min(old, qd);
400 			iolat->rq_depth.max_depth = old;
401 			wake_up_all(&iolat->rq_wait.wait);
402 		}
403 	} else {
404 		old >>= 1;
405 		iolat->rq_depth.max_depth = max(old, 1UL);
406 	}
407 }
408 
409 /* Check our parent and see if the scale cookie has changed. */
410 static void check_scale_change(struct iolatency_grp *iolat)
411 {
412 	struct iolatency_grp *parent;
413 	struct child_latency_info *lat_info;
414 	unsigned int cur_cookie;
415 	unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
416 	u64 scale_lat;
417 	unsigned int old;
418 	int direction = 0;
419 
420 	if (lat_to_blkg(iolat)->parent == NULL)
421 		return;
422 
423 	parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
424 	if (!parent)
425 		return;
426 
427 	lat_info = &parent->child_lat;
428 	cur_cookie = atomic_read(&lat_info->scale_cookie);
429 	scale_lat = READ_ONCE(lat_info->scale_lat);
430 
431 	if (cur_cookie < our_cookie)
432 		direction = -1;
433 	else if (cur_cookie > our_cookie)
434 		direction = 1;
435 	else
436 		return;
437 
438 	old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
439 
440 	/* Somebody beat us to the punch, just bail. */
441 	if (old != our_cookie)
442 		return;
443 
444 	if (direction < 0 && iolat->min_lat_nsec) {
445 		u64 samples_thresh;
446 
447 		if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
448 			return;
449 
450 		/*
451 		 * Sometimes high priority groups are their own worst enemy, so
452 		 * instead of taking it out on some poor other group that did 5%
453 		 * or less of the IO's for the last summation just skip this
454 		 * scale down event.
455 		 */
456 		samples_thresh = lat_info->nr_samples * 5;
457 		samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
458 		if (iolat->nr_samples <= samples_thresh)
459 			return;
460 	}
461 
462 	/* We're as low as we can go. */
463 	if (iolat->rq_depth.max_depth == 1 && direction < 0) {
464 		blkcg_use_delay(lat_to_blkg(iolat));
465 		return;
466 	}
467 
468 	/* We're back to the default cookie, unthrottle all the things. */
469 	if (cur_cookie == DEFAULT_SCALE_COOKIE) {
470 		blkcg_clear_delay(lat_to_blkg(iolat));
471 		iolat->rq_depth.max_depth = UINT_MAX;
472 		wake_up_all(&iolat->rq_wait.wait);
473 		return;
474 	}
475 
476 	scale_change(iolat, direction > 0);
477 }
478 
479 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio,
480 				     spinlock_t *lock)
481 {
482 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
483 	struct blkcg_gq *blkg = bio->bi_blkg;
484 	bool issue_as_root = bio_issue_as_root_blkg(bio);
485 
486 	if (!blk_iolatency_enabled(blkiolat))
487 		return;
488 
489 	while (blkg && blkg->parent) {
490 		struct iolatency_grp *iolat = blkg_to_lat(blkg);
491 		if (!iolat) {
492 			blkg = blkg->parent;
493 			continue;
494 		}
495 
496 		check_scale_change(iolat);
497 		__blkcg_iolatency_throttle(rqos, iolat, lock, issue_as_root,
498 				     (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
499 		blkg = blkg->parent;
500 	}
501 	if (!timer_pending(&blkiolat->timer))
502 		mod_timer(&blkiolat->timer, jiffies + HZ);
503 }
504 
505 static void iolatency_record_time(struct iolatency_grp *iolat,
506 				  struct bio_issue *issue, u64 now,
507 				  bool issue_as_root)
508 {
509 	u64 start = bio_issue_time(issue);
510 	u64 req_time;
511 
512 	/*
513 	 * Have to do this so we are truncated to the correct time that our
514 	 * issue is truncated to.
515 	 */
516 	now = __bio_issue_time(now);
517 
518 	if (now <= start)
519 		return;
520 
521 	req_time = now - start;
522 
523 	/*
524 	 * We don't want to count issue_as_root bio's in the cgroups latency
525 	 * statistics as it could skew the numbers downwards.
526 	 */
527 	if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
528 		u64 sub = iolat->min_lat_nsec;
529 		if (req_time < sub)
530 			blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
531 		return;
532 	}
533 
534 	latency_stat_record_time(iolat, req_time);
535 }
536 
537 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
538 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
539 
540 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
541 {
542 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
543 	struct iolatency_grp *parent;
544 	struct child_latency_info *lat_info;
545 	struct latency_stat stat;
546 	unsigned long flags;
547 	int cpu;
548 
549 	latency_stat_init(iolat, &stat);
550 	preempt_disable();
551 	for_each_online_cpu(cpu) {
552 		struct latency_stat *s;
553 		s = per_cpu_ptr(iolat->stats, cpu);
554 		latency_stat_sum(iolat, &stat, s);
555 		latency_stat_init(iolat, s);
556 	}
557 	preempt_enable();
558 
559 	parent = blkg_to_lat(blkg->parent);
560 	if (!parent)
561 		return;
562 
563 	lat_info = &parent->child_lat;
564 
565 	iolat_update_total_lat_avg(iolat, &stat);
566 
567 	/* Everything is ok and we don't need to adjust the scale. */
568 	if (latency_sum_ok(iolat, &stat) &&
569 	    atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
570 		return;
571 
572 	/* Somebody beat us to the punch, just bail. */
573 	spin_lock_irqsave(&lat_info->lock, flags);
574 
575 	latency_stat_sum(iolat, &iolat->cur_stat, &stat);
576 	lat_info->nr_samples -= iolat->nr_samples;
577 	lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
578 	iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
579 
580 	if ((lat_info->last_scale_event >= now ||
581 	    now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
582 		goto out;
583 
584 	if (latency_sum_ok(iolat, &iolat->cur_stat) &&
585 	    latency_sum_ok(iolat, &stat)) {
586 		if (latency_stat_samples(iolat, &iolat->cur_stat) <
587 		    BLKIOLATENCY_MIN_GOOD_SAMPLES)
588 			goto out;
589 		if (lat_info->scale_grp == iolat) {
590 			lat_info->last_scale_event = now;
591 			scale_cookie_change(iolat->blkiolat, lat_info, true);
592 		}
593 	} else if (lat_info->scale_lat == 0 ||
594 		   lat_info->scale_lat >= iolat->min_lat_nsec) {
595 		lat_info->last_scale_event = now;
596 		if (!lat_info->scale_grp ||
597 		    lat_info->scale_lat > iolat->min_lat_nsec) {
598 			WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
599 			lat_info->scale_grp = iolat;
600 		}
601 		scale_cookie_change(iolat->blkiolat, lat_info, false);
602 	}
603 	latency_stat_init(iolat, &iolat->cur_stat);
604 out:
605 	spin_unlock_irqrestore(&lat_info->lock, flags);
606 }
607 
608 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
609 {
610 	struct blkcg_gq *blkg;
611 	struct rq_wait *rqw;
612 	struct iolatency_grp *iolat;
613 	u64 window_start;
614 	u64 now = ktime_to_ns(ktime_get());
615 	bool issue_as_root = bio_issue_as_root_blkg(bio);
616 	bool enabled = false;
617 
618 	blkg = bio->bi_blkg;
619 	if (!blkg)
620 		return;
621 
622 	iolat = blkg_to_lat(bio->bi_blkg);
623 	if (!iolat)
624 		return;
625 
626 	enabled = blk_iolatency_enabled(iolat->blkiolat);
627 	while (blkg && blkg->parent) {
628 		iolat = blkg_to_lat(blkg);
629 		if (!iolat) {
630 			blkg = blkg->parent;
631 			continue;
632 		}
633 		rqw = &iolat->rq_wait;
634 
635 		atomic_dec(&rqw->inflight);
636 		if (!enabled || iolat->min_lat_nsec == 0)
637 			goto next;
638 		iolatency_record_time(iolat, &bio->bi_issue, now,
639 				      issue_as_root);
640 		window_start = atomic64_read(&iolat->window_start);
641 		if (now > window_start &&
642 		    (now - window_start) >= iolat->cur_win_nsec) {
643 			if (atomic64_cmpxchg(&iolat->window_start,
644 					window_start, now) == window_start)
645 				iolatency_check_latencies(iolat, now);
646 		}
647 next:
648 		wake_up(&rqw->wait);
649 		blkg = blkg->parent;
650 	}
651 }
652 
653 static void blkcg_iolatency_cleanup(struct rq_qos *rqos, struct bio *bio)
654 {
655 	struct blkcg_gq *blkg;
656 
657 	blkg = bio->bi_blkg;
658 	while (blkg && blkg->parent) {
659 		struct rq_wait *rqw;
660 		struct iolatency_grp *iolat;
661 
662 		iolat = blkg_to_lat(blkg);
663 		if (!iolat)
664 			goto next;
665 
666 		rqw = &iolat->rq_wait;
667 		atomic_dec(&rqw->inflight);
668 		wake_up(&rqw->wait);
669 next:
670 		blkg = blkg->parent;
671 	}
672 }
673 
674 static void blkcg_iolatency_exit(struct rq_qos *rqos)
675 {
676 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
677 
678 	del_timer_sync(&blkiolat->timer);
679 	blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
680 	kfree(blkiolat);
681 }
682 
683 static struct rq_qos_ops blkcg_iolatency_ops = {
684 	.throttle = blkcg_iolatency_throttle,
685 	.cleanup = blkcg_iolatency_cleanup,
686 	.done_bio = blkcg_iolatency_done_bio,
687 	.exit = blkcg_iolatency_exit,
688 };
689 
690 static void blkiolatency_timer_fn(struct timer_list *t)
691 {
692 	struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
693 	struct blkcg_gq *blkg;
694 	struct cgroup_subsys_state *pos_css;
695 	u64 now = ktime_to_ns(ktime_get());
696 
697 	rcu_read_lock();
698 	blkg_for_each_descendant_pre(blkg, pos_css,
699 				     blkiolat->rqos.q->root_blkg) {
700 		struct iolatency_grp *iolat;
701 		struct child_latency_info *lat_info;
702 		unsigned long flags;
703 		u64 cookie;
704 
705 		/*
706 		 * We could be exiting, don't access the pd unless we have a
707 		 * ref on the blkg.
708 		 */
709 		if (!blkg_tryget(blkg))
710 			continue;
711 
712 		iolat = blkg_to_lat(blkg);
713 		if (!iolat)
714 			goto next;
715 
716 		lat_info = &iolat->child_lat;
717 		cookie = atomic_read(&lat_info->scale_cookie);
718 
719 		if (cookie >= DEFAULT_SCALE_COOKIE)
720 			goto next;
721 
722 		spin_lock_irqsave(&lat_info->lock, flags);
723 		if (lat_info->last_scale_event >= now)
724 			goto next_lock;
725 
726 		/*
727 		 * We scaled down but don't have a scale_grp, scale up and carry
728 		 * on.
729 		 */
730 		if (lat_info->scale_grp == NULL) {
731 			scale_cookie_change(iolat->blkiolat, lat_info, true);
732 			goto next_lock;
733 		}
734 
735 		/*
736 		 * It's been 5 seconds since our last scale event, clear the
737 		 * scale grp in case the group that needed the scale down isn't
738 		 * doing any IO currently.
739 		 */
740 		if (now - lat_info->last_scale_event >=
741 		    ((u64)NSEC_PER_SEC * 5))
742 			lat_info->scale_grp = NULL;
743 next_lock:
744 		spin_unlock_irqrestore(&lat_info->lock, flags);
745 next:
746 		blkg_put(blkg);
747 	}
748 	rcu_read_unlock();
749 }
750 
751 int blk_iolatency_init(struct request_queue *q)
752 {
753 	struct blk_iolatency *blkiolat;
754 	struct rq_qos *rqos;
755 	int ret;
756 
757 	blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
758 	if (!blkiolat)
759 		return -ENOMEM;
760 
761 	rqos = &blkiolat->rqos;
762 	rqos->id = RQ_QOS_CGROUP;
763 	rqos->ops = &blkcg_iolatency_ops;
764 	rqos->q = q;
765 
766 	rq_qos_add(q, rqos);
767 
768 	ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
769 	if (ret) {
770 		rq_qos_del(q, rqos);
771 		kfree(blkiolat);
772 		return ret;
773 	}
774 
775 	timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
776 
777 	return 0;
778 }
779 
780 static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
781 {
782 	struct iolatency_grp *iolat = blkg_to_lat(blkg);
783 	struct blk_iolatency *blkiolat = iolat->blkiolat;
784 	u64 oldval = iolat->min_lat_nsec;
785 
786 	iolat->min_lat_nsec = val;
787 	iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
788 	iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
789 				    BLKIOLATENCY_MAX_WIN_SIZE);
790 
791 	if (!oldval && val)
792 		atomic_inc(&blkiolat->enabled);
793 	if (oldval && !val)
794 		atomic_dec(&blkiolat->enabled);
795 }
796 
797 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
798 {
799 	if (blkg->parent) {
800 		struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
801 		struct child_latency_info *lat_info;
802 		if (!iolat)
803 			return;
804 
805 		lat_info = &iolat->child_lat;
806 		spin_lock(&lat_info->lock);
807 		atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
808 		lat_info->last_scale_event = 0;
809 		lat_info->scale_grp = NULL;
810 		lat_info->scale_lat = 0;
811 		spin_unlock(&lat_info->lock);
812 	}
813 }
814 
815 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
816 			     size_t nbytes, loff_t off)
817 {
818 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
819 	struct blkcg_gq *blkg;
820 	struct blkg_conf_ctx ctx;
821 	struct iolatency_grp *iolat;
822 	char *p, *tok;
823 	u64 lat_val = 0;
824 	u64 oldval;
825 	int ret;
826 
827 	ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
828 	if (ret)
829 		return ret;
830 
831 	iolat = blkg_to_lat(ctx.blkg);
832 	p = ctx.body;
833 
834 	ret = -EINVAL;
835 	while ((tok = strsep(&p, " "))) {
836 		char key[16];
837 		char val[21];	/* 18446744073709551616 */
838 
839 		if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
840 			goto out;
841 
842 		if (!strcmp(key, "target")) {
843 			u64 v;
844 
845 			if (!strcmp(val, "max"))
846 				lat_val = 0;
847 			else if (sscanf(val, "%llu", &v) == 1)
848 				lat_val = v * NSEC_PER_USEC;
849 			else
850 				goto out;
851 		} else {
852 			goto out;
853 		}
854 	}
855 
856 	/* Walk up the tree to see if our new val is lower than it should be. */
857 	blkg = ctx.blkg;
858 	oldval = iolat->min_lat_nsec;
859 
860 	iolatency_set_min_lat_nsec(blkg, lat_val);
861 	if (oldval != iolat->min_lat_nsec) {
862 		iolatency_clear_scaling(blkg);
863 	}
864 
865 	ret = 0;
866 out:
867 	blkg_conf_finish(&ctx);
868 	return ret ?: nbytes;
869 }
870 
871 static u64 iolatency_prfill_limit(struct seq_file *sf,
872 				  struct blkg_policy_data *pd, int off)
873 {
874 	struct iolatency_grp *iolat = pd_to_lat(pd);
875 	const char *dname = blkg_dev_name(pd->blkg);
876 
877 	if (!dname || !iolat->min_lat_nsec)
878 		return 0;
879 	seq_printf(sf, "%s target=%llu\n",
880 		   dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
881 	return 0;
882 }
883 
884 static int iolatency_print_limit(struct seq_file *sf, void *v)
885 {
886 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
887 			  iolatency_prfill_limit,
888 			  &blkcg_policy_iolatency, seq_cft(sf)->private, false);
889 	return 0;
890 }
891 
892 static size_t iolatency_ssd_stat(struct iolatency_grp *iolat, char *buf,
893 				 size_t size)
894 {
895 	struct latency_stat stat;
896 	int cpu;
897 
898 	latency_stat_init(iolat, &stat);
899 	preempt_disable();
900 	for_each_online_cpu(cpu) {
901 		struct latency_stat *s;
902 		s = per_cpu_ptr(iolat->stats, cpu);
903 		latency_stat_sum(iolat, &stat, s);
904 	}
905 	preempt_enable();
906 
907 	if (iolat->rq_depth.max_depth == UINT_MAX)
908 		return scnprintf(buf, size, " missed=%llu total=%llu depth=max",
909 				 (unsigned long long)stat.ps.missed,
910 				 (unsigned long long)stat.ps.total);
911 	return scnprintf(buf, size, " missed=%llu total=%llu depth=%u",
912 			 (unsigned long long)stat.ps.missed,
913 			 (unsigned long long)stat.ps.total,
914 			 iolat->rq_depth.max_depth);
915 }
916 
917 static size_t iolatency_pd_stat(struct blkg_policy_data *pd, char *buf,
918 				size_t size)
919 {
920 	struct iolatency_grp *iolat = pd_to_lat(pd);
921 	unsigned long long avg_lat;
922 	unsigned long long cur_win;
923 
924 	if (iolat->ssd)
925 		return iolatency_ssd_stat(iolat, buf, size);
926 
927 	avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
928 	cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
929 	if (iolat->rq_depth.max_depth == UINT_MAX)
930 		return scnprintf(buf, size, " depth=max avg_lat=%llu win=%llu",
931 				 avg_lat, cur_win);
932 
933 	return scnprintf(buf, size, " depth=%u avg_lat=%llu win=%llu",
934 			 iolat->rq_depth.max_depth, avg_lat, cur_win);
935 }
936 
937 
938 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp, int node)
939 {
940 	struct iolatency_grp *iolat;
941 
942 	iolat = kzalloc_node(sizeof(*iolat), gfp, node);
943 	if (!iolat)
944 		return NULL;
945 	iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
946 				       __alignof__(struct latency_stat), gfp);
947 	if (!iolat->stats) {
948 		kfree(iolat);
949 		return NULL;
950 	}
951 	return &iolat->pd;
952 }
953 
954 static void iolatency_pd_init(struct blkg_policy_data *pd)
955 {
956 	struct iolatency_grp *iolat = pd_to_lat(pd);
957 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
958 	struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
959 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
960 	u64 now = ktime_to_ns(ktime_get());
961 	int cpu;
962 
963 	if (blk_queue_nonrot(blkg->q))
964 		iolat->ssd = true;
965 	else
966 		iolat->ssd = false;
967 
968 	for_each_possible_cpu(cpu) {
969 		struct latency_stat *stat;
970 		stat = per_cpu_ptr(iolat->stats, cpu);
971 		latency_stat_init(iolat, stat);
972 	}
973 
974 	latency_stat_init(iolat, &iolat->cur_stat);
975 	rq_wait_init(&iolat->rq_wait);
976 	spin_lock_init(&iolat->child_lat.lock);
977 	iolat->rq_depth.queue_depth = blkg->q->nr_requests;
978 	iolat->rq_depth.max_depth = UINT_MAX;
979 	iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
980 	iolat->blkiolat = blkiolat;
981 	iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
982 	atomic64_set(&iolat->window_start, now);
983 
984 	/*
985 	 * We init things in list order, so the pd for the parent may not be
986 	 * init'ed yet for whatever reason.
987 	 */
988 	if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
989 		struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
990 		atomic_set(&iolat->scale_cookie,
991 			   atomic_read(&parent->child_lat.scale_cookie));
992 	} else {
993 		atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
994 	}
995 
996 	atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
997 }
998 
999 static void iolatency_pd_offline(struct blkg_policy_data *pd)
1000 {
1001 	struct iolatency_grp *iolat = pd_to_lat(pd);
1002 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
1003 
1004 	iolatency_set_min_lat_nsec(blkg, 0);
1005 	iolatency_clear_scaling(blkg);
1006 }
1007 
1008 static void iolatency_pd_free(struct blkg_policy_data *pd)
1009 {
1010 	struct iolatency_grp *iolat = pd_to_lat(pd);
1011 	free_percpu(iolat->stats);
1012 	kfree(iolat);
1013 }
1014 
1015 static struct cftype iolatency_files[] = {
1016 	{
1017 		.name = "latency",
1018 		.flags = CFTYPE_NOT_ON_ROOT,
1019 		.seq_show = iolatency_print_limit,
1020 		.write = iolatency_set_limit,
1021 	},
1022 	{}
1023 };
1024 
1025 static struct blkcg_policy blkcg_policy_iolatency = {
1026 	.dfl_cftypes	= iolatency_files,
1027 	.pd_alloc_fn	= iolatency_pd_alloc,
1028 	.pd_init_fn	= iolatency_pd_init,
1029 	.pd_offline_fn	= iolatency_pd_offline,
1030 	.pd_free_fn	= iolatency_pd_free,
1031 	.pd_stat_fn	= iolatency_pd_stat,
1032 };
1033 
1034 static int __init iolatency_init(void)
1035 {
1036 	return blkcg_policy_register(&blkcg_policy_iolatency);
1037 }
1038 
1039 static void __exit iolatency_exit(void)
1040 {
1041 	return blkcg_policy_unregister(&blkcg_policy_iolatency);
1042 }
1043 
1044 module_init(iolatency_init);
1045 module_exit(iolatency_exit);
1046