xref: /freebsd/sys/cam/cam_iosched.c (revision eb69d1f144a6fcc765d1b9d44a5ae8082353e70b)
1 /*-
2  * CAM IO Scheduler Interface
3  *
4  * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
5  *
6  * Copyright (c) 2015 Netflix, Inc.
7  * All rights reserved.
8  *
9  * Redistribution and use in source and binary forms, with or without
10  * modification, are permitted provided that the following conditions
11  * are met:
12  * 1. Redistributions of source code must retain the above copyright
13  *    notice, this list of conditions, and the following disclaimer,
14  *    without modification, immediately at the beginning of the file.
15  * 2. The name of the author may not be used to endorse or promote products
16  *    derived from this software without specific prior written permission.
17  *
18  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
22  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
28  * SUCH DAMAGE.
29  *
30  * $FreeBSD$
31  */
32 
33 #include "opt_cam.h"
34 #include "opt_ddb.h"
35 
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
38 
39 #include <sys/param.h>
40 
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
43 #include <sys/bio.h>
44 #include <sys/lock.h>
45 #include <sys/malloc.h>
46 #include <sys/mutex.h>
47 #include <sys/sbuf.h>
48 #include <sys/sysctl.h>
49 
50 #include <cam/cam.h>
51 #include <cam/cam_ccb.h>
52 #include <cam/cam_periph.h>
53 #include <cam/cam_xpt_periph.h>
54 #include <cam/cam_xpt_internal.h>
55 #include <cam/cam_iosched.h>
56 
57 #include <ddb/ddb.h>
58 
59 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler",
60     "CAM I/O Scheduler buffers");
61 
62 /*
63  * Default I/O scheduler for FreeBSD. This implementation is just a thin-vineer
64  * over the bioq_* interface, with notions of separate calls for normal I/O and
65  * for trims.
66  *
67  * When CAM_IOSCHED_DYNAMIC is defined, the scheduler is enhanced to dynamically
68  * steer the rate of one type of traffic to help other types of traffic (eg
69  * limit writes when read latency deteriorates on SSDs).
70  */
71 
72 #ifdef CAM_IOSCHED_DYNAMIC
73 
74 static int do_dynamic_iosched = 1;
75 TUNABLE_INT("kern.cam.do_dynamic_iosched", &do_dynamic_iosched);
76 SYSCTL_INT(_kern_cam, OID_AUTO, do_dynamic_iosched, CTLFLAG_RD,
77     &do_dynamic_iosched, 1,
78     "Enable Dynamic I/O scheduler optimizations.");
79 
80 /*
81  * For an EMA, with an alpha of alpha, we know
82  * 	alpha = 2 / (N + 1)
83  * or
84  * 	N = 1 + (2 / alpha)
85  * where N is the number of samples that 86% of the current
86  * EMA is derived from.
87  *
88  * So we invent[*] alpha_bits:
89  *	alpha_bits = -log_2(alpha)
90  *	alpha = 2^-alpha_bits
91  * So
92  *	N = 1 + 2^(alpha_bits + 1)
93  *
94  * The default 9 gives a 1025 lookback for 86% of the data.
95  * For a brief intro: https://en.wikipedia.org/wiki/Moving_average
96  *
97  * [*] Steal from the load average code and many other places.
98  * Note: See computation of EMA and EMVAR for acceptable ranges of alpha.
99  */
100 static int alpha_bits = 9;
101 TUNABLE_INT("kern.cam.iosched_alpha_bits", &alpha_bits);
102 SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW,
103     &alpha_bits, 1,
104     "Bits in EMA's alpha.");
105 
106 struct iop_stats;
107 struct cam_iosched_softc;
108 
109 int iosched_debug = 0;
110 
111 typedef enum {
112 	none = 0,				/* No limits */
113 	queue_depth,			/* Limit how many ops we queue to SIM */
114 	iops,				/* Limit # of IOPS to the drive */
115 	bandwidth,			/* Limit bandwidth to the drive */
116 	limiter_max
117 } io_limiter;
118 
119 static const char *cam_iosched_limiter_names[] =
120     { "none", "queue_depth", "iops", "bandwidth" };
121 
122 /*
123  * Called to initialize the bits of the iop_stats structure relevant to the
124  * limiter. Called just after the limiter is set.
125  */
126 typedef int l_init_t(struct iop_stats *);
127 
128 /*
129  * Called every tick.
130  */
131 typedef int l_tick_t(struct iop_stats *);
132 
133 /*
134  * Called to see if the limiter thinks this IOP can be allowed to
135  * proceed. If so, the limiter assumes that the IOP proceeded
136  * and makes any accounting of it that's needed.
137  */
138 typedef int l_iop_t(struct iop_stats *, struct bio *);
139 
140 /*
141  * Called when an I/O completes so the limiter can update its
142  * accounting. Pending I/Os may complete in any order (even when
143  * sent to the hardware at the same time), so the limiter may not
144  * make any assumptions other than this I/O has completed. If it
145  * returns 1, then xpt_schedule() needs to be called again.
146  */
147 typedef int l_iodone_t(struct iop_stats *, struct bio *);
148 
149 static l_iop_t cam_iosched_qd_iop;
150 static l_iop_t cam_iosched_qd_caniop;
151 static l_iodone_t cam_iosched_qd_iodone;
152 
153 static l_init_t cam_iosched_iops_init;
154 static l_tick_t cam_iosched_iops_tick;
155 static l_iop_t cam_iosched_iops_caniop;
156 static l_iop_t cam_iosched_iops_iop;
157 
158 static l_init_t cam_iosched_bw_init;
159 static l_tick_t cam_iosched_bw_tick;
160 static l_iop_t cam_iosched_bw_caniop;
161 static l_iop_t cam_iosched_bw_iop;
162 
163 struct limswitch {
164 	l_init_t	*l_init;
165 	l_tick_t	*l_tick;
166 	l_iop_t		*l_iop;
167 	l_iop_t		*l_caniop;
168 	l_iodone_t	*l_iodone;
169 } limsw[] =
170 {
171 	{	/* none */
172 		.l_init = NULL,
173 		.l_tick = NULL,
174 		.l_iop = NULL,
175 		.l_iodone= NULL,
176 	},
177 	{	/* queue_depth */
178 		.l_init = NULL,
179 		.l_tick = NULL,
180 		.l_caniop = cam_iosched_qd_caniop,
181 		.l_iop = cam_iosched_qd_iop,
182 		.l_iodone= cam_iosched_qd_iodone,
183 	},
184 	{	/* iops */
185 		.l_init = cam_iosched_iops_init,
186 		.l_tick = cam_iosched_iops_tick,
187 		.l_caniop = cam_iosched_iops_caniop,
188 		.l_iop = cam_iosched_iops_iop,
189 		.l_iodone= NULL,
190 	},
191 	{	/* bandwidth */
192 		.l_init = cam_iosched_bw_init,
193 		.l_tick = cam_iosched_bw_tick,
194 		.l_caniop = cam_iosched_bw_caniop,
195 		.l_iop = cam_iosched_bw_iop,
196 		.l_iodone= NULL,
197 	},
198 };
199 
200 struct iop_stats {
201 	/*
202 	 * sysctl state for this subnode.
203 	 */
204 	struct sysctl_ctx_list	sysctl_ctx;
205 	struct sysctl_oid	*sysctl_tree;
206 
207 	/*
208 	 * Information about the current rate limiters, if any
209 	 */
210 	io_limiter	limiter;	/* How are I/Os being limited */
211 	int		min;		/* Low range of limit */
212 	int		max;		/* High range of limit */
213 	int		current;	/* Current rate limiter */
214 	int		l_value1;	/* per-limiter scratch value 1. */
215 	int		l_value2;	/* per-limiter scratch value 2. */
216 
217 	/*
218 	 * Debug information about counts of I/Os that have gone through the
219 	 * scheduler.
220 	 */
221 	int		pending;	/* I/Os pending in the hardware */
222 	int		queued;		/* number currently in the queue */
223 	int		total;		/* Total for all time -- wraps */
224 	int		in;		/* number queued all time -- wraps */
225 	int		out;		/* number completed all time -- wraps */
226 	int		errs;		/* Number of I/Os completed with error --  wraps */
227 
228 	/*
229 	 * Statistics on different bits of the process.
230 	 */
231 		/* Exp Moving Average, see alpha_bits for more details */
232 	sbintime_t      ema;
233 	sbintime_t      emvar;
234 	sbintime_t      sd;		/* Last computed sd */
235 
236 	uint32_t	state_flags;
237 #define IOP_RATE_LIMITED		1u
238 
239 #define LAT_BUCKETS 15			/* < 1ms < 2ms ... < 2^(n-1)ms >= 2^(n-1)ms*/
240 	uint64_t	latencies[LAT_BUCKETS];
241 
242 	struct cam_iosched_softc *softc;
243 };
244 
245 
246 typedef enum {
247 	set_max = 0,			/* current = max */
248 	read_latency,			/* Steer read latency by throttling writes */
249 	cl_max				/* Keep last */
250 } control_type;
251 
252 static const char *cam_iosched_control_type_names[] =
253     { "set_max", "read_latency" };
254 
255 struct control_loop {
256 	/*
257 	 * sysctl state for this subnode.
258 	 */
259 	struct sysctl_ctx_list	sysctl_ctx;
260 	struct sysctl_oid	*sysctl_tree;
261 
262 	sbintime_t	next_steer;		/* Time of next steer */
263 	sbintime_t	steer_interval;		/* How often do we steer? */
264 	sbintime_t	lolat;
265 	sbintime_t	hilat;
266 	int		alpha;
267 	control_type	type;			/* What type of control? */
268 	int		last_count;		/* Last I/O count */
269 
270 	struct cam_iosched_softc *softc;
271 };
272 
273 #endif
274 
275 struct cam_iosched_softc {
276 	struct bio_queue_head bio_queue;
277 	struct bio_queue_head trim_queue;
278 				/* scheduler flags < 16, user flags >= 16 */
279 	uint32_t	flags;
280 	int		sort_io_queue;
281 #ifdef CAM_IOSCHED_DYNAMIC
282 	int		read_bias;		/* Read bias setting */
283 	int		current_read_bias;	/* Current read bias state */
284 	int		total_ticks;
285 	int		load;			/* EMA of 'load average' of disk / 2^16 */
286 
287 	struct bio_queue_head write_queue;
288 	struct iop_stats read_stats, write_stats, trim_stats;
289 	struct sysctl_ctx_list	sysctl_ctx;
290 	struct sysctl_oid	*sysctl_tree;
291 
292 	int		quanta;			/* Number of quanta per second */
293 	struct callout	ticker;			/* Callout for our quota system */
294 	struct cam_periph *periph;		/* cam periph associated with this device */
295 	uint32_t	this_frac;		/* Fraction of a second (1024ths) for this tick */
296 	sbintime_t	last_time;		/* Last time we ticked */
297 	struct control_loop cl;
298 #endif
299 };
300 
301 #ifdef CAM_IOSCHED_DYNAMIC
302 /*
303  * helper functions to call the limsw functions.
304  */
305 static int
306 cam_iosched_limiter_init(struct iop_stats *ios)
307 {
308 	int lim = ios->limiter;
309 
310 	/* maybe this should be a kassert */
311 	if (lim < none || lim >= limiter_max)
312 		return EINVAL;
313 
314 	if (limsw[lim].l_init)
315 		return limsw[lim].l_init(ios);
316 
317 	return 0;
318 }
319 
320 static int
321 cam_iosched_limiter_tick(struct iop_stats *ios)
322 {
323 	int lim = ios->limiter;
324 
325 	/* maybe this should be a kassert */
326 	if (lim < none || lim >= limiter_max)
327 		return EINVAL;
328 
329 	if (limsw[lim].l_tick)
330 		return limsw[lim].l_tick(ios);
331 
332 	return 0;
333 }
334 
335 static int
336 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
337 {
338 	int lim = ios->limiter;
339 
340 	/* maybe this should be a kassert */
341 	if (lim < none || lim >= limiter_max)
342 		return EINVAL;
343 
344 	if (limsw[lim].l_iop)
345 		return limsw[lim].l_iop(ios, bp);
346 
347 	return 0;
348 }
349 
350 static int
351 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
352 {
353 	int lim = ios->limiter;
354 
355 	/* maybe this should be a kassert */
356 	if (lim < none || lim >= limiter_max)
357 		return EINVAL;
358 
359 	if (limsw[lim].l_caniop)
360 		return limsw[lim].l_caniop(ios, bp);
361 
362 	return 0;
363 }
364 
365 static int
366 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
367 {
368 	int lim = ios->limiter;
369 
370 	/* maybe this should be a kassert */
371 	if (lim < none || lim >= limiter_max)
372 		return 0;
373 
374 	if (limsw[lim].l_iodone)
375 		return limsw[lim].l_iodone(ios, bp);
376 
377 	return 0;
378 }
379 
380 /*
381  * Functions to implement the different kinds of limiters
382  */
383 
384 static int
385 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
386 {
387 
388 	if (ios->current <= 0 || ios->pending < ios->current)
389 		return 0;
390 
391 	return EAGAIN;
392 }
393 
394 static int
395 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
396 {
397 
398 	if (ios->current <= 0 || ios->pending < ios->current)
399 		return 0;
400 
401 	return EAGAIN;
402 }
403 
404 static int
405 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
406 {
407 
408 	if (ios->current <= 0 || ios->pending != ios->current)
409 		return 0;
410 
411 	return 1;
412 }
413 
414 static int
415 cam_iosched_iops_init(struct iop_stats *ios)
416 {
417 
418 	ios->l_value1 = ios->current / ios->softc->quanta;
419 	if (ios->l_value1 <= 0)
420 		ios->l_value1 = 1;
421 	ios->l_value2 = 0;
422 
423 	return 0;
424 }
425 
426 static int
427 cam_iosched_iops_tick(struct iop_stats *ios)
428 {
429 	int new_ios;
430 
431 	/*
432 	 * Allow at least one IO per tick until all
433 	 * the IOs for this interval have been spent.
434 	 */
435 	new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
436 	if (new_ios < 1 && ios->l_value2 < ios->current) {
437 		new_ios = 1;
438 		ios->l_value2++;
439 	}
440 
441 	/*
442 	 * If this a new accounting interval, discard any "unspent" ios
443 	 * granted in the previous interval.  Otherwise add the new ios to
444 	 * the previously granted ones that haven't been spent yet.
445 	 */
446 	if ((ios->softc->total_ticks % ios->softc->quanta) == 0) {
447 		ios->l_value1 = new_ios;
448 		ios->l_value2 = 1;
449 	} else {
450 		ios->l_value1 += new_ios;
451 	}
452 
453 
454 	return 0;
455 }
456 
457 static int
458 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
459 {
460 
461 	/*
462 	 * So if we have any more IOPs left, allow it,
463 	 * otherwise wait. If current iops is 0, treat that
464 	 * as unlimited as a failsafe.
465 	 */
466 	if (ios->current > 0 && ios->l_value1 <= 0)
467 		return EAGAIN;
468 	return 0;
469 }
470 
471 static int
472 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
473 {
474 	int rv;
475 
476 	rv = cam_iosched_limiter_caniop(ios, bp);
477 	if (rv == 0)
478 		ios->l_value1--;
479 
480 	return rv;
481 }
482 
483 static int
484 cam_iosched_bw_init(struct iop_stats *ios)
485 {
486 
487 	/* ios->current is in kB/s, so scale to bytes */
488 	ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
489 
490 	return 0;
491 }
492 
493 static int
494 cam_iosched_bw_tick(struct iop_stats *ios)
495 {
496 	int bw;
497 
498 	/*
499 	 * If we're in the hole for available quota from
500 	 * the last time, then add the quantum for this.
501 	 * If we have any left over from last quantum,
502 	 * then too bad, that's lost. Also, ios->current
503 	 * is in kB/s, so scale.
504 	 *
505 	 * We also allow up to 4 quanta of credits to
506 	 * accumulate to deal with burstiness. 4 is extremely
507 	 * arbitrary.
508 	 */
509 	bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
510 	if (ios->l_value1 < bw * 4)
511 		ios->l_value1 += bw;
512 
513 	return 0;
514 }
515 
516 static int
517 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
518 {
519 	/*
520 	 * So if we have any more bw quota left, allow it,
521 	 * otherwise wait. Note, we'll go negative and that's
522 	 * OK. We'll just get a little less next quota.
523 	 *
524 	 * Note on going negative: that allows us to process
525 	 * requests in order better, since we won't allow
526 	 * shorter reads to get around the long one that we
527 	 * don't have the quota to do just yet. It also prevents
528 	 * starvation by being a little more permissive about
529 	 * what we let through this quantum (to prevent the
530 	 * starvation), at the cost of getting a little less
531 	 * next quantum.
532 	 *
533 	 * Also note that if the current limit is <= 0,
534 	 * we treat it as unlimited as a failsafe.
535 	 */
536 	if (ios->current > 0 && ios->l_value1 <= 0)
537 		return EAGAIN;
538 
539 
540 	return 0;
541 }
542 
543 static int
544 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
545 {
546 	int rv;
547 
548 	rv = cam_iosched_limiter_caniop(ios, bp);
549 	if (rv == 0)
550 		ios->l_value1 -= bp->bio_length;
551 
552 	return rv;
553 }
554 
555 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
556 
557 static void
558 cam_iosched_ticker(void *arg)
559 {
560 	struct cam_iosched_softc *isc = arg;
561 	sbintime_t now, delta;
562 	int pending;
563 
564 	callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
565 
566 	now = sbinuptime();
567 	delta = now - isc->last_time;
568 	isc->this_frac = (uint32_t)delta >> 16;		/* Note: discards seconds -- should be 0 harmless if not */
569 	isc->last_time = now;
570 
571 	cam_iosched_cl_maybe_steer(&isc->cl);
572 
573 	cam_iosched_limiter_tick(&isc->read_stats);
574 	cam_iosched_limiter_tick(&isc->write_stats);
575 	cam_iosched_limiter_tick(&isc->trim_stats);
576 
577 	cam_iosched_schedule(isc, isc->periph);
578 
579 	/*
580 	 * isc->load is an EMA of the pending I/Os at each tick. The number of
581 	 * pending I/Os is the sum of the I/Os queued to the hardware, and those
582 	 * in the software queue that could be queued to the hardware if there
583 	 * were slots.
584 	 *
585 	 * ios_stats.pending is a count of requests in the SIM right now for
586 	 * each of these types of I/O. So the total pending count is the sum of
587 	 * these I/Os and the sum of the queued I/Os still in the software queue
588 	 * for those operations that aren't being rate limited at the moment.
589 	 *
590 	 * The reason for the rate limiting bit is because those I/Os
591 	 * aren't part of the software queued load (since we could
592 	 * give them to hardware, but choose not to).
593 	 *
594 	 * Note: due to a bug in counting pending TRIM in the device, we
595 	 * don't include them in this count. We count each BIO_DELETE in
596 	 * the pending count, but the periph drivers collapse them down
597 	 * into one TRIM command. That one trim command gets the completion
598 	 * so the counts get off.
599 	 */
600 	pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
601 	pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
602 	    !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
603 	    !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
604 	pending <<= 16;
605 	pending /= isc->periph->path->device->ccbq.total_openings;
606 
607 	isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
608 
609 	isc->total_ticks++;
610 }
611 
612 
613 static void
614 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
615 {
616 
617 	clp->next_steer = sbinuptime();
618 	clp->softc = isc;
619 	clp->steer_interval = SBT_1S * 5;	/* Let's start out steering every 5s */
620 	clp->lolat = 5 * SBT_1MS;
621 	clp->hilat = 15 * SBT_1MS;
622 	clp->alpha = 20;			/* Alpha == gain. 20 = .2 */
623 	clp->type = set_max;
624 }
625 
626 static void
627 cam_iosched_cl_maybe_steer(struct control_loop *clp)
628 {
629 	struct cam_iosched_softc *isc;
630 	sbintime_t now, lat;
631 	int old;
632 
633 	isc = clp->softc;
634 	now = isc->last_time;
635 	if (now < clp->next_steer)
636 		return;
637 
638 	clp->next_steer = now + clp->steer_interval;
639 	switch (clp->type) {
640 	case set_max:
641 		if (isc->write_stats.current != isc->write_stats.max)
642 			printf("Steering write from %d kBps to %d kBps\n",
643 			    isc->write_stats.current, isc->write_stats.max);
644 		isc->read_stats.current = isc->read_stats.max;
645 		isc->write_stats.current = isc->write_stats.max;
646 		isc->trim_stats.current = isc->trim_stats.max;
647 		break;
648 	case read_latency:
649 		old = isc->write_stats.current;
650 		lat = isc->read_stats.ema;
651 		/*
652 		 * Simple PLL-like engine. Since we're steering to a range for
653 		 * the SP (set point) that makes things a little more
654 		 * complicated. In addition, we're not directly controlling our
655 		 * PV (process variable), the read latency, but instead are
656 		 * manipulating the write bandwidth limit for our MV
657 		 * (manipulation variable), analysis of this code gets a bit
658 		 * messy. Also, the MV is a very noisy control surface for read
659 		 * latency since it is affected by many hidden processes inside
660 		 * the device which change how responsive read latency will be
661 		 * in reaction to changes in write bandwidth. Unlike the classic
662 		 * boiler control PLL. this may result in over-steering while
663 		 * the SSD takes its time to react to the new, lower load. This
664 		 * is why we use a relatively low alpha of between .1 and .25 to
665 		 * compensate for this effect. At .1, it takes ~22 steering
666 		 * intervals to back off by a factor of 10. At .2 it only takes
667 		 * ~10. At .25 it only takes ~8. However some preliminary data
668 		 * from the SSD drives suggests a reasponse time in 10's of
669 		 * seconds before latency drops regardless of the new write
670 		 * rate. Careful observation will be required to tune this
671 		 * effectively.
672 		 *
673 		 * Also, when there's no read traffic, we jack up the write
674 		 * limit too regardless of the last read latency.  10 is
675 		 * somewhat arbitrary.
676 		 */
677 		if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
678 			isc->write_stats.current = isc->write_stats.current *
679 			    (100 + clp->alpha) / 100;	/* Scale up */
680 		else if (lat > clp->hilat)
681 			isc->write_stats.current = isc->write_stats.current *
682 			    (100 - clp->alpha) / 100;	/* Scale down */
683 		clp->last_count = isc->read_stats.total;
684 
685 		/*
686 		 * Even if we don't steer, per se, enforce the min/max limits as
687 		 * those may have changed.
688 		 */
689 		if (isc->write_stats.current < isc->write_stats.min)
690 			isc->write_stats.current = isc->write_stats.min;
691 		if (isc->write_stats.current > isc->write_stats.max)
692 			isc->write_stats.current = isc->write_stats.max;
693 		if (old != isc->write_stats.current && 	iosched_debug)
694 			printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
695 			    old, isc->write_stats.current,
696 			    (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
697 		break;
698 	case cl_max:
699 		break;
700 	}
701 }
702 #endif
703 
704 /*
705  * Trim or similar currently pending completion. Should only be set for
706  * those drivers wishing only one Trim active at a time.
707  */
708 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE	(1ul << 0)
709 			/* Callout active, and needs to be torn down */
710 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
711 
712 			/* Periph drivers set these flags to indicate work */
713 #define CAM_IOSCHED_FLAG_WORK_FLAGS	((0xffffu) << 16)
714 
715 #ifdef CAM_IOSCHED_DYNAMIC
716 static void
717 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
718     sbintime_t sim_latency, int cmd, size_t size);
719 #endif
720 
721 static inline int
722 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
723 {
724 	return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
725 }
726 
727 static inline int
728 cam_iosched_has_io(struct cam_iosched_softc *isc)
729 {
730 #ifdef CAM_IOSCHED_DYNAMIC
731 	if (do_dynamic_iosched) {
732 		struct bio *rbp = bioq_first(&isc->bio_queue);
733 		struct bio *wbp = bioq_first(&isc->write_queue);
734 		int can_write = wbp != NULL &&
735 		    cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
736 		int can_read = rbp != NULL &&
737 		    cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
738 		if (iosched_debug > 2) {
739 			printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
740 			printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
741 			printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
742 		}
743 		return can_read || can_write;
744 	}
745 #endif
746 	return bioq_first(&isc->bio_queue) != NULL;
747 }
748 
749 static inline int
750 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
751 {
752 	return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) &&
753 	    bioq_first(&isc->trim_queue);
754 }
755 
756 #define cam_iosched_sort_queue(isc)	((isc)->sort_io_queue >= 0 ?	\
757     (isc)->sort_io_queue : cam_sort_io_queues)
758 
759 
760 static inline int
761 cam_iosched_has_work(struct cam_iosched_softc *isc)
762 {
763 #ifdef CAM_IOSCHED_DYNAMIC
764 	if (iosched_debug > 2)
765 		printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
766 		    cam_iosched_has_more_trim(isc),
767 		    cam_iosched_has_flagged_work(isc));
768 #endif
769 
770 	return cam_iosched_has_io(isc) ||
771 		cam_iosched_has_more_trim(isc) ||
772 		cam_iosched_has_flagged_work(isc);
773 }
774 
775 #ifdef CAM_IOSCHED_DYNAMIC
776 static void
777 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
778 {
779 
780 	ios->limiter = none;
781 	ios->in = 0;
782 	ios->max = ios->current = 300000;
783 	ios->min = 1;
784 	ios->out = 0;
785 	ios->errs = 0;
786 	ios->pending = 0;
787 	ios->queued = 0;
788 	ios->total = 0;
789 	ios->ema = 0;
790 	ios->emvar = 0;
791 	ios->softc = isc;
792 	cam_iosched_limiter_init(ios);
793 }
794 
795 static int
796 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
797 {
798 	char buf[16];
799 	struct iop_stats *ios;
800 	struct cam_iosched_softc *isc;
801 	int value, i, error;
802 	const char *p;
803 
804 	ios = arg1;
805 	isc = ios->softc;
806 	value = ios->limiter;
807 	if (value < none || value >= limiter_max)
808 		p = "UNKNOWN";
809 	else
810 		p = cam_iosched_limiter_names[value];
811 
812 	strlcpy(buf, p, sizeof(buf));
813 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
814 	if (error != 0 || req->newptr == NULL)
815 		return error;
816 
817 	cam_periph_lock(isc->periph);
818 
819 	for (i = none; i < limiter_max; i++) {
820 		if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
821 			continue;
822 		ios->limiter = i;
823 		error = cam_iosched_limiter_init(ios);
824 		if (error != 0) {
825 			ios->limiter = value;
826 			cam_periph_unlock(isc->periph);
827 			return error;
828 		}
829 		/* Note: disk load averate requires ticker to be always running */
830 		callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
831 		isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
832 
833 		cam_periph_unlock(isc->periph);
834 		return 0;
835 	}
836 
837 	cam_periph_unlock(isc->periph);
838 	return EINVAL;
839 }
840 
841 static int
842 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
843 {
844 	char buf[16];
845 	struct control_loop *clp;
846 	struct cam_iosched_softc *isc;
847 	int value, i, error;
848 	const char *p;
849 
850 	clp = arg1;
851 	isc = clp->softc;
852 	value = clp->type;
853 	if (value < none || value >= cl_max)
854 		p = "UNKNOWN";
855 	else
856 		p = cam_iosched_control_type_names[value];
857 
858 	strlcpy(buf, p, sizeof(buf));
859 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
860 	if (error != 0 || req->newptr == NULL)
861 		return error;
862 
863 	for (i = set_max; i < cl_max; i++) {
864 		if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
865 			continue;
866 		cam_periph_lock(isc->periph);
867 		clp->type = i;
868 		cam_periph_unlock(isc->periph);
869 		return 0;
870 	}
871 
872 	return EINVAL;
873 }
874 
875 static int
876 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
877 {
878 	char buf[16];
879 	sbintime_t value;
880 	int error;
881 	uint64_t us;
882 
883 	value = *(sbintime_t *)arg1;
884 	us = (uint64_t)value / SBT_1US;
885 	snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
886 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
887 	if (error != 0 || req->newptr == NULL)
888 		return error;
889 	us = strtoul(buf, NULL, 10);
890 	if (us == 0)
891 		return EINVAL;
892 	*(sbintime_t *)arg1 = us * SBT_1US;
893 	return 0;
894 }
895 
896 static int
897 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
898 {
899 	int i, error;
900 	struct sbuf sb;
901 	uint64_t *latencies;
902 
903 	latencies = arg1;
904 	sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
905 
906 	for (i = 0; i < LAT_BUCKETS - 1; i++)
907 		sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
908 	sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
909 	error = sbuf_finish(&sb);
910 	sbuf_delete(&sb);
911 
912 	return (error);
913 }
914 
915 static int
916 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS)
917 {
918 	int *quanta;
919 	int error, value;
920 
921 	quanta = (unsigned *)arg1;
922 	value = *quanta;
923 
924 	error = sysctl_handle_int(oidp, (int *)&value, 0, req);
925 	if ((error != 0) || (req->newptr == NULL))
926 		return (error);
927 
928 	if (value < 1 || value > hz)
929 		return (EINVAL);
930 
931 	*quanta = value;
932 
933 	return (0);
934 }
935 
936 static void
937 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
938 {
939 	struct sysctl_oid_list *n;
940 	struct sysctl_ctx_list *ctx;
941 
942 	ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
943 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
944 	    CTLFLAG_RD, 0, name);
945 	n = SYSCTL_CHILDREN(ios->sysctl_tree);
946 	ctx = &ios->sysctl_ctx;
947 
948 	SYSCTL_ADD_UQUAD(ctx, n,
949 	    OID_AUTO, "ema", CTLFLAG_RD,
950 	    &ios->ema,
951 	    "Fast Exponentially Weighted Moving Average");
952 	SYSCTL_ADD_UQUAD(ctx, n,
953 	    OID_AUTO, "emvar", CTLFLAG_RD,
954 	    &ios->emvar,
955 	    "Fast Exponentially Weighted Moving Variance");
956 
957 	SYSCTL_ADD_INT(ctx, n,
958 	    OID_AUTO, "pending", CTLFLAG_RD,
959 	    &ios->pending, 0,
960 	    "Instantaneous # of pending transactions");
961 	SYSCTL_ADD_INT(ctx, n,
962 	    OID_AUTO, "count", CTLFLAG_RD,
963 	    &ios->total, 0,
964 	    "# of transactions submitted to hardware");
965 	SYSCTL_ADD_INT(ctx, n,
966 	    OID_AUTO, "queued", CTLFLAG_RD,
967 	    &ios->queued, 0,
968 	    "# of transactions in the queue");
969 	SYSCTL_ADD_INT(ctx, n,
970 	    OID_AUTO, "in", CTLFLAG_RD,
971 	    &ios->in, 0,
972 	    "# of transactions queued to driver");
973 	SYSCTL_ADD_INT(ctx, n,
974 	    OID_AUTO, "out", CTLFLAG_RD,
975 	    &ios->out, 0,
976 	    "# of transactions completed (including with error)");
977 	SYSCTL_ADD_INT(ctx, n,
978 	    OID_AUTO, "errs", CTLFLAG_RD,
979 	    &ios->errs, 0,
980 	    "# of transactions completed with an error");
981 
982 	SYSCTL_ADD_PROC(ctx, n,
983 	    OID_AUTO, "limiter", CTLTYPE_STRING | CTLFLAG_RW,
984 	    ios, 0, cam_iosched_limiter_sysctl, "A",
985 	    "Current limiting type.");
986 	SYSCTL_ADD_INT(ctx, n,
987 	    OID_AUTO, "min", CTLFLAG_RW,
988 	    &ios->min, 0,
989 	    "min resource");
990 	SYSCTL_ADD_INT(ctx, n,
991 	    OID_AUTO, "max", CTLFLAG_RW,
992 	    &ios->max, 0,
993 	    "max resource");
994 	SYSCTL_ADD_INT(ctx, n,
995 	    OID_AUTO, "current", CTLFLAG_RW,
996 	    &ios->current, 0,
997 	    "current resource");
998 
999 	SYSCTL_ADD_PROC(ctx, n,
1000 	    OID_AUTO, "latencies", CTLTYPE_STRING | CTLFLAG_RD,
1001 	    &ios->latencies, 0,
1002 	    cam_iosched_sysctl_latencies, "A",
1003 	    "Array of power of 2 latency from 1ms to 1.024s");
1004 }
1005 
1006 static void
1007 cam_iosched_iop_stats_fini(struct iop_stats *ios)
1008 {
1009 	if (ios->sysctl_tree)
1010 		if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
1011 			printf("can't remove iosched sysctl stats context\n");
1012 }
1013 
1014 static void
1015 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
1016 {
1017 	struct sysctl_oid_list *n;
1018 	struct sysctl_ctx_list *ctx;
1019 	struct control_loop *clp;
1020 
1021 	clp = &isc->cl;
1022 	clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1023 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
1024 	    CTLFLAG_RD, 0, "Control loop info");
1025 	n = SYSCTL_CHILDREN(clp->sysctl_tree);
1026 	ctx = &clp->sysctl_ctx;
1027 
1028 	SYSCTL_ADD_PROC(ctx, n,
1029 	    OID_AUTO, "type", CTLTYPE_STRING | CTLFLAG_RW,
1030 	    clp, 0, cam_iosched_control_type_sysctl, "A",
1031 	    "Control loop algorithm");
1032 	SYSCTL_ADD_PROC(ctx, n,
1033 	    OID_AUTO, "steer_interval", CTLTYPE_STRING | CTLFLAG_RW,
1034 	    &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
1035 	    "How often to steer (in us)");
1036 	SYSCTL_ADD_PROC(ctx, n,
1037 	    OID_AUTO, "lolat", CTLTYPE_STRING | CTLFLAG_RW,
1038 	    &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
1039 	    "Low water mark for Latency (in us)");
1040 	SYSCTL_ADD_PROC(ctx, n,
1041 	    OID_AUTO, "hilat", CTLTYPE_STRING | CTLFLAG_RW,
1042 	    &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
1043 	    "Hi water mark for Latency (in us)");
1044 	SYSCTL_ADD_INT(ctx, n,
1045 	    OID_AUTO, "alpha", CTLFLAG_RW,
1046 	    &clp->alpha, 0,
1047 	    "Alpha for PLL (x100) aka gain");
1048 }
1049 
1050 static void
1051 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1052 {
1053 	if (clp->sysctl_tree)
1054 		if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1055 			printf("can't remove iosched sysctl control loop context\n");
1056 }
1057 #endif
1058 
1059 /*
1060  * Allocate the iosched structure. This also insulates callers from knowing
1061  * sizeof struct cam_iosched_softc.
1062  */
1063 int
1064 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1065 {
1066 
1067 	*iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1068 	if (*iscp == NULL)
1069 		return ENOMEM;
1070 #ifdef CAM_IOSCHED_DYNAMIC
1071 	if (iosched_debug)
1072 		printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1073 #endif
1074 	(*iscp)->sort_io_queue = -1;
1075 	bioq_init(&(*iscp)->bio_queue);
1076 	bioq_init(&(*iscp)->trim_queue);
1077 #ifdef CAM_IOSCHED_DYNAMIC
1078 	if (do_dynamic_iosched) {
1079 		bioq_init(&(*iscp)->write_queue);
1080 		(*iscp)->read_bias = 100;
1081 		(*iscp)->current_read_bias = 100;
1082 		(*iscp)->quanta = min(hz, 200);
1083 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1084 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1085 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1086 		(*iscp)->trim_stats.max = 1;	/* Trims are special: one at a time for now */
1087 		(*iscp)->last_time = sbinuptime();
1088 		callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1089 		(*iscp)->periph = periph;
1090 		cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1091 		callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp);
1092 		(*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1093 	}
1094 #endif
1095 
1096 	return 0;
1097 }
1098 
1099 /*
1100  * Reclaim all used resources. This assumes that other folks have
1101  * drained the requests in the hardware. Maybe an unwise assumption.
1102  */
1103 void
1104 cam_iosched_fini(struct cam_iosched_softc *isc)
1105 {
1106 	if (isc) {
1107 		cam_iosched_flush(isc, NULL, ENXIO);
1108 #ifdef CAM_IOSCHED_DYNAMIC
1109 		cam_iosched_iop_stats_fini(&isc->read_stats);
1110 		cam_iosched_iop_stats_fini(&isc->write_stats);
1111 		cam_iosched_iop_stats_fini(&isc->trim_stats);
1112 		cam_iosched_cl_sysctl_fini(&isc->cl);
1113 		if (isc->sysctl_tree)
1114 			if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1115 				printf("can't remove iosched sysctl stats context\n");
1116 		if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1117 			callout_drain(&isc->ticker);
1118 			isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1119 		}
1120 #endif
1121 		free(isc, M_CAMSCHED);
1122 	}
1123 }
1124 
1125 /*
1126  * After we're sure we're attaching a device, go ahead and add
1127  * hooks for any sysctl we may wish to honor.
1128  */
1129 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1130     struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1131 {
1132 #ifdef CAM_IOSCHED_DYNAMIC
1133 	struct sysctl_oid_list *n;
1134 #endif
1135 
1136 	SYSCTL_ADD_INT(ctx, SYSCTL_CHILDREN(node),
1137 		OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1138 		&isc->sort_io_queue, 0,
1139 		"Sort IO queue to try and optimise disk access patterns");
1140 
1141 #ifdef CAM_IOSCHED_DYNAMIC
1142 	if (!do_dynamic_iosched)
1143 		return;
1144 
1145 	isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1146 	    SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1147 	    CTLFLAG_RD, 0, "I/O scheduler statistics");
1148 	n = SYSCTL_CHILDREN(isc->sysctl_tree);
1149 	ctx = &isc->sysctl_ctx;
1150 
1151 	cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1152 	cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1153 	cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1154 	cam_iosched_cl_sysctl_init(isc);
1155 
1156 	SYSCTL_ADD_INT(ctx, n,
1157 	    OID_AUTO, "read_bias", CTLFLAG_RW,
1158 	    &isc->read_bias, 100,
1159 	    "How biased towards read should we be independent of limits");
1160 
1161 	SYSCTL_ADD_PROC(ctx, n,
1162 	    OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW,
1163 	    &isc->quanta, 0, cam_iosched_quanta_sysctl, "I",
1164 	    "How many quanta per second do we slice the I/O up into");
1165 
1166 	SYSCTL_ADD_INT(ctx, n,
1167 	    OID_AUTO, "total_ticks", CTLFLAG_RD,
1168 	    &isc->total_ticks, 0,
1169 	    "Total number of ticks we've done");
1170 
1171 	SYSCTL_ADD_INT(ctx, n,
1172 	    OID_AUTO, "load", CTLFLAG_RD,
1173 	    &isc->load, 0,
1174 	    "scaled load average / 100");
1175 #endif
1176 }
1177 
1178 /*
1179  * Flush outstanding I/O. Consumers of this library don't know all the
1180  * queues we may keep, so this allows all I/O to be flushed in one
1181  * convenient call.
1182  */
1183 void
1184 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1185 {
1186 	bioq_flush(&isc->bio_queue, stp, err);
1187 	bioq_flush(&isc->trim_queue, stp, err);
1188 #ifdef CAM_IOSCHED_DYNAMIC
1189 	if (do_dynamic_iosched)
1190 		bioq_flush(&isc->write_queue, stp, err);
1191 #endif
1192 }
1193 
1194 #ifdef CAM_IOSCHED_DYNAMIC
1195 static struct bio *
1196 cam_iosched_get_write(struct cam_iosched_softc *isc)
1197 {
1198 	struct bio *bp;
1199 
1200 	/*
1201 	 * We control the write rate by controlling how many requests we send
1202 	 * down to the drive at any one time. Fewer requests limits the
1203 	 * effects of both starvation when the requests take a while and write
1204 	 * amplification when each request is causing more than one write to
1205 	 * the NAND media. Limiting the queue depth like this will also limit
1206 	 * the write throughput and give and reads that want to compete to
1207 	 * compete unfairly.
1208 	 */
1209 	bp = bioq_first(&isc->write_queue);
1210 	if (bp == NULL) {
1211 		if (iosched_debug > 3)
1212 			printf("No writes present in write_queue\n");
1213 		return NULL;
1214 	}
1215 
1216 	/*
1217 	 * If pending read, prefer that based on current read bias
1218 	 * setting.
1219 	 */
1220 	if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1221 		if (iosched_debug)
1222 			printf("Reads present and current_read_bias is %d queued writes %d queued reads %d\n", isc->current_read_bias, isc->write_stats.queued, isc->read_stats.queued);
1223 		isc->current_read_bias--;
1224 		/* We're not limiting writes, per se, just doing reads first */
1225 		return NULL;
1226 	}
1227 
1228 	/*
1229 	 * See if our current limiter allows this I/O.
1230 	 */
1231 	if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1232 		if (iosched_debug)
1233 			printf("Can't write because limiter says no.\n");
1234 		isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1235 		return NULL;
1236 	}
1237 
1238 	/*
1239 	 * Let's do this: We've passed all the gates and we're a go
1240 	 * to schedule the I/O in the SIM.
1241 	 */
1242 	isc->current_read_bias = isc->read_bias;
1243 	bioq_remove(&isc->write_queue, bp);
1244 	if (bp->bio_cmd == BIO_WRITE) {
1245 		isc->write_stats.queued--;
1246 		isc->write_stats.total++;
1247 		isc->write_stats.pending++;
1248 	}
1249 	if (iosched_debug > 9)
1250 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1251 	isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1252 	return bp;
1253 }
1254 #endif
1255 
1256 /*
1257  * Put back a trim that you weren't able to actually schedule this time.
1258  */
1259 void
1260 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1261 {
1262 	bioq_insert_head(&isc->trim_queue, bp);
1263 #ifdef CAM_IOSCHED_DYNAMIC
1264 	isc->trim_stats.queued++;
1265 	isc->trim_stats.total--;		/* since we put it back, don't double count */
1266 	isc->trim_stats.pending--;
1267 #endif
1268 }
1269 
1270 /*
1271  * gets the next trim from the trim queue.
1272  *
1273  * Assumes we're called with the periph lock held.  It removes this
1274  * trim from the queue and the device must explicitly reinsert it
1275  * should the need arise.
1276  */
1277 struct bio *
1278 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1279 {
1280 	struct bio *bp;
1281 
1282 	bp  = bioq_first(&isc->trim_queue);
1283 	if (bp == NULL)
1284 		return NULL;
1285 	bioq_remove(&isc->trim_queue, bp);
1286 #ifdef CAM_IOSCHED_DYNAMIC
1287 	isc->trim_stats.queued--;
1288 	isc->trim_stats.total++;
1289 	isc->trim_stats.pending++;
1290 #endif
1291 	return bp;
1292 }
1293 
1294 /*
1295  * gets an available trim from the trim queue, if there's no trim
1296  * already pending. It removes this trim from the queue and the device
1297  * must explicitly reinsert it should the need arise.
1298  *
1299  * Assumes we're called with the periph lock held.
1300  */
1301 struct bio *
1302 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1303 {
1304 
1305 	if (!cam_iosched_has_more_trim(isc))
1306 		return NULL;
1307 
1308 	return cam_iosched_next_trim(isc);
1309 }
1310 
1311 /*
1312  * Determine what the next bit of work to do is for the periph. The
1313  * default implementation looks to see if we have trims to do, but no
1314  * trims outstanding. If so, we do that. Otherwise we see if we have
1315  * other work. If we do, then we do that. Otherwise why were we called?
1316  */
1317 struct bio *
1318 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1319 {
1320 	struct bio *bp;
1321 
1322 	/*
1323 	 * See if we have a trim that can be scheduled. We can only send one
1324 	 * at a time down, so this takes that into account.
1325 	 *
1326 	 * XXX newer TRIM commands are queueable. Revisit this when we
1327 	 * implement them.
1328 	 */
1329 	if ((bp = cam_iosched_get_trim(isc)) != NULL)
1330 		return bp;
1331 
1332 #ifdef CAM_IOSCHED_DYNAMIC
1333 	/*
1334 	 * See if we have any pending writes, and room in the queue for them,
1335 	 * and if so, those are next.
1336 	 */
1337 	if (do_dynamic_iosched) {
1338 		if ((bp = cam_iosched_get_write(isc)) != NULL)
1339 			return bp;
1340 	}
1341 #endif
1342 
1343 	/*
1344 	 * next, see if there's other, normal I/O waiting. If so return that.
1345 	 */
1346 	if ((bp = bioq_first(&isc->bio_queue)) == NULL)
1347 		return NULL;
1348 
1349 #ifdef CAM_IOSCHED_DYNAMIC
1350 	/*
1351 	 * For the dynamic scheduler, bio_queue is only for reads, so enforce
1352 	 * the limits here. Enforce only for reads.
1353 	 */
1354 	if (do_dynamic_iosched) {
1355 		if (bp->bio_cmd == BIO_READ &&
1356 		    cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1357 			isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1358 			return NULL;
1359 		}
1360 	}
1361 	isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1362 #endif
1363 	bioq_remove(&isc->bio_queue, bp);
1364 #ifdef CAM_IOSCHED_DYNAMIC
1365 	if (do_dynamic_iosched) {
1366 		if (bp->bio_cmd == BIO_READ) {
1367 			isc->read_stats.queued--;
1368 			isc->read_stats.total++;
1369 			isc->read_stats.pending++;
1370 		} else
1371 			printf("Found bio_cmd = %#x\n", bp->bio_cmd);
1372 	}
1373 	if (iosched_debug > 9)
1374 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1375 #endif
1376 	return bp;
1377 }
1378 
1379 /*
1380  * Driver has been given some work to do by the block layer. Tell the
1381  * scheduler about it and have it queue the work up. The scheduler module
1382  * will then return the currently most useful bit of work later, possibly
1383  * deferring work for various reasons.
1384  */
1385 void
1386 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1387 {
1388 
1389 	/*
1390 	 * Put all trims on the trim queue sorted, since we know
1391 	 * that the collapsing code requires this. Otherwise put
1392 	 * the work on the bio queue.
1393 	 */
1394 	if (bp->bio_cmd == BIO_DELETE) {
1395 		bioq_disksort(&isc->trim_queue, bp);
1396 #ifdef CAM_IOSCHED_DYNAMIC
1397 		isc->trim_stats.in++;
1398 		isc->trim_stats.queued++;
1399 #endif
1400 	}
1401 #ifdef CAM_IOSCHED_DYNAMIC
1402 	else if (do_dynamic_iosched && (bp->bio_cmd != BIO_READ)) {
1403 		if (cam_iosched_sort_queue(isc))
1404 			bioq_disksort(&isc->write_queue, bp);
1405 		else
1406 			bioq_insert_tail(&isc->write_queue, bp);
1407 		if (iosched_debug > 9)
1408 			printf("Qw  : %p %#x\n", bp, bp->bio_cmd);
1409 		if (bp->bio_cmd == BIO_WRITE) {
1410 			isc->write_stats.in++;
1411 			isc->write_stats.queued++;
1412 		}
1413 	}
1414 #endif
1415 	else {
1416 		if (cam_iosched_sort_queue(isc))
1417 			bioq_disksort(&isc->bio_queue, bp);
1418 		else
1419 			bioq_insert_tail(&isc->bio_queue, bp);
1420 #ifdef CAM_IOSCHED_DYNAMIC
1421 		if (iosched_debug > 9)
1422 			printf("Qr  : %p %#x\n", bp, bp->bio_cmd);
1423 		if (bp->bio_cmd == BIO_READ) {
1424 			isc->read_stats.in++;
1425 			isc->read_stats.queued++;
1426 		} else if (bp->bio_cmd == BIO_WRITE) {
1427 			isc->write_stats.in++;
1428 			isc->write_stats.queued++;
1429 		}
1430 #endif
1431 	}
1432 }
1433 
1434 /*
1435  * If we have work, get it scheduled. Called with the periph lock held.
1436  */
1437 void
1438 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1439 {
1440 
1441 	if (cam_iosched_has_work(isc))
1442 		xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1443 }
1444 
1445 /*
1446  * Complete a trim request. Mark that we no longer have one in flight.
1447  */
1448 void
1449 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1450 {
1451 
1452 	isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1453 }
1454 
1455 /*
1456  * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1457  * might use notes in the ccb for statistics.
1458  */
1459 int
1460 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1461     union ccb *done_ccb)
1462 {
1463 	int retval = 0;
1464 #ifdef CAM_IOSCHED_DYNAMIC
1465 	if (!do_dynamic_iosched)
1466 		return retval;
1467 
1468 	if (iosched_debug > 10)
1469 		printf("done: %p %#x\n", bp, bp->bio_cmd);
1470 	if (bp->bio_cmd == BIO_WRITE) {
1471 		retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1472 		if (!(bp->bio_flags & BIO_ERROR))
1473 			isc->write_stats.errs++;
1474 		isc->write_stats.out++;
1475 		isc->write_stats.pending--;
1476 	} else if (bp->bio_cmd == BIO_READ) {
1477 		retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1478 		if (!(bp->bio_flags & BIO_ERROR))
1479 			isc->read_stats.errs++;
1480 		isc->read_stats.out++;
1481 		isc->read_stats.pending--;
1482 	} else if (bp->bio_cmd == BIO_DELETE) {
1483 		if (!(bp->bio_flags & BIO_ERROR))
1484 			isc->trim_stats.errs++;
1485 		isc->trim_stats.out++;
1486 		isc->trim_stats.pending--;
1487 	} else if (bp->bio_cmd != BIO_FLUSH) {
1488 		if (iosched_debug)
1489 			printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1490 	}
1491 
1492 	if (!(bp->bio_flags & BIO_ERROR))
1493 		cam_iosched_io_metric_update(isc,
1494 		    cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data),
1495 		    bp->bio_cmd, bp->bio_bcount);
1496 #endif
1497 	return retval;
1498 }
1499 
1500 /*
1501  * Tell the io scheduler that you've pushed a trim down into the sim.
1502  * This also tells the I/O scheduler not to push any more trims down, so
1503  * some periphs do not call it if they can cope with multiple trims in flight.
1504  */
1505 void
1506 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1507 {
1508 
1509 	isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1510 }
1511 
1512 /*
1513  * Change the sorting policy hint for I/O transactions for this device.
1514  */
1515 void
1516 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1517 {
1518 
1519 	isc->sort_io_queue = val;
1520 }
1521 
1522 int
1523 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1524 {
1525 	return isc->flags & flags;
1526 }
1527 
1528 void
1529 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1530 {
1531 	isc->flags |= flags;
1532 }
1533 
1534 void
1535 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1536 {
1537 	isc->flags &= ~flags;
1538 }
1539 
1540 #ifdef CAM_IOSCHED_DYNAMIC
1541 /*
1542  * After the method presented in Jack Crenshaw's 1998 article "Integer
1543  * Square Roots," reprinted at
1544  * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1545  * and well worth the read. Briefly, we find the power of 4 that's the
1546  * largest smaller than val. We then check each smaller power of 4 to
1547  * see if val is still bigger. The right shifts at each step divide
1548  * the result by 2 which after successive application winds up
1549  * accumulating the right answer. It could also have been accumulated
1550  * using a separate root counter, but this code is smaller and faster
1551  * than that method. This method is also integer size invariant.
1552  * It returns floor(sqrt((float)val)), or the largest integer less than
1553  * or equal to the square root.
1554  */
1555 static uint64_t
1556 isqrt64(uint64_t val)
1557 {
1558 	uint64_t res = 0;
1559 	uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1560 
1561 	/*
1562 	 * Find the largest power of 4 smaller than val.
1563 	 */
1564 	while (bit > val)
1565 		bit >>= 2;
1566 
1567 	/*
1568 	 * Accumulate the answer, one bit at a time (we keep moving
1569 	 * them over since 2 is the square root of 4 and we test
1570 	 * powers of 4). We accumulate where we find the bit, but
1571 	 * the successive shifts land the bit in the right place
1572 	 * by the end.
1573 	 */
1574 	while (bit != 0) {
1575 		if (val >= res + bit) {
1576 			val -= res + bit;
1577 			res = (res >> 1) + bit;
1578 		} else
1579 			res >>= 1;
1580 		bit >>= 2;
1581 	}
1582 
1583 	return res;
1584 }
1585 
1586 static sbintime_t latencies[LAT_BUCKETS - 1] = {
1587 	SBT_1MS <<  0,
1588 	SBT_1MS <<  1,
1589 	SBT_1MS <<  2,
1590 	SBT_1MS <<  3,
1591 	SBT_1MS <<  4,
1592 	SBT_1MS <<  5,
1593 	SBT_1MS <<  6,
1594 	SBT_1MS <<  7,
1595 	SBT_1MS <<  8,
1596 	SBT_1MS <<  9,
1597 	SBT_1MS << 10,
1598 	SBT_1MS << 11,
1599 	SBT_1MS << 12,
1600 	SBT_1MS << 13		/* 8.192s */
1601 };
1602 
1603 static void
1604 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1605 {
1606 	sbintime_t y, deltasq, delta;
1607 	int i;
1608 
1609 	/*
1610 	 * Keep counts for latency. We do it by power of two buckets.
1611 	 * This helps us spot outlier behavior obscured by averages.
1612 	 */
1613 	for (i = 0; i < LAT_BUCKETS - 1; i++) {
1614 		if (sim_latency < latencies[i]) {
1615 			iop->latencies[i]++;
1616 			break;
1617 		}
1618 	}
1619 	if (i == LAT_BUCKETS - 1)
1620 		iop->latencies[i]++; 	 /* Put all > 1024ms values into the last bucket. */
1621 
1622 	/*
1623 	 * Classic exponentially decaying average with a tiny alpha
1624 	 * (2 ^ -alpha_bits). For more info see the NIST statistical
1625 	 * handbook.
1626 	 *
1627 	 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha)		[nist]
1628 	 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
1629 	 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1630 	 * alpha = 1 / (1 << alpha_bits)
1631 	 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
1632 	 *	= y_t/b - e/b + be/b
1633 	 *      = (y_t - e + be) / b
1634 	 *	= (e + d) / b
1635 	 *
1636 	 * Since alpha is a power of two, we can compute this w/o any mult or
1637 	 * division.
1638 	 *
1639 	 * Variance can also be computed. Usually, it would be expressed as follows:
1640 	 *	diff_t = y_t - ema_t-1
1641 	 *	emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
1642 	 *	  = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
1643 	 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
1644 	 *	  = e - e/b + dd/b + dd/bb
1645 	 *	  = (bbe - be + bdd + dd) / bb
1646 	 *	  = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
1647 	 */
1648 	/*
1649 	 * XXX possible numeric issues
1650 	 *	o We assume right shifted integers do the right thing, since that's
1651 	 *	  implementation defined. You can change the right shifts to / (1LL << alpha).
1652 	 *	o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
1653 	 *	  for emvar. This puts a ceiling of 13 bits on alpha since we need a
1654 	 *	  few tens of seconds of representation.
1655 	 *	o We mitigate alpha issues by never setting it too high.
1656 	 */
1657 	y = sim_latency;
1658 	delta = (y - iop->ema);					/* d */
1659 	iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1660 
1661 	/*
1662 	 * Were we to naively plow ahead at this point, we wind up with many numerical
1663 	 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
1664 	 * us with microsecond level precision in the input, so the same in the
1665 	 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
1666 	 * also means that emvar can be up 46 bits 40 of which are fraction, which
1667 	 * gives us a way to measure up to ~8s in the SD before the computation goes
1668 	 * unstable. Even the worst hard disk rarely has > 1s service time in the
1669 	 * drive. It does mean we have to shift left 12 bits after taking the
1670 	 * square root to compute the actual standard deviation estimate. This loss of
1671 	 * precision is preferable to needing int128 types to work. The above numbers
1672 	 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
1673 	 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1674 	 */
1675 	delta >>= 12;
1676 	deltasq = delta * delta;				/* dd */
1677 	iop->emvar = ((iop->emvar << (2 * alpha_bits)) +	/* bbe */
1678 	    ((deltasq - iop->emvar) << alpha_bits) +		/* b(dd-e) */
1679 	    deltasq)						/* dd */
1680 	    >> (2 * alpha_bits);				/* div bb */
1681 	iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1682 }
1683 
1684 static void
1685 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1686     sbintime_t sim_latency, int cmd, size_t size)
1687 {
1688 	/* xxx Do we need to scale based on the size of the I/O ? */
1689 	switch (cmd) {
1690 	case BIO_READ:
1691 		cam_iosched_update(&isc->read_stats, sim_latency);
1692 		break;
1693 	case BIO_WRITE:
1694 		cam_iosched_update(&isc->write_stats, sim_latency);
1695 		break;
1696 	case BIO_DELETE:
1697 		cam_iosched_update(&isc->trim_stats, sim_latency);
1698 		break;
1699 	default:
1700 		break;
1701 	}
1702 }
1703 
1704 #ifdef DDB
1705 static int biolen(struct bio_queue_head *bq)
1706 {
1707 	int i = 0;
1708 	struct bio *bp;
1709 
1710 	TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
1711 		i++;
1712 	}
1713 	return i;
1714 }
1715 
1716 /*
1717  * Show the internal state of the I/O scheduler.
1718  */
1719 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
1720 {
1721 	struct cam_iosched_softc *isc;
1722 
1723 	if (!have_addr) {
1724 		db_printf("Need addr\n");
1725 		return;
1726 	}
1727 	isc = (struct cam_iosched_softc *)addr;
1728 	db_printf("pending_reads:     %d\n", isc->read_stats.pending);
1729 	db_printf("min_reads:         %d\n", isc->read_stats.min);
1730 	db_printf("max_reads:         %d\n", isc->read_stats.max);
1731 	db_printf("reads:             %d\n", isc->read_stats.total);
1732 	db_printf("in_reads:          %d\n", isc->read_stats.in);
1733 	db_printf("out_reads:         %d\n", isc->read_stats.out);
1734 	db_printf("queued_reads:      %d\n", isc->read_stats.queued);
1735 	db_printf("Current Q len      %d\n", biolen(&isc->bio_queue));
1736 	db_printf("pending_writes:    %d\n", isc->write_stats.pending);
1737 	db_printf("min_writes:        %d\n", isc->write_stats.min);
1738 	db_printf("max_writes:        %d\n", isc->write_stats.max);
1739 	db_printf("writes:            %d\n", isc->write_stats.total);
1740 	db_printf("in_writes:         %d\n", isc->write_stats.in);
1741 	db_printf("out_writes:        %d\n", isc->write_stats.out);
1742 	db_printf("queued_writes:     %d\n", isc->write_stats.queued);
1743 	db_printf("Current Q len      %d\n", biolen(&isc->write_queue));
1744 	db_printf("pending_trims:     %d\n", isc->trim_stats.pending);
1745 	db_printf("min_trims:         %d\n", isc->trim_stats.min);
1746 	db_printf("max_trims:         %d\n", isc->trim_stats.max);
1747 	db_printf("trims:             %d\n", isc->trim_stats.total);
1748 	db_printf("in_trims:          %d\n", isc->trim_stats.in);
1749 	db_printf("out_trims:         %d\n", isc->trim_stats.out);
1750 	db_printf("queued_trims:      %d\n", isc->trim_stats.queued);
1751 	db_printf("Current Q len      %d\n", biolen(&isc->trim_queue));
1752 	db_printf("read_bias:         %d\n", isc->read_bias);
1753 	db_printf("current_read_bias: %d\n", isc->current_read_bias);
1754 	db_printf("Trim active?       %s\n",
1755 	    (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");
1756 }
1757 #endif
1758 #endif
1759