Linux Audio

Check our new training course

Loading...
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}
v4.10.11
 
  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		wbt_clear_state(stat);
182	} else {
183		WARN_ON_ONCE(stat == rwb->sync_cookie);
184		__wbt_done(rwb, wbt_stat_to_mask(stat));
185		wbt_clear_state(stat);
186	}
187}
188
189/*
190 * Return true, if we can't increase the depth further by scaling
191 */
192static bool calc_wb_limits(struct rq_wb *rwb)
193{
194	unsigned int depth;
195	bool ret = false;
196
197	if (!rwb->min_lat_nsec) {
198		rwb->wb_max = rwb->wb_normal = rwb->wb_background = 0;
199		return false;
200	}
201
202	/*
203	 * For QD=1 devices, this is a special case. It's important for those
204	 * to have one request ready when one completes, so force a depth of
205	 * 2 for those devices. On the backend, it'll be a depth of 1 anyway,
206	 * since the device can't have more than that in flight. If we're
207	 * scaling down, then keep a setting of 1/1/1.
208	 */
209	if (rwb->queue_depth == 1) {
210		if (rwb->scale_step > 0)
211			rwb->wb_max = rwb->wb_normal = 1;
212		else {
213			rwb->wb_max = rwb->wb_normal = 2;
214			ret = true;
215		}
216		rwb->wb_background = 1;
217	} else {
218		/*
219		 * scale_step == 0 is our default state. If we have suffered
220		 * latency spikes, step will be > 0, and we shrink the
221		 * allowed write depths. If step is < 0, we're only doing
222		 * writes, and we allow a temporarily higher depth to
223		 * increase performance.
224		 */
225		depth = min_t(unsigned int, RWB_DEF_DEPTH, rwb->queue_depth);
226		if (rwb->scale_step > 0)
227			depth = 1 + ((depth - 1) >> min(31, rwb->scale_step));
228		else if (rwb->scale_step < 0) {
229			unsigned int maxd = 3 * rwb->queue_depth / 4;
230
231			depth = 1 + ((depth - 1) << -rwb->scale_step);
232			if (depth > maxd) {
233				depth = maxd;
234				ret = true;
235			}
236		}
237
238		/*
239		 * Set our max/normal/bg queue depths based on how far
240		 * we have scaled down (->scale_step).
241		 */
242		rwb->wb_max = depth;
243		rwb->wb_normal = (rwb->wb_max + 1) / 2;
244		rwb->wb_background = (rwb->wb_max + 3) / 4;
245	}
246
247	return ret;
248}
249
250static inline bool stat_sample_valid(struct blk_rq_stat *stat)
251{
252	/*
253	 * We need at least one read sample, and a minimum of
254	 * RWB_MIN_WRITE_SAMPLES. We require some write samples to know
255	 * that it's writes impacting us, and not just some sole read on
256	 * a device that is in a lower power state.
257	 */
258	return stat[BLK_STAT_READ].nr_samples >= 1 &&
259		stat[BLK_STAT_WRITE].nr_samples >= RWB_MIN_WRITE_SAMPLES;
260}
261
262static u64 rwb_sync_issue_lat(struct rq_wb *rwb)
263{
264	u64 now, issue = ACCESS_ONCE(rwb->sync_issue);
265
266	if (!issue || !rwb->sync_cookie)
267		return 0;
268
269	now = ktime_to_ns(ktime_get());
270	return now - issue;
271}
272
273enum {
274	LAT_OK = 1,
275	LAT_UNKNOWN,
276	LAT_UNKNOWN_WRITES,
277	LAT_EXCEEDED,
278};
279
280static int __latency_exceeded(struct rq_wb *rwb, struct blk_rq_stat *stat)
281{
282	struct backing_dev_info *bdi = &rwb->queue->backing_dev_info;
 
283	u64 thislat;
284
285	/*
286	 * If our stored sync issue exceeds the window size, or it
287	 * exceeds our min target AND we haven't logged any entries,
288	 * flag the latency as exceeded. wbt works off completion latencies,
289	 * but for a flooded device, a single sync IO can take a long time
290	 * to complete after being issued. If this time exceeds our
291	 * monitoring window AND we didn't see any other completions in that
292	 * window, then count that sync IO as a violation of the latency.
293	 */
294	thislat = rwb_sync_issue_lat(rwb);
295	if (thislat > rwb->cur_win_nsec ||
296	    (thislat > rwb->min_lat_nsec && !stat[BLK_STAT_READ].nr_samples)) {
297		trace_wbt_lat(bdi, thislat);
298		return LAT_EXCEEDED;
299	}
300
301	/*
302	 * No read/write mix, if stat isn't valid
303	 */
304	if (!stat_sample_valid(stat)) {
305		/*
306		 * If we had writes in this stat window and the window is
307		 * current, we're only doing writes. If a task recently
308		 * waited or still has writes in flights, consider us doing
309		 * just writes as well.
310		 */
311		if ((stat[BLK_STAT_WRITE].nr_samples && blk_stat_is_current(stat)) ||
312		    wb_recent_wait(rwb) || wbt_inflight(rwb))
313			return LAT_UNKNOWN_WRITES;
314		return LAT_UNKNOWN;
315	}
316
317	/*
318	 * If the 'min' latency exceeds our target, step down.
319	 */
320	if (stat[BLK_STAT_READ].min > rwb->min_lat_nsec) {
321		trace_wbt_lat(bdi, stat[BLK_STAT_READ].min);
322		trace_wbt_stat(bdi, stat);
323		return LAT_EXCEEDED;
324	}
325
326	if (rwb->scale_step)
327		trace_wbt_stat(bdi, stat);
328
329	return LAT_OK;
330}
331
332static int latency_exceeded(struct rq_wb *rwb)
333{
334	struct blk_rq_stat stat[2];
 
335
336	blk_queue_stat_get(rwb->queue, stat);
337	return __latency_exceeded(rwb, stat);
338}
339
340static void rwb_trace_step(struct rq_wb *rwb, const char *msg)
341{
342	struct backing_dev_info *bdi = &rwb->queue->backing_dev_info;
343
344	trace_wbt_step(bdi, msg, rwb->scale_step, rwb->cur_win_nsec,
345			rwb->wb_background, rwb->wb_normal, rwb->wb_max);
 
 
 
 
 
346}
347
348static void scale_up(struct rq_wb *rwb)
349{
350	/*
351	 * Hit max in previous round, stop here
352	 */
353	if (rwb->scaled_max)
354		return;
355
356	rwb->scale_step--;
357	rwb->unknown_cnt = 0;
358	blk_stat_clear(rwb->queue);
359
360	rwb->scaled_max = calc_wb_limits(rwb);
361
362	rwb_wake_all(rwb);
363
364	rwb_trace_step(rwb, "step up");
365}
366
367/*
368 * Scale rwb down. If 'hard_throttle' is set, do it quicker, since we
369 * had a latency violation.
370 */
371static void scale_down(struct rq_wb *rwb, bool hard_throttle)
372{
373	/*
374	 * Stop scaling down when we've hit the limit. This also prevents
375	 * ->scale_step from going to crazy values, if the device can't
376	 * keep up.
377	 */
378	if (rwb->wb_max == 1)
379		return;
380
381	if (rwb->scale_step < 0 && hard_throttle)
382		rwb->scale_step = 0;
383	else
384		rwb->scale_step++;
385
386	rwb->scaled_max = false;
387	rwb->unknown_cnt = 0;
388	blk_stat_clear(rwb->queue);
389	calc_wb_limits(rwb);
390	rwb_trace_step(rwb, "step down");
391}
392
393static void rwb_arm_timer(struct rq_wb *rwb)
394{
395	unsigned long expires;
396
397	if (rwb->scale_step > 0) {
398		/*
399		 * We should speed this up, using some variant of a fast
400		 * integer inverse square root calculation. Since we only do
401		 * this for every window expiration, it's not a huge deal,
402		 * though.
403		 */
404		rwb->cur_win_nsec = div_u64(rwb->win_nsec << 4,
405					int_sqrt((rwb->scale_step + 1) << 8));
406	} else {
407		/*
408		 * For step < 0, we don't want to increase/decrease the
409		 * window size.
410		 */
411		rwb->cur_win_nsec = rwb->win_nsec;
412	}
413
414	expires = jiffies + nsecs_to_jiffies(rwb->cur_win_nsec);
415	mod_timer(&rwb->window_timer, expires);
416}
417
418static void wb_timer_fn(unsigned long data)
419{
420	struct rq_wb *rwb = (struct rq_wb *) data;
 
421	unsigned int inflight = wbt_inflight(rwb);
422	int status;
423
424	status = latency_exceeded(rwb);
 
 
 
425
426	trace_wbt_timer(&rwb->queue->backing_dev_info, status, rwb->scale_step,
427			inflight);
428
429	/*
430	 * If we exceeded the latency target, step down. If we did not,
431	 * step one level up. If we don't know enough to say either exceeded
432	 * or ok, then don't do anything.
433	 */
434	switch (status) {
435	case LAT_EXCEEDED:
436		scale_down(rwb, true);
437		break;
438	case LAT_OK:
439		scale_up(rwb);
440		break;
441	case LAT_UNKNOWN_WRITES:
442		/*
443		 * We started a the center step, but don't have a valid
444		 * read/write sample, but we do have writes going on.
445		 * Allow step to go negative, to increase write perf.
446		 */
447		scale_up(rwb);
448		break;
449	case LAT_UNKNOWN:
450		if (++rwb->unknown_cnt < RWB_UNKNOWN_BUMP)
451			break;
452		/*
453		 * We get here when previously scaled reduced depth, and we
454		 * currently don't have a valid read/write sample. For that
455		 * case, slowly return to center state (step == 0).
456		 */
457		if (rwb->scale_step > 0)
458			scale_up(rwb);
459		else if (rwb->scale_step < 0)
460			scale_down(rwb, false);
461		break;
462	default:
463		break;
464	}
465
466	/*
467	 * Re-arm timer, if we have IO in flight
468	 */
469	if (rwb->scale_step || inflight)
470		rwb_arm_timer(rwb);
471}
472
473void wbt_update_limits(struct rq_wb *rwb)
474{
475	rwb->scale_step = 0;
476	rwb->scaled_max = false;
 
 
 
 
477	calc_wb_limits(rwb);
478
479	rwb_wake_all(rwb);
480}
481
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
482static bool close_io(struct rq_wb *rwb)
483{
484	const unsigned long now = jiffies;
485
486	return time_before(now, rwb->last_issue + HZ / 10) ||
487		time_before(now, rwb->last_comp + HZ / 10);
488}
489
490#define REQ_HIPRIO	(REQ_SYNC | REQ_META | REQ_PRIO)
491
492static inline unsigned int get_limit(struct rq_wb *rwb, unsigned long rw)
493{
494	unsigned int limit;
495
496	/*
 
 
 
 
 
 
 
 
 
 
497	 * At this point we know it's a buffered write. If this is
498	 * kswapd trying to free memory, or REQ_SYNC is set, set, then
499	 * it's WB_SYNC_ALL writeback, and we'll use the max limit for
500	 * that. If the write is marked as a background write, then use
501	 * the idle limit, or go to normal if we haven't had competing
502	 * IO for a bit.
503	 */
504	if ((rw & REQ_HIPRIO) || wb_recent_wait(rwb) || current_is_kswapd())
505		limit = rwb->wb_max;
506	else if ((rw & REQ_BACKGROUND) || close_io(rwb)) {
507		/*
508		 * If less than 100ms since we completed unrelated IO,
509		 * limit us to half the depth for background writeback.
510		 */
511		limit = rwb->wb_background;
512	} else
513		limit = rwb->wb_normal;
514
515	return limit;
516}
517
518static inline bool may_queue(struct rq_wb *rwb, struct rq_wait *rqw,
519			     wait_queue_t *wait, unsigned long rw)
 
 
 
 
 
520{
521	/*
522	 * inc it here even if disabled, since we'll dec it at completion.
523	 * this only happens if the task was sleeping in __wbt_wait(),
524	 * and someone turned it off at the same time.
525	 */
526	if (!rwb_enabled(rwb)) {
527		atomic_inc(&rqw->inflight);
528		return true;
529	}
530
531	/*
532	 * If the waitqueue is already active and we are not the next
533	 * in line to be woken up, wait for our turn.
534	 */
535	if (waitqueue_active(&rqw->wait) &&
536	    rqw->wait.task_list.next != &wait->task_list)
537		return false;
538
539	return atomic_inc_below(&rqw->inflight, get_limit(rwb, rw));
540}
541
542/*
543 * Block if we will exceed our limit, or if we are currently waiting for
544 * the timer to kick off queuing again.
545 */
546static void __wbt_wait(struct rq_wb *rwb, unsigned long rw, spinlock_t *lock)
547	__releases(lock)
548	__acquires(lock)
549{
550	struct rq_wait *rqw = get_rq_wait(rwb, current_is_kswapd());
551	DEFINE_WAIT(wait);
 
 
 
 
552
553	if (may_queue(rwb, rqw, &wait, rw))
554		return;
555
556	do {
557		prepare_to_wait_exclusive(&rqw->wait, &wait,
558						TASK_UNINTERRUPTIBLE);
559
560		if (may_queue(rwb, rqw, &wait, rw))
561			break;
562
563		if (lock) {
564			spin_unlock_irq(lock);
565			io_schedule();
566			spin_lock_irq(lock);
567		} else
568			io_schedule();
569	} while (1);
570
571	finish_wait(&rqw->wait, &wait);
572}
573
574static inline bool wbt_should_throttle(struct rq_wb *rwb, struct bio *bio)
575{
576	const int op = bio_op(bio);
577
578	/*
579	 * If not a WRITE, do nothing
580	 */
581	if (op != REQ_OP_WRITE)
582		return false;
583
584	/*
585	 * Don't throttle WRITE_ODIRECT
586	 */
587	if ((bio->bi_opf & (REQ_SYNC | REQ_IDLE)) == (REQ_SYNC | REQ_IDLE))
588		return false;
589
590	return true;
 
 
 
 
 
 
 
 
 
 
591}
592
593/*
594 * Returns true if the IO request should be accounted, false if not.
595 * May sleep, if we have exceeded the writeback limits. Caller can pass
596 * in an irq held spinlock, if it holds one when calling this function.
597 * If we do sleep, we'll release and re-grab it.
598 */
599enum wbt_flags wbt_wait(struct rq_wb *rwb, struct bio *bio, spinlock_t *lock)
600{
601	unsigned int ret = 0;
602
603	if (!rwb_enabled(rwb))
604		return 0;
605
606	if (bio_op(bio) == REQ_OP_READ)
607		ret = WBT_READ;
608
609	if (!wbt_should_throttle(rwb, bio)) {
610		if (ret & WBT_READ)
 
611			wb_timestamp(rwb, &rwb->last_issue);
612		return ret;
613	}
614
615	__wbt_wait(rwb, bio->bi_opf, lock);
616
617	if (!timer_pending(&rwb->window_timer))
618		rwb_arm_timer(rwb);
 
619
620	if (current_is_kswapd())
621		ret |= WBT_KSWAPD;
622
623	return ret | WBT_TRACKED;
624}
625
626void wbt_issue(struct rq_wb *rwb, struct blk_issue_stat *stat)
627{
 
 
628	if (!rwb_enabled(rwb))
629		return;
630
631	/*
632	 * Track sync issue, in case it takes a long time to complete. Allows
633	 * us to react quicker, if a sync IO takes a long time to complete.
634	 * Note that this is just a hint. 'stat' can go away when the
635	 * request completes, so it's important we never dereference it. We
636	 * only use the address to compare with, which is why we store the
637	 * sync_issue time locally.
638	 */
639	if (wbt_is_read(stat) && !rwb->sync_issue) {
640		rwb->sync_cookie = stat;
641		rwb->sync_issue = blk_stat_time(stat);
642	}
643}
644
645void wbt_requeue(struct rq_wb *rwb, struct blk_issue_stat *stat)
646{
 
647	if (!rwb_enabled(rwb))
648		return;
649	if (stat == rwb->sync_cookie) {
650		rwb->sync_issue = 0;
651		rwb->sync_cookie = NULL;
652	}
653}
654
655void wbt_set_queue_depth(struct rq_wb *rwb, unsigned int depth)
656{
657	if (rwb) {
658		rwb->queue_depth = depth;
659		wbt_update_limits(rwb);
660	}
661}
662
663void wbt_set_write_cache(struct rq_wb *rwb, bool write_cache_on)
 
 
 
664{
665	if (rwb)
666		rwb->wc = write_cache_on;
667}
 
 
 
 
 
 
 
 
 
668
669 /*
670 * Disable wbt, if enabled by default. Only called from CFQ, if we have
671 * cgroups enabled
672 */
673void wbt_disable_default(struct request_queue *q)
674{
675	struct rq_wb *rwb = q->rq_wb;
676
677	if (rwb && rwb->enable_state == WBT_STATE_ON_DEFAULT) {
678		del_timer_sync(&rwb->window_timer);
679		rwb->win_nsec = rwb->min_lat_nsec = 0;
680		wbt_update_limits(rwb);
681	}
682}
683EXPORT_SYMBOL_GPL(wbt_disable_default);
684
685u64 wbt_default_latency_nsec(struct request_queue *q)
686{
687	/*
688	 * We default to 2msec for non-rotational storage, and 75msec
689	 * for rotational storage.
690	 */
691	if (blk_queue_nonrot(q))
692		return 2000000ULL;
693	else
694		return 75000000ULL;
695}
696
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
697int wbt_init(struct request_queue *q)
698{
699	struct rq_wb *rwb;
700	int i;
701
702	/*
703	 * For now, we depend on the stats window being larger than
704	 * our monitoring window. Ensure that this isn't inadvertently
705	 * violated.
706	 */
707	BUILD_BUG_ON(RWB_WINDOW_NSEC > BLK_STAT_NSEC);
708	BUILD_BUG_ON(WBT_NR_BITS > BLK_STAT_RES_BITS);
709
710	rwb = kzalloc(sizeof(*rwb), GFP_KERNEL);
711	if (!rwb)
712		return -ENOMEM;
713
714	for (i = 0; i < WBT_NUM_RWQ; i++) {
715		atomic_set(&rwb->rq_wait[i].inflight, 0);
716		init_waitqueue_head(&rwb->rq_wait[i].wait);
 
717	}
718
719	setup_timer(&rwb->window_timer, wb_timer_fn, (unsigned long) rwb);
720	rwb->wc = 1;
721	rwb->queue_depth = RWB_DEF_DEPTH;
 
 
 
722	rwb->last_comp = rwb->last_issue = jiffies;
723	rwb->queue = q;
724	rwb->win_nsec = RWB_WINDOW_NSEC;
725	rwb->enable_state = WBT_STATE_ON_DEFAULT;
726	wbt_update_limits(rwb);
 
 
 
 
727
728	/*
729	 * Assign rwb, and turn on stats tracking for this queue
730	 */
731	q->rq_wb = rwb;
732	blk_stat_enable(q);
 
733
734	rwb->min_lat_nsec = wbt_default_latency_nsec(q);
735
736	wbt_set_queue_depth(rwb, blk_queue_depth(q));
737	wbt_set_write_cache(rwb, test_bit(QUEUE_FLAG_WC, &q->queue_flags));
738
739	return 0;
740}
741
742void wbt_exit(struct request_queue *q)
743{
744	struct rq_wb *rwb = q->rq_wb;
 
745
746	if (rwb) {
747		del_timer_sync(&rwb->window_timer);
748		q->rq_wb = NULL;
749		kfree(rwb);
750	}
751}