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.
96 * Note: See computation of EMA and EMVAR for acceptable ranges of alpha.
98 static int alpha_bits = 9;
99 TUNABLE_INT("kern.cam.iosched_alpha_bits", &alpha_bits);
100 SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW,
102 "Bits in EMA's alpha.");
105 struct cam_iosched_softc;
107 int iosched_debug = 0;
110 none = 0, /* No limits */
111 queue_depth, /* Limit how many ops we queue to SIM */
112 iops, /* Limit # of IOPS to the drive */
113 bandwidth, /* Limit bandwidth to the drive */
117 static const char *cam_iosched_limiter_names[] =
118 { "none", "queue_depth", "iops", "bandwidth" };
121 * Called to initialize the bits of the iop_stats structure relevant to the
122 * limiter. Called just after the limiter is set.
124 typedef int l_init_t(struct iop_stats *);
129 typedef int l_tick_t(struct iop_stats *);
132 * Called to see if the limiter thinks this IOP can be allowed to
133 * proceed. If so, the limiter assumes that the IOP proceeded
134 * and makes any accounting of it that's needed.
136 typedef int l_iop_t(struct iop_stats *, struct bio *);
139 * Called when an I/O completes so the limiter can update its
140 * accounting. Pending I/Os may complete in any order (even when
141 * sent to the hardware at the same time), so the limiter may not
142 * make any assumptions other than this I/O has completed. If it
143 * returns 1, then xpt_schedule() needs to be called again.
145 typedef int l_iodone_t(struct iop_stats *, struct bio *);
147 static l_iop_t cam_iosched_qd_iop;
148 static l_iop_t cam_iosched_qd_caniop;
149 static l_iodone_t cam_iosched_qd_iodone;
151 static l_init_t cam_iosched_iops_init;
152 static l_tick_t cam_iosched_iops_tick;
153 static l_iop_t cam_iosched_iops_caniop;
154 static l_iop_t cam_iosched_iops_iop;
156 static l_init_t cam_iosched_bw_init;
157 static l_tick_t cam_iosched_bw_tick;
158 static l_iop_t cam_iosched_bw_caniop;
159 static l_iop_t cam_iosched_bw_iop;
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,
200 * sysctl state for this subnode.
202 struct sysctl_ctx_list sysctl_ctx;
203 struct sysctl_oid *sysctl_tree;
206 * Information about the current rate limiters, if any
208 io_limiter limiter; /* How are I/Os being limited */
209 int min; /* Low range of limit */
210 int max; /* High range of limit */
211 int current; /* Current rate limiter */
212 int l_value1; /* per-limiter scratch value 1. */
213 int l_value2; /* per-limiter scratch value 2. */
216 * Debug information about counts of I/Os that have gone through the
219 int pending; /* I/Os pending in the hardware */
220 int queued; /* number currently in the queue */
221 int total; /* Total for all time -- wraps */
222 int in; /* number queued all time -- wraps */
223 int out; /* number completed all time -- wraps */
226 * Statistics on different bits of the process.
228 /* Exp Moving Average, see alpha_bits for more details */
231 sbintime_t sd; /* Last computed sd */
233 uint32_t state_flags;
234 #define IOP_RATE_LIMITED 1u
236 #define LAT_BUCKETS 15 /* < 1ms < 2ms ... < 2^(n-1)ms >= 2^(n-1)ms*/
237 uint64_t latencies[LAT_BUCKETS];
239 struct cam_iosched_softc *softc;
244 set_max = 0, /* current = max */
245 read_latency, /* Steer read latency by throttling writes */
246 cl_max /* Keep last */
249 static const char *cam_iosched_control_type_names[] =
250 { "set_max", "read_latency" };
252 struct control_loop {
254 * sysctl state for this subnode.
256 struct sysctl_ctx_list sysctl_ctx;
257 struct sysctl_oid *sysctl_tree;
259 sbintime_t next_steer; /* Time of next steer */
260 sbintime_t steer_interval; /* How often do we steer? */
264 control_type type; /* What type of control? */
265 int last_count; /* Last I/O count */
267 struct cam_iosched_softc *softc;
272 struct cam_iosched_softc {
273 struct bio_queue_head bio_queue;
274 struct bio_queue_head trim_queue;
275 /* scheduler flags < 16, user flags >= 16 */
278 #ifdef CAM_IOSCHED_DYNAMIC
279 int read_bias; /* Read bias setting */
280 int current_read_bias; /* Current read bias state */
282 int load; /* EMA of 'load average' of disk / 2^16 */
284 struct bio_queue_head write_queue;
285 struct iop_stats read_stats, write_stats, trim_stats;
286 struct sysctl_ctx_list sysctl_ctx;
287 struct sysctl_oid *sysctl_tree;
289 int quanta; /* Number of quanta per second */
290 struct callout ticker; /* Callout for our quota system */
291 struct cam_periph *periph; /* cam periph associated with this device */
292 uint32_t this_frac; /* Fraction of a second (1024ths) for this tick */
293 sbintime_t last_time; /* Last time we ticked */
294 struct control_loop cl;
298 #ifdef CAM_IOSCHED_DYNAMIC
300 * helper functions to call the limsw functions.
303 cam_iosched_limiter_init(struct iop_stats *ios)
305 int lim = ios->limiter;
307 /* maybe this should be a kassert */
308 if (lim < none || lim >= limiter_max)
311 if (limsw[lim].l_init)
312 return limsw[lim].l_init(ios);
318 cam_iosched_limiter_tick(struct iop_stats *ios)
320 int lim = ios->limiter;
322 /* maybe this should be a kassert */
323 if (lim < none || lim >= limiter_max)
326 if (limsw[lim].l_tick)
327 return limsw[lim].l_tick(ios);
333 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
335 int lim = ios->limiter;
337 /* maybe this should be a kassert */
338 if (lim < none || lim >= limiter_max)
341 if (limsw[lim].l_iop)
342 return limsw[lim].l_iop(ios, bp);
348 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
350 int lim = ios->limiter;
352 /* maybe this should be a kassert */
353 if (lim < none || lim >= limiter_max)
356 if (limsw[lim].l_caniop)
357 return limsw[lim].l_caniop(ios, bp);
363 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
365 int lim = ios->limiter;
367 /* maybe this should be a kassert */
368 if (lim < none || lim >= limiter_max)
371 if (limsw[lim].l_iodone)
372 return limsw[lim].l_iodone(ios, bp);
378 * Functions to implement the different kinds of limiters
382 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
385 if (ios->current <= 0 || ios->pending < ios->current)
392 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
395 if (ios->current <= 0 || ios->pending < ios->current)
402 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
405 if (ios->current <= 0 || ios->pending != ios->current)
412 cam_iosched_iops_init(struct iop_stats *ios)
415 ios->l_value1 = ios->current / ios->softc->quanta;
416 if (ios->l_value1 <= 0)
423 cam_iosched_iops_tick(struct iop_stats *ios)
426 ios->l_value1 = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
427 if (ios->l_value1 <= 0)
434 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
438 * So if we have any more IOPs left, allow it,
441 if (ios->l_value1 <= 0)
447 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
451 rv = cam_iosched_limiter_caniop(ios, bp);
459 cam_iosched_bw_init(struct iop_stats *ios)
462 /* ios->current is in kB/s, so scale to bytes */
463 ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
469 cam_iosched_bw_tick(struct iop_stats *ios)
474 * If we're in the hole for available quota from
475 * the last time, then add the quantum for this.
476 * If we have any left over from last quantum,
477 * then too bad, that's lost. Also, ios->current
478 * is in kB/s, so scale.
480 * We also allow up to 4 quanta of credits to
481 * accumulate to deal with burstiness. 4 is extremely
484 bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
485 if (ios->l_value1 < bw * 4)
492 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
495 * So if we have any more bw quota left, allow it,
496 * otherwise wait. Note, we'll go negative and that's
497 * OK. We'll just get a little less next quota.
499 * Note on going negative: that allows us to process
500 * requests in order better, since we won't allow
501 * shorter reads to get around the long one that we
502 * don't have the quota to do just yet. It also prevents
503 * starvation by being a little more permissive about
504 * what we let through this quantum (to prevent the
505 * starvation), at the cost of getting a little less
508 if (ios->l_value1 <= 0)
516 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
520 rv = cam_iosched_limiter_caniop(ios, bp);
522 ios->l_value1 -= bp->bio_length;
527 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
530 cam_iosched_ticker(void *arg)
532 struct cam_iosched_softc *isc = arg;
533 sbintime_t now, delta;
536 callout_reset(&isc->ticker, hz / isc->quanta - 1, cam_iosched_ticker, isc);
539 delta = now - isc->last_time;
540 isc->this_frac = (uint32_t)delta >> 16; /* Note: discards seconds -- should be 0 harmless if not */
541 isc->last_time = now;
543 cam_iosched_cl_maybe_steer(&isc->cl);
545 cam_iosched_limiter_tick(&isc->read_stats);
546 cam_iosched_limiter_tick(&isc->write_stats);
547 cam_iosched_limiter_tick(&isc->trim_stats);
549 cam_iosched_schedule(isc, isc->periph);
552 * isc->load is an EMA of the pending I/Os at each tick. The number of
553 * pending I/Os is the sum of the I/Os queued to the hardware, and those
554 * in the software queue that could be queued to the hardware if there
557 * ios_stats.pending is a count of requests in the SIM right now for
558 * each of these types of I/O. So the total pending count is the sum of
559 * these I/Os and the sum of the queued I/Os still in the software queue
560 * for those operations that aren't being rate limited at the moment.
562 * The reason for the rate limiting bit is because those I/Os
563 * aren't part of the software queued load (since we could
564 * give them to hardware, but choose not to).
566 * Note: due to a bug in counting pending TRIM in the device, we
567 * don't include them in this count. We count each BIO_DELETE in
568 * the pending count, but the periph drivers collapse them down
569 * into one TRIM command. That one trim command gets the completion
570 * so the counts get off.
572 pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
573 pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
574 !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
575 !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
577 pending /= isc->periph->path->device->ccbq.total_openings;
579 isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
586 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
589 clp->next_steer = sbinuptime();
591 clp->steer_interval = SBT_1S * 5; /* Let's start out steering every 5s */
592 clp->lolat = 5 * SBT_1MS;
593 clp->hilat = 15 * SBT_1MS;
594 clp->alpha = 20; /* Alpha == gain. 20 = .2 */
599 cam_iosched_cl_maybe_steer(struct control_loop *clp)
601 struct cam_iosched_softc *isc;
606 now = isc->last_time;
607 if (now < clp->next_steer)
610 clp->next_steer = now + clp->steer_interval;
613 if (isc->write_stats.current != isc->write_stats.max)
614 printf("Steering write from %d kBps to %d kBps\n",
615 isc->write_stats.current, isc->write_stats.max);
616 isc->read_stats.current = isc->read_stats.max;
617 isc->write_stats.current = isc->write_stats.max;
618 isc->trim_stats.current = isc->trim_stats.max;
621 old = isc->write_stats.current;
622 lat = isc->read_stats.ema;
624 * Simple PLL-like engine. Since we're steering to a range for
625 * the SP (set point) that makes things a little more
626 * complicated. In addition, we're not directly controlling our
627 * PV (process variable), the read latency, but instead are
628 * manipulating the write bandwidth limit for our MV
629 * (manipulation variable), analysis of this code gets a bit
630 * messy. Also, the MV is a very noisy control surface for read
631 * latency since it is affected by many hidden processes inside
632 * the device which change how responsive read latency will be
633 * in reaction to changes in write bandwidth. Unlike the classic
634 * boiler control PLL. this may result in over-steering while
635 * the SSD takes its time to react to the new, lower load. This
636 * is why we use a relatively low alpha of between .1 and .25 to
637 * compensate for this effect. At .1, it takes ~22 steering
638 * intervals to back off by a factor of 10. At .2 it only takes
639 * ~10. At .25 it only takes ~8. However some preliminary data
640 * from the SSD drives suggests a reasponse time in 10's of
641 * seconds before latency drops regardless of the new write
642 * rate. Careful observation will be required to tune this
645 * Also, when there's no read traffic, we jack up the write
646 * limit too regardless of the last read latency. 10 is
647 * somewhat arbitrary.
649 if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
650 isc->write_stats.current = isc->write_stats.current *
651 (100 + clp->alpha) / 100; /* Scale up */
652 else if (lat > clp->hilat)
653 isc->write_stats.current = isc->write_stats.current *
654 (100 - clp->alpha) / 100; /* Scale down */
655 clp->last_count = isc->read_stats.total;
658 * Even if we don't steer, per se, enforce the min/max limits as
659 * those may have changed.
661 if (isc->write_stats.current < isc->write_stats.min)
662 isc->write_stats.current = isc->write_stats.min;
663 if (isc->write_stats.current > isc->write_stats.max)
664 isc->write_stats.current = isc->write_stats.max;
665 if (old != isc->write_stats.current && iosched_debug)
666 printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
667 old, isc->write_stats.current,
668 (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
677 * Trim or similar currently pending completion. Should only be set for
678 * those drivers wishing only one Trim active at a time.
680 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE (1ul << 0)
681 /* Callout active, and needs to be torn down */
682 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
684 /* Periph drivers set these flags to indicate work */
685 #define CAM_IOSCHED_FLAG_WORK_FLAGS ((0xffffu) << 16)
687 #ifdef CAM_IOSCHED_DYNAMIC
689 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
690 sbintime_t sim_latency, int cmd, size_t size);
694 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
696 return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
700 cam_iosched_has_io(struct cam_iosched_softc *isc)
702 #ifdef CAM_IOSCHED_DYNAMIC
703 if (do_dynamic_iosched) {
704 struct bio *rbp = bioq_first(&isc->bio_queue);
705 struct bio *wbp = bioq_first(&isc->write_queue);
706 int can_write = wbp != NULL &&
707 cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
708 int can_read = rbp != NULL &&
709 cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
710 if (iosched_debug > 2) {
711 printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
712 printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
713 printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
715 return can_read || can_write;
718 return bioq_first(&isc->bio_queue) != NULL;
722 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
724 return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) &&
725 bioq_first(&isc->trim_queue);
728 #define cam_iosched_sort_queue(isc) ((isc)->sort_io_queue >= 0 ? \
729 (isc)->sort_io_queue : cam_sort_io_queues)
733 cam_iosched_has_work(struct cam_iosched_softc *isc)
735 #ifdef CAM_IOSCHED_DYNAMIC
736 if (iosched_debug > 2)
737 printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
738 cam_iosched_has_more_trim(isc),
739 cam_iosched_has_flagged_work(isc));
742 return cam_iosched_has_io(isc) ||
743 cam_iosched_has_more_trim(isc) ||
744 cam_iosched_has_flagged_work(isc);
747 #ifdef CAM_IOSCHED_DYNAMIC
749 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
753 cam_iosched_limiter_init(ios);
767 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
770 struct iop_stats *ios;
771 struct cam_iosched_softc *isc;
777 value = ios->limiter;
778 if (value < none || value >= limiter_max)
781 p = cam_iosched_limiter_names[value];
783 strlcpy(buf, p, sizeof(buf));
784 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
785 if (error != 0 || req->newptr == NULL)
788 cam_periph_lock(isc->periph);
790 for (i = none; i < limiter_max; i++) {
791 if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
794 error = cam_iosched_limiter_init(ios);
796 ios->limiter = value;
797 cam_periph_unlock(isc->periph);
800 /* Note: disk load averate requires ticker to be always running */
801 callout_reset(&isc->ticker, hz / isc->quanta - 1, cam_iosched_ticker, isc);
802 isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
804 cam_periph_unlock(isc->periph);
808 cam_periph_unlock(isc->periph);
813 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
816 struct control_loop *clp;
817 struct cam_iosched_softc *isc;
824 if (value < none || value >= cl_max)
827 p = cam_iosched_control_type_names[value];
829 strlcpy(buf, p, sizeof(buf));
830 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
831 if (error != 0 || req->newptr == NULL)
834 for (i = set_max; i < cl_max; i++) {
835 if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
837 cam_periph_lock(isc->periph);
839 cam_periph_unlock(isc->periph);
847 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
854 value = *(sbintime_t *)arg1;
855 us = (uint64_t)value / SBT_1US;
856 snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
857 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
858 if (error != 0 || req->newptr == NULL)
860 us = strtoul(buf, NULL, 10);
863 *(sbintime_t *)arg1 = us * SBT_1US;
868 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
875 sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
877 for (i = 0; i < LAT_BUCKETS - 1; i++)
878 sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
879 sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
880 error = sbuf_finish(&sb);
887 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS)
892 quanta = (unsigned *)arg1;
895 error = sysctl_handle_int(oidp, (int *)&value, 0, req);
896 if ((error != 0) || (req->newptr == NULL))
899 if (value < 1 || value > hz)
908 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
910 struct sysctl_oid_list *n;
911 struct sysctl_ctx_list *ctx;
913 ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
914 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
915 CTLFLAG_RD, 0, name);
916 n = SYSCTL_CHILDREN(ios->sysctl_tree);
917 ctx = &ios->sysctl_ctx;
919 SYSCTL_ADD_UQUAD(ctx, n,
920 OID_AUTO, "ema", CTLFLAG_RD,
922 "Fast Exponentially Weighted Moving Average");
923 SYSCTL_ADD_UQUAD(ctx, n,
924 OID_AUTO, "emvar", CTLFLAG_RD,
926 "Fast Exponentially Weighted Moving Variance");
928 SYSCTL_ADD_INT(ctx, n,
929 OID_AUTO, "pending", CTLFLAG_RD,
931 "Instantaneous # of pending transactions");
932 SYSCTL_ADD_INT(ctx, n,
933 OID_AUTO, "count", CTLFLAG_RD,
935 "# of transactions submitted to hardware");
936 SYSCTL_ADD_INT(ctx, n,
937 OID_AUTO, "queued", CTLFLAG_RD,
939 "# of transactions in the queue");
940 SYSCTL_ADD_INT(ctx, n,
941 OID_AUTO, "in", CTLFLAG_RD,
943 "# of transactions queued to driver");
944 SYSCTL_ADD_INT(ctx, n,
945 OID_AUTO, "out", CTLFLAG_RD,
947 "# of transactions completed");
949 SYSCTL_ADD_PROC(ctx, n,
950 OID_AUTO, "limiter", CTLTYPE_STRING | CTLFLAG_RW,
951 ios, 0, cam_iosched_limiter_sysctl, "A",
952 "Current limiting type.");
953 SYSCTL_ADD_INT(ctx, n,
954 OID_AUTO, "min", CTLFLAG_RW,
957 SYSCTL_ADD_INT(ctx, n,
958 OID_AUTO, "max", CTLFLAG_RW,
961 SYSCTL_ADD_INT(ctx, n,
962 OID_AUTO, "current", CTLFLAG_RW,
966 SYSCTL_ADD_PROC(ctx, n,
967 OID_AUTO, "latencies", CTLTYPE_STRING | CTLFLAG_RD,
969 cam_iosched_sysctl_latencies, "A",
970 "Array of power of 2 latency from 1ms to 1.024s");
974 cam_iosched_iop_stats_fini(struct iop_stats *ios)
976 if (ios->sysctl_tree)
977 if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
978 printf("can't remove iosched sysctl stats context\n");
982 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
984 struct sysctl_oid_list *n;
985 struct sysctl_ctx_list *ctx;
986 struct control_loop *clp;
989 clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
990 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
991 CTLFLAG_RD, 0, "Control loop info");
992 n = SYSCTL_CHILDREN(clp->sysctl_tree);
993 ctx = &clp->sysctl_ctx;
995 SYSCTL_ADD_PROC(ctx, n,
996 OID_AUTO, "type", CTLTYPE_STRING | CTLFLAG_RW,
997 clp, 0, cam_iosched_control_type_sysctl, "A",
998 "Control loop algorithm");
999 SYSCTL_ADD_PROC(ctx, n,
1000 OID_AUTO, "steer_interval", CTLTYPE_STRING | CTLFLAG_RW,
1001 &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
1002 "How often to steer (in us)");
1003 SYSCTL_ADD_PROC(ctx, n,
1004 OID_AUTO, "lolat", CTLTYPE_STRING | CTLFLAG_RW,
1005 &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
1006 "Low water mark for Latency (in us)");
1007 SYSCTL_ADD_PROC(ctx, n,
1008 OID_AUTO, "hilat", CTLTYPE_STRING | CTLFLAG_RW,
1009 &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
1010 "Hi water mark for Latency (in us)");
1011 SYSCTL_ADD_INT(ctx, n,
1012 OID_AUTO, "alpha", CTLFLAG_RW,
1014 "Alpha for PLL (x100) aka gain");
1018 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1020 if (clp->sysctl_tree)
1021 if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1022 printf("can't remove iosched sysctl control loop context\n");
1027 * Allocate the iosched structure. This also insulates callers from knowing
1028 * sizeof struct cam_iosched_softc.
1031 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1034 *iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1037 #ifdef CAM_IOSCHED_DYNAMIC
1039 printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1041 (*iscp)->sort_io_queue = -1;
1042 bioq_init(&(*iscp)->bio_queue);
1043 bioq_init(&(*iscp)->trim_queue);
1044 #ifdef CAM_IOSCHED_DYNAMIC
1045 if (do_dynamic_iosched) {
1046 bioq_init(&(*iscp)->write_queue);
1047 (*iscp)->read_bias = 100;
1048 (*iscp)->current_read_bias = 100;
1049 (*iscp)->quanta = min(hz, 200);
1050 cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1051 cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1052 cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1053 (*iscp)->trim_stats.max = 1; /* Trims are special: one at a time for now */
1054 (*iscp)->last_time = sbinuptime();
1055 callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1056 (*iscp)->periph = periph;
1057 cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1058 callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta - 1, cam_iosched_ticker, *iscp);
1059 (*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1067 * Reclaim all used resources. This assumes that other folks have
1068 * drained the requests in the hardware. Maybe an unwise assumption.
1071 cam_iosched_fini(struct cam_iosched_softc *isc)
1074 cam_iosched_flush(isc, NULL, ENXIO);
1075 #ifdef CAM_IOSCHED_DYNAMIC
1076 cam_iosched_iop_stats_fini(&isc->read_stats);
1077 cam_iosched_iop_stats_fini(&isc->write_stats);
1078 cam_iosched_iop_stats_fini(&isc->trim_stats);
1079 cam_iosched_cl_sysctl_fini(&isc->cl);
1080 if (isc->sysctl_tree)
1081 if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1082 printf("can't remove iosched sysctl stats context\n");
1083 if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1084 callout_drain(&isc->ticker);
1085 isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1088 free(isc, M_CAMSCHED);
1093 * After we're sure we're attaching a device, go ahead and add
1094 * hooks for any sysctl we may wish to honor.
1096 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1097 struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1099 #ifdef CAM_IOSCHED_DYNAMIC
1100 struct sysctl_oid_list *n;
1103 SYSCTL_ADD_INT(ctx, SYSCTL_CHILDREN(node),
1104 OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1105 &isc->sort_io_queue, 0,
1106 "Sort IO queue to try and optimise disk access patterns");
1108 #ifdef CAM_IOSCHED_DYNAMIC
1109 if (!do_dynamic_iosched)
1112 isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1113 SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1114 CTLFLAG_RD, 0, "I/O scheduler statistics");
1115 n = SYSCTL_CHILDREN(isc->sysctl_tree);
1116 ctx = &isc->sysctl_ctx;
1118 cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1119 cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1120 cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1121 cam_iosched_cl_sysctl_init(isc);
1123 SYSCTL_ADD_INT(ctx, n,
1124 OID_AUTO, "read_bias", CTLFLAG_RW,
1125 &isc->read_bias, 100,
1126 "How biased towards read should we be independent of limits");
1128 SYSCTL_ADD_PROC(ctx, n,
1129 OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW,
1130 &isc->quanta, 0, cam_iosched_quanta_sysctl, "I",
1131 "How many quanta per second do we slice the I/O up into");
1133 SYSCTL_ADD_INT(ctx, n,
1134 OID_AUTO, "total_ticks", CTLFLAG_RD,
1135 &isc->total_ticks, 0,
1136 "Total number of ticks we've done");
1138 SYSCTL_ADD_INT(ctx, n,
1139 OID_AUTO, "load", CTLFLAG_RD,
1141 "scaled load average / 100");
1146 * Flush outstanding I/O. Consumers of this library don't know all the
1147 * queues we may keep, so this allows all I/O to be flushed in one
1151 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1153 bioq_flush(&isc->bio_queue, stp, err);
1154 bioq_flush(&isc->trim_queue, stp, err);
1155 #ifdef CAM_IOSCHED_DYNAMIC
1156 if (do_dynamic_iosched)
1157 bioq_flush(&isc->write_queue, stp, err);
1161 #ifdef CAM_IOSCHED_DYNAMIC
1163 cam_iosched_get_write(struct cam_iosched_softc *isc)
1168 * We control the write rate by controlling how many requests we send
1169 * down to the drive at any one time. Fewer requests limits the
1170 * effects of both starvation when the requests take a while and write
1171 * amplification when each request is causing more than one write to
1172 * the NAND media. Limiting the queue depth like this will also limit
1173 * the write throughput and give and reads that want to compete to
1176 bp = bioq_first(&isc->write_queue);
1178 if (iosched_debug > 3)
1179 printf("No writes present in write_queue\n");
1184 * If pending read, prefer that based on current read bias
1187 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1189 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);
1190 isc->current_read_bias--;
1191 /* We're not limiting writes, per se, just doing reads first */
1196 * See if our current limiter allows this I/O.
1198 if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1200 printf("Can't write because limiter says no.\n");
1201 isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1206 * Let's do this: We've passed all the gates and we're a go
1207 * to schedule the I/O in the SIM.
1209 isc->current_read_bias = isc->read_bias;
1210 bioq_remove(&isc->write_queue, bp);
1211 if (bp->bio_cmd == BIO_WRITE) {
1212 isc->write_stats.queued--;
1213 isc->write_stats.total++;
1214 isc->write_stats.pending++;
1216 if (iosched_debug > 9)
1217 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1218 isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1224 * Put back a trim that you weren't able to actually schedule this time.
1227 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1229 bioq_insert_head(&isc->trim_queue, bp);
1230 #ifdef CAM_IOSCHED_DYNAMIC
1231 isc->trim_stats.queued++;
1232 isc->trim_stats.total--; /* since we put it back, don't double count */
1233 isc->trim_stats.pending--;
1238 * gets the next trim from the trim queue.
1240 * Assumes we're called with the periph lock held. It removes this
1241 * trim from the queue and the device must explicitly reinsert it
1242 * should the need arise.
1245 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1249 bp = bioq_first(&isc->trim_queue);
1252 bioq_remove(&isc->trim_queue, bp);
1253 #ifdef CAM_IOSCHED_DYNAMIC
1254 isc->trim_stats.queued--;
1255 isc->trim_stats.total++;
1256 isc->trim_stats.pending++;
1262 * gets an available trim from the trim queue, if there's no trim
1263 * already pending. It removes this trim from the queue and the device
1264 * must explicitly reinsert it should the need arise.
1266 * Assumes we're called with the periph lock held.
1269 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1272 if (!cam_iosched_has_more_trim(isc))
1275 return cam_iosched_next_trim(isc);
1279 * Determine what the next bit of work to do is for the periph. The
1280 * default implementation looks to see if we have trims to do, but no
1281 * trims outstanding. If so, we do that. Otherwise we see if we have
1282 * other work. If we do, then we do that. Otherwise why were we called?
1285 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1290 * See if we have a trim that can be scheduled. We can only send one
1291 * at a time down, so this takes that into account.
1293 * XXX newer TRIM commands are queueable. Revisit this when we
1296 if ((bp = cam_iosched_get_trim(isc)) != NULL)
1299 #ifdef CAM_IOSCHED_DYNAMIC
1301 * See if we have any pending writes, and room in the queue for them,
1302 * and if so, those are next.
1304 if (do_dynamic_iosched) {
1305 if ((bp = cam_iosched_get_write(isc)) != NULL)
1311 * next, see if there's other, normal I/O waiting. If so return that.
1313 if ((bp = bioq_first(&isc->bio_queue)) == NULL)
1316 #ifdef CAM_IOSCHED_DYNAMIC
1318 * For the dynamic scheduler, bio_queue is only for reads, so enforce
1319 * the limits here. Enforce only for reads.
1321 if (do_dynamic_iosched) {
1322 if (bp->bio_cmd == BIO_READ &&
1323 cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1324 isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1328 isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1330 bioq_remove(&isc->bio_queue, bp);
1331 #ifdef CAM_IOSCHED_DYNAMIC
1332 if (do_dynamic_iosched) {
1333 if (bp->bio_cmd == BIO_READ) {
1334 isc->read_stats.queued--;
1335 isc->read_stats.total++;
1336 isc->read_stats.pending++;
1338 printf("Found bio_cmd = %#x\n", bp->bio_cmd);
1340 if (iosched_debug > 9)
1341 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1347 * Driver has been given some work to do by the block layer. Tell the
1348 * scheduler about it and have it queue the work up. The scheduler module
1349 * will then return the currently most useful bit of work later, possibly
1350 * deferring work for various reasons.
1353 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1357 * Put all trims on the trim queue sorted, since we know
1358 * that the collapsing code requires this. Otherwise put
1359 * the work on the bio queue.
1361 if (bp->bio_cmd == BIO_DELETE) {
1362 bioq_disksort(&isc->trim_queue, bp);
1363 #ifdef CAM_IOSCHED_DYNAMIC
1364 isc->trim_stats.in++;
1365 isc->trim_stats.queued++;
1368 #ifdef CAM_IOSCHED_DYNAMIC
1369 else if (do_dynamic_iosched && (bp->bio_cmd != BIO_READ)) {
1370 if (cam_iosched_sort_queue(isc))
1371 bioq_disksort(&isc->write_queue, bp);
1373 bioq_insert_tail(&isc->write_queue, bp);
1374 if (iosched_debug > 9)
1375 printf("Qw : %p %#x\n", bp, bp->bio_cmd);
1376 if (bp->bio_cmd == BIO_WRITE) {
1377 isc->write_stats.in++;
1378 isc->write_stats.queued++;
1383 if (cam_iosched_sort_queue(isc))
1384 bioq_disksort(&isc->bio_queue, bp);
1386 bioq_insert_tail(&isc->bio_queue, bp);
1387 #ifdef CAM_IOSCHED_DYNAMIC
1388 if (iosched_debug > 9)
1389 printf("Qr : %p %#x\n", bp, bp->bio_cmd);
1390 if (bp->bio_cmd == BIO_READ) {
1391 isc->read_stats.in++;
1392 isc->read_stats.queued++;
1393 } else if (bp->bio_cmd == BIO_WRITE) {
1394 isc->write_stats.in++;
1395 isc->write_stats.queued++;
1402 * If we have work, get it scheduled. Called with the periph lock held.
1405 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1408 if (cam_iosched_has_work(isc))
1409 xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1413 * Complete a trim request. Mark that we no longer have one in flight.
1416 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1419 isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1423 * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1424 * might use notes in the ccb for statistics.
1427 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1428 union ccb *done_ccb)
1431 #ifdef CAM_IOSCHED_DYNAMIC
1432 if (!do_dynamic_iosched)
1435 if (iosched_debug > 10)
1436 printf("done: %p %#x\n", bp, bp->bio_cmd);
1437 if (bp->bio_cmd == BIO_WRITE) {
1438 retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1439 isc->write_stats.out++;
1440 isc->write_stats.pending--;
1441 } else if (bp->bio_cmd == BIO_READ) {
1442 retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1443 isc->read_stats.out++;
1444 isc->read_stats.pending--;
1445 } else if (bp->bio_cmd == BIO_DELETE) {
1446 isc->trim_stats.out++;
1447 isc->trim_stats.pending--;
1448 } else if (bp->bio_cmd != BIO_FLUSH) {
1450 printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1453 if (!(bp->bio_flags & BIO_ERROR))
1454 cam_iosched_io_metric_update(isc,
1455 cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data),
1456 bp->bio_cmd, bp->bio_bcount);
1462 * Tell the io scheduler that you've pushed a trim down into the sim.
1463 * This also tells the I/O scheduler not to push any more trims down, so
1464 * some periphs do not call it if they can cope with multiple trims in flight.
1467 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1470 isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1474 * Change the sorting policy hint for I/O transactions for this device.
1477 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1480 isc->sort_io_queue = val;
1484 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1486 return isc->flags & flags;
1490 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1492 isc->flags |= flags;
1496 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1498 isc->flags &= ~flags;
1501 #ifdef CAM_IOSCHED_DYNAMIC
1503 * After the method presented in Jack Crenshaw's 1998 article "Integer
1504 * Square Roots," reprinted at
1505 * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1506 * and well worth the read. Briefly, we find the power of 4 that's the
1507 * largest smaller than val. We then check each smaller power of 4 to
1508 * see if val is still bigger. The right shifts at each step divide
1509 * the result by 2 which after successive application winds up
1510 * accumulating the right answer. It could also have been accumulated
1511 * using a separate root counter, but this code is smaller and faster
1512 * than that method. This method is also integer size invariant.
1513 * It returns floor(sqrt((float)val)), or the largest integer less than
1514 * or equal to the square root.
1517 isqrt64(uint64_t val)
1520 uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1523 * Find the largest power of 4 smaller than val.
1529 * Accumulate the answer, one bit at a time (we keep moving
1530 * them over since 2 is the square root of 4 and we test
1531 * powers of 4). We accumulate where we find the bit, but
1532 * the successive shifts land the bit in the right place
1536 if (val >= res + bit) {
1538 res = (res >> 1) + bit;
1547 static sbintime_t latencies[LAT_BUCKETS - 1] = {
1561 SBT_1MS << 13 /* 8.192s */
1565 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1567 sbintime_t y, deltasq, delta;
1571 * Keep counts for latency. We do it by power of two buckets.
1572 * This helps us spot outlier behavior obscured by averages.
1574 for (i = 0; i < LAT_BUCKETS - 1; i++) {
1575 if (sim_latency < latencies[i]) {
1576 iop->latencies[i]++;
1580 if (i == LAT_BUCKETS - 1)
1581 iop->latencies[i]++; /* Put all > 1024ms values into the last bucket. */
1584 * Classic exponentially decaying average with a tiny alpha
1585 * (2 ^ -alpha_bits). For more info see the NIST statistical
1588 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha) [nist]
1589 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
1590 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1591 * alpha = 1 / (1 << alpha_bits)
1592 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
1593 * = y_t/b - e/b + be/b
1594 * = (y_t - e + be) / b
1597 * Since alpha is a power of two, we can compute this w/o any mult or
1600 * Variance can also be computed. Usually, it would be expressed as follows:
1601 * diff_t = y_t - ema_t-1
1602 * emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
1603 * = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
1604 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
1605 * = e - e/b + dd/b + dd/bb
1606 * = (bbe - be + bdd + dd) / bb
1607 * = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
1610 * XXX possible numeric issues
1611 * o We assume right shifted integers do the right thing, since that's
1612 * implementation defined. You can change the right shifts to / (1LL << alpha).
1613 * o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
1614 * for emvar. This puts a ceiling of 13 bits on alpha since we need a
1615 * few tens of seconds of representation.
1616 * o We mitigate alpha issues by never setting it too high.
1619 delta = (y - iop->ema); /* d */
1620 iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1623 * Were we to naively plow ahead at this point, we wind up with many numerical
1624 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
1625 * us with microsecond level precision in the input, so the same in the
1626 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
1627 * also means that emvar can be up 46 bits 40 of which are fraction, which
1628 * gives us a way to measure up to ~8s in the SD before the computation goes
1629 * unstable. Even the worst hard disk rarely has > 1s service time in the
1630 * drive. It does mean we have to shift left 12 bits after taking the
1631 * square root to compute the actual standard deviation estimate. This loss of
1632 * precision is preferable to needing int128 types to work. The above numbers
1633 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
1634 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1637 deltasq = delta * delta; /* dd */
1638 iop->emvar = ((iop->emvar << (2 * alpha_bits)) + /* bbe */
1639 ((deltasq - iop->emvar) << alpha_bits) + /* b(dd-e) */
1641 >> (2 * alpha_bits); /* div bb */
1642 iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1646 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1647 sbintime_t sim_latency, int cmd, size_t size)
1649 /* xxx Do we need to scale based on the size of the I/O ? */
1652 cam_iosched_update(&isc->read_stats, sim_latency);
1655 cam_iosched_update(&isc->write_stats, sim_latency);
1658 cam_iosched_update(&isc->trim_stats, sim_latency);
1666 static int biolen(struct bio_queue_head *bq)
1671 TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
1678 * Show the internal state of the I/O scheduler.
1680 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
1682 struct cam_iosched_softc *isc;
1685 db_printf("Need addr\n");
1688 isc = (struct cam_iosched_softc *)addr;
1689 db_printf("pending_reads: %d\n", isc->read_stats.pending);
1690 db_printf("min_reads: %d\n", isc->read_stats.min);
1691 db_printf("max_reads: %d\n", isc->read_stats.max);
1692 db_printf("reads: %d\n", isc->read_stats.total);
1693 db_printf("in_reads: %d\n", isc->read_stats.in);
1694 db_printf("out_reads: %d\n", isc->read_stats.out);
1695 db_printf("queued_reads: %d\n", isc->read_stats.queued);
1696 db_printf("Current Q len %d\n", biolen(&isc->bio_queue));
1697 db_printf("pending_writes: %d\n", isc->write_stats.pending);
1698 db_printf("min_writes: %d\n", isc->write_stats.min);
1699 db_printf("max_writes: %d\n", isc->write_stats.max);
1700 db_printf("writes: %d\n", isc->write_stats.total);
1701 db_printf("in_writes: %d\n", isc->write_stats.in);
1702 db_printf("out_writes: %d\n", isc->write_stats.out);
1703 db_printf("queued_writes: %d\n", isc->write_stats.queued);
1704 db_printf("Current Q len %d\n", biolen(&isc->write_queue));
1705 db_printf("pending_trims: %d\n", isc->trim_stats.pending);
1706 db_printf("min_trims: %d\n", isc->trim_stats.min);
1707 db_printf("max_trims: %d\n", isc->trim_stats.max);
1708 db_printf("trims: %d\n", isc->trim_stats.total);
1709 db_printf("in_trims: %d\n", isc->trim_stats.in);
1710 db_printf("out_trims: %d\n", isc->trim_stats.out);
1711 db_printf("queued_trims: %d\n", isc->trim_stats.queued);
1712 db_printf("Current Q len %d\n", biolen(&isc->trim_queue));
1713 db_printf("read_bias: %d\n", isc->read_bias);
1714 db_printf("current_read_bias: %d\n", isc->current_read_bias);
1715 db_printf("Trim active? %s\n",
1716 (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");