xref: /linux/block/blk-iolatency.c (revision 03f7c1d2a49acd30e38789cd809d3300721e9b0e)
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 
91 	/*
92 	 * ->enabled is the master enable switch gating the throttling logic and
93 	 * inflight tracking. The number of cgroups which have iolat enabled is
94 	 * tracked in ->enable_cnt, and ->enable is flipped on/off accordingly
95 	 * from ->enable_work with the request_queue frozen. For details, See
96 	 * blkiolatency_enable_work_fn().
97 	 */
98 	bool enabled;
99 	atomic_t enable_cnt;
100 	struct work_struct enable_work;
101 };
102 
103 static inline struct blk_iolatency *BLKIOLATENCY(struct rq_qos *rqos)
104 {
105 	return container_of(rqos, struct blk_iolatency, rqos);
106 }
107 
108 struct child_latency_info {
109 	spinlock_t lock;
110 
111 	/* Last time we adjusted the scale of everybody. */
112 	u64 last_scale_event;
113 
114 	/* The latency that we missed. */
115 	u64 scale_lat;
116 
117 	/* Total io's from all of our children for the last summation. */
118 	u64 nr_samples;
119 
120 	/* The guy who actually changed the latency numbers. */
121 	struct iolatency_grp *scale_grp;
122 
123 	/* Cookie to tell if we need to scale up or down. */
124 	atomic_t scale_cookie;
125 };
126 
127 struct percentile_stats {
128 	u64 total;
129 	u64 missed;
130 };
131 
132 struct latency_stat {
133 	union {
134 		struct percentile_stats ps;
135 		struct blk_rq_stat rqs;
136 	};
137 };
138 
139 struct iolatency_grp {
140 	struct blkg_policy_data pd;
141 	struct latency_stat __percpu *stats;
142 	struct latency_stat cur_stat;
143 	struct blk_iolatency *blkiolat;
144 	struct rq_depth rq_depth;
145 	struct rq_wait rq_wait;
146 	atomic64_t window_start;
147 	atomic_t scale_cookie;
148 	u64 min_lat_nsec;
149 	u64 cur_win_nsec;
150 
151 	/* total running average of our io latency. */
152 	u64 lat_avg;
153 
154 	/* Our current number of IO's for the last summation. */
155 	u64 nr_samples;
156 
157 	bool ssd;
158 	struct child_latency_info child_lat;
159 };
160 
161 #define BLKIOLATENCY_MIN_WIN_SIZE (100 * NSEC_PER_MSEC)
162 #define BLKIOLATENCY_MAX_WIN_SIZE NSEC_PER_SEC
163 /*
164  * These are the constants used to fake the fixed-point moving average
165  * calculation just like load average.  The call to calc_load() folds
166  * (FIXED_1 (2048) - exp_factor) * new_sample into lat_avg.  The sampling
167  * window size is bucketed to try to approximately calculate average
168  * latency such that 1/exp (decay rate) is [1 min, 2.5 min) when windows
169  * elapse immediately.  Note, windows only elapse with IO activity.  Idle
170  * periods extend the most recent window.
171  */
172 #define BLKIOLATENCY_NR_EXP_FACTORS 5
173 #define BLKIOLATENCY_EXP_BUCKET_SIZE (BLKIOLATENCY_MAX_WIN_SIZE / \
174 				      (BLKIOLATENCY_NR_EXP_FACTORS - 1))
175 static const u64 iolatency_exp_factors[BLKIOLATENCY_NR_EXP_FACTORS] = {
176 	2045, // exp(1/600) - 600 samples
177 	2039, // exp(1/240) - 240 samples
178 	2031, // exp(1/120) - 120 samples
179 	2023, // exp(1/80)  - 80 samples
180 	2014, // exp(1/60)  - 60 samples
181 };
182 
183 static inline struct iolatency_grp *pd_to_lat(struct blkg_policy_data *pd)
184 {
185 	return pd ? container_of(pd, struct iolatency_grp, pd) : NULL;
186 }
187 
188 static inline struct iolatency_grp *blkg_to_lat(struct blkcg_gq *blkg)
189 {
190 	return pd_to_lat(blkg_to_pd(blkg, &blkcg_policy_iolatency));
191 }
192 
193 static inline struct blkcg_gq *lat_to_blkg(struct iolatency_grp *iolat)
194 {
195 	return pd_to_blkg(&iolat->pd);
196 }
197 
198 static inline void latency_stat_init(struct iolatency_grp *iolat,
199 				     struct latency_stat *stat)
200 {
201 	if (iolat->ssd) {
202 		stat->ps.total = 0;
203 		stat->ps.missed = 0;
204 	} else
205 		blk_rq_stat_init(&stat->rqs);
206 }
207 
208 static inline void latency_stat_sum(struct iolatency_grp *iolat,
209 				    struct latency_stat *sum,
210 				    struct latency_stat *stat)
211 {
212 	if (iolat->ssd) {
213 		sum->ps.total += stat->ps.total;
214 		sum->ps.missed += stat->ps.missed;
215 	} else
216 		blk_rq_stat_sum(&sum->rqs, &stat->rqs);
217 }
218 
219 static inline void latency_stat_record_time(struct iolatency_grp *iolat,
220 					    u64 req_time)
221 {
222 	struct latency_stat *stat = get_cpu_ptr(iolat->stats);
223 	if (iolat->ssd) {
224 		if (req_time >= iolat->min_lat_nsec)
225 			stat->ps.missed++;
226 		stat->ps.total++;
227 	} else
228 		blk_rq_stat_add(&stat->rqs, req_time);
229 	put_cpu_ptr(stat);
230 }
231 
232 static inline bool latency_sum_ok(struct iolatency_grp *iolat,
233 				  struct latency_stat *stat)
234 {
235 	if (iolat->ssd) {
236 		u64 thresh = div64_u64(stat->ps.total, 10);
237 		thresh = max(thresh, 1ULL);
238 		return stat->ps.missed < thresh;
239 	}
240 	return stat->rqs.mean <= iolat->min_lat_nsec;
241 }
242 
243 static inline u64 latency_stat_samples(struct iolatency_grp *iolat,
244 				       struct latency_stat *stat)
245 {
246 	if (iolat->ssd)
247 		return stat->ps.total;
248 	return stat->rqs.nr_samples;
249 }
250 
251 static inline void iolat_update_total_lat_avg(struct iolatency_grp *iolat,
252 					      struct latency_stat *stat)
253 {
254 	int exp_idx;
255 
256 	if (iolat->ssd)
257 		return;
258 
259 	/*
260 	 * calc_load() takes in a number stored in fixed point representation.
261 	 * Because we are using this for IO time in ns, the values stored
262 	 * are significantly larger than the FIXED_1 denominator (2048).
263 	 * Therefore, rounding errors in the calculation are negligible and
264 	 * can be ignored.
265 	 */
266 	exp_idx = min_t(int, BLKIOLATENCY_NR_EXP_FACTORS - 1,
267 			div64_u64(iolat->cur_win_nsec,
268 				  BLKIOLATENCY_EXP_BUCKET_SIZE));
269 	iolat->lat_avg = calc_load(iolat->lat_avg,
270 				   iolatency_exp_factors[exp_idx],
271 				   stat->rqs.mean);
272 }
273 
274 static void iolat_cleanup_cb(struct rq_wait *rqw, void *private_data)
275 {
276 	atomic_dec(&rqw->inflight);
277 	wake_up(&rqw->wait);
278 }
279 
280 static bool iolat_acquire_inflight(struct rq_wait *rqw, void *private_data)
281 {
282 	struct iolatency_grp *iolat = private_data;
283 	return rq_wait_inc_below(rqw, iolat->rq_depth.max_depth);
284 }
285 
286 static void __blkcg_iolatency_throttle(struct rq_qos *rqos,
287 				       struct iolatency_grp *iolat,
288 				       bool issue_as_root,
289 				       bool use_memdelay)
290 {
291 	struct rq_wait *rqw = &iolat->rq_wait;
292 	unsigned use_delay = atomic_read(&lat_to_blkg(iolat)->use_delay);
293 
294 	if (use_delay)
295 		blkcg_schedule_throttle(rqos->q->disk, use_memdelay);
296 
297 	/*
298 	 * To avoid priority inversions we want to just take a slot if we are
299 	 * issuing as root.  If we're being killed off there's no point in
300 	 * delaying things, we may have been killed by OOM so throttling may
301 	 * make recovery take even longer, so just let the IO's through so the
302 	 * task can go away.
303 	 */
304 	if (issue_as_root || fatal_signal_pending(current)) {
305 		atomic_inc(&rqw->inflight);
306 		return;
307 	}
308 
309 	rq_qos_wait(rqw, iolat, iolat_acquire_inflight, iolat_cleanup_cb);
310 }
311 
312 #define SCALE_DOWN_FACTOR 2
313 #define SCALE_UP_FACTOR 4
314 
315 static inline unsigned long scale_amount(unsigned long qd, bool up)
316 {
317 	return max(up ? qd >> SCALE_UP_FACTOR : qd >> SCALE_DOWN_FACTOR, 1UL);
318 }
319 
320 /*
321  * We scale the qd down faster than we scale up, so we need to use this helper
322  * to adjust the scale_cookie accordingly so we don't prematurely get
323  * scale_cookie at DEFAULT_SCALE_COOKIE and unthrottle too much.
324  *
325  * Each group has their own local copy of the last scale cookie they saw, so if
326  * the global scale cookie goes up or down they know which way they need to go
327  * based on their last knowledge of it.
328  */
329 static void scale_cookie_change(struct blk_iolatency *blkiolat,
330 				struct child_latency_info *lat_info,
331 				bool up)
332 {
333 	unsigned long qd = blkiolat->rqos.q->nr_requests;
334 	unsigned long scale = scale_amount(qd, up);
335 	unsigned long old = atomic_read(&lat_info->scale_cookie);
336 	unsigned long max_scale = qd << 1;
337 	unsigned long diff = 0;
338 
339 	if (old < DEFAULT_SCALE_COOKIE)
340 		diff = DEFAULT_SCALE_COOKIE - old;
341 
342 	if (up) {
343 		if (scale + old > DEFAULT_SCALE_COOKIE)
344 			atomic_set(&lat_info->scale_cookie,
345 				   DEFAULT_SCALE_COOKIE);
346 		else if (diff > qd)
347 			atomic_inc(&lat_info->scale_cookie);
348 		else
349 			atomic_add(scale, &lat_info->scale_cookie);
350 	} else {
351 		/*
352 		 * We don't want to dig a hole so deep that it takes us hours to
353 		 * dig out of it.  Just enough that we don't throttle/unthrottle
354 		 * with jagged workloads but can still unthrottle once pressure
355 		 * has sufficiently dissipated.
356 		 */
357 		if (diff > qd) {
358 			if (diff < max_scale)
359 				atomic_dec(&lat_info->scale_cookie);
360 		} else {
361 			atomic_sub(scale, &lat_info->scale_cookie);
362 		}
363 	}
364 }
365 
366 /*
367  * Change the queue depth of the iolatency_grp.  We add/subtract 1/16th of the
368  * queue depth at a time so we don't get wild swings and hopefully dial in to
369  * fairer distribution of the overall queue depth.
370  */
371 static void scale_change(struct iolatency_grp *iolat, bool up)
372 {
373 	unsigned long qd = iolat->blkiolat->rqos.q->nr_requests;
374 	unsigned long scale = scale_amount(qd, up);
375 	unsigned long old = iolat->rq_depth.max_depth;
376 
377 	if (old > qd)
378 		old = qd;
379 
380 	if (up) {
381 		if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
382 			return;
383 
384 		if (old < qd) {
385 			old += scale;
386 			old = min(old, qd);
387 			iolat->rq_depth.max_depth = old;
388 			wake_up_all(&iolat->rq_wait.wait);
389 		}
390 	} else {
391 		old >>= 1;
392 		iolat->rq_depth.max_depth = max(old, 1UL);
393 	}
394 }
395 
396 /* Check our parent and see if the scale cookie has changed. */
397 static void check_scale_change(struct iolatency_grp *iolat)
398 {
399 	struct iolatency_grp *parent;
400 	struct child_latency_info *lat_info;
401 	unsigned int cur_cookie;
402 	unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
403 	u64 scale_lat;
404 	int direction = 0;
405 
406 	if (lat_to_blkg(iolat)->parent == NULL)
407 		return;
408 
409 	parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
410 	if (!parent)
411 		return;
412 
413 	lat_info = &parent->child_lat;
414 	cur_cookie = atomic_read(&lat_info->scale_cookie);
415 	scale_lat = READ_ONCE(lat_info->scale_lat);
416 
417 	if (cur_cookie < our_cookie)
418 		direction = -1;
419 	else if (cur_cookie > our_cookie)
420 		direction = 1;
421 	else
422 		return;
423 
424 	if (!atomic_try_cmpxchg(&iolat->scale_cookie, &our_cookie, cur_cookie)) {
425 		/* Somebody beat us to the punch, just bail. */
426 		return;
427 	}
428 
429 	if (direction < 0 && iolat->min_lat_nsec) {
430 		u64 samples_thresh;
431 
432 		if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
433 			return;
434 
435 		/*
436 		 * Sometimes high priority groups are their own worst enemy, so
437 		 * instead of taking it out on some poor other group that did 5%
438 		 * or less of the IO's for the last summation just skip this
439 		 * scale down event.
440 		 */
441 		samples_thresh = lat_info->nr_samples * 5;
442 		samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
443 		if (iolat->nr_samples <= samples_thresh)
444 			return;
445 	}
446 
447 	/* We're as low as we can go. */
448 	if (iolat->rq_depth.max_depth == 1 && direction < 0) {
449 		blkcg_use_delay(lat_to_blkg(iolat));
450 		return;
451 	}
452 
453 	/* We're back to the default cookie, unthrottle all the things. */
454 	if (cur_cookie == DEFAULT_SCALE_COOKIE) {
455 		blkcg_clear_delay(lat_to_blkg(iolat));
456 		iolat->rq_depth.max_depth = UINT_MAX;
457 		wake_up_all(&iolat->rq_wait.wait);
458 		return;
459 	}
460 
461 	scale_change(iolat, direction > 0);
462 }
463 
464 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
465 {
466 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
467 	struct blkcg_gq *blkg = bio->bi_blkg;
468 	bool issue_as_root = bio_issue_as_root_blkg(bio);
469 
470 	if (!blkiolat->enabled)
471 		return;
472 
473 	while (blkg && blkg->parent) {
474 		struct iolatency_grp *iolat = blkg_to_lat(blkg);
475 		if (!iolat) {
476 			blkg = blkg->parent;
477 			continue;
478 		}
479 
480 		check_scale_change(iolat);
481 		__blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
482 				     (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
483 		blkg = blkg->parent;
484 	}
485 	if (!timer_pending(&blkiolat->timer))
486 		mod_timer(&blkiolat->timer, jiffies + HZ);
487 }
488 
489 static void iolatency_record_time(struct iolatency_grp *iolat,
490 				  struct bio_issue *issue, u64 now,
491 				  bool issue_as_root)
492 {
493 	u64 start = bio_issue_time(issue);
494 	u64 req_time;
495 
496 	/*
497 	 * Have to do this so we are truncated to the correct time that our
498 	 * issue is truncated to.
499 	 */
500 	now = __bio_issue_time(now);
501 
502 	if (now <= start)
503 		return;
504 
505 	req_time = now - start;
506 
507 	/*
508 	 * We don't want to count issue_as_root bio's in the cgroups latency
509 	 * statistics as it could skew the numbers downwards.
510 	 */
511 	if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
512 		u64 sub = iolat->min_lat_nsec;
513 		if (req_time < sub)
514 			blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
515 		return;
516 	}
517 
518 	latency_stat_record_time(iolat, req_time);
519 }
520 
521 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
522 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
523 
524 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
525 {
526 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
527 	struct iolatency_grp *parent;
528 	struct child_latency_info *lat_info;
529 	struct latency_stat stat;
530 	unsigned long flags;
531 	int cpu;
532 
533 	latency_stat_init(iolat, &stat);
534 	preempt_disable();
535 	for_each_online_cpu(cpu) {
536 		struct latency_stat *s;
537 		s = per_cpu_ptr(iolat->stats, cpu);
538 		latency_stat_sum(iolat, &stat, s);
539 		latency_stat_init(iolat, s);
540 	}
541 	preempt_enable();
542 
543 	parent = blkg_to_lat(blkg->parent);
544 	if (!parent)
545 		return;
546 
547 	lat_info = &parent->child_lat;
548 
549 	iolat_update_total_lat_avg(iolat, &stat);
550 
551 	/* Everything is ok and we don't need to adjust the scale. */
552 	if (latency_sum_ok(iolat, &stat) &&
553 	    atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
554 		return;
555 
556 	/* Somebody beat us to the punch, just bail. */
557 	spin_lock_irqsave(&lat_info->lock, flags);
558 
559 	latency_stat_sum(iolat, &iolat->cur_stat, &stat);
560 	lat_info->nr_samples -= iolat->nr_samples;
561 	lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
562 	iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
563 
564 	if ((lat_info->last_scale_event >= now ||
565 	    now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
566 		goto out;
567 
568 	if (latency_sum_ok(iolat, &iolat->cur_stat) &&
569 	    latency_sum_ok(iolat, &stat)) {
570 		if (latency_stat_samples(iolat, &iolat->cur_stat) <
571 		    BLKIOLATENCY_MIN_GOOD_SAMPLES)
572 			goto out;
573 		if (lat_info->scale_grp == iolat) {
574 			lat_info->last_scale_event = now;
575 			scale_cookie_change(iolat->blkiolat, lat_info, true);
576 		}
577 	} else if (lat_info->scale_lat == 0 ||
578 		   lat_info->scale_lat >= iolat->min_lat_nsec) {
579 		lat_info->last_scale_event = now;
580 		if (!lat_info->scale_grp ||
581 		    lat_info->scale_lat > iolat->min_lat_nsec) {
582 			WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
583 			lat_info->scale_grp = iolat;
584 		}
585 		scale_cookie_change(iolat->blkiolat, lat_info, false);
586 	}
587 	latency_stat_init(iolat, &iolat->cur_stat);
588 out:
589 	spin_unlock_irqrestore(&lat_info->lock, flags);
590 }
591 
592 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
593 {
594 	struct blkcg_gq *blkg;
595 	struct rq_wait *rqw;
596 	struct iolatency_grp *iolat;
597 	u64 window_start;
598 	u64 now;
599 	bool issue_as_root = bio_issue_as_root_blkg(bio);
600 	int inflight = 0;
601 
602 	blkg = bio->bi_blkg;
603 	if (!blkg || !bio_flagged(bio, BIO_QOS_THROTTLED))
604 		return;
605 
606 	iolat = blkg_to_lat(bio->bi_blkg);
607 	if (!iolat)
608 		return;
609 
610 	if (!iolat->blkiolat->enabled)
611 		return;
612 
613 	now = ktime_to_ns(ktime_get());
614 	while (blkg && blkg->parent) {
615 		iolat = blkg_to_lat(blkg);
616 		if (!iolat) {
617 			blkg = blkg->parent;
618 			continue;
619 		}
620 		rqw = &iolat->rq_wait;
621 
622 		inflight = atomic_dec_return(&rqw->inflight);
623 		WARN_ON_ONCE(inflight < 0);
624 		/*
625 		 * If bi_status is BLK_STS_AGAIN, the bio wasn't actually
626 		 * submitted, so do not account for it.
627 		 */
628 		if (iolat->min_lat_nsec && bio->bi_status != BLK_STS_AGAIN) {
629 			iolatency_record_time(iolat, &bio->bi_issue, now,
630 					      issue_as_root);
631 			window_start = atomic64_read(&iolat->window_start);
632 			if (now > window_start &&
633 			    (now - window_start) >= iolat->cur_win_nsec) {
634 				if (atomic64_try_cmpxchg(&iolat->window_start,
635 							 &window_start, now))
636 					iolatency_check_latencies(iolat, now);
637 			}
638 		}
639 		wake_up(&rqw->wait);
640 		blkg = blkg->parent;
641 	}
642 }
643 
644 static void blkcg_iolatency_exit(struct rq_qos *rqos)
645 {
646 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
647 
648 	del_timer_sync(&blkiolat->timer);
649 	flush_work(&blkiolat->enable_work);
650 	blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
651 	kfree(blkiolat);
652 }
653 
654 static struct rq_qos_ops blkcg_iolatency_ops = {
655 	.throttle = blkcg_iolatency_throttle,
656 	.done_bio = blkcg_iolatency_done_bio,
657 	.exit = blkcg_iolatency_exit,
658 };
659 
660 static void blkiolatency_timer_fn(struct timer_list *t)
661 {
662 	struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
663 	struct blkcg_gq *blkg;
664 	struct cgroup_subsys_state *pos_css;
665 	u64 now = ktime_to_ns(ktime_get());
666 
667 	rcu_read_lock();
668 	blkg_for_each_descendant_pre(blkg, pos_css,
669 				     blkiolat->rqos.q->root_blkg) {
670 		struct iolatency_grp *iolat;
671 		struct child_latency_info *lat_info;
672 		unsigned long flags;
673 		u64 cookie;
674 
675 		/*
676 		 * We could be exiting, don't access the pd unless we have a
677 		 * ref on the blkg.
678 		 */
679 		if (!blkg_tryget(blkg))
680 			continue;
681 
682 		iolat = blkg_to_lat(blkg);
683 		if (!iolat)
684 			goto next;
685 
686 		lat_info = &iolat->child_lat;
687 		cookie = atomic_read(&lat_info->scale_cookie);
688 
689 		if (cookie >= DEFAULT_SCALE_COOKIE)
690 			goto next;
691 
692 		spin_lock_irqsave(&lat_info->lock, flags);
693 		if (lat_info->last_scale_event >= now)
694 			goto next_lock;
695 
696 		/*
697 		 * We scaled down but don't have a scale_grp, scale up and carry
698 		 * on.
699 		 */
700 		if (lat_info->scale_grp == NULL) {
701 			scale_cookie_change(iolat->blkiolat, lat_info, true);
702 			goto next_lock;
703 		}
704 
705 		/*
706 		 * It's been 5 seconds since our last scale event, clear the
707 		 * scale grp in case the group that needed the scale down isn't
708 		 * doing any IO currently.
709 		 */
710 		if (now - lat_info->last_scale_event >=
711 		    ((u64)NSEC_PER_SEC * 5))
712 			lat_info->scale_grp = NULL;
713 next_lock:
714 		spin_unlock_irqrestore(&lat_info->lock, flags);
715 next:
716 		blkg_put(blkg);
717 	}
718 	rcu_read_unlock();
719 }
720 
721 /**
722  * blkiolatency_enable_work_fn - Enable or disable iolatency on the device
723  * @work: enable_work of the blk_iolatency of interest
724  *
725  * iolatency needs to keep track of the number of in-flight IOs per cgroup. This
726  * is relatively expensive as it involves walking up the hierarchy twice for
727  * every IO. Thus, if iolatency is not enabled in any cgroup for the device, we
728  * want to disable the in-flight tracking.
729  *
730  * We have to make sure that the counting is balanced - we don't want to leak
731  * the in-flight counts by disabling accounting in the completion path while IOs
732  * are in flight. This is achieved by ensuring that no IO is in flight by
733  * freezing the queue while flipping ->enabled. As this requires a sleepable
734  * context, ->enabled flipping is punted to this work function.
735  */
736 static void blkiolatency_enable_work_fn(struct work_struct *work)
737 {
738 	struct blk_iolatency *blkiolat = container_of(work, struct blk_iolatency,
739 						      enable_work);
740 	bool enabled;
741 
742 	/*
743 	 * There can only be one instance of this function running for @blkiolat
744 	 * and it's guaranteed to be executed at least once after the latest
745 	 * ->enabled_cnt modification. Acting on the latest ->enable_cnt is
746 	 * sufficient.
747 	 *
748 	 * Also, we know @blkiolat is safe to access as ->enable_work is flushed
749 	 * in blkcg_iolatency_exit().
750 	 */
751 	enabled = atomic_read(&blkiolat->enable_cnt);
752 	if (enabled != blkiolat->enabled) {
753 		blk_mq_freeze_queue(blkiolat->rqos.q);
754 		blkiolat->enabled = enabled;
755 		blk_mq_unfreeze_queue(blkiolat->rqos.q);
756 	}
757 }
758 
759 int blk_iolatency_init(struct gendisk *disk)
760 {
761 	struct request_queue *q = disk->queue;
762 	struct blk_iolatency *blkiolat;
763 	struct rq_qos *rqos;
764 	int ret;
765 
766 	blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
767 	if (!blkiolat)
768 		return -ENOMEM;
769 
770 	rqos = &blkiolat->rqos;
771 	rqos->id = RQ_QOS_LATENCY;
772 	rqos->ops = &blkcg_iolatency_ops;
773 	rqos->q = q;
774 
775 	ret = rq_qos_add(q, rqos);
776 	if (ret)
777 		goto err_free;
778 	ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
779 	if (ret)
780 		goto err_qos_del;
781 
782 	timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
783 	INIT_WORK(&blkiolat->enable_work, blkiolatency_enable_work_fn);
784 
785 	return 0;
786 
787 err_qos_del:
788 	rq_qos_del(q, rqos);
789 err_free:
790 	kfree(blkiolat);
791 	return ret;
792 }
793 
794 static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
795 {
796 	struct iolatency_grp *iolat = blkg_to_lat(blkg);
797 	struct blk_iolatency *blkiolat = iolat->blkiolat;
798 	u64 oldval = iolat->min_lat_nsec;
799 
800 	iolat->min_lat_nsec = val;
801 	iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
802 	iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
803 				    BLKIOLATENCY_MAX_WIN_SIZE);
804 
805 	if (!oldval && val) {
806 		if (atomic_inc_return(&blkiolat->enable_cnt) == 1)
807 			schedule_work(&blkiolat->enable_work);
808 	}
809 	if (oldval && !val) {
810 		blkcg_clear_delay(blkg);
811 		if (atomic_dec_return(&blkiolat->enable_cnt) == 0)
812 			schedule_work(&blkiolat->enable_work);
813 	}
814 }
815 
816 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
817 {
818 	if (blkg->parent) {
819 		struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
820 		struct child_latency_info *lat_info;
821 		if (!iolat)
822 			return;
823 
824 		lat_info = &iolat->child_lat;
825 		spin_lock(&lat_info->lock);
826 		atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
827 		lat_info->last_scale_event = 0;
828 		lat_info->scale_grp = NULL;
829 		lat_info->scale_lat = 0;
830 		spin_unlock(&lat_info->lock);
831 	}
832 }
833 
834 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
835 			     size_t nbytes, loff_t off)
836 {
837 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
838 	struct blkcg_gq *blkg;
839 	struct blkg_conf_ctx ctx;
840 	struct iolatency_grp *iolat;
841 	char *p, *tok;
842 	u64 lat_val = 0;
843 	u64 oldval;
844 	int ret;
845 
846 	ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
847 	if (ret)
848 		return ret;
849 
850 	iolat = blkg_to_lat(ctx.blkg);
851 	p = ctx.body;
852 
853 	ret = -EINVAL;
854 	while ((tok = strsep(&p, " "))) {
855 		char key[16];
856 		char val[21];	/* 18446744073709551616 */
857 
858 		if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
859 			goto out;
860 
861 		if (!strcmp(key, "target")) {
862 			u64 v;
863 
864 			if (!strcmp(val, "max"))
865 				lat_val = 0;
866 			else if (sscanf(val, "%llu", &v) == 1)
867 				lat_val = v * NSEC_PER_USEC;
868 			else
869 				goto out;
870 		} else {
871 			goto out;
872 		}
873 	}
874 
875 	/* Walk up the tree to see if our new val is lower than it should be. */
876 	blkg = ctx.blkg;
877 	oldval = iolat->min_lat_nsec;
878 
879 	iolatency_set_min_lat_nsec(blkg, lat_val);
880 	if (oldval != iolat->min_lat_nsec)
881 		iolatency_clear_scaling(blkg);
882 	ret = 0;
883 out:
884 	blkg_conf_finish(&ctx);
885 	return ret ?: nbytes;
886 }
887 
888 static u64 iolatency_prfill_limit(struct seq_file *sf,
889 				  struct blkg_policy_data *pd, int off)
890 {
891 	struct iolatency_grp *iolat = pd_to_lat(pd);
892 	const char *dname = blkg_dev_name(pd->blkg);
893 
894 	if (!dname || !iolat->min_lat_nsec)
895 		return 0;
896 	seq_printf(sf, "%s target=%llu\n",
897 		   dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
898 	return 0;
899 }
900 
901 static int iolatency_print_limit(struct seq_file *sf, void *v)
902 {
903 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
904 			  iolatency_prfill_limit,
905 			  &blkcg_policy_iolatency, seq_cft(sf)->private, false);
906 	return 0;
907 }
908 
909 static void iolatency_ssd_stat(struct iolatency_grp *iolat, struct seq_file *s)
910 {
911 	struct latency_stat stat;
912 	int cpu;
913 
914 	latency_stat_init(iolat, &stat);
915 	preempt_disable();
916 	for_each_online_cpu(cpu) {
917 		struct latency_stat *s;
918 		s = per_cpu_ptr(iolat->stats, cpu);
919 		latency_stat_sum(iolat, &stat, s);
920 	}
921 	preempt_enable();
922 
923 	if (iolat->rq_depth.max_depth == UINT_MAX)
924 		seq_printf(s, " missed=%llu total=%llu depth=max",
925 			(unsigned long long)stat.ps.missed,
926 			(unsigned long long)stat.ps.total);
927 	else
928 		seq_printf(s, " missed=%llu total=%llu depth=%u",
929 			(unsigned long long)stat.ps.missed,
930 			(unsigned long long)stat.ps.total,
931 			iolat->rq_depth.max_depth);
932 }
933 
934 static void iolatency_pd_stat(struct blkg_policy_data *pd, struct seq_file *s)
935 {
936 	struct iolatency_grp *iolat = pd_to_lat(pd);
937 	unsigned long long avg_lat;
938 	unsigned long long cur_win;
939 
940 	if (!blkcg_debug_stats)
941 		return;
942 
943 	if (iolat->ssd)
944 		return iolatency_ssd_stat(iolat, s);
945 
946 	avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
947 	cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
948 	if (iolat->rq_depth.max_depth == UINT_MAX)
949 		seq_printf(s, " depth=max avg_lat=%llu win=%llu",
950 			avg_lat, cur_win);
951 	else
952 		seq_printf(s, " depth=%u avg_lat=%llu win=%llu",
953 			iolat->rq_depth.max_depth, avg_lat, cur_win);
954 }
955 
956 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp,
957 						   struct request_queue *q,
958 						   struct blkcg *blkcg)
959 {
960 	struct iolatency_grp *iolat;
961 
962 	iolat = kzalloc_node(sizeof(*iolat), gfp, q->node);
963 	if (!iolat)
964 		return NULL;
965 	iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
966 				       __alignof__(struct latency_stat), gfp);
967 	if (!iolat->stats) {
968 		kfree(iolat);
969 		return NULL;
970 	}
971 	return &iolat->pd;
972 }
973 
974 static void iolatency_pd_init(struct blkg_policy_data *pd)
975 {
976 	struct iolatency_grp *iolat = pd_to_lat(pd);
977 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
978 	struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
979 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
980 	u64 now = ktime_to_ns(ktime_get());
981 	int cpu;
982 
983 	if (blk_queue_nonrot(blkg->q))
984 		iolat->ssd = true;
985 	else
986 		iolat->ssd = false;
987 
988 	for_each_possible_cpu(cpu) {
989 		struct latency_stat *stat;
990 		stat = per_cpu_ptr(iolat->stats, cpu);
991 		latency_stat_init(iolat, stat);
992 	}
993 
994 	latency_stat_init(iolat, &iolat->cur_stat);
995 	rq_wait_init(&iolat->rq_wait);
996 	spin_lock_init(&iolat->child_lat.lock);
997 	iolat->rq_depth.queue_depth = blkg->q->nr_requests;
998 	iolat->rq_depth.max_depth = UINT_MAX;
999 	iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
1000 	iolat->blkiolat = blkiolat;
1001 	iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
1002 	atomic64_set(&iolat->window_start, now);
1003 
1004 	/*
1005 	 * We init things in list order, so the pd for the parent may not be
1006 	 * init'ed yet for whatever reason.
1007 	 */
1008 	if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
1009 		struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
1010 		atomic_set(&iolat->scale_cookie,
1011 			   atomic_read(&parent->child_lat.scale_cookie));
1012 	} else {
1013 		atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
1014 	}
1015 
1016 	atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
1017 }
1018 
1019 static void iolatency_pd_offline(struct blkg_policy_data *pd)
1020 {
1021 	struct iolatency_grp *iolat = pd_to_lat(pd);
1022 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
1023 
1024 	iolatency_set_min_lat_nsec(blkg, 0);
1025 	iolatency_clear_scaling(blkg);
1026 }
1027 
1028 static void iolatency_pd_free(struct blkg_policy_data *pd)
1029 {
1030 	struct iolatency_grp *iolat = pd_to_lat(pd);
1031 	free_percpu(iolat->stats);
1032 	kfree(iolat);
1033 }
1034 
1035 static struct cftype iolatency_files[] = {
1036 	{
1037 		.name = "latency",
1038 		.flags = CFTYPE_NOT_ON_ROOT,
1039 		.seq_show = iolatency_print_limit,
1040 		.write = iolatency_set_limit,
1041 	},
1042 	{}
1043 };
1044 
1045 static struct blkcg_policy blkcg_policy_iolatency = {
1046 	.dfl_cftypes	= iolatency_files,
1047 	.pd_alloc_fn	= iolatency_pd_alloc,
1048 	.pd_init_fn	= iolatency_pd_init,
1049 	.pd_offline_fn	= iolatency_pd_offline,
1050 	.pd_free_fn	= iolatency_pd_free,
1051 	.pd_stat_fn	= iolatency_pd_stat,
1052 };
1053 
1054 static int __init iolatency_init(void)
1055 {
1056 	return blkcg_policy_register(&blkcg_policy_iolatency);
1057 }
1058 
1059 static void __exit iolatency_exit(void)
1060 {
1061 	blkcg_policy_unregister(&blkcg_policy_iolatency);
1062 }
1063 
1064 module_init(iolatency_init);
1065 module_exit(iolatency_exit);
1066