xref: /linux/block/blk-wbt.c (revision 9a379e77033f02c4a071891afdf0f0a01eff8ccb)
1 /*
2  * buffered writeback throttling. loosely based on CoDel. We can't drop
3  * packets for IO scheduling, so the logic is something like this:
4  *
5  * - Monitor latencies in a defined window of time.
6  * - If the minimum latency in the above window exceeds some target, increment
7  *   scaling step and scale down queue depth by a factor of 2x. The monitoring
8  *   window is then shrunk to 100 / sqrt(scaling step + 1).
9  * - For any window where we don't have solid data on what the latencies
10  *   look like, retain status quo.
11  * - If latencies look good, decrement scaling step.
12  * - If we're only doing writes, allow the scaling step to go negative. This
13  *   will temporarily boost write performance, snapping back to a stable
14  *   scaling step of 0 if reads show up or the heavy writers finish. Unlike
15  *   positive scaling steps where we shrink the monitoring window, a negative
16  *   scaling step retains the default step==0 window size.
17  *
18  * Copyright (C) 2016 Jens Axboe
19  *
20  */
21 #include <linux/kernel.h>
22 #include <linux/blk_types.h>
23 #include <linux/slab.h>
24 #include <linux/backing-dev.h>
25 #include <linux/swap.h>
26 
27 #include "blk-wbt.h"
28 
29 #define CREATE_TRACE_POINTS
30 #include <trace/events/wbt.h>
31 
32 enum {
33 	/*
34 	 * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
35 	 * from here depending on device stats
36 	 */
37 	RWB_DEF_DEPTH	= 16,
38 
39 	/*
40 	 * 100msec window
41 	 */
42 	RWB_WINDOW_NSEC		= 100 * 1000 * 1000ULL,
43 
44 	/*
45 	 * Disregard stats, if we don't meet this minimum
46 	 */
47 	RWB_MIN_WRITE_SAMPLES	= 3,
48 
49 	/*
50 	 * If we have this number of consecutive windows with not enough
51 	 * information to scale up or down, scale up.
52 	 */
53 	RWB_UNKNOWN_BUMP	= 5,
54 };
55 
56 static inline bool rwb_enabled(struct rq_wb *rwb)
57 {
58 	return rwb && rwb->wb_normal != 0;
59 }
60 
61 /*
62  * Increment 'v', if 'v' is below 'below'. Returns true if we succeeded,
63  * false if 'v' + 1 would be bigger than 'below'.
64  */
65 static bool atomic_inc_below(atomic_t *v, int below)
66 {
67 	int cur = atomic_read(v);
68 
69 	for (;;) {
70 		int old;
71 
72 		if (cur >= below)
73 			return false;
74 		old = atomic_cmpxchg(v, cur, cur + 1);
75 		if (old == cur)
76 			break;
77 		cur = old;
78 	}
79 
80 	return true;
81 }
82 
83 static void wb_timestamp(struct rq_wb *rwb, unsigned long *var)
84 {
85 	if (rwb_enabled(rwb)) {
86 		const unsigned long cur = jiffies;
87 
88 		if (cur != *var)
89 			*var = cur;
90 	}
91 }
92 
93 /*
94  * If a task was rate throttled in balance_dirty_pages() within the last
95  * second or so, use that to indicate a higher cleaning rate.
96  */
97 static bool wb_recent_wait(struct rq_wb *rwb)
98 {
99 	struct bdi_writeback *wb = &rwb->queue->backing_dev_info->wb;
100 
101 	return time_before(jiffies, wb->dirty_sleep + HZ);
102 }
103 
104 static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, bool is_kswapd)
105 {
106 	return &rwb->rq_wait[is_kswapd];
107 }
108 
109 static void rwb_wake_all(struct rq_wb *rwb)
110 {
111 	int i;
112 
113 	for (i = 0; i < WBT_NUM_RWQ; i++) {
114 		struct rq_wait *rqw = &rwb->rq_wait[i];
115 
116 		if (waitqueue_active(&rqw->wait))
117 			wake_up_all(&rqw->wait);
118 	}
119 }
120 
121 void __wbt_done(struct rq_wb *rwb, enum wbt_flags wb_acct)
122 {
123 	struct rq_wait *rqw;
124 	int inflight, limit;
125 
126 	if (!(wb_acct & WBT_TRACKED))
127 		return;
128 
129 	rqw = get_rq_wait(rwb, wb_acct & WBT_KSWAPD);
130 	inflight = atomic_dec_return(&rqw->inflight);
131 
132 	/*
133 	 * wbt got disabled with IO in flight. Wake up any potential
134 	 * waiters, we don't have to do more than that.
135 	 */
136 	if (unlikely(!rwb_enabled(rwb))) {
137 		rwb_wake_all(rwb);
138 		return;
139 	}
140 
141 	/*
142 	 * If the device does write back caching, drop further down
143 	 * before we wake people up.
144 	 */
145 	if (rwb->wc && !wb_recent_wait(rwb))
146 		limit = 0;
147 	else
148 		limit = rwb->wb_normal;
149 
150 	/*
151 	 * Don't wake anyone up if we are above the normal limit.
152 	 */
153 	if (inflight && inflight >= limit)
154 		return;
155 
156 	if (waitqueue_active(&rqw->wait)) {
157 		int diff = limit - inflight;
158 
159 		if (!inflight || diff >= rwb->wb_background / 2)
160 			wake_up_all(&rqw->wait);
161 	}
162 }
163 
164 /*
165  * Called on completion of a request. Note that it's also called when
166  * a request is merged, when the request gets freed.
167  */
168 void wbt_done(struct rq_wb *rwb, struct blk_issue_stat *stat)
169 {
170 	if (!rwb)
171 		return;
172 
173 	if (!wbt_is_tracked(stat)) {
174 		if (rwb->sync_cookie == stat) {
175 			rwb->sync_issue = 0;
176 			rwb->sync_cookie = NULL;
177 		}
178 
179 		if (wbt_is_read(stat))
180 			wb_timestamp(rwb, &rwb->last_comp);
181 	} else {
182 		WARN_ON_ONCE(stat == rwb->sync_cookie);
183 		__wbt_done(rwb, wbt_stat_to_mask(stat));
184 	}
185 	wbt_clear_state(stat);
186 }
187 
188 /*
189  * Return true, if we can't increase the depth further by scaling
190  */
191 static bool calc_wb_limits(struct rq_wb *rwb)
192 {
193 	unsigned int depth;
194 	bool ret = false;
195 
196 	if (!rwb->min_lat_nsec) {
197 		rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0;
198 		return false;
199 	}
200 
201 	/*
202 	 * For QD=1 devices, this is a special case. It's important for those
203 	 * to have one request ready when one completes, so force a depth of
204 	 * 2 for those devices. On the backend, it'll be a depth of 1 anyway,
205 	 * since the device can't have more than that in flight. If we're
206 	 * scaling down, then keep a setting of 1/1/1.
207 	 */
208 	if (rwb->queue_depth == 1) {
209 		if (rwb->scale_step > 0)
210 			rwb->wb_max = rwb->wb_normal = 1;
211 		else {
212 			rwb->wb_max = rwb->wb_normal = 2;
213 			ret = true;
214 		}
215 		rwb->wb_background = 1;
216 	} else {
217 		/*
218 		 * scale_step == 0 is our default state. If we have suffered
219 		 * latency spikes, step will be > 0, and we shrink the
220 		 * allowed write depths. If step is < 0, we're only doing
221 		 * writes, and we allow a temporarily higher depth to
222 		 * increase performance.
223 		 */
224 		depth = min_t(unsigned int, RWB_DEF_DEPTH, rwb->queue_depth);
225 		if (rwb->scale_step > 0)
226 			depth = 1 + ((depth - 1) >> min(31, rwb->scale_step));
227 		else if (rwb->scale_step < 0) {
228 			unsigned int maxd = 3 * rwb->queue_depth / 4;
229 
230 			depth = 1 + ((depth - 1) << -rwb->scale_step);
231 			if (depth > maxd) {
232 				depth = maxd;
233 				ret = true;
234 			}
235 		}
236 
237 		/*
238 		 * Set our max/normal/bg queue depths based on how far
239 		 * we have scaled down (->scale_step).
240 		 */
241 		rwb->wb_max = depth;
242 		rwb->wb_normal = (rwb->wb_max + 1) / 2;
243 		rwb->wb_background = (rwb->wb_max + 3) / 4;
244 	}
245 
246 	return ret;
247 }
248 
249 static inline bool stat_sample_valid(struct blk_rq_stat *stat)
250 {
251 	/*
252 	 * We need at least one read sample, and a minimum of
253 	 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
254 	 * that it's writes impacting us, and not just some sole read on
255 	 * a device that is in a lower power state.
256 	 */
257 	return (stat[READ].nr_samples >= 1 &&
258 		stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
259 }
260 
261 static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
262 {
263 	u64 now, issue = READ_ONCE(rwb->sync_issue);
264 
265 	if (!issue || !rwb->sync_cookie)
266 		return 0;
267 
268 	now = ktime_to_ns(ktime_get());
269 	return now - issue;
270 }
271 
272 enum {
273 	LAT_OK = 1,
274 	LAT_UNKNOWN,
275 	LAT_UNKNOWN_WRITES,
276 	LAT_EXCEEDED,
277 };
278 
279 static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
280 {
281 	struct backing_dev_info *bdi = rwb->queue->backing_dev_info;
282 	u64 thislat;
283 
284 	/*
285 	 * If our stored sync issue exceeds the window size, or it
286 	 * exceeds our min target AND we haven't logged any entries,
287 	 * flag the latency as exceeded. wbt works off completion latencies,
288 	 * but for a flooded device, a single sync IO can take a long time
289 	 * to complete after being issued. If this time exceeds our
290 	 * monitoring window AND we didn't see any other completions in that
291 	 * window, then count that sync IO as a violation of the latency.
292 	 */
293 	thislat = rwb_sync_issue_lat(rwb);
294 	if (thislat > rwb->cur_win_nsec ||
295 	    (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
296 		trace_wbt_lat(bdi, thislat);
297 		return LAT_EXCEEDED;
298 	}
299 
300 	/*
301 	 * No read/write mix, if stat isn't valid
302 	 */
303 	if (!stat_sample_valid(stat)) {
304 		/*
305 		 * If we had writes in this stat window and the window is
306 		 * current, we're only doing writes. If a task recently
307 		 * waited or still has writes in flights, consider us doing
308 		 * just writes as well.
309 		 */
310 		if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
311 		    wbt_inflight(rwb))
312 			return LAT_UNKNOWN_WRITES;
313 		return LAT_UNKNOWN;
314 	}
315 
316 	/*
317 	 * If the 'min' latency exceeds our target, step down.
318 	 */
319 	if (stat[READ].min > rwb->min_lat_nsec) {
320 		trace_wbt_lat(bdi, stat[READ].min);
321 		trace_wbt_stat(bdi, stat);
322 		return LAT_EXCEEDED;
323 	}
324 
325 	if (rwb->scale_step)
326 		trace_wbt_stat(bdi, stat);
327 
328 	return LAT_OK;
329 }
330 
331 static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
332 {
333 	struct backing_dev_info *bdi = rwb->queue->backing_dev_info;
334 
335 	trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec,
336 			rwb->wb_background, rwb->wb_normal, rwb->wb_max);
337 }
338 
339 static void scale_up(struct rq_wb *rwb)
340 {
341 	/*
342 	 * Hit max in previous round, stop here
343 	 */
344 	if (rwb->scaled_max)
345 		return;
346 
347 	rwb->scale_step--;
348 	rwb->unknown_cnt = 0;
349 
350 	rwb->scaled_max = calc_wb_limits(rwb);
351 
352 	rwb_wake_all(rwb);
353 
354 	rwb_trace_step(rwb, "step up");
355 }
356 
357 /*
358  * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we
359  * had a latency violation.
360  */
361 static void scale_down(struct rq_wb *rwb, bool hard_throttle)
362 {
363 	/*
364 	 * Stop scaling down when we've hit the limit. This also prevents
365 	 * ->scale_step from going to crazy values, if the device can't
366 	 * keep up.
367 	 */
368 	if (rwb->wb_max == 1)
369 		return;
370 
371 	if (rwb->scale_step < 0 && hard_throttle)
372 		rwb->scale_step = 0;
373 	else
374 		rwb->scale_step++;
375 
376 	rwb->scaled_max = false;
377 	rwb->unknown_cnt = 0;
378 	calc_wb_limits(rwb);
379 	rwb_trace_step(rwb, "step down");
380 }
381 
382 static void rwb_arm_timer(struct rq_wb *rwb)
383 {
384 	if (rwb->scale_step > 0) {
385 		/*
386 		 * We should speed this up, using some variant of a fast
387 		 * integer inverse square root calculation. Since we only do
388 		 * this for every window expiration, it's not a huge deal,
389 		 * though.
390 		 */
391 		rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
392 					int_sqrt((rwb->scale_step + 1) << 8));
393 	} else {
394 		/*
395 		 * For step < 0, we don't want to increase/decrease the
396 		 * window size.
397 		 */
398 		rwb->cur_win_nsec = rwb->win_nsec;
399 	}
400 
401 	blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
402 }
403 
404 static void wb_timer_fn(struct blk_stat_callback *cb)
405 {
406 	struct rq_wb *rwb = cb->data;
407 	unsigned int inflight = wbt_inflight(rwb);
408 	int status;
409 
410 	status = latency_exceeded(rwb, cb->stat);
411 
412 	trace_wbt_timer(rwb->queue->backing_dev_info, status, rwb->scale_step,
413 			inflight);
414 
415 	/*
416 	 * If we exceeded the latency target, step down. If we did not,
417 	 * step one level up. If we don't know enough to say either exceeded
418 	 * or ok, then don't do anything.
419 	 */
420 	switch (status) {
421 	case LAT_EXCEEDED:
422 		scale_down(rwb, true);
423 		break;
424 	case LAT_OK:
425 		scale_up(rwb);
426 		break;
427 	case LAT_UNKNOWN_WRITES:
428 		/*
429 		 * We started a the center step, but don't have a valid
430 		 * read/write sample, but we do have writes going on.
431 		 * Allow step to go negative, to increase write perf.
432 		 */
433 		scale_up(rwb);
434 		break;
435 	case LAT_UNKNOWN:
436 		if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
437 			break;
438 		/*
439 		 * We get here when previously scaled reduced depth, and we
440 		 * currently don't have a valid read/write sample. For that
441 		 * case, slowly return to center state (step == 0).
442 		 */
443 		if (rwb->scale_step > 0)
444 			scale_up(rwb);
445 		else if (rwb->scale_step < 0)
446 			scale_down(rwb, false);
447 		break;
448 	default:
449 		break;
450 	}
451 
452 	/*
453 	 * Re-arm timer, if we have IO in flight
454 	 */
455 	if (rwb->scale_step || inflight)
456 		rwb_arm_timer(rwb);
457 }
458 
459 void wbt_update_limits(struct rq_wb *rwb)
460 {
461 	rwb->scale_step = 0;
462 	rwb->scaled_max = false;
463 	calc_wb_limits(rwb);
464 
465 	rwb_wake_all(rwb);
466 }
467 
468 static bool close_io(struct rq_wb *rwb)
469 {
470 	const unsigned long now = jiffies;
471 
472 	return time_before(now, rwb->last_issue + HZ / 10) ||
473 		time_before(now, rwb->last_comp + HZ / 10);
474 }
475 
476 #define REQ_HIPRIO	(REQ_SYNC | REQ_META | REQ_PRIO)
477 
478 static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw)
479 {
480 	unsigned int limit;
481 
482 	/*
483 	 * At this point we know it's a buffered write. If this is
484 	 * kswapd trying to free memory, or REQ_SYNC is set, then
485 	 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
486 	 * that. If the write is marked as a background write, then use
487 	 * the idle limit, or go to normal if we haven't had competing
488 	 * IO for a bit.
489 	 */
490 	if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
491 		limit = rwb->wb_max;
492 	else if ((rw & REQ_BACKGROUND) || close_io(rwb)) {
493 		/*
494 		 * If less than 100ms since we completed unrelated IO,
495 		 * limit us to half the depth for background writeback.
496 		 */
497 		limit = rwb->wb_background;
498 	} else
499 		limit = rwb->wb_normal;
500 
501 	return limit;
502 }
503 
504 static inline bool may_queue(struct rq_wb *rwb, struct rq_wait *rqw,
505 			     wait_queue_entry_t *wait, unsigned long rw)
506 {
507 	/*
508 	 * inc it here even if disabled, since we'll dec it at completion.
509 	 * this only happens if the task was sleeping in __wbt_wait(),
510 	 * and someone turned it off at the same time.
511 	 */
512 	if (!rwb_enabled(rwb)) {
513 		atomic_inc(&rqw->inflight);
514 		return true;
515 	}
516 
517 	/*
518 	 * If the waitqueue is already active and we are not the next
519 	 * in line to be woken up, wait for our turn.
520 	 */
521 	if (waitqueue_active(&rqw->wait) &&
522 	    rqw->wait.head.next != &wait->entry)
523 		return false;
524 
525 	return atomic_inc_below(&rqw->inflight, get_limit(rwb, rw));
526 }
527 
528 /*
529  * Block if we will exceed our limit, or if we are currently waiting for
530  * the timer to kick off queuing again.
531  */
532 static void __wbt_wait(struct rq_wb *rwb, unsigned long rw, spinlock_t *lock)
533 	__releases(lock)
534 	__acquires(lock)
535 {
536 	struct rq_wait *rqw = get_rq_wait(rwb, current_is_kswapd());
537 	DEFINE_WAIT(wait);
538 
539 	if (may_queue(rwb, rqw, &wait, rw))
540 		return;
541 
542 	do {
543 		prepare_to_wait_exclusive(&rqw->wait, &wait,
544 						TASK_UNINTERRUPTIBLE);
545 
546 		if (may_queue(rwb, rqw, &wait, rw))
547 			break;
548 
549 		if (lock) {
550 			spin_unlock_irq(lock);
551 			io_schedule();
552 			spin_lock_irq(lock);
553 		} else
554 			io_schedule();
555 	} while (1);
556 
557 	finish_wait(&rqw->wait, &wait);
558 }
559 
560 static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio)
561 {
562 	const int op = bio_op(bio);
563 
564 	/*
565 	 * If not a WRITE, do nothing
566 	 */
567 	if (op != REQ_OP_WRITE)
568 		return false;
569 
570 	/*
571 	 * Don't throttle WRITE_ODIRECT
572 	 */
573 	if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == (REQ_SYNC | REQ_IDLE))
574 		return false;
575 
576 	return true;
577 }
578 
579 /*
580  * Returns true if the IO request should be accounted, false if not.
581  * May sleep, if we have exceeded the writeback limits. Caller can pass
582  * in an irq held spinlock, if it holds one when calling this function.
583  * If we do sleep, we'll release and re-grab it.
584  */
585 enum wbt_flags wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock)
586 {
587 	unsigned int ret = 0;
588 
589 	if (!rwb_enabled(rwb))
590 		return 0;
591 
592 	if (bio_op(bio) == REQ_OP_READ)
593 		ret = WBT_READ;
594 
595 	if (!wbt_should_throttle(rwb, bio)) {
596 		if (ret & WBT_READ)
597 			wb_timestamp(rwb, &rwb->last_issue);
598 		return ret;
599 	}
600 
601 	__wbt_wait(rwb, bio->bi_opf, lock);
602 
603 	if (!blk_stat_is_active(rwb->cb))
604 		rwb_arm_timer(rwb);
605 
606 	if (current_is_kswapd())
607 		ret |= WBT_KSWAPD;
608 
609 	return ret | WBT_TRACKED;
610 }
611 
612 void wbt_issue(struct rq_wb *rwb, struct blk_issue_stat *stat)
613 {
614 	if (!rwb_enabled(rwb))
615 		return;
616 
617 	/*
618 	 * Track sync issue, in case it takes a long time to complete. Allows
619 	 * us to react quicker, if a sync IO takes a long time to complete.
620 	 * Note that this is just a hint. 'stat' can go away when the
621 	 * request completes, so it's important we never dereference it. We
622 	 * only use the address to compare with, which is why we store the
623 	 * sync_issue time locally.
624 	 */
625 	if (wbt_is_read(stat) && !rwb->sync_issue) {
626 		rwb->sync_cookie = stat;
627 		rwb->sync_issue = blk_stat_time(stat);
628 	}
629 }
630 
631 void wbt_requeue(struct rq_wb *rwb, struct blk_issue_stat *stat)
632 {
633 	if (!rwb_enabled(rwb))
634 		return;
635 	if (stat == rwb->sync_cookie) {
636 		rwb->sync_issue = 0;
637 		rwb->sync_cookie = NULL;
638 	}
639 }
640 
641 void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth)
642 {
643 	if (rwb) {
644 		rwb->queue_depth = depth;
645 		wbt_update_limits(rwb);
646 	}
647 }
648 
649 void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on)
650 {
651 	if (rwb)
652 		rwb->wc = write_cache_on;
653 }
654 
655 /*
656  * Disable wbt, if enabled by default.
657  */
658 void wbt_disable_default(struct request_queue *q)
659 {
660 	struct rq_wb *rwb = q->rq_wb;
661 
662 	if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT)
663 		wbt_exit(q);
664 }
665 EXPORT_SYMBOL_GPL(wbt_disable_default);
666 
667 /*
668  * Enable wbt if defaults are configured that way
669  */
670 void wbt_enable_default(struct request_queue *q)
671 {
672 	/* Throttling already enabled? */
673 	if (q->rq_wb)
674 		return;
675 
676 	/* Queue not registered? Maybe shutting down... */
677 	if (!test_bit(QUEUE_FLAG_REGISTERED, &q->queue_flags))
678 		return;
679 
680 	if ((q->mq_ops && IS_ENABLED(CONFIG_BLK_WBT_MQ)) ||
681 	    (q->request_fn && IS_ENABLED(CONFIG_BLK_WBT_SQ)))
682 		wbt_init(q);
683 }
684 EXPORT_SYMBOL_GPL(wbt_enable_default);
685 
686 u64 wbt_default_latency_nsec(struct request_queue *q)
687 {
688 	/*
689 	 * We default to 2msec for non-rotational storage, and 75msec
690 	 * for rotational storage.
691 	 */
692 	if (blk_queue_nonrot(q))
693 		return 2000000ULL;
694 	else
695 		return 75000000ULL;
696 }
697 
698 static int wbt_data_dir(const struct request *rq)
699 {
700 	const int op = req_op(rq);
701 
702 	if (op == REQ_OP_READ)
703 		return READ;
704 	else if (op == REQ_OP_WRITE || op == REQ_OP_FLUSH)
705 		return WRITE;
706 
707 	/* don't account */
708 	return -1;
709 }
710 
711 int wbt_init(struct request_queue *q)
712 {
713 	struct rq_wb *rwb;
714 	int i;
715 
716 	BUILD_BUG_ON(WBT_NR_BITS > BLK_STAT_RES_BITS);
717 
718 	rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
719 	if (!rwb)
720 		return -ENOMEM;
721 
722 	rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
723 	if (!rwb->cb) {
724 		kfree(rwb);
725 		return -ENOMEM;
726 	}
727 
728 	for (i = 0; i < WBT_NUM_RWQ; i++) {
729 		atomic_set(&rwb->rq_wait[i].inflight, 0);
730 		init_waitqueue_head(&rwb->rq_wait[i].wait);
731 	}
732 
733 	rwb->last_comp = rwb->last_issue = jiffies;
734 	rwb->queue = q;
735 	rwb->win_nsec = RWB_WINDOW_NSEC;
736 	rwb->enable_state = WBT_STATE_ON_DEFAULT;
737 	wbt_update_limits(rwb);
738 
739 	/*
740 	 * Assign rwb and add the stats callback.
741 	 */
742 	q->rq_wb = rwb;
743 	blk_stat_add_callback(q, rwb->cb);
744 
745 	rwb->min_lat_nsec = wbt_default_latency_nsec(q);
746 
747 	wbt_set_queue_depth(rwb, blk_queue_depth(q));
748 	wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags));
749 
750 	return 0;
751 }
752 
753 void wbt_exit(struct request_queue *q)
754 {
755 	struct rq_wb *rwb = q->rq_wb;
756 
757 	if (rwb) {
758 		blk_stat_remove_callback(q, rwb->cb);
759 		blk_stat_free_callback(rwb->cb);
760 		q->rq_wb = NULL;
761 		kfree(rwb);
762 	}
763 }
764