xref: /freebsd/sys/cam/cam_iosched.c (revision 3416500aef140042c64bc149cb1ec6620483bc44)
1 /*-
2  * CAM IO Scheduler Interface
3  *
4  * Copyright (c) 2015 Netflix, Inc.
5  * All rights reserved.
6  *
7  * Redistribution and use in source and binary forms, with or without
8  * modification, are permitted provided that the following conditions
9  * are met:
10  * 1. Redistributions of source code must retain the above copyright
11  *    notice, this list of conditions, and the following disclaimer,
12  *    without modification, immediately at the beginning of the file.
13  * 2. The name of the author may not be used to endorse or promote products
14  *    derived from this software without specific prior written permission.
15  *
16  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
17  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
18  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
19  * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
20  * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
21  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
22  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
23  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
24  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
25  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
26  * SUCH DAMAGE.
27  *
28  * $FreeBSD$
29  */
30 
31 #include "opt_cam.h"
32 #include "opt_ddb.h"
33 
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
36 
37 #include <sys/param.h>
38 
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
41 #include <sys/bio.h>
42 #include <sys/lock.h>
43 #include <sys/malloc.h>
44 #include <sys/mutex.h>
45 #include <sys/sbuf.h>
46 #include <sys/sysctl.h>
47 
48 #include <cam/cam.h>
49 #include <cam/cam_ccb.h>
50 #include <cam/cam_periph.h>
51 #include <cam/cam_xpt_periph.h>
52 #include <cam/cam_xpt_internal.h>
53 #include <cam/cam_iosched.h>
54 
55 #include <ddb/ddb.h>
56 
57 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler",
58     "CAM I/O Scheduler buffers");
59 
60 /*
61  * Default I/O scheduler for FreeBSD. This implementation is just a thin-vineer
62  * over the bioq_* interface, with notions of separate calls for normal I/O and
63  * for trims.
64  *
65  * When CAM_IOSCHED_DYNAMIC is defined, the scheduler is enhanced to dynamically
66  * steer the rate of one type of traffic to help other types of traffic (eg
67  * limit writes when read latency deteriorates on SSDs).
68  */
69 
70 #ifdef CAM_IOSCHED_DYNAMIC
71 
72 static int do_dynamic_iosched = 1;
73 TUNABLE_INT("kern.cam.do_dynamic_iosched", &do_dynamic_iosched);
74 SYSCTL_INT(_kern_cam, OID_AUTO, do_dynamic_iosched, CTLFLAG_RD,
75     &do_dynamic_iosched, 1,
76     "Enable Dynamic I/O scheduler optimizations.");
77 
78 /*
79  * For an EMA, with an alpha of alpha, we know
80  * 	alpha = 2 / (N + 1)
81  * or
82  * 	N = 1 + (2 / alpha)
83  * where N is the number of samples that 86% of the current
84  * EMA is derived from.
85  *
86  * So we invent[*] alpha_bits:
87  *	alpha_bits = -log_2(alpha)
88  *	alpha = 2^-alpha_bits
89  * So
90  *	N = 1 + 2^(alpha_bits + 1)
91  *
92  * The default 9 gives a 1025 lookback for 86% of the data.
93  * For a brief intro: https://en.wikipedia.org/wiki/Moving_average
94  *
95  * [*] Steal from the load average code and many other places.
96  * Note: See computation of EMA and EMVAR for acceptable ranges of alpha.
97  */
98 static int alpha_bits = 9;
99 TUNABLE_INT("kern.cam.iosched_alpha_bits", &alpha_bits);
100 SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW,
101     &alpha_bits, 1,
102     "Bits in EMA's alpha.");
103 
104 struct iop_stats;
105 struct cam_iosched_softc;
106 
107 int iosched_debug = 0;
108 
109 typedef enum {
110 	none = 0,				/* No limits */
111 	queue_depth,			/* Limit how many ops we queue to SIM */
112 	iops,				/* Limit # of IOPS to the drive */
113 	bandwidth,			/* Limit bandwidth to the drive */
114 	limiter_max
115 } io_limiter;
116 
117 static const char *cam_iosched_limiter_names[] =
118     { "none", "queue_depth", "iops", "bandwidth" };
119 
120 /*
121  * Called to initialize the bits of the iop_stats structure relevant to the
122  * limiter. Called just after the limiter is set.
123  */
124 typedef int l_init_t(struct iop_stats *);
125 
126 /*
127  * Called every tick.
128  */
129 typedef int l_tick_t(struct iop_stats *);
130 
131 /*
132  * Called to see if the limiter thinks this IOP can be allowed to
133  * proceed. If so, the limiter assumes that the while IOP proceeded
134  * and makes any accounting of it that's needed.
135  */
136 typedef int l_iop_t(struct iop_stats *, struct bio *);
137 
138 /*
139  * Called when an I/O completes so the limiter can updates its
140  * accounting. Pending I/Os may complete in any order (even when
141  * sent to the hardware at the same time), so the limiter may not
142  * make any assumptions other than this I/O has completed. If it
143  * returns 1, then xpt_schedule() needs to be called again.
144  */
145 typedef int l_iodone_t(struct iop_stats *, struct bio *);
146 
147 static l_iop_t cam_iosched_qd_iop;
148 static l_iop_t cam_iosched_qd_caniop;
149 static l_iodone_t cam_iosched_qd_iodone;
150 
151 static l_init_t cam_iosched_iops_init;
152 static l_tick_t cam_iosched_iops_tick;
153 static l_iop_t cam_iosched_iops_caniop;
154 static l_iop_t cam_iosched_iops_iop;
155 
156 static l_init_t cam_iosched_bw_init;
157 static l_tick_t cam_iosched_bw_tick;
158 static l_iop_t cam_iosched_bw_caniop;
159 static l_iop_t cam_iosched_bw_iop;
160 
161 struct limswitch {
162 	l_init_t	*l_init;
163 	l_tick_t	*l_tick;
164 	l_iop_t		*l_iop;
165 	l_iop_t		*l_caniop;
166 	l_iodone_t	*l_iodone;
167 } limsw[] =
168 {
169 	{	/* none */
170 		.l_init = NULL,
171 		.l_tick = NULL,
172 		.l_iop = NULL,
173 		.l_iodone= NULL,
174 	},
175 	{	/* queue_depth */
176 		.l_init = NULL,
177 		.l_tick = NULL,
178 		.l_caniop = cam_iosched_qd_caniop,
179 		.l_iop = cam_iosched_qd_iop,
180 		.l_iodone= cam_iosched_qd_iodone,
181 	},
182 	{	/* iops */
183 		.l_init = cam_iosched_iops_init,
184 		.l_tick = cam_iosched_iops_tick,
185 		.l_caniop = cam_iosched_iops_caniop,
186 		.l_iop = cam_iosched_iops_iop,
187 		.l_iodone= NULL,
188 	},
189 	{	/* bandwidth */
190 		.l_init = cam_iosched_bw_init,
191 		.l_tick = cam_iosched_bw_tick,
192 		.l_caniop = cam_iosched_bw_caniop,
193 		.l_iop = cam_iosched_bw_iop,
194 		.l_iodone= NULL,
195 	},
196 };
197 
198 struct iop_stats {
199 	/*
200 	 * sysctl state for this subnode.
201 	 */
202 	struct sysctl_ctx_list	sysctl_ctx;
203 	struct sysctl_oid	*sysctl_tree;
204 
205 	/*
206 	 * Information about the current rate limiters, if any
207 	 */
208 	io_limiter	limiter;	/* How are I/Os being limited */
209 	int		min;		/* Low range of limit */
210 	int		max;		/* High range of limit */
211 	int		current;	/* Current rate limiter */
212 	int		l_value1;	/* per-limiter scratch value 1. */
213 	int		l_value2;	/* per-limiter scratch value 2. */
214 
215 	/*
216 	 * Debug information about counts of I/Os that have gone through the
217 	 * scheduler.
218 	 */
219 	int		pending;	/* I/Os pending in the hardware */
220 	int		queued;		/* number currently in the queue */
221 	int		total;		/* Total for all time -- wraps */
222 	int		in;		/* number queued all time -- wraps */
223 	int		out;		/* number completed all time -- wraps */
224 
225 	/*
226 	 * Statistics on different bits of the process.
227 	 */
228 		/* Exp Moving Average, see alpha_bits for more details */
229 	sbintime_t      ema;
230 	sbintime_t      emvar;
231 	sbintime_t      sd;		/* Last computed sd */
232 
233 	uint32_t	state_flags;
234 #define IOP_RATE_LIMITED		1u
235 
236 #define LAT_BUCKETS 12			/* < 1ms < 2ms ... 512ms < 1024ms > 1024ms */
237 	uint64_t	latencies[LAT_BUCKETS];
238 
239 	struct cam_iosched_softc *softc;
240 };
241 
242 
243 typedef enum {
244 	set_max = 0,			/* current = max */
245 	read_latency,			/* Steer read latency by throttling writes */
246 	cl_max				/* Keep last */
247 } control_type;
248 
249 static const char *cam_iosched_control_type_names[] =
250     { "set_max", "read_latency" };
251 
252 struct control_loop {
253 	/*
254 	 * sysctl state for this subnode.
255 	 */
256 	struct sysctl_ctx_list	sysctl_ctx;
257 	struct sysctl_oid	*sysctl_tree;
258 
259 	sbintime_t	next_steer;		/* Time of next steer */
260 	sbintime_t	steer_interval;		/* How often do we steer? */
261 	sbintime_t	lolat;
262 	sbintime_t	hilat;
263 	int		alpha;
264 	control_type	type;			/* What type of control? */
265 	int		last_count;		/* Last I/O count */
266 
267 	struct cam_iosched_softc *softc;
268 };
269 
270 #endif
271 
272 struct cam_iosched_softc {
273 	struct bio_queue_head bio_queue;
274 	struct bio_queue_head trim_queue;
275 				/* scheduler flags < 16, user flags >= 16 */
276 	uint32_t	flags;
277 	int		sort_io_queue;
278 #ifdef CAM_IOSCHED_DYNAMIC
279 	int		read_bias;		/* Read bias setting */
280 	int		current_read_bias;	/* Current read bias state */
281 	int		total_ticks;
282 	int		load;			/* EMA of 'load average' of disk / 2^16 */
283 
284 	struct bio_queue_head write_queue;
285 	struct iop_stats read_stats, write_stats, trim_stats;
286 	struct sysctl_ctx_list	sysctl_ctx;
287 	struct sysctl_oid	*sysctl_tree;
288 
289 	int		quanta;			/* Number of quanta per second */
290 	struct callout	ticker;			/* Callout for our quota system */
291 	struct cam_periph *periph;		/* cam periph associated with this device */
292 	uint32_t	this_frac;		/* Fraction of a second (1024ths) for this tick */
293 	sbintime_t	last_time;		/* Last time we ticked */
294 	struct control_loop cl;
295 #endif
296 };
297 
298 #ifdef CAM_IOSCHED_DYNAMIC
299 /*
300  * helper functions to call the limsw functions.
301  */
302 static int
303 cam_iosched_limiter_init(struct iop_stats *ios)
304 {
305 	int lim = ios->limiter;
306 
307 	/* maybe this should be a kassert */
308 	if (lim < none || lim >= limiter_max)
309 		return EINVAL;
310 
311 	if (limsw[lim].l_init)
312 		return limsw[lim].l_init(ios);
313 
314 	return 0;
315 }
316 
317 static int
318 cam_iosched_limiter_tick(struct iop_stats *ios)
319 {
320 	int lim = ios->limiter;
321 
322 	/* maybe this should be a kassert */
323 	if (lim < none || lim >= limiter_max)
324 		return EINVAL;
325 
326 	if (limsw[lim].l_tick)
327 		return limsw[lim].l_tick(ios);
328 
329 	return 0;
330 }
331 
332 static int
333 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
334 {
335 	int lim = ios->limiter;
336 
337 	/* maybe this should be a kassert */
338 	if (lim < none || lim >= limiter_max)
339 		return EINVAL;
340 
341 	if (limsw[lim].l_iop)
342 		return limsw[lim].l_iop(ios, bp);
343 
344 	return 0;
345 }
346 
347 static int
348 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
349 {
350 	int lim = ios->limiter;
351 
352 	/* maybe this should be a kassert */
353 	if (lim < none || lim >= limiter_max)
354 		return EINVAL;
355 
356 	if (limsw[lim].l_caniop)
357 		return limsw[lim].l_caniop(ios, bp);
358 
359 	return 0;
360 }
361 
362 static int
363 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
364 {
365 	int lim = ios->limiter;
366 
367 	/* maybe this should be a kassert */
368 	if (lim < none || lim >= limiter_max)
369 		return 0;
370 
371 	if (limsw[lim].l_iodone)
372 		return limsw[lim].l_iodone(ios, bp);
373 
374 	return 0;
375 }
376 
377 /*
378  * Functions to implement the different kinds of limiters
379  */
380 
381 static int
382 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
383 {
384 
385 	if (ios->current <= 0 || ios->pending < ios->current)
386 		return 0;
387 
388 	return EAGAIN;
389 }
390 
391 static int
392 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
393 {
394 
395 	if (ios->current <= 0 || ios->pending < ios->current)
396 		return 0;
397 
398 	return EAGAIN;
399 }
400 
401 static int
402 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
403 {
404 
405 	if (ios->current <= 0 || ios->pending != ios->current)
406 		return 0;
407 
408 	return 1;
409 }
410 
411 static int
412 cam_iosched_iops_init(struct iop_stats *ios)
413 {
414 
415 	ios->l_value1 = ios->current / ios->softc->quanta;
416 	if (ios->l_value1 <= 0)
417 		ios->l_value1 = 1;
418 
419 	return 0;
420 }
421 
422 static int
423 cam_iosched_iops_tick(struct iop_stats *ios)
424 {
425 
426 	ios->l_value1 = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
427 	if (ios->l_value1 <= 0)
428 		ios->l_value1 = 1;
429 
430 	return 0;
431 }
432 
433 static int
434 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
435 {
436 
437 	/*
438 	 * So if we have any more IOPs left, allow it,
439 	 * otherwise wait.
440 	 */
441 	if (ios->l_value1 <= 0)
442 		return EAGAIN;
443 	return 0;
444 }
445 
446 static int
447 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
448 {
449 	int rv;
450 
451 	rv = cam_iosched_limiter_caniop(ios, bp);
452 	if (rv == 0)
453 		ios->l_value1--;
454 
455 	return rv;
456 }
457 
458 static int
459 cam_iosched_bw_init(struct iop_stats *ios)
460 {
461 
462 	/* ios->current is in kB/s, so scale to bytes */
463 	ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
464 
465 	return 0;
466 }
467 
468 static int
469 cam_iosched_bw_tick(struct iop_stats *ios)
470 {
471 	int bw;
472 
473 	/*
474 	 * If we're in the hole for available quota from
475 	 * the last time, then add the quantum for this.
476 	 * If we have any left over from last quantum,
477 	 * then too bad, that's lost. Also, ios->current
478 	 * is in kB/s, so scale.
479 	 *
480 	 * We also allow up to 4 quanta of credits to
481 	 * accumulate to deal with burstiness. 4 is extremely
482 	 * arbitrary.
483 	 */
484 	bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
485 	if (ios->l_value1 < bw * 4)
486 		ios->l_value1 += bw;
487 
488 	return 0;
489 }
490 
491 static int
492 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
493 {
494 	/*
495 	 * So if we have any more bw quota left, allow it,
496 	 * otherwise wait. Not, we'll go negative and that's
497 	 * OK. We'll just get a lettle less next quota.
498 	 *
499 	 * Note on going negative: that allows us to process
500 	 * requests in order better, since we won't allow
501 	 * shorter reads to get around the long one that we
502 	 * don't have the quota to do just yet. It also prevents
503 	 * starvation by being a little more permissive about
504 	 * what we let through this quantum (to prevent the
505 	 * starvation), at the cost of getting a little less
506 	 * next quantum.
507 	 */
508 	if (ios->l_value1 <= 0)
509 		return EAGAIN;
510 
511 
512 	return 0;
513 }
514 
515 static int
516 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
517 {
518 	int rv;
519 
520 	rv = cam_iosched_limiter_caniop(ios, bp);
521 	if (rv == 0)
522 		ios->l_value1 -= bp->bio_length;
523 
524 	return rv;
525 }
526 
527 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
528 
529 static void
530 cam_iosched_ticker(void *arg)
531 {
532 	struct cam_iosched_softc *isc = arg;
533 	sbintime_t now, delta;
534 	int pending;
535 
536 	callout_reset(&isc->ticker, hz / isc->quanta - 1, cam_iosched_ticker, isc);
537 
538 	now = sbinuptime();
539 	delta = now - isc->last_time;
540 	isc->this_frac = (uint32_t)delta >> 16;		/* Note: discards seconds -- should be 0 harmless if not */
541 	isc->last_time = now;
542 
543 	cam_iosched_cl_maybe_steer(&isc->cl);
544 
545 	cam_iosched_limiter_tick(&isc->read_stats);
546 	cam_iosched_limiter_tick(&isc->write_stats);
547 	cam_iosched_limiter_tick(&isc->trim_stats);
548 
549 	cam_iosched_schedule(isc, isc->periph);
550 
551 	/*
552 	 * isc->load is an EMA of the pending I/Os at each tick. The number of
553 	 * pending I/Os is the sum of the I/Os queued to the hardware, and those
554 	 * in the software queue that could be queued to the hardware if there
555 	 * were slots.
556 	 *
557 	 * ios_stats.pending is a count of requests in the SIM right now for
558 	 * each of these types of I/O. So the total pending count is the sum of
559 	 * these I/Os and the sum of the queued I/Os still in the software queue
560 	 * for those operations that aren't being rate limited at the moment.
561 	 *
562 	 * The reason for the rate limiting bit is because those I/Os
563 	 * aren't part of the software queued load (since we could
564 	 * give them to hardware, but choose not to).
565 	 *
566 	 * Note: due to a bug in counting pending TRIM in the device, we
567 	 * don't include them in this count. We count each BIO_DELETE in
568 	 * the pending count, but the periph drivers collapse them down
569 	 * into one TRIM command. That one trim command gets the completion
570 	 * so the counts get off.
571 	 */
572 	pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
573 	pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
574 	    !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
575 	    !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
576 	pending <<= 16;
577 	pending /= isc->periph->path->device->ccbq.total_openings;
578 
579 	isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
580 
581 	isc->total_ticks++;
582 }
583 
584 
585 static void
586 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
587 {
588 
589 	clp->next_steer = sbinuptime();
590 	clp->softc = isc;
591 	clp->steer_interval = SBT_1S * 5;	/* Let's start out steering every 5s */
592 	clp->lolat = 5 * SBT_1MS;
593 	clp->hilat = 15 * SBT_1MS;
594 	clp->alpha = 20;			/* Alpha == gain. 20 = .2 */
595 	clp->type = set_max;
596 }
597 
598 static void
599 cam_iosched_cl_maybe_steer(struct control_loop *clp)
600 {
601 	struct cam_iosched_softc *isc;
602 	sbintime_t now, lat;
603 	int old;
604 
605 	isc = clp->softc;
606 	now = isc->last_time;
607 	if (now < clp->next_steer)
608 		return;
609 
610 	clp->next_steer = now + clp->steer_interval;
611 	switch (clp->type) {
612 	case set_max:
613 		if (isc->write_stats.current != isc->write_stats.max)
614 			printf("Steering write from %d kBps to %d kBps\n",
615 			    isc->write_stats.current, isc->write_stats.max);
616 		isc->read_stats.current = isc->read_stats.max;
617 		isc->write_stats.current = isc->write_stats.max;
618 		isc->trim_stats.current = isc->trim_stats.max;
619 		break;
620 	case read_latency:
621 		old = isc->write_stats.current;
622 		lat = isc->read_stats.ema;
623 		/*
624 		 * Simple PLL-like engine. Since we're steering to a range for
625 		 * the SP (set point) that makes things a little more
626 		 * complicated. In addition, we're not directly controlling our
627 		 * PV (process variable), the read latency, but instead are
628 		 * manipulating the write bandwidth limit for our MV
629 		 * (manipulation variable), analysis of this code gets a bit
630 		 * messy. Also, the MV is a very noisy control surface for read
631 		 * latency since it is affected by many hidden processes inside
632 		 * the device which change how responsive read latency will be
633 		 * in reaction to changes in write bandwidth. Unlike the classic
634 		 * boiler control PLL. this may result in over-steering while
635 		 * the SSD takes its time to react to the new, lower load. This
636 		 * is why we use a relatively low alpha of between .1 and .25 to
637 		 * compensate for this effect. At .1, it takes ~22 steering
638 		 * intervals to back off by a factor of 10. At .2 it only takes
639 		 * ~10. At .25 it only takes ~8. However some preliminary data
640 		 * from the SSD drives suggests a reasponse time in 10's of
641 		 * seconds before latency drops regardless of the new write
642 		 * rate. Careful observation will be reqiured to tune this
643 		 * effectively.
644 		 *
645 		 * Also, when there's no read traffic, we jack up the write
646 		 * limit too regardless of the last read latency.  10 is
647 		 * somewhat arbitrary.
648 		 */
649 		if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
650 			isc->write_stats.current = isc->write_stats.current *
651 			    (100 + clp->alpha) / 100;	/* Scale up */
652 		else if (lat > clp->hilat)
653 			isc->write_stats.current = isc->write_stats.current *
654 			    (100 - clp->alpha) / 100;	/* Scale down */
655 		clp->last_count = isc->read_stats.total;
656 
657 		/*
658 		 * Even if we don't steer, per se, enforce the min/max limits as
659 		 * those may have changed.
660 		 */
661 		if (isc->write_stats.current < isc->write_stats.min)
662 			isc->write_stats.current = isc->write_stats.min;
663 		if (isc->write_stats.current > isc->write_stats.max)
664 			isc->write_stats.current = isc->write_stats.max;
665 		if (old != isc->write_stats.current && 	iosched_debug)
666 			printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
667 			    old, isc->write_stats.current,
668 			    (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
669 		break;
670 	case cl_max:
671 		break;
672 	}
673 }
674 #endif
675 
676 			/* Trim or similar currently pending completion */
677 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE	(1ul << 0)
678 			/* Callout active, and needs to be torn down */
679 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
680 
681 			/* Periph drivers set these flags to indicate work */
682 #define CAM_IOSCHED_FLAG_WORK_FLAGS	((0xffffu) << 16)
683 
684 #ifdef CAM_IOSCHED_DYNAMIC
685 static void
686 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
687     sbintime_t sim_latency, int cmd, size_t size);
688 #endif
689 
690 static inline int
691 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
692 {
693 	return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
694 }
695 
696 static inline int
697 cam_iosched_has_io(struct cam_iosched_softc *isc)
698 {
699 #ifdef CAM_IOSCHED_DYNAMIC
700 	if (do_dynamic_iosched) {
701 		struct bio *rbp = bioq_first(&isc->bio_queue);
702 		struct bio *wbp = bioq_first(&isc->write_queue);
703 		int can_write = wbp != NULL &&
704 		    cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
705 		int can_read = rbp != NULL &&
706 		    cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
707 		if (iosched_debug > 2) {
708 			printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
709 			printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
710 			printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
711 		}
712 		return can_read || can_write;
713 	}
714 #endif
715 	return bioq_first(&isc->bio_queue) != NULL;
716 }
717 
718 static inline int
719 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
720 {
721 	return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) &&
722 	    bioq_first(&isc->trim_queue);
723 }
724 
725 #define cam_iosched_sort_queue(isc)	((isc)->sort_io_queue >= 0 ?	\
726     (isc)->sort_io_queue : cam_sort_io_queues)
727 
728 
729 static inline int
730 cam_iosched_has_work(struct cam_iosched_softc *isc)
731 {
732 #ifdef CAM_IOSCHED_DYNAMIC
733 	if (iosched_debug > 2)
734 		printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
735 		    cam_iosched_has_more_trim(isc),
736 		    cam_iosched_has_flagged_work(isc));
737 #endif
738 
739 	return cam_iosched_has_io(isc) ||
740 		cam_iosched_has_more_trim(isc) ||
741 		cam_iosched_has_flagged_work(isc);
742 }
743 
744 #ifdef CAM_IOSCHED_DYNAMIC
745 static void
746 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
747 {
748 
749 	ios->limiter = none;
750 	cam_iosched_limiter_init(ios);
751 	ios->in = 0;
752 	ios->max = 300000;
753 	ios->min = 1;
754 	ios->out = 0;
755 	ios->pending = 0;
756 	ios->queued = 0;
757 	ios->total = 0;
758 	ios->ema = 0;
759 	ios->emvar = 0;
760 	ios->softc = isc;
761 }
762 
763 static int
764 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
765 {
766 	char buf[16];
767 	struct iop_stats *ios;
768 	struct cam_iosched_softc *isc;
769 	int value, i, error;
770 	const char *p;
771 
772 	ios = arg1;
773 	isc = ios->softc;
774 	value = ios->limiter;
775 	if (value < none || value >= limiter_max)
776 		p = "UNKNOWN";
777 	else
778 		p = cam_iosched_limiter_names[value];
779 
780 	strlcpy(buf, p, sizeof(buf));
781 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
782 	if (error != 0 || req->newptr == NULL)
783 		return error;
784 
785 	cam_periph_lock(isc->periph);
786 
787 	for (i = none; i < limiter_max; i++) {
788 		if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
789 			continue;
790 		ios->limiter = i;
791 		error = cam_iosched_limiter_init(ios);
792 		if (error != 0) {
793 			ios->limiter = value;
794 			cam_periph_unlock(isc->periph);
795 			return error;
796 		}
797 		/* Note: disk load averate requires ticker to be always running */
798 		callout_reset(&isc->ticker, hz / isc->quanta - 1, cam_iosched_ticker, isc);
799 		isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
800 
801 		cam_periph_unlock(isc->periph);
802 		return 0;
803 	}
804 
805 	cam_periph_unlock(isc->periph);
806 	return EINVAL;
807 }
808 
809 static int
810 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
811 {
812 	char buf[16];
813 	struct control_loop *clp;
814 	struct cam_iosched_softc *isc;
815 	int value, i, error;
816 	const char *p;
817 
818 	clp = arg1;
819 	isc = clp->softc;
820 	value = clp->type;
821 	if (value < none || value >= cl_max)
822 		p = "UNKNOWN";
823 	else
824 		p = cam_iosched_control_type_names[value];
825 
826 	strlcpy(buf, p, sizeof(buf));
827 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
828 	if (error != 0 || req->newptr == NULL)
829 		return error;
830 
831 	for (i = set_max; i < cl_max; i++) {
832 		if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
833 			continue;
834 		cam_periph_lock(isc->periph);
835 		clp->type = i;
836 		cam_periph_unlock(isc->periph);
837 		return 0;
838 	}
839 
840 	return EINVAL;
841 }
842 
843 static int
844 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
845 {
846 	char buf[16];
847 	sbintime_t value;
848 	int error;
849 	uint64_t us;
850 
851 	value = *(sbintime_t *)arg1;
852 	us = (uint64_t)value / SBT_1US;
853 	snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
854 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
855 	if (error != 0 || req->newptr == NULL)
856 		return error;
857 	us = strtoul(buf, NULL, 10);
858 	if (us == 0)
859 		return EINVAL;
860 	*(sbintime_t *)arg1 = us * SBT_1US;
861 	return 0;
862 }
863 
864 static int
865 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
866 {
867 	int i, error;
868 	struct sbuf sb;
869 	uint64_t *latencies;
870 
871 	latencies = arg1;
872 	sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
873 
874 	for (i = 0; i < LAT_BUCKETS - 1; i++)
875 		sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
876 	sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
877 	error = sbuf_finish(&sb);
878 	sbuf_delete(&sb);
879 
880 	return (error);
881 }
882 
883 static void
884 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
885 {
886 	struct sysctl_oid_list *n;
887 	struct sysctl_ctx_list *ctx;
888 
889 	ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
890 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
891 	    CTLFLAG_RD, 0, name);
892 	n = SYSCTL_CHILDREN(ios->sysctl_tree);
893 	ctx = &ios->sysctl_ctx;
894 
895 	SYSCTL_ADD_UQUAD(ctx, n,
896 	    OID_AUTO, "ema", CTLFLAG_RD,
897 	    &ios->ema,
898 	    "Fast Exponentially Weighted Moving Average");
899 	SYSCTL_ADD_UQUAD(ctx, n,
900 	    OID_AUTO, "emvar", CTLFLAG_RD,
901 	    &ios->emvar,
902 	    "Fast Exponentially Weighted Moving Variance");
903 
904 	SYSCTL_ADD_INT(ctx, n,
905 	    OID_AUTO, "pending", CTLFLAG_RD,
906 	    &ios->pending, 0,
907 	    "Instantaneous # of pending transactions");
908 	SYSCTL_ADD_INT(ctx, n,
909 	    OID_AUTO, "count", CTLFLAG_RD,
910 	    &ios->total, 0,
911 	    "# of transactions submitted to hardware");
912 	SYSCTL_ADD_INT(ctx, n,
913 	    OID_AUTO, "queued", CTLFLAG_RD,
914 	    &ios->queued, 0,
915 	    "# of transactions in the queue");
916 	SYSCTL_ADD_INT(ctx, n,
917 	    OID_AUTO, "in", CTLFLAG_RD,
918 	    &ios->in, 0,
919 	    "# of transactions queued to driver");
920 	SYSCTL_ADD_INT(ctx, n,
921 	    OID_AUTO, "out", CTLFLAG_RD,
922 	    &ios->out, 0,
923 	    "# of transactions completed");
924 
925 	SYSCTL_ADD_PROC(ctx, n,
926 	    OID_AUTO, "limiter", CTLTYPE_STRING | CTLFLAG_RW,
927 	    ios, 0, cam_iosched_limiter_sysctl, "A",
928 	    "Current limiting type.");
929 	SYSCTL_ADD_INT(ctx, n,
930 	    OID_AUTO, "min", CTLFLAG_RW,
931 	    &ios->min, 0,
932 	    "min resource");
933 	SYSCTL_ADD_INT(ctx, n,
934 	    OID_AUTO, "max", CTLFLAG_RW,
935 	    &ios->max, 0,
936 	    "max resource");
937 	SYSCTL_ADD_INT(ctx, n,
938 	    OID_AUTO, "current", CTLFLAG_RW,
939 	    &ios->current, 0,
940 	    "current resource");
941 
942 	SYSCTL_ADD_PROC(ctx, n,
943 	    OID_AUTO, "latencies", CTLTYPE_STRING | CTLFLAG_RD,
944 	    &ios->latencies, 0,
945 	    cam_iosched_sysctl_latencies, "A",
946 	    "Array of power of 2 latency from 1ms to 1.024s");
947 }
948 
949 static void
950 cam_iosched_iop_stats_fini(struct iop_stats *ios)
951 {
952 	if (ios->sysctl_tree)
953 		if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
954 			printf("can't remove iosched sysctl stats context\n");
955 }
956 
957 static void
958 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
959 {
960 	struct sysctl_oid_list *n;
961 	struct sysctl_ctx_list *ctx;
962 	struct control_loop *clp;
963 
964 	clp = &isc->cl;
965 	clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
966 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
967 	    CTLFLAG_RD, 0, "Control loop info");
968 	n = SYSCTL_CHILDREN(clp->sysctl_tree);
969 	ctx = &clp->sysctl_ctx;
970 
971 	SYSCTL_ADD_PROC(ctx, n,
972 	    OID_AUTO, "type", CTLTYPE_STRING | CTLFLAG_RW,
973 	    clp, 0, cam_iosched_control_type_sysctl, "A",
974 	    "Control loop algorithm");
975 	SYSCTL_ADD_PROC(ctx, n,
976 	    OID_AUTO, "steer_interval", CTLTYPE_STRING | CTLFLAG_RW,
977 	    &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
978 	    "How often to steer (in us)");
979 	SYSCTL_ADD_PROC(ctx, n,
980 	    OID_AUTO, "lolat", CTLTYPE_STRING | CTLFLAG_RW,
981 	    &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
982 	    "Low water mark for Latency (in us)");
983 	SYSCTL_ADD_PROC(ctx, n,
984 	    OID_AUTO, "hilat", CTLTYPE_STRING | CTLFLAG_RW,
985 	    &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
986 	    "Hi water mark for Latency (in us)");
987 	SYSCTL_ADD_INT(ctx, n,
988 	    OID_AUTO, "alpha", CTLFLAG_RW,
989 	    &clp->alpha, 0,
990 	    "Alpha for PLL (x100) aka gain");
991 }
992 
993 static void
994 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
995 {
996 	if (clp->sysctl_tree)
997 		if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
998 			printf("can't remove iosched sysctl control loop context\n");
999 }
1000 #endif
1001 
1002 /*
1003  * Allocate the iosched structure. This also insulates callers from knowing
1004  * sizeof struct cam_iosched_softc.
1005  */
1006 int
1007 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1008 {
1009 
1010 	*iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1011 	if (*iscp == NULL)
1012 		return ENOMEM;
1013 #ifdef CAM_IOSCHED_DYNAMIC
1014 	if (iosched_debug)
1015 		printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1016 #endif
1017 	(*iscp)->sort_io_queue = -1;
1018 	bioq_init(&(*iscp)->bio_queue);
1019 	bioq_init(&(*iscp)->trim_queue);
1020 #ifdef CAM_IOSCHED_DYNAMIC
1021 	if (do_dynamic_iosched) {
1022 		bioq_init(&(*iscp)->write_queue);
1023 		(*iscp)->read_bias = 100;
1024 		(*iscp)->current_read_bias = 100;
1025 		(*iscp)->quanta = 200;
1026 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1027 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1028 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1029 		(*iscp)->trim_stats.max = 1;	/* Trims are special: one at a time for now */
1030 		(*iscp)->last_time = sbinuptime();
1031 		callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1032 		(*iscp)->periph = periph;
1033 		cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1034 		callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta - 1, cam_iosched_ticker, *iscp);
1035 		(*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1036 	}
1037 #endif
1038 
1039 	return 0;
1040 }
1041 
1042 /*
1043  * Reclaim all used resources. This assumes that other folks have
1044  * drained the requests in the hardware. Maybe an unwise assumption.
1045  */
1046 void
1047 cam_iosched_fini(struct cam_iosched_softc *isc)
1048 {
1049 	if (isc) {
1050 		cam_iosched_flush(isc, NULL, ENXIO);
1051 #ifdef CAM_IOSCHED_DYNAMIC
1052 		cam_iosched_iop_stats_fini(&isc->read_stats);
1053 		cam_iosched_iop_stats_fini(&isc->write_stats);
1054 		cam_iosched_iop_stats_fini(&isc->trim_stats);
1055 		cam_iosched_cl_sysctl_fini(&isc->cl);
1056 		if (isc->sysctl_tree)
1057 			if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1058 				printf("can't remove iosched sysctl stats context\n");
1059 		if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1060 			callout_drain(&isc->ticker);
1061 			isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1062 		}
1063 #endif
1064 		free(isc, M_CAMSCHED);
1065 	}
1066 }
1067 
1068 /*
1069  * After we're sure we're attaching a device, go ahead and add
1070  * hooks for any sysctl we may wish to honor.
1071  */
1072 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1073     struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1074 {
1075 #ifdef CAM_IOSCHED_DYNAMIC
1076 	struct sysctl_oid_list *n;
1077 #endif
1078 
1079 	SYSCTL_ADD_INT(ctx, SYSCTL_CHILDREN(node),
1080 		OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1081 		&isc->sort_io_queue, 0,
1082 		"Sort IO queue to try and optimise disk access patterns");
1083 
1084 #ifdef CAM_IOSCHED_DYNAMIC
1085 	if (!do_dynamic_iosched)
1086 		return;
1087 
1088 	isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1089 	    SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1090 	    CTLFLAG_RD, 0, "I/O scheduler statistics");
1091 	n = SYSCTL_CHILDREN(isc->sysctl_tree);
1092 	ctx = &isc->sysctl_ctx;
1093 
1094 	cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1095 	cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1096 	cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1097 	cam_iosched_cl_sysctl_init(isc);
1098 
1099 	SYSCTL_ADD_INT(ctx, n,
1100 	    OID_AUTO, "read_bias", CTLFLAG_RW,
1101 	    &isc->read_bias, 100,
1102 	    "How biased towards read should we be independent of limits");
1103 
1104 	SYSCTL_ADD_INT(ctx, n,
1105 	    OID_AUTO, "quanta", CTLFLAG_RW,
1106 	    &isc->quanta, 200,
1107 	    "How many quanta per second do we slice the I/O up into");
1108 
1109 	SYSCTL_ADD_INT(ctx, n,
1110 	    OID_AUTO, "total_ticks", CTLFLAG_RD,
1111 	    &isc->total_ticks, 0,
1112 	    "Total number of ticks we've done");
1113 
1114 	SYSCTL_ADD_INT(ctx, n,
1115 	    OID_AUTO, "load", CTLFLAG_RD,
1116 	    &isc->load, 0,
1117 	    "scaled load average / 100");
1118 #endif
1119 }
1120 
1121 /*
1122  * Flush outstanding I/O. Consumers of this library don't know all the
1123  * queues we may keep, so this allows all I/O to be flushed in one
1124  * convenient call.
1125  */
1126 void
1127 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1128 {
1129 	bioq_flush(&isc->bio_queue, stp, err);
1130 	bioq_flush(&isc->trim_queue, stp, err);
1131 #ifdef CAM_IOSCHED_DYNAMIC
1132 	if (do_dynamic_iosched)
1133 		bioq_flush(&isc->write_queue, stp, err);
1134 #endif
1135 }
1136 
1137 #ifdef CAM_IOSCHED_DYNAMIC
1138 static struct bio *
1139 cam_iosched_get_write(struct cam_iosched_softc *isc)
1140 {
1141 	struct bio *bp;
1142 
1143 	/*
1144 	 * We control the write rate by controlling how many requests we send
1145 	 * down to the drive at any one time. Fewer requests limits the
1146 	 * effects of both starvation when the requests take a while and write
1147 	 * amplification when each request is causing more than one write to
1148 	 * the NAND media. Limiting the queue depth like this will also limit
1149 	 * the write throughput and give and reads that want to compete to
1150 	 * compete unfairly.
1151 	 */
1152 	bp = bioq_first(&isc->write_queue);
1153 	if (bp == NULL) {
1154 		if (iosched_debug > 3)
1155 			printf("No writes present in write_queue\n");
1156 		return NULL;
1157 	}
1158 
1159 	/*
1160 	 * If pending read, prefer that based on current read bias
1161 	 * setting.
1162 	 */
1163 	if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1164 		if (iosched_debug)
1165 			printf("Reads present and current_read_bias is %d queued writes %d queued reads %d\n", isc->current_read_bias, isc->write_stats.queued, isc->read_stats.queued);
1166 		isc->current_read_bias--;
1167 		/* We're not limiting writes, per se, just doing reads first */
1168 		return NULL;
1169 	}
1170 
1171 	/*
1172 	 * See if our current limiter allows this I/O.
1173 	 */
1174 	if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1175 		if (iosched_debug)
1176 			printf("Can't write because limiter says no.\n");
1177 		isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1178 		return NULL;
1179 	}
1180 
1181 	/*
1182 	 * Let's do this: We've passed all the gates and we're a go
1183 	 * to schedule the I/O in the SIM.
1184 	 */
1185 	isc->current_read_bias = isc->read_bias;
1186 	bioq_remove(&isc->write_queue, bp);
1187 	if (bp->bio_cmd == BIO_WRITE) {
1188 		isc->write_stats.queued--;
1189 		isc->write_stats.total++;
1190 		isc->write_stats.pending++;
1191 	}
1192 	if (iosched_debug > 9)
1193 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1194 	isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1195 	return bp;
1196 }
1197 #endif
1198 
1199 /*
1200  * Put back a trim that you weren't able to actually schedule this time.
1201  */
1202 void
1203 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1204 {
1205 	bioq_insert_head(&isc->trim_queue, bp);
1206 #ifdef CAM_IOSCHED_DYNAMIC
1207 	isc->trim_stats.queued++;
1208 	isc->trim_stats.total--;		/* since we put it back, don't double count */
1209 	isc->trim_stats.pending--;
1210 #endif
1211 }
1212 
1213 /*
1214  * gets the next trim from the trim queue.
1215  *
1216  * Assumes we're called with the periph lock held.  It removes this
1217  * trim from the queue and the device must explicitly reinstert it
1218  * should the need arise.
1219  */
1220 struct bio *
1221 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1222 {
1223 	struct bio *bp;
1224 
1225 	bp  = bioq_first(&isc->trim_queue);
1226 	if (bp == NULL)
1227 		return NULL;
1228 	bioq_remove(&isc->trim_queue, bp);
1229 #ifdef CAM_IOSCHED_DYNAMIC
1230 	isc->trim_stats.queued--;
1231 	isc->trim_stats.total++;
1232 	isc->trim_stats.pending++;
1233 #endif
1234 	return bp;
1235 }
1236 
1237 /*
1238  * gets the an available trim from the trim queue, if there's no trim
1239  * already pending. It removes this trim from the queue and the device
1240  * must explicitly reinstert it should the need arise.
1241  *
1242  * Assumes we're called with the periph lock held.
1243  */
1244 struct bio *
1245 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1246 {
1247 
1248 	if (!cam_iosched_has_more_trim(isc))
1249 		return NULL;
1250 
1251 	return cam_iosched_next_trim(isc);
1252 }
1253 
1254 /*
1255  * Determine what the next bit of work to do is for the periph. The
1256  * default implementation looks to see if we have trims to do, but no
1257  * trims outstanding. If so, we do that. Otherwise we see if we have
1258  * other work. If we do, then we do that. Otherwise why were we called?
1259  */
1260 struct bio *
1261 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1262 {
1263 	struct bio *bp;
1264 
1265 	/*
1266 	 * See if we have a trim that can be scheduled. We can only send one
1267 	 * at a time down, so this takes that into account.
1268 	 *
1269 	 * XXX newer TRIM commands are queueable. Revisit this when we
1270 	 * implement them.
1271 	 */
1272 	if ((bp = cam_iosched_get_trim(isc)) != NULL)
1273 		return bp;
1274 
1275 #ifdef CAM_IOSCHED_DYNAMIC
1276 	/*
1277 	 * See if we have any pending writes, and room in the queue for them,
1278 	 * and if so, those are next.
1279 	 */
1280 	if (do_dynamic_iosched) {
1281 		if ((bp = cam_iosched_get_write(isc)) != NULL)
1282 			return bp;
1283 	}
1284 #endif
1285 
1286 	/*
1287 	 * next, see if there's other, normal I/O waiting. If so return that.
1288 	 */
1289 	if ((bp = bioq_first(&isc->bio_queue)) == NULL)
1290 		return NULL;
1291 
1292 #ifdef CAM_IOSCHED_DYNAMIC
1293 	/*
1294 	 * For the dynamic scheduler, bio_queue is only for reads, so enforce
1295 	 * the limits here. Enforce only for reads.
1296 	 */
1297 	if (do_dynamic_iosched) {
1298 		if (bp->bio_cmd == BIO_READ &&
1299 		    cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1300 			isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1301 			return NULL;
1302 		}
1303 	}
1304 	isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1305 #endif
1306 	bioq_remove(&isc->bio_queue, bp);
1307 #ifdef CAM_IOSCHED_DYNAMIC
1308 	if (do_dynamic_iosched) {
1309 		if (bp->bio_cmd == BIO_READ) {
1310 			isc->read_stats.queued--;
1311 			isc->read_stats.total++;
1312 			isc->read_stats.pending++;
1313 		} else
1314 			printf("Found bio_cmd = %#x\n", bp->bio_cmd);
1315 	}
1316 	if (iosched_debug > 9)
1317 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1318 #endif
1319 	return bp;
1320 }
1321 
1322 /*
1323  * Driver has been given some work to do by the block layer. Tell the
1324  * scheduler about it and have it queue the work up. The scheduler module
1325  * will then return the currently most useful bit of work later, possibly
1326  * deferring work for various reasons.
1327  */
1328 void
1329 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1330 {
1331 
1332 	/*
1333 	 * Put all trims on the trim queue sorted, since we know
1334 	 * that the collapsing code requires this. Otherwise put
1335 	 * the work on the bio queue.
1336 	 */
1337 	if (bp->bio_cmd == BIO_DELETE) {
1338 		bioq_disksort(&isc->trim_queue, bp);
1339 #ifdef CAM_IOSCHED_DYNAMIC
1340 		isc->trim_stats.in++;
1341 		isc->trim_stats.queued++;
1342 #endif
1343 	}
1344 #ifdef CAM_IOSCHED_DYNAMIC
1345 	else if (do_dynamic_iosched &&
1346 	    (bp->bio_cmd == BIO_WRITE || bp->bio_cmd == BIO_FLUSH)) {
1347 		if (cam_iosched_sort_queue(isc))
1348 			bioq_disksort(&isc->write_queue, bp);
1349 		else
1350 			bioq_insert_tail(&isc->write_queue, bp);
1351 		if (iosched_debug > 9)
1352 			printf("Qw  : %p %#x\n", bp, bp->bio_cmd);
1353 		if (bp->bio_cmd == BIO_WRITE) {
1354 			isc->write_stats.in++;
1355 			isc->write_stats.queued++;
1356 		}
1357 	}
1358 #endif
1359 	else {
1360 		if (cam_iosched_sort_queue(isc))
1361 			bioq_disksort(&isc->bio_queue, bp);
1362 		else
1363 			bioq_insert_tail(&isc->bio_queue, bp);
1364 #ifdef CAM_IOSCHED_DYNAMIC
1365 		if (iosched_debug > 9)
1366 			printf("Qr  : %p %#x\n", bp, bp->bio_cmd);
1367 		if (bp->bio_cmd == BIO_READ) {
1368 			isc->read_stats.in++;
1369 			isc->read_stats.queued++;
1370 		} else if (bp->bio_cmd == BIO_WRITE) {
1371 			isc->write_stats.in++;
1372 			isc->write_stats.queued++;
1373 		}
1374 #endif
1375 	}
1376 }
1377 
1378 /*
1379  * If we have work, get it scheduled. Called with the periph lock held.
1380  */
1381 void
1382 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1383 {
1384 
1385 	if (cam_iosched_has_work(isc))
1386 		xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1387 }
1388 
1389 /*
1390  * Complete a trim request
1391  */
1392 void
1393 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1394 {
1395 
1396 	isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1397 }
1398 
1399 /*
1400  * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1401  * might use notes in the ccb for statistics.
1402  */
1403 int
1404 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1405     union ccb *done_ccb)
1406 {
1407 	int retval = 0;
1408 #ifdef CAM_IOSCHED_DYNAMIC
1409 	if (!do_dynamic_iosched)
1410 		return retval;
1411 
1412 	if (iosched_debug > 10)
1413 		printf("done: %p %#x\n", bp, bp->bio_cmd);
1414 	if (bp->bio_cmd == BIO_WRITE) {
1415 		retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1416 		isc->write_stats.out++;
1417 		isc->write_stats.pending--;
1418 	} else if (bp->bio_cmd == BIO_READ) {
1419 		retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1420 		isc->read_stats.out++;
1421 		isc->read_stats.pending--;
1422 	} else if (bp->bio_cmd == BIO_DELETE) {
1423 		isc->trim_stats.out++;
1424 		isc->trim_stats.pending--;
1425 	} else if (bp->bio_cmd != BIO_FLUSH) {
1426 		if (iosched_debug)
1427 			printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1428 	}
1429 
1430 	if (!(bp->bio_flags & BIO_ERROR))
1431 		cam_iosched_io_metric_update(isc, done_ccb->ccb_h.qos.sim_data,
1432 		    bp->bio_cmd, bp->bio_bcount);
1433 #endif
1434 	return retval;
1435 }
1436 
1437 /*
1438  * Tell the io scheduler that you've pushed a trim down into the sim.
1439  * xxx better place for this?
1440  */
1441 void
1442 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1443 {
1444 
1445 	isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1446 }
1447 
1448 /*
1449  * Change the sorting policy hint for I/O transactions for this device.
1450  */
1451 void
1452 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1453 {
1454 
1455 	isc->sort_io_queue = val;
1456 }
1457 
1458 int
1459 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1460 {
1461 	return isc->flags & flags;
1462 }
1463 
1464 void
1465 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1466 {
1467 	isc->flags |= flags;
1468 }
1469 
1470 void
1471 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1472 {
1473 	isc->flags &= ~flags;
1474 }
1475 
1476 #ifdef CAM_IOSCHED_DYNAMIC
1477 /*
1478  * After the method presented in Jack Crenshaw's 1998 article "Integer
1479  * Suqare Roots," reprinted at
1480  * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1481  * and well worth the read. Briefly, we find the power of 4 that's the
1482  * largest smaller than val. We then check each smaller power of 4 to
1483  * see if val is still bigger. The right shifts at each step divide
1484  * the result by 2 which after successive application winds up
1485  * accumulating the right answer. It could also have been accumulated
1486  * using a separate root counter, but this code is smaller and faster
1487  * than that method. This method is also integer size invariant.
1488  * It returns floor(sqrt((float)val)), or the larget integer less than
1489  * or equal to the square root.
1490  */
1491 static uint64_t
1492 isqrt64(uint64_t val)
1493 {
1494 	uint64_t res = 0;
1495 	uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1496 
1497 	/*
1498 	 * Find the largest power of 4 smaller than val.
1499 	 */
1500 	while (bit > val)
1501 		bit >>= 2;
1502 
1503 	/*
1504 	 * Accumulate the answer, one bit at a time (we keep moving
1505 	 * them over since 2 is the square root of 4 and we test
1506 	 * powers of 4). We accumulate where we find the bit, but
1507 	 * the successive shifts land the bit in the right place
1508 	 * by the end.
1509 	 */
1510 	while (bit != 0) {
1511 		if (val >= res + bit) {
1512 			val -= res + bit;
1513 			res = (res >> 1) + bit;
1514 		} else
1515 			res >>= 1;
1516 		bit >>= 2;
1517 	}
1518 
1519 	return res;
1520 }
1521 
1522 static sbintime_t latencies[] = {
1523 	SBT_1MS <<  0,
1524 	SBT_1MS <<  1,
1525 	SBT_1MS <<  2,
1526 	SBT_1MS <<  3,
1527 	SBT_1MS <<  4,
1528 	SBT_1MS <<  5,
1529 	SBT_1MS <<  6,
1530 	SBT_1MS <<  7,
1531 	SBT_1MS <<  8,
1532 	SBT_1MS <<  9,
1533 	SBT_1MS << 10
1534 };
1535 
1536 static void
1537 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1538 {
1539 	sbintime_t y, deltasq, delta;
1540 	int i;
1541 
1542 	/*
1543 	 * Keep counts for latency. We do it by power of two buckets.
1544 	 * This helps us spot outlier behavior obscured by averages.
1545 	 */
1546 	for (i = 0; i < LAT_BUCKETS - 1; i++) {
1547 		if (sim_latency < latencies[i]) {
1548 			iop->latencies[i]++;
1549 			break;
1550 		}
1551 	}
1552 	if (i == LAT_BUCKETS - 1)
1553 		iop->latencies[i]++; 	 /* Put all > 1024ms values into the last bucket. */
1554 
1555 	/*
1556 	 * Classic expoentially decaying average with a tiny alpha
1557 	 * (2 ^ -alpha_bits). For more info see the NIST statistical
1558 	 * handbook.
1559 	 *
1560 	 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha)		[nist]
1561 	 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
1562 	 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1563 	 * alpha = 1 / (1 << alpha_bits)
1564 	 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
1565 	 *	= y_t/b - e/b + be/b
1566 	 *      = (y_t - e + be) / b
1567 	 *	= (e + d) / b
1568 	 *
1569 	 * Since alpha is a power of two, we can compute this w/o any mult or
1570 	 * division.
1571 	 *
1572 	 * Variance can also be computed. Usually, it would be expressed as follows:
1573 	 *	diff_t = y_t - ema_t-1
1574 	 *	emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
1575 	 *	  = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
1576 	 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
1577 	 *	  = e - e/b + dd/b + dd/bb
1578 	 *	  = (bbe - be + bdd + dd) / bb
1579 	 *	  = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
1580 	 */
1581 	/*
1582 	 * XXX possible numeric issues
1583 	 *	o We assume right shifted integers do the right thing, since that's
1584 	 *	  implementation defined. You can change the right shifts to / (1LL << alpha).
1585 	 *	o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
1586 	 *	  for emvar. This puts a ceiling of 13 bits on alpha since we need a
1587 	 *	  few tens of seconds of representation.
1588 	 *	o We mitigate alpha issues by never setting it too high.
1589 	 */
1590 	y = sim_latency;
1591 	delta = (y - iop->ema);					/* d */
1592 	iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1593 
1594 	/*
1595 	 * Were we to naively plow ahead at this point, we wind up with many numerical
1596 	 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
1597 	 * us with microsecond level precision in the input, so the same in the
1598 	 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
1599 	 * also means that emvar can be up 46 bits 40 of which are fraction, which
1600 	 * gives us a way to measure up to ~8s in the SD before the computation goes
1601 	 * unstable. Even the worst hard disk rarely has > 1s service time in the
1602 	 * drive. It does mean we have to shift left 12 bits after taking the
1603 	 * square root to compute the actual standard deviation estimate. This loss of
1604 	 * precision is preferable to needing int128 types to work. The above numbers
1605 	 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
1606 	 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1607 	 */
1608 	delta >>= 12;
1609 	deltasq = delta * delta;				/* dd */
1610 	iop->emvar = ((iop->emvar << (2 * alpha_bits)) +	/* bbe */
1611 	    ((deltasq - iop->emvar) << alpha_bits) +		/* b(dd-e) */
1612 	    deltasq)						/* dd */
1613 	    >> (2 * alpha_bits);				/* div bb */
1614 	iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1615 }
1616 
1617 static void
1618 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1619     sbintime_t sim_latency, int cmd, size_t size)
1620 {
1621 	/* xxx Do we need to scale based on the size of the I/O ? */
1622 	switch (cmd) {
1623 	case BIO_READ:
1624 		cam_iosched_update(&isc->read_stats, sim_latency);
1625 		break;
1626 	case BIO_WRITE:
1627 		cam_iosched_update(&isc->write_stats, sim_latency);
1628 		break;
1629 	case BIO_DELETE:
1630 		cam_iosched_update(&isc->trim_stats, sim_latency);
1631 		break;
1632 	default:
1633 		break;
1634 	}
1635 }
1636 
1637 #ifdef DDB
1638 static int biolen(struct bio_queue_head *bq)
1639 {
1640 	int i = 0;
1641 	struct bio *bp;
1642 
1643 	TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
1644 		i++;
1645 	}
1646 	return i;
1647 }
1648 
1649 /*
1650  * Show the internal state of the I/O scheduler.
1651  */
1652 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
1653 {
1654 	struct cam_iosched_softc *isc;
1655 
1656 	if (!have_addr) {
1657 		db_printf("Need addr\n");
1658 		return;
1659 	}
1660 	isc = (struct cam_iosched_softc *)addr;
1661 	db_printf("pending_reads:     %d\n", isc->read_stats.pending);
1662 	db_printf("min_reads:         %d\n", isc->read_stats.min);
1663 	db_printf("max_reads:         %d\n", isc->read_stats.max);
1664 	db_printf("reads:             %d\n", isc->read_stats.total);
1665 	db_printf("in_reads:          %d\n", isc->read_stats.in);
1666 	db_printf("out_reads:         %d\n", isc->read_stats.out);
1667 	db_printf("queued_reads:      %d\n", isc->read_stats.queued);
1668 	db_printf("Current Q len      %d\n", biolen(&isc->bio_queue));
1669 	db_printf("pending_writes:    %d\n", isc->write_stats.pending);
1670 	db_printf("min_writes:        %d\n", isc->write_stats.min);
1671 	db_printf("max_writes:        %d\n", isc->write_stats.max);
1672 	db_printf("writes:            %d\n", isc->write_stats.total);
1673 	db_printf("in_writes:         %d\n", isc->write_stats.in);
1674 	db_printf("out_writes:        %d\n", isc->write_stats.out);
1675 	db_printf("queued_writes:     %d\n", isc->write_stats.queued);
1676 	db_printf("Current Q len      %d\n", biolen(&isc->write_queue));
1677 	db_printf("pending_trims:     %d\n", isc->trim_stats.pending);
1678 	db_printf("min_trims:         %d\n", isc->trim_stats.min);
1679 	db_printf("max_trims:         %d\n", isc->trim_stats.max);
1680 	db_printf("trims:             %d\n", isc->trim_stats.total);
1681 	db_printf("in_trims:          %d\n", isc->trim_stats.in);
1682 	db_printf("out_trims:         %d\n", isc->trim_stats.out);
1683 	db_printf("queued_trims:      %d\n", isc->trim_stats.queued);
1684 	db_printf("Current Q len      %d\n", biolen(&isc->trim_queue));
1685 	db_printf("read_bias:         %d\n", isc->read_bias);
1686 	db_printf("current_read_bias: %d\n", isc->current_read_bias);
1687 	db_printf("Trim active?       %s\n",
1688 	    (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");
1689 }
1690 #endif
1691 #endif
1692