2 * CAM IO Scheduler Interface
4 * Copyright (c) 2015 Netflix, Inc.
7 * Redistribution and use in source and binary forms, with or without
8 * modification, are permitted provided that the following conditions
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.
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
34 #include <sys/cdefs.h>
35 __FBSDID("$FreeBSD$");
37 #include <sys/param.h>
39 #include <sys/systm.h>
40 #include <sys/kernel.h>
43 #include <sys/malloc.h>
44 #include <sys/mutex.h>
46 #include <sys/sysctl.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>
57 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler",
58 "CAM I/O Scheduler buffers");
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
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).
70 #ifdef CAM_IOSCHED_DYNAMIC
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.");
79 * For an EMA, with an alpha of alpha, we know
83 * where N is the number of samples that 86% of the current
84 * EMA is derived from.
86 * So we invent[*] alpha_bits:
87 * alpha_bits = -log_2(alpha)
88 * alpha = 2^-alpha_bits
90 * N = 1 + 2^(alpha_bits + 1)
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
95 * [*] Steal from the load average code and many other places.
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,
101 "Bits in EMA's alpha.");
104 struct cam_iosched_softc;
106 int iosched_debug = 0;
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 */
116 static const char *cam_iosched_limiter_names[] =
117 { "none", "queue_depth", "iops", "bandwidth" };
120 * Called to initialize the bits of the iop_stats structure relevant to the
121 * limiter. Called just after the limiter is set.
123 typedef int l_init_t(struct iop_stats *);
128 typedef int l_tick_t(struct iop_stats *);
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.
135 typedef int l_iop_t(struct iop_stats *, struct bio *);
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.
144 typedef int l_iodone_t(struct iop_stats *, struct bio *);
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;
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;
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;
166 l_iodone_t *l_iodone;
178 .l_caniop = cam_iosched_qd_caniop,
179 .l_iop = cam_iosched_qd_iop,
180 .l_iodone= cam_iosched_qd_iodone,
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,
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,
201 * sysctl state for this subnode.
203 struct sysctl_ctx_list sysctl_ctx;
204 struct sysctl_oid *sysctl_tree;
207 * Information about the current rate limiters, if any
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. */
218 * Debug information about counts of I/Os that have gone through the
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 */
228 * Statistics on different bits of the process.
230 /* Exp Moving Average, see alpha_bits for more details */
232 sbintime_t emss; /* Exp Moving sum of the squares */
233 sbintime_t sd; /* Last computed sd */
235 uint32_t state_flags;
236 #define IOP_RATE_LIMITED 1u
238 #define LAT_BUCKETS 12 /* < 1ms < 2ms ... 512ms < 1024ms > 1024ms */
239 uint64_t latencies[LAT_BUCKETS];
241 struct cam_iosched_softc *softc;
246 set_max = 0, /* current = max */
247 read_latency, /* Steer read latency by throttling writes */
248 cl_max /* Keep last */
251 static const char *cam_iosched_control_type_names[] =
252 { "set_max", "read_latency" };
257 * sysctl state for this subnode.
259 struct sysctl_ctx_list sysctl_ctx;
260 struct sysctl_oid *sysctl_tree;
262 sbintime_t next_steer; /* Time of next steer */
263 sbintime_t steer_interval; /* How often do we steer? */
267 control_type type; /* What type of control? */
268 int last_count; /* Last I/O count */
270 struct cam_iosched_softc *softc;
275 struct cam_iosched_softc
277 struct bio_queue_head bio_queue;
278 struct bio_queue_head trim_queue;
279 /* scheduler flags < 16, user flags >= 16 */
282 #ifdef CAM_IOSCHED_DYNAMIC
283 int read_bias; /* Read bias setting */
284 int current_read_bias; /* Current read bias state */
286 int load; /* EMA of 'load average' of disk / 2^16 */
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;
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;
302 #ifdef CAM_IOSCHED_DYNAMIC
304 * helper functions to call the limsw functions.
307 cam_iosched_limiter_init(struct iop_stats *ios)
309 int lim = ios->limiter;
311 /* maybe this should be a kassert */
312 if (lim < none || lim >= limiter_max)
315 if (limsw[lim].l_init)
316 return limsw[lim].l_init(ios);
322 cam_iosched_limiter_tick(struct iop_stats *ios)
324 int lim = ios->limiter;
326 /* maybe this should be a kassert */
327 if (lim < none || lim >= limiter_max)
330 if (limsw[lim].l_tick)
331 return limsw[lim].l_tick(ios);
337 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
339 int lim = ios->limiter;
341 /* maybe this should be a kassert */
342 if (lim < none || lim >= limiter_max)
345 if (limsw[lim].l_iop)
346 return limsw[lim].l_iop(ios, bp);
352 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
354 int lim = ios->limiter;
356 /* maybe this should be a kassert */
357 if (lim < none || lim >= limiter_max)
360 if (limsw[lim].l_caniop)
361 return limsw[lim].l_caniop(ios, bp);
367 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
369 int lim = ios->limiter;
371 /* maybe this should be a kassert */
372 if (lim < none || lim >= limiter_max)
375 if (limsw[lim].l_iodone)
376 return limsw[lim].l_iodone(ios, bp);
382 * Functions to implement the different kinds of limiters
386 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
389 if (ios->current <= 0 || ios->pending < ios->current)
396 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
399 if (ios->current <= 0 || ios->pending < ios->current)
406 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
409 if (ios->current <= 0 || ios->pending != ios->current)
416 cam_iosched_iops_init(struct iop_stats *ios)
419 ios->l_value1 = ios->current / ios->softc->quanta;
420 if (ios->l_value1 <= 0)
427 cam_iosched_iops_tick(struct iop_stats *ios)
430 ios->l_value1 = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
431 if (ios->l_value1 <= 0)
438 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
442 * So if we have any more IOPs left, allow it,
445 if (ios->l_value1 <= 0)
451 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
455 rv = cam_iosched_limiter_caniop(ios, bp);
463 cam_iosched_bw_init(struct iop_stats *ios)
466 /* ios->current is in kB/s, so scale to bytes */
467 ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
473 cam_iosched_bw_tick(struct iop_stats *ios)
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.
484 * We also allow up to 4 quanta of credits to
485 * accumulate to deal with burstiness. 4 is extremely
488 bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
489 if (ios->l_value1 < bw * 4)
496 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
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.
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
512 if (ios->l_value1 <= 0)
520 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
524 rv = cam_iosched_limiter_caniop(ios, bp);
526 ios->l_value1 -= bp->bio_length;
531 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
534 cam_iosched_ticker(void *arg)
536 struct cam_iosched_softc *isc = arg;
537 sbintime_t now, delta;
540 callout_reset(&isc->ticker, hz / isc->quanta - 1, cam_iosched_ticker, isc);
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;
547 cam_iosched_cl_maybe_steer(&isc->cl);
549 cam_iosched_limiter_tick(&isc->read_stats);
550 cam_iosched_limiter_tick(&isc->write_stats);
551 cam_iosched_limiter_tick(&isc->trim_stats);
553 cam_iosched_schedule(isc, isc->periph);
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
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.
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).
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.
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 */ ;
581 pending /= isc->periph->path->device->ccbq.total_openings;
583 isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
590 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
593 clp->next_steer = sbinuptime();
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 */
603 cam_iosched_cl_maybe_steer(struct control_loop *clp)
605 struct cam_iosched_softc *isc;
610 now = isc->last_time;
611 if (now < clp->next_steer)
614 clp->next_steer = now + clp->steer_interval;
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;
625 old = isc->write_stats.current;
626 lat = isc->read_stats.ema;
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
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.
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;
662 * Even if we don't steer, per se, enforce the min/max limits as
663 * those may have changed.
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);
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)
685 /* Periph drivers set these flags to indicate work */
686 #define CAM_IOSCHED_FLAG_WORK_FLAGS ((0xffffu) << 16)
688 #ifdef CAM_IOSCHED_DYNAMIC
690 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
691 sbintime_t sim_latency, int cmd, size_t size);
695 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
697 return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
701 cam_iosched_has_io(struct cam_iosched_softc *isc)
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);
716 return can_read || can_write;
719 return bioq_first(&isc->bio_queue) != NULL;
723 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
725 return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) &&
726 bioq_first(&isc->trim_queue);
729 #define cam_iosched_sort_queue(isc) ((isc)->sort_io_queue >= 0 ? \
730 (isc)->sort_io_queue : cam_sort_io_queues)
734 cam_iosched_has_work(struct cam_iosched_softc *isc)
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));
743 return cam_iosched_has_io(isc) ||
744 cam_iosched_has_more_trim(isc) ||
745 cam_iosched_has_flagged_work(isc);
748 #ifdef CAM_IOSCHED_DYNAMIC
750 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
754 cam_iosched_limiter_init(ios);
769 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
772 struct iop_stats *ios;
773 struct cam_iosched_softc *isc;
779 value = ios->limiter;
780 if (value < none || value >= limiter_max)
783 p = cam_iosched_limiter_names[value];
785 strlcpy(buf, p, sizeof(buf));
786 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
787 if (error != 0 || req->newptr == NULL)
790 cam_periph_lock(isc->periph);
792 for (i = none; i < limiter_max; i++) {
793 if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
796 error = cam_iosched_limiter_init(ios);
798 ios->limiter = value;
799 cam_periph_unlock(isc->periph);
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;
806 cam_periph_unlock(isc->periph);
810 cam_periph_unlock(isc->periph);
815 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
818 struct control_loop *clp;
819 struct cam_iosched_softc *isc;
826 if (value < none || value >= cl_max)
829 p = cam_iosched_control_type_names[value];
831 strlcpy(buf, p, sizeof(buf));
832 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
833 if (error != 0 || req->newptr == NULL)
836 for (i = set_max; i < cl_max; i++) {
837 if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
839 cam_periph_lock(isc->periph);
841 cam_periph_unlock(isc->periph);
849 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
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)
862 us = strtoul(buf, NULL, 10);
865 *(sbintime_t *)arg1 = us * SBT_1US;
870 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
877 sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
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);
889 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
891 struct sysctl_oid_list *n;
892 struct sysctl_ctx_list *ctx;
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;
900 SYSCTL_ADD_UQUAD(ctx, n,
901 OID_AUTO, "ema", CTLFLAG_RD,
903 "Fast Exponentially Weighted Moving Average");
904 SYSCTL_ADD_UQUAD(ctx, n,
905 OID_AUTO, "emss", CTLFLAG_RD,
907 "Fast Exponentially Weighted Moving Sum of Squares (maybe wrong)");
908 SYSCTL_ADD_UQUAD(ctx, n,
909 OID_AUTO, "sd", CTLFLAG_RD,
911 "Estimated SD for fast ema (may be wrong)");
913 SYSCTL_ADD_INT(ctx, n,
914 OID_AUTO, "pending", CTLFLAG_RD,
916 "Instantaneous # of pending transactions");
917 SYSCTL_ADD_INT(ctx, n,
918 OID_AUTO, "count", CTLFLAG_RD,
920 "# of transactions submitted to hardware");
921 SYSCTL_ADD_INT(ctx, n,
922 OID_AUTO, "queued", CTLFLAG_RD,
924 "# of transactions in the queue");
925 SYSCTL_ADD_INT(ctx, n,
926 OID_AUTO, "in", CTLFLAG_RD,
928 "# of transactions queued to driver");
929 SYSCTL_ADD_INT(ctx, n,
930 OID_AUTO, "out", CTLFLAG_RD,
932 "# of transactions completed");
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,
942 SYSCTL_ADD_INT(ctx, n,
943 OID_AUTO, "max", CTLFLAG_RW,
946 SYSCTL_ADD_INT(ctx, n,
947 OID_AUTO, "current", CTLFLAG_RW,
951 SYSCTL_ADD_PROC(ctx, n,
952 OID_AUTO, "latencies", CTLTYPE_STRING | CTLFLAG_RD,
954 cam_iosched_sysctl_latencies, "A",
955 "Array of power of 2 latency from 1ms to 1.024s");
959 cam_iosched_iop_stats_fini(struct iop_stats *ios)
961 if (ios->sysctl_tree)
962 if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
963 printf("can't remove iosched sysctl stats context\n");
967 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
969 struct sysctl_oid_list *n;
970 struct sysctl_ctx_list *ctx;
971 struct control_loop *clp;
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;
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,
999 "Alpha for PLL (x100) aka gain");
1003 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
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");
1012 * Allocate the iosched structure. This also insulates callers from knowing
1013 * sizeof struct cam_iosched_softc.
1016 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1019 *iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1022 #ifdef CAM_IOSCHED_DYNAMIC
1024 printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
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;
1052 * Reclaim all used resources. This assumes that other folks have
1053 * drained the requests in the hardware. Maybe an unwise assumption.
1056 cam_iosched_fini(struct cam_iosched_softc *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;
1074 free(isc, M_CAMSCHED);
1079 * After we're sure we're attaching a device, go ahead and add
1080 * hooks for any sysctl we may wish to honor.
1082 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1083 struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1085 #ifdef CAM_IOSCHED_DYNAMIC
1086 struct sysctl_oid_list *n;
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");
1094 #ifdef CAM_IOSCHED_DYNAMIC
1095 if (!do_dynamic_iosched)
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;
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);
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");
1114 SYSCTL_ADD_INT(ctx, n,
1115 OID_AUTO, "quanta", CTLFLAG_RW,
1117 "How many quanta per second do we slice the I/O up into");
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");
1124 SYSCTL_ADD_INT(ctx, n,
1125 OID_AUTO, "load", CTLFLAG_RD,
1127 "scaled load average / 100");
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
1137 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
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);
1147 #ifdef CAM_IOSCHED_DYNAMIC
1149 cam_iosched_get_write(struct cam_iosched_softc *isc)
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
1162 bp = bioq_first(&isc->write_queue);
1164 if (iosched_debug > 3)
1165 printf("No writes present in write_queue\n");
1170 * If pending read, prefer that based on current read bias
1173 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
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 */
1182 * See if our current limiter allows this I/O.
1184 if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1186 printf("Can't write because limiter says no.\n");
1187 isc->write_stats.state_flags |= IOP_RATE_LIMITED;
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.
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++;
1202 if (iosched_debug > 9)
1203 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1204 isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1210 * Put back a trim that you weren't able to actually schedule this time.
1213 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
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--;
1224 * gets the next trim from the trim queue.
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.
1231 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1235 bp = bioq_first(&isc->trim_queue);
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++;
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.
1252 * Assumes we're called with the periph lock held.
1255 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1258 if (!cam_iosched_has_more_trim(isc))
1261 return cam_iosched_next_trim(isc);
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?
1271 cam_iosched_next_bio(struct cam_iosched_softc *isc)
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.
1279 * XXX newer TRIM commands are queueable. Revisit this when we
1282 if ((bp = cam_iosched_get_trim(isc)) != NULL)
1285 #ifdef CAM_IOSCHED_DYNAMIC
1287 * See if we have any pending writes, and room in the queue for them,
1288 * and if so, those are next.
1290 if (do_dynamic_iosched) {
1291 if ((bp = cam_iosched_get_write(isc)) != NULL)
1297 * next, see if there's other, normal I/O waiting. If so return that.
1299 if ((bp = bioq_first(&isc->bio_queue)) == NULL)
1302 #ifdef CAM_IOSCHED_DYNAMIC
1304 * For the dynamic scheduler, bio_queue is only for reads, so enforce
1305 * the limits here. Enforce only for reads.
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;
1314 isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
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++;
1324 printf("Found bio_cmd = %#x\n", bp->bio_cmd);
1326 if (iosched_debug > 9)
1327 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
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.
1339 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
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.
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++;
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);
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++;
1370 if (cam_iosched_sort_queue(isc))
1371 bioq_disksort(&isc->bio_queue, bp);
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++;
1389 * If we have work, get it scheduled. Called with the periph lock held.
1392 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1395 if (cam_iosched_has_work(isc))
1396 xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1400 * Complete a trim request
1403 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1406 isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
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.
1414 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1415 union ccb *done_ccb)
1418 #ifdef CAM_IOSCHED_DYNAMIC
1419 if (!do_dynamic_iosched)
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) {
1437 printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
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);
1448 * Tell the io scheduler that you've pushed a trim down into the sim.
1449 * xxx better place for this?
1452 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1455 isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1459 * Change the sorting policy hint for I/O transactions for this device.
1462 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1465 isc->sort_io_queue = val;
1469 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1471 return isc->flags & flags;
1475 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1477 isc->flags |= flags;
1481 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1483 isc->flags &= ~flags;
1486 #ifdef CAM_IOSCHED_DYNAMIC
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.
1502 isqrt64(uint64_t val)
1505 uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1508 * Find the largest power of 4 smaller than val.
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
1521 if (val >= res + bit) {
1523 res = (res >> 1) + bit;
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
1539 * (ah + al / R) * (bh + bl / R)
1540 * ah * bh + (al * bh + ah * bl) / R + al * bl / R^2
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
1551 mul(uint64_t a, uint64_t b)
1553 uint64_t al, ah, bl, bh;
1554 al = a & 0xffffffff;
1556 bl = b & 0xffffffff;
1558 return ((ah * bh) << 32) + al * bh + ah * bl + ((al * bl) >> 32);
1561 static sbintime_t latencies[] = {
1576 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1583 * Keep counts for latency. We do it by power of two buckets.
1584 * This helps us spot outlier behavior obscured by averages.
1586 for (i = 0; i < LAT_BUCKETS - 1; i++) {
1587 if (sim_latency < latencies[i]) {
1588 iop->latencies[i]++;
1592 if (i == LAT_BUCKETS - 1)
1593 iop->latencies[i]++; /* Put all > 1024ms values into the last bucket. */
1596 * Classic expoentially decaying average with a tiny alpha
1597 * (2 ^ -alpha_bits). For more info see the NIST statistical
1600 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha)
1601 * alpha = 1 / (1 << alpha_bits)
1603 * Since alpha is a power of two, we can compute this w/o any mult or
1607 iop->ema = (y + (iop->ema << alpha_bits) - iop->ema) >> alpha_bits;
1610 iop->emss = (yy + (iop->emss << alpha_bits) - iop->emss) >> alpha_bits;
1614 * s_2 = sum of data * data
1615 * ema ~ mean (or s_1 / N)
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))))
1622 * alpha < 1 / 16 (typically much less)
1623 * N > 31 --> N large so N * (N - 1) is approx N * N
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.
1634 var = iop->emss - mul(iop->ema, iop->ema);
1635 iop->sd = (int64_t)var < 0 ? 0 : isqrt64(var);
1638 #ifdef CAM_IOSCHED_DYNAMIC
1640 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1641 sbintime_t sim_latency, int cmd, size_t size)
1643 /* xxx Do we need to scale based on the size of the I/O ? */
1646 cam_iosched_update(&isc->read_stats, sim_latency);
1649 cam_iosched_update(&isc->write_stats, sim_latency);
1652 cam_iosched_update(&isc->trim_stats, sim_latency);
1661 static int biolen(struct bio_queue_head *bq)
1666 TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
1673 * Show the internal state of the I/O scheduler.
1675 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
1677 struct cam_iosched_softc *isc;
1680 db_printf("Need addr\n");
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");