xref: /freebsd/sys/cam/cam_iosched.c (revision a0ca4af9455b844c5e094fc1b09b1390ffa979fc)
1 /*-
2  * CAM IO Scheduler Interface
3  *
4  * SPDX-License-Identifier: BSD-2-Clause
5  *
6  * Copyright (c) 2015 Netflix, Inc.
7  *
8  * Redistribution and use in source and binary forms, with or without
9  * modification, are permitted provided that the following conditions
10  * are met:
11  * 1. Redistributions of source code must retain the above copyright
12  *    notice, this list of conditions and the following disclaimer.
13  * 2. Redistributions in binary form must reproduce the above copyright
14  *    notice, this list of conditions and the following disclaimer in the
15  *    documentation and/or other materials provided with the distribution.
16  *
17  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
21  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
27  * SUCH DAMAGE.
28  */
29 
30 #include "opt_ddb.h"
31 
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/kernel.h>
35 #include <sys/bio.h>
36 #include <sys/lock.h>
37 #include <sys/malloc.h>
38 #include <sys/mutex.h>
39 #include <sys/sbuf.h>
40 #include <sys/sysctl.h>
41 
42 #include <cam/cam.h>
43 #include <cam/cam_ccb.h>
44 #include <cam/cam_periph.h>
45 #include <cam/cam_xpt_periph.h>
46 #include <cam/cam_xpt_internal.h>
47 #include <cam/cam_iosched.h>
48 
49 #include <ddb/ddb.h>
50 
51 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler",
52     "CAM I/O Scheduler buffers");
53 
54 static SYSCTL_NODE(_kern_cam, OID_AUTO, iosched, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
55     "CAM I/O Scheduler parameters");
56 
57 /*
58  * Default I/O scheduler for FreeBSD. This implementation is just a thin-vineer
59  * over the bioq_* interface, with notions of separate calls for normal I/O and
60  * for trims.
61  *
62  * When CAM_IOSCHED_DYNAMIC is defined, the scheduler is enhanced to dynamically
63  * steer the rate of one type of traffic to help other types of traffic (eg
64  * limit writes when read latency deteriorates on SSDs).
65  */
66 
67 #ifdef CAM_IOSCHED_DYNAMIC
68 
69 static bool do_dynamic_iosched = true;
70 SYSCTL_BOOL(_kern_cam_iosched, OID_AUTO, dynamic, CTLFLAG_RDTUN,
71     &do_dynamic_iosched, 1,
72     "Enable Dynamic I/O scheduler optimizations.");
73 
74 /*
75  * For an EMA, with an alpha of alpha, we know
76  * 	alpha = 2 / (N + 1)
77  * or
78  * 	N = 1 + (2 / alpha)
79  * where N is the number of samples that 86% of the current
80  * EMA is derived from.
81  *
82  * So we invent[*] alpha_bits:
83  *	alpha_bits = -log_2(alpha)
84  *	alpha = 2^-alpha_bits
85  * So
86  *	N = 1 + 2^(alpha_bits + 1)
87  *
88  * The default 9 gives a 1025 lookback for 86% of the data.
89  * For a brief intro: https://en.wikipedia.org/wiki/Moving_average
90  *
91  * [*] Steal from the load average code and many other places.
92  * Note: See computation of EMA and EMVAR for acceptable ranges of alpha.
93  */
94 static int alpha_bits = 9;
95 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, alpha_bits, CTLFLAG_RWTUN,
96     &alpha_bits, 1,
97     "Bits in EMA's alpha.");
98 
99 /*
100  * Different parameters for the buckets of latency we keep track of. These are all
101  * published read-only since at present they are compile time constants.
102  *
103  * Bucket base is the upper bounds of the first latency bucket. It's currently 20us.
104  * With 20 buckets (see below), that leads to a geometric progression with a max size
105  * of 5.2s which is safeily larger than 1s to help diagnose extreme outliers better.
106  */
107 #ifndef BUCKET_BASE
108 #define BUCKET_BASE ((SBT_1S / 50000) + 1)	/* 20us */
109 #endif
110 static sbintime_t bucket_base = BUCKET_BASE;
111 SYSCTL_SBINTIME_USEC(_kern_cam_iosched, OID_AUTO, bucket_base_us, CTLFLAG_RD,
112     &bucket_base,
113     "Size of the smallest latency bucket");
114 
115 /*
116  * Bucket ratio is the geometric progression for the bucket. For a bucket b_n
117  * the size of bucket b_n+1 is b_n * bucket_ratio / 100.
118  */
119 static int bucket_ratio = 200;	/* Rather hard coded at the moment */
120 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, bucket_ratio, CTLFLAG_RD,
121     &bucket_ratio, 200,
122     "Latency Bucket Ratio for geometric progression.");
123 
124 /*
125  * Number of total buckets. Starting at BUCKET_BASE, each one is a power of 2.
126  */
127 #ifndef LAT_BUCKETS
128 #define LAT_BUCKETS 20	/* < 20us < 40us ... < 2^(n-1)*20us >= 2^(n-1)*20us */
129 #endif
130 static int lat_buckets = LAT_BUCKETS;
131 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, buckets, CTLFLAG_RD,
132     &lat_buckets, LAT_BUCKETS,
133     "Total number of latency buckets published");
134 
135 /*
136  * Read bias: how many reads do we favor before scheduling a write
137  * when we have a choice.
138  */
139 static int default_read_bias = 0;
140 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, read_bias, CTLFLAG_RWTUN,
141     &default_read_bias, 0,
142     "Default read bias for new devices.");
143 
144 struct iop_stats;
145 struct cam_iosched_softc;
146 
147 int iosched_debug = 0;
148 
149 typedef enum {
150 	none = 0,				/* No limits */
151 	queue_depth,			/* Limit how many ops we queue to SIM */
152 	iops,				/* Limit # of IOPS to the drive */
153 	bandwidth,			/* Limit bandwidth to the drive */
154 	limiter_max
155 } io_limiter;
156 
157 static const char *cam_iosched_limiter_names[] =
158     { "none", "queue_depth", "iops", "bandwidth" };
159 
160 /*
161  * Called to initialize the bits of the iop_stats structure relevant to the
162  * limiter. Called just after the limiter is set.
163  */
164 typedef int l_init_t(struct iop_stats *);
165 
166 /*
167  * Called every tick.
168  */
169 typedef int l_tick_t(struct iop_stats *);
170 
171 /*
172  * Called to see if the limiter thinks this IOP can be allowed to
173  * proceed. If so, the limiter assumes that the IOP proceeded
174  * and makes any accounting of it that's needed.
175  */
176 typedef int l_iop_t(struct iop_stats *, struct bio *);
177 
178 /*
179  * Called when an I/O completes so the limiter can update its
180  * accounting. Pending I/Os may complete in any order (even when
181  * sent to the hardware at the same time), so the limiter may not
182  * make any assumptions other than this I/O has completed. If it
183  * returns 1, then xpt_schedule() needs to be called again.
184  */
185 typedef int l_iodone_t(struct iop_stats *, struct bio *);
186 
187 static l_iop_t cam_iosched_qd_iop;
188 static l_iop_t cam_iosched_qd_caniop;
189 static l_iodone_t cam_iosched_qd_iodone;
190 
191 static l_init_t cam_iosched_iops_init;
192 static l_tick_t cam_iosched_iops_tick;
193 static l_iop_t cam_iosched_iops_caniop;
194 static l_iop_t cam_iosched_iops_iop;
195 
196 static l_init_t cam_iosched_bw_init;
197 static l_tick_t cam_iosched_bw_tick;
198 static l_iop_t cam_iosched_bw_caniop;
199 static l_iop_t cam_iosched_bw_iop;
200 
201 struct limswitch {
202 	l_init_t	*l_init;
203 	l_tick_t	*l_tick;
204 	l_iop_t		*l_iop;
205 	l_iop_t		*l_caniop;
206 	l_iodone_t	*l_iodone;
207 } limsw[] =
208 {
209 	{	/* none */
210 		.l_init = NULL,
211 		.l_tick = NULL,
212 		.l_iop = NULL,
213 		.l_iodone= NULL,
214 	},
215 	{	/* queue_depth */
216 		.l_init = NULL,
217 		.l_tick = NULL,
218 		.l_caniop = cam_iosched_qd_caniop,
219 		.l_iop = cam_iosched_qd_iop,
220 		.l_iodone= cam_iosched_qd_iodone,
221 	},
222 	{	/* iops */
223 		.l_init = cam_iosched_iops_init,
224 		.l_tick = cam_iosched_iops_tick,
225 		.l_caniop = cam_iosched_iops_caniop,
226 		.l_iop = cam_iosched_iops_iop,
227 		.l_iodone= NULL,
228 	},
229 	{	/* bandwidth */
230 		.l_init = cam_iosched_bw_init,
231 		.l_tick = cam_iosched_bw_tick,
232 		.l_caniop = cam_iosched_bw_caniop,
233 		.l_iop = cam_iosched_bw_iop,
234 		.l_iodone= NULL,
235 	},
236 };
237 
238 struct iop_stats {
239 	/*
240 	 * sysctl state for this subnode.
241 	 */
242 	struct sysctl_ctx_list	sysctl_ctx;
243 	struct sysctl_oid	*sysctl_tree;
244 
245 	/*
246 	 * Information about the current rate limiters, if any
247 	 */
248 	io_limiter	limiter;	/* How are I/Os being limited */
249 	int		min;		/* Low range of limit */
250 	int		max;		/* High range of limit */
251 	int		current;	/* Current rate limiter */
252 	int		l_value1;	/* per-limiter scratch value 1. */
253 	int		l_value2;	/* per-limiter scratch value 2. */
254 
255 	/*
256 	 * Debug information about counts of I/Os that have gone through the
257 	 * scheduler.
258 	 */
259 	int		pending;	/* I/Os pending in the hardware */
260 	int		queued;		/* number currently in the queue */
261 	int		total;		/* Total for all time -- wraps */
262 	int		in;		/* number queued all time -- wraps */
263 	int		out;		/* number completed all time -- wraps */
264 	int		errs;		/* Number of I/Os completed with error --  wraps */
265 
266 	/*
267 	 * Statistics on different bits of the process.
268 	 */
269 		/* Exp Moving Average, see alpha_bits for more details */
270 	sbintime_t      ema;
271 	sbintime_t      emvar;
272 	sbintime_t      sd;		/* Last computed sd */
273 
274 	uint32_t	state_flags;
275 #define IOP_RATE_LIMITED		1u
276 
277 	uint64_t	latencies[LAT_BUCKETS];
278 
279 	struct cam_iosched_softc *softc;
280 };
281 
282 typedef enum {
283 	set_max = 0,			/* current = max */
284 	read_latency,			/* Steer read latency by throttling writes */
285 	cl_max				/* Keep last */
286 } control_type;
287 
288 static const char *cam_iosched_control_type_names[] =
289     { "set_max", "read_latency" };
290 
291 struct control_loop {
292 	/*
293 	 * sysctl state for this subnode.
294 	 */
295 	struct sysctl_ctx_list	sysctl_ctx;
296 	struct sysctl_oid	*sysctl_tree;
297 
298 	sbintime_t	next_steer;		/* Time of next steer */
299 	sbintime_t	steer_interval;		/* How often do we steer? */
300 	sbintime_t	lolat;
301 	sbintime_t	hilat;
302 	int		alpha;
303 	control_type	type;			/* What type of control? */
304 	int		last_count;		/* Last I/O count */
305 
306 	struct cam_iosched_softc *softc;
307 };
308 
309 #endif
310 
311 struct cam_iosched_softc {
312 	struct bio_queue_head bio_queue;
313 	struct bio_queue_head trim_queue;
314 				/* scheduler flags < 16, user flags >= 16 */
315 	uint32_t	flags;
316 	int		sort_io_queue;
317 	int		trim_goal;		/* # of trims to queue before sending */
318 	int		trim_ticks;		/* Max ticks to hold trims */
319 	int		last_trim_tick;		/* Last 'tick' time ld a trim */
320 	int		queued_trims;		/* Number of trims in the queue */
321 #ifdef CAM_IOSCHED_DYNAMIC
322 	int		read_bias;		/* Read bias setting */
323 	int		current_read_bias;	/* Current read bias state */
324 	int		total_ticks;
325 	int		load;			/* EMA of 'load average' of disk / 2^16 */
326 
327 	struct bio_queue_head write_queue;
328 	struct iop_stats read_stats, write_stats, trim_stats;
329 	struct sysctl_ctx_list	sysctl_ctx;
330 	struct sysctl_oid	*sysctl_tree;
331 
332 	int		quanta;			/* Number of quanta per second */
333 	struct callout	ticker;			/* Callout for our quota system */
334 	struct cam_periph *periph;		/* cam periph associated with this device */
335 	uint32_t	this_frac;		/* Fraction of a second (1024ths) for this tick */
336 	sbintime_t	last_time;		/* Last time we ticked */
337 	struct control_loop cl;
338 	sbintime_t	max_lat;		/* when != 0, if iop latency > max_lat, call max_lat_fcn */
339 	cam_iosched_latfcn_t	latfcn;
340 	void		*latarg;
341 #endif
342 };
343 
344 #ifdef CAM_IOSCHED_DYNAMIC
345 /*
346  * helper functions to call the limsw functions.
347  */
348 static int
349 cam_iosched_limiter_init(struct iop_stats *ios)
350 {
351 	int lim = ios->limiter;
352 
353 	/* maybe this should be a kassert */
354 	if (lim < none || lim >= limiter_max)
355 		return EINVAL;
356 
357 	if (limsw[lim].l_init)
358 		return limsw[lim].l_init(ios);
359 
360 	return 0;
361 }
362 
363 static int
364 cam_iosched_limiter_tick(struct iop_stats *ios)
365 {
366 	int lim = ios->limiter;
367 
368 	/* maybe this should be a kassert */
369 	if (lim < none || lim >= limiter_max)
370 		return EINVAL;
371 
372 	if (limsw[lim].l_tick)
373 		return limsw[lim].l_tick(ios);
374 
375 	return 0;
376 }
377 
378 static int
379 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
380 {
381 	int lim = ios->limiter;
382 
383 	/* maybe this should be a kassert */
384 	if (lim < none || lim >= limiter_max)
385 		return EINVAL;
386 
387 	if (limsw[lim].l_iop)
388 		return limsw[lim].l_iop(ios, bp);
389 
390 	return 0;
391 }
392 
393 static int
394 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
395 {
396 	int lim = ios->limiter;
397 
398 	/* maybe this should be a kassert */
399 	if (lim < none || lim >= limiter_max)
400 		return EINVAL;
401 
402 	if (limsw[lim].l_caniop)
403 		return limsw[lim].l_caniop(ios, bp);
404 
405 	return 0;
406 }
407 
408 static int
409 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
410 {
411 	int lim = ios->limiter;
412 
413 	/* maybe this should be a kassert */
414 	if (lim < none || lim >= limiter_max)
415 		return 0;
416 
417 	if (limsw[lim].l_iodone)
418 		return limsw[lim].l_iodone(ios, bp);
419 
420 	return 0;
421 }
422 
423 /*
424  * Functions to implement the different kinds of limiters
425  */
426 
427 static int
428 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
429 {
430 
431 	if (ios->current <= 0 || ios->pending < ios->current)
432 		return 0;
433 
434 	return EAGAIN;
435 }
436 
437 static int
438 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
439 {
440 
441 	if (ios->current <= 0 || ios->pending < ios->current)
442 		return 0;
443 
444 	return EAGAIN;
445 }
446 
447 static int
448 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
449 {
450 
451 	if (ios->current <= 0 || ios->pending != ios->current)
452 		return 0;
453 
454 	return 1;
455 }
456 
457 static int
458 cam_iosched_iops_init(struct iop_stats *ios)
459 {
460 
461 	ios->l_value1 = ios->current / ios->softc->quanta;
462 	if (ios->l_value1 <= 0)
463 		ios->l_value1 = 1;
464 	ios->l_value2 = 0;
465 
466 	return 0;
467 }
468 
469 static int
470 cam_iosched_iops_tick(struct iop_stats *ios)
471 {
472 	int new_ios;
473 
474 	/*
475 	 * Allow at least one IO per tick until all
476 	 * the IOs for this interval have been spent.
477 	 */
478 	new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
479 	if (new_ios < 1 && ios->l_value2 < ios->current) {
480 		new_ios = 1;
481 		ios->l_value2++;
482 	}
483 
484 	/*
485 	 * If this a new accounting interval, discard any "unspent" ios
486 	 * granted in the previous interval.  Otherwise add the new ios to
487 	 * the previously granted ones that haven't been spent yet.
488 	 */
489 	if ((ios->softc->total_ticks % ios->softc->quanta) == 0) {
490 		ios->l_value1 = new_ios;
491 		ios->l_value2 = 1;
492 	} else {
493 		ios->l_value1 += new_ios;
494 	}
495 
496 	return 0;
497 }
498 
499 static int
500 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
501 {
502 
503 	/*
504 	 * So if we have any more IOPs left, allow it,
505 	 * otherwise wait. If current iops is 0, treat that
506 	 * as unlimited as a failsafe.
507 	 */
508 	if (ios->current > 0 && ios->l_value1 <= 0)
509 		return EAGAIN;
510 	return 0;
511 }
512 
513 static int
514 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
515 {
516 	int rv;
517 
518 	rv = cam_iosched_limiter_caniop(ios, bp);
519 	if (rv == 0)
520 		ios->l_value1--;
521 
522 	return rv;
523 }
524 
525 static int
526 cam_iosched_bw_init(struct iop_stats *ios)
527 {
528 
529 	/* ios->current is in kB/s, so scale to bytes */
530 	ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
531 
532 	return 0;
533 }
534 
535 static int
536 cam_iosched_bw_tick(struct iop_stats *ios)
537 {
538 	int bw;
539 
540 	/*
541 	 * If we're in the hole for available quota from
542 	 * the last time, then add the quantum for this.
543 	 * If we have any left over from last quantum,
544 	 * then too bad, that's lost. Also, ios->current
545 	 * is in kB/s, so scale.
546 	 *
547 	 * We also allow up to 4 quanta of credits to
548 	 * accumulate to deal with burstiness. 4 is extremely
549 	 * arbitrary.
550 	 */
551 	bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
552 	if (ios->l_value1 < bw * 4)
553 		ios->l_value1 += bw;
554 
555 	return 0;
556 }
557 
558 static int
559 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
560 {
561 	/*
562 	 * So if we have any more bw quota left, allow it,
563 	 * otherwise wait. Note, we'll go negative and that's
564 	 * OK. We'll just get a little less next quota.
565 	 *
566 	 * Note on going negative: that allows us to process
567 	 * requests in order better, since we won't allow
568 	 * shorter reads to get around the long one that we
569 	 * don't have the quota to do just yet. It also prevents
570 	 * starvation by being a little more permissive about
571 	 * what we let through this quantum (to prevent the
572 	 * starvation), at the cost of getting a little less
573 	 * next quantum.
574 	 *
575 	 * Also note that if the current limit is <= 0,
576 	 * we treat it as unlimited as a failsafe.
577 	 */
578 	if (ios->current > 0 && ios->l_value1 <= 0)
579 		return EAGAIN;
580 
581 	return 0;
582 }
583 
584 static int
585 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
586 {
587 	int rv;
588 
589 	rv = cam_iosched_limiter_caniop(ios, bp);
590 	if (rv == 0)
591 		ios->l_value1 -= bp->bio_length;
592 
593 	return rv;
594 }
595 
596 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
597 
598 static void
599 cam_iosched_ticker(void *arg)
600 {
601 	struct cam_iosched_softc *isc = arg;
602 	sbintime_t now, delta;
603 	int pending;
604 
605 	callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
606 
607 	now = sbinuptime();
608 	delta = now - isc->last_time;
609 	isc->this_frac = (uint32_t)delta >> 16;		/* Note: discards seconds -- should be 0 harmless if not */
610 	isc->last_time = now;
611 
612 	cam_iosched_cl_maybe_steer(&isc->cl);
613 
614 	cam_iosched_limiter_tick(&isc->read_stats);
615 	cam_iosched_limiter_tick(&isc->write_stats);
616 	cam_iosched_limiter_tick(&isc->trim_stats);
617 
618 	cam_iosched_schedule(isc, isc->periph);
619 
620 	/*
621 	 * isc->load is an EMA of the pending I/Os at each tick. The number of
622 	 * pending I/Os is the sum of the I/Os queued to the hardware, and those
623 	 * in the software queue that could be queued to the hardware if there
624 	 * were slots.
625 	 *
626 	 * ios_stats.pending is a count of requests in the SIM right now for
627 	 * each of these types of I/O. So the total pending count is the sum of
628 	 * these I/Os and the sum of the queued I/Os still in the software queue
629 	 * for those operations that aren't being rate limited at the moment.
630 	 *
631 	 * The reason for the rate limiting bit is because those I/Os
632 	 * aren't part of the software queued load (since we could
633 	 * give them to hardware, but choose not to).
634 	 *
635 	 * Note: due to a bug in counting pending TRIM in the device, we
636 	 * don't include them in this count. We count each BIO_DELETE in
637 	 * the pending count, but the periph drivers collapse them down
638 	 * into one TRIM command. That one trim command gets the completion
639 	 * so the counts get off.
640 	 */
641 	pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
642 	pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
643 	    !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
644 	    !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
645 	pending <<= 16;
646 	pending /= isc->periph->path->device->ccbq.total_openings;
647 
648 	isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
649 
650 	isc->total_ticks++;
651 }
652 
653 static void
654 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
655 {
656 
657 	clp->next_steer = sbinuptime();
658 	clp->softc = isc;
659 	clp->steer_interval = SBT_1S * 5;	/* Let's start out steering every 5s */
660 	clp->lolat = 5 * SBT_1MS;
661 	clp->hilat = 15 * SBT_1MS;
662 	clp->alpha = 20;			/* Alpha == gain. 20 = .2 */
663 	clp->type = set_max;
664 }
665 
666 static void
667 cam_iosched_cl_maybe_steer(struct control_loop *clp)
668 {
669 	struct cam_iosched_softc *isc;
670 	sbintime_t now, lat;
671 	int old;
672 
673 	isc = clp->softc;
674 	now = isc->last_time;
675 	if (now < clp->next_steer)
676 		return;
677 
678 	clp->next_steer = now + clp->steer_interval;
679 	switch (clp->type) {
680 	case set_max:
681 		if (isc->write_stats.current != isc->write_stats.max)
682 			printf("Steering write from %d kBps to %d kBps\n",
683 			    isc->write_stats.current, isc->write_stats.max);
684 		isc->read_stats.current = isc->read_stats.max;
685 		isc->write_stats.current = isc->write_stats.max;
686 		isc->trim_stats.current = isc->trim_stats.max;
687 		break;
688 	case read_latency:
689 		old = isc->write_stats.current;
690 		lat = isc->read_stats.ema;
691 		/*
692 		 * Simple PLL-like engine. Since we're steering to a range for
693 		 * the SP (set point) that makes things a little more
694 		 * complicated. In addition, we're not directly controlling our
695 		 * PV (process variable), the read latency, but instead are
696 		 * manipulating the write bandwidth limit for our MV
697 		 * (manipulation variable), analysis of this code gets a bit
698 		 * messy. Also, the MV is a very noisy control surface for read
699 		 * latency since it is affected by many hidden processes inside
700 		 * the device which change how responsive read latency will be
701 		 * in reaction to changes in write bandwidth. Unlike the classic
702 		 * boiler control PLL. this may result in over-steering while
703 		 * the SSD takes its time to react to the new, lower load. This
704 		 * is why we use a relatively low alpha of between .1 and .25 to
705 		 * compensate for this effect. At .1, it takes ~22 steering
706 		 * intervals to back off by a factor of 10. At .2 it only takes
707 		 * ~10. At .25 it only takes ~8. However some preliminary data
708 		 * from the SSD drives suggests a reasponse time in 10's of
709 		 * seconds before latency drops regardless of the new write
710 		 * rate. Careful observation will be required to tune this
711 		 * effectively.
712 		 *
713 		 * Also, when there's no read traffic, we jack up the write
714 		 * limit too regardless of the last read latency.  10 is
715 		 * somewhat arbitrary.
716 		 */
717 		if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
718 			isc->write_stats.current = isc->write_stats.current *
719 			    (100 + clp->alpha) / 100;	/* Scale up */
720 		else if (lat > clp->hilat)
721 			isc->write_stats.current = isc->write_stats.current *
722 			    (100 - clp->alpha) / 100;	/* Scale down */
723 		clp->last_count = isc->read_stats.total;
724 
725 		/*
726 		 * Even if we don't steer, per se, enforce the min/max limits as
727 		 * those may have changed.
728 		 */
729 		if (isc->write_stats.current < isc->write_stats.min)
730 			isc->write_stats.current = isc->write_stats.min;
731 		if (isc->write_stats.current > isc->write_stats.max)
732 			isc->write_stats.current = isc->write_stats.max;
733 		if (old != isc->write_stats.current && 	iosched_debug)
734 			printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
735 			    old, isc->write_stats.current,
736 			    (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
737 		break;
738 	case cl_max:
739 		break;
740 	}
741 }
742 #endif
743 
744 /*
745  * Trim or similar currently pending completion. Should only be set for
746  * those drivers wishing only one Trim active at a time.
747  */
748 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE	(1ul << 0)
749 			/* Callout active, and needs to be torn down */
750 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
751 
752 			/* Periph drivers set these flags to indicate work */
753 #define CAM_IOSCHED_FLAG_WORK_FLAGS	((0xffffu) << 16)
754 
755 #ifdef CAM_IOSCHED_DYNAMIC
756 static void
757 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
758     sbintime_t sim_latency, int cmd, size_t size);
759 #endif
760 
761 static inline bool
762 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
763 {
764 	return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
765 }
766 
767 static inline bool
768 cam_iosched_has_io(struct cam_iosched_softc *isc)
769 {
770 #ifdef CAM_IOSCHED_DYNAMIC
771 	if (do_dynamic_iosched) {
772 		struct bio *rbp = bioq_first(&isc->bio_queue);
773 		struct bio *wbp = bioq_first(&isc->write_queue);
774 		bool can_write = wbp != NULL &&
775 		    cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
776 		bool can_read = rbp != NULL &&
777 		    cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
778 		if (iosched_debug > 2) {
779 			printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
780 			printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
781 			printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
782 		}
783 		return can_read || can_write;
784 	}
785 #endif
786 	return bioq_first(&isc->bio_queue) != NULL;
787 }
788 
789 static inline bool
790 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
791 {
792 	struct bio *bp;
793 
794 	bp = bioq_first(&isc->trim_queue);
795 #ifdef CAM_IOSCHED_DYNAMIC
796 	if (do_dynamic_iosched) {
797 		/*
798 		 * If we're limiting trims, then defer action on trims
799 		 * for a bit.
800 		 */
801 		if (bp == NULL || cam_iosched_limiter_caniop(&isc->trim_stats, bp) != 0)
802 			return false;
803 	}
804 #endif
805 
806 	/*
807 	 * If we've set a trim_goal, then if we exceed that allow trims
808 	 * to be passed back to the driver. If we've also set a tick timeout
809 	 * allow trims back to the driver. Otherwise, don't allow trims yet.
810 	 */
811 	if (isc->trim_goal > 0) {
812 		if (isc->queued_trims >= isc->trim_goal)
813 			return true;
814 		if (isc->queued_trims > 0 &&
815 		    isc->trim_ticks > 0 &&
816 		    ticks - isc->last_trim_tick > isc->trim_ticks)
817 			return true;
818 		return false;
819 	}
820 
821 	/* NB: Should perhaps have a max trim active independent of I/O limiters */
822 	return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) && bp != NULL;
823 }
824 
825 #define cam_iosched_sort_queue(isc)	((isc)->sort_io_queue >= 0 ?	\
826     (isc)->sort_io_queue : cam_sort_io_queues)
827 
828 static inline bool
829 cam_iosched_has_work(struct cam_iosched_softc *isc)
830 {
831 #ifdef CAM_IOSCHED_DYNAMIC
832 	if (iosched_debug > 2)
833 		printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
834 		    cam_iosched_has_more_trim(isc),
835 		    cam_iosched_has_flagged_work(isc));
836 #endif
837 
838 	return cam_iosched_has_io(isc) ||
839 		cam_iosched_has_more_trim(isc) ||
840 		cam_iosched_has_flagged_work(isc);
841 }
842 
843 #ifdef CAM_IOSCHED_DYNAMIC
844 static void
845 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
846 {
847 
848 	ios->limiter = none;
849 	ios->in = 0;
850 	ios->max = ios->current = 300000;
851 	ios->min = 1;
852 	ios->out = 0;
853 	ios->errs = 0;
854 	ios->pending = 0;
855 	ios->queued = 0;
856 	ios->total = 0;
857 	ios->ema = 0;
858 	ios->emvar = 0;
859 	ios->softc = isc;
860 	cam_iosched_limiter_init(ios);
861 }
862 
863 static int
864 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
865 {
866 	char buf[16];
867 	struct iop_stats *ios;
868 	struct cam_iosched_softc *isc;
869 	int value, i, error;
870 	const char *p;
871 
872 	ios = arg1;
873 	isc = ios->softc;
874 	value = ios->limiter;
875 	if (value < none || value >= limiter_max)
876 		p = "UNKNOWN";
877 	else
878 		p = cam_iosched_limiter_names[value];
879 
880 	strlcpy(buf, p, sizeof(buf));
881 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
882 	if (error != 0 || req->newptr == NULL)
883 		return error;
884 
885 	cam_periph_lock(isc->periph);
886 
887 	for (i = none; i < limiter_max; i++) {
888 		if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
889 			continue;
890 		ios->limiter = i;
891 		error = cam_iosched_limiter_init(ios);
892 		if (error != 0) {
893 			ios->limiter = value;
894 			cam_periph_unlock(isc->periph);
895 			return error;
896 		}
897 		/* Note: disk load averate requires ticker to be always running */
898 		callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
899 		isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
900 
901 		cam_periph_unlock(isc->periph);
902 		return 0;
903 	}
904 
905 	cam_periph_unlock(isc->periph);
906 	return EINVAL;
907 }
908 
909 static int
910 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
911 {
912 	char buf[16];
913 	struct control_loop *clp;
914 	struct cam_iosched_softc *isc;
915 	int value, i, error;
916 	const char *p;
917 
918 	clp = arg1;
919 	isc = clp->softc;
920 	value = clp->type;
921 	if (value < none || value >= cl_max)
922 		p = "UNKNOWN";
923 	else
924 		p = cam_iosched_control_type_names[value];
925 
926 	strlcpy(buf, p, sizeof(buf));
927 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
928 	if (error != 0 || req->newptr == NULL)
929 		return error;
930 
931 	for (i = set_max; i < cl_max; i++) {
932 		if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
933 			continue;
934 		cam_periph_lock(isc->periph);
935 		clp->type = i;
936 		cam_periph_unlock(isc->periph);
937 		return 0;
938 	}
939 
940 	return EINVAL;
941 }
942 
943 static int
944 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
945 {
946 	char buf[16];
947 	sbintime_t value;
948 	int error;
949 	uint64_t us;
950 
951 	value = *(sbintime_t *)arg1;
952 	us = (uint64_t)value / SBT_1US;
953 	snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
954 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
955 	if (error != 0 || req->newptr == NULL)
956 		return error;
957 	us = strtoul(buf, NULL, 10);
958 	if (us == 0)
959 		return EINVAL;
960 	*(sbintime_t *)arg1 = us * SBT_1US;
961 	return 0;
962 }
963 
964 static int
965 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
966 {
967 	int i, error;
968 	struct sbuf sb;
969 	uint64_t *latencies;
970 
971 	latencies = arg1;
972 	sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
973 
974 	for (i = 0; i < LAT_BUCKETS - 1; i++)
975 		sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
976 	sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
977 	error = sbuf_finish(&sb);
978 	sbuf_delete(&sb);
979 
980 	return (error);
981 }
982 
983 static int
984 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS)
985 {
986 	int *quanta;
987 	int error, value;
988 
989 	quanta = (unsigned *)arg1;
990 	value = *quanta;
991 
992 	error = sysctl_handle_int(oidp, (int *)&value, 0, req);
993 	if ((error != 0) || (req->newptr == NULL))
994 		return (error);
995 
996 	if (value < 1 || value > hz)
997 		return (EINVAL);
998 
999 	*quanta = value;
1000 
1001 	return (0);
1002 }
1003 
1004 static void
1005 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
1006 {
1007 	struct sysctl_oid_list *n;
1008 	struct sysctl_ctx_list *ctx;
1009 
1010 	ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1011 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
1012 	    CTLFLAG_RD | CTLFLAG_MPSAFE, 0, name);
1013 	n = SYSCTL_CHILDREN(ios->sysctl_tree);
1014 	ctx = &ios->sysctl_ctx;
1015 
1016 	SYSCTL_ADD_UQUAD(ctx, n,
1017 	    OID_AUTO, "ema", CTLFLAG_RD,
1018 	    &ios->ema,
1019 	    "Fast Exponentially Weighted Moving Average");
1020 	SYSCTL_ADD_UQUAD(ctx, n,
1021 	    OID_AUTO, "emvar", CTLFLAG_RD,
1022 	    &ios->emvar,
1023 	    "Fast Exponentially Weighted Moving Variance");
1024 
1025 	SYSCTL_ADD_INT(ctx, n,
1026 	    OID_AUTO, "pending", CTLFLAG_RD,
1027 	    &ios->pending, 0,
1028 	    "Instantaneous # of pending transactions");
1029 	SYSCTL_ADD_INT(ctx, n,
1030 	    OID_AUTO, "count", CTLFLAG_RD,
1031 	    &ios->total, 0,
1032 	    "# of transactions submitted to hardware");
1033 	SYSCTL_ADD_INT(ctx, n,
1034 	    OID_AUTO, "queued", CTLFLAG_RD,
1035 	    &ios->queued, 0,
1036 	    "# of transactions in the queue");
1037 	SYSCTL_ADD_INT(ctx, n,
1038 	    OID_AUTO, "in", CTLFLAG_RD,
1039 	    &ios->in, 0,
1040 	    "# of transactions queued to driver");
1041 	SYSCTL_ADD_INT(ctx, n,
1042 	    OID_AUTO, "out", CTLFLAG_RD,
1043 	    &ios->out, 0,
1044 	    "# of transactions completed (including with error)");
1045 	SYSCTL_ADD_INT(ctx, n,
1046 	    OID_AUTO, "errs", CTLFLAG_RD,
1047 	    &ios->errs, 0,
1048 	    "# of transactions completed with an error");
1049 
1050 	SYSCTL_ADD_PROC(ctx, n,
1051 	    OID_AUTO, "limiter",
1052 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1053 	    ios, 0, cam_iosched_limiter_sysctl, "A",
1054 	    "Current limiting type.");
1055 	SYSCTL_ADD_INT(ctx, n,
1056 	    OID_AUTO, "min", CTLFLAG_RW,
1057 	    &ios->min, 0,
1058 	    "min resource");
1059 	SYSCTL_ADD_INT(ctx, n,
1060 	    OID_AUTO, "max", CTLFLAG_RW,
1061 	    &ios->max, 0,
1062 	    "max resource");
1063 	SYSCTL_ADD_INT(ctx, n,
1064 	    OID_AUTO, "current", CTLFLAG_RW,
1065 	    &ios->current, 0,
1066 	    "current resource");
1067 
1068 	SYSCTL_ADD_PROC(ctx, n,
1069 	    OID_AUTO, "latencies",
1070 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
1071 	    &ios->latencies, 0,
1072 	    cam_iosched_sysctl_latencies, "A",
1073 	    "Array of latencies, a geometric progresson from\n"
1074 	    "kern.cam.iosched.bucket_base_us with a ratio of\n"
1075 	    "kern.cam.iosched.bucket_ration / 100 from one to\n"
1076 	    "the next. By default 20 steps from 20us to 10.485s\n"
1077 	    "by doubling.");
1078 
1079 }
1080 
1081 static void
1082 cam_iosched_iop_stats_fini(struct iop_stats *ios)
1083 {
1084 	if (ios->sysctl_tree)
1085 		if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
1086 			printf("can't remove iosched sysctl stats context\n");
1087 }
1088 
1089 static void
1090 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
1091 {
1092 	struct sysctl_oid_list *n;
1093 	struct sysctl_ctx_list *ctx;
1094 	struct control_loop *clp;
1095 
1096 	clp = &isc->cl;
1097 	clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1098 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
1099 	    CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Control loop info");
1100 	n = SYSCTL_CHILDREN(clp->sysctl_tree);
1101 	ctx = &clp->sysctl_ctx;
1102 
1103 	SYSCTL_ADD_PROC(ctx, n,
1104 	    OID_AUTO, "type",
1105 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1106 	    clp, 0, cam_iosched_control_type_sysctl, "A",
1107 	    "Control loop algorithm");
1108 	SYSCTL_ADD_PROC(ctx, n,
1109 	    OID_AUTO, "steer_interval",
1110 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1111 	    &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
1112 	    "How often to steer (in us)");
1113 	SYSCTL_ADD_PROC(ctx, n,
1114 	    OID_AUTO, "lolat",
1115 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1116 	    &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
1117 	    "Low water mark for Latency (in us)");
1118 	SYSCTL_ADD_PROC(ctx, n,
1119 	    OID_AUTO, "hilat",
1120 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1121 	    &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
1122 	    "Hi water mark for Latency (in us)");
1123 	SYSCTL_ADD_INT(ctx, n,
1124 	    OID_AUTO, "alpha", CTLFLAG_RW,
1125 	    &clp->alpha, 0,
1126 	    "Alpha for PLL (x100) aka gain");
1127 }
1128 
1129 static void
1130 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1131 {
1132 	if (clp->sysctl_tree)
1133 		if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1134 			printf("can't remove iosched sysctl control loop context\n");
1135 }
1136 #endif
1137 
1138 /*
1139  * Allocate the iosched structure. This also insulates callers from knowing
1140  * sizeof struct cam_iosched_softc.
1141  */
1142 int
1143 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1144 {
1145 
1146 	*iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1147 	if (*iscp == NULL)
1148 		return ENOMEM;
1149 #ifdef CAM_IOSCHED_DYNAMIC
1150 	if (iosched_debug)
1151 		printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1152 #endif
1153 	(*iscp)->sort_io_queue = -1;
1154 	bioq_init(&(*iscp)->bio_queue);
1155 	bioq_init(&(*iscp)->trim_queue);
1156 #ifdef CAM_IOSCHED_DYNAMIC
1157 	if (do_dynamic_iosched) {
1158 		bioq_init(&(*iscp)->write_queue);
1159 		(*iscp)->read_bias = default_read_bias;
1160 		(*iscp)->current_read_bias = 0;
1161 		(*iscp)->quanta = min(hz, 200);
1162 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1163 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1164 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1165 		(*iscp)->trim_stats.max = 1;	/* Trims are special: one at a time for now */
1166 		(*iscp)->last_time = sbinuptime();
1167 		callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1168 		(*iscp)->periph = periph;
1169 		cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1170 		callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp);
1171 		(*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1172 	}
1173 #endif
1174 
1175 	return 0;
1176 }
1177 
1178 /*
1179  * Reclaim all used resources. This assumes that other folks have
1180  * drained the requests in the hardware. Maybe an unwise assumption.
1181  */
1182 void
1183 cam_iosched_fini(struct cam_iosched_softc *isc)
1184 {
1185 	if (isc) {
1186 		cam_iosched_flush(isc, NULL, ENXIO);
1187 #ifdef CAM_IOSCHED_DYNAMIC
1188 		cam_iosched_iop_stats_fini(&isc->read_stats);
1189 		cam_iosched_iop_stats_fini(&isc->write_stats);
1190 		cam_iosched_iop_stats_fini(&isc->trim_stats);
1191 		cam_iosched_cl_sysctl_fini(&isc->cl);
1192 		if (isc->sysctl_tree)
1193 			if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1194 				printf("can't remove iosched sysctl stats context\n");
1195 		if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1196 			callout_drain(&isc->ticker);
1197 			isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1198 		}
1199 #endif
1200 		free(isc, M_CAMSCHED);
1201 	}
1202 }
1203 
1204 /*
1205  * After we're sure we're attaching a device, go ahead and add
1206  * hooks for any sysctl we may wish to honor.
1207  */
1208 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1209     struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1210 {
1211 	struct sysctl_oid_list *n;
1212 
1213 	n = SYSCTL_CHILDREN(node);
1214 	SYSCTL_ADD_INT(ctx, n,
1215 		OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1216 		&isc->sort_io_queue, 0,
1217 		"Sort IO queue to try and optimise disk access patterns");
1218 	SYSCTL_ADD_INT(ctx, n,
1219 	    OID_AUTO, "trim_goal", CTLFLAG_RW,
1220 	    &isc->trim_goal, 0,
1221 	    "Number of trims to try to accumulate before sending to hardware");
1222 	SYSCTL_ADD_INT(ctx, n,
1223 	    OID_AUTO, "trim_ticks", CTLFLAG_RW,
1224 	    &isc->trim_goal, 0,
1225 	    "IO Schedul qaunta to hold back trims for when accumulating");
1226 
1227 #ifdef CAM_IOSCHED_DYNAMIC
1228 	if (!do_dynamic_iosched)
1229 		return;
1230 
1231 	isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1232 	    SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1233 	    CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "I/O scheduler statistics");
1234 	n = SYSCTL_CHILDREN(isc->sysctl_tree);
1235 	ctx = &isc->sysctl_ctx;
1236 
1237 	cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1238 	cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1239 	cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1240 	cam_iosched_cl_sysctl_init(isc);
1241 
1242 	SYSCTL_ADD_INT(ctx, n,
1243 	    OID_AUTO, "read_bias", CTLFLAG_RW,
1244 	    &isc->read_bias, default_read_bias,
1245 	    "How biased towards read should we be independent of limits");
1246 
1247 	SYSCTL_ADD_PROC(ctx, n,
1248 	    OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1249 	    &isc->quanta, 0, cam_iosched_quanta_sysctl, "I",
1250 	    "How many quanta per second do we slice the I/O up into");
1251 
1252 	SYSCTL_ADD_INT(ctx, n,
1253 	    OID_AUTO, "total_ticks", CTLFLAG_RD,
1254 	    &isc->total_ticks, 0,
1255 	    "Total number of ticks we've done");
1256 
1257 	SYSCTL_ADD_INT(ctx, n,
1258 	    OID_AUTO, "load", CTLFLAG_RD,
1259 	    &isc->load, 0,
1260 	    "scaled load average / 100");
1261 
1262 	SYSCTL_ADD_U64(ctx, n,
1263 	    OID_AUTO, "latency_trigger", CTLFLAG_RW,
1264 	    &isc->max_lat, 0,
1265 	    "Latency treshold to trigger callbacks");
1266 #endif
1267 }
1268 
1269 void
1270 cam_iosched_set_latfcn(struct cam_iosched_softc *isc,
1271     cam_iosched_latfcn_t fnp, void *argp)
1272 {
1273 #ifdef CAM_IOSCHED_DYNAMIC
1274 	isc->latfcn = fnp;
1275 	isc->latarg = argp;
1276 #endif
1277 }
1278 
1279 /*
1280  * Client drivers can set two parameters. "goal" is the number of BIO_DELETEs
1281  * that will be queued up before iosched will "release" the trims to the client
1282  * driver to wo with what they will (usually combine as many as possible). If we
1283  * don't get this many, after trim_ticks we'll submit the I/O anyway with
1284  * whatever we have.  We do need an I/O of some kind of to clock the deferred
1285  * trims out to disk. Since we will eventually get a write for the super block
1286  * or something before we shutdown, the trims will complete. To be safe, when a
1287  * BIO_FLUSH is presented to the iosched work queue, we set the ticks time far
1288  * enough in the past so we'll present the BIO_DELETEs to the client driver.
1289  * There might be a race if no BIO_DELETESs were queued, a BIO_FLUSH comes in
1290  * and then a BIO_DELETE is sent down. No know client does this, and there's
1291  * already a race between an ordered BIO_FLUSH and any BIO_DELETEs in flight,
1292  * but no client depends on the ordering being honored.
1293  *
1294  * XXX I'm not sure what the interaction between UFS direct BIOs and the BUF
1295  * flushing on shutdown. I think there's bufs that would be dependent on the BIO
1296  * finishing to write out at least metadata, so we'll be fine. To be safe, keep
1297  * the number of ticks low (less than maybe 10s) to avoid shutdown races.
1298  */
1299 
1300 void
1301 cam_iosched_set_trim_goal(struct cam_iosched_softc *isc, int goal)
1302 {
1303 
1304 	isc->trim_goal = goal;
1305 }
1306 
1307 void
1308 cam_iosched_set_trim_ticks(struct cam_iosched_softc *isc, int trim_ticks)
1309 {
1310 
1311 	isc->trim_ticks = trim_ticks;
1312 }
1313 
1314 /*
1315  * Flush outstanding I/O. Consumers of this library don't know all the
1316  * queues we may keep, so this allows all I/O to be flushed in one
1317  * convenient call.
1318  */
1319 void
1320 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1321 {
1322 	bioq_flush(&isc->bio_queue, stp, err);
1323 	bioq_flush(&isc->trim_queue, stp, err);
1324 #ifdef CAM_IOSCHED_DYNAMIC
1325 	if (do_dynamic_iosched)
1326 		bioq_flush(&isc->write_queue, stp, err);
1327 #endif
1328 }
1329 
1330 #ifdef CAM_IOSCHED_DYNAMIC
1331 static struct bio *
1332 cam_iosched_get_write(struct cam_iosched_softc *isc)
1333 {
1334 	struct bio *bp;
1335 
1336 	/*
1337 	 * We control the write rate by controlling how many requests we send
1338 	 * down to the drive at any one time. Fewer requests limits the
1339 	 * effects of both starvation when the requests take a while and write
1340 	 * amplification when each request is causing more than one write to
1341 	 * the NAND media. Limiting the queue depth like this will also limit
1342 	 * the write throughput and give and reads that want to compete to
1343 	 * compete unfairly.
1344 	 */
1345 	bp = bioq_first(&isc->write_queue);
1346 	if (bp == NULL) {
1347 		if (iosched_debug > 3)
1348 			printf("No writes present in write_queue\n");
1349 		return NULL;
1350 	}
1351 
1352 	/*
1353 	 * If pending read, prefer that based on current read bias
1354 	 * setting.
1355 	 */
1356 	if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1357 		if (iosched_debug)
1358 			printf(
1359 			    "Reads present and current_read_bias is %d queued "
1360 			    "writes %d queued reads %d\n",
1361 			    isc->current_read_bias, isc->write_stats.queued,
1362 			    isc->read_stats.queued);
1363 		isc->current_read_bias--;
1364 		/* We're not limiting writes, per se, just doing reads first */
1365 		return NULL;
1366 	}
1367 
1368 	/*
1369 	 * See if our current limiter allows this I/O.
1370 	 */
1371 	if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1372 		if (iosched_debug)
1373 			printf("Can't write because limiter says no.\n");
1374 		isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1375 		return NULL;
1376 	}
1377 
1378 	/*
1379 	 * Let's do this: We've passed all the gates and we're a go
1380 	 * to schedule the I/O in the SIM.
1381 	 */
1382 	isc->current_read_bias = isc->read_bias;
1383 	bioq_remove(&isc->write_queue, bp);
1384 	if (bp->bio_cmd == BIO_WRITE) {
1385 		isc->write_stats.queued--;
1386 		isc->write_stats.total++;
1387 		isc->write_stats.pending++;
1388 	}
1389 	if (iosched_debug > 9)
1390 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1391 	isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1392 	return bp;
1393 }
1394 #endif
1395 
1396 /*
1397  * Put back a trim that you weren't able to actually schedule this time.
1398  */
1399 void
1400 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1401 {
1402 	bioq_insert_head(&isc->trim_queue, bp);
1403 	if (isc->queued_trims == 0)
1404 		isc->last_trim_tick = ticks;
1405 	isc->queued_trims++;
1406 #ifdef CAM_IOSCHED_DYNAMIC
1407 	isc->trim_stats.queued++;
1408 	isc->trim_stats.total--;		/* since we put it back, don't double count */
1409 	isc->trim_stats.pending--;
1410 #endif
1411 }
1412 
1413 /*
1414  * gets the next trim from the trim queue.
1415  *
1416  * Assumes we're called with the periph lock held.  It removes this
1417  * trim from the queue and the device must explicitly reinsert it
1418  * should the need arise.
1419  */
1420 struct bio *
1421 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1422 {
1423 	struct bio *bp;
1424 
1425 	bp  = bioq_first(&isc->trim_queue);
1426 	if (bp == NULL)
1427 		return NULL;
1428 	bioq_remove(&isc->trim_queue, bp);
1429 	isc->queued_trims--;
1430 	isc->last_trim_tick = ticks;	/* Reset the tick timer when we take trims */
1431 #ifdef CAM_IOSCHED_DYNAMIC
1432 	isc->trim_stats.queued--;
1433 	isc->trim_stats.total++;
1434 	isc->trim_stats.pending++;
1435 #endif
1436 	return bp;
1437 }
1438 
1439 /*
1440  * gets an available trim from the trim queue, if there's no trim
1441  * already pending. It removes this trim from the queue and the device
1442  * must explicitly reinsert it should the need arise.
1443  *
1444  * Assumes we're called with the periph lock held.
1445  */
1446 struct bio *
1447 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1448 {
1449 #ifdef CAM_IOSCHED_DYNAMIC
1450 	struct bio *bp;
1451 #endif
1452 
1453 	if (!cam_iosched_has_more_trim(isc))
1454 		return NULL;
1455 #ifdef CAM_IOSCHED_DYNAMIC
1456 	bp  = bioq_first(&isc->trim_queue);
1457 	if (bp == NULL)
1458 		return NULL;
1459 
1460 	/*
1461 	 * If pending read, prefer that based on current read bias setting. The
1462 	 * read bias is shared for both writes and TRIMs, but on TRIMs the bias
1463 	 * is for a combined TRIM not a single TRIM request that's come in.
1464 	 */
1465 	if (do_dynamic_iosched) {
1466 		if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1467 			if (iosched_debug)
1468 				printf("Reads present and current_read_bias is %d"
1469 				    " queued trims %d queued reads %d\n",
1470 				    isc->current_read_bias, isc->trim_stats.queued,
1471 				    isc->read_stats.queued);
1472 			isc->current_read_bias--;
1473 			/* We're not limiting TRIMS, per se, just doing reads first */
1474 			return NULL;
1475 		}
1476 		/*
1477 		 * We're going to do a trim, so reset the bias.
1478 		 */
1479 		isc->current_read_bias = isc->read_bias;
1480 	}
1481 
1482 	/*
1483 	 * See if our current limiter allows this I/O. Because we only call this
1484 	 * here, and not in next_trim, the 'bandwidth' limits for trims won't
1485 	 * work, while the iops or max queued limits will work. It's tricky
1486 	 * because we want the limits to be from the perspective of the
1487 	 * "commands sent to the device." To make iops work, we need to check
1488 	 * only here (since we want all the ops we combine to count as one). To
1489 	 * make bw limits work, we'd need to check in next_trim, but that would
1490 	 * have the effect of limiting the iops as seen from the upper layers.
1491 	 */
1492 	if (cam_iosched_limiter_iop(&isc->trim_stats, bp) != 0) {
1493 		if (iosched_debug)
1494 			printf("Can't trim because limiter says no.\n");
1495 		isc->trim_stats.state_flags |= IOP_RATE_LIMITED;
1496 		return NULL;
1497 	}
1498 	isc->current_read_bias = isc->read_bias;
1499 	isc->trim_stats.state_flags &= ~IOP_RATE_LIMITED;
1500 	/* cam_iosched_next_trim below keeps proper book */
1501 #endif
1502 	return cam_iosched_next_trim(isc);
1503 }
1504 
1505 
1506 #ifdef CAM_IOSCHED_DYNAMIC
1507 static struct bio *
1508 bio_next(struct bio *bp)
1509 {
1510 	bp = TAILQ_NEXT(bp, bio_queue);
1511 	/*
1512 	 * After the first commands, the ordered bit terminates
1513 	 * our search because BIO_ORDERED acts like a barrier.
1514 	 */
1515 	if (bp == NULL || bp->bio_flags & BIO_ORDERED)
1516 		return NULL;
1517 	return bp;
1518 }
1519 
1520 static bool
1521 cam_iosched_rate_limited(struct iop_stats *ios)
1522 {
1523 	return ios->state_flags & IOP_RATE_LIMITED;
1524 }
1525 #endif
1526 
1527 /*
1528  * Determine what the next bit of work to do is for the periph. The
1529  * default implementation looks to see if we have trims to do, but no
1530  * trims outstanding. If so, we do that. Otherwise we see if we have
1531  * other work. If we do, then we do that. Otherwise why were we called?
1532  */
1533 struct bio *
1534 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1535 {
1536 	struct bio *bp;
1537 
1538 	/*
1539 	 * See if we have a trim that can be scheduled. We can only send one
1540 	 * at a time down, so this takes that into account.
1541 	 *
1542 	 * XXX newer TRIM commands are queueable. Revisit this when we
1543 	 * implement them.
1544 	 */
1545 	if ((bp = cam_iosched_get_trim(isc)) != NULL)
1546 		return bp;
1547 
1548 #ifdef CAM_IOSCHED_DYNAMIC
1549 	/*
1550 	 * See if we have any pending writes, room in the queue for them,
1551 	 * and no pending reads (unless we've scheduled too many).
1552 	 * if so, those are next.
1553 	 */
1554 	if (do_dynamic_iosched) {
1555 		if ((bp = cam_iosched_get_write(isc)) != NULL)
1556 			return bp;
1557 	}
1558 #endif
1559 	/*
1560 	 * next, see if there's other, normal I/O waiting. If so return that.
1561 	 */
1562 #ifdef CAM_IOSCHED_DYNAMIC
1563 	if (do_dynamic_iosched) {
1564 		for (bp = bioq_first(&isc->bio_queue); bp != NULL;
1565 		     bp = bio_next(bp)) {
1566 			/*
1567 			 * For the dynamic scheduler with a read bias, bio_queue
1568 			 * is only for reads. However, without one, all
1569 			 * operations are queued. Enforce limits here for any
1570 			 * operation we find here.
1571 			 */
1572 			if (bp->bio_cmd == BIO_READ) {
1573 				if (cam_iosched_rate_limited(&isc->read_stats) ||
1574 				    cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1575 					isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1576 					continue;
1577 				}
1578 				isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1579 			}
1580 			/*
1581 			 * There can only be write requests on the queue when
1582 			 * the read bias is 0, but we need to process them
1583 			 * here. We do not assert for read bias == 0, however,
1584 			 * since it is dynamic and we can have WRITE operations
1585 			 * in the queue after we transition from 0 to non-zero.
1586 			 */
1587 			if (bp->bio_cmd == BIO_WRITE) {
1588 				if (cam_iosched_rate_limited(&isc->write_stats) ||
1589 				    cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1590 					isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1591 					continue;
1592 				}
1593 				isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1594 			}
1595 			/*
1596 			 * here we know we have a bp that's != NULL, that's not rate limited
1597 			 * and can be the next I/O.
1598 			 */
1599 			break;
1600 		}
1601 	} else
1602 #endif
1603 		bp = bioq_first(&isc->bio_queue);
1604 
1605 	if (bp == NULL)
1606 		return (NULL);
1607 	bioq_remove(&isc->bio_queue, bp);
1608 #ifdef CAM_IOSCHED_DYNAMIC
1609 	if (do_dynamic_iosched) {
1610 		if (bp->bio_cmd == BIO_READ) {
1611 			isc->read_stats.queued--;
1612 			isc->read_stats.total++;
1613 			isc->read_stats.pending++;
1614 		} else if (bp->bio_cmd == BIO_WRITE) {
1615 			isc->write_stats.queued--;
1616 			isc->write_stats.total++;
1617 			isc->write_stats.pending++;
1618 		}
1619 	}
1620 	if (iosched_debug > 9)
1621 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1622 #endif
1623 	return bp;
1624 }
1625 
1626 /*
1627  * Driver has been given some work to do by the block layer. Tell the
1628  * scheduler about it and have it queue the work up. The scheduler module
1629  * will then return the currently most useful bit of work later, possibly
1630  * deferring work for various reasons.
1631  */
1632 void
1633 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1634 {
1635 
1636 	/*
1637 	 * A BIO_SPEEDUP from the upper layers means that they have a block
1638 	 * shortage. At the present, this is only sent when we're trying to
1639 	 * allocate blocks, but have a shortage before giving up. bio_length is
1640 	 * the size of their shortage. We will complete just enough BIO_DELETEs
1641 	 * in the queue to satisfy the need. If bio_length is 0, we'll complete
1642 	 * them all. This allows the scheduler to delay BIO_DELETEs to improve
1643 	 * read/write performance without worrying about the upper layers. When
1644 	 * it's possibly a problem, we respond by pretending the BIO_DELETEs
1645 	 * just worked. We can't do anything about the BIO_DELETEs in the
1646 	 * hardware, though. We have to wait for them to complete.
1647 	 */
1648 	if (bp->bio_cmd == BIO_SPEEDUP) {
1649 		off_t len;
1650 		struct bio *nbp;
1651 
1652 		len = 0;
1653 		while (bioq_first(&isc->trim_queue) &&
1654 		    (bp->bio_length == 0 || len < bp->bio_length)) {
1655 			nbp = bioq_takefirst(&isc->trim_queue);
1656 			len += nbp->bio_length;
1657 			nbp->bio_error = 0;
1658 			biodone(nbp);
1659 		}
1660 		if (bp->bio_length > 0) {
1661 			if (bp->bio_length > len)
1662 				bp->bio_resid = bp->bio_length - len;
1663 			else
1664 				bp->bio_resid = 0;
1665 		}
1666 		bp->bio_error = 0;
1667 		biodone(bp);
1668 		return;
1669 	}
1670 
1671 	/*
1672 	 * If we get a BIO_FLUSH, and we're doing delayed BIO_DELETEs then we
1673 	 * set the last tick time to one less than the current ticks minus the
1674 	 * delay to force the BIO_DELETEs to be presented to the client driver.
1675 	 */
1676 	if (bp->bio_cmd == BIO_FLUSH && isc->trim_ticks > 0)
1677 		isc->last_trim_tick = ticks - isc->trim_ticks - 1;
1678 
1679 	/*
1680 	 * Put all trims on the trim queue. Otherwise put the work on the bio
1681 	 * queue.
1682 	 */
1683 	if (bp->bio_cmd == BIO_DELETE) {
1684 		bioq_insert_tail(&isc->trim_queue, bp);
1685 		if (isc->queued_trims == 0)
1686 			isc->last_trim_tick = ticks;
1687 		isc->queued_trims++;
1688 #ifdef CAM_IOSCHED_DYNAMIC
1689 		isc->trim_stats.in++;
1690 		isc->trim_stats.queued++;
1691 #endif
1692 	}
1693 #ifdef CAM_IOSCHED_DYNAMIC
1694 	else if (do_dynamic_iosched && isc->read_bias != 0 &&
1695 	    (bp->bio_cmd != BIO_READ)) {
1696 		if (cam_iosched_sort_queue(isc))
1697 			bioq_disksort(&isc->write_queue, bp);
1698 		else
1699 			bioq_insert_tail(&isc->write_queue, bp);
1700 		if (iosched_debug > 9)
1701 			printf("Qw  : %p %#x\n", bp, bp->bio_cmd);
1702 		if (bp->bio_cmd == BIO_WRITE) {
1703 			isc->write_stats.in++;
1704 			isc->write_stats.queued++;
1705 		}
1706 	}
1707 #endif
1708 	else {
1709 		if (cam_iosched_sort_queue(isc))
1710 			bioq_disksort(&isc->bio_queue, bp);
1711 		else
1712 			bioq_insert_tail(&isc->bio_queue, bp);
1713 #ifdef CAM_IOSCHED_DYNAMIC
1714 		if (iosched_debug > 9)
1715 			printf("Qr  : %p %#x\n", bp, bp->bio_cmd);
1716 		if (bp->bio_cmd == BIO_READ) {
1717 			isc->read_stats.in++;
1718 			isc->read_stats.queued++;
1719 		} else if (bp->bio_cmd == BIO_WRITE) {
1720 			isc->write_stats.in++;
1721 			isc->write_stats.queued++;
1722 		}
1723 #endif
1724 	}
1725 }
1726 
1727 /*
1728  * If we have work, get it scheduled. Called with the periph lock held.
1729  */
1730 void
1731 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1732 {
1733 
1734 	if (cam_iosched_has_work(isc))
1735 		xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1736 }
1737 
1738 /*
1739  * Complete a trim request. Mark that we no longer have one in flight.
1740  */
1741 void
1742 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1743 {
1744 
1745 	isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1746 }
1747 
1748 /*
1749  * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1750  * might use notes in the ccb for statistics.
1751  */
1752 int
1753 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1754     union ccb *done_ccb)
1755 {
1756 	int retval = 0;
1757 #ifdef CAM_IOSCHED_DYNAMIC
1758 	if (!do_dynamic_iosched)
1759 		return retval;
1760 
1761 	if (iosched_debug > 10)
1762 		printf("done: %p %#x\n", bp, bp->bio_cmd);
1763 	if (bp->bio_cmd == BIO_WRITE) {
1764 		retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1765 		if ((bp->bio_flags & BIO_ERROR) != 0)
1766 			isc->write_stats.errs++;
1767 		isc->write_stats.out++;
1768 		isc->write_stats.pending--;
1769 	} else if (bp->bio_cmd == BIO_READ) {
1770 		retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1771 		if ((bp->bio_flags & BIO_ERROR) != 0)
1772 			isc->read_stats.errs++;
1773 		isc->read_stats.out++;
1774 		isc->read_stats.pending--;
1775 	} else if (bp->bio_cmd == BIO_DELETE) {
1776 		if ((bp->bio_flags & BIO_ERROR) != 0)
1777 			isc->trim_stats.errs++;
1778 		isc->trim_stats.out++;
1779 		isc->trim_stats.pending--;
1780 	} else if (bp->bio_cmd != BIO_FLUSH) {
1781 		if (iosched_debug)
1782 			printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1783 	}
1784 
1785 	if ((bp->bio_flags & BIO_ERROR) == 0 && done_ccb != NULL &&
1786 	    (done_ccb->ccb_h.status & CAM_QOS_VALID) != 0) {
1787 		sbintime_t sim_latency;
1788 
1789 		sim_latency = cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data);
1790 
1791 		cam_iosched_io_metric_update(isc, sim_latency,
1792 		    bp->bio_cmd, bp->bio_bcount);
1793 		/*
1794 		 * Debugging code: allow callbacks to the periph driver when latency max
1795 		 * is exceeded. This can be useful for triggering external debugging actions.
1796 		 */
1797 		if (isc->latfcn && isc->max_lat != 0 && sim_latency > isc->max_lat)
1798 			isc->latfcn(isc->latarg, sim_latency, bp);
1799 	}
1800 
1801 #endif
1802 	return retval;
1803 }
1804 
1805 /*
1806  * Tell the io scheduler that you've pushed a trim down into the sim.
1807  * This also tells the I/O scheduler not to push any more trims down, so
1808  * some periphs do not call it if they can cope with multiple trims in flight.
1809  */
1810 void
1811 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1812 {
1813 
1814 	isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1815 }
1816 
1817 /*
1818  * Change the sorting policy hint for I/O transactions for this device.
1819  */
1820 void
1821 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1822 {
1823 
1824 	isc->sort_io_queue = val;
1825 }
1826 
1827 int
1828 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1829 {
1830 	return isc->flags & flags;
1831 }
1832 
1833 void
1834 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1835 {
1836 	isc->flags |= flags;
1837 }
1838 
1839 void
1840 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1841 {
1842 	isc->flags &= ~flags;
1843 }
1844 
1845 #ifdef CAM_IOSCHED_DYNAMIC
1846 /*
1847  * After the method presented in Jack Crenshaw's 1998 article "Integer
1848  * Square Roots," reprinted at
1849  * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1850  * and well worth the read. Briefly, we find the power of 4 that's the
1851  * largest smaller than val. We then check each smaller power of 4 to
1852  * see if val is still bigger. The right shifts at each step divide
1853  * the result by 2 which after successive application winds up
1854  * accumulating the right answer. It could also have been accumulated
1855  * using a separate root counter, but this code is smaller and faster
1856  * than that method. This method is also integer size invariant.
1857  * It returns floor(sqrt((float)val)), or the largest integer less than
1858  * or equal to the square root.
1859  */
1860 static uint64_t
1861 isqrt64(uint64_t val)
1862 {
1863 	uint64_t res = 0;
1864 	uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1865 
1866 	/*
1867 	 * Find the largest power of 4 smaller than val.
1868 	 */
1869 	while (bit > val)
1870 		bit >>= 2;
1871 
1872 	/*
1873 	 * Accumulate the answer, one bit at a time (we keep moving
1874 	 * them over since 2 is the square root of 4 and we test
1875 	 * powers of 4). We accumulate where we find the bit, but
1876 	 * the successive shifts land the bit in the right place
1877 	 * by the end.
1878 	 */
1879 	while (bit != 0) {
1880 		if (val >= res + bit) {
1881 			val -= res + bit;
1882 			res = (res >> 1) + bit;
1883 		} else
1884 			res >>= 1;
1885 		bit >>= 2;
1886 	}
1887 
1888 	return res;
1889 }
1890 
1891 static sbintime_t latencies[LAT_BUCKETS - 1] = {
1892 	BUCKET_BASE <<  0,	/* 20us */
1893 	BUCKET_BASE <<  1,
1894 	BUCKET_BASE <<  2,
1895 	BUCKET_BASE <<  3,
1896 	BUCKET_BASE <<  4,
1897 	BUCKET_BASE <<  5,
1898 	BUCKET_BASE <<  6,
1899 	BUCKET_BASE <<  7,
1900 	BUCKET_BASE <<  8,
1901 	BUCKET_BASE <<  9,
1902 	BUCKET_BASE << 10,
1903 	BUCKET_BASE << 11,
1904 	BUCKET_BASE << 12,
1905 	BUCKET_BASE << 13,
1906 	BUCKET_BASE << 14,
1907 	BUCKET_BASE << 15,
1908 	BUCKET_BASE << 16,
1909 	BUCKET_BASE << 17,
1910 	BUCKET_BASE << 18	/* 5,242,880us */
1911 };
1912 
1913 static void
1914 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1915 {
1916 	sbintime_t y, deltasq, delta;
1917 	int i;
1918 
1919 	/*
1920 	 * Keep counts for latency. We do it by power of two buckets.
1921 	 * This helps us spot outlier behavior obscured by averages.
1922 	 */
1923 	for (i = 0; i < LAT_BUCKETS - 1; i++) {
1924 		if (sim_latency < latencies[i]) {
1925 			iop->latencies[i]++;
1926 			break;
1927 		}
1928 	}
1929 	if (i == LAT_BUCKETS - 1)
1930 		iop->latencies[i]++; 	 /* Put all > 8192ms values into the last bucket. */
1931 
1932 	/*
1933 	 * Classic exponentially decaying average with a tiny alpha
1934 	 * (2 ^ -alpha_bits). For more info see the NIST statistical
1935 	 * handbook.
1936 	 *
1937 	 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha)		[nist]
1938 	 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
1939 	 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1940 	 * alpha = 1 / (1 << alpha_bits)
1941 	 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
1942 	 *	= y_t/b - e/b + be/b
1943 	 *      = (y_t - e + be) / b
1944 	 *	= (e + d) / b
1945 	 *
1946 	 * Since alpha is a power of two, we can compute this w/o any mult or
1947 	 * division.
1948 	 *
1949 	 * Variance can also be computed. Usually, it would be expressed as follows:
1950 	 *	diff_t = y_t - ema_t-1
1951 	 *	emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
1952 	 *	  = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
1953 	 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
1954 	 *	  = e - e/b + dd/b + dd/bb
1955 	 *	  = (bbe - be + bdd + dd) / bb
1956 	 *	  = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
1957 	 */
1958 	/*
1959 	 * XXX possible numeric issues
1960 	 *	o We assume right shifted integers do the right thing, since that's
1961 	 *	  implementation defined. You can change the right shifts to / (1LL << alpha).
1962 	 *	o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
1963 	 *	  for emvar. This puts a ceiling of 13 bits on alpha since we need a
1964 	 *	  few tens of seconds of representation.
1965 	 *	o We mitigate alpha issues by never setting it too high.
1966 	 */
1967 	y = sim_latency;
1968 	delta = (y - iop->ema);					/* d */
1969 	iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1970 
1971 	/*
1972 	 * Were we to naively plow ahead at this point, we wind up with many numerical
1973 	 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
1974 	 * us with microsecond level precision in the input, so the same in the
1975 	 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
1976 	 * also means that emvar can be up 46 bits 40 of which are fraction, which
1977 	 * gives us a way to measure up to ~8s in the SD before the computation goes
1978 	 * unstable. Even the worst hard disk rarely has > 1s service time in the
1979 	 * drive. It does mean we have to shift left 12 bits after taking the
1980 	 * square root to compute the actual standard deviation estimate. This loss of
1981 	 * precision is preferable to needing int128 types to work. The above numbers
1982 	 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
1983 	 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1984 	 */
1985 	delta >>= 12;
1986 	deltasq = delta * delta;				/* dd */
1987 	iop->emvar = ((iop->emvar << (2 * alpha_bits)) +	/* bbe */
1988 	    ((deltasq - iop->emvar) << alpha_bits) +		/* b(dd-e) */
1989 	    deltasq)						/* dd */
1990 	    >> (2 * alpha_bits);				/* div bb */
1991 	iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1992 }
1993 
1994 static void
1995 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1996     sbintime_t sim_latency, int cmd, size_t size)
1997 {
1998 	/* xxx Do we need to scale based on the size of the I/O ? */
1999 	switch (cmd) {
2000 	case BIO_READ:
2001 		cam_iosched_update(&isc->read_stats, sim_latency);
2002 		break;
2003 	case BIO_WRITE:
2004 		cam_iosched_update(&isc->write_stats, sim_latency);
2005 		break;
2006 	case BIO_DELETE:
2007 		cam_iosched_update(&isc->trim_stats, sim_latency);
2008 		break;
2009 	default:
2010 		break;
2011 	}
2012 }
2013 
2014 #ifdef DDB
2015 static int biolen(struct bio_queue_head *bq)
2016 {
2017 	int i = 0;
2018 	struct bio *bp;
2019 
2020 	TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
2021 		i++;
2022 	}
2023 	return i;
2024 }
2025 
2026 /*
2027  * Show the internal state of the I/O scheduler.
2028  */
2029 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
2030 {
2031 	struct cam_iosched_softc *isc;
2032 
2033 	if (!have_addr) {
2034 		db_printf("Need addr\n");
2035 		return;
2036 	}
2037 	isc = (struct cam_iosched_softc *)addr;
2038 	db_printf("pending_reads:     %d\n", isc->read_stats.pending);
2039 	db_printf("min_reads:         %d\n", isc->read_stats.min);
2040 	db_printf("max_reads:         %d\n", isc->read_stats.max);
2041 	db_printf("reads:             %d\n", isc->read_stats.total);
2042 	db_printf("in_reads:          %d\n", isc->read_stats.in);
2043 	db_printf("out_reads:         %d\n", isc->read_stats.out);
2044 	db_printf("queued_reads:      %d\n", isc->read_stats.queued);
2045 	db_printf("Read Q len         %d\n", biolen(&isc->bio_queue));
2046 	db_printf("pending_writes:    %d\n", isc->write_stats.pending);
2047 	db_printf("min_writes:        %d\n", isc->write_stats.min);
2048 	db_printf("max_writes:        %d\n", isc->write_stats.max);
2049 	db_printf("writes:            %d\n", isc->write_stats.total);
2050 	db_printf("in_writes:         %d\n", isc->write_stats.in);
2051 	db_printf("out_writes:        %d\n", isc->write_stats.out);
2052 	db_printf("queued_writes:     %d\n", isc->write_stats.queued);
2053 	db_printf("Write Q len        %d\n", biolen(&isc->write_queue));
2054 	db_printf("pending_trims:     %d\n", isc->trim_stats.pending);
2055 	db_printf("min_trims:         %d\n", isc->trim_stats.min);
2056 	db_printf("max_trims:         %d\n", isc->trim_stats.max);
2057 	db_printf("trims:             %d\n", isc->trim_stats.total);
2058 	db_printf("in_trims:          %d\n", isc->trim_stats.in);
2059 	db_printf("out_trims:         %d\n", isc->trim_stats.out);
2060 	db_printf("queued_trims:      %d\n", isc->trim_stats.queued);
2061 	db_printf("Trim Q len         %d\n", biolen(&isc->trim_queue));
2062 	db_printf("read_bias:         %d\n", isc->read_bias);
2063 	db_printf("current_read_bias: %d\n", isc->current_read_bias);
2064 	db_printf("Trim active?       %s\n",
2065 	    (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");
2066 }
2067 #endif
2068 #endif
2069