Linux Audio

Check our new training course

Loading...
v4.17
 
  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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 32enum {
 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
 56static 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 */
 65static 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
 83static 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 */
 97static 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
104static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb, bool is_kswapd)
 
105{
106	return &rwb->rq_wait[is_kswapd];
 
 
 
 
 
107}
108
109static 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
121void __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 */
168void 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 */
191static 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
249static 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
261static 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
272enum {
273	LAT_OK = 1,
274	LAT_UNKNOWN,
275	LAT_UNKNOWN_WRITES,
276	LAT_EXCEEDED,
277};
278
279static 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
331static 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
339static 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 */
361static 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
382static 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
404static 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
459void 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
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
468static 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
478static 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
504static 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 */
532static 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
560static 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 */
585enum 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
612void 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
631void 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
641void 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
649void 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 */
658void 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}
665EXPORT_SYMBOL_GPL(wbt_disable_default);
666
667/*
668 * Enable wbt if defaults are configured that way
669 */
670void 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}
684EXPORT_SYMBOL_GPL(wbt_enable_default);
685
686u64 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
698static 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
711int 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
753void 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}
v6.2
  1// SPDX-License-Identifier: GPL-2.0
  2/*
  3 * buffered writeback throttling. loosely based on CoDel. We can't drop
  4 * packets for IO scheduling, so the logic is something like this:
  5 *
  6 * - Monitor latencies in a defined window of time.
  7 * - If the minimum latency in the above window exceeds some target, increment
  8 *   scaling step and scale down queue depth by a factor of 2x. The monitoring
  9 *   window is then shrunk to 100 / sqrt(scaling step + 1).
 10 * - For any window where we don't have solid data on what the latencies
 11 *   look like, retain status quo.
 12 * - If latencies look good, decrement scaling step.
 13 * - If we're only doing writes, allow the scaling step to go negative. This
 14 *   will temporarily boost write performance, snapping back to a stable
 15 *   scaling step of 0 if reads show up or the heavy writers finish. Unlike
 16 *   positive scaling steps where we shrink the monitoring window, a negative
 17 *   scaling step retains the default step==0 window size.
 18 *
 19 * Copyright (C) 2016 Jens Axboe
 20 *
 21 */
 22#include <linux/kernel.h>
 23#include <linux/blk_types.h>
 24#include <linux/slab.h>
 25#include <linux/backing-dev.h>
 26#include <linux/swap.h>
 27
 28#include "blk-wbt.h"
 29#include "blk-rq-qos.h"
 30#include "elevator.h"
 31
 32#define CREATE_TRACE_POINTS
 33#include <trace/events/wbt.h>
 34
 35static inline void wbt_clear_state(struct request *rq)
 36{
 37	rq->wbt_flags = 0;
 38}
 39
 40static inline enum wbt_flags wbt_flags(struct request *rq)
 41{
 42	return rq->wbt_flags;
 43}
 44
 45static inline bool wbt_is_tracked(struct request *rq)
 46{
 47	return rq->wbt_flags & WBT_TRACKED;
 48}
 49
 50static inline bool wbt_is_read(struct request *rq)
 51{
 52	return rq->wbt_flags & WBT_READ;
 53}
 54
 55enum {
 56	/*
 57	 * Default setting, we'll scale up (to 75% of QD max) or down (min 1)
 58	 * from here depending on device stats
 59	 */
 60	RWB_DEF_DEPTH	= 16,
 61
 62	/*
 63	 * 100msec window
 64	 */
 65	RWB_WINDOW_NSEC		= 100 * 1000 * 1000ULL,
 66
 67	/*
 68	 * Disregard stats, if we don't meet this minimum
 69	 */
 70	RWB_MIN_WRITE_SAMPLES	= 3,
 71
 72	/*
 73	 * If we have this number of consecutive windows with not enough
 74	 * information to scale up or down, scale up.
 75	 */
 76	RWB_UNKNOWN_BUMP	= 5,
 77};
 78
 79static inline bool rwb_enabled(struct rq_wb *rwb)
 80{
 81	return rwb && rwb->enable_state != WBT_STATE_OFF_DEFAULT &&
 82		      rwb->wb_normal != 0;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 83}
 84
 85static void wb_timestamp(struct rq_wb *rwb, unsigned long *var)
 86{
 87	if (rwb_enabled(rwb)) {
 88		const unsigned long cur = jiffies;
 89
 90		if (cur != *var)
 91			*var = cur;
 92	}
 93}
 94
 95/*
 96 * If a task was rate throttled in balance_dirty_pages() within the last
 97 * second or so, use that to indicate a higher cleaning rate.
 98 */
 99static bool wb_recent_wait(struct rq_wb *rwb)
100{
101	struct bdi_writeback *wb = &rwb->rqos.q->disk->bdi->wb;
102
103	return time_before(jiffies, wb->dirty_sleep + HZ);
104}
105
106static inline struct rq_wait *get_rq_wait(struct rq_wb *rwb,
107					  enum wbt_flags wb_acct)
108{
109	if (wb_acct & WBT_KSWAPD)
110		return &rwb->rq_wait[WBT_RWQ_KSWAPD];
111	else if (wb_acct & WBT_DISCARD)
112		return &rwb->rq_wait[WBT_RWQ_DISCARD];
113
114	return &rwb->rq_wait[WBT_RWQ_BG];
115}
116
117static void rwb_wake_all(struct rq_wb *rwb)
118{
119	int i;
120
121	for (i = 0; i < WBT_NUM_RWQ; i++) {
122		struct rq_wait *rqw = &rwb->rq_wait[i];
123
124		if (wq_has_sleeper(&rqw->wait))
125			wake_up_all(&rqw->wait);
126	}
127}
128
129static void wbt_rqw_done(struct rq_wb *rwb, struct rq_wait *rqw,
130			 enum wbt_flags wb_acct)
131{
 
132	int inflight, limit;
133
 
 
 
 
134	inflight = atomic_dec_return(&rqw->inflight);
135
136	/*
137	 * wbt got disabled with IO in flight. Wake up any potential
138	 * waiters, we don't have to do more than that.
139	 */
140	if (unlikely(!rwb_enabled(rwb))) {
141		rwb_wake_all(rwb);
142		return;
143	}
144
145	/*
146	 * For discards, our limit is always the background. For writes, if
147	 * the device does write back caching, drop further down before we
148	 * wake people up.
149	 */
150	if (wb_acct & WBT_DISCARD)
151		limit = rwb->wb_background;
152	else if (rwb->wc && !wb_recent_wait(rwb))
153		limit = 0;
154	else
155		limit = rwb->wb_normal;
156
157	/*
158	 * Don't wake anyone up if we are above the normal limit.
159	 */
160	if (inflight && inflight >= limit)
161		return;
162
163	if (wq_has_sleeper(&rqw->wait)) {
164		int diff = limit - inflight;
165
166		if (!inflight || diff >= rwb->wb_background / 2)
167			wake_up_all(&rqw->wait);
168	}
169}
170
171static void __wbt_done(struct rq_qos *rqos, enum wbt_flags wb_acct)
172{
173	struct rq_wb *rwb = RQWB(rqos);
174	struct rq_wait *rqw;
175
176	if (!(wb_acct & WBT_TRACKED))
177		return;
178
179	rqw = get_rq_wait(rwb, wb_acct);
180	wbt_rqw_done(rwb, rqw, wb_acct);
181}
182
183/*
184 * Called on completion of a request. Note that it's also called when
185 * a request is merged, when the request gets freed.
186 */
187static void wbt_done(struct rq_qos *rqos, struct request *rq)
188{
189	struct rq_wb *rwb = RQWB(rqos);
 
190
191	if (!wbt_is_tracked(rq)) {
192		if (rwb->sync_cookie == rq) {
193			rwb->sync_issue = 0;
194			rwb->sync_cookie = NULL;
195		}
196
197		if (wbt_is_read(rq))
198			wb_timestamp(rwb, &rwb->last_comp);
199	} else {
200		WARN_ON_ONCE(rq == rwb->sync_cookie);
201		__wbt_done(rqos, wbt_flags(rq));
202	}
203	wbt_clear_state(rq);
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
204}
205
206static inline bool stat_sample_valid(struct blk_rq_stat *stat)
207{
208	/*
209	 * We need at least one read sample, and a minimum of
210	 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
211	 * that it's writes impacting us, and not just some sole read on
212	 * a device that is in a lower power state.
213	 */
214	return (stat[READ].nr_samples >= 1 &&
215		stat[WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES);
216}
217
218static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
219{
220	u64 now, issue = READ_ONCE(rwb->sync_issue);
221
222	if (!issue || !rwb->sync_cookie)
223		return 0;
224
225	now = ktime_to_ns(ktime_get());
226	return now - issue;
227}
228
229enum {
230	LAT_OK = 1,
231	LAT_UNKNOWN,
232	LAT_UNKNOWN_WRITES,
233	LAT_EXCEEDED,
234};
235
236static int latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
237{
238	struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi;
239	struct rq_depth *rqd = &rwb->rq_depth;
240	u64 thislat;
241
242	/*
243	 * If our stored sync issue exceeds the window size, or it
244	 * exceeds our min target AND we haven't logged any entries,
245	 * flag the latency as exceeded. wbt works off completion latencies,
246	 * but for a flooded device, a single sync IO can take a long time
247	 * to complete after being issued. If this time exceeds our
248	 * monitoring window AND we didn't see any other completions in that
249	 * window, then count that sync IO as a violation of the latency.
250	 */
251	thislat = rwb_sync_issue_lat(rwb);
252	if (thislat > rwb->cur_win_nsec ||
253	    (thislat > rwb->min_lat_nsec && !stat[READ].nr_samples)) {
254		trace_wbt_lat(bdi, thislat);
255		return LAT_EXCEEDED;
256	}
257
258	/*
259	 * No read/write mix, if stat isn't valid
260	 */
261	if (!stat_sample_valid(stat)) {
262		/*
263		 * If we had writes in this stat window and the window is
264		 * current, we're only doing writes. If a task recently
265		 * waited or still has writes in flights, consider us doing
266		 * just writes as well.
267		 */
268		if (stat[WRITE].nr_samples || wb_recent_wait(rwb) ||
269		    wbt_inflight(rwb))
270			return LAT_UNKNOWN_WRITES;
271		return LAT_UNKNOWN;
272	}
273
274	/*
275	 * If the 'min' latency exceeds our target, step down.
276	 */
277	if (stat[READ].min > rwb->min_lat_nsec) {
278		trace_wbt_lat(bdi, stat[READ].min);
279		trace_wbt_stat(bdi, stat);
280		return LAT_EXCEEDED;
281	}
282
283	if (rqd->scale_step)
284		trace_wbt_stat(bdi, stat);
285
286	return LAT_OK;
287}
288
289static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
290{
291	struct backing_dev_info *bdi = rwb->rqos.q->disk->bdi;
292	struct rq_depth *rqd = &rwb->rq_depth;
293
294	trace_wbt_step(bdi, msg, rqd->scale_step, rwb->cur_win_nsec,
295			rwb->wb_background, rwb->wb_normal, rqd->max_depth);
296}
297
298static void calc_wb_limits(struct rq_wb *rwb)
299{
300	if (rwb->min_lat_nsec == 0) {
301		rwb->wb_normal = rwb->wb_background = 0;
302	} else if (rwb->rq_depth.max_depth <= 2) {
303		rwb->wb_normal = rwb->rq_depth.max_depth;
304		rwb->wb_background = 1;
305	} else {
306		rwb->wb_normal = (rwb->rq_depth.max_depth + 1) / 2;
307		rwb->wb_background = (rwb->rq_depth.max_depth + 3) / 4;
308	}
309}
310
311static void scale_up(struct rq_wb *rwb)
312{
313	if (!rq_depth_scale_up(&rwb->rq_depth))
 
 
 
314		return;
315	calc_wb_limits(rwb);
 
316	rwb->unknown_cnt = 0;
 
 
 
317	rwb_wake_all(rwb);
318	rwb_trace_step(rwb, tracepoint_string("scale up"));
 
319}
320
 
 
 
 
321static void scale_down(struct rq_wb *rwb, bool hard_throttle)
322{
323	if (!rq_depth_scale_down(&rwb->rq_depth, hard_throttle))
 
 
 
 
 
324		return;
 
 
 
 
 
 
 
 
325	calc_wb_limits(rwb);
326	rwb->unknown_cnt = 0;
327	rwb_trace_step(rwb, tracepoint_string("scale down"));
328}
329
330static void rwb_arm_timer(struct rq_wb *rwb)
331{
332	struct rq_depth *rqd = &rwb->rq_depth;
333
334	if (rqd->scale_step > 0) {
335		/*
336		 * We should speed this up, using some variant of a fast
337		 * integer inverse square root calculation. Since we only do
338		 * this for every window expiration, it's not a huge deal,
339		 * though.
340		 */
341		rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
342					int_sqrt((rqd->scale_step + 1) << 8));
343	} else {
344		/*
345		 * For step < 0, we don't want to increase/decrease the
346		 * window size.
347		 */
348		rwb->cur_win_nsec = rwb->win_nsec;
349	}
350
351	blk_stat_activate_nsecs(rwb->cb, rwb->cur_win_nsec);
352}
353
354static void wb_timer_fn(struct blk_stat_callback *cb)
355{
356	struct rq_wb *rwb = cb->data;
357	struct rq_depth *rqd = &rwb->rq_depth;
358	unsigned int inflight = wbt_inflight(rwb);
359	int status;
360
361	if (!rwb->rqos.q->disk)
362		return;
363
364	status = latency_exceeded(rwb, cb->stat);
365
366	trace_wbt_timer(rwb->rqos.q->disk->bdi, status, rqd->scale_step,
367			inflight);
368
369	/*
370	 * If we exceeded the latency target, step down. If we did not,
371	 * step one level up. If we don't know enough to say either exceeded
372	 * or ok, then don't do anything.
373	 */
374	switch (status) {
375	case LAT_EXCEEDED:
376		scale_down(rwb, true);
377		break;
378	case LAT_OK:
379		scale_up(rwb);
380		break;
381	case LAT_UNKNOWN_WRITES:
382		/*
383		 * We started a the center step, but don't have a valid
384		 * read/write sample, but we do have writes going on.
385		 * Allow step to go negative, to increase write perf.
386		 */
387		scale_up(rwb);
388		break;
389	case LAT_UNKNOWN:
390		if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
391			break;
392		/*
393		 * We get here when previously scaled reduced depth, and we
394		 * currently don't have a valid read/write sample. For that
395		 * case, slowly return to center state (step == 0).
396		 */
397		if (rqd->scale_step > 0)
398			scale_up(rwb);
399		else if (rqd->scale_step < 0)
400			scale_down(rwb, false);
401		break;
402	default:
403		break;
404	}
405
406	/*
407	 * Re-arm timer, if we have IO in flight
408	 */
409	if (rqd->scale_step || inflight)
410		rwb_arm_timer(rwb);
411}
412
413static void wbt_update_limits(struct rq_wb *rwb)
414{
415	struct rq_depth *rqd = &rwb->rq_depth;
416
417	rqd->scale_step = 0;
418	rqd->scaled_max = false;
419
420	rq_depth_calc_max_depth(rqd);
421	calc_wb_limits(rwb);
422
423	rwb_wake_all(rwb);
424}
425
426bool wbt_disabled(struct request_queue *q)
427{
428	struct rq_qos *rqos = wbt_rq_qos(q);
429
430	return !rqos || RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT ||
431	       RQWB(rqos)->enable_state == WBT_STATE_OFF_MANUAL;
432}
433
434u64 wbt_get_min_lat(struct request_queue *q)
435{
436	struct rq_qos *rqos = wbt_rq_qos(q);
437	if (!rqos)
438		return 0;
439	return RQWB(rqos)->min_lat_nsec;
440}
441
442void wbt_set_min_lat(struct request_queue *q, u64 val)
443{
444	struct rq_qos *rqos = wbt_rq_qos(q);
445	if (!rqos)
446		return;
447
448	RQWB(rqos)->min_lat_nsec = val;
449	if (val)
450		RQWB(rqos)->enable_state = WBT_STATE_ON_MANUAL;
451	else
452		RQWB(rqos)->enable_state = WBT_STATE_OFF_MANUAL;
453
454	wbt_update_limits(RQWB(rqos));
455}
456
457
458static bool close_io(struct rq_wb *rwb)
459{
460	const unsigned long now = jiffies;
461
462	return time_before(now, rwb->last_issue + HZ / 10) ||
463		time_before(now, rwb->last_comp + HZ / 10);
464}
465
466#define REQ_HIPRIO	(REQ_SYNC | REQ_META | REQ_PRIO)
467
468static inline unsigned int get_limit(struct rq_wb *rwb, blk_opf_t opf)
469{
470	unsigned int limit;
471
472	/*
473	 * If we got disabled, just return UINT_MAX. This ensures that
474	 * we'll properly inc a new IO, and dec+wakeup at the end.
475	 */
476	if (!rwb_enabled(rwb))
477		return UINT_MAX;
478
479	if ((opf & REQ_OP_MASK) == REQ_OP_DISCARD)
480		return rwb->wb_background;
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 ((opf & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
491		limit = rwb->rq_depth.max_depth;
492	else if ((opf & 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
504struct wbt_wait_data {
505	struct rq_wb *rwb;
506	enum wbt_flags wb_acct;
507	blk_opf_t opf;
508};
 
 
 
 
 
 
 
509
510static bool wbt_inflight_cb(struct rq_wait *rqw, void *private_data)
511{
512	struct wbt_wait_data *data = private_data;
513	return rq_wait_inc_below(rqw, get_limit(data->rwb, data->opf));
514}
 
 
515
516static void wbt_cleanup_cb(struct rq_wait *rqw, void *private_data)
517{
518	struct wbt_wait_data *data = private_data;
519	wbt_rqw_done(data->rwb, rqw, data->wb_acct);
520}
521
522/*
523 * Block if we will exceed our limit, or if we are currently waiting for
524 * the timer to kick off queuing again.
525 */
526static void __wbt_wait(struct rq_wb *rwb, enum wbt_flags wb_acct,
527		       blk_opf_t opf)
 
528{
529	struct rq_wait *rqw = get_rq_wait(rwb, wb_acct);
530	struct wbt_wait_data data = {
531		.rwb = rwb,
532		.wb_acct = wb_acct,
533		.opf = opf,
534	};
 
 
 
 
 
 
535
536	rq_qos_wait(rqw, &data, wbt_inflight_cb, wbt_cleanup_cb);
 
 
 
 
 
 
 
 
537}
538
539static inline bool wbt_should_throttle(struct bio *bio)
540{
541	switch (bio_op(bio)) {
542	case REQ_OP_WRITE:
543		/*
544		 * Don't throttle WRITE_ODIRECT
545		 */
546		if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) ==
547		    (REQ_SYNC | REQ_IDLE))
548			return false;
549		fallthrough;
550	case REQ_OP_DISCARD:
551		return true;
552	default:
553		return false;
554	}
555}
556
557static enum wbt_flags bio_to_wbt_flags(struct rq_wb *rwb, struct bio *bio)
558{
559	enum wbt_flags flags = 0;
560
561	if (!rwb_enabled(rwb))
562		return 0;
563
564	if (bio_op(bio) == REQ_OP_READ) {
565		flags = WBT_READ;
566	} else if (wbt_should_throttle(bio)) {
567		if (current_is_kswapd())
568			flags |= WBT_KSWAPD;
569		if (bio_op(bio) == REQ_OP_DISCARD)
570			flags |= WBT_DISCARD;
571		flags |= WBT_TRACKED;
572	}
573	return flags;
574}
575
576static void wbt_cleanup(struct rq_qos *rqos, struct bio *bio)
577{
578	struct rq_wb *rwb = RQWB(rqos);
579	enum wbt_flags flags = bio_to_wbt_flags(rwb, bio);
580	__wbt_done(rqos, flags);
581}
582
583/*
 
584 * May sleep, if we have exceeded the writeback limits. Caller can pass
585 * in an irq held spinlock, if it holds one when calling this function.
586 * If we do sleep, we'll release and re-grab it.
587 */
588static void wbt_wait(struct rq_qos *rqos, struct bio *bio)
589{
590	struct rq_wb *rwb = RQWB(rqos);
591	enum wbt_flags flags;
592
593	flags = bio_to_wbt_flags(rwb, bio);
594	if (!(flags & WBT_TRACKED)) {
595		if (flags & WBT_READ)
 
 
 
 
 
596			wb_timestamp(rwb, &rwb->last_issue);
597		return;
598	}
599
600	__wbt_wait(rwb, flags, bio->bi_opf);
601
602	if (!blk_stat_is_active(rwb->cb))
603		rwb_arm_timer(rwb);
604}
605
606static void wbt_track(struct rq_qos *rqos, struct request *rq, struct bio *bio)
607{
608	struct rq_wb *rwb = RQWB(rqos);
609	rq->wbt_flags |= bio_to_wbt_flags(rwb, bio);
610}
611
612static void wbt_issue(struct rq_qos *rqos, struct request *rq)
613{
614	struct rq_wb *rwb = RQWB(rqos);
615
616	if (!rwb_enabled(rwb))
617		return;
618
619	/*
620	 * Track sync issue, in case it takes a long time to complete. Allows us
621	 * to react quicker, if a sync IO takes a long time to complete. Note
622	 * that this is just a hint. The request can go away when it completes,
623	 * so it's important we never dereference it. We only use the address to
624	 * compare with, which is why we store the sync_issue time locally.
 
625	 */
626	if (wbt_is_read(rq) && !rwb->sync_issue) {
627		rwb->sync_cookie = rq;
628		rwb->sync_issue = rq->io_start_time_ns;
629	}
630}
631
632static void wbt_requeue(struct rq_qos *rqos, struct request *rq)
633{
634	struct rq_wb *rwb = RQWB(rqos);
635	if (!rwb_enabled(rwb))
636		return;
637	if (rq == rwb->sync_cookie) {
638		rwb->sync_issue = 0;
639		rwb->sync_cookie = NULL;
640	}
641}
642
643void wbt_set_write_cache(struct request_queue *q, bool write_cache_on)
644{
645	struct rq_qos *rqos = wbt_rq_qos(q);
646	if (rqos)
647		RQWB(rqos)->wc = write_cache_on;
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
648}
 
649
650/*
651 * Enable wbt if defaults are configured that way
652 */
653void wbt_enable_default(struct request_queue *q)
654{
655	struct rq_qos *rqos;
656	bool disable_flag = q->elevator &&
657		    test_bit(ELEVATOR_FLAG_DISABLE_WBT, &q->elevator->flags);
658
659	/* Throttling already enabled? */
660	rqos = wbt_rq_qos(q);
661	if (rqos) {
662		if (!disable_flag &&
663		    RQWB(rqos)->enable_state == WBT_STATE_OFF_DEFAULT)
664			RQWB(rqos)->enable_state = WBT_STATE_ON_DEFAULT;
665		return;
666	}
667
668	/* Queue not registered? Maybe shutting down... */
669	if (!blk_queue_registered(q))
670		return;
671
672	if (queue_is_mq(q) && !disable_flag)
 
673		wbt_init(q);
674}
675EXPORT_SYMBOL_GPL(wbt_enable_default);
676
677u64 wbt_default_latency_nsec(struct request_queue *q)
678{
679	/*
680	 * We default to 2msec for non-rotational storage, and 75msec
681	 * for rotational storage.
682	 */
683	if (blk_queue_nonrot(q))
684		return 2000000ULL;
685	else
686		return 75000000ULL;
687}
688
689static int wbt_data_dir(const struct request *rq)
690{
691	const enum req_op op = req_op(rq);
692
693	if (op == REQ_OP_READ)
694		return READ;
695	else if (op_is_write(op))
696		return WRITE;
697
698	/* don't account */
699	return -1;
700}
701
702static void wbt_queue_depth_changed(struct rq_qos *rqos)
703{
704	RQWB(rqos)->rq_depth.queue_depth = blk_queue_depth(rqos->q);
705	wbt_update_limits(RQWB(rqos));
706}
707
708static void wbt_exit(struct rq_qos *rqos)
709{
710	struct rq_wb *rwb = RQWB(rqos);
711	struct request_queue *q = rqos->q;
712
713	blk_stat_remove_callback(q, rwb->cb);
714	blk_stat_free_callback(rwb->cb);
715	kfree(rwb);
716}
717
718/*
719 * Disable wbt, if enabled by default.
720 */
721void wbt_disable_default(struct request_queue *q)
722{
723	struct rq_qos *rqos = wbt_rq_qos(q);
724	struct rq_wb *rwb;
725	if (!rqos)
726		return;
727	rwb = RQWB(rqos);
728	if (rwb->enable_state == WBT_STATE_ON_DEFAULT) {
729		blk_stat_deactivate(rwb->cb);
730		rwb->enable_state = WBT_STATE_OFF_DEFAULT;
731	}
732}
733EXPORT_SYMBOL_GPL(wbt_disable_default);
734
735#ifdef CONFIG_BLK_DEBUG_FS
736static int wbt_curr_win_nsec_show(void *data, struct seq_file *m)
737{
738	struct rq_qos *rqos = data;
739	struct rq_wb *rwb = RQWB(rqos);
740
741	seq_printf(m, "%llu\n", rwb->cur_win_nsec);
742	return 0;
743}
744
745static int wbt_enabled_show(void *data, struct seq_file *m)
746{
747	struct rq_qos *rqos = data;
748	struct rq_wb *rwb = RQWB(rqos);
749
750	seq_printf(m, "%d\n", rwb->enable_state);
751	return 0;
752}
753
754static int wbt_id_show(void *data, struct seq_file *m)
755{
756	struct rq_qos *rqos = data;
757
758	seq_printf(m, "%u\n", rqos->id);
759	return 0;
760}
761
762static int wbt_inflight_show(void *data, struct seq_file *m)
763{
764	struct rq_qos *rqos = data;
765	struct rq_wb *rwb = RQWB(rqos);
766	int i;
767
768	for (i = 0; i < WBT_NUM_RWQ; i++)
769		seq_printf(m, "%d: inflight %d\n", i,
770			   atomic_read(&rwb->rq_wait[i].inflight));
771	return 0;
772}
773
774static int wbt_min_lat_nsec_show(void *data, struct seq_file *m)
775{
776	struct rq_qos *rqos = data;
777	struct rq_wb *rwb = RQWB(rqos);
778
779	seq_printf(m, "%lu\n", rwb->min_lat_nsec);
780	return 0;
781}
782
783static int wbt_unknown_cnt_show(void *data, struct seq_file *m)
784{
785	struct rq_qos *rqos = data;
786	struct rq_wb *rwb = RQWB(rqos);
787
788	seq_printf(m, "%u\n", rwb->unknown_cnt);
789	return 0;
790}
791
792static int wbt_normal_show(void *data, struct seq_file *m)
793{
794	struct rq_qos *rqos = data;
795	struct rq_wb *rwb = RQWB(rqos);
796
797	seq_printf(m, "%u\n", rwb->wb_normal);
798	return 0;
799}
800
801static int wbt_background_show(void *data, struct seq_file *m)
802{
803	struct rq_qos *rqos = data;
804	struct rq_wb *rwb = RQWB(rqos);
805
806	seq_printf(m, "%u\n", rwb->wb_background);
807	return 0;
808}
809
810static const struct blk_mq_debugfs_attr wbt_debugfs_attrs[] = {
811	{"curr_win_nsec", 0400, wbt_curr_win_nsec_show},
812	{"enabled", 0400, wbt_enabled_show},
813	{"id", 0400, wbt_id_show},
814	{"inflight", 0400, wbt_inflight_show},
815	{"min_lat_nsec", 0400, wbt_min_lat_nsec_show},
816	{"unknown_cnt", 0400, wbt_unknown_cnt_show},
817	{"wb_normal", 0400, wbt_normal_show},
818	{"wb_background", 0400, wbt_background_show},
819	{},
820};
821#endif
822
823static struct rq_qos_ops wbt_rqos_ops = {
824	.throttle = wbt_wait,
825	.issue = wbt_issue,
826	.track = wbt_track,
827	.requeue = wbt_requeue,
828	.done = wbt_done,
829	.cleanup = wbt_cleanup,
830	.queue_depth_changed = wbt_queue_depth_changed,
831	.exit = wbt_exit,
832#ifdef CONFIG_BLK_DEBUG_FS
833	.debugfs_attrs = wbt_debugfs_attrs,
834#endif
835};
836
837int wbt_init(struct request_queue *q)
838{
839	struct rq_wb *rwb;
840	int i;
841	int ret;
842
843	rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
844	if (!rwb)
845		return -ENOMEM;
846
847	rwb->cb = blk_stat_alloc_callback(wb_timer_fn, wbt_data_dir, 2, rwb);
848	if (!rwb->cb) {
849		kfree(rwb);
850		return -ENOMEM;
851	}
852
853	for (i = 0; i < WBT_NUM_RWQ; i++)
854		rq_wait_init(&rwb->rq_wait[i]);
 
 
855
856	rwb->rqos.id = RQ_QOS_WBT;
857	rwb->rqos.ops = &wbt_rqos_ops;
858	rwb->rqos.q = q;
859	rwb->last_comp = rwb->last_issue = jiffies;
 
860	rwb->win_nsec = RWB_WINDOW_NSEC;
861	rwb->enable_state = WBT_STATE_ON_DEFAULT;
862	rwb->wc = test_bit(QUEUE_FLAG_WC, &q->queue_flags);
863	rwb->rq_depth.default_depth = RWB_DEF_DEPTH;
864	rwb->min_lat_nsec = wbt_default_latency_nsec(q);
865
866	wbt_queue_depth_changed(&rwb->rqos);
867
868	/*
869	 * Assign rwb and add the stats callback.
870	 */
871	ret = rq_qos_add(q, &rwb->rqos);
872	if (ret)
873		goto err_free;
874
875	blk_stat_add_callback(q, rwb->cb);
 
 
 
876
877	return 0;
 
878
879err_free:
880	blk_stat_free_callback(rwb->cb);
881	kfree(rwb);
882	return ret;
883
 
 
 
 
 
 
884}