xref: /openbmc/linux/block/blk-iolatency.c (revision 2cc39179)
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, 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 	unsigned int old;
405 	int direction = 0;
406 
407 	if (lat_to_blkg(iolat)->parent == NULL)
408 		return;
409 
410 	parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
411 	if (!parent)
412 		return;
413 
414 	lat_info = &parent->child_lat;
415 	cur_cookie = atomic_read(&lat_info->scale_cookie);
416 	scale_lat = READ_ONCE(lat_info->scale_lat);
417 
418 	if (cur_cookie < our_cookie)
419 		direction = -1;
420 	else if (cur_cookie > our_cookie)
421 		direction = 1;
422 	else
423 		return;
424 
425 	old = atomic_cmpxchg(&iolat->scale_cookie, our_cookie, cur_cookie);
426 
427 	/* Somebody beat us to the punch, just bail. */
428 	if (old != our_cookie)
429 		return;
430 
431 	if (direction < 0 && iolat->min_lat_nsec) {
432 		u64 samples_thresh;
433 
434 		if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
435 			return;
436 
437 		/*
438 		 * Sometimes high priority groups are their own worst enemy, so
439 		 * instead of taking it out on some poor other group that did 5%
440 		 * or less of the IO's for the last summation just skip this
441 		 * scale down event.
442 		 */
443 		samples_thresh = lat_info->nr_samples * 5;
444 		samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
445 		if (iolat->nr_samples <= samples_thresh)
446 			return;
447 	}
448 
449 	/* We're as low as we can go. */
450 	if (iolat->rq_depth.max_depth == 1 && direction < 0) {
451 		blkcg_use_delay(lat_to_blkg(iolat));
452 		return;
453 	}
454 
455 	/* We're back to the default cookie, unthrottle all the things. */
456 	if (cur_cookie == DEFAULT_SCALE_COOKIE) {
457 		blkcg_clear_delay(lat_to_blkg(iolat));
458 		iolat->rq_depth.max_depth = UINT_MAX;
459 		wake_up_all(&iolat->rq_wait.wait);
460 		return;
461 	}
462 
463 	scale_change(iolat, direction > 0);
464 }
465 
466 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
467 {
468 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
469 	struct blkcg_gq *blkg = bio->bi_blkg;
470 	bool issue_as_root = bio_issue_as_root_blkg(bio);
471 
472 	if (!blkiolat->enabled)
473 		return;
474 
475 	while (blkg && blkg->parent) {
476 		struct iolatency_grp *iolat = blkg_to_lat(blkg);
477 		if (!iolat) {
478 			blkg = blkg->parent;
479 			continue;
480 		}
481 
482 		check_scale_change(iolat);
483 		__blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
484 				     (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
485 		blkg = blkg->parent;
486 	}
487 	if (!timer_pending(&blkiolat->timer))
488 		mod_timer(&blkiolat->timer, jiffies + HZ);
489 }
490 
491 static void iolatency_record_time(struct iolatency_grp *iolat,
492 				  struct bio_issue *issue, u64 now,
493 				  bool issue_as_root)
494 {
495 	u64 start = bio_issue_time(issue);
496 	u64 req_time;
497 
498 	/*
499 	 * Have to do this so we are truncated to the correct time that our
500 	 * issue is truncated to.
501 	 */
502 	now = __bio_issue_time(now);
503 
504 	if (now <= start)
505 		return;
506 
507 	req_time = now - start;
508 
509 	/*
510 	 * We don't want to count issue_as_root bio's in the cgroups latency
511 	 * statistics as it could skew the numbers downwards.
512 	 */
513 	if (unlikely(issue_as_root && iolat->rq_depth.max_depth != UINT_MAX)) {
514 		u64 sub = iolat->min_lat_nsec;
515 		if (req_time < sub)
516 			blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
517 		return;
518 	}
519 
520 	latency_stat_record_time(iolat, req_time);
521 }
522 
523 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
524 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
525 
526 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
527 {
528 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
529 	struct iolatency_grp *parent;
530 	struct child_latency_info *lat_info;
531 	struct latency_stat stat;
532 	unsigned long flags;
533 	int cpu;
534 
535 	latency_stat_init(iolat, &stat);
536 	preempt_disable();
537 	for_each_online_cpu(cpu) {
538 		struct latency_stat *s;
539 		s = per_cpu_ptr(iolat->stats, cpu);
540 		latency_stat_sum(iolat, &stat, s);
541 		latency_stat_init(iolat, s);
542 	}
543 	preempt_enable();
544 
545 	parent = blkg_to_lat(blkg->parent);
546 	if (!parent)
547 		return;
548 
549 	lat_info = &parent->child_lat;
550 
551 	iolat_update_total_lat_avg(iolat, &stat);
552 
553 	/* Everything is ok and we don't need to adjust the scale. */
554 	if (latency_sum_ok(iolat, &stat) &&
555 	    atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
556 		return;
557 
558 	/* Somebody beat us to the punch, just bail. */
559 	spin_lock_irqsave(&lat_info->lock, flags);
560 
561 	latency_stat_sum(iolat, &iolat->cur_stat, &stat);
562 	lat_info->nr_samples -= iolat->nr_samples;
563 	lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
564 	iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
565 
566 	if ((lat_info->last_scale_event >= now ||
567 	    now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
568 		goto out;
569 
570 	if (latency_sum_ok(iolat, &iolat->cur_stat) &&
571 	    latency_sum_ok(iolat, &stat)) {
572 		if (latency_stat_samples(iolat, &iolat->cur_stat) <
573 		    BLKIOLATENCY_MIN_GOOD_SAMPLES)
574 			goto out;
575 		if (lat_info->scale_grp == iolat) {
576 			lat_info->last_scale_event = now;
577 			scale_cookie_change(iolat->blkiolat, lat_info, true);
578 		}
579 	} else if (lat_info->scale_lat == 0 ||
580 		   lat_info->scale_lat >= iolat->min_lat_nsec) {
581 		lat_info->last_scale_event = now;
582 		if (!lat_info->scale_grp ||
583 		    lat_info->scale_lat > iolat->min_lat_nsec) {
584 			WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
585 			lat_info->scale_grp = iolat;
586 		}
587 		scale_cookie_change(iolat->blkiolat, lat_info, false);
588 	}
589 	latency_stat_init(iolat, &iolat->cur_stat);
590 out:
591 	spin_unlock_irqrestore(&lat_info->lock, flags);
592 }
593 
594 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
595 {
596 	struct blkcg_gq *blkg;
597 	struct rq_wait *rqw;
598 	struct iolatency_grp *iolat;
599 	u64 window_start;
600 	u64 now;
601 	bool issue_as_root = bio_issue_as_root_blkg(bio);
602 	int inflight = 0;
603 
604 	blkg = bio->bi_blkg;
605 	if (!blkg || !bio_flagged(bio, BIO_QOS_THROTTLED))
606 		return;
607 
608 	iolat = blkg_to_lat(bio->bi_blkg);
609 	if (!iolat)
610 		return;
611 
612 	if (!iolat->blkiolat->enabled)
613 		return;
614 
615 	now = ktime_to_ns(ktime_get());
616 	while (blkg && blkg->parent) {
617 		iolat = blkg_to_lat(blkg);
618 		if (!iolat) {
619 			blkg = blkg->parent;
620 			continue;
621 		}
622 		rqw = &iolat->rq_wait;
623 
624 		inflight = atomic_dec_return(&rqw->inflight);
625 		WARN_ON_ONCE(inflight < 0);
626 		/*
627 		 * If bi_status is BLK_STS_AGAIN, the bio wasn't actually
628 		 * submitted, so do not account for it.
629 		 */
630 		if (iolat->min_lat_nsec && bio->bi_status != BLK_STS_AGAIN) {
631 			iolatency_record_time(iolat, &bio->bi_issue, now,
632 					      issue_as_root);
633 			window_start = atomic64_read(&iolat->window_start);
634 			if (now > window_start &&
635 			    (now - window_start) >= iolat->cur_win_nsec) {
636 				if (atomic64_cmpxchg(&iolat->window_start,
637 					     window_start, now) == window_start)
638 					iolatency_check_latencies(iolat, now);
639 			}
640 		}
641 		wake_up(&rqw->wait);
642 		blkg = blkg->parent;
643 	}
644 }
645 
646 static void blkcg_iolatency_exit(struct rq_qos *rqos)
647 {
648 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
649 
650 	del_timer_sync(&blkiolat->timer);
651 	flush_work(&blkiolat->enable_work);
652 	blkcg_deactivate_policy(rqos->q, &blkcg_policy_iolatency);
653 	kfree(blkiolat);
654 }
655 
656 static struct rq_qos_ops blkcg_iolatency_ops = {
657 	.throttle = blkcg_iolatency_throttle,
658 	.done_bio = blkcg_iolatency_done_bio,
659 	.exit = blkcg_iolatency_exit,
660 };
661 
662 static void blkiolatency_timer_fn(struct timer_list *t)
663 {
664 	struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
665 	struct blkcg_gq *blkg;
666 	struct cgroup_subsys_state *pos_css;
667 	u64 now = ktime_to_ns(ktime_get());
668 
669 	rcu_read_lock();
670 	blkg_for_each_descendant_pre(blkg, pos_css,
671 				     blkiolat->rqos.q->root_blkg) {
672 		struct iolatency_grp *iolat;
673 		struct child_latency_info *lat_info;
674 		unsigned long flags;
675 		u64 cookie;
676 
677 		/*
678 		 * We could be exiting, don't access the pd unless we have a
679 		 * ref on the blkg.
680 		 */
681 		if (!blkg_tryget(blkg))
682 			continue;
683 
684 		iolat = blkg_to_lat(blkg);
685 		if (!iolat)
686 			goto next;
687 
688 		lat_info = &iolat->child_lat;
689 		cookie = atomic_read(&lat_info->scale_cookie);
690 
691 		if (cookie >= DEFAULT_SCALE_COOKIE)
692 			goto next;
693 
694 		spin_lock_irqsave(&lat_info->lock, flags);
695 		if (lat_info->last_scale_event >= now)
696 			goto next_lock;
697 
698 		/*
699 		 * We scaled down but don't have a scale_grp, scale up and carry
700 		 * on.
701 		 */
702 		if (lat_info->scale_grp == NULL) {
703 			scale_cookie_change(iolat->blkiolat, lat_info, true);
704 			goto next_lock;
705 		}
706 
707 		/*
708 		 * It's been 5 seconds since our last scale event, clear the
709 		 * scale grp in case the group that needed the scale down isn't
710 		 * doing any IO currently.
711 		 */
712 		if (now - lat_info->last_scale_event >=
713 		    ((u64)NSEC_PER_SEC * 5))
714 			lat_info->scale_grp = NULL;
715 next_lock:
716 		spin_unlock_irqrestore(&lat_info->lock, flags);
717 next:
718 		blkg_put(blkg);
719 	}
720 	rcu_read_unlock();
721 }
722 
723 /**
724  * blkiolatency_enable_work_fn - Enable or disable iolatency on the device
725  * @work: enable_work of the blk_iolatency of interest
726  *
727  * iolatency needs to keep track of the number of in-flight IOs per cgroup. This
728  * is relatively expensive as it involves walking up the hierarchy twice for
729  * every IO. Thus, if iolatency is not enabled in any cgroup for the device, we
730  * want to disable the in-flight tracking.
731  *
732  * We have to make sure that the counting is balanced - we don't want to leak
733  * the in-flight counts by disabling accounting in the completion path while IOs
734  * are in flight. This is achieved by ensuring that no IO is in flight by
735  * freezing the queue while flipping ->enabled. As this requires a sleepable
736  * context, ->enabled flipping is punted to this work function.
737  */
738 static void blkiolatency_enable_work_fn(struct work_struct *work)
739 {
740 	struct blk_iolatency *blkiolat = container_of(work, struct blk_iolatency,
741 						      enable_work);
742 	bool enabled;
743 
744 	/*
745 	 * There can only be one instance of this function running for @blkiolat
746 	 * and it's guaranteed to be executed at least once after the latest
747 	 * ->enabled_cnt modification. Acting on the latest ->enable_cnt is
748 	 * sufficient.
749 	 *
750 	 * Also, we know @blkiolat is safe to access as ->enable_work is flushed
751 	 * in blkcg_iolatency_exit().
752 	 */
753 	enabled = atomic_read(&blkiolat->enable_cnt);
754 	if (enabled != blkiolat->enabled) {
755 		blk_mq_freeze_queue(blkiolat->rqos.q);
756 		blkiolat->enabled = enabled;
757 		blk_mq_unfreeze_queue(blkiolat->rqos.q);
758 	}
759 }
760 
761 int blk_iolatency_init(struct request_queue *q)
762 {
763 	struct blk_iolatency *blkiolat;
764 	struct rq_qos *rqos;
765 	int ret;
766 
767 	blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
768 	if (!blkiolat)
769 		return -ENOMEM;
770 
771 	rqos = &blkiolat->rqos;
772 	rqos->id = RQ_QOS_LATENCY;
773 	rqos->ops = &blkcg_iolatency_ops;
774 	rqos->q = q;
775 
776 	rq_qos_add(q, rqos);
777 
778 	ret = blkcg_activate_policy(q, &blkcg_policy_iolatency);
779 	if (ret) {
780 		rq_qos_del(q, rqos);
781 		kfree(blkiolat);
782 		return ret;
783 	}
784 
785 	timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
786 	INIT_WORK(&blkiolat->enable_work, blkiolatency_enable_work_fn);
787 
788 	return 0;
789 }
790 
791 static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
792 {
793 	struct iolatency_grp *iolat = blkg_to_lat(blkg);
794 	struct blk_iolatency *blkiolat = iolat->blkiolat;
795 	u64 oldval = iolat->min_lat_nsec;
796 
797 	iolat->min_lat_nsec = val;
798 	iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
799 	iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
800 				    BLKIOLATENCY_MAX_WIN_SIZE);
801 
802 	if (!oldval && val) {
803 		if (atomic_inc_return(&blkiolat->enable_cnt) == 1)
804 			schedule_work(&blkiolat->enable_work);
805 	}
806 	if (oldval && !val) {
807 		blkcg_clear_delay(blkg);
808 		if (atomic_dec_return(&blkiolat->enable_cnt) == 0)
809 			schedule_work(&blkiolat->enable_work);
810 	}
811 }
812 
813 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
814 {
815 	if (blkg->parent) {
816 		struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
817 		struct child_latency_info *lat_info;
818 		if (!iolat)
819 			return;
820 
821 		lat_info = &iolat->child_lat;
822 		spin_lock(&lat_info->lock);
823 		atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
824 		lat_info->last_scale_event = 0;
825 		lat_info->scale_grp = NULL;
826 		lat_info->scale_lat = 0;
827 		spin_unlock(&lat_info->lock);
828 	}
829 }
830 
831 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
832 			     size_t nbytes, loff_t off)
833 {
834 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
835 	struct blkcg_gq *blkg;
836 	struct blkg_conf_ctx ctx;
837 	struct iolatency_grp *iolat;
838 	char *p, *tok;
839 	u64 lat_val = 0;
840 	u64 oldval;
841 	int ret;
842 
843 	ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, buf, &ctx);
844 	if (ret)
845 		return ret;
846 
847 	iolat = blkg_to_lat(ctx.blkg);
848 	p = ctx.body;
849 
850 	ret = -EINVAL;
851 	while ((tok = strsep(&p, " "))) {
852 		char key[16];
853 		char val[21];	/* 18446744073709551616 */
854 
855 		if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
856 			goto out;
857 
858 		if (!strcmp(key, "target")) {
859 			u64 v;
860 
861 			if (!strcmp(val, "max"))
862 				lat_val = 0;
863 			else if (sscanf(val, "%llu", &v) == 1)
864 				lat_val = v * NSEC_PER_USEC;
865 			else
866 				goto out;
867 		} else {
868 			goto out;
869 		}
870 	}
871 
872 	/* Walk up the tree to see if our new val is lower than it should be. */
873 	blkg = ctx.blkg;
874 	oldval = iolat->min_lat_nsec;
875 
876 	iolatency_set_min_lat_nsec(blkg, lat_val);
877 	if (oldval != iolat->min_lat_nsec)
878 		iolatency_clear_scaling(blkg);
879 	ret = 0;
880 out:
881 	blkg_conf_finish(&ctx);
882 	return ret ?: nbytes;
883 }
884 
885 static u64 iolatency_prfill_limit(struct seq_file *sf,
886 				  struct blkg_policy_data *pd, int off)
887 {
888 	struct iolatency_grp *iolat = pd_to_lat(pd);
889 	const char *dname = blkg_dev_name(pd->blkg);
890 
891 	if (!dname || !iolat->min_lat_nsec)
892 		return 0;
893 	seq_printf(sf, "%s target=%llu\n",
894 		   dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
895 	return 0;
896 }
897 
898 static int iolatency_print_limit(struct seq_file *sf, void *v)
899 {
900 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
901 			  iolatency_prfill_limit,
902 			  &blkcg_policy_iolatency, seq_cft(sf)->private, false);
903 	return 0;
904 }
905 
906 static void iolatency_ssd_stat(struct iolatency_grp *iolat, struct seq_file *s)
907 {
908 	struct latency_stat stat;
909 	int cpu;
910 
911 	latency_stat_init(iolat, &stat);
912 	preempt_disable();
913 	for_each_online_cpu(cpu) {
914 		struct latency_stat *s;
915 		s = per_cpu_ptr(iolat->stats, cpu);
916 		latency_stat_sum(iolat, &stat, s);
917 	}
918 	preempt_enable();
919 
920 	if (iolat->rq_depth.max_depth == UINT_MAX)
921 		seq_printf(s, " missed=%llu total=%llu depth=max",
922 			(unsigned long long)stat.ps.missed,
923 			(unsigned long long)stat.ps.total);
924 	else
925 		seq_printf(s, " missed=%llu total=%llu depth=%u",
926 			(unsigned long long)stat.ps.missed,
927 			(unsigned long long)stat.ps.total,
928 			iolat->rq_depth.max_depth);
929 }
930 
931 static void iolatency_pd_stat(struct blkg_policy_data *pd, struct seq_file *s)
932 {
933 	struct iolatency_grp *iolat = pd_to_lat(pd);
934 	unsigned long long avg_lat;
935 	unsigned long long cur_win;
936 
937 	if (!blkcg_debug_stats)
938 		return;
939 
940 	if (iolat->ssd)
941 		return iolatency_ssd_stat(iolat, s);
942 
943 	avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
944 	cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
945 	if (iolat->rq_depth.max_depth == UINT_MAX)
946 		seq_printf(s, " depth=max avg_lat=%llu win=%llu",
947 			avg_lat, cur_win);
948 	else
949 		seq_printf(s, " depth=%u avg_lat=%llu win=%llu",
950 			iolat->rq_depth.max_depth, avg_lat, cur_win);
951 }
952 
953 static struct blkg_policy_data *iolatency_pd_alloc(gfp_t gfp,
954 						   struct request_queue *q,
955 						   struct blkcg *blkcg)
956 {
957 	struct iolatency_grp *iolat;
958 
959 	iolat = kzalloc_node(sizeof(*iolat), gfp, q->node);
960 	if (!iolat)
961 		return NULL;
962 	iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
963 				       __alignof__(struct latency_stat), gfp);
964 	if (!iolat->stats) {
965 		kfree(iolat);
966 		return NULL;
967 	}
968 	return &iolat->pd;
969 }
970 
971 static void iolatency_pd_init(struct blkg_policy_data *pd)
972 {
973 	struct iolatency_grp *iolat = pd_to_lat(pd);
974 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
975 	struct rq_qos *rqos = blkcg_rq_qos(blkg->q);
976 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
977 	u64 now = ktime_to_ns(ktime_get());
978 	int cpu;
979 
980 	if (blk_queue_nonrot(blkg->q))
981 		iolat->ssd = true;
982 	else
983 		iolat->ssd = false;
984 
985 	for_each_possible_cpu(cpu) {
986 		struct latency_stat *stat;
987 		stat = per_cpu_ptr(iolat->stats, cpu);
988 		latency_stat_init(iolat, stat);
989 	}
990 
991 	latency_stat_init(iolat, &iolat->cur_stat);
992 	rq_wait_init(&iolat->rq_wait);
993 	spin_lock_init(&iolat->child_lat.lock);
994 	iolat->rq_depth.queue_depth = blkg->q->nr_requests;
995 	iolat->rq_depth.max_depth = UINT_MAX;
996 	iolat->rq_depth.default_depth = iolat->rq_depth.queue_depth;
997 	iolat->blkiolat = blkiolat;
998 	iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
999 	atomic64_set(&iolat->window_start, now);
1000 
1001 	/*
1002 	 * We init things in list order, so the pd for the parent may not be
1003 	 * init'ed yet for whatever reason.
1004 	 */
1005 	if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
1006 		struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
1007 		atomic_set(&iolat->scale_cookie,
1008 			   atomic_read(&parent->child_lat.scale_cookie));
1009 	} else {
1010 		atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
1011 	}
1012 
1013 	atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
1014 }
1015 
1016 static void iolatency_pd_offline(struct blkg_policy_data *pd)
1017 {
1018 	struct iolatency_grp *iolat = pd_to_lat(pd);
1019 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
1020 
1021 	iolatency_set_min_lat_nsec(blkg, 0);
1022 	iolatency_clear_scaling(blkg);
1023 }
1024 
1025 static void iolatency_pd_free(struct blkg_policy_data *pd)
1026 {
1027 	struct iolatency_grp *iolat = pd_to_lat(pd);
1028 	free_percpu(iolat->stats);
1029 	kfree(iolat);
1030 }
1031 
1032 static struct cftype iolatency_files[] = {
1033 	{
1034 		.name = "latency",
1035 		.flags = CFTYPE_NOT_ON_ROOT,
1036 		.seq_show = iolatency_print_limit,
1037 		.write = iolatency_set_limit,
1038 	},
1039 	{}
1040 };
1041 
1042 static struct blkcg_policy blkcg_policy_iolatency = {
1043 	.dfl_cftypes	= iolatency_files,
1044 	.pd_alloc_fn	= iolatency_pd_alloc,
1045 	.pd_init_fn	= iolatency_pd_init,
1046 	.pd_offline_fn	= iolatency_pd_offline,
1047 	.pd_free_fn	= iolatency_pd_free,
1048 	.pd_stat_fn	= iolatency_pd_stat,
1049 };
1050 
1051 static int __init iolatency_init(void)
1052 {
1053 	return blkcg_policy_register(&blkcg_policy_iolatency);
1054 }
1055 
1056 static void __exit iolatency_exit(void)
1057 {
1058 	blkcg_policy_unregister(&blkcg_policy_iolatency);
1059 }
1060 
1061 module_init(iolatency_init);
1062 module_exit(iolatency_exit);
1063