2 * CAM IO Scheduler Interface
4 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
6 * Copyright (c) 2015 Netflix, Inc.
9 * Redistribution and use in source and binary forms, with or without
10 * modification, are permitted provided that the following conditions
12 * 1. Redistributions of source code must retain the above copyright
13 * notice, this list of conditions, and the following disclaimer,
14 * without modification, immediately at the beginning of the file.
15 * 2. The name of the author may not be used to endorse or promote products
16 * derived from this software without specific prior written permission.
18 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
19 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
21 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
22 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
24 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
25 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
26 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
27 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36 #include <sys/cdefs.h>
37 __FBSDID("$FreeBSD$");
39 #include <sys/param.h>
41 #include <sys/systm.h>
42 #include <sys/kernel.h>
45 #include <sys/malloc.h>
46 #include <sys/mutex.h>
48 #include <sys/sysctl.h>
51 #include <cam/cam_ccb.h>
52 #include <cam/cam_periph.h>
53 #include <cam/cam_xpt_periph.h>
54 #include <cam/cam_xpt_internal.h>
55 #include <cam/cam_iosched.h>
59 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler",
60 "CAM I/O Scheduler buffers");
63 * Trim or similar currently pending completion. Should only be set for
64 * those drivers wishing only one Trim active at a time.
66 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE (1ul << 0)
67 /* Callout active, and needs to be torn down */
68 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
69 /* Timer has just ticked */
70 #define CAM_IOSCHED_FLAG_TICK (1ul << 2)
72 /* Periph drivers set these flags to indicate work */
73 #define CAM_IOSCHED_FLAG_WORK_FLAGS ((0xffffu) << 16)
76 * Default I/O scheduler for FreeBSD. This implementation is just a thin-vineer
77 * over the bioq_* interface, with notions of separate calls for normal I/O and
80 * When CAM_IOSCHED_DYNAMIC is defined, the scheduler is enhanced to dynamically
81 * steer the rate of one type of traffic to help other types of traffic (eg
82 * limit writes when read latency deteriorates on SSDs).
85 #ifdef CAM_IOSCHED_DYNAMIC
87 static int do_dynamic_iosched = 1;
88 TUNABLE_INT("kern.cam.do_dynamic_iosched", &do_dynamic_iosched);
89 SYSCTL_INT(_kern_cam, OID_AUTO, do_dynamic_iosched, CTLFLAG_RD,
90 &do_dynamic_iosched, 1,
91 "Enable Dynamic I/O scheduler optimizations.");
94 * For an EMA, with an alpha of alpha, we know
98 * where N is the number of samples that 86% of the current
99 * EMA is derived from.
101 * So we invent[*] alpha_bits:
102 * alpha_bits = -log_2(alpha)
103 * alpha = 2^-alpha_bits
105 * N = 1 + 2^(alpha_bits + 1)
107 * The default 9 gives a 1025 lookback for 86% of the data.
108 * For a brief intro: https://en.wikipedia.org/wiki/Moving_average
110 * [*] Steal from the load average code and many other places.
111 * Note: See computation of EMA and EMVAR for acceptable ranges of alpha.
113 static int alpha_bits = 9;
114 TUNABLE_INT("kern.cam.iosched_alpha_bits", &alpha_bits);
115 SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW,
117 "Bits in EMA's alpha.");
120 struct cam_iosched_softc;
122 int iosched_debug = 0;
125 none = 0, /* No limits */
126 queue_depth, /* Limit how many ops we queue to SIM */
127 iops, /* Limit # of IOPS to the drive */
128 bandwidth, /* Limit bandwidth to the drive */
132 static const char *cam_iosched_limiter_names[] =
133 { "none", "queue_depth", "iops", "bandwidth" };
136 * Called to initialize the bits of the iop_stats structure relevant to the
137 * limiter. Called just after the limiter is set.
139 typedef int l_init_t(struct iop_stats *);
144 typedef int l_tick_t(struct iop_stats *);
147 * Called to see if the limiter thinks this IOP can be allowed to
148 * proceed. If so, the limiter assumes that the IOP proceeded
149 * and makes any accounting of it that's needed.
151 typedef int l_iop_t(struct iop_stats *, struct bio *);
154 * Called when an I/O completes so the limiter can update its
155 * accounting. Pending I/Os may complete in any order (even when
156 * sent to the hardware at the same time), so the limiter may not
157 * make any assumptions other than this I/O has completed. If it
158 * returns 1, then xpt_schedule() needs to be called again.
160 typedef int l_iodone_t(struct iop_stats *, struct bio *);
162 static l_iop_t cam_iosched_qd_iop;
163 static l_iop_t cam_iosched_qd_caniop;
164 static l_iodone_t cam_iosched_qd_iodone;
166 static l_init_t cam_iosched_iops_init;
167 static l_tick_t cam_iosched_iops_tick;
168 static l_iop_t cam_iosched_iops_caniop;
169 static l_iop_t cam_iosched_iops_iop;
171 static l_init_t cam_iosched_bw_init;
172 static l_tick_t cam_iosched_bw_tick;
173 static l_iop_t cam_iosched_bw_caniop;
174 static l_iop_t cam_iosched_bw_iop;
181 l_iodone_t *l_iodone;
193 .l_caniop = cam_iosched_qd_caniop,
194 .l_iop = cam_iosched_qd_iop,
195 .l_iodone= cam_iosched_qd_iodone,
198 .l_init = cam_iosched_iops_init,
199 .l_tick = cam_iosched_iops_tick,
200 .l_caniop = cam_iosched_iops_caniop,
201 .l_iop = cam_iosched_iops_iop,
205 .l_init = cam_iosched_bw_init,
206 .l_tick = cam_iosched_bw_tick,
207 .l_caniop = cam_iosched_bw_caniop,
208 .l_iop = cam_iosched_bw_iop,
215 * sysctl state for this subnode.
217 struct sysctl_ctx_list sysctl_ctx;
218 struct sysctl_oid *sysctl_tree;
221 * Information about the current rate limiters, if any
223 io_limiter limiter; /* How are I/Os being limited */
224 int min; /* Low range of limit */
225 int max; /* High range of limit */
226 int current; /* Current rate limiter */
227 int l_value1; /* per-limiter scratch value 1. */
228 int l_value2; /* per-limiter scratch value 2. */
231 * Debug information about counts of I/Os that have gone through the
234 int pending; /* I/Os pending in the hardware */
235 int queued; /* number currently in the queue */
236 int total; /* Total for all time -- wraps */
237 int in; /* number queued all time -- wraps */
238 int out; /* number completed all time -- wraps */
239 int errs; /* Number of I/Os completed with error -- wraps */
242 * Statistics on different bits of the process.
244 /* Exp Moving Average, see alpha_bits for more details */
247 sbintime_t sd; /* Last computed sd */
249 uint32_t state_flags;
250 #define IOP_RATE_LIMITED 1u
252 #define LAT_BUCKETS 15 /* < 1ms < 2ms ... < 2^(n-1)ms >= 2^(n-1)ms*/
253 uint64_t latencies[LAT_BUCKETS];
255 struct cam_iosched_softc *softc;
260 set_max = 0, /* current = max */
261 read_latency, /* Steer read latency by throttling writes */
262 cl_max /* Keep last */
265 static const char *cam_iosched_control_type_names[] =
266 { "set_max", "read_latency" };
268 struct control_loop {
270 * sysctl state for this subnode.
272 struct sysctl_ctx_list sysctl_ctx;
273 struct sysctl_oid *sysctl_tree;
275 sbintime_t next_steer; /* Time of next steer */
276 sbintime_t steer_interval; /* How often do we steer? */
280 control_type type; /* What type of control? */
281 int last_count; /* Last I/O count */
283 struct cam_iosched_softc *softc;
288 struct cam_iosched_softc {
289 struct bio_queue_head bio_queue;
290 struct bio_queue_head trim_queue;
291 /* scheduler flags < 16, user flags >= 16 */
294 #ifdef CAM_IOSCHED_DYNAMIC
295 int read_bias; /* Read bias setting */
296 int current_read_bias; /* Current read bias state */
298 int load; /* EMA of 'load average' of disk / 2^16 */
300 struct bio_queue_head write_queue;
301 struct iop_stats read_stats, write_stats, trim_stats;
302 struct sysctl_ctx_list sysctl_ctx;
303 struct sysctl_oid *sysctl_tree;
305 int quanta; /* Number of quanta per second */
306 struct callout ticker; /* Callout for our quota system */
307 struct cam_periph *periph; /* cam periph associated with this device */
308 uint32_t this_frac; /* Fraction of a second (1024ths) for this tick */
309 sbintime_t last_time; /* Last time we ticked */
310 struct control_loop cl;
314 #ifdef CAM_IOSCHED_DYNAMIC
316 * helper functions to call the limsw functions.
319 cam_iosched_limiter_init(struct iop_stats *ios)
321 int lim = ios->limiter;
323 /* maybe this should be a kassert */
324 if (lim < none || lim >= limiter_max)
327 if (limsw[lim].l_init)
328 return limsw[lim].l_init(ios);
334 cam_iosched_limiter_tick(struct iop_stats *ios)
336 int lim = ios->limiter;
338 /* maybe this should be a kassert */
339 if (lim < none || lim >= limiter_max)
342 if (limsw[lim].l_tick)
343 return limsw[lim].l_tick(ios);
349 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
351 int lim = ios->limiter;
353 /* maybe this should be a kassert */
354 if (lim < none || lim >= limiter_max)
357 if (limsw[lim].l_iop)
358 return limsw[lim].l_iop(ios, bp);
364 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
366 int lim = ios->limiter;
368 /* maybe this should be a kassert */
369 if (lim < none || lim >= limiter_max)
372 if (limsw[lim].l_caniop)
373 return limsw[lim].l_caniop(ios, bp);
379 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
381 int lim = ios->limiter;
383 /* maybe this should be a kassert */
384 if (lim < none || lim >= limiter_max)
387 if (limsw[lim].l_iodone)
388 return limsw[lim].l_iodone(ios, bp);
394 * Functions to implement the different kinds of limiters
398 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
401 if (ios->current <= 0 || ios->pending < ios->current)
408 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
411 if (ios->current <= 0 || ios->pending < ios->current)
418 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
421 if (ios->current <= 0 || ios->pending != ios->current)
428 cam_iosched_iops_init(struct iop_stats *ios)
431 ios->l_value1 = ios->current / ios->softc->quanta;
432 if (ios->l_value1 <= 0)
440 cam_iosched_iops_tick(struct iop_stats *ios)
445 * Allow at least one IO per tick until all
446 * the IOs for this interval have been spent.
448 new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
449 if (new_ios < 1 && ios->l_value2 < ios->current) {
455 * If this a new accounting interval, discard any "unspent" ios
456 * granted in the previous interval. Otherwise add the new ios to
457 * the previously granted ones that haven't been spent yet.
459 if ((ios->softc->total_ticks % ios->softc->quanta) == 0) {
460 ios->l_value1 = new_ios;
463 ios->l_value1 += new_ios;
471 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
475 * So if we have any more IOPs left, allow it,
476 * otherwise wait. If current iops is 0, treat that
477 * as unlimited as a failsafe.
479 if (ios->current > 0 && ios->l_value1 <= 0)
485 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
489 rv = cam_iosched_limiter_caniop(ios, bp);
497 cam_iosched_bw_init(struct iop_stats *ios)
500 /* ios->current is in kB/s, so scale to bytes */
501 ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
507 cam_iosched_bw_tick(struct iop_stats *ios)
512 * If we're in the hole for available quota from
513 * the last time, then add the quantum for this.
514 * If we have any left over from last quantum,
515 * then too bad, that's lost. Also, ios->current
516 * is in kB/s, so scale.
518 * We also allow up to 4 quanta of credits to
519 * accumulate to deal with burstiness. 4 is extremely
522 bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
523 if (ios->l_value1 < bw * 4)
530 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
533 * So if we have any more bw quota left, allow it,
534 * otherwise wait. Note, we'll go negative and that's
535 * OK. We'll just get a little less next quota.
537 * Note on going negative: that allows us to process
538 * requests in order better, since we won't allow
539 * shorter reads to get around the long one that we
540 * don't have the quota to do just yet. It also prevents
541 * starvation by being a little more permissive about
542 * what we let through this quantum (to prevent the
543 * starvation), at the cost of getting a little less
546 * Also note that if the current limit is <= 0,
547 * we treat it as unlimited as a failsafe.
549 if (ios->current > 0 && ios->l_value1 <= 0)
557 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
561 rv = cam_iosched_limiter_caniop(ios, bp);
563 ios->l_value1 -= bp->bio_length;
568 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
571 cam_iosched_ticker(void *arg)
573 struct cam_iosched_softc *isc = arg;
574 sbintime_t now, delta;
577 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
580 delta = now - isc->last_time;
581 isc->this_frac = (uint32_t)delta >> 16; /* Note: discards seconds -- should be 0 harmless if not */
582 isc->last_time = now;
584 cam_iosched_cl_maybe_steer(&isc->cl);
586 cam_iosched_limiter_tick(&isc->read_stats);
587 cam_iosched_limiter_tick(&isc->write_stats);
588 cam_iosched_limiter_tick(&isc->trim_stats);
590 isc->flags |= CAM_IOSCHED_FLAGS_TICK;
591 cam_iosched_schedule(isc, isc->periph);
594 * isc->load is an EMA of the pending I/Os at each tick. The number of
595 * pending I/Os is the sum of the I/Os queued to the hardware, and those
596 * in the software queue that could be queued to the hardware if there
599 * ios_stats.pending is a count of requests in the SIM right now for
600 * each of these types of I/O. So the total pending count is the sum of
601 * these I/Os and the sum of the queued I/Os still in the software queue
602 * for those operations that aren't being rate limited at the moment.
604 * The reason for the rate limiting bit is because those I/Os
605 * aren't part of the software queued load (since we could
606 * give them to hardware, but choose not to).
608 * Note: due to a bug in counting pending TRIM in the device, we
609 * don't include them in this count. We count each BIO_DELETE in
610 * the pending count, but the periph drivers collapse them down
611 * into one TRIM command. That one trim command gets the completion
612 * so the counts get off.
614 pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
615 pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
616 !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
617 !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
619 pending /= isc->periph->path->device->ccbq.total_openings;
621 isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
628 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
631 clp->next_steer = sbinuptime();
633 clp->steer_interval = SBT_1S * 5; /* Let's start out steering every 5s */
634 clp->lolat = 5 * SBT_1MS;
635 clp->hilat = 15 * SBT_1MS;
636 clp->alpha = 20; /* Alpha == gain. 20 = .2 */
641 cam_iosched_cl_maybe_steer(struct control_loop *clp)
643 struct cam_iosched_softc *isc;
648 now = isc->last_time;
649 if (now < clp->next_steer)
652 clp->next_steer = now + clp->steer_interval;
655 if (isc->write_stats.current != isc->write_stats.max)
656 printf("Steering write from %d kBps to %d kBps\n",
657 isc->write_stats.current, isc->write_stats.max);
658 isc->read_stats.current = isc->read_stats.max;
659 isc->write_stats.current = isc->write_stats.max;
660 isc->trim_stats.current = isc->trim_stats.max;
663 old = isc->write_stats.current;
664 lat = isc->read_stats.ema;
666 * Simple PLL-like engine. Since we're steering to a range for
667 * the SP (set point) that makes things a little more
668 * complicated. In addition, we're not directly controlling our
669 * PV (process variable), the read latency, but instead are
670 * manipulating the write bandwidth limit for our MV
671 * (manipulation variable), analysis of this code gets a bit
672 * messy. Also, the MV is a very noisy control surface for read
673 * latency since it is affected by many hidden processes inside
674 * the device which change how responsive read latency will be
675 * in reaction to changes in write bandwidth. Unlike the classic
676 * boiler control PLL. this may result in over-steering while
677 * the SSD takes its time to react to the new, lower load. This
678 * is why we use a relatively low alpha of between .1 and .25 to
679 * compensate for this effect. At .1, it takes ~22 steering
680 * intervals to back off by a factor of 10. At .2 it only takes
681 * ~10. At .25 it only takes ~8. However some preliminary data
682 * from the SSD drives suggests a reasponse time in 10's of
683 * seconds before latency drops regardless of the new write
684 * rate. Careful observation will be required to tune this
687 * Also, when there's no read traffic, we jack up the write
688 * limit too regardless of the last read latency. 10 is
689 * somewhat arbitrary.
691 if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
692 isc->write_stats.current = isc->write_stats.current *
693 (100 + clp->alpha) / 100; /* Scale up */
694 else if (lat > clp->hilat)
695 isc->write_stats.current = isc->write_stats.current *
696 (100 - clp->alpha) / 100; /* Scale down */
697 clp->last_count = isc->read_stats.total;
700 * Even if we don't steer, per se, enforce the min/max limits as
701 * those may have changed.
703 if (isc->write_stats.current < isc->write_stats.min)
704 isc->write_stats.current = isc->write_stats.min;
705 if (isc->write_stats.current > isc->write_stats.max)
706 isc->write_stats.current = isc->write_stats.max;
707 if (old != isc->write_stats.current && iosched_debug)
708 printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
709 old, isc->write_stats.current,
710 (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
718 #ifdef CAM_IOSCHED_DYNAMIC
720 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
721 sbintime_t sim_latency, int cmd, size_t size);
725 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
727 return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
731 cam_iosched_has_io(struct cam_iosched_softc *isc)
733 #ifdef CAM_IOSCHED_DYNAMIC
734 if (do_dynamic_iosched) {
735 struct bio *rbp = bioq_first(&isc->bio_queue);
736 struct bio *wbp = bioq_first(&isc->write_queue);
737 int can_write = wbp != NULL &&
738 cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
739 int can_read = rbp != NULL &&
740 cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
741 if (iosched_debug > 2) {
742 printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
743 printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
744 printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
746 return can_read || can_write;
749 return bioq_first(&isc->bio_queue) != NULL;
753 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
755 return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) &&
756 bioq_first(&isc->trim_queue);
759 #define cam_iosched_sort_queue(isc) ((isc)->sort_io_queue >= 0 ? \
760 (isc)->sort_io_queue : cam_sort_io_queues)
764 cam_iosched_has_work(struct cam_iosched_softc *isc)
766 #ifdef CAM_IOSCHED_DYNAMIC
767 if (iosched_debug > 2)
768 printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
769 cam_iosched_has_more_trim(isc),
770 cam_iosched_has_flagged_work(isc));
773 return cam_iosched_has_io(isc) ||
774 cam_iosched_has_more_trim(isc) ||
775 cam_iosched_has_flagged_work(isc);
778 #ifdef CAM_IOSCHED_DYNAMIC
780 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
785 ios->max = ios->current = 300000;
795 cam_iosched_limiter_init(ios);
799 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
802 struct iop_stats *ios;
803 struct cam_iosched_softc *isc;
809 value = ios->limiter;
810 if (value < none || value >= limiter_max)
813 p = cam_iosched_limiter_names[value];
815 strlcpy(buf, p, sizeof(buf));
816 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
817 if (error != 0 || req->newptr == NULL)
820 cam_periph_lock(isc->periph);
822 for (i = none; i < limiter_max; i++) {
823 if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
826 error = cam_iosched_limiter_init(ios);
828 ios->limiter = value;
829 cam_periph_unlock(isc->periph);
832 /* Note: disk load averate requires ticker to be always running */
833 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
834 isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
836 cam_periph_unlock(isc->periph);
840 cam_periph_unlock(isc->periph);
845 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
848 struct control_loop *clp;
849 struct cam_iosched_softc *isc;
856 if (value < none || value >= cl_max)
859 p = cam_iosched_control_type_names[value];
861 strlcpy(buf, p, sizeof(buf));
862 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
863 if (error != 0 || req->newptr == NULL)
866 for (i = set_max; i < cl_max; i++) {
867 if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
869 cam_periph_lock(isc->periph);
871 cam_periph_unlock(isc->periph);
879 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
886 value = *(sbintime_t *)arg1;
887 us = (uint64_t)value / SBT_1US;
888 snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
889 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
890 if (error != 0 || req->newptr == NULL)
892 us = strtoul(buf, NULL, 10);
895 *(sbintime_t *)arg1 = us * SBT_1US;
900 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
907 sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
909 for (i = 0; i < LAT_BUCKETS - 1; i++)
910 sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
911 sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
912 error = sbuf_finish(&sb);
919 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS)
924 quanta = (unsigned *)arg1;
927 error = sysctl_handle_int(oidp, (int *)&value, 0, req);
928 if ((error != 0) || (req->newptr == NULL))
931 if (value < 1 || value > hz)
940 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
942 struct sysctl_oid_list *n;
943 struct sysctl_ctx_list *ctx;
945 ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
946 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
947 CTLFLAG_RD, 0, name);
948 n = SYSCTL_CHILDREN(ios->sysctl_tree);
949 ctx = &ios->sysctl_ctx;
951 SYSCTL_ADD_UQUAD(ctx, n,
952 OID_AUTO, "ema", CTLFLAG_RD,
954 "Fast Exponentially Weighted Moving Average");
955 SYSCTL_ADD_UQUAD(ctx, n,
956 OID_AUTO, "emvar", CTLFLAG_RD,
958 "Fast Exponentially Weighted Moving Variance");
960 SYSCTL_ADD_INT(ctx, n,
961 OID_AUTO, "pending", CTLFLAG_RD,
963 "Instantaneous # of pending transactions");
964 SYSCTL_ADD_INT(ctx, n,
965 OID_AUTO, "count", CTLFLAG_RD,
967 "# of transactions submitted to hardware");
968 SYSCTL_ADD_INT(ctx, n,
969 OID_AUTO, "queued", CTLFLAG_RD,
971 "# of transactions in the queue");
972 SYSCTL_ADD_INT(ctx, n,
973 OID_AUTO, "in", CTLFLAG_RD,
975 "# of transactions queued to driver");
976 SYSCTL_ADD_INT(ctx, n,
977 OID_AUTO, "out", CTLFLAG_RD,
979 "# of transactions completed (including with error)");
980 SYSCTL_ADD_INT(ctx, n,
981 OID_AUTO, "errs", CTLFLAG_RD,
983 "# of transactions completed with an error");
985 SYSCTL_ADD_PROC(ctx, n,
986 OID_AUTO, "limiter", CTLTYPE_STRING | CTLFLAG_RW,
987 ios, 0, cam_iosched_limiter_sysctl, "A",
988 "Current limiting type.");
989 SYSCTL_ADD_INT(ctx, n,
990 OID_AUTO, "min", CTLFLAG_RW,
993 SYSCTL_ADD_INT(ctx, n,
994 OID_AUTO, "max", CTLFLAG_RW,
997 SYSCTL_ADD_INT(ctx, n,
998 OID_AUTO, "current", CTLFLAG_RW,
1000 "current resource");
1002 SYSCTL_ADD_PROC(ctx, n,
1003 OID_AUTO, "latencies", CTLTYPE_STRING | CTLFLAG_RD,
1005 cam_iosched_sysctl_latencies, "A",
1006 "Array of power of 2 latency from 1ms to 1.024s");
1010 cam_iosched_iop_stats_fini(struct iop_stats *ios)
1012 if (ios->sysctl_tree)
1013 if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
1014 printf("can't remove iosched sysctl stats context\n");
1018 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
1020 struct sysctl_oid_list *n;
1021 struct sysctl_ctx_list *ctx;
1022 struct control_loop *clp;
1025 clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1026 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
1027 CTLFLAG_RD, 0, "Control loop info");
1028 n = SYSCTL_CHILDREN(clp->sysctl_tree);
1029 ctx = &clp->sysctl_ctx;
1031 SYSCTL_ADD_PROC(ctx, n,
1032 OID_AUTO, "type", CTLTYPE_STRING | CTLFLAG_RW,
1033 clp, 0, cam_iosched_control_type_sysctl, "A",
1034 "Control loop algorithm");
1035 SYSCTL_ADD_PROC(ctx, n,
1036 OID_AUTO, "steer_interval", CTLTYPE_STRING | CTLFLAG_RW,
1037 &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
1038 "How often to steer (in us)");
1039 SYSCTL_ADD_PROC(ctx, n,
1040 OID_AUTO, "lolat", CTLTYPE_STRING | CTLFLAG_RW,
1041 &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
1042 "Low water mark for Latency (in us)");
1043 SYSCTL_ADD_PROC(ctx, n,
1044 OID_AUTO, "hilat", CTLTYPE_STRING | CTLFLAG_RW,
1045 &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
1046 "Hi water mark for Latency (in us)");
1047 SYSCTL_ADD_INT(ctx, n,
1048 OID_AUTO, "alpha", CTLFLAG_RW,
1050 "Alpha for PLL (x100) aka gain");
1054 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1056 if (clp->sysctl_tree)
1057 if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1058 printf("can't remove iosched sysctl control loop context\n");
1063 * Allocate the iosched structure. This also insulates callers from knowing
1064 * sizeof struct cam_iosched_softc.
1067 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1070 *iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1073 #ifdef CAM_IOSCHED_DYNAMIC
1075 printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1077 (*iscp)->sort_io_queue = -1;
1078 bioq_init(&(*iscp)->bio_queue);
1079 bioq_init(&(*iscp)->trim_queue);
1080 #ifdef CAM_IOSCHED_DYNAMIC
1081 if (do_dynamic_iosched) {
1082 bioq_init(&(*iscp)->write_queue);
1083 (*iscp)->read_bias = 100;
1084 (*iscp)->current_read_bias = 100;
1085 (*iscp)->quanta = min(hz, 200);
1086 cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1087 cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1088 cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1089 (*iscp)->trim_stats.max = 1; /* Trims are special: one at a time for now */
1090 (*iscp)->last_time = sbinuptime();
1091 callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1092 (*iscp)->periph = periph;
1093 cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1094 callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp);
1095 (*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1103 * Reclaim all used resources. This assumes that other folks have
1104 * drained the requests in the hardware. Maybe an unwise assumption.
1107 cam_iosched_fini(struct cam_iosched_softc *isc)
1110 cam_iosched_flush(isc, NULL, ENXIO);
1111 #ifdef CAM_IOSCHED_DYNAMIC
1112 cam_iosched_iop_stats_fini(&isc->read_stats);
1113 cam_iosched_iop_stats_fini(&isc->write_stats);
1114 cam_iosched_iop_stats_fini(&isc->trim_stats);
1115 cam_iosched_cl_sysctl_fini(&isc->cl);
1116 if (isc->sysctl_tree)
1117 if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1118 printf("can't remove iosched sysctl stats context\n");
1119 if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1120 callout_drain(&isc->ticker);
1121 isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1124 free(isc, M_CAMSCHED);
1129 * After we're sure we're attaching a device, go ahead and add
1130 * hooks for any sysctl we may wish to honor.
1132 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1133 struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1135 #ifdef CAM_IOSCHED_DYNAMIC
1136 struct sysctl_oid_list *n;
1139 SYSCTL_ADD_INT(ctx, SYSCTL_CHILDREN(node),
1140 OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1141 &isc->sort_io_queue, 0,
1142 "Sort IO queue to try and optimise disk access patterns");
1144 #ifdef CAM_IOSCHED_DYNAMIC
1145 if (!do_dynamic_iosched)
1148 isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1149 SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1150 CTLFLAG_RD, 0, "I/O scheduler statistics");
1151 n = SYSCTL_CHILDREN(isc->sysctl_tree);
1152 ctx = &isc->sysctl_ctx;
1154 cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1155 cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1156 cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1157 cam_iosched_cl_sysctl_init(isc);
1159 SYSCTL_ADD_INT(ctx, n,
1160 OID_AUTO, "read_bias", CTLFLAG_RW,
1161 &isc->read_bias, 100,
1162 "How biased towards read should we be independent of limits");
1164 SYSCTL_ADD_PROC(ctx, n,
1165 OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW,
1166 &isc->quanta, 0, cam_iosched_quanta_sysctl, "I",
1167 "How many quanta per second do we slice the I/O up into");
1169 SYSCTL_ADD_INT(ctx, n,
1170 OID_AUTO, "total_ticks", CTLFLAG_RD,
1171 &isc->total_ticks, 0,
1172 "Total number of ticks we've done");
1174 SYSCTL_ADD_INT(ctx, n,
1175 OID_AUTO, "load", CTLFLAG_RD,
1177 "scaled load average / 100");
1182 * Flush outstanding I/O. Consumers of this library don't know all the
1183 * queues we may keep, so this allows all I/O to be flushed in one
1187 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1189 bioq_flush(&isc->bio_queue, stp, err);
1190 bioq_flush(&isc->trim_queue, stp, err);
1191 #ifdef CAM_IOSCHED_DYNAMIC
1192 if (do_dynamic_iosched)
1193 bioq_flush(&isc->write_queue, stp, err);
1197 #ifdef CAM_IOSCHED_DYNAMIC
1199 cam_iosched_get_write(struct cam_iosched_softc *isc)
1204 * We control the write rate by controlling how many requests we send
1205 * down to the drive at any one time. Fewer requests limits the
1206 * effects of both starvation when the requests take a while and write
1207 * amplification when each request is causing more than one write to
1208 * the NAND media. Limiting the queue depth like this will also limit
1209 * the write throughput and give and reads that want to compete to
1212 bp = bioq_first(&isc->write_queue);
1214 if (iosched_debug > 3)
1215 printf("No writes present in write_queue\n");
1220 * If pending read, prefer that based on current read bias
1223 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1226 "Reads present and current_read_bias is %d queued "
1227 "writes %d queued reads %d\n",
1228 isc->current_read_bias, isc->write_stats.queued,
1229 isc->read_stats.queued);
1230 isc->current_read_bias--;
1231 /* We're not limiting writes, per se, just doing reads first */
1236 * See if our current limiter allows this I/O.
1238 if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1240 printf("Can't write because limiter says no.\n");
1241 isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1246 * Let's do this: We've passed all the gates and we're a go
1247 * to schedule the I/O in the SIM.
1249 isc->current_read_bias = isc->read_bias;
1250 bioq_remove(&isc->write_queue, bp);
1251 if (bp->bio_cmd == BIO_WRITE) {
1252 isc->write_stats.queued--;
1253 isc->write_stats.total++;
1254 isc->write_stats.pending++;
1256 if (iosched_debug > 9)
1257 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1258 isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1264 * Put back a trim that you weren't able to actually schedule this time.
1267 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1269 bioq_insert_head(&isc->trim_queue, bp);
1270 #ifdef CAM_IOSCHED_DYNAMIC
1271 isc->trim_stats.queued++;
1272 isc->trim_stats.total--; /* since we put it back, don't double count */
1273 isc->trim_stats.pending--;
1278 * gets the next trim from the trim queue.
1280 * Assumes we're called with the periph lock held. It removes this
1281 * trim from the queue and the device must explicitly reinsert it
1282 * should the need arise.
1285 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1289 bp = bioq_first(&isc->trim_queue);
1292 bioq_remove(&isc->trim_queue, bp);
1293 #ifdef CAM_IOSCHED_DYNAMIC
1294 isc->trim_stats.queued--;
1295 isc->trim_stats.total++;
1296 isc->trim_stats.pending++;
1302 * gets an available trim from the trim queue, if there's no trim
1303 * already pending. It removes this trim from the queue and the device
1304 * must explicitly reinsert it should the need arise.
1306 * Assumes we're called with the periph lock held.
1309 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1312 if (!cam_iosched_has_more_trim(isc))
1315 return cam_iosched_next_trim(isc);
1319 * Determine what the next bit of work to do is for the periph. The
1320 * default implementation looks to see if we have trims to do, but no
1321 * trims outstanding. If so, we do that. Otherwise we see if we have
1322 * other work. If we do, then we do that. Otherwise why were we called?
1325 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1330 wastick = !!(isc->flags & CAM_IOSCHED_FLAGS_TICK);
1331 isc->flags &= ~CAM_IOSCHED_FLAGS_TICK;
1334 * See if we have a trim that can be scheduled. We can only send one
1335 * at a time down, so this takes that into account.
1337 * XXX newer TRIM commands are queueable. Revisit this when we
1340 if ((bp = cam_iosched_get_trim(isc)) != NULL)
1343 #ifdef CAM_IOSCHED_DYNAMIC
1345 * See if we have any pending writes, and room in the queue for them,
1346 * and if so, those are next.
1348 if (do_dynamic_iosched) {
1349 if ((bp = cam_iosched_get_write(isc)) != NULL)
1355 * next, see if there's other, normal I/O waiting. If so return that.
1357 if ((bp = bioq_first(&isc->bio_queue)) == NULL)
1360 #ifdef CAM_IOSCHED_DYNAMIC
1362 * For the dynamic scheduler, bio_queue is only for reads, so enforce
1363 * the limits here. Enforce only for reads.
1365 if (do_dynamic_iosched) {
1366 if (bp->bio_cmd == BIO_READ &&
1367 cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1368 isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1372 isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1374 bioq_remove(&isc->bio_queue, bp);
1375 #ifdef CAM_IOSCHED_DYNAMIC
1376 if (do_dynamic_iosched) {
1377 if (bp->bio_cmd == BIO_READ) {
1378 isc->read_stats.queued--;
1379 isc->read_stats.total++;
1380 isc->read_stats.pending++;
1382 printf("Found bio_cmd = %#x\n", bp->bio_cmd);
1384 if (iosched_debug > 9)
1385 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1391 * Driver has been given some work to do by the block layer. Tell the
1392 * scheduler about it and have it queue the work up. The scheduler module
1393 * will then return the currently most useful bit of work later, possibly
1394 * deferring work for various reasons.
1397 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1401 * Put all trims on the trim queue sorted, since we know
1402 * that the collapsing code requires this. Otherwise put
1403 * the work on the bio queue.
1405 if (bp->bio_cmd == BIO_DELETE) {
1406 bioq_insert_tail(&isc->trim_queue, bp);
1407 #ifdef CAM_IOSCHED_DYNAMIC
1408 isc->trim_stats.in++;
1409 isc->trim_stats.queued++;
1412 #ifdef CAM_IOSCHED_DYNAMIC
1413 else if (do_dynamic_iosched && (bp->bio_cmd != BIO_READ)) {
1414 if (cam_iosched_sort_queue(isc))
1415 bioq_disksort(&isc->write_queue, bp);
1417 bioq_insert_tail(&isc->write_queue, bp);
1418 if (iosched_debug > 9)
1419 printf("Qw : %p %#x\n", bp, bp->bio_cmd);
1420 if (bp->bio_cmd == BIO_WRITE) {
1421 isc->write_stats.in++;
1422 isc->write_stats.queued++;
1427 if (cam_iosched_sort_queue(isc))
1428 bioq_disksort(&isc->bio_queue, bp);
1430 bioq_insert_tail(&isc->bio_queue, bp);
1431 #ifdef CAM_IOSCHED_DYNAMIC
1432 if (iosched_debug > 9)
1433 printf("Qr : %p %#x\n", bp, bp->bio_cmd);
1434 if (bp->bio_cmd == BIO_READ) {
1435 isc->read_stats.in++;
1436 isc->read_stats.queued++;
1437 } else if (bp->bio_cmd == BIO_WRITE) {
1438 isc->write_stats.in++;
1439 isc->write_stats.queued++;
1446 * If we have work, get it scheduled. Called with the periph lock held.
1449 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1452 if (cam_iosched_has_work(isc))
1453 xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1457 * Complete a trim request. Mark that we no longer have one in flight.
1460 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1463 isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1467 * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1468 * might use notes in the ccb for statistics.
1471 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1472 union ccb *done_ccb)
1475 #ifdef CAM_IOSCHED_DYNAMIC
1476 if (!do_dynamic_iosched)
1479 if (iosched_debug > 10)
1480 printf("done: %p %#x\n", bp, bp->bio_cmd);
1481 if (bp->bio_cmd == BIO_WRITE) {
1482 retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1483 if (!(bp->bio_flags & BIO_ERROR))
1484 isc->write_stats.errs++;
1485 isc->write_stats.out++;
1486 isc->write_stats.pending--;
1487 } else if (bp->bio_cmd == BIO_READ) {
1488 retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1489 if (!(bp->bio_flags & BIO_ERROR))
1490 isc->read_stats.errs++;
1491 isc->read_stats.out++;
1492 isc->read_stats.pending--;
1493 } else if (bp->bio_cmd == BIO_DELETE) {
1494 if (!(bp->bio_flags & BIO_ERROR))
1495 isc->trim_stats.errs++;
1496 isc->trim_stats.out++;
1497 isc->trim_stats.pending--;
1498 } else if (bp->bio_cmd != BIO_FLUSH) {
1500 printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1503 if (!(bp->bio_flags & BIO_ERROR))
1504 cam_iosched_io_metric_update(isc,
1505 cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data),
1506 bp->bio_cmd, bp->bio_bcount);
1512 * Tell the io scheduler that you've pushed a trim down into the sim.
1513 * This also tells the I/O scheduler not to push any more trims down, so
1514 * some periphs do not call it if they can cope with multiple trims in flight.
1517 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1520 isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1524 * Change the sorting policy hint for I/O transactions for this device.
1527 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1530 isc->sort_io_queue = val;
1534 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1536 return isc->flags & flags;
1540 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1542 isc->flags |= flags;
1546 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1548 isc->flags &= ~flags;
1551 #ifdef CAM_IOSCHED_DYNAMIC
1553 * After the method presented in Jack Crenshaw's 1998 article "Integer
1554 * Square Roots," reprinted at
1555 * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1556 * and well worth the read. Briefly, we find the power of 4 that's the
1557 * largest smaller than val. We then check each smaller power of 4 to
1558 * see if val is still bigger. The right shifts at each step divide
1559 * the result by 2 which after successive application winds up
1560 * accumulating the right answer. It could also have been accumulated
1561 * using a separate root counter, but this code is smaller and faster
1562 * than that method. This method is also integer size invariant.
1563 * It returns floor(sqrt((float)val)), or the largest integer less than
1564 * or equal to the square root.
1567 isqrt64(uint64_t val)
1570 uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1573 * Find the largest power of 4 smaller than val.
1579 * Accumulate the answer, one bit at a time (we keep moving
1580 * them over since 2 is the square root of 4 and we test
1581 * powers of 4). We accumulate where we find the bit, but
1582 * the successive shifts land the bit in the right place
1586 if (val >= res + bit) {
1588 res = (res >> 1) + bit;
1597 static sbintime_t latencies[LAT_BUCKETS - 1] = {
1611 SBT_1MS << 13 /* 8.192s */
1615 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1617 sbintime_t y, deltasq, delta;
1621 * Keep counts for latency. We do it by power of two buckets.
1622 * This helps us spot outlier behavior obscured by averages.
1624 for (i = 0; i < LAT_BUCKETS - 1; i++) {
1625 if (sim_latency < latencies[i]) {
1626 iop->latencies[i]++;
1630 if (i == LAT_BUCKETS - 1)
1631 iop->latencies[i]++; /* Put all > 1024ms values into the last bucket. */
1634 * Classic exponentially decaying average with a tiny alpha
1635 * (2 ^ -alpha_bits). For more info see the NIST statistical
1638 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha) [nist]
1639 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
1640 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1641 * alpha = 1 / (1 << alpha_bits)
1642 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
1643 * = y_t/b - e/b + be/b
1644 * = (y_t - e + be) / b
1647 * Since alpha is a power of two, we can compute this w/o any mult or
1650 * Variance can also be computed. Usually, it would be expressed as follows:
1651 * diff_t = y_t - ema_t-1
1652 * emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
1653 * = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
1654 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
1655 * = e - e/b + dd/b + dd/bb
1656 * = (bbe - be + bdd + dd) / bb
1657 * = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
1660 * XXX possible numeric issues
1661 * o We assume right shifted integers do the right thing, since that's
1662 * implementation defined. You can change the right shifts to / (1LL << alpha).
1663 * o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
1664 * for emvar. This puts a ceiling of 13 bits on alpha since we need a
1665 * few tens of seconds of representation.
1666 * o We mitigate alpha issues by never setting it too high.
1669 delta = (y - iop->ema); /* d */
1670 iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1673 * Were we to naively plow ahead at this point, we wind up with many numerical
1674 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
1675 * us with microsecond level precision in the input, so the same in the
1676 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
1677 * also means that emvar can be up 46 bits 40 of which are fraction, which
1678 * gives us a way to measure up to ~8s in the SD before the computation goes
1679 * unstable. Even the worst hard disk rarely has > 1s service time in the
1680 * drive. It does mean we have to shift left 12 bits after taking the
1681 * square root to compute the actual standard deviation estimate. This loss of
1682 * precision is preferable to needing int128 types to work. The above numbers
1683 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
1684 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1687 deltasq = delta * delta; /* dd */
1688 iop->emvar = ((iop->emvar << (2 * alpha_bits)) + /* bbe */
1689 ((deltasq - iop->emvar) << alpha_bits) + /* b(dd-e) */
1691 >> (2 * alpha_bits); /* div bb */
1692 iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1696 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1697 sbintime_t sim_latency, int cmd, size_t size)
1699 /* xxx Do we need to scale based on the size of the I/O ? */
1702 cam_iosched_update(&isc->read_stats, sim_latency);
1705 cam_iosched_update(&isc->write_stats, sim_latency);
1708 cam_iosched_update(&isc->trim_stats, sim_latency);
1716 static int biolen(struct bio_queue_head *bq)
1721 TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
1728 * Show the internal state of the I/O scheduler.
1730 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
1732 struct cam_iosched_softc *isc;
1735 db_printf("Need addr\n");
1738 isc = (struct cam_iosched_softc *)addr;
1739 db_printf("pending_reads: %d\n", isc->read_stats.pending);
1740 db_printf("min_reads: %d\n", isc->read_stats.min);
1741 db_printf("max_reads: %d\n", isc->read_stats.max);
1742 db_printf("reads: %d\n", isc->read_stats.total);
1743 db_printf("in_reads: %d\n", isc->read_stats.in);
1744 db_printf("out_reads: %d\n", isc->read_stats.out);
1745 db_printf("queued_reads: %d\n", isc->read_stats.queued);
1746 db_printf("Current Q len %d\n", biolen(&isc->bio_queue));
1747 db_printf("pending_writes: %d\n", isc->write_stats.pending);
1748 db_printf("min_writes: %d\n", isc->write_stats.min);
1749 db_printf("max_writes: %d\n", isc->write_stats.max);
1750 db_printf("writes: %d\n", isc->write_stats.total);
1751 db_printf("in_writes: %d\n", isc->write_stats.in);
1752 db_printf("out_writes: %d\n", isc->write_stats.out);
1753 db_printf("queued_writes: %d\n", isc->write_stats.queued);
1754 db_printf("Current Q len %d\n", biolen(&isc->write_queue));
1755 db_printf("pending_trims: %d\n", isc->trim_stats.pending);
1756 db_printf("min_trims: %d\n", isc->trim_stats.min);
1757 db_printf("max_trims: %d\n", isc->trim_stats.max);
1758 db_printf("trims: %d\n", isc->trim_stats.total);
1759 db_printf("in_trims: %d\n", isc->trim_stats.in);
1760 db_printf("out_trims: %d\n", isc->trim_stats.out);
1761 db_printf("queued_trims: %d\n", isc->trim_stats.queued);
1762 db_printf("Current Q len %d\n", biolen(&isc->trim_queue));
1763 db_printf("read_bias: %d\n", isc->read_bias);
1764 db_printf("current_read_bias: %d\n", isc->current_read_bias);
1765 db_printf("Trim active? %s\n",
1766 (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");