xref: /freebsd/sys/cam/cam_iosched.c (revision d06955f9bdb1416d9196043ed781f9b36dae9adc)
1 /*-
2  * CAM IO Scheduler Interface
3  *
4  * Copyright (c) 2015 Netflix, Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions, and the following disclaimer,
12  *    without modification, immediately at the beginning of the file.
13  * 2. The name of the author may not be used to endorse or promote products
14  *    derived from this software without specific prior written permission.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
20  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * $FreeBSD$
29  */
30 
31 #include "opt_cam.h"
32 #include "opt_ddb.h"
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include <sys/param.h>
38 
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
41 #include <sys/bio.h>
42 #include <sys/lock.h>
43 #include <sys/malloc.h>
44 #include <sys/mutex.h>
45 #include <sys/sbuf.h>
46 #include <sys/sysctl.h>
47 
48 #include <cam/cam.h>
49 #include <cam/cam_ccb.h>
50 #include <cam/cam_periph.h>
51 #include <cam/cam_xpt_periph.h>
52 #include <cam/cam_xpt_internal.h>
53 #include <cam/cam_iosched.h>
54 
55 #include <ddb/ddb.h>
56 
57 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler",
58     "CAM I/O Scheduler buffers");
59 
60 /*
61  * Default I/O scheduler for FreeBSD. This implementation is just a thin-vineer
62  * over the bioq_* interface, with notions of separate calls for normal I/O and
63  * for trims.
64  *
65  * When CAM_IOSCHED_DYNAMIC is defined, the scheduler is enhanced to dynamically
66  * steer the rate of one type of traffic to help other types of traffic (eg
67  * limit writes when read latency deteriorates on SSDs).
68  */
69 
70 #ifdef CAM_IOSCHED_DYNAMIC
71 
72 static int do_dynamic_iosched = 1;
73 TUNABLE_INT("kern.cam.do_dynamic_iosched", &do_dynamic_iosched);
74 SYSCTL_INT(_kern_cam, OID_AUTO, do_dynamic_iosched, CTLFLAG_RD,
75     &do_dynamic_iosched, 1,
76     "Enable Dynamic I/O scheduler optimizations.");
77 
78 /*
79  * For an EMA, with an alpha of alpha, we know
80  * 	alpha = 2 / (N + 1)
81  * or
82  * 	N = 1 + (2 / alpha)
83  * where N is the number of samples that 86% of the current
84  * EMA is derived from.
85  *
86  * So we invent[*] alpha_bits:
87  *	alpha_bits = -log_2(alpha)
88  *	alpha = 2^-alpha_bits
89  * So
90  *	N = 1 + 2^(alpha_bits + 1)
91  *
92  * The default 9 gives a 1025 lookback for 86% of the data.
93  * For a brief intro: https://en.wikipedia.org/wiki/Moving_average
94  *
95  * [*] Steal from the load average code and many other places.
96  * Note: See computation of EMA and EMVAR for acceptable ranges of alpha.
97  */
98 static int alpha_bits = 9;
99 TUNABLE_INT("kern.cam.iosched_alpha_bits", &alpha_bits);
100 SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW,
101     &alpha_bits, 1,
102     "Bits in EMA's alpha.");
103 
104 struct iop_stats;
105 struct cam_iosched_softc;
106 
107 int iosched_debug = 0;
108 
109 typedef enum {
110 	none = 0,				/* No limits */
111 	queue_depth,			/* Limit how many ops we queue to SIM */
112 	iops,				/* Limit # of IOPS to the drive */
113 	bandwidth,			/* Limit bandwidth to the drive */
114 	limiter_max
115 } io_limiter;
116 
117 static const char *cam_iosched_limiter_names[] =
118     { "none", "queue_depth", "iops", "bandwidth" };
119 
120 /*
121  * Called to initialize the bits of the iop_stats structure relevant to the
122  * limiter. Called just after the limiter is set.
123  */
124 typedef int l_init_t(struct iop_stats *);
125 
126 /*
127  * Called every tick.
128  */
129 typedef int l_tick_t(struct iop_stats *);
130 
131 /*
132  * Called to see if the limiter thinks this IOP can be allowed to
133  * proceed. If so, the limiter assumes that the IOP proceeded
134  * and makes any accounting of it that's needed.
135  */
136 typedef int l_iop_t(struct iop_stats *, struct bio *);
137 
138 /*
139  * Called when an I/O completes so the limiter can update its
140  * accounting. Pending I/Os may complete in any order (even when
141  * sent to the hardware at the same time), so the limiter may not
142  * make any assumptions other than this I/O has completed. If it
143  * returns 1, then xpt_schedule() needs to be called again.
144  */
145 typedef int l_iodone_t(struct iop_stats *, struct bio *);
146 
147 static l_iop_t cam_iosched_qd_iop;
148 static l_iop_t cam_iosched_qd_caniop;
149 static l_iodone_t cam_iosched_qd_iodone;
150 
151 static l_init_t cam_iosched_iops_init;
152 static l_tick_t cam_iosched_iops_tick;
153 static l_iop_t cam_iosched_iops_caniop;
154 static l_iop_t cam_iosched_iops_iop;
155 
156 static l_init_t cam_iosched_bw_init;
157 static l_tick_t cam_iosched_bw_tick;
158 static l_iop_t cam_iosched_bw_caniop;
159 static l_iop_t cam_iosched_bw_iop;
160 
161 struct limswitch {
162 	l_init_t	*l_init;
163 	l_tick_t	*l_tick;
164 	l_iop_t		*l_iop;
165 	l_iop_t		*l_caniop;
166 	l_iodone_t	*l_iodone;
167 } limsw[] =
168 {
169 	{	/* none */
170 		.l_init = NULL,
171 		.l_tick = NULL,
172 		.l_iop = NULL,
173 		.l_iodone= NULL,
174 	},
175 	{	/* queue_depth */
176 		.l_init = NULL,
177 		.l_tick = NULL,
178 		.l_caniop = cam_iosched_qd_caniop,
179 		.l_iop = cam_iosched_qd_iop,
180 		.l_iodone= cam_iosched_qd_iodone,
181 	},
182 	{	/* iops */
183 		.l_init = cam_iosched_iops_init,
184 		.l_tick = cam_iosched_iops_tick,
185 		.l_caniop = cam_iosched_iops_caniop,
186 		.l_iop = cam_iosched_iops_iop,
187 		.l_iodone= NULL,
188 	},
189 	{	/* bandwidth */
190 		.l_init = cam_iosched_bw_init,
191 		.l_tick = cam_iosched_bw_tick,
192 		.l_caniop = cam_iosched_bw_caniop,
193 		.l_iop = cam_iosched_bw_iop,
194 		.l_iodone= NULL,
195 	},
196 };
197 
198 struct iop_stats {
199 	/*
200 	 * sysctl state for this subnode.
201 	 */
202 	struct sysctl_ctx_list	sysctl_ctx;
203 	struct sysctl_oid	*sysctl_tree;
204 
205 	/*
206 	 * Information about the current rate limiters, if any
207 	 */
208 	io_limiter	limiter;	/* How are I/Os being limited */
209 	int		min;		/* Low range of limit */
210 	int		max;		/* High range of limit */
211 	int		current;	/* Current rate limiter */
212 	int		l_value1;	/* per-limiter scratch value 1. */
213 	int		l_value2;	/* per-limiter scratch value 2. */
214 
215 	/*
216 	 * Debug information about counts of I/Os that have gone through the
217 	 * scheduler.
218 	 */
219 	int		pending;	/* I/Os pending in the hardware */
220 	int		queued;		/* number currently in the queue */
221 	int		total;		/* Total for all time -- wraps */
222 	int		in;		/* number queued all time -- wraps */
223 	int		out;		/* number completed all time -- wraps */
224 
225 	/*
226 	 * Statistics on different bits of the process.
227 	 */
228 		/* Exp Moving Average, see alpha_bits for more details */
229 	sbintime_t      ema;
230 	sbintime_t      emvar;
231 	sbintime_t      sd;		/* Last computed sd */
232 
233 	uint32_t	state_flags;
234 #define IOP_RATE_LIMITED		1u
235 
236 #define LAT_BUCKETS 15			/* < 1ms < 2ms ... < 2^(n-1)ms >= 2^(n-1)ms*/
237 	uint64_t	latencies[LAT_BUCKETS];
238 
239 	struct cam_iosched_softc *softc;
240 };
241 
242 
243 typedef enum {
244 	set_max = 0,			/* current = max */
245 	read_latency,			/* Steer read latency by throttling writes */
246 	cl_max				/* Keep last */
247 } control_type;
248 
249 static const char *cam_iosched_control_type_names[] =
250     { "set_max", "read_latency" };
251 
252 struct control_loop {
253 	/*
254 	 * sysctl state for this subnode.
255 	 */
256 	struct sysctl_ctx_list	sysctl_ctx;
257 	struct sysctl_oid	*sysctl_tree;
258 
259 	sbintime_t	next_steer;		/* Time of next steer */
260 	sbintime_t	steer_interval;		/* How often do we steer? */
261 	sbintime_t	lolat;
262 	sbintime_t	hilat;
263 	int		alpha;
264 	control_type	type;			/* What type of control? */
265 	int		last_count;		/* Last I/O count */
266 
267 	struct cam_iosched_softc *softc;
268 };
269 
270 #endif
271 
272 struct cam_iosched_softc {
273 	struct bio_queue_head bio_queue;
274 	struct bio_queue_head trim_queue;
275 				/* scheduler flags < 16, user flags >= 16 */
276 	uint32_t	flags;
277 	int		sort_io_queue;
278 #ifdef CAM_IOSCHED_DYNAMIC
279 	int		read_bias;		/* Read bias setting */
280 	int		current_read_bias;	/* Current read bias state */
281 	int		total_ticks;
282 	int		load;			/* EMA of 'load average' of disk / 2^16 */
283 
284 	struct bio_queue_head write_queue;
285 	struct iop_stats read_stats, write_stats, trim_stats;
286 	struct sysctl_ctx_list	sysctl_ctx;
287 	struct sysctl_oid	*sysctl_tree;
288 
289 	int		quanta;			/* Number of quanta per second */
290 	struct callout	ticker;			/* Callout for our quota system */
291 	struct cam_periph *periph;		/* cam periph associated with this device */
292 	uint32_t	this_frac;		/* Fraction of a second (1024ths) for this tick */
293 	sbintime_t	last_time;		/* Last time we ticked */
294 	struct control_loop cl;
295 #endif
296 };
297 
298 #ifdef CAM_IOSCHED_DYNAMIC
299 /*
300  * helper functions to call the limsw functions.
301  */
302 static int
303 cam_iosched_limiter_init(struct iop_stats *ios)
304 {
305 	int lim = ios->limiter;
306 
307 	/* maybe this should be a kassert */
308 	if (lim < none || lim >= limiter_max)
309 		return EINVAL;
310 
311 	if (limsw[lim].l_init)
312 		return limsw[lim].l_init(ios);
313 
314 	return 0;
315 }
316 
317 static int
318 cam_iosched_limiter_tick(struct iop_stats *ios)
319 {
320 	int lim = ios->limiter;
321 
322 	/* maybe this should be a kassert */
323 	if (lim < none || lim >= limiter_max)
324 		return EINVAL;
325 
326 	if (limsw[lim].l_tick)
327 		return limsw[lim].l_tick(ios);
328 
329 	return 0;
330 }
331 
332 static int
333 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
334 {
335 	int lim = ios->limiter;
336 
337 	/* maybe this should be a kassert */
338 	if (lim < none || lim >= limiter_max)
339 		return EINVAL;
340 
341 	if (limsw[lim].l_iop)
342 		return limsw[lim].l_iop(ios, bp);
343 
344 	return 0;
345 }
346 
347 static int
348 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
349 {
350 	int lim = ios->limiter;
351 
352 	/* maybe this should be a kassert */
353 	if (lim < none || lim >= limiter_max)
354 		return EINVAL;
355 
356 	if (limsw[lim].l_caniop)
357 		return limsw[lim].l_caniop(ios, bp);
358 
359 	return 0;
360 }
361 
362 static int
363 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
364 {
365 	int lim = ios->limiter;
366 
367 	/* maybe this should be a kassert */
368 	if (lim < none || lim >= limiter_max)
369 		return 0;
370 
371 	if (limsw[lim].l_iodone)
372 		return limsw[lim].l_iodone(ios, bp);
373 
374 	return 0;
375 }
376 
377 /*
378  * Functions to implement the different kinds of limiters
379  */
380 
381 static int
382 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
383 {
384 
385 	if (ios->current <= 0 || ios->pending < ios->current)
386 		return 0;
387 
388 	return EAGAIN;
389 }
390 
391 static int
392 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
393 {
394 
395 	if (ios->current <= 0 || ios->pending < ios->current)
396 		return 0;
397 
398 	return EAGAIN;
399 }
400 
401 static int
402 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
403 {
404 
405 	if (ios->current <= 0 || ios->pending != ios->current)
406 		return 0;
407 
408 	return 1;
409 }
410 
411 static int
412 cam_iosched_iops_init(struct iop_stats *ios)
413 {
414 
415 	ios->l_value1 = ios->current / ios->softc->quanta;
416 	if (ios->l_value1 <= 0)
417 		ios->l_value1 = 1;
418 	ios->l_value2 = 0;
419 
420 	return 0;
421 }
422 
423 static int
424 cam_iosched_iops_tick(struct iop_stats *ios)
425 {
426 	int new_ios;
427 
428 	/*
429 	 * Allow at least one IO per tick until all
430 	 * the IOs for this interval have been spent.
431 	 */
432 	new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
433 	if (new_ios < 1 && ios->l_value2 < ios->current) {
434 		new_ios = 1;
435 		ios->l_value2++;
436 	}
437 
438 	/*
439 	 * If this a new accounting interval, discard any "unspent" ios
440 	 * granted in the previous interval.  Otherwise add the new ios to
441 	 * the previously granted ones that haven't been spent yet.
442 	 */
443 	if ((ios->softc->total_ticks % ios->softc->quanta) == 0) {
444 		ios->l_value1 = new_ios;
445 		ios->l_value2 = 1;
446 	} else {
447 		ios->l_value1 += new_ios;
448 	}
449 
450 
451 	return 0;
452 }
453 
454 static int
455 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
456 {
457 
458 	/*
459 	 * So if we have any more IOPs left, allow it,
460 	 * otherwise wait. If current iops is 0, treat that
461 	 * as unlimited as a failsafe.
462 	 */
463 	if (ios->current > 0 && ios->l_value1 <= 0)
464 		return EAGAIN;
465 	return 0;
466 }
467 
468 static int
469 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
470 {
471 	int rv;
472 
473 	rv = cam_iosched_limiter_caniop(ios, bp);
474 	if (rv == 0)
475 		ios->l_value1--;
476 
477 	return rv;
478 }
479 
480 static int
481 cam_iosched_bw_init(struct iop_stats *ios)
482 {
483 
484 	/* ios->current is in kB/s, so scale to bytes */
485 	ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
486 
487 	return 0;
488 }
489 
490 static int
491 cam_iosched_bw_tick(struct iop_stats *ios)
492 {
493 	int bw;
494 
495 	/*
496 	 * If we're in the hole for available quota from
497 	 * the last time, then add the quantum for this.
498 	 * If we have any left over from last quantum,
499 	 * then too bad, that's lost. Also, ios->current
500 	 * is in kB/s, so scale.
501 	 *
502 	 * We also allow up to 4 quanta of credits to
503 	 * accumulate to deal with burstiness. 4 is extremely
504 	 * arbitrary.
505 	 */
506 	bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
507 	if (ios->l_value1 < bw * 4)
508 		ios->l_value1 += bw;
509 
510 	return 0;
511 }
512 
513 static int
514 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
515 {
516 	/*
517 	 * So if we have any more bw quota left, allow it,
518 	 * otherwise wait. Note, we'll go negative and that's
519 	 * OK. We'll just get a little less next quota.
520 	 *
521 	 * Note on going negative: that allows us to process
522 	 * requests in order better, since we won't allow
523 	 * shorter reads to get around the long one that we
524 	 * don't have the quota to do just yet. It also prevents
525 	 * starvation by being a little more permissive about
526 	 * what we let through this quantum (to prevent the
527 	 * starvation), at the cost of getting a little less
528 	 * next quantum.
529 	 *
530 	 * Also note that if the current limit is <= 0,
531 	 * we treat it as unlimited as a failsafe.
532 	 */
533 	if (ios->current > 0 && ios->l_value1 <= 0)
534 		return EAGAIN;
535 
536 
537 	return 0;
538 }
539 
540 static int
541 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
542 {
543 	int rv;
544 
545 	rv = cam_iosched_limiter_caniop(ios, bp);
546 	if (rv == 0)
547 		ios->l_value1 -= bp->bio_length;
548 
549 	return rv;
550 }
551 
552 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
553 
554 static void
555 cam_iosched_ticker(void *arg)
556 {
557 	struct cam_iosched_softc *isc = arg;
558 	sbintime_t now, delta;
559 	int pending;
560 
561 	callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
562 
563 	now = sbinuptime();
564 	delta = now - isc->last_time;
565 	isc->this_frac = (uint32_t)delta >> 16;		/* Note: discards seconds -- should be 0 harmless if not */
566 	isc->last_time = now;
567 
568 	cam_iosched_cl_maybe_steer(&isc->cl);
569 
570 	cam_iosched_limiter_tick(&isc->read_stats);
571 	cam_iosched_limiter_tick(&isc->write_stats);
572 	cam_iosched_limiter_tick(&isc->trim_stats);
573 
574 	cam_iosched_schedule(isc, isc->periph);
575 
576 	/*
577 	 * isc->load is an EMA of the pending I/Os at each tick. The number of
578 	 * pending I/Os is the sum of the I/Os queued to the hardware, and those
579 	 * in the software queue that could be queued to the hardware if there
580 	 * were slots.
581 	 *
582 	 * ios_stats.pending is a count of requests in the SIM right now for
583 	 * each of these types of I/O. So the total pending count is the sum of
584 	 * these I/Os and the sum of the queued I/Os still in the software queue
585 	 * for those operations that aren't being rate limited at the moment.
586 	 *
587 	 * The reason for the rate limiting bit is because those I/Os
588 	 * aren't part of the software queued load (since we could
589 	 * give them to hardware, but choose not to).
590 	 *
591 	 * Note: due to a bug in counting pending TRIM in the device, we
592 	 * don't include them in this count. We count each BIO_DELETE in
593 	 * the pending count, but the periph drivers collapse them down
594 	 * into one TRIM command. That one trim command gets the completion
595 	 * so the counts get off.
596 	 */
597 	pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
598 	pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
599 	    !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
600 	    !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
601 	pending <<= 16;
602 	pending /= isc->periph->path->device->ccbq.total_openings;
603 
604 	isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
605 
606 	isc->total_ticks++;
607 }
608 
609 
610 static void
611 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
612 {
613 
614 	clp->next_steer = sbinuptime();
615 	clp->softc = isc;
616 	clp->steer_interval = SBT_1S * 5;	/* Let's start out steering every 5s */
617 	clp->lolat = 5 * SBT_1MS;
618 	clp->hilat = 15 * SBT_1MS;
619 	clp->alpha = 20;			/* Alpha == gain. 20 = .2 */
620 	clp->type = set_max;
621 }
622 
623 static void
624 cam_iosched_cl_maybe_steer(struct control_loop *clp)
625 {
626 	struct cam_iosched_softc *isc;
627 	sbintime_t now, lat;
628 	int old;
629 
630 	isc = clp->softc;
631 	now = isc->last_time;
632 	if (now < clp->next_steer)
633 		return;
634 
635 	clp->next_steer = now + clp->steer_interval;
636 	switch (clp->type) {
637 	case set_max:
638 		if (isc->write_stats.current != isc->write_stats.max)
639 			printf("Steering write from %d kBps to %d kBps\n",
640 			    isc->write_stats.current, isc->write_stats.max);
641 		isc->read_stats.current = isc->read_stats.max;
642 		isc->write_stats.current = isc->write_stats.max;
643 		isc->trim_stats.current = isc->trim_stats.max;
644 		break;
645 	case read_latency:
646 		old = isc->write_stats.current;
647 		lat = isc->read_stats.ema;
648 		/*
649 		 * Simple PLL-like engine. Since we're steering to a range for
650 		 * the SP (set point) that makes things a little more
651 		 * complicated. In addition, we're not directly controlling our
652 		 * PV (process variable), the read latency, but instead are
653 		 * manipulating the write bandwidth limit for our MV
654 		 * (manipulation variable), analysis of this code gets a bit
655 		 * messy. Also, the MV is a very noisy control surface for read
656 		 * latency since it is affected by many hidden processes inside
657 		 * the device which change how responsive read latency will be
658 		 * in reaction to changes in write bandwidth. Unlike the classic
659 		 * boiler control PLL. this may result in over-steering while
660 		 * the SSD takes its time to react to the new, lower load. This
661 		 * is why we use a relatively low alpha of between .1 and .25 to
662 		 * compensate for this effect. At .1, it takes ~22 steering
663 		 * intervals to back off by a factor of 10. At .2 it only takes
664 		 * ~10. At .25 it only takes ~8. However some preliminary data
665 		 * from the SSD drives suggests a reasponse time in 10's of
666 		 * seconds before latency drops regardless of the new write
667 		 * rate. Careful observation will be required to tune this
668 		 * effectively.
669 		 *
670 		 * Also, when there's no read traffic, we jack up the write
671 		 * limit too regardless of the last read latency.  10 is
672 		 * somewhat arbitrary.
673 		 */
674 		if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
675 			isc->write_stats.current = isc->write_stats.current *
676 			    (100 + clp->alpha) / 100;	/* Scale up */
677 		else if (lat > clp->hilat)
678 			isc->write_stats.current = isc->write_stats.current *
679 			    (100 - clp->alpha) / 100;	/* Scale down */
680 		clp->last_count = isc->read_stats.total;
681 
682 		/*
683 		 * Even if we don't steer, per se, enforce the min/max limits as
684 		 * those may have changed.
685 		 */
686 		if (isc->write_stats.current < isc->write_stats.min)
687 			isc->write_stats.current = isc->write_stats.min;
688 		if (isc->write_stats.current > isc->write_stats.max)
689 			isc->write_stats.current = isc->write_stats.max;
690 		if (old != isc->write_stats.current && 	iosched_debug)
691 			printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
692 			    old, isc->write_stats.current,
693 			    (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
694 		break;
695 	case cl_max:
696 		break;
697 	}
698 }
699 #endif
700 
701 /*
702  * Trim or similar currently pending completion. Should only be set for
703  * those drivers wishing only one Trim active at a time.
704  */
705 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE	(1ul << 0)
706 			/* Callout active, and needs to be torn down */
707 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
708 
709 			/* Periph drivers set these flags to indicate work */
710 #define CAM_IOSCHED_FLAG_WORK_FLAGS	((0xffffu) << 16)
711 
712 #ifdef CAM_IOSCHED_DYNAMIC
713 static void
714 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
715     sbintime_t sim_latency, int cmd, size_t size);
716 #endif
717 
718 static inline int
719 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
720 {
721 	return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
722 }
723 
724 static inline int
725 cam_iosched_has_io(struct cam_iosched_softc *isc)
726 {
727 #ifdef CAM_IOSCHED_DYNAMIC
728 	if (do_dynamic_iosched) {
729 		struct bio *rbp = bioq_first(&isc->bio_queue);
730 		struct bio *wbp = bioq_first(&isc->write_queue);
731 		int can_write = wbp != NULL &&
732 		    cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
733 		int can_read = rbp != NULL &&
734 		    cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
735 		if (iosched_debug > 2) {
736 			printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
737 			printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
738 			printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
739 		}
740 		return can_read || can_write;
741 	}
742 #endif
743 	return bioq_first(&isc->bio_queue) != NULL;
744 }
745 
746 static inline int
747 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
748 {
749 	return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) &&
750 	    bioq_first(&isc->trim_queue);
751 }
752 
753 #define cam_iosched_sort_queue(isc)	((isc)->sort_io_queue >= 0 ?	\
754     (isc)->sort_io_queue : cam_sort_io_queues)
755 
756 
757 static inline int
758 cam_iosched_has_work(struct cam_iosched_softc *isc)
759 {
760 #ifdef CAM_IOSCHED_DYNAMIC
761 	if (iosched_debug > 2)
762 		printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
763 		    cam_iosched_has_more_trim(isc),
764 		    cam_iosched_has_flagged_work(isc));
765 #endif
766 
767 	return cam_iosched_has_io(isc) ||
768 		cam_iosched_has_more_trim(isc) ||
769 		cam_iosched_has_flagged_work(isc);
770 }
771 
772 #ifdef CAM_IOSCHED_DYNAMIC
773 static void
774 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
775 {
776 
777 	ios->limiter = none;
778 	ios->in = 0;
779 	ios->max = ios->current = 300000;
780 	ios->min = 1;
781 	ios->out = 0;
782 	ios->pending = 0;
783 	ios->queued = 0;
784 	ios->total = 0;
785 	ios->ema = 0;
786 	ios->emvar = 0;
787 	ios->softc = isc;
788 	cam_iosched_limiter_init(ios);
789 }
790 
791 static int
792 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
793 {
794 	char buf[16];
795 	struct iop_stats *ios;
796 	struct cam_iosched_softc *isc;
797 	int value, i, error;
798 	const char *p;
799 
800 	ios = arg1;
801 	isc = ios->softc;
802 	value = ios->limiter;
803 	if (value < none || value >= limiter_max)
804 		p = "UNKNOWN";
805 	else
806 		p = cam_iosched_limiter_names[value];
807 
808 	strlcpy(buf, p, sizeof(buf));
809 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
810 	if (error != 0 || req->newptr == NULL)
811 		return error;
812 
813 	cam_periph_lock(isc->periph);
814 
815 	for (i = none; i < limiter_max; i++) {
816 		if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
817 			continue;
818 		ios->limiter = i;
819 		error = cam_iosched_limiter_init(ios);
820 		if (error != 0) {
821 			ios->limiter = value;
822 			cam_periph_unlock(isc->periph);
823 			return error;
824 		}
825 		/* Note: disk load averate requires ticker to be always running */
826 		callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
827 		isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
828 
829 		cam_periph_unlock(isc->periph);
830 		return 0;
831 	}
832 
833 	cam_periph_unlock(isc->periph);
834 	return EINVAL;
835 }
836 
837 static int
838 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
839 {
840 	char buf[16];
841 	struct control_loop *clp;
842 	struct cam_iosched_softc *isc;
843 	int value, i, error;
844 	const char *p;
845 
846 	clp = arg1;
847 	isc = clp->softc;
848 	value = clp->type;
849 	if (value < none || value >= cl_max)
850 		p = "UNKNOWN";
851 	else
852 		p = cam_iosched_control_type_names[value];
853 
854 	strlcpy(buf, p, sizeof(buf));
855 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
856 	if (error != 0 || req->newptr == NULL)
857 		return error;
858 
859 	for (i = set_max; i < cl_max; i++) {
860 		if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
861 			continue;
862 		cam_periph_lock(isc->periph);
863 		clp->type = i;
864 		cam_periph_unlock(isc->periph);
865 		return 0;
866 	}
867 
868 	return EINVAL;
869 }
870 
871 static int
872 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
873 {
874 	char buf[16];
875 	sbintime_t value;
876 	int error;
877 	uint64_t us;
878 
879 	value = *(sbintime_t *)arg1;
880 	us = (uint64_t)value / SBT_1US;
881 	snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
882 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
883 	if (error != 0 || req->newptr == NULL)
884 		return error;
885 	us = strtoul(buf, NULL, 10);
886 	if (us == 0)
887 		return EINVAL;
888 	*(sbintime_t *)arg1 = us * SBT_1US;
889 	return 0;
890 }
891 
892 static int
893 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
894 {
895 	int i, error;
896 	struct sbuf sb;
897 	uint64_t *latencies;
898 
899 	latencies = arg1;
900 	sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
901 
902 	for (i = 0; i < LAT_BUCKETS - 1; i++)
903 		sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
904 	sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
905 	error = sbuf_finish(&sb);
906 	sbuf_delete(&sb);
907 
908 	return (error);
909 }
910 
911 static int
912 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS)
913 {
914 	int *quanta;
915 	int error, value;
916 
917 	quanta = (unsigned *)arg1;
918 	value = *quanta;
919 
920 	error = sysctl_handle_int(oidp, (int *)&value, 0, req);
921 	if ((error != 0) || (req->newptr == NULL))
922 		return (error);
923 
924 	if (value < 1 || value > hz)
925 		return (EINVAL);
926 
927 	*quanta = value;
928 
929 	return (0);
930 }
931 
932 static void
933 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
934 {
935 	struct sysctl_oid_list *n;
936 	struct sysctl_ctx_list *ctx;
937 
938 	ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
939 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
940 	    CTLFLAG_RD, 0, name);
941 	n = SYSCTL_CHILDREN(ios->sysctl_tree);
942 	ctx = &ios->sysctl_ctx;
943 
944 	SYSCTL_ADD_UQUAD(ctx, n,
945 	    OID_AUTO, "ema", CTLFLAG_RD,
946 	    &ios->ema,
947 	    "Fast Exponentially Weighted Moving Average");
948 	SYSCTL_ADD_UQUAD(ctx, n,
949 	    OID_AUTO, "emvar", CTLFLAG_RD,
950 	    &ios->emvar,
951 	    "Fast Exponentially Weighted Moving Variance");
952 
953 	SYSCTL_ADD_INT(ctx, n,
954 	    OID_AUTO, "pending", CTLFLAG_RD,
955 	    &ios->pending, 0,
956 	    "Instantaneous # of pending transactions");
957 	SYSCTL_ADD_INT(ctx, n,
958 	    OID_AUTO, "count", CTLFLAG_RD,
959 	    &ios->total, 0,
960 	    "# of transactions submitted to hardware");
961 	SYSCTL_ADD_INT(ctx, n,
962 	    OID_AUTO, "queued", CTLFLAG_RD,
963 	    &ios->queued, 0,
964 	    "# of transactions in the queue");
965 	SYSCTL_ADD_INT(ctx, n,
966 	    OID_AUTO, "in", CTLFLAG_RD,
967 	    &ios->in, 0,
968 	    "# of transactions queued to driver");
969 	SYSCTL_ADD_INT(ctx, n,
970 	    OID_AUTO, "out", CTLFLAG_RD,
971 	    &ios->out, 0,
972 	    "# of transactions completed");
973 
974 	SYSCTL_ADD_PROC(ctx, n,
975 	    OID_AUTO, "limiter", CTLTYPE_STRING | CTLFLAG_RW,
976 	    ios, 0, cam_iosched_limiter_sysctl, "A",
977 	    "Current limiting type.");
978 	SYSCTL_ADD_INT(ctx, n,
979 	    OID_AUTO, "min", CTLFLAG_RW,
980 	    &ios->min, 0,
981 	    "min resource");
982 	SYSCTL_ADD_INT(ctx, n,
983 	    OID_AUTO, "max", CTLFLAG_RW,
984 	    &ios->max, 0,
985 	    "max resource");
986 	SYSCTL_ADD_INT(ctx, n,
987 	    OID_AUTO, "current", CTLFLAG_RW,
988 	    &ios->current, 0,
989 	    "current resource");
990 
991 	SYSCTL_ADD_PROC(ctx, n,
992 	    OID_AUTO, "latencies", CTLTYPE_STRING | CTLFLAG_RD,
993 	    &ios->latencies, 0,
994 	    cam_iosched_sysctl_latencies, "A",
995 	    "Array of power of 2 latency from 1ms to 1.024s");
996 }
997 
998 static void
999 cam_iosched_iop_stats_fini(struct iop_stats *ios)
1000 {
1001 	if (ios->sysctl_tree)
1002 		if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
1003 			printf("can't remove iosched sysctl stats context\n");
1004 }
1005 
1006 static void
1007 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
1008 {
1009 	struct sysctl_oid_list *n;
1010 	struct sysctl_ctx_list *ctx;
1011 	struct control_loop *clp;
1012 
1013 	clp = &isc->cl;
1014 	clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1015 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
1016 	    CTLFLAG_RD, 0, "Control loop info");
1017 	n = SYSCTL_CHILDREN(clp->sysctl_tree);
1018 	ctx = &clp->sysctl_ctx;
1019 
1020 	SYSCTL_ADD_PROC(ctx, n,
1021 	    OID_AUTO, "type", CTLTYPE_STRING | CTLFLAG_RW,
1022 	    clp, 0, cam_iosched_control_type_sysctl, "A",
1023 	    "Control loop algorithm");
1024 	SYSCTL_ADD_PROC(ctx, n,
1025 	    OID_AUTO, "steer_interval", CTLTYPE_STRING | CTLFLAG_RW,
1026 	    &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
1027 	    "How often to steer (in us)");
1028 	SYSCTL_ADD_PROC(ctx, n,
1029 	    OID_AUTO, "lolat", CTLTYPE_STRING | CTLFLAG_RW,
1030 	    &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
1031 	    "Low water mark for Latency (in us)");
1032 	SYSCTL_ADD_PROC(ctx, n,
1033 	    OID_AUTO, "hilat", CTLTYPE_STRING | CTLFLAG_RW,
1034 	    &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
1035 	    "Hi water mark for Latency (in us)");
1036 	SYSCTL_ADD_INT(ctx, n,
1037 	    OID_AUTO, "alpha", CTLFLAG_RW,
1038 	    &clp->alpha, 0,
1039 	    "Alpha for PLL (x100) aka gain");
1040 }
1041 
1042 static void
1043 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1044 {
1045 	if (clp->sysctl_tree)
1046 		if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1047 			printf("can't remove iosched sysctl control loop context\n");
1048 }
1049 #endif
1050 
1051 /*
1052  * Allocate the iosched structure. This also insulates callers from knowing
1053  * sizeof struct cam_iosched_softc.
1054  */
1055 int
1056 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1057 {
1058 
1059 	*iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1060 	if (*iscp == NULL)
1061 		return ENOMEM;
1062 #ifdef CAM_IOSCHED_DYNAMIC
1063 	if (iosched_debug)
1064 		printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1065 #endif
1066 	(*iscp)->sort_io_queue = -1;
1067 	bioq_init(&(*iscp)->bio_queue);
1068 	bioq_init(&(*iscp)->trim_queue);
1069 #ifdef CAM_IOSCHED_DYNAMIC
1070 	if (do_dynamic_iosched) {
1071 		bioq_init(&(*iscp)->write_queue);
1072 		(*iscp)->read_bias = 100;
1073 		(*iscp)->current_read_bias = 100;
1074 		(*iscp)->quanta = min(hz, 200);
1075 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1076 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1077 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1078 		(*iscp)->trim_stats.max = 1;	/* Trims are special: one at a time for now */
1079 		(*iscp)->last_time = sbinuptime();
1080 		callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1081 		(*iscp)->periph = periph;
1082 		cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1083 		callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp);
1084 		(*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1085 	}
1086 #endif
1087 
1088 	return 0;
1089 }
1090 
1091 /*
1092  * Reclaim all used resources. This assumes that other folks have
1093  * drained the requests in the hardware. Maybe an unwise assumption.
1094  */
1095 void
1096 cam_iosched_fini(struct cam_iosched_softc *isc)
1097 {
1098 	if (isc) {
1099 		cam_iosched_flush(isc, NULL, ENXIO);
1100 #ifdef CAM_IOSCHED_DYNAMIC
1101 		cam_iosched_iop_stats_fini(&isc->read_stats);
1102 		cam_iosched_iop_stats_fini(&isc->write_stats);
1103 		cam_iosched_iop_stats_fini(&isc->trim_stats);
1104 		cam_iosched_cl_sysctl_fini(&isc->cl);
1105 		if (isc->sysctl_tree)
1106 			if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1107 				printf("can't remove iosched sysctl stats context\n");
1108 		if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1109 			callout_drain(&isc->ticker);
1110 			isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1111 		}
1112 #endif
1113 		free(isc, M_CAMSCHED);
1114 	}
1115 }
1116 
1117 /*
1118  * After we're sure we're attaching a device, go ahead and add
1119  * hooks for any sysctl we may wish to honor.
1120  */
1121 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1122     struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1123 {
1124 #ifdef CAM_IOSCHED_DYNAMIC
1125 	struct sysctl_oid_list *n;
1126 #endif
1127 
1128 	SYSCTL_ADD_INT(ctx, SYSCTL_CHILDREN(node),
1129 		OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1130 		&isc->sort_io_queue, 0,
1131 		"Sort IO queue to try and optimise disk access patterns");
1132 
1133 #ifdef CAM_IOSCHED_DYNAMIC
1134 	if (!do_dynamic_iosched)
1135 		return;
1136 
1137 	isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1138 	    SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1139 	    CTLFLAG_RD, 0, "I/O scheduler statistics");
1140 	n = SYSCTL_CHILDREN(isc->sysctl_tree);
1141 	ctx = &isc->sysctl_ctx;
1142 
1143 	cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1144 	cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1145 	cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1146 	cam_iosched_cl_sysctl_init(isc);
1147 
1148 	SYSCTL_ADD_INT(ctx, n,
1149 	    OID_AUTO, "read_bias", CTLFLAG_RW,
1150 	    &isc->read_bias, 100,
1151 	    "How biased towards read should we be independent of limits");
1152 
1153 	SYSCTL_ADD_PROC(ctx, n,
1154 	    OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW,
1155 	    &isc->quanta, 0, cam_iosched_quanta_sysctl, "I",
1156 	    "How many quanta per second do we slice the I/O up into");
1157 
1158 	SYSCTL_ADD_INT(ctx, n,
1159 	    OID_AUTO, "total_ticks", CTLFLAG_RD,
1160 	    &isc->total_ticks, 0,
1161 	    "Total number of ticks we've done");
1162 
1163 	SYSCTL_ADD_INT(ctx, n,
1164 	    OID_AUTO, "load", CTLFLAG_RD,
1165 	    &isc->load, 0,
1166 	    "scaled load average / 100");
1167 #endif
1168 }
1169 
1170 /*
1171  * Flush outstanding I/O. Consumers of this library don't know all the
1172  * queues we may keep, so this allows all I/O to be flushed in one
1173  * convenient call.
1174  */
1175 void
1176 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1177 {
1178 	bioq_flush(&isc->bio_queue, stp, err);
1179 	bioq_flush(&isc->trim_queue, stp, err);
1180 #ifdef CAM_IOSCHED_DYNAMIC
1181 	if (do_dynamic_iosched)
1182 		bioq_flush(&isc->write_queue, stp, err);
1183 #endif
1184 }
1185 
1186 #ifdef CAM_IOSCHED_DYNAMIC
1187 static struct bio *
1188 cam_iosched_get_write(struct cam_iosched_softc *isc)
1189 {
1190 	struct bio *bp;
1191 
1192 	/*
1193 	 * We control the write rate by controlling how many requests we send
1194 	 * down to the drive at any one time. Fewer requests limits the
1195 	 * effects of both starvation when the requests take a while and write
1196 	 * amplification when each request is causing more than one write to
1197 	 * the NAND media. Limiting the queue depth like this will also limit
1198 	 * the write throughput and give and reads that want to compete to
1199 	 * compete unfairly.
1200 	 */
1201 	bp = bioq_first(&isc->write_queue);
1202 	if (bp == NULL) {
1203 		if (iosched_debug > 3)
1204 			printf("No writes present in write_queue\n");
1205 		return NULL;
1206 	}
1207 
1208 	/*
1209 	 * If pending read, prefer that based on current read bias
1210 	 * setting.
1211 	 */
1212 	if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1213 		if (iosched_debug)
1214 			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);
1215 		isc->current_read_bias--;
1216 		/* We're not limiting writes, per se, just doing reads first */
1217 		return NULL;
1218 	}
1219 
1220 	/*
1221 	 * See if our current limiter allows this I/O.
1222 	 */
1223 	if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1224 		if (iosched_debug)
1225 			printf("Can't write because limiter says no.\n");
1226 		isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1227 		return NULL;
1228 	}
1229 
1230 	/*
1231 	 * Let's do this: We've passed all the gates and we're a go
1232 	 * to schedule the I/O in the SIM.
1233 	 */
1234 	isc->current_read_bias = isc->read_bias;
1235 	bioq_remove(&isc->write_queue, bp);
1236 	if (bp->bio_cmd == BIO_WRITE) {
1237 		isc->write_stats.queued--;
1238 		isc->write_stats.total++;
1239 		isc->write_stats.pending++;
1240 	}
1241 	if (iosched_debug > 9)
1242 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1243 	isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1244 	return bp;
1245 }
1246 #endif
1247 
1248 /*
1249  * Put back a trim that you weren't able to actually schedule this time.
1250  */
1251 void
1252 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1253 {
1254 	bioq_insert_head(&isc->trim_queue, bp);
1255 #ifdef CAM_IOSCHED_DYNAMIC
1256 	isc->trim_stats.queued++;
1257 	isc->trim_stats.total--;		/* since we put it back, don't double count */
1258 	isc->trim_stats.pending--;
1259 #endif
1260 }
1261 
1262 /*
1263  * gets the next trim from the trim queue.
1264  *
1265  * Assumes we're called with the periph lock held.  It removes this
1266  * trim from the queue and the device must explicitly reinsert it
1267  * should the need arise.
1268  */
1269 struct bio *
1270 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1271 {
1272 	struct bio *bp;
1273 
1274 	bp  = bioq_first(&isc->trim_queue);
1275 	if (bp == NULL)
1276 		return NULL;
1277 	bioq_remove(&isc->trim_queue, bp);
1278 #ifdef CAM_IOSCHED_DYNAMIC
1279 	isc->trim_stats.queued--;
1280 	isc->trim_stats.total++;
1281 	isc->trim_stats.pending++;
1282 #endif
1283 	return bp;
1284 }
1285 
1286 /*
1287  * gets an available trim from the trim queue, if there's no trim
1288  * already pending. It removes this trim from the queue and the device
1289  * must explicitly reinsert it should the need arise.
1290  *
1291  * Assumes we're called with the periph lock held.
1292  */
1293 struct bio *
1294 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1295 {
1296 
1297 	if (!cam_iosched_has_more_trim(isc))
1298 		return NULL;
1299 
1300 	return cam_iosched_next_trim(isc);
1301 }
1302 
1303 /*
1304  * Determine what the next bit of work to do is for the periph. The
1305  * default implementation looks to see if we have trims to do, but no
1306  * trims outstanding. If so, we do that. Otherwise we see if we have
1307  * other work. If we do, then we do that. Otherwise why were we called?
1308  */
1309 struct bio *
1310 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1311 {
1312 	struct bio *bp;
1313 
1314 	/*
1315 	 * See if we have a trim that can be scheduled. We can only send one
1316 	 * at a time down, so this takes that into account.
1317 	 *
1318 	 * XXX newer TRIM commands are queueable. Revisit this when we
1319 	 * implement them.
1320 	 */
1321 	if ((bp = cam_iosched_get_trim(isc)) != NULL)
1322 		return bp;
1323 
1324 #ifdef CAM_IOSCHED_DYNAMIC
1325 	/*
1326 	 * See if we have any pending writes, and room in the queue for them,
1327 	 * and if so, those are next.
1328 	 */
1329 	if (do_dynamic_iosched) {
1330 		if ((bp = cam_iosched_get_write(isc)) != NULL)
1331 			return bp;
1332 	}
1333 #endif
1334 
1335 	/*
1336 	 * next, see if there's other, normal I/O waiting. If so return that.
1337 	 */
1338 	if ((bp = bioq_first(&isc->bio_queue)) == NULL)
1339 		return NULL;
1340 
1341 #ifdef CAM_IOSCHED_DYNAMIC
1342 	/*
1343 	 * For the dynamic scheduler, bio_queue is only for reads, so enforce
1344 	 * the limits here. Enforce only for reads.
1345 	 */
1346 	if (do_dynamic_iosched) {
1347 		if (bp->bio_cmd == BIO_READ &&
1348 		    cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1349 			isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1350 			return NULL;
1351 		}
1352 	}
1353 	isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1354 #endif
1355 	bioq_remove(&isc->bio_queue, bp);
1356 #ifdef CAM_IOSCHED_DYNAMIC
1357 	if (do_dynamic_iosched) {
1358 		if (bp->bio_cmd == BIO_READ) {
1359 			isc->read_stats.queued--;
1360 			isc->read_stats.total++;
1361 			isc->read_stats.pending++;
1362 		} else
1363 			printf("Found bio_cmd = %#x\n", bp->bio_cmd);
1364 	}
1365 	if (iosched_debug > 9)
1366 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1367 #endif
1368 	return bp;
1369 }
1370 
1371 /*
1372  * Driver has been given some work to do by the block layer. Tell the
1373  * scheduler about it and have it queue the work up. The scheduler module
1374  * will then return the currently most useful bit of work later, possibly
1375  * deferring work for various reasons.
1376  */
1377 void
1378 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1379 {
1380 
1381 	/*
1382 	 * Put all trims on the trim queue sorted, since we know
1383 	 * that the collapsing code requires this. Otherwise put
1384 	 * the work on the bio queue.
1385 	 */
1386 	if (bp->bio_cmd == BIO_DELETE) {
1387 		bioq_disksort(&isc->trim_queue, bp);
1388 #ifdef CAM_IOSCHED_DYNAMIC
1389 		isc->trim_stats.in++;
1390 		isc->trim_stats.queued++;
1391 #endif
1392 	}
1393 #ifdef CAM_IOSCHED_DYNAMIC
1394 	else if (do_dynamic_iosched && (bp->bio_cmd != BIO_READ)) {
1395 		if (cam_iosched_sort_queue(isc))
1396 			bioq_disksort(&isc->write_queue, bp);
1397 		else
1398 			bioq_insert_tail(&isc->write_queue, bp);
1399 		if (iosched_debug > 9)
1400 			printf("Qw  : %p %#x\n", bp, bp->bio_cmd);
1401 		if (bp->bio_cmd == BIO_WRITE) {
1402 			isc->write_stats.in++;
1403 			isc->write_stats.queued++;
1404 		}
1405 	}
1406 #endif
1407 	else {
1408 		if (cam_iosched_sort_queue(isc))
1409 			bioq_disksort(&isc->bio_queue, bp);
1410 		else
1411 			bioq_insert_tail(&isc->bio_queue, bp);
1412 #ifdef CAM_IOSCHED_DYNAMIC
1413 		if (iosched_debug > 9)
1414 			printf("Qr  : %p %#x\n", bp, bp->bio_cmd);
1415 		if (bp->bio_cmd == BIO_READ) {
1416 			isc->read_stats.in++;
1417 			isc->read_stats.queued++;
1418 		} else if (bp->bio_cmd == BIO_WRITE) {
1419 			isc->write_stats.in++;
1420 			isc->write_stats.queued++;
1421 		}
1422 #endif
1423 	}
1424 }
1425 
1426 /*
1427  * If we have work, get it scheduled. Called with the periph lock held.
1428  */
1429 void
1430 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1431 {
1432 
1433 	if (cam_iosched_has_work(isc))
1434 		xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1435 }
1436 
1437 /*
1438  * Complete a trim request. Mark that we no longer have one in flight.
1439  */
1440 void
1441 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1442 {
1443 
1444 	isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1445 }
1446 
1447 /*
1448  * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1449  * might use notes in the ccb for statistics.
1450  */
1451 int
1452 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1453     union ccb *done_ccb)
1454 {
1455 	int retval = 0;
1456 #ifdef CAM_IOSCHED_DYNAMIC
1457 	if (!do_dynamic_iosched)
1458 		return retval;
1459 
1460 	if (iosched_debug > 10)
1461 		printf("done: %p %#x\n", bp, bp->bio_cmd);
1462 	if (bp->bio_cmd == BIO_WRITE) {
1463 		retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1464 		isc->write_stats.out++;
1465 		isc->write_stats.pending--;
1466 	} else if (bp->bio_cmd == BIO_READ) {
1467 		retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1468 		isc->read_stats.out++;
1469 		isc->read_stats.pending--;
1470 	} else if (bp->bio_cmd == BIO_DELETE) {
1471 		isc->trim_stats.out++;
1472 		isc->trim_stats.pending--;
1473 	} else if (bp->bio_cmd != BIO_FLUSH) {
1474 		if (iosched_debug)
1475 			printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1476 	}
1477 
1478 	if (!(bp->bio_flags & BIO_ERROR))
1479 		cam_iosched_io_metric_update(isc,
1480 		    cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data),
1481 		    bp->bio_cmd, bp->bio_bcount);
1482 #endif
1483 	return retval;
1484 }
1485 
1486 /*
1487  * Tell the io scheduler that you've pushed a trim down into the sim.
1488  * This also tells the I/O scheduler not to push any more trims down, so
1489  * some periphs do not call it if they can cope with multiple trims in flight.
1490  */
1491 void
1492 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1493 {
1494 
1495 	isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1496 }
1497 
1498 /*
1499  * Change the sorting policy hint for I/O transactions for this device.
1500  */
1501 void
1502 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1503 {
1504 
1505 	isc->sort_io_queue = val;
1506 }
1507 
1508 int
1509 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1510 {
1511 	return isc->flags & flags;
1512 }
1513 
1514 void
1515 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1516 {
1517 	isc->flags |= flags;
1518 }
1519 
1520 void
1521 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1522 {
1523 	isc->flags &= ~flags;
1524 }
1525 
1526 #ifdef CAM_IOSCHED_DYNAMIC
1527 /*
1528  * After the method presented in Jack Crenshaw's 1998 article "Integer
1529  * Square Roots," reprinted at
1530  * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1531  * and well worth the read. Briefly, we find the power of 4 that's the
1532  * largest smaller than val. We then check each smaller power of 4 to
1533  * see if val is still bigger. The right shifts at each step divide
1534  * the result by 2 which after successive application winds up
1535  * accumulating the right answer. It could also have been accumulated
1536  * using a separate root counter, but this code is smaller and faster
1537  * than that method. This method is also integer size invariant.
1538  * It returns floor(sqrt((float)val)), or the largest integer less than
1539  * or equal to the square root.
1540  */
1541 static uint64_t
1542 isqrt64(uint64_t val)
1543 {
1544 	uint64_t res = 0;
1545 	uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1546 
1547 	/*
1548 	 * Find the largest power of 4 smaller than val.
1549 	 */
1550 	while (bit > val)
1551 		bit >>= 2;
1552 
1553 	/*
1554 	 * Accumulate the answer, one bit at a time (we keep moving
1555 	 * them over since 2 is the square root of 4 and we test
1556 	 * powers of 4). We accumulate where we find the bit, but
1557 	 * the successive shifts land the bit in the right place
1558 	 * by the end.
1559 	 */
1560 	while (bit != 0) {
1561 		if (val >= res + bit) {
1562 			val -= res + bit;
1563 			res = (res >> 1) + bit;
1564 		} else
1565 			res >>= 1;
1566 		bit >>= 2;
1567 	}
1568 
1569 	return res;
1570 }
1571 
1572 static sbintime_t latencies[LAT_BUCKETS - 1] = {
1573 	SBT_1MS <<  0,
1574 	SBT_1MS <<  1,
1575 	SBT_1MS <<  2,
1576 	SBT_1MS <<  3,
1577 	SBT_1MS <<  4,
1578 	SBT_1MS <<  5,
1579 	SBT_1MS <<  6,
1580 	SBT_1MS <<  7,
1581 	SBT_1MS <<  8,
1582 	SBT_1MS <<  9,
1583 	SBT_1MS << 10,
1584 	SBT_1MS << 11,
1585 	SBT_1MS << 12,
1586 	SBT_1MS << 13		/* 8.192s */
1587 };
1588 
1589 static void
1590 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1591 {
1592 	sbintime_t y, deltasq, delta;
1593 	int i;
1594 
1595 	/*
1596 	 * Keep counts for latency. We do it by power of two buckets.
1597 	 * This helps us spot outlier behavior obscured by averages.
1598 	 */
1599 	for (i = 0; i < LAT_BUCKETS - 1; i++) {
1600 		if (sim_latency < latencies[i]) {
1601 			iop->latencies[i]++;
1602 			break;
1603 		}
1604 	}
1605 	if (i == LAT_BUCKETS - 1)
1606 		iop->latencies[i]++; 	 /* Put all > 1024ms values into the last bucket. */
1607 
1608 	/*
1609 	 * Classic exponentially decaying average with a tiny alpha
1610 	 * (2 ^ -alpha_bits). For more info see the NIST statistical
1611 	 * handbook.
1612 	 *
1613 	 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha)		[nist]
1614 	 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
1615 	 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1616 	 * alpha = 1 / (1 << alpha_bits)
1617 	 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
1618 	 *	= y_t/b - e/b + be/b
1619 	 *      = (y_t - e + be) / b
1620 	 *	= (e + d) / b
1621 	 *
1622 	 * Since alpha is a power of two, we can compute this w/o any mult or
1623 	 * division.
1624 	 *
1625 	 * Variance can also be computed. Usually, it would be expressed as follows:
1626 	 *	diff_t = y_t - ema_t-1
1627 	 *	emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
1628 	 *	  = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
1629 	 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
1630 	 *	  = e - e/b + dd/b + dd/bb
1631 	 *	  = (bbe - be + bdd + dd) / bb
1632 	 *	  = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
1633 	 */
1634 	/*
1635 	 * XXX possible numeric issues
1636 	 *	o We assume right shifted integers do the right thing, since that's
1637 	 *	  implementation defined. You can change the right shifts to / (1LL << alpha).
1638 	 *	o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
1639 	 *	  for emvar. This puts a ceiling of 13 bits on alpha since we need a
1640 	 *	  few tens of seconds of representation.
1641 	 *	o We mitigate alpha issues by never setting it too high.
1642 	 */
1643 	y = sim_latency;
1644 	delta = (y - iop->ema);					/* d */
1645 	iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1646 
1647 	/*
1648 	 * Were we to naively plow ahead at this point, we wind up with many numerical
1649 	 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
1650 	 * us with microsecond level precision in the input, so the same in the
1651 	 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
1652 	 * also means that emvar can be up 46 bits 40 of which are fraction, which
1653 	 * gives us a way to measure up to ~8s in the SD before the computation goes
1654 	 * unstable. Even the worst hard disk rarely has > 1s service time in the
1655 	 * drive. It does mean we have to shift left 12 bits after taking the
1656 	 * square root to compute the actual standard deviation estimate. This loss of
1657 	 * precision is preferable to needing int128 types to work. The above numbers
1658 	 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
1659 	 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1660 	 */
1661 	delta >>= 12;
1662 	deltasq = delta * delta;				/* dd */
1663 	iop->emvar = ((iop->emvar << (2 * alpha_bits)) +	/* bbe */
1664 	    ((deltasq - iop->emvar) << alpha_bits) +		/* b(dd-e) */
1665 	    deltasq)						/* dd */
1666 	    >> (2 * alpha_bits);				/* div bb */
1667 	iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1668 }
1669 
1670 static void
1671 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1672     sbintime_t sim_latency, int cmd, size_t size)
1673 {
1674 	/* xxx Do we need to scale based on the size of the I/O ? */
1675 	switch (cmd) {
1676 	case BIO_READ:
1677 		cam_iosched_update(&isc->read_stats, sim_latency);
1678 		break;
1679 	case BIO_WRITE:
1680 		cam_iosched_update(&isc->write_stats, sim_latency);
1681 		break;
1682 	case BIO_DELETE:
1683 		cam_iosched_update(&isc->trim_stats, sim_latency);
1684 		break;
1685 	default:
1686 		break;
1687 	}
1688 }
1689 
1690 #ifdef DDB
1691 static int biolen(struct bio_queue_head *bq)
1692 {
1693 	int i = 0;
1694 	struct bio *bp;
1695 
1696 	TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
1697 		i++;
1698 	}
1699 	return i;
1700 }
1701 
1702 /*
1703  * Show the internal state of the I/O scheduler.
1704  */
1705 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
1706 {
1707 	struct cam_iosched_softc *isc;
1708 
1709 	if (!have_addr) {
1710 		db_printf("Need addr\n");
1711 		return;
1712 	}
1713 	isc = (struct cam_iosched_softc *)addr;
1714 	db_printf("pending_reads:     %d\n", isc->read_stats.pending);
1715 	db_printf("min_reads:         %d\n", isc->read_stats.min);
1716 	db_printf("max_reads:         %d\n", isc->read_stats.max);
1717 	db_printf("reads:             %d\n", isc->read_stats.total);
1718 	db_printf("in_reads:          %d\n", isc->read_stats.in);
1719 	db_printf("out_reads:         %d\n", isc->read_stats.out);
1720 	db_printf("queued_reads:      %d\n", isc->read_stats.queued);
1721 	db_printf("Current Q len      %d\n", biolen(&isc->bio_queue));
1722 	db_printf("pending_writes:    %d\n", isc->write_stats.pending);
1723 	db_printf("min_writes:        %d\n", isc->write_stats.min);
1724 	db_printf("max_writes:        %d\n", isc->write_stats.max);
1725 	db_printf("writes:            %d\n", isc->write_stats.total);
1726 	db_printf("in_writes:         %d\n", isc->write_stats.in);
1727 	db_printf("out_writes:        %d\n", isc->write_stats.out);
1728 	db_printf("queued_writes:     %d\n", isc->write_stats.queued);
1729 	db_printf("Current Q len      %d\n", biolen(&isc->write_queue));
1730 	db_printf("pending_trims:     %d\n", isc->trim_stats.pending);
1731 	db_printf("min_trims:         %d\n", isc->trim_stats.min);
1732 	db_printf("max_trims:         %d\n", isc->trim_stats.max);
1733 	db_printf("trims:             %d\n", isc->trim_stats.total);
1734 	db_printf("in_trims:          %d\n", isc->trim_stats.in);
1735 	db_printf("out_trims:         %d\n", isc->trim_stats.out);
1736 	db_printf("queued_trims:      %d\n", isc->trim_stats.queued);
1737 	db_printf("Current Q len      %d\n", biolen(&isc->trim_queue));
1738 	db_printf("read_bias:         %d\n", isc->read_bias);
1739 	db_printf("current_read_bias: %d\n", isc->current_read_bias);
1740 	db_printf("Trim active?       %s\n",
1741 	    (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");
1742 }
1743 #endif
1744 #endif
1745