xref: /freebsd/sys/cam/cam_iosched.c (revision 23f6875a43f7ce365f2d52cf857da010c47fb03b)
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  */
97 static int alpha_bits = 9;
98 TUNABLE_INT("kern.cam.iosched_alpha_bits", &alpha_bits);
99 SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW,
100     &alpha_bits, 1,
101     "Bits in EMA's alpha.");
102 
103 struct iop_stats;
104 struct cam_iosched_softc;
105 
106 int iosched_debug = 0;
107 
108 typedef enum {
109 	none = 0,				/* No limits */
110 	queue_depth,			/* Limit how many ops we queue to SIM */
111 	iops,				/* Limit # of IOPS to the drive */
112 	bandwidth,			/* Limit bandwidth to the drive */
113 	limiter_max
114 } io_limiter;
115 
116 static const char *cam_iosched_limiter_names[] =
117     { "none", "queue_depth", "iops", "bandwidth" };
118 
119 /*
120  * Called to initialize the bits of the iop_stats structure relevant to the
121  * limiter. Called just after the limiter is set.
122  */
123 typedef int l_init_t(struct iop_stats *);
124 
125 /*
126  * Called every tick.
127  */
128 typedef int l_tick_t(struct iop_stats *);
129 
130 /*
131  * Called to see if the limiter thinks this IOP can be allowed to
132  * proceed. If so, the limiter assumes that the while IOP proceeded
133  * and makes any accounting of it that's needed.
134  */
135 typedef int l_iop_t(struct iop_stats *, struct bio *);
136 
137 /*
138  * Called when an I/O completes so the limiter can updates its
139  * accounting. Pending I/Os may complete in any order (even when
140  * sent to the hardware at the same time), so the limiter may not
141  * make any assumptions other than this I/O has completed. If it
142  * returns 1, then xpt_schedule() needs to be called again.
143  */
144 typedef int l_iodone_t(struct iop_stats *, struct bio *);
145 
146 static l_iop_t cam_iosched_qd_iop;
147 static l_iop_t cam_iosched_qd_caniop;
148 static l_iodone_t cam_iosched_qd_iodone;
149 
150 static l_init_t cam_iosched_iops_init;
151 static l_tick_t cam_iosched_iops_tick;
152 static l_iop_t cam_iosched_iops_caniop;
153 static l_iop_t cam_iosched_iops_iop;
154 
155 static l_init_t cam_iosched_bw_init;
156 static l_tick_t cam_iosched_bw_tick;
157 static l_iop_t cam_iosched_bw_caniop;
158 static l_iop_t cam_iosched_bw_iop;
159 
160 struct limswitch
161 {
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 	/*
201 	 * sysctl state for this subnode.
202 	 */
203 	struct sysctl_ctx_list	sysctl_ctx;
204 	struct sysctl_oid	*sysctl_tree;
205 
206 	/*
207 	 * Information about the current rate limiters, if any
208 	 */
209 	io_limiter	limiter;	/* How are I/Os being limited */
210 	int		min;		/* Low range of limit */
211 	int		max;		/* High range of limit */
212 	int		current;	/* Current rate limiter */
213 	int		l_value1;	/* per-limiter scratch value 1. */
214 	int		l_value2;	/* per-limiter scratch value 2. */
215 
216 
217 	/*
218 	 * Debug information about counts of I/Os that have gone through the
219 	 * scheduler.
220 	 */
221 	int		pending;	/* I/Os pending in the hardware */
222 	int		queued;		/* number currently in the queue */
223 	int		total;		/* Total for all time -- wraps */
224 	int		in;		/* number queued all time -- wraps */
225 	int		out;		/* number completed all time -- wraps */
226 
227 	/*
228 	 * Statistics on different bits of the process.
229 	 */
230 		/* Exp Moving Average, see alpha_bits for more details */
231 	sbintime_t      ema;
232 	sbintime_t      emss;		/* Exp Moving sum of the squares */
233 	sbintime_t      sd;		/* Last computed sd */
234 
235 	uint32_t	state_flags;
236 #define IOP_RATE_LIMITED		1u
237 
238 #define LAT_BUCKETS 12			/* < 1ms < 2ms ... 512ms < 1024ms > 1024ms */
239 	uint64_t	latencies[LAT_BUCKETS];
240 
241 	struct cam_iosched_softc *softc;
242 };
243 
244 
245 typedef enum {
246 	set_max = 0,			/* current = max */
247 	read_latency,			/* Steer read latency by throttling writes */
248 	cl_max				/* Keep last */
249 } control_type;
250 
251 static const char *cam_iosched_control_type_names[] =
252     { "set_max", "read_latency" };
253 
254 struct control_loop
255 {
256 	/*
257 	 * sysctl state for this subnode.
258 	 */
259 	struct sysctl_ctx_list	sysctl_ctx;
260 	struct sysctl_oid	*sysctl_tree;
261 
262 	sbintime_t	next_steer;		/* Time of next steer */
263 	sbintime_t	steer_interval;		/* How often do we steer? */
264 	sbintime_t	lolat;
265 	sbintime_t	hilat;
266 	int		alpha;
267 	control_type	type;			/* What type of control? */
268 	int		last_count;		/* Last I/O count */
269 
270 	struct cam_iosched_softc *softc;
271 };
272 
273 #endif
274 
275 struct cam_iosched_softc
276 {
277 	struct bio_queue_head bio_queue;
278 	struct bio_queue_head trim_queue;
279 				/* scheduler flags < 16, user flags >= 16 */
280 	uint32_t	flags;
281 	int		sort_io_queue;
282 #ifdef CAM_IOSCHED_DYNAMIC
283 	int		read_bias;		/* Read bias setting */
284 	int		current_read_bias;	/* Current read bias state */
285 	int		total_ticks;
286 	int		load;			/* EMA of 'load average' of disk / 2^16 */
287 
288 	struct bio_queue_head write_queue;
289 	struct iop_stats read_stats, write_stats, trim_stats;
290 	struct sysctl_ctx_list	sysctl_ctx;
291 	struct sysctl_oid	*sysctl_tree;
292 
293 	int		quanta;			/* Number of quanta per second */
294 	struct callout	ticker;			/* Callout for our quota system */
295 	struct cam_periph *periph;		/* cam periph associated with this device */
296 	uint32_t	this_frac;		/* Fraction of a second (1024ths) for this tick */
297 	sbintime_t	last_time;		/* Last time we ticked */
298 	struct control_loop cl;
299 #endif
300 };
301 
302 #ifdef CAM_IOSCHED_DYNAMIC
303 /*
304  * helper functions to call the limsw functions.
305  */
306 static int
307 cam_iosched_limiter_init(struct iop_stats *ios)
308 {
309 	int lim = ios->limiter;
310 
311 	/* maybe this should be a kassert */
312 	if (lim < none || lim >= limiter_max)
313 		return EINVAL;
314 
315 	if (limsw[lim].l_init)
316 		return limsw[lim].l_init(ios);
317 
318 	return 0;
319 }
320 
321 static int
322 cam_iosched_limiter_tick(struct iop_stats *ios)
323 {
324 	int lim = ios->limiter;
325 
326 	/* maybe this should be a kassert */
327 	if (lim < none || lim >= limiter_max)
328 		return EINVAL;
329 
330 	if (limsw[lim].l_tick)
331 		return limsw[lim].l_tick(ios);
332 
333 	return 0;
334 }
335 
336 static int
337 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
338 {
339 	int lim = ios->limiter;
340 
341 	/* maybe this should be a kassert */
342 	if (lim < none || lim >= limiter_max)
343 		return EINVAL;
344 
345 	if (limsw[lim].l_iop)
346 		return limsw[lim].l_iop(ios, bp);
347 
348 	return 0;
349 }
350 
351 static int
352 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
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_caniop)
361 		return limsw[lim].l_caniop(ios, bp);
362 
363 	return 0;
364 }
365 
366 static int
367 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
368 {
369 	int lim = ios->limiter;
370 
371 	/* maybe this should be a kassert */
372 	if (lim < none || lim >= limiter_max)
373 		return 0;
374 
375 	if (limsw[lim].l_iodone)
376 		return limsw[lim].l_iodone(ios, bp);
377 
378 	return 0;
379 }
380 
381 /*
382  * Functions to implement the different kinds of limiters
383  */
384 
385 static int
386 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
387 {
388 
389 	if (ios->current <= 0 || ios->pending < ios->current)
390 		return 0;
391 
392 	return EAGAIN;
393 }
394 
395 static int
396 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
397 {
398 
399 	if (ios->current <= 0 || ios->pending < ios->current)
400 		return 0;
401 
402 	return EAGAIN;
403 }
404 
405 static int
406 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
407 {
408 
409 	if (ios->current <= 0 || ios->pending != ios->current)
410 		return 0;
411 
412 	return 1;
413 }
414 
415 static int
416 cam_iosched_iops_init(struct iop_stats *ios)
417 {
418 
419 	ios->l_value1 = ios->current / ios->softc->quanta;
420 	if (ios->l_value1 <= 0)
421 		ios->l_value1 = 1;
422 
423 	return 0;
424 }
425 
426 static int
427 cam_iosched_iops_tick(struct iop_stats *ios)
428 {
429 
430 	ios->l_value1 = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
431 	if (ios->l_value1 <= 0)
432 		ios->l_value1 = 1;
433 
434 	return 0;
435 }
436 
437 static int
438 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
439 {
440 
441 	/*
442 	 * So if we have any more IOPs left, allow it,
443 	 * otherwise wait.
444 	 */
445 	if (ios->l_value1 <= 0)
446 		return EAGAIN;
447 	return 0;
448 }
449 
450 static int
451 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
452 {
453 	int rv;
454 
455 	rv = cam_iosched_limiter_caniop(ios, bp);
456 	if (rv == 0)
457 		ios->l_value1--;
458 
459 	return rv;
460 }
461 
462 static int
463 cam_iosched_bw_init(struct iop_stats *ios)
464 {
465 
466 	/* ios->current is in kB/s, so scale to bytes */
467 	ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
468 
469 	return 0;
470 }
471 
472 static int
473 cam_iosched_bw_tick(struct iop_stats *ios)
474 {
475 	int bw;
476 
477 	/*
478 	 * If we're in the hole for available quota from
479 	 * the last time, then add the quantum for this.
480 	 * If we have any left over from last quantum,
481 	 * then too bad, that's lost. Also, ios->current
482 	 * is in kB/s, so scale.
483 	 *
484 	 * We also allow up to 4 quanta of credits to
485 	 * accumulate to deal with burstiness. 4 is extremely
486 	 * arbitrary.
487 	 */
488 	bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
489 	if (ios->l_value1 < bw * 4)
490 		ios->l_value1 += bw;
491 
492 	return 0;
493 }
494 
495 static int
496 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
497 {
498 	/*
499 	 * So if we have any more bw quota left, allow it,
500 	 * otherwise wait. Not, we'll go negative and that's
501 	 * OK. We'll just get a lettle less next quota.
502 	 *
503 	 * Note on going negative: that allows us to process
504 	 * requests in order better, since we won't allow
505 	 * shorter reads to get around the long one that we
506 	 * don't have the quota to do just yet. It also prevents
507 	 * starvation by being a little more permissive about
508 	 * what we let through this quantum (to prevent the
509 	 * starvation), at the cost of getting a little less
510 	 * next quantum.
511 	 */
512 	if (ios->l_value1 <= 0)
513 		return EAGAIN;
514 
515 
516 	return 0;
517 }
518 
519 static int
520 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
521 {
522 	int rv;
523 
524 	rv = cam_iosched_limiter_caniop(ios, bp);
525 	if (rv == 0)
526 		ios->l_value1 -= bp->bio_length;
527 
528 	return rv;
529 }
530 
531 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
532 
533 static void
534 cam_iosched_ticker(void *arg)
535 {
536 	struct cam_iosched_softc *isc = arg;
537 	sbintime_t now, delta;
538 	int pending;
539 
540 	callout_reset(&isc->ticker, hz / isc->quanta - 1, cam_iosched_ticker, isc);
541 
542 	now = sbinuptime();
543 	delta = now - isc->last_time;
544 	isc->this_frac = (uint32_t)delta >> 16;		/* Note: discards seconds -- should be 0 harmless if not */
545 	isc->last_time = now;
546 
547 	cam_iosched_cl_maybe_steer(&isc->cl);
548 
549 	cam_iosched_limiter_tick(&isc->read_stats);
550 	cam_iosched_limiter_tick(&isc->write_stats);
551 	cam_iosched_limiter_tick(&isc->trim_stats);
552 
553 	cam_iosched_schedule(isc, isc->periph);
554 
555 	/*
556 	 * isc->load is an EMA of the pending I/Os at each tick. The number of
557 	 * pending I/Os is the sum of the I/Os queued to the hardware, and those
558 	 * in the software queue that could be queued to the hardware if there
559 	 * were slots.
560 	 *
561 	 * ios_stats.pending is a count of requests in the SIM right now for
562 	 * each of these types of I/O. So the total pending count is the sum of
563 	 * these I/Os and the sum of the queued I/Os still in the software queue
564 	 * for those operations that aren't being rate limited at the moment.
565 	 *
566 	 * The reason for the rate limiting bit is because those I/Os
567 	 * aren't part of the software queued load (since we could
568 	 * give them to hardware, but choose not to).
569 	 *
570 	 * Note: due to a bug in counting pending TRIM in the device, we
571 	 * don't include them in this count. We count each BIO_DELETE in
572 	 * the pending count, but the periph drivers collapse them down
573 	 * into one TRIM command. That one trim command gets the completion
574 	 * so the counts get off.
575 	 */
576 	pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
577 	pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
578 	    !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
579 	    !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
580 	pending <<= 16;
581 	pending /= isc->periph->path->device->ccbq.total_openings;
582 
583 	isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
584 
585 	isc->total_ticks++;
586 }
587 
588 
589 static void
590 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
591 {
592 
593 	clp->next_steer = sbinuptime();
594 	clp->softc = isc;
595 	clp->steer_interval = SBT_1S * 5;	/* Let's start out steering every 5s */
596 	clp->lolat = 5 * SBT_1MS;
597 	clp->hilat = 15 * SBT_1MS;
598 	clp->alpha = 20;			/* Alpha == gain. 20 = .2 */
599 	clp->type = set_max;
600 }
601 
602 static void
603 cam_iosched_cl_maybe_steer(struct control_loop *clp)
604 {
605 	struct cam_iosched_softc *isc;
606 	sbintime_t now, lat;
607 	int old;
608 
609 	isc = clp->softc;
610 	now = isc->last_time;
611 	if (now < clp->next_steer)
612 		return;
613 
614 	clp->next_steer = now + clp->steer_interval;
615 	switch (clp->type) {
616 	case set_max:
617 		if (isc->write_stats.current != isc->write_stats.max)
618 			printf("Steering write from %d kBps to %d kBps\n",
619 			    isc->write_stats.current, isc->write_stats.max);
620 		isc->read_stats.current = isc->read_stats.max;
621 		isc->write_stats.current = isc->write_stats.max;
622 		isc->trim_stats.current = isc->trim_stats.max;
623 		break;
624 	case read_latency:
625 		old = isc->write_stats.current;
626 		lat = isc->read_stats.ema;
627 		/*
628 		 * Simple PLL-like engine. Since we're steering to a range for
629 		 * the SP (set point) that makes things a little more
630 		 * complicated. In addition, we're not directly controlling our
631 		 * PV (process variable), the read latency, but instead are
632 		 * manipulating the write bandwidth limit for our MV
633 		 * (manipulation variable), analysis of this code gets a bit
634 		 * messy. Also, the MV is a very noisy control surface for read
635 		 * latency since it is affected by many hidden processes inside
636 		 * the device which change how responsive read latency will be
637 		 * in reaction to changes in write bandwidth. Unlike the classic
638 		 * boiler control PLL. this may result in over-steering while
639 		 * the SSD takes its time to react to the new, lower load. This
640 		 * is why we use a relatively low alpha of between .1 and .25 to
641 		 * compensate for this effect. At .1, it takes ~22 steering
642 		 * intervals to back off by a factor of 10. At .2 it only takes
643 		 * ~10. At .25 it only takes ~8. However some preliminary data
644 		 * from the SSD drives suggests a reasponse time in 10's of
645 		 * seconds before latency drops regardless of the new write
646 		 * rate. Careful observation will be reqiured to tune this
647 		 * effectively.
648 		 *
649 		 * Also, when there's no read traffic, we jack up the write
650 		 * limit too regardless of the last read latency.  10 is
651 		 * somewhat arbitrary.
652 		 */
653 		if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
654 			isc->write_stats.current = isc->write_stats.current *
655 			    (100 + clp->alpha) / 100;	/* Scale up */
656 		else if (lat > clp->hilat)
657 			isc->write_stats.current = isc->write_stats.current *
658 			    (100 - clp->alpha) / 100;	/* Scale down */
659 		clp->last_count = isc->read_stats.total;
660 
661 		/*
662 		 * Even if we don't steer, per se, enforce the min/max limits as
663 		 * those may have changed.
664 		 */
665 		if (isc->write_stats.current < isc->write_stats.min)
666 			isc->write_stats.current = isc->write_stats.min;
667 		if (isc->write_stats.current > isc->write_stats.max)
668 			isc->write_stats.current = isc->write_stats.max;
669 		if (old != isc->write_stats.current && 	iosched_debug)
670 			printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
671 			    old, isc->write_stats.current,
672 			    (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
673 		break;
674 	case cl_max:
675 		break;
676 	}
677 }
678 #endif
679 
680 			/* Trim or similar currently pending completion */
681 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE	(1ul << 0)
682 			/* Callout active, and needs to be torn down */
683 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
684 
685 			/* Periph drivers set these flags to indicate work */
686 #define CAM_IOSCHED_FLAG_WORK_FLAGS	((0xffffu) << 16)
687 
688 #ifdef CAM_IOSCHED_DYNAMIC
689 static void
690 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
691     sbintime_t sim_latency, int cmd, size_t size);
692 #endif
693 
694 static inline int
695 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
696 {
697 	return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
698 }
699 
700 static inline int
701 cam_iosched_has_io(struct cam_iosched_softc *isc)
702 {
703 #ifdef CAM_IOSCHED_DYNAMIC
704 	if (do_dynamic_iosched) {
705 		struct bio *rbp = bioq_first(&isc->bio_queue);
706 		struct bio *wbp = bioq_first(&isc->write_queue);
707 		int can_write = wbp != NULL &&
708 		    cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
709 		int can_read = rbp != NULL &&
710 		    cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
711 		if (iosched_debug > 2) {
712 			printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
713 			printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
714 			printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
715 		}
716 		return can_read || can_write;
717 	}
718 #endif
719 	return bioq_first(&isc->bio_queue) != NULL;
720 }
721 
722 static inline int
723 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
724 {
725 	return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) &&
726 	    bioq_first(&isc->trim_queue);
727 }
728 
729 #define cam_iosched_sort_queue(isc)	((isc)->sort_io_queue >= 0 ?	\
730     (isc)->sort_io_queue : cam_sort_io_queues)
731 
732 
733 static inline int
734 cam_iosched_has_work(struct cam_iosched_softc *isc)
735 {
736 #ifdef CAM_IOSCHED_DYNAMIC
737 	if (iosched_debug > 2)
738 		printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
739 		    cam_iosched_has_more_trim(isc),
740 		    cam_iosched_has_flagged_work(isc));
741 #endif
742 
743 	return cam_iosched_has_io(isc) ||
744 		cam_iosched_has_more_trim(isc) ||
745 		cam_iosched_has_flagged_work(isc);
746 }
747 
748 #ifdef CAM_IOSCHED_DYNAMIC
749 static void
750 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
751 {
752 
753 	ios->limiter = none;
754 	cam_iosched_limiter_init(ios);
755 	ios->in = 0;
756 	ios->max = 300000;
757 	ios->min = 1;
758 	ios->out = 0;
759 	ios->pending = 0;
760 	ios->queued = 0;
761 	ios->total = 0;
762 	ios->ema = 0;
763 	ios->emss = 0;
764 	ios->sd = 0;
765 	ios->softc = isc;
766 }
767 
768 static int
769 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
770 {
771 	char buf[16];
772 	struct iop_stats *ios;
773 	struct cam_iosched_softc *isc;
774 	int value, i, error;
775 	const char *p;
776 
777 	ios = arg1;
778 	isc = ios->softc;
779 	value = ios->limiter;
780 	if (value < none || value >= limiter_max)
781 		p = "UNKNOWN";
782 	else
783 		p = cam_iosched_limiter_names[value];
784 
785 	strlcpy(buf, p, sizeof(buf));
786 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
787 	if (error != 0 || req->newptr == NULL)
788 		return error;
789 
790 	cam_periph_lock(isc->periph);
791 
792 	for (i = none; i < limiter_max; i++) {
793 		if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
794 			continue;
795 		ios->limiter = i;
796 		error = cam_iosched_limiter_init(ios);
797 		if (error != 0) {
798 			ios->limiter = value;
799 			cam_periph_unlock(isc->periph);
800 			return error;
801 		}
802 		/* Note: disk load averate requires ticker to be always running */
803 		callout_reset(&isc->ticker, hz / isc->quanta - 1, cam_iosched_ticker, isc);
804 		isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
805 
806 		cam_periph_unlock(isc->periph);
807 		return 0;
808 	}
809 
810 	cam_periph_unlock(isc->periph);
811 	return EINVAL;
812 }
813 
814 static int
815 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
816 {
817 	char buf[16];
818 	struct control_loop *clp;
819 	struct cam_iosched_softc *isc;
820 	int value, i, error;
821 	const char *p;
822 
823 	clp = arg1;
824 	isc = clp->softc;
825 	value = clp->type;
826 	if (value < none || value >= cl_max)
827 		p = "UNKNOWN";
828 	else
829 		p = cam_iosched_control_type_names[value];
830 
831 	strlcpy(buf, p, sizeof(buf));
832 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
833 	if (error != 0 || req->newptr == NULL)
834 		return error;
835 
836 	for (i = set_max; i < cl_max; i++) {
837 		if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
838 			continue;
839 		cam_periph_lock(isc->periph);
840 		clp->type = i;
841 		cam_periph_unlock(isc->periph);
842 		return 0;
843 	}
844 
845 	return EINVAL;
846 }
847 
848 static int
849 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
850 {
851 	char buf[16];
852 	sbintime_t value;
853 	int error;
854 	uint64_t us;
855 
856 	value = *(sbintime_t *)arg1;
857 	us = (uint64_t)value / SBT_1US;
858 	snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
859 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
860 	if (error != 0 || req->newptr == NULL)
861 		return error;
862 	us = strtoul(buf, NULL, 10);
863 	if (us == 0)
864 		return EINVAL;
865 	*(sbintime_t *)arg1 = us * SBT_1US;
866 	return 0;
867 }
868 
869 static int
870 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
871 {
872 	int i, error;
873 	struct sbuf sb;
874 	uint64_t *latencies;
875 
876 	latencies = arg1;
877 	sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
878 
879 	for (i = 0; i < LAT_BUCKETS - 1; i++)
880 		sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
881 	sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
882 	error = sbuf_finish(&sb);
883 	sbuf_delete(&sb);
884 
885 	return (error);
886 }
887 
888 static void
889 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
890 {
891 	struct sysctl_oid_list *n;
892 	struct sysctl_ctx_list *ctx;
893 
894 	ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
895 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
896 	    CTLFLAG_RD, 0, name);
897 	n = SYSCTL_CHILDREN(ios->sysctl_tree);
898 	ctx = &ios->sysctl_ctx;
899 
900 	SYSCTL_ADD_UQUAD(ctx, n,
901 	    OID_AUTO, "ema", CTLFLAG_RD,
902 	    &ios->ema,
903 	    "Fast Exponentially Weighted Moving Average");
904 	SYSCTL_ADD_UQUAD(ctx, n,
905 	    OID_AUTO, "emss", CTLFLAG_RD,
906 	    &ios->emss,
907 	    "Fast Exponentially Weighted Moving Sum of Squares (maybe wrong)");
908 	SYSCTL_ADD_UQUAD(ctx, n,
909 	    OID_AUTO, "sd", CTLFLAG_RD,
910 	    &ios->sd,
911 	    "Estimated SD for fast ema (may be wrong)");
912 
913 	SYSCTL_ADD_INT(ctx, n,
914 	    OID_AUTO, "pending", CTLFLAG_RD,
915 	    &ios->pending, 0,
916 	    "Instantaneous # of pending transactions");
917 	SYSCTL_ADD_INT(ctx, n,
918 	    OID_AUTO, "count", CTLFLAG_RD,
919 	    &ios->total, 0,
920 	    "# of transactions submitted to hardware");
921 	SYSCTL_ADD_INT(ctx, n,
922 	    OID_AUTO, "queued", CTLFLAG_RD,
923 	    &ios->queued, 0,
924 	    "# of transactions in the queue");
925 	SYSCTL_ADD_INT(ctx, n,
926 	    OID_AUTO, "in", CTLFLAG_RD,
927 	    &ios->in, 0,
928 	    "# of transactions queued to driver");
929 	SYSCTL_ADD_INT(ctx, n,
930 	    OID_AUTO, "out", CTLFLAG_RD,
931 	    &ios->out, 0,
932 	    "# of transactions completed");
933 
934 	SYSCTL_ADD_PROC(ctx, n,
935 	    OID_AUTO, "limiter", CTLTYPE_STRING | CTLFLAG_RW,
936 	    ios, 0, cam_iosched_limiter_sysctl, "A",
937 	    "Current limiting type.");
938 	SYSCTL_ADD_INT(ctx, n,
939 	    OID_AUTO, "min", CTLFLAG_RW,
940 	    &ios->min, 0,
941 	    "min resource");
942 	SYSCTL_ADD_INT(ctx, n,
943 	    OID_AUTO, "max", CTLFLAG_RW,
944 	    &ios->max, 0,
945 	    "max resource");
946 	SYSCTL_ADD_INT(ctx, n,
947 	    OID_AUTO, "current", CTLFLAG_RW,
948 	    &ios->current, 0,
949 	    "current resource");
950 
951 	SYSCTL_ADD_PROC(ctx, n,
952 	    OID_AUTO, "latencies", CTLTYPE_STRING | CTLFLAG_RD,
953 	    &ios->latencies, 0,
954 	    cam_iosched_sysctl_latencies, "A",
955 	    "Array of power of 2 latency from 1ms to 1.024s");
956 }
957 
958 static void
959 cam_iosched_iop_stats_fini(struct iop_stats *ios)
960 {
961 	if (ios->sysctl_tree)
962 		if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
963 			printf("can't remove iosched sysctl stats context\n");
964 }
965 
966 static void
967 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
968 {
969 	struct sysctl_oid_list *n;
970 	struct sysctl_ctx_list *ctx;
971 	struct control_loop *clp;
972 
973 	clp = &isc->cl;
974 	clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
975 	    SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
976 	    CTLFLAG_RD, 0, "Control loop info");
977 	n = SYSCTL_CHILDREN(clp->sysctl_tree);
978 	ctx = &clp->sysctl_ctx;
979 
980 	SYSCTL_ADD_PROC(ctx, n,
981 	    OID_AUTO, "type", CTLTYPE_STRING | CTLFLAG_RW,
982 	    clp, 0, cam_iosched_control_type_sysctl, "A",
983 	    "Control loop algorithm");
984 	SYSCTL_ADD_PROC(ctx, n,
985 	    OID_AUTO, "steer_interval", CTLTYPE_STRING | CTLFLAG_RW,
986 	    &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
987 	    "How often to steer (in us)");
988 	SYSCTL_ADD_PROC(ctx, n,
989 	    OID_AUTO, "lolat", CTLTYPE_STRING | CTLFLAG_RW,
990 	    &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
991 	    "Low water mark for Latency (in us)");
992 	SYSCTL_ADD_PROC(ctx, n,
993 	    OID_AUTO, "hilat", CTLTYPE_STRING | CTLFLAG_RW,
994 	    &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
995 	    "Hi water mark for Latency (in us)");
996 	SYSCTL_ADD_INT(ctx, n,
997 	    OID_AUTO, "alpha", CTLFLAG_RW,
998 	    &clp->alpha, 0,
999 	    "Alpha for PLL (x100) aka gain");
1000 }
1001 
1002 static void
1003 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1004 {
1005 	if (clp->sysctl_tree)
1006 		if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1007 			printf("can't remove iosched sysctl control loop context\n");
1008 }
1009 #endif
1010 
1011 /*
1012  * Allocate the iosched structure. This also insulates callers from knowing
1013  * sizeof struct cam_iosched_softc.
1014  */
1015 int
1016 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1017 {
1018 
1019 	*iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1020 	if (*iscp == NULL)
1021 		return ENOMEM;
1022 #ifdef CAM_IOSCHED_DYNAMIC
1023 	if (iosched_debug)
1024 		printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1025 #endif
1026 	(*iscp)->sort_io_queue = -1;
1027 	bioq_init(&(*iscp)->bio_queue);
1028 	bioq_init(&(*iscp)->trim_queue);
1029 #ifdef CAM_IOSCHED_DYNAMIC
1030 	if (do_dynamic_iosched) {
1031 		bioq_init(&(*iscp)->write_queue);
1032 		(*iscp)->read_bias = 100;
1033 		(*iscp)->current_read_bias = 100;
1034 		(*iscp)->quanta = 200;
1035 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1036 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1037 		cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1038 		(*iscp)->trim_stats.max = 1;	/* Trims are special: one at a time for now */
1039 		(*iscp)->last_time = sbinuptime();
1040 		callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1041 		(*iscp)->periph = periph;
1042 		cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1043 		callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta - 1, cam_iosched_ticker, *iscp);
1044 		(*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1045 	}
1046 #endif
1047 
1048 	return 0;
1049 }
1050 
1051 /*
1052  * Reclaim all used resources. This assumes that other folks have
1053  * drained the requests in the hardware. Maybe an unwise assumption.
1054  */
1055 void
1056 cam_iosched_fini(struct cam_iosched_softc *isc)
1057 {
1058 	if (isc) {
1059 		cam_iosched_flush(isc, NULL, ENXIO);
1060 #ifdef CAM_IOSCHED_DYNAMIC
1061 		cam_iosched_iop_stats_fini(&isc->read_stats);
1062 		cam_iosched_iop_stats_fini(&isc->write_stats);
1063 		cam_iosched_iop_stats_fini(&isc->trim_stats);
1064 		cam_iosched_cl_sysctl_fini(&isc->cl);
1065 		if (isc->sysctl_tree)
1066 			if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1067 				printf("can't remove iosched sysctl stats context\n");
1068 		if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1069 			callout_drain(&isc->ticker);
1070 			isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1071 		}
1072 
1073 #endif
1074 		free(isc, M_CAMSCHED);
1075 	}
1076 }
1077 
1078 /*
1079  * After we're sure we're attaching a device, go ahead and add
1080  * hooks for any sysctl we may wish to honor.
1081  */
1082 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1083     struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1084 {
1085 #ifdef CAM_IOSCHED_DYNAMIC
1086 	struct sysctl_oid_list *n;
1087 #endif
1088 
1089 	SYSCTL_ADD_INT(ctx, SYSCTL_CHILDREN(node),
1090 		OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1091 		&isc->sort_io_queue, 0,
1092 		"Sort IO queue to try and optimise disk access patterns");
1093 
1094 #ifdef CAM_IOSCHED_DYNAMIC
1095 	if (!do_dynamic_iosched)
1096 		return;
1097 
1098 	isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1099 	    SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1100 	    CTLFLAG_RD, 0, "I/O scheduler statistics");
1101 	n = SYSCTL_CHILDREN(isc->sysctl_tree);
1102 	ctx = &isc->sysctl_ctx;
1103 
1104 	cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1105 	cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1106 	cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1107 	cam_iosched_cl_sysctl_init(isc);
1108 
1109 	SYSCTL_ADD_INT(ctx, n,
1110 	    OID_AUTO, "read_bias", CTLFLAG_RW,
1111 	    &isc->read_bias, 100,
1112 	    "How biased towards read should we be independent of limits");
1113 
1114 	SYSCTL_ADD_INT(ctx, n,
1115 	    OID_AUTO, "quanta", CTLFLAG_RW,
1116 	    &isc->quanta, 200,
1117 	    "How many quanta per second do we slice the I/O up into");
1118 
1119 	SYSCTL_ADD_INT(ctx, n,
1120 	    OID_AUTO, "total_ticks", CTLFLAG_RD,
1121 	    &isc->total_ticks, 0,
1122 	    "Total number of ticks we've done");
1123 
1124 	SYSCTL_ADD_INT(ctx, n,
1125 	    OID_AUTO, "load", CTLFLAG_RD,
1126 	    &isc->load, 0,
1127 	    "scaled load average / 100");
1128 #endif
1129 }
1130 
1131 /*
1132  * Flush outstanding I/O. Consumers of this library don't know all the
1133  * queues we may keep, so this allows all I/O to be flushed in one
1134  * convenient call.
1135  */
1136 void
1137 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1138 {
1139 	bioq_flush(&isc->bio_queue, stp, err);
1140 	bioq_flush(&isc->trim_queue, stp, err);
1141 #ifdef CAM_IOSCHED_DYNAMIC
1142 	if (do_dynamic_iosched)
1143 		bioq_flush(&isc->write_queue, stp, err);
1144 #endif
1145 }
1146 
1147 #ifdef CAM_IOSCHED_DYNAMIC
1148 static struct bio *
1149 cam_iosched_get_write(struct cam_iosched_softc *isc)
1150 {
1151 	struct bio *bp;
1152 
1153 	/*
1154 	 * We control the write rate by controlling how many requests we send
1155 	 * down to the drive at any one time. Fewer requests limits the
1156 	 * effects of both starvation when the requests take a while and write
1157 	 * amplification when each request is causing more than one write to
1158 	 * the NAND media. Limiting the queue depth like this will also limit
1159 	 * the write throughput and give and reads that want to compete to
1160 	 * compete unfairly.
1161 	 */
1162 	bp = bioq_first(&isc->write_queue);
1163 	if (bp == NULL) {
1164 		if (iosched_debug > 3)
1165 			printf("No writes present in write_queue\n");
1166 		return NULL;
1167 	}
1168 
1169 	/*
1170 	 * If pending read, prefer that based on current read bias
1171 	 * setting.
1172 	 */
1173 	if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1174 		if (iosched_debug)
1175 			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);
1176 		isc->current_read_bias--;
1177 		/* We're not limiting writes, per se, just doing reads first */
1178 		return NULL;
1179 	}
1180 
1181 	/*
1182 	 * See if our current limiter allows this I/O.
1183 	 */
1184 	if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1185 		if (iosched_debug)
1186 			printf("Can't write because limiter says no.\n");
1187 		isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1188 		return NULL;
1189 	}
1190 
1191 	/*
1192 	 * Let's do this: We've passed all the gates and we're a go
1193 	 * to schedule the I/O in the SIM.
1194 	 */
1195 	isc->current_read_bias = isc->read_bias;
1196 	bioq_remove(&isc->write_queue, bp);
1197 	if (bp->bio_cmd == BIO_WRITE) {
1198 		isc->write_stats.queued--;
1199 		isc->write_stats.total++;
1200 		isc->write_stats.pending++;
1201 	}
1202 	if (iosched_debug > 9)
1203 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1204 	isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1205 	return bp;
1206 }
1207 #endif
1208 
1209 /*
1210  * Put back a trim that you weren't able to actually schedule this time.
1211  */
1212 void
1213 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1214 {
1215 	bioq_insert_head(&isc->trim_queue, bp);
1216 #ifdef CAM_IOSCHED_DYNAMIC
1217 	isc->trim_stats.queued++;
1218 	isc->trim_stats.total--;		/* since we put it back, don't double count */
1219 	isc->trim_stats.pending--;
1220 #endif
1221 }
1222 
1223 /*
1224  * gets the next trim from the trim queue.
1225  *
1226  * Assumes we're called with the periph lock held.  It removes this
1227  * trim from the queue and the device must explicitly reinstert it
1228  * should the need arise.
1229  */
1230 struct bio *
1231 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1232 {
1233 	struct bio *bp;
1234 
1235 	bp  = bioq_first(&isc->trim_queue);
1236 	if (bp == NULL)
1237 		return NULL;
1238 	bioq_remove(&isc->trim_queue, bp);
1239 #ifdef CAM_IOSCHED_DYNAMIC
1240 	isc->trim_stats.queued--;
1241 	isc->trim_stats.total++;
1242 	isc->trim_stats.pending++;
1243 #endif
1244 	return bp;
1245 }
1246 
1247 /*
1248  * gets the an available trim from the trim queue, if there's no trim
1249  * already pending. It removes this trim from the queue and the device
1250  * must explicitly reinstert it should the need arise.
1251  *
1252  * Assumes we're called with the periph lock held.
1253  */
1254 struct bio *
1255 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1256 {
1257 
1258 	if (!cam_iosched_has_more_trim(isc))
1259 		return NULL;
1260 
1261 	return cam_iosched_next_trim(isc);
1262 }
1263 
1264 /*
1265  * Determine what the next bit of work to do is for the periph. The
1266  * default implementation looks to see if we have trims to do, but no
1267  * trims outstanding. If so, we do that. Otherwise we see if we have
1268  * other work. If we do, then we do that. Otherwise why were we called?
1269  */
1270 struct bio *
1271 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1272 {
1273 	struct bio *bp;
1274 
1275 	/*
1276 	 * See if we have a trim that can be scheduled. We can only send one
1277 	 * at a time down, so this takes that into account.
1278 	 *
1279 	 * XXX newer TRIM commands are queueable. Revisit this when we
1280 	 * implement them.
1281 	 */
1282 	if ((bp = cam_iosched_get_trim(isc)) != NULL)
1283 		return bp;
1284 
1285 #ifdef CAM_IOSCHED_DYNAMIC
1286 	/*
1287 	 * See if we have any pending writes, and room in the queue for them,
1288 	 * and if so, those are next.
1289 	 */
1290 	if (do_dynamic_iosched) {
1291 		if ((bp = cam_iosched_get_write(isc)) != NULL)
1292 			return bp;
1293 	}
1294 #endif
1295 
1296 	/*
1297 	 * next, see if there's other, normal I/O waiting. If so return that.
1298 	 */
1299 	if ((bp = bioq_first(&isc->bio_queue)) == NULL)
1300 		return NULL;
1301 
1302 #ifdef CAM_IOSCHED_DYNAMIC
1303 	/*
1304 	 * For the dynamic scheduler, bio_queue is only for reads, so enforce
1305 	 * the limits here. Enforce only for reads.
1306 	 */
1307 	if (do_dynamic_iosched) {
1308 		if (bp->bio_cmd == BIO_READ &&
1309 		    cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1310 			isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1311 			return NULL;
1312 		}
1313 	}
1314 	isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1315 #endif
1316 	bioq_remove(&isc->bio_queue, bp);
1317 #ifdef CAM_IOSCHED_DYNAMIC
1318 	if (do_dynamic_iosched) {
1319 		if (bp->bio_cmd == BIO_READ) {
1320 			isc->read_stats.queued--;
1321 			isc->read_stats.total++;
1322 			isc->read_stats.pending++;
1323 		} else
1324 			printf("Found bio_cmd = %#x\n", bp->bio_cmd);
1325 	}
1326 	if (iosched_debug > 9)
1327 		printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1328 #endif
1329 	return bp;
1330 }
1331 
1332 /*
1333  * Driver has been given some work to do by the block layer. Tell the
1334  * scheduler about it and have it queue the work up. The scheduler module
1335  * will then return the currently most useful bit of work later, possibly
1336  * deferring work for various reasons.
1337  */
1338 void
1339 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1340 {
1341 
1342 	/*
1343 	 * Put all trims on the trim queue sorted, since we know
1344 	 * that the collapsing code requires this. Otherwise put
1345 	 * the work on the bio queue.
1346 	 */
1347 	if (bp->bio_cmd == BIO_DELETE) {
1348 		bioq_disksort(&isc->trim_queue, bp);
1349 #ifdef CAM_IOSCHED_DYNAMIC
1350 		isc->trim_stats.in++;
1351 		isc->trim_stats.queued++;
1352 #endif
1353 	}
1354 #ifdef CAM_IOSCHED_DYNAMIC
1355 	else if (do_dynamic_iosched &&
1356 	    (bp->bio_cmd == BIO_WRITE || bp->bio_cmd == BIO_FLUSH)) {
1357 		if (cam_iosched_sort_queue(isc))
1358 			bioq_disksort(&isc->write_queue, bp);
1359 		else
1360 			bioq_insert_tail(&isc->write_queue, bp);
1361 		if (iosched_debug > 9)
1362 			printf("Qw  : %p %#x\n", bp, bp->bio_cmd);
1363 		if (bp->bio_cmd == BIO_WRITE) {
1364 			isc->write_stats.in++;
1365 			isc->write_stats.queued++;
1366 		}
1367 	}
1368 #endif
1369 	else {
1370 		if (cam_iosched_sort_queue(isc))
1371 			bioq_disksort(&isc->bio_queue, bp);
1372 		else
1373 			bioq_insert_tail(&isc->bio_queue, bp);
1374 #ifdef CAM_IOSCHED_DYNAMIC
1375 		if (iosched_debug > 9)
1376 			printf("Qr  : %p %#x\n", bp, bp->bio_cmd);
1377 		if (bp->bio_cmd == BIO_READ) {
1378 			isc->read_stats.in++;
1379 			isc->read_stats.queued++;
1380 		} else if (bp->bio_cmd == BIO_WRITE) {
1381 			isc->write_stats.in++;
1382 			isc->write_stats.queued++;
1383 		}
1384 #endif
1385 	}
1386 }
1387 
1388 /*
1389  * If we have work, get it scheduled. Called with the periph lock held.
1390  */
1391 void
1392 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1393 {
1394 
1395 	if (cam_iosched_has_work(isc))
1396 		xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1397 }
1398 
1399 /*
1400  * Complete a trim request
1401  */
1402 void
1403 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1404 {
1405 
1406 	isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1407 }
1408 
1409 /*
1410  * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1411  * might use notes in the ccb for statistics.
1412  */
1413 int
1414 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1415     union ccb *done_ccb)
1416 {
1417 	int retval = 0;
1418 #ifdef CAM_IOSCHED_DYNAMIC
1419 	if (!do_dynamic_iosched)
1420 		return retval;
1421 
1422 	if (iosched_debug > 10)
1423 		printf("done: %p %#x\n", bp, bp->bio_cmd);
1424 	if (bp->bio_cmd == BIO_WRITE) {
1425 		retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1426 		isc->write_stats.out++;
1427 		isc->write_stats.pending--;
1428 	} else if (bp->bio_cmd == BIO_READ) {
1429 		retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1430 		isc->read_stats.out++;
1431 		isc->read_stats.pending--;
1432 	} else if (bp->bio_cmd == BIO_DELETE) {
1433 		isc->trim_stats.out++;
1434 		isc->trim_stats.pending--;
1435 	} else if (bp->bio_cmd != BIO_FLUSH) {
1436 		if (iosched_debug)
1437 			printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1438 	}
1439 
1440 	if (!(bp->bio_flags & BIO_ERROR))
1441 		cam_iosched_io_metric_update(isc, done_ccb->ccb_h.qos.sim_data,
1442 		    bp->bio_cmd, bp->bio_bcount);
1443 #endif
1444 	return retval;
1445 }
1446 
1447 /*
1448  * Tell the io scheduler that you've pushed a trim down into the sim.
1449  * xxx better place for this?
1450  */
1451 void
1452 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1453 {
1454 
1455 	isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1456 }
1457 
1458 /*
1459  * Change the sorting policy hint for I/O transactions for this device.
1460  */
1461 void
1462 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1463 {
1464 
1465 	isc->sort_io_queue = val;
1466 }
1467 
1468 int
1469 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1470 {
1471 	return isc->flags & flags;
1472 }
1473 
1474 void
1475 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1476 {
1477 	isc->flags |= flags;
1478 }
1479 
1480 void
1481 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1482 {
1483 	isc->flags &= ~flags;
1484 }
1485 
1486 #ifdef CAM_IOSCHED_DYNAMIC
1487 /*
1488  * After the method presented in Jack Crenshaw's 1998 article "Integer
1489  * Suqare Roots," reprinted at
1490  * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1491  * and well worth the read. Briefly, we find the power of 4 that's the
1492  * largest smaller than val. We then check each smaller power of 4 to
1493  * see if val is still bigger. The right shifts at each step divide
1494  * the result by 2 which after successive application winds up
1495  * accumulating the right answer. It could also have been accumulated
1496  * using a separate root counter, but this code is smaller and faster
1497  * than that method. This method is also integer size invariant.
1498  * It returns floor(sqrt((float)val)), or the larget integer less than
1499  * or equal to the square root.
1500  */
1501 static uint64_t
1502 isqrt64(uint64_t val)
1503 {
1504 	uint64_t res = 0;
1505 	uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1506 
1507 	/*
1508 	 * Find the largest power of 4 smaller than val.
1509 	 */
1510 	while (bit > val)
1511 		bit >>= 2;
1512 
1513 	/*
1514 	 * Accumulate the answer, one bit at a time (we keep moving
1515 	 * them over since 2 is the square root of 4 and we test
1516 	 * powers of 4). We accumulate where we find the bit, but
1517 	 * the successive shifts land the bit in the right place
1518 	 * by the end.
1519 	 */
1520 	while (bit != 0) {
1521 		if (val >= res + bit) {
1522 			val -= res + bit;
1523 			res = (res >> 1) + bit;
1524 		} else
1525 			res >>= 1;
1526 		bit >>= 2;
1527 	}
1528 
1529 	return res;
1530 }
1531 
1532 /*
1533  * a and b are 32.32 fixed point stored in a 64-bit word.
1534  * Let al and bl be the .32 part of a and b.
1535  * Let ah and bh be the 32 part of a and b.
1536  * R is the radix and is 1 << 32
1537  *
1538  * a * b
1539  * (ah + al / R) * (bh + bl / R)
1540  * ah * bh + (al * bh + ah * bl) / R + al * bl / R^2
1541  *
1542  * After multiplicaiton, we have to renormalize by multiply by
1543  * R, so we wind up with
1544  *	ah * bh * R + al * bh + ah * bl + al * bl / R
1545  * which turns out to be a very nice way to compute this value
1546  * so long as ah and bh are < 65536 there's no loss of high bits
1547  * and the low order bits are below the threshold of caring for
1548  * this application.
1549  */
1550 static uint64_t
1551 mul(uint64_t a, uint64_t b)
1552 {
1553 	uint64_t al, ah, bl, bh;
1554 	al = a & 0xffffffff;
1555 	ah = a >> 32;
1556 	bl = b & 0xffffffff;
1557 	bh = b >> 32;
1558 	return ((ah * bh) << 32) + al * bh + ah * bl + ((al * bl) >> 32);
1559 }
1560 
1561 static sbintime_t latencies[] = {
1562 	SBT_1MS <<  0,
1563 	SBT_1MS <<  1,
1564 	SBT_1MS <<  2,
1565 	SBT_1MS <<  3,
1566 	SBT_1MS <<  4,
1567 	SBT_1MS <<  5,
1568 	SBT_1MS <<  6,
1569 	SBT_1MS <<  7,
1570 	SBT_1MS <<  8,
1571 	SBT_1MS <<  9,
1572 	SBT_1MS << 10
1573 };
1574 
1575 static void
1576 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1577 {
1578 	sbintime_t y, yy;
1579 	uint64_t var;
1580 	int i;
1581 
1582 	/*
1583 	 * Keep counts for latency. We do it by power of two buckets.
1584 	 * This helps us spot outlier behavior obscured by averages.
1585 	 */
1586 	for (i = 0; i < LAT_BUCKETS - 1; i++) {
1587 		if (sim_latency < latencies[i]) {
1588 			iop->latencies[i]++;
1589 			break;
1590 		}
1591 	}
1592 	if (i == LAT_BUCKETS - 1)
1593 		iop->latencies[i]++; 	 /* Put all > 1024ms values into the last bucket. */
1594 
1595 	/*
1596 	 * Classic expoentially decaying average with a tiny alpha
1597 	 * (2 ^ -alpha_bits). For more info see the NIST statistical
1598 	 * handbook.
1599 	 *
1600 	 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha)
1601 	 * alpha = 1 / (1 << alpha_bits)
1602 	 *
1603 	 * Since alpha is a power of two, we can compute this w/o any mult or
1604 	 * division.
1605 	 */
1606 	y = sim_latency;
1607 	iop->ema = (y + (iop->ema << alpha_bits) - iop->ema) >> alpha_bits;
1608 
1609 	yy = mul(y, y);
1610 	iop->emss = (yy + (iop->emss << alpha_bits) - iop->emss) >> alpha_bits;
1611 
1612 	/*
1613          * s_1 = sum of data
1614 	 * s_2 = sum of data * data
1615 	 * ema ~ mean (or s_1 / N)
1616 	 * emss ~ s_2 / N
1617 	 *
1618 	 * sd = sqrt((N * s_2 - s_1 ^ 2) / (N * (N - 1)))
1619 	 * sd = sqrt((N * s_2 / N * (N - 1)) - (s_1 ^ 2 / (N * (N - 1))))
1620 	 *
1621 	 * N ~ 2 / alpha - 1
1622 	 * alpha < 1 / 16 (typically much less)
1623 	 * N > 31 --> N large so N * (N - 1) is approx N * N
1624 	 *
1625 	 * substituting and rearranging:
1626 	 * sd ~ sqrt(s_2 / N - (s_1 / N) ^ 2)
1627 	 *    ~ sqrt(emss - ema ^ 2);
1628 	 * which is the formula used here to get a decent estimate of sd which
1629 	 * we use to detect outliers. Note that when first starting up, it
1630 	 * takes a while for emss sum of squares estimator to converge on a
1631 	 * good value.  during this time, it can be less than ema^2. We
1632 	 * compute a sd of 0 in that case, and ignore outliers.
1633 	 */
1634 	var = iop->emss - mul(iop->ema, iop->ema);
1635 	iop->sd = (int64_t)var < 0 ? 0 : isqrt64(var);
1636 }
1637 
1638 #ifdef CAM_IOSCHED_DYNAMIC
1639 static void
1640 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1641     sbintime_t sim_latency, int cmd, size_t size)
1642 {
1643 	/* xxx Do we need to scale based on the size of the I/O ? */
1644 	switch (cmd) {
1645 	case BIO_READ:
1646 		cam_iosched_update(&isc->read_stats, sim_latency);
1647 		break;
1648 	case BIO_WRITE:
1649 		cam_iosched_update(&isc->write_stats, sim_latency);
1650 		break;
1651 	case BIO_DELETE:
1652 		cam_iosched_update(&isc->trim_stats, sim_latency);
1653 		break;
1654 	default:
1655 		break;
1656 	}
1657 }
1658 #endif
1659 
1660 #ifdef DDB
1661 static int biolen(struct bio_queue_head *bq)
1662 {
1663 	int i = 0;
1664 	struct bio *bp;
1665 
1666 	TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
1667 		i++;
1668 	}
1669 	return i;
1670 }
1671 
1672 /*
1673  * Show the internal state of the I/O scheduler.
1674  */
1675 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
1676 {
1677 	struct cam_iosched_softc *isc;
1678 
1679 	if (!have_addr) {
1680 		db_printf("Need addr\n");
1681 		return;
1682 	}
1683 	isc = (struct cam_iosched_softc *)addr;
1684 	db_printf("pending_reads:     %d\n", isc->read_stats.pending);
1685 	db_printf("min_reads:         %d\n", isc->read_stats.min);
1686 	db_printf("max_reads:         %d\n", isc->read_stats.max);
1687 	db_printf("reads:             %d\n", isc->read_stats.total);
1688 	db_printf("in_reads:          %d\n", isc->read_stats.in);
1689 	db_printf("out_reads:         %d\n", isc->read_stats.out);
1690 	db_printf("queued_reads:      %d\n", isc->read_stats.queued);
1691 	db_printf("Current Q len      %d\n", biolen(&isc->bio_queue));
1692 	db_printf("pending_writes:    %d\n", isc->write_stats.pending);
1693 	db_printf("min_writes:        %d\n", isc->write_stats.min);
1694 	db_printf("max_writes:        %d\n", isc->write_stats.max);
1695 	db_printf("writes:            %d\n", isc->write_stats.total);
1696 	db_printf("in_writes:         %d\n", isc->write_stats.in);
1697 	db_printf("out_writes:        %d\n", isc->write_stats.out);
1698 	db_printf("queued_writes:     %d\n", isc->write_stats.queued);
1699 	db_printf("Current Q len      %d\n", biolen(&isc->write_queue));
1700 	db_printf("pending_trims:     %d\n", isc->trim_stats.pending);
1701 	db_printf("min_trims:         %d\n", isc->trim_stats.min);
1702 	db_printf("max_trims:         %d\n", isc->trim_stats.max);
1703 	db_printf("trims:             %d\n", isc->trim_stats.total);
1704 	db_printf("in_trims:          %d\n", isc->trim_stats.in);
1705 	db_printf("out_trims:         %d\n", isc->trim_stats.out);
1706 	db_printf("queued_trims:      %d\n", isc->trim_stats.queued);
1707 	db_printf("Current Q len      %d\n", biolen(&isc->trim_queue));
1708 	db_printf("read_bias:         %d\n", isc->read_bias);
1709 	db_printf("current_read_bias: %d\n", isc->current_read_bias);
1710 	db_printf("Trim active?       %s\n",
1711 	    (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");
1712 }
1713 #endif
1714 #endif
1715