xref: /linux/block/blk-iolatency.c (revision 3f0a50f345f78183f6e9b39c2f45ca5dcaa511ca)
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * Block rq-qos base io controller
4  *
5  * This works similar to wbt with a few exceptions
6  *
7  * - It's bio based, so the latency covers the whole block layer in addition to
8  *   the actual io.
9  * - We will throttle all IO that comes in here if we need to.
10  * - We use the mean latency over the 100ms window.  This is because writes can
11  *   be particularly fast, which could give us a false sense of the impact of
12  *   other workloads on our protected workload.
13  * - By default there's no throttling, we set the queue_depth to UINT_MAX so
14  *   that we can have as many outstanding bio's as we're allowed to.  Only at
15  *   throttle time do we pay attention to the actual queue depth.
16  *
17  * The hierarchy works like the cpu controller does, we track the latency at
18  * every configured node, and each configured node has it's own independent
19  * queue depth.  This means that we only care about our latency targets at the
20  * peer level.  Some group at the bottom of the hierarchy isn't going to affect
21  * a group at the end of some other path if we're only configred at leaf level.
22  *
23  * Consider the following
24  *
25  *                   root blkg
26  *             /                     \
27  *        fast (target=5ms)     slow (target=10ms)
28  *         /     \                  /        \
29  *       a        b          normal(15ms)   unloved
30  *
31  * "a" and "b" have no target, but their combined io under "fast" cannot exceed
32  * an average latency of 5ms.  If it does then we will throttle the "slow"
33  * group.  In the case of "normal", if it exceeds its 15ms target, we will
34  * throttle "unloved", but nobody else.
35  *
36  * In this example "fast", "slow", and "normal" will be the only groups actually
37  * accounting their io latencies.  We have to walk up the heirarchy to the root
38  * on every submit and complete so we can do the appropriate stat recording and
39  * adjust the queue depth of ourselves if needed.
40  *
41  * There are 2 ways we throttle IO.
42  *
43  * 1) Queue depth throttling.  As we throttle down we will adjust the maximum
44  * number of IO's we're allowed to have in flight.  This starts at (u64)-1 down
45  * to 1.  If the group is only ever submitting IO for itself then this is the
46  * only way we throttle.
47  *
48  * 2) Induced delay throttling.  This is for the case that a group is generating
49  * IO that has to be issued by the root cg to avoid priority inversion. So think
50  * REQ_META or REQ_SWAP.  If we are already at qd == 1 and we're getting a lot
51  * of work done for us on behalf of the root cg and are being asked to scale
52  * down more then we induce a latency at userspace return.  We accumulate the
53  * total amount of time we need to be punished by doing
54  *
55  * total_time += min_lat_nsec - actual_io_completion
56  *
57  * and then at throttle time will do
58  *
59  * throttle_time = min(total_time, NSEC_PER_SEC)
60  *
61  * This induced delay will throttle back the activity that is generating the
62  * root cg issued io's, wethere that's some metadata intensive operation or the
63  * group is using so much memory that it is pushing us into swap.
64  *
65  * Copyright (C) 2018 Josef Bacik
66  */
67 #include <linux/kernel.h>
68 #include <linux/blk_types.h>
69 #include <linux/backing-dev.h>
70 #include <linux/module.h>
71 #include <linux/timer.h>
72 #include <linux/memcontrol.h>
73 #include <linux/sched/loadavg.h>
74 #include <linux/sched/signal.h>
75 #include <trace/events/block.h>
76 #include <linux/blk-mq.h>
77 #include "blk-rq-qos.h"
78 #include "blk-stat.h"
79 #include "blk-cgroup.h"
80 #include "blk.h"
81 
82 #define DEFAULT_SCALE_COOKIE 1000000U
83 
84 static struct blkcg_policy blkcg_policy_iolatency;
85 struct iolatency_grp;
86 
87 struct blk_iolatency {
88 	struct rq_qos rqos;
89 	struct timer_list timer;
90 	atomic_t enabled;
91 };
92 
93 static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
94 {
95 	return container_of(rqos, struct blk_iolatency, rqos);
96 }
97 
98 static inline bool blk_iolatency_enabled(struct blk_iolatency *blkiolat)
99 {
100 	return atomic_read(&blkiolat->enabled) > 0;
101 }
102 
103 struct child_latency_info {
104 	spinlock_t lock;
105 
106 	/* Last time we adjusted the scale of everybody. */
107 	u64 last_scale_event;
108 
109 	/* The latency that we missed. */
110 	u64 scale_lat;
111 
112 	/* Total io's from all of our children for the last summation. */
113 	u64 nr_samples;
114 
115 	/* The guy who actually changed the latency numbers. */
116 	struct iolatency_grp *scale_grp;
117 
118 	/* Cookie to tell if we need to scale up or down. */
119 	atomic_t scale_cookie;
120 };
121 
122 struct percentile_stats {
123 	u64 total;
124 	u64 missed;
125 };
126 
127 struct latency_stat {
128 	union {
129 		struct percentile_stats ps;
130 		struct blk_rq_stat rqs;
131 	};
132 };
133 
134 struct iolatency_grp {
135 	struct blkg_policy_data pd;
136 	struct latency_stat __percpu *stats;
137 	struct latency_stat cur_stat;
138 	struct blk_iolatency *blkiolat;
139 	struct rq_depth rq_depth;
140 	struct rq_wait rq_wait;
141 	atomic64_t window_start;
142 	atomic_t scale_cookie;
143 	u64 min_lat_nsec;
144 	u64 cur_win_nsec;
145 
146 	/* total running average of our io latency. */
147 	u64 lat_avg;
148 
149 	/* Our current number of IO's for the last summation. */
150 	u64 nr_samples;
151 
152 	bool ssd;
153 	struct child_latency_info child_lat;
154 };
155 
156 #define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
157 #define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
158 /*
159  * These are the constants used to fake the fixed-point moving average
160  * calculation just like load average.  The call to calc_load() folds
161  * (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg.  The sampling
162  * window size is bucketed to try to approximately calculate average
163  * latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
164  * elapse immediately.  Note, windows only elapse with IO activity.  Idle
165  * periods extend the most recent window.
166  */
167 #define BLKIOLATENCY_NR_EXP_FACTORS 5
168 #define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
169 				      (BLKIOLATENCY_NR_EXP_FACTORS - 1))
170 static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
171 	2045, // exp(1/600) - 600 samples
172 	2039, // exp(1/240) - 240 samples
173 	2031, // exp(1/120) - 120 samples
174 	2023, // exp(1/80)  - 80 samples
175 	2014, // exp(1/60)  - 60 samples
176 };
177 
178 static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
179 {
180 	return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
181 }
182 
183 static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
184 {
185 	return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
186 }
187 
188 static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
189 {
190 	return pd_to_blkg(&iolat->pd);
191 }
192 
193 static inline void latency_stat_init(struct iolatency_grp *iolat,
194 				     struct latency_stat *stat)
195 {
196 	if (iolat->ssd) {
197 		stat->ps.total = 0;
198 		stat->ps.missed = 0;
199 	} else
200 		blk_rq_stat_init(&stat->rqs);
201 }
202 
203 static inline void latency_stat_sum(struct iolatency_grp *iolat,
204 				    struct latency_stat *sum,
205 				    struct latency_stat *stat)
206 {
207 	if (iolat->ssd) {
208 		sum->ps.total += stat->ps.total;
209 		sum->ps.missed += stat->ps.missed;
210 	} else
211 		blk_rq_stat_sum(&sum->rqs, &stat->rqs);
212 }
213 
214 static inline void latency_stat_record_time(struct iolatency_grp *iolat,
215 					    u64 req_time)
216 {
217 	struct latency_stat *stat = get_cpu_ptr(iolat->stats);
218 	if (iolat->ssd) {
219 		if (req_time >= iolat->min_lat_nsec)
220 			stat->ps.missed++;
221 		stat->ps.total++;
222 	} else
223 		blk_rq_stat_add(&stat->rqs, req_time);
224 	put_cpu_ptr(stat);
225 }
226 
227 static inline bool latency_sum_ok(struct iolatency_grp *iolat,
228 				  struct latency_stat *stat)
229 {
230 	if (iolat->ssd) {
231 		u64 thresh = div64_u64(stat->ps.total, 10);
232 		thresh = max(thresh, 1ULL);
233 		return stat->ps.missed < thresh;
234 	}
235 	return stat->rqs.mean <= iolat->min_lat_nsec;
236 }
237 
238 static inline u64 latency_stat_samples(struct iolatency_grp *iolat,
239 				       struct latency_stat *stat)
240 {
241 	if (iolat->ssd)
242 		return stat->ps.total;
243 	return stat->rqs.nr_samples;
244 }
245 
246 static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat,
247 					      struct latency_stat *stat)
248 {
249 	int exp_idx;
250 
251 	if (iolat->ssd)
252 		return;
253 
254 	/*
255 	 * calc_load() takes in a number stored in fixed point representation.
256 	 * Because we are using this for IO time in ns, the values stored
257 	 * are significantly larger than the FIXED_1 denominator (2048).
258 	 * Therefore, rounding errors in the calculation are negligible and
259 	 * can be ignored.
260 	 */
261 	exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
262 			div64_u64(iolat->cur_win_nsec,
263 				  BLKIOLATENCY_EXP_BUCKET_SIZE));
264 	iolat->lat_avg = calc_load(iolat->lat_avg,
265 				   iolatency_exp_factors[exp_idx],
266 				   stat->rqs.mean);
267 }
268 
269 static void iolat_cleanup_cb(struct rq_wait *rqw, void *private_data)
270 {
271 	atomic_dec(&rqw->inflight);
272 	wake_up(&rqw->wait);
273 }
274 
275 static bool iolat_acquire_inflight(struct rq_wait *rqw, void *private_data)
276 {
277 	struct iolatency_grp *iolat = private_data;
278 	return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
279 }
280 
281 static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
282 				       struct iolatency_grp *iolat,
283 				       bool issue_as_root,
284 				       bool use_memdelay)
285 {
286 	struct rq_wait *rqw = &iolat->rq_wait;
287 	unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
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 	rq_qos_wait(rqw, iolat, iolat_acquire_inflight, iolat_cleanup_cb);
305 }
306 
307 #define SCALE_DOWN_FACTOR 2
308 #define SCALE_UP_FACTOR 4
309 
310 static inline unsigned long scale_amount(unsigned long qd, bool up)
311 {
312 	return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
313 }
314 
315 /*
316  * We scale the qd down faster than we scale up, so we need to use this helper
317  * to adjust the scale_cookie accordingly so we don't prematurely get
318  * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
319  *
320  * Each group has their own local copy of the last scale cookie they saw, so if
321  * the global scale cookie goes up or down they know which way they need to go
322  * based on their last knowledge of it.
323  */
324 static void scale_cookie_change(struct blk_iolatency *blkiolat,
325 				struct child_latency_info *lat_info,
326 				bool up)
327 {
328 	unsigned long qd = blkiolat->rqos.q->nr_requests;
329 	unsigned long scale = scale_amount(qd, up);
330 	unsigned long old = atomic_read(&lat_info->scale_cookie);
331 	unsigned long max_scale = qd << 1;
332 	unsigned long diff = 0;
333 
334 	if (old < DEFAULT_SCALE_COOKIE)
335 		diff = DEFAULT_SCALE_COOKIE - old;
336 
337 	if (up) {
338 		if (scale + old > DEFAULT_SCALE_COOKIE)
339 			atomic_set(&lat_info->scale_cookie,
340 				   DEFAULT_SCALE_COOKIE);
341 		else if (diff > qd)
342 			atomic_inc(&lat_info->scale_cookie);
343 		else
344 			atomic_add(scale, &lat_info->scale_cookie);
345 	} else {
346 		/*
347 		 * We don't want to dig a hole so deep that it takes us hours to
348 		 * dig out of it.  Just enough that we don't throttle/unthrottle
349 		 * with jagged workloads but can still unthrottle once pressure
350 		 * has sufficiently dissipated.
351 		 */
352 		if (diff > qd) {
353 			if (diff < max_scale)
354 				atomic_dec(&lat_info->scale_cookie);
355 		} else {
356 			atomic_sub(scale, &lat_info->scale_cookie);
357 		}
358 	}
359 }
360 
361 /*
362  * Change the queue depth of the iolatency_grp.  We add/subtract 1/16th of the
363  * queue depth at a time so we don't get wild swings and hopefully dial in to
364  * fairer distribution of the overall queue depth.
365  */
366 static void scale_change(struct iolatency_grp *iolat, bool up)
367 {
368 	unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
369 	unsigned long scale = scale_amount(qd, up);
370 	unsigned long old = iolat->rq_depth.max_depth;
371 
372 	if (old > qd)
373 		old = qd;
374 
375 	if (up) {
376 		if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
377 			return;
378 
379 		if (old < qd) {
380 			old += scale;
381 			old = min(old, qd);
382 			iolat->rq_depth.max_depth = old;
383 			wake_up_all(&iolat->rq_wait.wait);
384 		}
385 	} else {
386 		old >>= 1;
387 		iolat->rq_depth.max_depth = max(old, 1UL);
388 	}
389 }
390 
391 /* Check our parent and see if the scale cookie has changed. */
392 static void check_scale_change(struct iolatency_grp *iolat)
393 {
394 	struct iolatency_grp *parent;
395 	struct child_latency_info *lat_info;
396 	unsigned int cur_cookie;
397 	unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
398 	u64 scale_lat;
399 	unsigned int old;
400 	int direction = 0;
401 
402 	if (lat_to_blkg(iolat)->parent == NULL)
403 		return;
404 
405 	parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
406 	if (!parent)
407 		return;
408 
409 	lat_info = &parent->child_lat;
410 	cur_cookie = atomic_read(&lat_info->scale_cookie);
411 	scale_lat = READ_ONCE(lat_info->scale_lat);
412 
413 	if (cur_cookie < our_cookie)
414 		direction = -1;
415 	else if (cur_cookie > our_cookie)
416 		direction = 1;
417 	else
418 		return;
419 
420 	old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
421 
422 	/* Somebody beat us to the punch, just bail. */
423 	if (old != our_cookie)
424 		return;
425 
426 	if (direction < 0 && iolat->min_lat_nsec) {
427 		u64 samples_thresh;
428 
429 		if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
430 			return;
431 
432 		/*
433 		 * Sometimes high priority groups are their own worst enemy, so
434 		 * instead of taking it out on some poor other group that did 5%
435 		 * or less of the IO's for the last summation just skip this
436 		 * scale down event.
437 		 */
438 		samples_thresh = lat_info->nr_samples * 5;
439 		samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
440 		if (iolat->nr_samples <= samples_thresh)
441 			return;
442 	}
443 
444 	/* We're as low as we can go. */
445 	if (iolat->rq_depth.max_depth == 1 && direction < 0) {
446 		blkcg_use_delay(lat_to_blkg(iolat));
447 		return;
448 	}
449 
450 	/* We're back to the default cookie, unthrottle all the things. */
451 	if (cur_cookie == DEFAULT_SCALE_COOKIE) {
452 		blkcg_clear_delay(lat_to_blkg(iolat));
453 		iolat->rq_depth.max_depth = UINT_MAX;
454 		wake_up_all(&iolat->rq_wait.wait);
455 		return;
456 	}
457 
458 	scale_change(iolat, direction > 0);
459 }
460 
461 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
462 {
463 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
464 	struct blkcg_gq *blkg = bio->bi_blkg;
465 	bool issue_as_root = bio_issue_as_root_blkg(bio);
466 
467 	if (!blk_iolatency_enabled(blkiolat))
468 		return;
469 
470 	while (blkg && blkg->parent) {
471 		struct iolatency_grp *iolat = blkg_to_lat(blkg);
472 		if (!iolat) {
473 			blkg = blkg->parent;
474 			continue;
475 		}
476 
477 		check_scale_change(iolat);
478 		__blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
479 				     (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
480 		blkg = blkg->parent;
481 	}
482 	if (!timer_pending(&blkiolat->timer))
483 		mod_timer(&blkiolat->timer, jiffies + HZ);
484 }
485 
486 static void iolatency_record_time(struct iolatency_grp *iolat,
487 				  struct bio_issue *issue, u64 now,
488 				  bool issue_as_root)
489 {
490 	u64 start = bio_issue_time(issue);
491 	u64 req_time;
492 
493 	/*
494 	 * Have to do this so we are truncated to the correct time that our
495 	 * issue is truncated to.
496 	 */
497 	now = __bio_issue_time(now);
498 
499 	if (now <= start)
500 		return;
501 
502 	req_time = now - start;
503 
504 	/*
505 	 * We don't want to count issue_as_root bio's in the cgroups latency
506 	 * statistics as it could skew the numbers downwards.
507 	 */
508 	if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
509 		u64 sub = iolat->min_lat_nsec;
510 		if (req_time < sub)
511 			blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
512 		return;
513 	}
514 
515 	latency_stat_record_time(iolat, req_time);
516 }
517 
518 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
519 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
520 
521 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
522 {
523 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
524 	struct iolatency_grp *parent;
525 	struct child_latency_info *lat_info;
526 	struct latency_stat stat;
527 	unsigned long flags;
528 	int cpu;
529 
530 	latency_stat_init(iolat, &stat);
531 	preempt_disable();
532 	for_each_online_cpu(cpu) {
533 		struct latency_stat *s;
534 		s = per_cpu_ptr(iolat->stats, cpu);
535 		latency_stat_sum(iolat, &stat, s);
536 		latency_stat_init(iolat, s);
537 	}
538 	preempt_enable();
539 
540 	parent = blkg_to_lat(blkg->parent);
541 	if (!parent)
542 		return;
543 
544 	lat_info = &parent->child_lat;
545 
546 	iolat_update_total_lat_avg(iolat, &stat);
547 
548 	/* Everything is ok and we don't need to adjust the scale. */
549 	if (latency_sum_ok(iolat, &stat) &&
550 	    atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
551 		return;
552 
553 	/* Somebody beat us to the punch, just bail. */
554 	spin_lock_irqsave(&lat_info->lock, flags);
555 
556 	latency_stat_sum(iolat, &iolat->cur_stat, &stat);
557 	lat_info->nr_samples -= iolat->nr_samples;
558 	lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
559 	iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
560 
561 	if ((lat_info->last_scale_event >= now ||
562 	    now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
563 		goto out;
564 
565 	if (latency_sum_ok(iolat, &iolat->cur_stat) &&
566 	    latency_sum_ok(iolat, &stat)) {
567 		if (latency_stat_samples(iolat, &iolat->cur_stat) <
568 		    BLKIOLATENCY_MIN_GOOD_SAMPLES)
569 			goto out;
570 		if (lat_info->scale_grp == iolat) {
571 			lat_info->last_scale_event = now;
572 			scale_cookie_change(iolat->blkiolat, lat_info, true);
573 		}
574 	} else if (lat_info->scale_lat == 0 ||
575 		   lat_info->scale_lat >= iolat->min_lat_nsec) {
576 		lat_info->last_scale_event = now;
577 		if (!lat_info->scale_grp ||
578 		    lat_info->scale_lat > iolat->min_lat_nsec) {
579 			WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
580 			lat_info->scale_grp = iolat;
581 		}
582 		scale_cookie_change(iolat->blkiolat, lat_info, false);
583 	}
584 	latency_stat_init(iolat, &iolat->cur_stat);
585 out:
586 	spin_unlock_irqrestore(&lat_info->lock, flags);
587 }
588 
589 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
590 {
591 	struct blkcg_gq *blkg;
592 	struct rq_wait *rqw;
593 	struct iolatency_grp *iolat;
594 	u64 window_start;
595 	u64 now;
596 	bool issue_as_root = bio_issue_as_root_blkg(bio);
597 	bool enabled = false;
598 	int inflight = 0;
599 
600 	blkg = bio->bi_blkg;
601 	if (!blkg || !bio_flagged(bio, BIO_QOS_THROTTLED))
602 		return;
603 
604 	iolat = blkg_to_lat(bio->bi_blkg);
605 	if (!iolat)
606 		return;
607 
608 	enabled = blk_iolatency_enabled(iolat->blkiolat);
609 	if (!enabled)
610 		return;
611 
612 	now = ktime_to_ns(ktime_get());
613 	while (blkg && blkg->parent) {
614 		iolat = blkg_to_lat(blkg);
615 		if (!iolat) {
616 			blkg = blkg->parent;
617 			continue;
618 		}
619 		rqw = &iolat->rq_wait;
620 
621 		inflight = atomic_dec_return(&rqw->inflight);
622 		WARN_ON_ONCE(inflight < 0);
623 		/*
624 		 * If bi_status is BLK_STS_AGAIN, the bio wasn't actually
625 		 * submitted, so do not account for it.
626 		 */
627 		if (iolat->min_lat_nsec && bio->bi_status != BLK_STS_AGAIN) {
628 			iolatency_record_time(iolat, &bio->bi_issue, now,
629 					      issue_as_root);
630 			window_start = atomic64_read(&iolat->window_start);
631 			if (now > window_start &&
632 			    (now - window_start) >= iolat->cur_win_nsec) {
633 				if (atomic64_cmpxchg(&iolat->window_start,
634 					     window_start, now) == window_start)
635 					iolatency_check_latencies(iolat, now);
636 			}
637 		}
638 		wake_up(&rqw->wait);
639 		blkg = blkg->parent;
640 	}
641 }
642 
643 static void blkcg_iolatency_exit(struct rq_qos *rqos)
644 {
645 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
646 
647 	del_timer_sync(&blkiolat->timer);
648 	blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
649 	kfree(blkiolat);
650 }
651 
652 static struct rq_qos_ops blkcg_iolatency_ops = {
653 	.throttle = blkcg_iolatency_throttle,
654 	.done_bio = blkcg_iolatency_done_bio,
655 	.exit = blkcg_iolatency_exit,
656 };
657 
658 static void blkiolatency_timer_fn(struct timer_list *t)
659 {
660 	struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
661 	struct blkcg_gq *blkg;
662 	struct cgroup_subsys_state *pos_css;
663 	u64 now = ktime_to_ns(ktime_get());
664 
665 	rcu_read_lock();
666 	blkg_for_each_descendant_pre(blkg, pos_css,
667 				     blkiolat->rqos.q->root_blkg) {
668 		struct iolatency_grp *iolat;
669 		struct child_latency_info *lat_info;
670 		unsigned long flags;
671 		u64 cookie;
672 
673 		/*
674 		 * We could be exiting, don't access the pd unless we have a
675 		 * ref on the blkg.
676 		 */
677 		if (!blkg_tryget(blkg))
678 			continue;
679 
680 		iolat = blkg_to_lat(blkg);
681 		if (!iolat)
682 			goto next;
683 
684 		lat_info = &iolat->child_lat;
685 		cookie = atomic_read(&lat_info->scale_cookie);
686 
687 		if (cookie >= DEFAULT_SCALE_COOKIE)
688 			goto next;
689 
690 		spin_lock_irqsave(&lat_info->lock, flags);
691 		if (lat_info->last_scale_event >= now)
692 			goto next_lock;
693 
694 		/*
695 		 * We scaled down but don't have a scale_grp, scale up and carry
696 		 * on.
697 		 */
698 		if (lat_info->scale_grp == NULL) {
699 			scale_cookie_change(iolat->blkiolat, lat_info, true);
700 			goto next_lock;
701 		}
702 
703 		/*
704 		 * It's been 5 seconds since our last scale event, clear the
705 		 * scale grp in case the group that needed the scale down isn't
706 		 * doing any IO currently.
707 		 */
708 		if (now - lat_info->last_scale_event >=
709 		    ((u64)NSEC_PER_SEC * 5))
710 			lat_info->scale_grp = NULL;
711 next_lock:
712 		spin_unlock_irqrestore(&lat_info->lock, flags);
713 next:
714 		blkg_put(blkg);
715 	}
716 	rcu_read_unlock();
717 }
718 
719 int blk_iolatency_init(struct request_queue *q)
720 {
721 	struct blk_iolatency *blkiolat;
722 	struct rq_qos *rqos;
723 	int ret;
724 
725 	blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
726 	if (!blkiolat)
727 		return -ENOMEM;
728 
729 	rqos = &blkiolat->rqos;
730 	rqos->id = RQ_QOS_LATENCY;
731 	rqos->ops = &blkcg_iolatency_ops;
732 	rqos->q = q;
733 
734 	rq_qos_add(q, rqos);
735 
736 	ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
737 	if (ret) {
738 		rq_qos_del(q, rqos);
739 		kfree(blkiolat);
740 		return ret;
741 	}
742 
743 	timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
744 
745 	return 0;
746 }
747 
748 /*
749  * return 1 for enabling iolatency, return -1 for disabling iolatency, otherwise
750  * return 0.
751  */
752 static int iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
753 {
754 	struct iolatency_grp *iolat = blkg_to_lat(blkg);
755 	u64 oldval = iolat->min_lat_nsec;
756 
757 	iolat->min_lat_nsec = val;
758 	iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
759 	iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
760 				    BLKIOLATENCY_MAX_WIN_SIZE);
761 
762 	if (!oldval && val)
763 		return 1;
764 	if (oldval && !val) {
765 		blkcg_clear_delay(blkg);
766 		return -1;
767 	}
768 	return 0;
769 }
770 
771 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
772 {
773 	if (blkg->parent) {
774 		struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
775 		struct child_latency_info *lat_info;
776 		if (!iolat)
777 			return;
778 
779 		lat_info = &iolat->child_lat;
780 		spin_lock(&lat_info->lock);
781 		atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
782 		lat_info->last_scale_event = 0;
783 		lat_info->scale_grp = NULL;
784 		lat_info->scale_lat = 0;
785 		spin_unlock(&lat_info->lock);
786 	}
787 }
788 
789 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
790 			     size_t nbytes, loff_t off)
791 {
792 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
793 	struct blkcg_gq *blkg;
794 	struct blkg_conf_ctx ctx;
795 	struct iolatency_grp *iolat;
796 	char *p, *tok;
797 	u64 lat_val = 0;
798 	u64 oldval;
799 	int ret;
800 	int enable = 0;
801 
802 	ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
803 	if (ret)
804 		return ret;
805 
806 	iolat = blkg_to_lat(ctx.blkg);
807 	p = ctx.body;
808 
809 	ret = -EINVAL;
810 	while ((tok = strsep(&p, " "))) {
811 		char key[16];
812 		char val[21];	/* 18446744073709551616 */
813 
814 		if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
815 			goto out;
816 
817 		if (!strcmp(key, "target")) {
818 			u64 v;
819 
820 			if (!strcmp(val, "max"))
821 				lat_val = 0;
822 			else if (sscanf(val, "%llu", &v) == 1)
823 				lat_val = v * NSEC_PER_USEC;
824 			else
825 				goto out;
826 		} else {
827 			goto out;
828 		}
829 	}
830 
831 	/* Walk up the tree to see if our new val is lower than it should be. */
832 	blkg = ctx.blkg;
833 	oldval = iolat->min_lat_nsec;
834 
835 	enable = iolatency_set_min_lat_nsec(blkg, lat_val);
836 	if (enable) {
837 		if (!blk_get_queue(blkg->q)) {
838 			ret = -ENODEV;
839 			goto out;
840 		}
841 
842 		blkg_get(blkg);
843 	}
844 
845 	if (oldval != iolat->min_lat_nsec) {
846 		iolatency_clear_scaling(blkg);
847 	}
848 
849 	ret = 0;
850 out:
851 	blkg_conf_finish(&ctx);
852 	if (ret == 0 && enable) {
853 		struct iolatency_grp *tmp = blkg_to_lat(blkg);
854 		struct blk_iolatency *blkiolat = tmp->blkiolat;
855 
856 		blk_mq_freeze_queue(blkg->q);
857 
858 		if (enable == 1)
859 			atomic_inc(&blkiolat->enabled);
860 		else if (enable == -1)
861 			atomic_dec(&blkiolat->enabled);
862 		else
863 			WARN_ON_ONCE(1);
864 
865 		blk_mq_unfreeze_queue(blkg->q);
866 
867 		blkg_put(blkg);
868 		blk_put_queue(blkg->q);
869 	}
870 	return ret ?: nbytes;
871 }
872 
873 static u64 iolatency_prfill_limit(struct seq_file *sf,
874 				  struct blkg_policy_data *pd, int off)
875 {
876 	struct iolatency_grp *iolat = pd_to_lat(pd);
877 	const char *dname = blkg_dev_name(pd->blkg);
878 
879 	if (!dname || !iolat->min_lat_nsec)
880 		return 0;
881 	seq_printf(sf, "%s target=%llu\n",
882 		   dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
883 	return 0;
884 }
885 
886 static int iolatency_print_limit(struct seq_file *sf, void *v)
887 {
888 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
889 			  iolatency_prfill_limit,
890 			  &blkcg_policy_iolatency, seq_cft(sf)->private, false);
891 	return 0;
892 }
893 
894 static bool iolatency_ssd_stat(struct iolatency_grp *iolat, struct seq_file *s)
895 {
896 	struct latency_stat stat;
897 	int cpu;
898 
899 	latency_stat_init(iolat, &stat);
900 	preempt_disable();
901 	for_each_online_cpu(cpu) {
902 		struct latency_stat *s;
903 		s = per_cpu_ptr(iolat->stats, cpu);
904 		latency_stat_sum(iolat, &stat, s);
905 	}
906 	preempt_enable();
907 
908 	if (iolat->rq_depth.max_depth == UINT_MAX)
909 		seq_printf(s, " missed=%llu total=%llu depth=max",
910 			(unsigned long long)stat.ps.missed,
911 			(unsigned long long)stat.ps.total);
912 	else
913 		seq_printf(s, " missed=%llu total=%llu depth=%u",
914 			(unsigned long long)stat.ps.missed,
915 			(unsigned long long)stat.ps.total,
916 			iolat->rq_depth.max_depth);
917 	return true;
918 }
919 
920 static bool iolatency_pd_stat(struct blkg_policy_data *pd, struct seq_file *s)
921 {
922 	struct iolatency_grp *iolat = pd_to_lat(pd);
923 	unsigned long long avg_lat;
924 	unsigned long long cur_win;
925 
926 	if (!blkcg_debug_stats)
927 		return false;
928 
929 	if (iolat->ssd)
930 		return iolatency_ssd_stat(iolat, s);
931 
932 	avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
933 	cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
934 	if (iolat->rq_depth.max_depth == UINT_MAX)
935 		seq_printf(s, " depth=max avg_lat=%llu win=%llu",
936 			avg_lat, cur_win);
937 	else
938 		seq_printf(s, " depth=%u avg_lat=%llu win=%llu",
939 			iolat->rq_depth.max_depth, avg_lat, cur_win);
940 	return true;
941 }
942 
943 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp,
944 						   struct request_queue *q,
945 						   struct blkcg *blkcg)
946 {
947 	struct iolatency_grp *iolat;
948 
949 	iolat = kzalloc_node(sizeof(*iolat), gfp, q->node);
950 	if (!iolat)
951 		return NULL;
952 	iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
953 				       __alignof__(struct latency_stat), gfp);
954 	if (!iolat->stats) {
955 		kfree(iolat);
956 		return NULL;
957 	}
958 	return &iolat->pd;
959 }
960 
961 static void iolatency_pd_init(struct blkg_policy_data *pd)
962 {
963 	struct iolatency_grp *iolat = pd_to_lat(pd);
964 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
965 	struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
966 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
967 	u64 now = ktime_to_ns(ktime_get());
968 	int cpu;
969 
970 	if (blk_queue_nonrot(blkg->q))
971 		iolat->ssd = true;
972 	else
973 		iolat->ssd = false;
974 
975 	for_each_possible_cpu(cpu) {
976 		struct latency_stat *stat;
977 		stat = per_cpu_ptr(iolat->stats, cpu);
978 		latency_stat_init(iolat, stat);
979 	}
980 
981 	latency_stat_init(iolat, &iolat->cur_stat);
982 	rq_wait_init(&iolat->rq_wait);
983 	spin_lock_init(&iolat->child_lat.lock);
984 	iolat->rq_depth.queue_depth = blkg->q->nr_requests;
985 	iolat->rq_depth.max_depth = UINT_MAX;
986 	iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
987 	iolat->blkiolat = blkiolat;
988 	iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
989 	atomic64_set(&iolat->window_start, now);
990 
991 	/*
992 	 * We init things in list order, so the pd for the parent may not be
993 	 * init'ed yet for whatever reason.
994 	 */
995 	if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
996 		struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
997 		atomic_set(&iolat->scale_cookie,
998 			   atomic_read(&parent->child_lat.scale_cookie));
999 	} else {
1000 		atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
1001 	}
1002 
1003 	atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
1004 }
1005 
1006 static void iolatency_pd_offline(struct blkg_policy_data *pd)
1007 {
1008 	struct iolatency_grp *iolat = pd_to_lat(pd);
1009 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
1010 	struct blk_iolatency *blkiolat = iolat->blkiolat;
1011 	int ret;
1012 
1013 	ret = iolatency_set_min_lat_nsec(blkg, 0);
1014 	if (ret == 1)
1015 		atomic_inc(&blkiolat->enabled);
1016 	if (ret == -1)
1017 		atomic_dec(&blkiolat->enabled);
1018 	iolatency_clear_scaling(blkg);
1019 }
1020 
1021 static void iolatency_pd_free(struct blkg_policy_data *pd)
1022 {
1023 	struct iolatency_grp *iolat = pd_to_lat(pd);
1024 	free_percpu(iolat->stats);
1025 	kfree(iolat);
1026 }
1027 
1028 static struct cftype iolatency_files[] = {
1029 	{
1030 		.name = "latency",
1031 		.flags = CFTYPE_NOT_ON_ROOT,
1032 		.seq_show = iolatency_print_limit,
1033 		.write = iolatency_set_limit,
1034 	},
1035 	{}
1036 };
1037 
1038 static struct blkcg_policy blkcg_policy_iolatency = {
1039 	.dfl_cftypes	= iolatency_files,
1040 	.pd_alloc_fn	= iolatency_pd_alloc,
1041 	.pd_init_fn	= iolatency_pd_init,
1042 	.pd_offline_fn	= iolatency_pd_offline,
1043 	.pd_free_fn	= iolatency_pd_free,
1044 	.pd_stat_fn	= iolatency_pd_stat,
1045 };
1046 
1047 static int __init iolatency_init(void)
1048 {
1049 	return blkcg_policy_register(&blkcg_policy_iolatency);
1050 }
1051 
1052 static void __exit iolatency_exit(void)
1053 {
1054 	blkcg_policy_unregister(&blkcg_policy_iolatency);
1055 }
1056 
1057 module_init(iolatency_init);
1058 module_exit(iolatency_exit);
1059