xref: /openbmc/linux/block/blk-iolatency.c (revision 0545810f)
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 	unsigned int max_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->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->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.disk->queue->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 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.  We halve the queue depth
370  * at a time so we can scale down queue depth quickly from default unlimited
371  * to target.
372  */
373 static void scale_change(struct iolatency_grp *iolat, bool up)
374 {
375 	unsigned long qd = iolat->blkiolat->rqos.disk->queue->nr_requests;
376 	unsigned long scale = scale_amount(qd, up);
377 	unsigned long old = iolat->max_depth;
378 
379 	if (old > qd)
380 		old = qd;
381 
382 	if (up) {
383 		if (old == 1 && blkcg_unuse_delay(lat_to_blkg(iolat)))
384 			return;
385 
386 		if (old < qd) {
387 			old += scale;
388 			old = min(old, qd);
389 			iolat->max_depth = old;
390 			wake_up_all(&iolat->rq_wait.wait);
391 		}
392 	} else {
393 		old >>= 1;
394 		iolat->max_depth = max(old, 1UL);
395 	}
396 }
397 
398 /* Check our parent and see if the scale cookie has changed. */
399 static void check_scale_change(struct iolatency_grp *iolat)
400 {
401 	struct iolatency_grp *parent;
402 	struct child_latency_info *lat_info;
403 	unsigned int cur_cookie;
404 	unsigned int our_cookie = atomic_read(&iolat->scale_cookie);
405 	u64 scale_lat;
406 	int direction = 0;
407 
408 	parent = blkg_to_lat(lat_to_blkg(iolat)->parent);
409 	if (!parent)
410 		return;
411 
412 	lat_info = &parent->child_lat;
413 	cur_cookie = atomic_read(&lat_info->scale_cookie);
414 	scale_lat = READ_ONCE(lat_info->scale_lat);
415 
416 	if (cur_cookie < our_cookie)
417 		direction = -1;
418 	else if (cur_cookie > our_cookie)
419 		direction = 1;
420 	else
421 		return;
422 
423 	if (!atomic_try_cmpxchg(&iolat->scale_cookie, &our_cookie, cur_cookie)) {
424 		/* Somebody beat us to the punch, just bail. */
425 		return;
426 	}
427 
428 	if (direction < 0 && iolat->min_lat_nsec) {
429 		u64 samples_thresh;
430 
431 		if (!scale_lat || iolat->min_lat_nsec <= scale_lat)
432 			return;
433 
434 		/*
435 		 * Sometimes high priority groups are their own worst enemy, so
436 		 * instead of taking it out on some poor other group that did 5%
437 		 * or less of the IO's for the last summation just skip this
438 		 * scale down event.
439 		 */
440 		samples_thresh = lat_info->nr_samples * 5;
441 		samples_thresh = max(1ULL, div64_u64(samples_thresh, 100));
442 		if (iolat->nr_samples <= samples_thresh)
443 			return;
444 	}
445 
446 	/* We're as low as we can go. */
447 	if (iolat->max_depth == 1 && direction < 0) {
448 		blkcg_use_delay(lat_to_blkg(iolat));
449 		return;
450 	}
451 
452 	/* We're back to the default cookie, unthrottle all the things. */
453 	if (cur_cookie == DEFAULT_SCALE_COOKIE) {
454 		blkcg_clear_delay(lat_to_blkg(iolat));
455 		iolat->max_depth = UINT_MAX;
456 		wake_up_all(&iolat->rq_wait.wait);
457 		return;
458 	}
459 
460 	scale_change(iolat, direction > 0);
461 }
462 
463 static void blkcg_iolatency_throttle(struct rq_qos *rqos, struct bio *bio)
464 {
465 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
466 	struct blkcg_gq *blkg = bio->bi_blkg;
467 	bool issue_as_root = bio_issue_as_root_blkg(bio);
468 
469 	if (!blkiolat->enabled)
470 		return;
471 
472 	while (blkg && blkg->parent) {
473 		struct iolatency_grp *iolat = blkg_to_lat(blkg);
474 		if (!iolat) {
475 			blkg = blkg->parent;
476 			continue;
477 		}
478 
479 		check_scale_change(iolat);
480 		__blkcg_iolatency_throttle(rqos, iolat, issue_as_root,
481 				     (bio->bi_opf & REQ_SWAP) == REQ_SWAP);
482 		blkg = blkg->parent;
483 	}
484 	if (!timer_pending(&blkiolat->timer))
485 		mod_timer(&blkiolat->timer, jiffies + HZ);
486 }
487 
488 static void iolatency_record_time(struct iolatency_grp *iolat,
489 				  struct bio_issue *issue, u64 now,
490 				  bool issue_as_root)
491 {
492 	u64 start = bio_issue_time(issue);
493 	u64 req_time;
494 
495 	/*
496 	 * Have to do this so we are truncated to the correct time that our
497 	 * issue is truncated to.
498 	 */
499 	now = __bio_issue_time(now);
500 
501 	if (now <= start)
502 		return;
503 
504 	req_time = now - start;
505 
506 	/*
507 	 * We don't want to count issue_as_root bio's in the cgroups latency
508 	 * statistics as it could skew the numbers downwards.
509 	 */
510 	if (unlikely(issue_as_root && iolat->max_depth != UINT_MAX)) {
511 		u64 sub = iolat->min_lat_nsec;
512 		if (req_time < sub)
513 			blkcg_add_delay(lat_to_blkg(iolat), now, sub - req_time);
514 		return;
515 	}
516 
517 	latency_stat_record_time(iolat, req_time);
518 }
519 
520 #define BLKIOLATENCY_MIN_ADJUST_TIME (500 * NSEC_PER_MSEC)
521 #define BLKIOLATENCY_MIN_GOOD_SAMPLES 5
522 
523 static void iolatency_check_latencies(struct iolatency_grp *iolat, u64 now)
524 {
525 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
526 	struct iolatency_grp *parent;
527 	struct child_latency_info *lat_info;
528 	struct latency_stat stat;
529 	unsigned long flags;
530 	int cpu;
531 
532 	latency_stat_init(iolat, &stat);
533 	preempt_disable();
534 	for_each_online_cpu(cpu) {
535 		struct latency_stat *s;
536 		s = per_cpu_ptr(iolat->stats, cpu);
537 		latency_stat_sum(iolat, &stat, s);
538 		latency_stat_init(iolat, s);
539 	}
540 	preempt_enable();
541 
542 	parent = blkg_to_lat(blkg->parent);
543 	if (!parent)
544 		return;
545 
546 	lat_info = &parent->child_lat;
547 
548 	iolat_update_total_lat_avg(iolat, &stat);
549 
550 	/* Everything is ok and we don't need to adjust the scale. */
551 	if (latency_sum_ok(iolat, &stat) &&
552 	    atomic_read(&lat_info->scale_cookie) == DEFAULT_SCALE_COOKIE)
553 		return;
554 
555 	/* Somebody beat us to the punch, just bail. */
556 	spin_lock_irqsave(&lat_info->lock, flags);
557 
558 	latency_stat_sum(iolat, &iolat->cur_stat, &stat);
559 	lat_info->nr_samples -= iolat->nr_samples;
560 	lat_info->nr_samples += latency_stat_samples(iolat, &iolat->cur_stat);
561 	iolat->nr_samples = latency_stat_samples(iolat, &iolat->cur_stat);
562 
563 	if ((lat_info->last_scale_event >= now ||
564 	    now - lat_info->last_scale_event < BLKIOLATENCY_MIN_ADJUST_TIME))
565 		goto out;
566 
567 	if (latency_sum_ok(iolat, &iolat->cur_stat) &&
568 	    latency_sum_ok(iolat, &stat)) {
569 		if (latency_stat_samples(iolat, &iolat->cur_stat) <
570 		    BLKIOLATENCY_MIN_GOOD_SAMPLES)
571 			goto out;
572 		if (lat_info->scale_grp == iolat) {
573 			lat_info->last_scale_event = now;
574 			scale_cookie_change(iolat->blkiolat, lat_info, true);
575 		}
576 	} else if (lat_info->scale_lat == 0 ||
577 		   lat_info->scale_lat >= iolat->min_lat_nsec) {
578 		lat_info->last_scale_event = now;
579 		if (!lat_info->scale_grp ||
580 		    lat_info->scale_lat > iolat->min_lat_nsec) {
581 			WRITE_ONCE(lat_info->scale_lat, iolat->min_lat_nsec);
582 			lat_info->scale_grp = iolat;
583 		}
584 		scale_cookie_change(iolat->blkiolat, lat_info, false);
585 	}
586 	latency_stat_init(iolat, &iolat->cur_stat);
587 out:
588 	spin_unlock_irqrestore(&lat_info->lock, flags);
589 }
590 
591 static void blkcg_iolatency_done_bio(struct rq_qos *rqos, struct bio *bio)
592 {
593 	struct blkcg_gq *blkg;
594 	struct rq_wait *rqw;
595 	struct iolatency_grp *iolat;
596 	u64 window_start;
597 	u64 now;
598 	bool issue_as_root = bio_issue_as_root_blkg(bio);
599 	int inflight = 0;
600 
601 	blkg = bio->bi_blkg;
602 	if (!blkg || !bio_flagged(bio, BIO_QOS_THROTTLED))
603 		return;
604 
605 	iolat = blkg_to_lat(bio->bi_blkg);
606 	if (!iolat)
607 		return;
608 
609 	if (!iolat->blkiolat->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_try_cmpxchg(&iolat->window_start,
634 							 &window_start, now))
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 	timer_shutdown_sync(&blkiolat->timer);
648 	flush_work(&blkiolat->enable_work);
649 	blkcg_deactivate_policy(rqos->disk, &blkcg_policy_iolatency);
650 	kfree(blkiolat);
651 }
652 
653 static const struct rq_qos_ops blkcg_iolatency_ops = {
654 	.throttle = blkcg_iolatency_throttle,
655 	.done_bio = blkcg_iolatency_done_bio,
656 	.exit = blkcg_iolatency_exit,
657 };
658 
659 static void blkiolatency_timer_fn(struct timer_list *t)
660 {
661 	struct blk_iolatency *blkiolat = from_timer(blkiolat, t, timer);
662 	struct blkcg_gq *blkg;
663 	struct cgroup_subsys_state *pos_css;
664 	u64 now = ktime_to_ns(ktime_get());
665 
666 	rcu_read_lock();
667 	blkg_for_each_descendant_pre(blkg, pos_css,
668 				     blkiolat->rqos.disk->queue->root_blkg) {
669 		struct iolatency_grp *iolat;
670 		struct child_latency_info *lat_info;
671 		unsigned long flags;
672 		u64 cookie;
673 
674 		/*
675 		 * We could be exiting, don't access the pd unless we have a
676 		 * ref on the blkg.
677 		 */
678 		if (!blkg_tryget(blkg))
679 			continue;
680 
681 		iolat = blkg_to_lat(blkg);
682 		if (!iolat)
683 			goto next;
684 
685 		lat_info = &iolat->child_lat;
686 		cookie = atomic_read(&lat_info->scale_cookie);
687 
688 		if (cookie >= DEFAULT_SCALE_COOKIE)
689 			goto next;
690 
691 		spin_lock_irqsave(&lat_info->lock, flags);
692 		if (lat_info->last_scale_event >= now)
693 			goto next_lock;
694 
695 		/*
696 		 * We scaled down but don't have a scale_grp, scale up and carry
697 		 * on.
698 		 */
699 		if (lat_info->scale_grp == NULL) {
700 			scale_cookie_change(iolat->blkiolat, lat_info, true);
701 			goto next_lock;
702 		}
703 
704 		/*
705 		 * It's been 5 seconds since our last scale event, clear the
706 		 * scale grp in case the group that needed the scale down isn't
707 		 * doing any IO currently.
708 		 */
709 		if (now - lat_info->last_scale_event >=
710 		    ((u64)NSEC_PER_SEC * 5))
711 			lat_info->scale_grp = NULL;
712 next_lock:
713 		spin_unlock_irqrestore(&lat_info->lock, flags);
714 next:
715 		blkg_put(blkg);
716 	}
717 	rcu_read_unlock();
718 }
719 
720 /**
721  * blkiolatency_enable_work_fn - Enable or disable iolatency on the device
722  * @work: enable_work of the blk_iolatency of interest
723  *
724  * iolatency needs to keep track of the number of in-flight IOs per cgroup. This
725  * is relatively expensive as it involves walking up the hierarchy twice for
726  * every IO. Thus, if iolatency is not enabled in any cgroup for the device, we
727  * want to disable the in-flight tracking.
728  *
729  * We have to make sure that the counting is balanced - we don't want to leak
730  * the in-flight counts by disabling accounting in the completion path while IOs
731  * are in flight. This is achieved by ensuring that no IO is in flight by
732  * freezing the queue while flipping ->enabled. As this requires a sleepable
733  * context, ->enabled flipping is punted to this work function.
734  */
735 static void blkiolatency_enable_work_fn(struct work_struct *work)
736 {
737 	struct blk_iolatency *blkiolat = container_of(work, struct blk_iolatency,
738 						      enable_work);
739 	bool enabled;
740 
741 	/*
742 	 * There can only be one instance of this function running for @blkiolat
743 	 * and it's guaranteed to be executed at least once after the latest
744 	 * ->enabled_cnt modification. Acting on the latest ->enable_cnt is
745 	 * sufficient.
746 	 *
747 	 * Also, we know @blkiolat is safe to access as ->enable_work is flushed
748 	 * in blkcg_iolatency_exit().
749 	 */
750 	enabled = atomic_read(&blkiolat->enable_cnt);
751 	if (enabled != blkiolat->enabled) {
752 		blk_mq_freeze_queue(blkiolat->rqos.disk->queue);
753 		blkiolat->enabled = enabled;
754 		blk_mq_unfreeze_queue(blkiolat->rqos.disk->queue);
755 	}
756 }
757 
758 static int blk_iolatency_init(struct gendisk *disk)
759 {
760 	struct blk_iolatency *blkiolat;
761 	int ret;
762 
763 	blkiolat = kzalloc(sizeof(*blkiolat), GFP_KERNEL);
764 	if (!blkiolat)
765 		return -ENOMEM;
766 
767 	ret = rq_qos_add(&blkiolat->rqos, disk, RQ_QOS_LATENCY,
768 			 &blkcg_iolatency_ops);
769 	if (ret)
770 		goto err_free;
771 	ret = blkcg_activate_policy(disk, &blkcg_policy_iolatency);
772 	if (ret)
773 		goto err_qos_del;
774 
775 	timer_setup(&blkiolat->timer, blkiolatency_timer_fn, 0);
776 	INIT_WORK(&blkiolat->enable_work, blkiolatency_enable_work_fn);
777 
778 	return 0;
779 
780 err_qos_del:
781 	rq_qos_del(&blkiolat->rqos);
782 err_free:
783 	kfree(blkiolat);
784 	return ret;
785 }
786 
787 static void iolatency_set_min_lat_nsec(struct blkcg_gq *blkg, u64 val)
788 {
789 	struct iolatency_grp *iolat = blkg_to_lat(blkg);
790 	struct blk_iolatency *blkiolat = iolat->blkiolat;
791 	u64 oldval = iolat->min_lat_nsec;
792 
793 	iolat->min_lat_nsec = val;
794 	iolat->cur_win_nsec = max_t(u64, val << 4, BLKIOLATENCY_MIN_WIN_SIZE);
795 	iolat->cur_win_nsec = min_t(u64, iolat->cur_win_nsec,
796 				    BLKIOLATENCY_MAX_WIN_SIZE);
797 
798 	if (!oldval && val) {
799 		if (atomic_inc_return(&blkiolat->enable_cnt) == 1)
800 			schedule_work(&blkiolat->enable_work);
801 	}
802 	if (oldval && !val) {
803 		blkcg_clear_delay(blkg);
804 		if (atomic_dec_return(&blkiolat->enable_cnt) == 0)
805 			schedule_work(&blkiolat->enable_work);
806 	}
807 }
808 
809 static void iolatency_clear_scaling(struct blkcg_gq *blkg)
810 {
811 	if (blkg->parent) {
812 		struct iolatency_grp *iolat = blkg_to_lat(blkg->parent);
813 		struct child_latency_info *lat_info;
814 		if (!iolat)
815 			return;
816 
817 		lat_info = &iolat->child_lat;
818 		spin_lock(&lat_info->lock);
819 		atomic_set(&lat_info->scale_cookie, DEFAULT_SCALE_COOKIE);
820 		lat_info->last_scale_event = 0;
821 		lat_info->scale_grp = NULL;
822 		lat_info->scale_lat = 0;
823 		spin_unlock(&lat_info->lock);
824 	}
825 }
826 
827 static int blk_iolatency_try_init(struct blkg_conf_ctx *ctx)
828 {
829 	static DEFINE_MUTEX(init_mutex);
830 	int ret;
831 
832 	ret = blkg_conf_open_bdev(ctx);
833 	if (ret)
834 		return ret;
835 
836 	/*
837 	 * blk_iolatency_init() may fail after rq_qos_add() succeeds which can
838 	 * confuse iolat_rq_qos() test. Make the test and init atomic.
839 	 */
840 	mutex_lock(&init_mutex);
841 
842 	if (!iolat_rq_qos(ctx->bdev->bd_queue))
843 		ret = blk_iolatency_init(ctx->bdev->bd_disk);
844 
845 	mutex_unlock(&init_mutex);
846 
847 	return ret;
848 }
849 
850 static ssize_t iolatency_set_limit(struct kernfs_open_file *of, char *buf,
851 			     size_t nbytes, loff_t off)
852 {
853 	struct blkcg *blkcg = css_to_blkcg(of_css(of));
854 	struct blkcg_gq *blkg;
855 	struct blkg_conf_ctx ctx;
856 	struct iolatency_grp *iolat;
857 	char *p, *tok;
858 	u64 lat_val = 0;
859 	u64 oldval;
860 	int ret;
861 
862 	blkg_conf_init(&ctx, buf);
863 
864 	ret = blk_iolatency_try_init(&ctx);
865 	if (ret)
866 		goto out;
867 
868 	ret = blkg_conf_prep(blkcg, &blkcg_policy_iolatency, &ctx);
869 	if (ret)
870 		goto out;
871 
872 	iolat = blkg_to_lat(ctx.blkg);
873 	p = ctx.body;
874 
875 	ret = -EINVAL;
876 	while ((tok = strsep(&p, " "))) {
877 		char key[16];
878 		char val[21];	/* 18446744073709551616 */
879 
880 		if (sscanf(tok, "%15[^=]=%20s", key, val) != 2)
881 			goto out;
882 
883 		if (!strcmp(key, "target")) {
884 			u64 v;
885 
886 			if (!strcmp(val, "max"))
887 				lat_val = 0;
888 			else if (sscanf(val, "%llu", &v) == 1)
889 				lat_val = v * NSEC_PER_USEC;
890 			else
891 				goto out;
892 		} else {
893 			goto out;
894 		}
895 	}
896 
897 	/* Walk up the tree to see if our new val is lower than it should be. */
898 	blkg = ctx.blkg;
899 	oldval = iolat->min_lat_nsec;
900 
901 	iolatency_set_min_lat_nsec(blkg, lat_val);
902 	if (oldval != iolat->min_lat_nsec)
903 		iolatency_clear_scaling(blkg);
904 	ret = 0;
905 out:
906 	blkg_conf_exit(&ctx);
907 	return ret ?: nbytes;
908 }
909 
910 static u64 iolatency_prfill_limit(struct seq_file *sf,
911 				  struct blkg_policy_data *pd, int off)
912 {
913 	struct iolatency_grp *iolat = pd_to_lat(pd);
914 	const char *dname = blkg_dev_name(pd->blkg);
915 
916 	if (!dname || !iolat->min_lat_nsec)
917 		return 0;
918 	seq_printf(sf, "%s target=%llu\n",
919 		   dname, div_u64(iolat->min_lat_nsec, NSEC_PER_USEC));
920 	return 0;
921 }
922 
923 static int iolatency_print_limit(struct seq_file *sf, void *v)
924 {
925 	blkcg_print_blkgs(sf, css_to_blkcg(seq_css(sf)),
926 			  iolatency_prfill_limit,
927 			  &blkcg_policy_iolatency, seq_cft(sf)->private, false);
928 	return 0;
929 }
930 
931 static void iolatency_ssd_stat(struct iolatency_grp *iolat, struct seq_file *s)
932 {
933 	struct latency_stat stat;
934 	int cpu;
935 
936 	latency_stat_init(iolat, &stat);
937 	preempt_disable();
938 	for_each_online_cpu(cpu) {
939 		struct latency_stat *s;
940 		s = per_cpu_ptr(iolat->stats, cpu);
941 		latency_stat_sum(iolat, &stat, s);
942 	}
943 	preempt_enable();
944 
945 	if (iolat->max_depth == UINT_MAX)
946 		seq_printf(s, " missed=%llu total=%llu depth=max",
947 			(unsigned long long)stat.ps.missed,
948 			(unsigned long long)stat.ps.total);
949 	else
950 		seq_printf(s, " missed=%llu total=%llu depth=%u",
951 			(unsigned long long)stat.ps.missed,
952 			(unsigned long long)stat.ps.total,
953 			iolat->max_depth);
954 }
955 
956 static void iolatency_pd_stat(struct blkg_policy_data *pd, struct seq_file *s)
957 {
958 	struct iolatency_grp *iolat = pd_to_lat(pd);
959 	unsigned long long avg_lat;
960 	unsigned long long cur_win;
961 
962 	if (!blkcg_debug_stats)
963 		return;
964 
965 	if (iolat->ssd)
966 		return iolatency_ssd_stat(iolat, s);
967 
968 	avg_lat = div64_u64(iolat->lat_avg, NSEC_PER_USEC);
969 	cur_win = div64_u64(iolat->cur_win_nsec, NSEC_PER_MSEC);
970 	if (iolat->max_depth == UINT_MAX)
971 		seq_printf(s, " depth=max avg_lat=%llu win=%llu",
972 			avg_lat, cur_win);
973 	else
974 		seq_printf(s, " depth=%u avg_lat=%llu win=%llu",
975 			iolat->max_depth, avg_lat, cur_win);
976 }
977 
978 static struct blkg_policy_data *iolatency_pd_alloc(struct gendisk *disk,
979 		struct blkcg *blkcg, gfp_t gfp)
980 {
981 	struct iolatency_grp *iolat;
982 
983 	iolat = kzalloc_node(sizeof(*iolat), gfp, disk->node_id);
984 	if (!iolat)
985 		return NULL;
986 	iolat->stats = __alloc_percpu_gfp(sizeof(struct latency_stat),
987 				       __alignof__(struct latency_stat), gfp);
988 	if (!iolat->stats) {
989 		kfree(iolat);
990 		return NULL;
991 	}
992 	return &iolat->pd;
993 }
994 
995 static void iolatency_pd_init(struct blkg_policy_data *pd)
996 {
997 	struct iolatency_grp *iolat = pd_to_lat(pd);
998 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
999 	struct rq_qos *rqos = iolat_rq_qos(blkg->q);
1000 	struct blk_iolatency *blkiolat = BLKIOLATENCY(rqos);
1001 	u64 now = ktime_to_ns(ktime_get());
1002 	int cpu;
1003 
1004 	if (blk_queue_nonrot(blkg->q))
1005 		iolat->ssd = true;
1006 	else
1007 		iolat->ssd = false;
1008 
1009 	for_each_possible_cpu(cpu) {
1010 		struct latency_stat *stat;
1011 		stat = per_cpu_ptr(iolat->stats, cpu);
1012 		latency_stat_init(iolat, stat);
1013 	}
1014 
1015 	latency_stat_init(iolat, &iolat->cur_stat);
1016 	rq_wait_init(&iolat->rq_wait);
1017 	spin_lock_init(&iolat->child_lat.lock);
1018 	iolat->max_depth = UINT_MAX;
1019 	iolat->blkiolat = blkiolat;
1020 	iolat->cur_win_nsec = 100 * NSEC_PER_MSEC;
1021 	atomic64_set(&iolat->window_start, now);
1022 
1023 	/*
1024 	 * We init things in list order, so the pd for the parent may not be
1025 	 * init'ed yet for whatever reason.
1026 	 */
1027 	if (blkg->parent && blkg_to_pd(blkg->parent, &blkcg_policy_iolatency)) {
1028 		struct iolatency_grp *parent = blkg_to_lat(blkg->parent);
1029 		atomic_set(&iolat->scale_cookie,
1030 			   atomic_read(&parent->child_lat.scale_cookie));
1031 	} else {
1032 		atomic_set(&iolat->scale_cookie, DEFAULT_SCALE_COOKIE);
1033 	}
1034 
1035 	atomic_set(&iolat->child_lat.scale_cookie, DEFAULT_SCALE_COOKIE);
1036 }
1037 
1038 static void iolatency_pd_offline(struct blkg_policy_data *pd)
1039 {
1040 	struct iolatency_grp *iolat = pd_to_lat(pd);
1041 	struct blkcg_gq *blkg = lat_to_blkg(iolat);
1042 
1043 	iolatency_set_min_lat_nsec(blkg, 0);
1044 	iolatency_clear_scaling(blkg);
1045 }
1046 
1047 static void iolatency_pd_free(struct blkg_policy_data *pd)
1048 {
1049 	struct iolatency_grp *iolat = pd_to_lat(pd);
1050 	free_percpu(iolat->stats);
1051 	kfree(iolat);
1052 }
1053 
1054 static struct cftype iolatency_files[] = {
1055 	{
1056 		.name = "latency",
1057 		.flags = CFTYPE_NOT_ON_ROOT,
1058 		.seq_show = iolatency_print_limit,
1059 		.write = iolatency_set_limit,
1060 	},
1061 	{}
1062 };
1063 
1064 static struct blkcg_policy blkcg_policy_iolatency = {
1065 	.dfl_cftypes	= iolatency_files,
1066 	.pd_alloc_fn	= iolatency_pd_alloc,
1067 	.pd_init_fn	= iolatency_pd_init,
1068 	.pd_offline_fn	= iolatency_pd_offline,
1069 	.pd_free_fn	= iolatency_pd_free,
1070 	.pd_stat_fn	= iolatency_pd_stat,
1071 };
1072 
1073 static int __init iolatency_init(void)
1074 {
1075 	return blkcg_policy_register(&blkcg_policy_iolatency);
1076 }
1077 
1078 static void __exit iolatency_exit(void)
1079 {
1080 	blkcg_policy_unregister(&blkcg_policy_iolatency);
1081 }
1082 
1083 module_init(iolatency_init);
1084 module_exit(iolatency_exit);
1085