xref: /freebsd/sys/cam/cam_iosched.c (revision 22cf89c938886d14f5796fc49f9f020c23ea8eaf)
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_cam.h"
31 #include "opt_ddb.h"
32 
33 #include <sys/cdefs.h>
34 #include <sys/param.h>
35 
36 #include <sys/systm.h>
37 #include <sys/kernel.h>
38 #include <sys/bio.h>
39 #include <sys/lock.h>
40 #include <sys/malloc.h>
41 #include <sys/mutex.h>
42 #include <sys/sbuf.h>
43 #include <sys/sysctl.h>
44 
45 #include <cam/cam.h>
46 #include <cam/cam_ccb.h>
47 #include <cam/cam_periph.h>
48 #include <cam/cam_xpt_periph.h>
49 #include <cam/cam_xpt_internal.h>
50 #include <cam/cam_iosched.h>
51 
52 #include <ddb/ddb.h>
53 
54 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler",
55     "CAM I/O Scheduler buffers");
56 
57 static SYSCTL_NODE(_kern_cam, OID_AUTO, iosched, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
58     "CAM I/O Scheduler parameters");
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 bool do_dynamic_iosched = true;
73 SYSCTL_BOOL(_kern_cam_iosched, OID_AUTO, dynamic, CTLFLAG_RDTUN,
74     &do_dynamic_iosched, 1,
75     "Enable Dynamic I/O scheduler optimizations.");
76 
77 /*
78  * For an EMA, with an alpha of alpha, we know
79  * 	alpha = 2 / (N + 1)
80  * or
81  * 	N = 1 + (2 / alpha)
82  * where N is the number of samples that 86% of the current
83  * EMA is derived from.
84  *
85  * So we invent[*] alpha_bits:
86  *	alpha_bits = -log_2(alpha)
87  *	alpha = 2^-alpha_bits
88  * So
89  *	N = 1 + 2^(alpha_bits + 1)
90  *
91  * The default 9 gives a 1025 lookback for 86% of the data.
92  * For a brief intro: https://en.wikipedia.org/wiki/Moving_average
93  *
94  * [*] Steal from the load average code and many other places.
95  * Note: See computation of EMA and EMVAR for acceptable ranges of alpha.
96  */
97 static int alpha_bits = 9;
98 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, alpha_bits, CTLFLAG_RWTUN,
99     &alpha_bits, 1,
100     "Bits in EMA's alpha.");
101 
102 /*
103  * Different parameters for the buckets of latency we keep track of. These are all
104  * published read-only since at present they are compile time constants.
105  *
106  * Bucket base is the upper bounds of the first latency bucket. It's currently 20us.
107  * With 20 buckets (see below), that leads to a geometric progression with a max size
108  * of 5.2s which is safeily larger than 1s to help diagnose extreme outliers better.
109  */
110 #ifndef BUCKET_BASE
111 #define BUCKET_BASE ((SBT_1S / 50000) + 1)	/* 20us */
112 #endif
113 static sbintime_t bucket_base = BUCKET_BASE;
114 SYSCTL_SBINTIME_USEC(_kern_cam_iosched, OID_AUTO, bucket_base_us, CTLFLAG_RD,
115     &bucket_base,
116     "Size of the smallest latency bucket");
117 
118 /*
119  * Bucket ratio is the geometric progression for the bucket. For a bucket b_n
120  * the size of bucket b_n+1 is b_n * bucket_ratio / 100.
121  */
122 static int bucket_ratio = 200;	/* Rather hard coded at the moment */
123 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, bucket_ratio, CTLFLAG_RD,
124     &bucket_ratio, 200,
125     "Latency Bucket Ratio for geometric progression.");
126 
127 /*
128  * Number of total buckets. Starting at BUCKET_BASE, each one is a power of 2.
129  */
130 #ifndef LAT_BUCKETS
131 #define LAT_BUCKETS 20	/* < 20us < 40us ... < 2^(n-1)*20us >= 2^(n-1)*20us */
132 #endif
133 static int lat_buckets = LAT_BUCKETS;
134 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, buckets, CTLFLAG_RD,
135     &lat_buckets, LAT_BUCKETS,
136     "Total number of latency buckets published");
137 
138 /*
139  * Read bias: how many reads do we favor before scheduling a write
140  * when we have a choice.
141  */
142 static int default_read_bias = 0;
143 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, read_bias, CTLFLAG_RWTUN,
144     &default_read_bias, 0,
145     "Default read bias for new devices.");
146 
147 struct iop_stats;
148 struct cam_iosched_softc;
149 
150 int iosched_debug = 0;
151 
152 typedef enum {
153 	none = 0,				/* No limits */
154 	queue_depth,			/* Limit how many ops we queue to SIM */
155 	iops,				/* Limit # of IOPS to the drive */
156 	bandwidth,			/* Limit bandwidth to the drive */
157 	limiter_max
158 } io_limiter;
159 
160 static const char *cam_iosched_limiter_names[] =
161     { "none", "queue_depth", "iops", "bandwidth" };
162 
163 /*
164  * Called to initialize the bits of the iop_stats structure relevant to the
165  * limiter. Called just after the limiter is set.
166  */
167 typedef int l_init_t(struct iop_stats *);
168 
169 /*
170  * Called every tick.
171  */
172 typedef int l_tick_t(struct iop_stats *);
173 
174 /*
175  * Called to see if the limiter thinks this IOP can be allowed to
176  * proceed. If so, the limiter assumes that the IOP proceeded
177  * and makes any accounting of it that's needed.
178  */
179 typedef int l_iop_t(struct iop_stats *, struct bio *);
180 
181 /*
182  * Called when an I/O completes so the limiter can update its
183  * accounting. Pending I/Os may complete in any order (even when
184  * sent to the hardware at the same time), so the limiter may not
185  * make any assumptions other than this I/O has completed. If it
186  * returns 1, then xpt_schedule() needs to be called again.
187  */
188 typedef int l_iodone_t(struct iop_stats *, struct bio *);
189 
190 static l_iop_t cam_iosched_qd_iop;
191 static l_iop_t cam_iosched_qd_caniop;
192 static l_iodone_t cam_iosched_qd_iodone;
193 
194 static l_init_t cam_iosched_iops_init;
195 static l_tick_t cam_iosched_iops_tick;
196 static l_iop_t cam_iosched_iops_caniop;
197 static l_iop_t cam_iosched_iops_iop;
198 
199 static l_init_t cam_iosched_bw_init;
200 static l_tick_t cam_iosched_bw_tick;
201 static l_iop_t cam_iosched_bw_caniop;
202 static l_iop_t cam_iosched_bw_iop;
203 
204 struct limswitch {
205 	l_init_t	*l_init;
206 	l_tick_t	*l_tick;
207 	l_iop_t		*l_iop;
208 	l_iop_t		*l_caniop;
209 	l_iodone_t	*l_iodone;
210 } limsw[] =
211 {
212 	{	/* none */
213 		.l_init = NULL,
214 		.l_tick = NULL,
215 		.l_iop = NULL,
216 		.l_iodone= NULL,
217 	},
218 	{	/* queue_depth */
219 		.l_init = NULL,
220 		.l_tick = NULL,
221 		.l_caniop = cam_iosched_qd_caniop,
222 		.l_iop = cam_iosched_qd_iop,
223 		.l_iodone= cam_iosched_qd_iodone,
224 	},
225 	{	/* iops */
226 		.l_init = cam_iosched_iops_init,
227 		.l_tick = cam_iosched_iops_tick,
228 		.l_caniop = cam_iosched_iops_caniop,
229 		.l_iop = cam_iosched_iops_iop,
230 		.l_iodone= NULL,
231 	},
232 	{	/* bandwidth */
233 		.l_init = cam_iosched_bw_init,
234 		.l_tick = cam_iosched_bw_tick,
235 		.l_caniop = cam_iosched_bw_caniop,
236 		.l_iop = cam_iosched_bw_iop,
237 		.l_iodone= NULL,
238 	},
239 };
240 
241 struct iop_stats {
242 	/*
243 	 * sysctl state for this subnode.
244 	 */
245 	struct sysctl_ctx_list	sysctl_ctx;
246 	struct sysctl_oid	*sysctl_tree;
247 
248 	/*
249 	 * Information about the current rate limiters, if any
250 	 */
251 	io_limiter	limiter;	/* How are I/Os being limited */
252 	int		min;		/* Low range of limit */
253 	int		max;		/* High range of limit */
254 	int		current;	/* Current rate limiter */
255 	int		l_value1;	/* per-limiter scratch value 1. */
256 	int		l_value2;	/* per-limiter scratch value 2. */
257 
258 	/*
259 	 * Debug information about counts of I/Os that have gone through the
260 	 * scheduler.
261 	 */
262 	int		pending;	/* I/Os pending in the hardware */
263 	int		queued;		/* number currently in the queue */
264 	int		total;		/* Total for all time -- wraps */
265 	int		in;		/* number queued all time -- wraps */
266 	int		out;		/* number completed all time -- wraps */
267 	int		errs;		/* Number of I/Os completed with error --  wraps */
268 
269 	/*
270 	 * Statistics on different bits of the process.
271 	 */
272 		/* Exp Moving Average, see alpha_bits for more details */
273 	sbintime_t      ema;
274 	sbintime_t      emvar;
275 	sbintime_t      sd;		/* Last computed sd */
276 
277 	uint32_t	state_flags;
278 #define IOP_RATE_LIMITED		1u
279 
280 	uint64_t	latencies[LAT_BUCKETS];
281 
282 	struct cam_iosched_softc *softc;
283 };
284 
285 typedef enum {
286 	set_max = 0,			/* current = max */
287 	read_latency,			/* Steer read latency by throttling writes */
288 	cl_max				/* Keep last */
289 } control_type;
290 
291 static const char *cam_iosched_control_type_names[] =
292     { "set_max", "read_latency" };
293 
294 struct control_loop {
295 	/*
296 	 * sysctl state for this subnode.
297 	 */
298 	struct sysctl_ctx_list	sysctl_ctx;
299 	struct sysctl_oid	*sysctl_tree;
300 
301 	sbintime_t	next_steer;		/* Time of next steer */
302 	sbintime_t	steer_interval;		/* How often do we steer? */
303 	sbintime_t	lolat;
304 	sbintime_t	hilat;
305 	int		alpha;
306 	control_type	type;			/* What type of control? */
307 	int		last_count;		/* Last I/O count */
308 
309 	struct cam_iosched_softc *softc;
310 };
311 
312 #endif
313 
314 struct cam_iosched_softc {
315 	struct bio_queue_head bio_queue;
316 	struct bio_queue_head trim_queue;
317 				/* scheduler flags < 16, user flags >= 16 */
318 	uint32_t	flags;
319 	int		sort_io_queue;
320 	int		trim_goal;		/* # of trims to queue before sending */
321 	int		trim_ticks;		/* Max ticks to hold trims */
322 	int		last_trim_tick;		/* Last 'tick' time ld a trim */
323 	int		queued_trims;		/* Number of trims in the queue */
324 #ifdef CAM_IOSCHED_DYNAMIC
325 	int		read_bias;		/* Read bias setting */
326 	int		current_read_bias;	/* Current read bias state */
327 	int		total_ticks;
328 	int		load;			/* EMA of 'load average' of disk / 2^16 */
329 
330 	struct bio_queue_head write_queue;
331 	struct iop_stats read_stats, write_stats, trim_stats;
332 	struct sysctl_ctx_list	sysctl_ctx;
333 	struct sysctl_oid	*sysctl_tree;
334 
335 	int		quanta;			/* Number of quanta per second */
336 	struct callout	ticker;			/* Callout for our quota system */
337 	struct cam_periph *periph;		/* cam periph associated with this device */
338 	uint32_t	this_frac;		/* Fraction of a second (1024ths) for this tick */
339 	sbintime_t	last_time;		/* Last time we ticked */
340 	struct control_loop cl;
341 	sbintime_t	max_lat;		/* when != 0, if iop latency > max_lat, call max_lat_fcn */
342 	cam_iosched_latfcn_t	latfcn;
343 	void		*latarg;
344 #endif
345 };
346 
347 #ifdef CAM_IOSCHED_DYNAMIC
348 /*
349  * helper functions to call the limsw functions.
350  */
351 static int
352 cam_iosched_limiter_init(struct iop_stats *ios)
353 {
354 	int lim = ios->limiter;
355 
356 	/* maybe this should be a kassert */
357 	if (lim < none || lim >= limiter_max)
358 		return EINVAL;
359 
360 	if (limsw[lim].l_init)
361 		return limsw[lim].l_init(ios);
362 
363 	return 0;
364 }
365 
366 static int
367 cam_iosched_limiter_tick(struct iop_stats *ios)
368 {
369 	int lim = ios->limiter;
370 
371 	/* maybe this should be a kassert */
372 	if (lim < none || lim >= limiter_max)
373 		return EINVAL;
374 
375 	if (limsw[lim].l_tick)
376 		return limsw[lim].l_tick(ios);
377 
378 	return 0;
379 }
380 
381 static int
382 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
383 {
384 	int lim = ios->limiter;
385 
386 	/* maybe this should be a kassert */
387 	if (lim < none || lim >= limiter_max)
388 		return EINVAL;
389 
390 	if (limsw[lim].l_iop)
391 		return limsw[lim].l_iop(ios, bp);
392 
393 	return 0;
394 }
395 
396 static int
397 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
398 {
399 	int lim = ios->limiter;
400 
401 	/* maybe this should be a kassert */
402 	if (lim < none || lim >= limiter_max)
403 		return EINVAL;
404 
405 	if (limsw[lim].l_caniop)
406 		return limsw[lim].l_caniop(ios, bp);
407 
408 	return 0;
409 }
410 
411 static int
412 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
413 {
414 	int lim = ios->limiter;
415 
416 	/* maybe this should be a kassert */
417 	if (lim < none || lim >= limiter_max)
418 		return 0;
419 
420 	if (limsw[lim].l_iodone)
421 		return limsw[lim].l_iodone(ios, bp);
422 
423 	return 0;
424 }
425 
426 /*
427  * Functions to implement the different kinds of limiters
428  */
429 
430 static int
431 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
432 {
433 
434 	if (ios->current <= 0 || ios->pending < ios->current)
435 		return 0;
436 
437 	return EAGAIN;
438 }
439 
440 static int
441 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
442 {
443 
444 	if (ios->current <= 0 || ios->pending < ios->current)
445 		return 0;
446 
447 	return EAGAIN;
448 }
449 
450 static int
451 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
452 {
453 
454 	if (ios->current <= 0 || ios->pending != ios->current)
455 		return 0;
456 
457 	return 1;
458 }
459 
460 static int
461 cam_iosched_iops_init(struct iop_stats *ios)
462 {
463 
464 	ios->l_value1 = ios->current / ios->softc->quanta;
465 	if (ios->l_value1 <= 0)
466 		ios->l_value1 = 1;
467 	ios->l_value2 = 0;
468 
469 	return 0;
470 }
471 
472 static int
473 cam_iosched_iops_tick(struct iop_stats *ios)
474 {
475 	int new_ios;
476 
477 	/*
478 	 * Allow at least one IO per tick until all
479 	 * the IOs for this interval have been spent.
480 	 */
481 	new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
482 	if (new_ios < 1 && ios->l_value2 < ios->current) {
483 		new_ios = 1;
484 		ios->l_value2++;
485 	}
486 
487 	/*
488 	 * If this a new accounting interval, discard any "unspent" ios
489 	 * granted in the previous interval.  Otherwise add the new ios to
490 	 * the previously granted ones that haven't been spent yet.
491 	 */
492 	if ((ios->softc->total_ticks % ios->softc->quanta) == 0) {
493 		ios->l_value1 = new_ios;
494 		ios->l_value2 = 1;
495 	} else {
496 		ios->l_value1 += new_ios;
497 	}
498 
499 	return 0;
500 }
501 
502 static int
503 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
504 {
505 
506 	/*
507 	 * So if we have any more IOPs left, allow it,
508 	 * otherwise wait. If current iops is 0, treat that
509 	 * as unlimited as a failsafe.
510 	 */
511 	if (ios->current > 0 && ios->l_value1 <= 0)
512 		return EAGAIN;
513 	return 0;
514 }
515 
516 static int
517 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
518 {
519 	int rv;
520 
521 	rv = cam_iosched_limiter_caniop(ios, bp);
522 	if (rv == 0)
523 		ios->l_value1--;
524 
525 	return rv;
526 }
527 
528 static int
529 cam_iosched_bw_init(struct iop_stats *ios)
530 {
531 
532 	/* ios->current is in kB/s, so scale to bytes */
533 	ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
534 
535 	return 0;
536 }
537 
538 static int
539 cam_iosched_bw_tick(struct iop_stats *ios)
540 {
541 	int bw;
542 
543 	/*
544 	 * If we're in the hole for available quota from
545 	 * the last time, then add the quantum for this.
546 	 * If we have any left over from last quantum,
547 	 * then too bad, that's lost. Also, ios->current
548 	 * is in kB/s, so scale.
549 	 *
550 	 * We also allow up to 4 quanta of credits to
551 	 * accumulate to deal with burstiness. 4 is extremely
552 	 * arbitrary.
553 	 */
554 	bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
555 	if (ios->l_value1 < bw * 4)
556 		ios->l_value1 += bw;
557 
558 	return 0;
559 }
560 
561 static int
562 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
563 {
564 	/*
565 	 * So if we have any more bw quota left, allow it,
566 	 * otherwise wait. Note, we'll go negative and that's
567 	 * OK. We'll just get a little less next quota.
568 	 *
569 	 * Note on going negative: that allows us to process
570 	 * requests in order better, since we won't allow
571 	 * shorter reads to get around the long one that we
572 	 * don't have the quota to do just yet. It also prevents
573 	 * starvation by being a little more permissive about
574 	 * what we let through this quantum (to prevent the
575 	 * starvation), at the cost of getting a little less
576 	 * next quantum.
577 	 *
578 	 * Also note that if the current limit is <= 0,
579 	 * we treat it as unlimited as a failsafe.
580 	 */
581 	if (ios->current > 0 && ios->l_value1 <= 0)
582 		return EAGAIN;
583 
584 	return 0;
585 }
586 
587 static int
588 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
589 {
590 	int rv;
591 
592 	rv = cam_iosched_limiter_caniop(ios, bp);
593 	if (rv == 0)
594 		ios->l_value1 -= bp->bio_length;
595 
596 	return rv;
597 }
598 
599 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
600 
601 static void
602 cam_iosched_ticker(void *arg)
603 {
604 	struct cam_iosched_softc *isc = arg;
605 	sbintime_t now, delta;
606 	int pending;
607 
608 	callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
609 
610 	now = sbinuptime();
611 	delta = now - isc->last_time;
612 	isc->this_frac = (uint32_t)delta >> 16;		/* Note: discards seconds -- should be 0 harmless if not */
613 	isc->last_time = now;
614 
615 	cam_iosched_cl_maybe_steer(&isc->cl);
616 
617 	cam_iosched_limiter_tick(&isc->read_stats);
618 	cam_iosched_limiter_tick(&isc->write_stats);
619 	cam_iosched_limiter_tick(&isc->trim_stats);
620 
621 	cam_iosched_schedule(isc, isc->periph);
622 
623 	/*
624 	 * isc->load is an EMA of the pending I/Os at each tick. The number of
625 	 * pending I/Os is the sum of the I/Os queued to the hardware, and those
626 	 * in the software queue that could be queued to the hardware if there
627 	 * were slots.
628 	 *
629 	 * ios_stats.pending is a count of requests in the SIM right now for
630 	 * each of these types of I/O. So the total pending count is the sum of
631 	 * these I/Os and the sum of the queued I/Os still in the software queue
632 	 * for those operations that aren't being rate limited at the moment.
633 	 *
634 	 * The reason for the rate limiting bit is because those I/Os
635 	 * aren't part of the software queued load (since we could
636 	 * give them to hardware, but choose not to).
637 	 *
638 	 * Note: due to a bug in counting pending TRIM in the device, we
639 	 * don't include them in this count. We count each BIO_DELETE in
640 	 * the pending count, but the periph drivers collapse them down
641 	 * into one TRIM command. That one trim command gets the completion
642 	 * so the counts get off.
643 	 */
644 	pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
645 	pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
646 	    !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
647 	    !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
648 	pending <<= 16;
649 	pending /= isc->periph->path->device->ccbq.total_openings;
650 
651 	isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
652 
653 	isc->total_ticks++;
654 }
655 
656 static void
657 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
658 {
659 
660 	clp->next_steer = sbinuptime();
661 	clp->softc = isc;
662 	clp->steer_interval = SBT_1S * 5;	/* Let's start out steering every 5s */
663 	clp->lolat = 5 * SBT_1MS;
664 	clp->hilat = 15 * SBT_1MS;
665 	clp->alpha = 20;			/* Alpha == gain. 20 = .2 */
666 	clp->type = set_max;
667 }
668 
669 static void
670 cam_iosched_cl_maybe_steer(struct control_loop *clp)
671 {
672 	struct cam_iosched_softc *isc;
673 	sbintime_t now, lat;
674 	int old;
675 
676 	isc = clp->softc;
677 	now = isc->last_time;
678 	if (now < clp->next_steer)
679 		return;
680 
681 	clp->next_steer = now + clp->steer_interval;
682 	switch (clp->type) {
683 	case set_max:
684 		if (isc->write_stats.current != isc->write_stats.max)
685 			printf("Steering write from %d kBps to %d kBps\n",
686 			    isc->write_stats.current, isc->write_stats.max);
687 		isc->read_stats.current = isc->read_stats.max;
688 		isc->write_stats.current = isc->write_stats.max;
689 		isc->trim_stats.current = isc->trim_stats.max;
690 		break;
691 	case read_latency:
692 		old = isc->write_stats.current;
693 		lat = isc->read_stats.ema;
694 		/*
695 		 * Simple PLL-like engine. Since we're steering to a range for
696 		 * the SP (set point) that makes things a little more
697 		 * complicated. In addition, we're not directly controlling our
698 		 * PV (process variable), the read latency, but instead are
699 		 * manipulating the write bandwidth limit for our MV
700 		 * (manipulation variable), analysis of this code gets a bit
701 		 * messy. Also, the MV is a very noisy control surface for read
702 		 * latency since it is affected by many hidden processes inside
703 		 * the device which change how responsive read latency will be
704 		 * in reaction to changes in write bandwidth. Unlike the classic
705 		 * boiler control PLL. this may result in over-steering while
706 		 * the SSD takes its time to react to the new, lower load. This
707 		 * is why we use a relatively low alpha of between .1 and .25 to
708 		 * compensate for this effect. At .1, it takes ~22 steering
709 		 * intervals to back off by a factor of 10. At .2 it only takes
710 		 * ~10. At .25 it only takes ~8. However some preliminary data
711 		 * from the SSD drives suggests a reasponse time in 10's of
712 		 * seconds before latency drops regardless of the new write
713 		 * rate. Careful observation will be required to tune this
714 		 * effectively.
715 		 *
716 		 * Also, when there's no read traffic, we jack up the write
717 		 * limit too regardless of the last read latency.  10 is
718 		 * somewhat arbitrary.
719 		 */
720 		if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
721 			isc->write_stats.current = isc->write_stats.current *
722 			    (100 + clp->alpha) / 100;	/* Scale up */
723 		else if (lat > clp->hilat)
724 			isc->write_stats.current = isc->write_stats.current *
725 			    (100 - clp->alpha) / 100;	/* Scale down */
726 		clp->last_count = isc->read_stats.total;
727 
728 		/*
729 		 * Even if we don't steer, per se, enforce the min/max limits as
730 		 * those may have changed.
731 		 */
732 		if (isc->write_stats.current < isc->write_stats.min)
733 			isc->write_stats.current = isc->write_stats.min;
734 		if (isc->write_stats.current > isc->write_stats.max)
735 			isc->write_stats.current = isc->write_stats.max;
736 		if (old != isc->write_stats.current && 	iosched_debug)
737 			printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
738 			    old, isc->write_stats.current,
739 			    (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
740 		break;
741 	case cl_max:
742 		break;
743 	}
744 }
745 #endif
746 
747 /*
748  * Trim or similar currently pending completion. Should only be set for
749  * those drivers wishing only one Trim active at a time.
750  */
751 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE	(1ul << 0)
752 			/* Callout active, and needs to be torn down */
753 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
754 
755 			/* Periph drivers set these flags to indicate work */
756 #define CAM_IOSCHED_FLAG_WORK_FLAGS	((0xffffu) << 16)
757 
758 #ifdef CAM_IOSCHED_DYNAMIC
759 static void
760 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
761     sbintime_t sim_latency, int cmd, size_t size);
762 #endif
763 
764 static inline bool
765 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
766 {
767 	return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
768 }
769 
770 static inline bool
771 cam_iosched_has_io(struct cam_iosched_softc *isc)
772 {
773 #ifdef CAM_IOSCHED_DYNAMIC
774 	if (do_dynamic_iosched) {
775 		struct bio *rbp = bioq_first(&isc->bio_queue);
776 		struct bio *wbp = bioq_first(&isc->write_queue);
777 		bool can_write = wbp != NULL &&
778 		    cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
779 		bool can_read = rbp != NULL &&
780 		    cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
781 		if (iosched_debug > 2) {
782 			printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
783 			printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
784 			printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
785 		}
786 		return can_read || can_write;
787 	}
788 #endif
789 	return bioq_first(&isc->bio_queue) != NULL;
790 }
791 
792 static inline bool
793 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
794 {
795 	struct bio *bp;
796 
797 	bp = bioq_first(&isc->trim_queue);
798 #ifdef CAM_IOSCHED_DYNAMIC
799 	if (do_dynamic_iosched) {
800 		/*
801 		 * If we're limiting trims, then defer action on trims
802 		 * for a bit.
803 		 */
804 		if (bp == NULL || cam_iosched_limiter_caniop(&isc->trim_stats, bp) != 0)
805 			return false;
806 	}
807 #endif
808 
809 	/*
810 	 * If we've set a trim_goal, then if we exceed that allow trims
811 	 * to be passed back to the driver. If we've also set a tick timeout
812 	 * allow trims back to the driver. Otherwise, don't allow trims yet.
813 	 */
814 	if (isc->trim_goal > 0) {
815 		if (isc->queued_trims >= isc->trim_goal)
816 			return true;
817 		if (isc->queued_trims > 0 &&
818 		    isc->trim_ticks > 0 &&
819 		    ticks - isc->last_trim_tick > isc->trim_ticks)
820 			return true;
821 		return false;
822 	}
823 
824 	/* NB: Should perhaps have a max trim active independent of I/O limiters */
825 	return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) && bp != NULL;
826 }
827 
828 #define cam_iosched_sort_queue(isc)	((isc)->sort_io_queue >= 0 ?	\
829     (isc)->sort_io_queue : cam_sort_io_queues)
830 
831 static inline bool
832 cam_iosched_has_work(struct cam_iosched_softc *isc)
833 {
834 #ifdef CAM_IOSCHED_DYNAMIC
835 	if (iosched_debug > 2)
836 		printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
837 		    cam_iosched_has_more_trim(isc),
838 		    cam_iosched_has_flagged_work(isc));
839 #endif
840 
841 	return cam_iosched_has_io(isc) ||
842 		cam_iosched_has_more_trim(isc) ||
843 		cam_iosched_has_flagged_work(isc);
844 }
845 
846 #ifdef CAM_IOSCHED_DYNAMIC
847 static void
848 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
849 {
850 
851 	ios->limiter = none;
852 	ios->in = 0;
853 	ios->max = ios->current = 300000;
854 	ios->min = 1;
855 	ios->out = 0;
856 	ios->errs = 0;
857 	ios->pending = 0;
858 	ios->queued = 0;
859 	ios->total = 0;
860 	ios->ema = 0;
861 	ios->emvar = 0;
862 	ios->softc = isc;
863 	cam_iosched_limiter_init(ios);
864 }
865 
866 static int
867 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
868 {
869 	char buf[16];
870 	struct iop_stats *ios;
871 	struct cam_iosched_softc *isc;
872 	int value, i, error;
873 	const char *p;
874 
875 	ios = arg1;
876 	isc = ios->softc;
877 	value = ios->limiter;
878 	if (value < none || value >= limiter_max)
879 		p = "UNKNOWN";
880 	else
881 		p = cam_iosched_limiter_names[value];
882 
883 	strlcpy(buf, p, sizeof(buf));
884 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
885 	if (error != 0 || req->newptr == NULL)
886 		return error;
887 
888 	cam_periph_lock(isc->periph);
889 
890 	for (i = none; i < limiter_max; i++) {
891 		if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
892 			continue;
893 		ios->limiter = i;
894 		error = cam_iosched_limiter_init(ios);
895 		if (error != 0) {
896 			ios->limiter = value;
897 			cam_periph_unlock(isc->periph);
898 			return error;
899 		}
900 		/* Note: disk load averate requires ticker to be always running */
901 		callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
902 		isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
903 
904 		cam_periph_unlock(isc->periph);
905 		return 0;
906 	}
907 
908 	cam_periph_unlock(isc->periph);
909 	return EINVAL;
910 }
911 
912 static int
913 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
914 {
915 	char buf[16];
916 	struct control_loop *clp;
917 	struct cam_iosched_softc *isc;
918 	int value, i, error;
919 	const char *p;
920 
921 	clp = arg1;
922 	isc = clp->softc;
923 	value = clp->type;
924 	if (value < none || value >= cl_max)
925 		p = "UNKNOWN";
926 	else
927 		p = cam_iosched_control_type_names[value];
928 
929 	strlcpy(buf, p, sizeof(buf));
930 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
931 	if (error != 0 || req->newptr == NULL)
932 		return error;
933 
934 	for (i = set_max; i < cl_max; i++) {
935 		if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
936 			continue;
937 		cam_periph_lock(isc->periph);
938 		clp->type = i;
939 		cam_periph_unlock(isc->periph);
940 		return 0;
941 	}
942 
943 	return EINVAL;
944 }
945 
946 static int
947 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
948 {
949 	char buf[16];
950 	sbintime_t value;
951 	int error;
952 	uint64_t us;
953 
954 	value = *(sbintime_t *)arg1;
955 	us = (uint64_t)value / SBT_1US;
956 	snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
957 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
958 	if (error != 0 || req->newptr == NULL)
959 		return error;
960 	us = strtoul(buf, NULL, 10);
961 	if (us == 0)
962 		return EINVAL;
963 	*(sbintime_t *)arg1 = us * SBT_1US;
964 	return 0;
965 }
966 
967 static int
968 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
969 {
970 	int i, error;
971 	struct sbuf sb;
972 	uint64_t *latencies;
973 
974 	latencies = arg1;
975 	sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
976 
977 	for (i = 0; i < LAT_BUCKETS - 1; i++)
978 		sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
979 	sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
980 	error = sbuf_finish(&sb);
981 	sbuf_delete(&sb);
982 
983 	return (error);
984 }
985 
986 static int
987 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS)
988 {
989 	int *quanta;
990 	int error, value;
991 
992 	quanta = (unsigned *)arg1;
993 	value = *quanta;
994 
995 	error = sysctl_handle_int(oidp, (int *)&value, 0, req);
996 	if ((error != 0) || (req->newptr == NULL))
997 		return (error);
998 
999 	if (value < 1 || value > hz)
1000 		return (EINVAL);
1001 
1002 	*quanta = value;
1003 
1004 	return (0);
1005 }
1006 
1007 static void
1008 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
1009 {
1010 	struct sysctl_oid_list *n;
1011 	struct sysctl_ctx_list *ctx;
1012 
1013 	ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1014 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
1015 	    CTLFLAG_RD | CTLFLAG_MPSAFE, 0, name);
1016 	n = SYSCTL_CHILDREN(ios->sysctl_tree);
1017 	ctx = &ios->sysctl_ctx;
1018 
1019 	SYSCTL_ADD_UQUAD(ctx, n,
1020 	    OID_AUTO, "ema", CTLFLAG_RD,
1021 	    &ios->ema,
1022 	    "Fast Exponentially Weighted Moving Average");
1023 	SYSCTL_ADD_UQUAD(ctx, n,
1024 	    OID_AUTO, "emvar", CTLFLAG_RD,
1025 	    &ios->emvar,
1026 	    "Fast Exponentially Weighted Moving Variance");
1027 
1028 	SYSCTL_ADD_INT(ctx, n,
1029 	    OID_AUTO, "pending", CTLFLAG_RD,
1030 	    &ios->pending, 0,
1031 	    "Instantaneous # of pending transactions");
1032 	SYSCTL_ADD_INT(ctx, n,
1033 	    OID_AUTO, "count", CTLFLAG_RD,
1034 	    &ios->total, 0,
1035 	    "# of transactions submitted to hardware");
1036 	SYSCTL_ADD_INT(ctx, n,
1037 	    OID_AUTO, "queued", CTLFLAG_RD,
1038 	    &ios->queued, 0,
1039 	    "# of transactions in the queue");
1040 	SYSCTL_ADD_INT(ctx, n,
1041 	    OID_AUTO, "in", CTLFLAG_RD,
1042 	    &ios->in, 0,
1043 	    "# of transactions queued to driver");
1044 	SYSCTL_ADD_INT(ctx, n,
1045 	    OID_AUTO, "out", CTLFLAG_RD,
1046 	    &ios->out, 0,
1047 	    "# of transactions completed (including with error)");
1048 	SYSCTL_ADD_INT(ctx, n,
1049 	    OID_AUTO, "errs", CTLFLAG_RD,
1050 	    &ios->errs, 0,
1051 	    "# of transactions completed with an error");
1052 
1053 	SYSCTL_ADD_PROC(ctx, n,
1054 	    OID_AUTO, "limiter",
1055 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1056 	    ios, 0, cam_iosched_limiter_sysctl, "A",
1057 	    "Current limiting type.");
1058 	SYSCTL_ADD_INT(ctx, n,
1059 	    OID_AUTO, "min", CTLFLAG_RW,
1060 	    &ios->min, 0,
1061 	    "min resource");
1062 	SYSCTL_ADD_INT(ctx, n,
1063 	    OID_AUTO, "max", CTLFLAG_RW,
1064 	    &ios->max, 0,
1065 	    "max resource");
1066 	SYSCTL_ADD_INT(ctx, n,
1067 	    OID_AUTO, "current", CTLFLAG_RW,
1068 	    &ios->current, 0,
1069 	    "current resource");
1070 
1071 	SYSCTL_ADD_PROC(ctx, n,
1072 	    OID_AUTO, "latencies",
1073 	    CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
1074 	    &ios->latencies, 0,
1075 	    cam_iosched_sysctl_latencies, "A",
1076 	    "Array of power of 2 latency from 1ms to 1.024s");
1077 }
1078 
1079 static void
1080 cam_iosched_iop_stats_fini(struct iop_stats *ios)
1081 {
1082 	if (ios->sysctl_tree)
1083 		if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
1084 			printf("can't remove iosched sysctl stats context\n");
1085 }
1086 
1087 static void
1088 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
1089 {
1090 	struct sysctl_oid_list *n;
1091 	struct sysctl_ctx_list *ctx;
1092 	struct control_loop *clp;
1093 
1094 	clp = &isc->cl;
1095 	clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1096 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
1097 	    CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Control loop info");
1098 	n = SYSCTL_CHILDREN(clp->sysctl_tree);
1099 	ctx = &clp->sysctl_ctx;
1100 
1101 	SYSCTL_ADD_PROC(ctx, n,
1102 	    OID_AUTO, "type",
1103 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1104 	    clp, 0, cam_iosched_control_type_sysctl, "A",
1105 	    "Control loop algorithm");
1106 	SYSCTL_ADD_PROC(ctx, n,
1107 	    OID_AUTO, "steer_interval",
1108 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1109 	    &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
1110 	    "How often to steer (in us)");
1111 	SYSCTL_ADD_PROC(ctx, n,
1112 	    OID_AUTO, "lolat",
1113 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1114 	    &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
1115 	    "Low water mark for Latency (in us)");
1116 	SYSCTL_ADD_PROC(ctx, n,
1117 	    OID_AUTO, "hilat",
1118 	    CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1119 	    &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
1120 	    "Hi water mark for Latency (in us)");
1121 	SYSCTL_ADD_INT(ctx, n,
1122 	    OID_AUTO, "alpha", CTLFLAG_RW,
1123 	    &clp->alpha, 0,
1124 	    "Alpha for PLL (x100) aka gain");
1125 }
1126 
1127 static void
1128 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1129 {
1130 	if (clp->sysctl_tree)
1131 		if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1132 			printf("can't remove iosched sysctl control loop context\n");
1133 }
1134 #endif
1135 
1136 /*
1137  * Allocate the iosched structure. This also insulates callers from knowing
1138  * sizeof struct cam_iosched_softc.
1139  */
1140 int
1141 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1142 {
1143 
1144 	*iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1145 	if (*iscp == NULL)
1146 		return ENOMEM;
1147 #ifdef CAM_IOSCHED_DYNAMIC
1148 	if (iosched_debug)
1149 		printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1150 #endif
1151 	(*iscp)->sort_io_queue = -1;
1152 	bioq_init(&(*iscp)->bio_queue);
1153 	bioq_init(&(*iscp)->trim_queue);
1154 #ifdef CAM_IOSCHED_DYNAMIC
1155 	if (do_dynamic_iosched) {
1156 		bioq_init(&(*iscp)->write_queue);
1157 		(*iscp)->read_bias = default_read_bias;
1158 		(*iscp)->current_read_bias = 0;
1159 		(*iscp)->quanta = min(hz, 200);
1160 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1161 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1162 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1163 		(*iscp)->trim_stats.max = 1;	/* Trims are special: one at a time for now */
1164 		(*iscp)->last_time = sbinuptime();
1165 		callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1166 		(*iscp)->periph = periph;
1167 		cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1168 		callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp);
1169 		(*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1170 	}
1171 #endif
1172 
1173 	return 0;
1174 }
1175 
1176 /*
1177  * Reclaim all used resources. This assumes that other folks have
1178  * drained the requests in the hardware. Maybe an unwise assumption.
1179  */
1180 void
1181 cam_iosched_fini(struct cam_iosched_softc *isc)
1182 {
1183 	if (isc) {
1184 		cam_iosched_flush(isc, NULL, ENXIO);
1185 #ifdef CAM_IOSCHED_DYNAMIC
1186 		cam_iosched_iop_stats_fini(&isc->read_stats);
1187 		cam_iosched_iop_stats_fini(&isc->write_stats);
1188 		cam_iosched_iop_stats_fini(&isc->trim_stats);
1189 		cam_iosched_cl_sysctl_fini(&isc->cl);
1190 		if (isc->sysctl_tree)
1191 			if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1192 				printf("can't remove iosched sysctl stats context\n");
1193 		if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1194 			callout_drain(&isc->ticker);
1195 			isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1196 		}
1197 #endif
1198 		free(isc, M_CAMSCHED);
1199 	}
1200 }
1201 
1202 /*
1203  * After we're sure we're attaching a device, go ahead and add
1204  * hooks for any sysctl we may wish to honor.
1205  */
1206 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1207     struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1208 {
1209 	struct sysctl_oid_list *n;
1210 
1211 	n = SYSCTL_CHILDREN(node);
1212 	SYSCTL_ADD_INT(ctx, n,
1213 		OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1214 		&isc->sort_io_queue, 0,
1215 		"Sort IO queue to try and optimise disk access patterns");
1216 	SYSCTL_ADD_INT(ctx, n,
1217 	    OID_AUTO, "trim_goal", CTLFLAG_RW,
1218 	    &isc->trim_goal, 0,
1219 	    "Number of trims to try to accumulate before sending to hardware");
1220 	SYSCTL_ADD_INT(ctx, n,
1221 	    OID_AUTO, "trim_ticks", CTLFLAG_RW,
1222 	    &isc->trim_goal, 0,
1223 	    "IO Schedul qaunta to hold back trims for when accumulating");
1224 
1225 #ifdef CAM_IOSCHED_DYNAMIC
1226 	if (!do_dynamic_iosched)
1227 		return;
1228 
1229 	isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1230 	    SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1231 	    CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "I/O scheduler statistics");
1232 	n = SYSCTL_CHILDREN(isc->sysctl_tree);
1233 	ctx = &isc->sysctl_ctx;
1234 
1235 	cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1236 	cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1237 	cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1238 	cam_iosched_cl_sysctl_init(isc);
1239 
1240 	SYSCTL_ADD_INT(ctx, n,
1241 	    OID_AUTO, "read_bias", CTLFLAG_RW,
1242 	    &isc->read_bias, default_read_bias,
1243 	    "How biased towards read should we be independent of limits");
1244 
1245 	SYSCTL_ADD_PROC(ctx, n,
1246 	    OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1247 	    &isc->quanta, 0, cam_iosched_quanta_sysctl, "I",
1248 	    "How many quanta per second do we slice the I/O up into");
1249 
1250 	SYSCTL_ADD_INT(ctx, n,
1251 	    OID_AUTO, "total_ticks", CTLFLAG_RD,
1252 	    &isc->total_ticks, 0,
1253 	    "Total number of ticks we've done");
1254 
1255 	SYSCTL_ADD_INT(ctx, n,
1256 	    OID_AUTO, "load", CTLFLAG_RD,
1257 	    &isc->load, 0,
1258 	    "scaled load average / 100");
1259 
1260 	SYSCTL_ADD_U64(ctx, n,
1261 	    OID_AUTO, "latency_trigger", CTLFLAG_RW,
1262 	    &isc->max_lat, 0,
1263 	    "Latency treshold to trigger callbacks");
1264 #endif
1265 }
1266 
1267 void
1268 cam_iosched_set_latfcn(struct cam_iosched_softc *isc,
1269     cam_iosched_latfcn_t fnp, void *argp)
1270 {
1271 #ifdef CAM_IOSCHED_DYNAMIC
1272 	isc->latfcn = fnp;
1273 	isc->latarg = argp;
1274 #endif
1275 }
1276 
1277 /*
1278  * Client drivers can set two parameters. "goal" is the number of BIO_DELETEs
1279  * that will be queued up before iosched will "release" the trims to the client
1280  * driver to wo with what they will (usually combine as many as possible). If we
1281  * don't get this many, after trim_ticks we'll submit the I/O anyway with
1282  * whatever we have.  We do need an I/O of some kind of to clock the deferred
1283  * trims out to disk. Since we will eventually get a write for the super block
1284  * or something before we shutdown, the trims will complete. To be safe, when a
1285  * BIO_FLUSH is presented to the iosched work queue, we set the ticks time far
1286  * enough in the past so we'll present the BIO_DELETEs to the client driver.
1287  * There might be a race if no BIO_DELETESs were queued, a BIO_FLUSH comes in
1288  * and then a BIO_DELETE is sent down. No know client does this, and there's
1289  * already a race between an ordered BIO_FLUSH and any BIO_DELETEs in flight,
1290  * but no client depends on the ordering being honored.
1291  *
1292  * XXX I'm not sure what the interaction between UFS direct BIOs and the BUF
1293  * flushing on shutdown. I think there's bufs that would be dependent on the BIO
1294  * finishing to write out at least metadata, so we'll be fine. To be safe, keep
1295  * the number of ticks low (less than maybe 10s) to avoid shutdown races.
1296  */
1297 
1298 void
1299 cam_iosched_set_trim_goal(struct cam_iosched_softc *isc, int goal)
1300 {
1301 
1302 	isc->trim_goal = goal;
1303 }
1304 
1305 void
1306 cam_iosched_set_trim_ticks(struct cam_iosched_softc *isc, int trim_ticks)
1307 {
1308 
1309 	isc->trim_ticks = trim_ticks;
1310 }
1311 
1312 /*
1313  * Flush outstanding I/O. Consumers of this library don't know all the
1314  * queues we may keep, so this allows all I/O to be flushed in one
1315  * convenient call.
1316  */
1317 void
1318 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1319 {
1320 	bioq_flush(&isc->bio_queue, stp, err);
1321 	bioq_flush(&isc->trim_queue, stp, err);
1322 #ifdef CAM_IOSCHED_DYNAMIC
1323 	if (do_dynamic_iosched)
1324 		bioq_flush(&isc->write_queue, stp, err);
1325 #endif
1326 }
1327 
1328 #ifdef CAM_IOSCHED_DYNAMIC
1329 static struct bio *
1330 cam_iosched_get_write(struct cam_iosched_softc *isc)
1331 {
1332 	struct bio *bp;
1333 
1334 	/*
1335 	 * We control the write rate by controlling how many requests we send
1336 	 * down to the drive at any one time. Fewer requests limits the
1337 	 * effects of both starvation when the requests take a while and write
1338 	 * amplification when each request is causing more than one write to
1339 	 * the NAND media. Limiting the queue depth like this will also limit
1340 	 * the write throughput and give and reads that want to compete to
1341 	 * compete unfairly.
1342 	 */
1343 	bp = bioq_first(&isc->write_queue);
1344 	if (bp == NULL) {
1345 		if (iosched_debug > 3)
1346 			printf("No writes present in write_queue\n");
1347 		return NULL;
1348 	}
1349 
1350 	/*
1351 	 * If pending read, prefer that based on current read bias
1352 	 * setting.
1353 	 */
1354 	if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1355 		if (iosched_debug)
1356 			printf(
1357 			    "Reads present and current_read_bias is %d queued "
1358 			    "writes %d queued reads %d\n",
1359 			    isc->current_read_bias, isc->write_stats.queued,
1360 			    isc->read_stats.queued);
1361 		isc->current_read_bias--;
1362 		/* We're not limiting writes, per se, just doing reads first */
1363 		return NULL;
1364 	}
1365 
1366 	/*
1367 	 * See if our current limiter allows this I/O.
1368 	 */
1369 	if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1370 		if (iosched_debug)
1371 			printf("Can't write because limiter says no.\n");
1372 		isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1373 		return NULL;
1374 	}
1375 
1376 	/*
1377 	 * Let's do this: We've passed all the gates and we're a go
1378 	 * to schedule the I/O in the SIM.
1379 	 */
1380 	isc->current_read_bias = isc->read_bias;
1381 	bioq_remove(&isc->write_queue, bp);
1382 	if (bp->bio_cmd == BIO_WRITE) {
1383 		isc->write_stats.queued--;
1384 		isc->write_stats.total++;
1385 		isc->write_stats.pending++;
1386 	}
1387 	if (iosched_debug > 9)
1388 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1389 	isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1390 	return bp;
1391 }
1392 #endif
1393 
1394 /*
1395  * Put back a trim that you weren't able to actually schedule this time.
1396  */
1397 void
1398 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1399 {
1400 	bioq_insert_head(&isc->trim_queue, bp);
1401 	if (isc->queued_trims == 0)
1402 		isc->last_trim_tick = ticks;
1403 	isc->queued_trims++;
1404 #ifdef CAM_IOSCHED_DYNAMIC
1405 	isc->trim_stats.queued++;
1406 	isc->trim_stats.total--;		/* since we put it back, don't double count */
1407 	isc->trim_stats.pending--;
1408 #endif
1409 }
1410 
1411 /*
1412  * gets the next trim from the trim queue.
1413  *
1414  * Assumes we're called with the periph lock held.  It removes this
1415  * trim from the queue and the device must explicitly reinsert it
1416  * should the need arise.
1417  */
1418 struct bio *
1419 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1420 {
1421 	struct bio *bp;
1422 
1423 	bp  = bioq_first(&isc->trim_queue);
1424 	if (bp == NULL)
1425 		return NULL;
1426 	bioq_remove(&isc->trim_queue, bp);
1427 	isc->queued_trims--;
1428 	isc->last_trim_tick = ticks;	/* Reset the tick timer when we take trims */
1429 #ifdef CAM_IOSCHED_DYNAMIC
1430 	isc->trim_stats.queued--;
1431 	isc->trim_stats.total++;
1432 	isc->trim_stats.pending++;
1433 #endif
1434 	return bp;
1435 }
1436 
1437 /*
1438  * gets an available trim from the trim queue, if there's no trim
1439  * already pending. It removes this trim from the queue and the device
1440  * must explicitly reinsert it should the need arise.
1441  *
1442  * Assumes we're called with the periph lock held.
1443  */
1444 struct bio *
1445 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1446 {
1447 #ifdef CAM_IOSCHED_DYNAMIC
1448 	struct bio *bp;
1449 #endif
1450 
1451 	if (!cam_iosched_has_more_trim(isc))
1452 		return NULL;
1453 #ifdef CAM_IOSCHED_DYNAMIC
1454 	bp  = bioq_first(&isc->trim_queue);
1455 	if (bp == NULL)
1456 		return NULL;
1457 
1458 	/*
1459 	 * If pending read, prefer that based on current read bias setting. The
1460 	 * read bias is shared for both writes and TRIMs, but on TRIMs the bias
1461 	 * is for a combined TRIM not a single TRIM request that's come in.
1462 	 */
1463 	if (do_dynamic_iosched) {
1464 		if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1465 			if (iosched_debug)
1466 				printf("Reads present and current_read_bias is %d"
1467 				    " queued trims %d queued reads %d\n",
1468 				    isc->current_read_bias, isc->trim_stats.queued,
1469 				    isc->read_stats.queued);
1470 			isc->current_read_bias--;
1471 			/* We're not limiting TRIMS, per se, just doing reads first */
1472 			return NULL;
1473 		}
1474 		/*
1475 		 * We're going to do a trim, so reset the bias.
1476 		 */
1477 		isc->current_read_bias = isc->read_bias;
1478 	}
1479 
1480 	/*
1481 	 * See if our current limiter allows this I/O. Because we only call this
1482 	 * here, and not in next_trim, the 'bandwidth' limits for trims won't
1483 	 * work, while the iops or max queued limits will work. It's tricky
1484 	 * because we want the limits to be from the perspective of the
1485 	 * "commands sent to the device." To make iops work, we need to check
1486 	 * only here (since we want all the ops we combine to count as one). To
1487 	 * make bw limits work, we'd need to check in next_trim, but that would
1488 	 * have the effect of limiting the iops as seen from the upper layers.
1489 	 */
1490 	if (cam_iosched_limiter_iop(&isc->trim_stats, bp) != 0) {
1491 		if (iosched_debug)
1492 			printf("Can't trim because limiter says no.\n");
1493 		isc->trim_stats.state_flags |= IOP_RATE_LIMITED;
1494 		return NULL;
1495 	}
1496 	isc->current_read_bias = isc->read_bias;
1497 	isc->trim_stats.state_flags &= ~IOP_RATE_LIMITED;
1498 	/* cam_iosched_next_trim below keeps proper book */
1499 #endif
1500 	return cam_iosched_next_trim(isc);
1501 }
1502 
1503 
1504 #ifdef CAM_IOSCHED_DYNAMIC
1505 static struct bio *
1506 bio_next(struct bio *bp)
1507 {
1508 	bp = TAILQ_NEXT(bp, bio_queue);
1509 	/*
1510 	 * After the first commands, the ordered bit terminates
1511 	 * our search because BIO_ORDERED acts like a barrier.
1512 	 */
1513 	if (bp == NULL || bp->bio_flags & BIO_ORDERED)
1514 		return NULL;
1515 	return bp;
1516 }
1517 
1518 static bool
1519 cam_iosched_rate_limited(struct iop_stats *ios)
1520 {
1521 	return ios->state_flags & IOP_RATE_LIMITED;
1522 }
1523 #endif
1524 
1525 /*
1526  * Determine what the next bit of work to do is for the periph. The
1527  * default implementation looks to see if we have trims to do, but no
1528  * trims outstanding. If so, we do that. Otherwise we see if we have
1529  * other work. If we do, then we do that. Otherwise why were we called?
1530  */
1531 struct bio *
1532 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1533 {
1534 	struct bio *bp;
1535 
1536 	/*
1537 	 * See if we have a trim that can be scheduled. We can only send one
1538 	 * at a time down, so this takes that into account.
1539 	 *
1540 	 * XXX newer TRIM commands are queueable. Revisit this when we
1541 	 * implement them.
1542 	 */
1543 	if ((bp = cam_iosched_get_trim(isc)) != NULL)
1544 		return bp;
1545 
1546 #ifdef CAM_IOSCHED_DYNAMIC
1547 	/*
1548 	 * See if we have any pending writes, room in the queue for them,
1549 	 * and no pending reads (unless we've scheduled too many).
1550 	 * if so, those are next.
1551 	 */
1552 	if (do_dynamic_iosched) {
1553 		if ((bp = cam_iosched_get_write(isc)) != NULL)
1554 			return bp;
1555 	}
1556 #endif
1557 	/*
1558 	 * next, see if there's other, normal I/O waiting. If so return that.
1559 	 */
1560 #ifdef CAM_IOSCHED_DYNAMIC
1561 	if (do_dynamic_iosched) {
1562 		for (bp = bioq_first(&isc->bio_queue); bp != NULL;
1563 		     bp = bio_next(bp)) {
1564 			/*
1565 			 * For the dynamic scheduler with a read bias, bio_queue
1566 			 * is only for reads. However, without one, all
1567 			 * operations are queued. Enforce limits here for any
1568 			 * operation we find here.
1569 			 */
1570 			if (bp->bio_cmd == BIO_READ) {
1571 				if (cam_iosched_rate_limited(&isc->read_stats) ||
1572 				    cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1573 					isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1574 					continue;
1575 				}
1576 				isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1577 			}
1578 			/*
1579 			 * There can only be write requests on the queue when
1580 			 * the read bias is 0, but we need to process them
1581 			 * here. We do not assert for read bias == 0, however,
1582 			 * since it is dynamic and we can have WRITE operations
1583 			 * in the queue after we transition from 0 to non-zero.
1584 			 */
1585 			if (bp->bio_cmd == BIO_WRITE) {
1586 				if (cam_iosched_rate_limited(&isc->write_stats) ||
1587 				    cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1588 					isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1589 					continue;
1590 				}
1591 				isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1592 			}
1593 			/*
1594 			 * here we know we have a bp that's != NULL, that's not rate limited
1595 			 * and can be the next I/O.
1596 			 */
1597 			break;
1598 		}
1599 	} else
1600 #endif
1601 		bp = bioq_first(&isc->bio_queue);
1602 
1603 	if (bp == NULL)
1604 		return (NULL);
1605 	bioq_remove(&isc->bio_queue, bp);
1606 #ifdef CAM_IOSCHED_DYNAMIC
1607 	if (do_dynamic_iosched) {
1608 		if (bp->bio_cmd == BIO_READ) {
1609 			isc->read_stats.queued--;
1610 			isc->read_stats.total++;
1611 			isc->read_stats.pending++;
1612 		} else if (bp->bio_cmd == BIO_WRITE) {
1613 			isc->write_stats.queued--;
1614 			isc->write_stats.total++;
1615 			isc->write_stats.pending++;
1616 		}
1617 	}
1618 	if (iosched_debug > 9)
1619 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1620 #endif
1621 	return bp;
1622 }
1623 
1624 /*
1625  * Driver has been given some work to do by the block layer. Tell the
1626  * scheduler about it and have it queue the work up. The scheduler module
1627  * will then return the currently most useful bit of work later, possibly
1628  * deferring work for various reasons.
1629  */
1630 void
1631 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1632 {
1633 
1634 	/*
1635 	 * A BIO_SPEEDUP from the upper layers means that they have a block
1636 	 * shortage. At the present, this is only sent when we're trying to
1637 	 * allocate blocks, but have a shortage before giving up. bio_length is
1638 	 * the size of their shortage. We will complete just enough BIO_DELETEs
1639 	 * in the queue to satisfy the need. If bio_length is 0, we'll complete
1640 	 * them all. This allows the scheduler to delay BIO_DELETEs to improve
1641 	 * read/write performance without worrying about the upper layers. When
1642 	 * it's possibly a problem, we respond by pretending the BIO_DELETEs
1643 	 * just worked. We can't do anything about the BIO_DELETEs in the
1644 	 * hardware, though. We have to wait for them to complete.
1645 	 */
1646 	if (bp->bio_cmd == BIO_SPEEDUP) {
1647 		off_t len;
1648 		struct bio *nbp;
1649 
1650 		len = 0;
1651 		while (bioq_first(&isc->trim_queue) &&
1652 		    (bp->bio_length == 0 || len < bp->bio_length)) {
1653 			nbp = bioq_takefirst(&isc->trim_queue);
1654 			len += nbp->bio_length;
1655 			nbp->bio_error = 0;
1656 			biodone(nbp);
1657 		}
1658 		if (bp->bio_length > 0) {
1659 			if (bp->bio_length > len)
1660 				bp->bio_resid = bp->bio_length - len;
1661 			else
1662 				bp->bio_resid = 0;
1663 		}
1664 		bp->bio_error = 0;
1665 		biodone(bp);
1666 		return;
1667 	}
1668 
1669 	/*
1670 	 * If we get a BIO_FLUSH, and we're doing delayed BIO_DELETEs then we
1671 	 * set the last tick time to one less than the current ticks minus the
1672 	 * delay to force the BIO_DELETEs to be presented to the client driver.
1673 	 */
1674 	if (bp->bio_cmd == BIO_FLUSH && isc->trim_ticks > 0)
1675 		isc->last_trim_tick = ticks - isc->trim_ticks - 1;
1676 
1677 	/*
1678 	 * Put all trims on the trim queue. Otherwise put the work on the bio
1679 	 * queue.
1680 	 */
1681 	if (bp->bio_cmd == BIO_DELETE) {
1682 		bioq_insert_tail(&isc->trim_queue, bp);
1683 		if (isc->queued_trims == 0)
1684 			isc->last_trim_tick = ticks;
1685 		isc->queued_trims++;
1686 #ifdef CAM_IOSCHED_DYNAMIC
1687 		isc->trim_stats.in++;
1688 		isc->trim_stats.queued++;
1689 #endif
1690 	}
1691 #ifdef CAM_IOSCHED_DYNAMIC
1692 	else if (do_dynamic_iosched && isc->read_bias != 0 &&
1693 	    (bp->bio_cmd != BIO_READ)) {
1694 		if (cam_iosched_sort_queue(isc))
1695 			bioq_disksort(&isc->write_queue, bp);
1696 		else
1697 			bioq_insert_tail(&isc->write_queue, bp);
1698 		if (iosched_debug > 9)
1699 			printf("Qw  : %p %#x\n", bp, bp->bio_cmd);
1700 		if (bp->bio_cmd == BIO_WRITE) {
1701 			isc->write_stats.in++;
1702 			isc->write_stats.queued++;
1703 		}
1704 	}
1705 #endif
1706 	else {
1707 		if (cam_iosched_sort_queue(isc))
1708 			bioq_disksort(&isc->bio_queue, bp);
1709 		else
1710 			bioq_insert_tail(&isc->bio_queue, bp);
1711 #ifdef CAM_IOSCHED_DYNAMIC
1712 		if (iosched_debug > 9)
1713 			printf("Qr  : %p %#x\n", bp, bp->bio_cmd);
1714 		if (bp->bio_cmd == BIO_READ) {
1715 			isc->read_stats.in++;
1716 			isc->read_stats.queued++;
1717 		} else if (bp->bio_cmd == BIO_WRITE) {
1718 			isc->write_stats.in++;
1719 			isc->write_stats.queued++;
1720 		}
1721 #endif
1722 	}
1723 }
1724 
1725 /*
1726  * If we have work, get it scheduled. Called with the periph lock held.
1727  */
1728 void
1729 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1730 {
1731 
1732 	if (cam_iosched_has_work(isc))
1733 		xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1734 }
1735 
1736 /*
1737  * Complete a trim request. Mark that we no longer have one in flight.
1738  */
1739 void
1740 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1741 {
1742 
1743 	isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1744 }
1745 
1746 /*
1747  * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1748  * might use notes in the ccb for statistics.
1749  */
1750 int
1751 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1752     union ccb *done_ccb)
1753 {
1754 	int retval = 0;
1755 #ifdef CAM_IOSCHED_DYNAMIC
1756 	if (!do_dynamic_iosched)
1757 		return retval;
1758 
1759 	if (iosched_debug > 10)
1760 		printf("done: %p %#x\n", bp, bp->bio_cmd);
1761 	if (bp->bio_cmd == BIO_WRITE) {
1762 		retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1763 		if ((bp->bio_flags & BIO_ERROR) != 0)
1764 			isc->write_stats.errs++;
1765 		isc->write_stats.out++;
1766 		isc->write_stats.pending--;
1767 	} else if (bp->bio_cmd == BIO_READ) {
1768 		retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1769 		if ((bp->bio_flags & BIO_ERROR) != 0)
1770 			isc->read_stats.errs++;
1771 		isc->read_stats.out++;
1772 		isc->read_stats.pending--;
1773 	} else if (bp->bio_cmd == BIO_DELETE) {
1774 		if ((bp->bio_flags & BIO_ERROR) != 0)
1775 			isc->trim_stats.errs++;
1776 		isc->trim_stats.out++;
1777 		isc->trim_stats.pending--;
1778 	} else if (bp->bio_cmd != BIO_FLUSH) {
1779 		if (iosched_debug)
1780 			printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1781 	}
1782 
1783 	if ((bp->bio_flags & BIO_ERROR) == 0 && done_ccb != NULL &&
1784 	    (done_ccb->ccb_h.status & CAM_QOS_VALID) != 0) {
1785 		sbintime_t sim_latency;
1786 
1787 		sim_latency = cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data);
1788 
1789 		cam_iosched_io_metric_update(isc, sim_latency,
1790 		    bp->bio_cmd, bp->bio_bcount);
1791 		/*
1792 		 * Debugging code: allow callbacks to the periph driver when latency max
1793 		 * is exceeded. This can be useful for triggering external debugging actions.
1794 		 */
1795 		if (isc->latfcn && isc->max_lat != 0 && sim_latency > isc->max_lat)
1796 			isc->latfcn(isc->latarg, sim_latency, bp);
1797 	}
1798 
1799 #endif
1800 	return retval;
1801 }
1802 
1803 /*
1804  * Tell the io scheduler that you've pushed a trim down into the sim.
1805  * This also tells the I/O scheduler not to push any more trims down, so
1806  * some periphs do not call it if they can cope with multiple trims in flight.
1807  */
1808 void
1809 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1810 {
1811 
1812 	isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1813 }
1814 
1815 /*
1816  * Change the sorting policy hint for I/O transactions for this device.
1817  */
1818 void
1819 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1820 {
1821 
1822 	isc->sort_io_queue = val;
1823 }
1824 
1825 int
1826 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1827 {
1828 	return isc->flags & flags;
1829 }
1830 
1831 void
1832 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1833 {
1834 	isc->flags |= flags;
1835 }
1836 
1837 void
1838 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1839 {
1840 	isc->flags &= ~flags;
1841 }
1842 
1843 #ifdef CAM_IOSCHED_DYNAMIC
1844 /*
1845  * After the method presented in Jack Crenshaw's 1998 article "Integer
1846  * Square Roots," reprinted at
1847  * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1848  * and well worth the read. Briefly, we find the power of 4 that's the
1849  * largest smaller than val. We then check each smaller power of 4 to
1850  * see if val is still bigger. The right shifts at each step divide
1851  * the result by 2 which after successive application winds up
1852  * accumulating the right answer. It could also have been accumulated
1853  * using a separate root counter, but this code is smaller and faster
1854  * than that method. This method is also integer size invariant.
1855  * It returns floor(sqrt((float)val)), or the largest integer less than
1856  * or equal to the square root.
1857  */
1858 static uint64_t
1859 isqrt64(uint64_t val)
1860 {
1861 	uint64_t res = 0;
1862 	uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1863 
1864 	/*
1865 	 * Find the largest power of 4 smaller than val.
1866 	 */
1867 	while (bit > val)
1868 		bit >>= 2;
1869 
1870 	/*
1871 	 * Accumulate the answer, one bit at a time (we keep moving
1872 	 * them over since 2 is the square root of 4 and we test
1873 	 * powers of 4). We accumulate where we find the bit, but
1874 	 * the successive shifts land the bit in the right place
1875 	 * by the end.
1876 	 */
1877 	while (bit != 0) {
1878 		if (val >= res + bit) {
1879 			val -= res + bit;
1880 			res = (res >> 1) + bit;
1881 		} else
1882 			res >>= 1;
1883 		bit >>= 2;
1884 	}
1885 
1886 	return res;
1887 }
1888 
1889 static sbintime_t latencies[LAT_BUCKETS - 1] = {
1890 	BUCKET_BASE <<  0,	/* 20us */
1891 	BUCKET_BASE <<  1,
1892 	BUCKET_BASE <<  2,
1893 	BUCKET_BASE <<  3,
1894 	BUCKET_BASE <<  4,
1895 	BUCKET_BASE <<  5,
1896 	BUCKET_BASE <<  6,
1897 	BUCKET_BASE <<  7,
1898 	BUCKET_BASE <<  8,
1899 	BUCKET_BASE <<  9,
1900 	BUCKET_BASE << 10,
1901 	BUCKET_BASE << 11,
1902 	BUCKET_BASE << 12,
1903 	BUCKET_BASE << 13,
1904 	BUCKET_BASE << 14,
1905 	BUCKET_BASE << 15,
1906 	BUCKET_BASE << 16,
1907 	BUCKET_BASE << 17,
1908 	BUCKET_BASE << 18	/* 5,242,880us */
1909 };
1910 
1911 static void
1912 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1913 {
1914 	sbintime_t y, deltasq, delta;
1915 	int i;
1916 
1917 	/*
1918 	 * Keep counts for latency. We do it by power of two buckets.
1919 	 * This helps us spot outlier behavior obscured by averages.
1920 	 */
1921 	for (i = 0; i < LAT_BUCKETS - 1; i++) {
1922 		if (sim_latency < latencies[i]) {
1923 			iop->latencies[i]++;
1924 			break;
1925 		}
1926 	}
1927 	if (i == LAT_BUCKETS - 1)
1928 		iop->latencies[i]++; 	 /* Put all > 8192ms values into the last bucket. */
1929 
1930 	/*
1931 	 * Classic exponentially decaying average with a tiny alpha
1932 	 * (2 ^ -alpha_bits). For more info see the NIST statistical
1933 	 * handbook.
1934 	 *
1935 	 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha)		[nist]
1936 	 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
1937 	 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1938 	 * alpha = 1 / (1 << alpha_bits)
1939 	 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
1940 	 *	= y_t/b - e/b + be/b
1941 	 *      = (y_t - e + be) / b
1942 	 *	= (e + d) / b
1943 	 *
1944 	 * Since alpha is a power of two, we can compute this w/o any mult or
1945 	 * division.
1946 	 *
1947 	 * Variance can also be computed. Usually, it would be expressed as follows:
1948 	 *	diff_t = y_t - ema_t-1
1949 	 *	emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
1950 	 *	  = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
1951 	 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
1952 	 *	  = e - e/b + dd/b + dd/bb
1953 	 *	  = (bbe - be + bdd + dd) / bb
1954 	 *	  = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
1955 	 */
1956 	/*
1957 	 * XXX possible numeric issues
1958 	 *	o We assume right shifted integers do the right thing, since that's
1959 	 *	  implementation defined. You can change the right shifts to / (1LL << alpha).
1960 	 *	o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
1961 	 *	  for emvar. This puts a ceiling of 13 bits on alpha since we need a
1962 	 *	  few tens of seconds of representation.
1963 	 *	o We mitigate alpha issues by never setting it too high.
1964 	 */
1965 	y = sim_latency;
1966 	delta = (y - iop->ema);					/* d */
1967 	iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1968 
1969 	/*
1970 	 * Were we to naively plow ahead at this point, we wind up with many numerical
1971 	 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
1972 	 * us with microsecond level precision in the input, so the same in the
1973 	 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
1974 	 * also means that emvar can be up 46 bits 40 of which are fraction, which
1975 	 * gives us a way to measure up to ~8s in the SD before the computation goes
1976 	 * unstable. Even the worst hard disk rarely has > 1s service time in the
1977 	 * drive. It does mean we have to shift left 12 bits after taking the
1978 	 * square root to compute the actual standard deviation estimate. This loss of
1979 	 * precision is preferable to needing int128 types to work. The above numbers
1980 	 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
1981 	 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1982 	 */
1983 	delta >>= 12;
1984 	deltasq = delta * delta;				/* dd */
1985 	iop->emvar = ((iop->emvar << (2 * alpha_bits)) +	/* bbe */
1986 	    ((deltasq - iop->emvar) << alpha_bits) +		/* b(dd-e) */
1987 	    deltasq)						/* dd */
1988 	    >> (2 * alpha_bits);				/* div bb */
1989 	iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1990 }
1991 
1992 static void
1993 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1994     sbintime_t sim_latency, int cmd, size_t size)
1995 {
1996 	/* xxx Do we need to scale based on the size of the I/O ? */
1997 	switch (cmd) {
1998 	case BIO_READ:
1999 		cam_iosched_update(&isc->read_stats, sim_latency);
2000 		break;
2001 	case BIO_WRITE:
2002 		cam_iosched_update(&isc->write_stats, sim_latency);
2003 		break;
2004 	case BIO_DELETE:
2005 		cam_iosched_update(&isc->trim_stats, sim_latency);
2006 		break;
2007 	default:
2008 		break;
2009 	}
2010 }
2011 
2012 #ifdef DDB
2013 static int biolen(struct bio_queue_head *bq)
2014 {
2015 	int i = 0;
2016 	struct bio *bp;
2017 
2018 	TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
2019 		i++;
2020 	}
2021 	return i;
2022 }
2023 
2024 /*
2025  * Show the internal state of the I/O scheduler.
2026  */
2027 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
2028 {
2029 	struct cam_iosched_softc *isc;
2030 
2031 	if (!have_addr) {
2032 		db_printf("Need addr\n");
2033 		return;
2034 	}
2035 	isc = (struct cam_iosched_softc *)addr;
2036 	db_printf("pending_reads:     %d\n", isc->read_stats.pending);
2037 	db_printf("min_reads:         %d\n", isc->read_stats.min);
2038 	db_printf("max_reads:         %d\n", isc->read_stats.max);
2039 	db_printf("reads:             %d\n", isc->read_stats.total);
2040 	db_printf("in_reads:          %d\n", isc->read_stats.in);
2041 	db_printf("out_reads:         %d\n", isc->read_stats.out);
2042 	db_printf("queued_reads:      %d\n", isc->read_stats.queued);
2043 	db_printf("Read Q len         %d\n", biolen(&isc->bio_queue));
2044 	db_printf("pending_writes:    %d\n", isc->write_stats.pending);
2045 	db_printf("min_writes:        %d\n", isc->write_stats.min);
2046 	db_printf("max_writes:        %d\n", isc->write_stats.max);
2047 	db_printf("writes:            %d\n", isc->write_stats.total);
2048 	db_printf("in_writes:         %d\n", isc->write_stats.in);
2049 	db_printf("out_writes:        %d\n", isc->write_stats.out);
2050 	db_printf("queued_writes:     %d\n", isc->write_stats.queued);
2051 	db_printf("Write Q len        %d\n", biolen(&isc->write_queue));
2052 	db_printf("pending_trims:     %d\n", isc->trim_stats.pending);
2053 	db_printf("min_trims:         %d\n", isc->trim_stats.min);
2054 	db_printf("max_trims:         %d\n", isc->trim_stats.max);
2055 	db_printf("trims:             %d\n", isc->trim_stats.total);
2056 	db_printf("in_trims:          %d\n", isc->trim_stats.in);
2057 	db_printf("out_trims:         %d\n", isc->trim_stats.out);
2058 	db_printf("queued_trims:      %d\n", isc->trim_stats.queued);
2059 	db_printf("Trim Q len         %d\n", biolen(&isc->trim_queue));
2060 	db_printf("read_bias:         %d\n", isc->read_bias);
2061 	db_printf("current_read_bias: %d\n", isc->current_read_bias);
2062 	db_printf("Trim active?       %s\n",
2063 	    (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");
2064 }
2065 #endif
2066 #endif
2067