2 * CAM IO Scheduler Interface
4 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
6 * Copyright (c) 2015 Netflix, Inc.
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
11 * 1. Redistributions of source code must retain the above copyright
12 * notice, this list of conditions, and the following disclaimer,
13 * without modification, immediately at the beginning of the file.
14 * 2. The name of the author may not be used to endorse or promote products
15 * derived from this software without specific prior written permission.
17 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
18 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
19 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
20 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR
21 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
22 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
23 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
24 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
25 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
26 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
38 #include <sys/param.h>
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
44 #include <sys/malloc.h>
45 #include <sys/mutex.h>
47 #include <sys/sysctl.h>
50 #include <cam/cam_ccb.h>
51 #include <cam/cam_periph.h>
52 #include <cam/cam_xpt_periph.h>
53 #include <cam/cam_xpt_internal.h>
54 #include <cam/cam_iosched.h>
58 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler",
59 "CAM I/O Scheduler buffers");
62 * Default I/O scheduler for FreeBSD. This implementation is just a thin-vineer
63 * over the bioq_* interface, with notions of separate calls for normal I/O and
66 * When CAM_IOSCHED_DYNAMIC is defined, the scheduler is enhanced to dynamically
67 * steer the rate of one type of traffic to help other types of traffic (eg
68 * limit writes when read latency deteriorates on SSDs).
71 #ifdef CAM_IOSCHED_DYNAMIC
73 static int do_dynamic_iosched = 1;
74 TUNABLE_INT("kern.cam.do_dynamic_iosched", &do_dynamic_iosched);
75 SYSCTL_INT(_kern_cam, OID_AUTO, do_dynamic_iosched, CTLFLAG_RD,
76 &do_dynamic_iosched, 1,
77 "Enable Dynamic I/O scheduler optimizations.");
80 * For an EMA, with an alpha of alpha, we know
84 * where N is the number of samples that 86% of the current
85 * EMA is derived from.
87 * So we invent[*] alpha_bits:
88 * alpha_bits = -log_2(alpha)
89 * alpha = 2^-alpha_bits
91 * N = 1 + 2^(alpha_bits + 1)
93 * The default 9 gives a 1025 lookback for 86% of the data.
94 * For a brief intro: https://en.wikipedia.org/wiki/Moving_average
96 * [*] Steal from the load average code and many other places.
97 * Note: See computation of EMA and EMVAR for acceptable ranges of alpha.
99 static int alpha_bits = 9;
100 TUNABLE_INT("kern.cam.iosched_alpha_bits", &alpha_bits);
101 SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW,
103 "Bits in EMA's alpha.");
106 struct cam_iosched_softc;
108 int iosched_debug = 0;
111 none = 0, /* No limits */
112 queue_depth, /* Limit how many ops we queue to SIM */
113 iops, /* Limit # of IOPS to the drive */
114 bandwidth, /* Limit bandwidth to the drive */
118 static const char *cam_iosched_limiter_names[] =
119 { "none", "queue_depth", "iops", "bandwidth" };
122 * Called to initialize the bits of the iop_stats structure relevant to the
123 * limiter. Called just after the limiter is set.
125 typedef int l_init_t(struct iop_stats *);
130 typedef int l_tick_t(struct iop_stats *);
133 * Called to see if the limiter thinks this IOP can be allowed to
134 * proceed. If so, the limiter assumes that the IOP proceeded
135 * and makes any accounting of it that's needed.
137 typedef int l_iop_t(struct iop_stats *, struct bio *);
140 * Called when an I/O completes so the limiter can update its
141 * accounting. Pending I/Os may complete in any order (even when
142 * sent to the hardware at the same time), so the limiter may not
143 * make any assumptions other than this I/O has completed. If it
144 * returns 1, then xpt_schedule() needs to be called again.
146 typedef int l_iodone_t(struct iop_stats *, struct bio *);
148 static l_iop_t cam_iosched_qd_iop;
149 static l_iop_t cam_iosched_qd_caniop;
150 static l_iodone_t cam_iosched_qd_iodone;
152 static l_init_t cam_iosched_iops_init;
153 static l_tick_t cam_iosched_iops_tick;
154 static l_iop_t cam_iosched_iops_caniop;
155 static l_iop_t cam_iosched_iops_iop;
157 static l_init_t cam_iosched_bw_init;
158 static l_tick_t cam_iosched_bw_tick;
159 static l_iop_t cam_iosched_bw_caniop;
160 static l_iop_t cam_iosched_bw_iop;
167 l_iodone_t *l_iodone;
179 .l_caniop = cam_iosched_qd_caniop,
180 .l_iop = cam_iosched_qd_iop,
181 .l_iodone= cam_iosched_qd_iodone,
184 .l_init = cam_iosched_iops_init,
185 .l_tick = cam_iosched_iops_tick,
186 .l_caniop = cam_iosched_iops_caniop,
187 .l_iop = cam_iosched_iops_iop,
191 .l_init = cam_iosched_bw_init,
192 .l_tick = cam_iosched_bw_tick,
193 .l_caniop = cam_iosched_bw_caniop,
194 .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. */
217 * Debug information about counts of I/Os that have gone through the
220 int pending; /* I/Os pending in the hardware */
221 int queued; /* number currently in the queue */
222 int total; /* Total for all time -- wraps */
223 int in; /* number queued all time -- wraps */
224 int out; /* number completed all time -- wraps */
225 int errs; /* Number of I/Os completed with error -- wraps */
228 * Statistics on different bits of the process.
230 /* Exp Moving Average, see alpha_bits for more details */
233 sbintime_t sd; /* Last computed sd */
235 uint32_t state_flags;
236 #define IOP_RATE_LIMITED 1u
238 #define LAT_BUCKETS 15 /* < 1ms < 2ms ... < 2^(n-1)ms >= 2^(n-1)ms*/
239 uint64_t latencies[LAT_BUCKETS];
241 struct cam_iosched_softc *softc;
245 set_max = 0, /* current = max */
246 read_latency, /* Steer read latency by throttling writes */
247 cl_max /* Keep last */
250 static const char *cam_iosched_control_type_names[] =
251 { "set_max", "read_latency" };
253 struct control_loop {
255 * sysctl state for this subnode.
257 struct sysctl_ctx_list sysctl_ctx;
258 struct sysctl_oid *sysctl_tree;
260 sbintime_t next_steer; /* Time of next steer */
261 sbintime_t steer_interval; /* How often do we steer? */
265 control_type type; /* What type of control? */
266 int last_count; /* Last I/O count */
268 struct cam_iosched_softc *softc;
273 struct cam_iosched_softc {
274 struct bio_queue_head bio_queue;
275 struct bio_queue_head trim_queue;
276 /* scheduler flags < 16, user flags >= 16 */
279 int trim_goal; /* # of trims to queue before sending */
280 int trim_ticks; /* Max ticks to hold trims */
281 int last_trim_tick; /* Last 'tick' time ld a trim */
282 int queued_trims; /* Number of trims in the queue */
283 #ifdef CAM_IOSCHED_DYNAMIC
284 int read_bias; /* Read bias setting */
285 int current_read_bias; /* Current read bias state */
287 int load; /* EMA of 'load average' of disk / 2^16 */
289 struct bio_queue_head write_queue;
290 struct iop_stats read_stats, write_stats, trim_stats;
291 struct sysctl_ctx_list sysctl_ctx;
292 struct sysctl_oid *sysctl_tree;
294 int quanta; /* Number of quanta per second */
295 struct callout ticker; /* Callout for our quota system */
296 struct cam_periph *periph; /* cam periph associated with this device */
297 uint32_t this_frac; /* Fraction of a second (1024ths) for this tick */
298 sbintime_t last_time; /* Last time we ticked */
299 struct control_loop cl;
300 sbintime_t max_lat; /* when != 0, if iop latency > max_lat, call max_lat_fcn */
301 cam_iosched_latfcn_t latfcn;
306 #ifdef CAM_IOSCHED_DYNAMIC
308 * helper functions to call the limsw functions.
311 cam_iosched_limiter_init(struct iop_stats *ios)
313 int lim = ios->limiter;
315 /* maybe this should be a kassert */
316 if (lim < none || lim >= limiter_max)
319 if (limsw[lim].l_init)
320 return limsw[lim].l_init(ios);
326 cam_iosched_limiter_tick(struct iop_stats *ios)
328 int lim = ios->limiter;
330 /* maybe this should be a kassert */
331 if (lim < none || lim >= limiter_max)
334 if (limsw[lim].l_tick)
335 return limsw[lim].l_tick(ios);
341 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
343 int lim = ios->limiter;
345 /* maybe this should be a kassert */
346 if (lim < none || lim >= limiter_max)
349 if (limsw[lim].l_iop)
350 return limsw[lim].l_iop(ios, bp);
356 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
358 int lim = ios->limiter;
360 /* maybe this should be a kassert */
361 if (lim < none || lim >= limiter_max)
364 if (limsw[lim].l_caniop)
365 return limsw[lim].l_caniop(ios, bp);
371 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
373 int lim = ios->limiter;
375 /* maybe this should be a kassert */
376 if (lim < none || lim >= limiter_max)
379 if (limsw[lim].l_iodone)
380 return limsw[lim].l_iodone(ios, bp);
386 * Functions to implement the different kinds of limiters
390 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
393 if (ios->current <= 0 || ios->pending < ios->current)
400 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
403 if (ios->current <= 0 || ios->pending < ios->current)
410 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
413 if (ios->current <= 0 || ios->pending != ios->current)
420 cam_iosched_iops_init(struct iop_stats *ios)
423 ios->l_value1 = ios->current / ios->softc->quanta;
424 if (ios->l_value1 <= 0)
432 cam_iosched_iops_tick(struct iop_stats *ios)
437 * Allow at least one IO per tick until all
438 * the IOs for this interval have been spent.
440 new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
441 if (new_ios < 1 && ios->l_value2 < ios->current) {
447 * If this a new accounting interval, discard any "unspent" ios
448 * granted in the previous interval. Otherwise add the new ios to
449 * the previously granted ones that haven't been spent yet.
451 if ((ios->softc->total_ticks % ios->softc->quanta) == 0) {
452 ios->l_value1 = new_ios;
455 ios->l_value1 += new_ios;
462 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
466 * So if we have any more IOPs left, allow it,
467 * otherwise wait. If current iops is 0, treat that
468 * as unlimited as a failsafe.
470 if (ios->current > 0 && ios->l_value1 <= 0)
476 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
480 rv = cam_iosched_limiter_caniop(ios, bp);
488 cam_iosched_bw_init(struct iop_stats *ios)
491 /* ios->current is in kB/s, so scale to bytes */
492 ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
498 cam_iosched_bw_tick(struct iop_stats *ios)
503 * If we're in the hole for available quota from
504 * the last time, then add the quantum for this.
505 * If we have any left over from last quantum,
506 * then too bad, that's lost. Also, ios->current
507 * is in kB/s, so scale.
509 * We also allow up to 4 quanta of credits to
510 * accumulate to deal with burstiness. 4 is extremely
513 bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
514 if (ios->l_value1 < bw * 4)
521 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
524 * So if we have any more bw quota left, allow it,
525 * otherwise wait. Note, we'll go negative and that's
526 * OK. We'll just get a little less next quota.
528 * Note on going negative: that allows us to process
529 * requests in order better, since we won't allow
530 * shorter reads to get around the long one that we
531 * don't have the quota to do just yet. It also prevents
532 * starvation by being a little more permissive about
533 * what we let through this quantum (to prevent the
534 * starvation), at the cost of getting a little less
537 * Also note that if the current limit is <= 0,
538 * we treat it as unlimited as a failsafe.
540 if (ios->current > 0 && ios->l_value1 <= 0)
547 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
551 rv = cam_iosched_limiter_caniop(ios, bp);
553 ios->l_value1 -= bp->bio_length;
558 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
561 cam_iosched_ticker(void *arg)
563 struct cam_iosched_softc *isc = arg;
564 sbintime_t now, delta;
567 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
570 delta = now - isc->last_time;
571 isc->this_frac = (uint32_t)delta >> 16; /* Note: discards seconds -- should be 0 harmless if not */
572 isc->last_time = now;
574 cam_iosched_cl_maybe_steer(&isc->cl);
576 cam_iosched_limiter_tick(&isc->read_stats);
577 cam_iosched_limiter_tick(&isc->write_stats);
578 cam_iosched_limiter_tick(&isc->trim_stats);
580 cam_iosched_schedule(isc, isc->periph);
583 * isc->load is an EMA of the pending I/Os at each tick. The number of
584 * pending I/Os is the sum of the I/Os queued to the hardware, and those
585 * in the software queue that could be queued to the hardware if there
588 * ios_stats.pending is a count of requests in the SIM right now for
589 * each of these types of I/O. So the total pending count is the sum of
590 * these I/Os and the sum of the queued I/Os still in the software queue
591 * for those operations that aren't being rate limited at the moment.
593 * The reason for the rate limiting bit is because those I/Os
594 * aren't part of the software queued load (since we could
595 * give them to hardware, but choose not to).
597 * Note: due to a bug in counting pending TRIM in the device, we
598 * don't include them in this count. We count each BIO_DELETE in
599 * the pending count, but the periph drivers collapse them down
600 * into one TRIM command. That one trim command gets the completion
601 * so the counts get off.
603 pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
604 pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
605 !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
606 !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
608 pending /= isc->periph->path->device->ccbq.total_openings;
610 isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
616 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
619 clp->next_steer = sbinuptime();
621 clp->steer_interval = SBT_1S * 5; /* Let's start out steering every 5s */
622 clp->lolat = 5 * SBT_1MS;
623 clp->hilat = 15 * SBT_1MS;
624 clp->alpha = 20; /* Alpha == gain. 20 = .2 */
629 cam_iosched_cl_maybe_steer(struct control_loop *clp)
631 struct cam_iosched_softc *isc;
636 now = isc->last_time;
637 if (now < clp->next_steer)
640 clp->next_steer = now + clp->steer_interval;
643 if (isc->write_stats.current != isc->write_stats.max)
644 printf("Steering write from %d kBps to %d kBps\n",
645 isc->write_stats.current, isc->write_stats.max);
646 isc->read_stats.current = isc->read_stats.max;
647 isc->write_stats.current = isc->write_stats.max;
648 isc->trim_stats.current = isc->trim_stats.max;
651 old = isc->write_stats.current;
652 lat = isc->read_stats.ema;
654 * Simple PLL-like engine. Since we're steering to a range for
655 * the SP (set point) that makes things a little more
656 * complicated. In addition, we're not directly controlling our
657 * PV (process variable), the read latency, but instead are
658 * manipulating the write bandwidth limit for our MV
659 * (manipulation variable), analysis of this code gets a bit
660 * messy. Also, the MV is a very noisy control surface for read
661 * latency since it is affected by many hidden processes inside
662 * the device which change how responsive read latency will be
663 * in reaction to changes in write bandwidth. Unlike the classic
664 * boiler control PLL. this may result in over-steering while
665 * the SSD takes its time to react to the new, lower load. This
666 * is why we use a relatively low alpha of between .1 and .25 to
667 * compensate for this effect. At .1, it takes ~22 steering
668 * intervals to back off by a factor of 10. At .2 it only takes
669 * ~10. At .25 it only takes ~8. However some preliminary data
670 * from the SSD drives suggests a reasponse time in 10's of
671 * seconds before latency drops regardless of the new write
672 * rate. Careful observation will be required to tune this
675 * Also, when there's no read traffic, we jack up the write
676 * limit too regardless of the last read latency. 10 is
677 * somewhat arbitrary.
679 if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
680 isc->write_stats.current = isc->write_stats.current *
681 (100 + clp->alpha) / 100; /* Scale up */
682 else if (lat > clp->hilat)
683 isc->write_stats.current = isc->write_stats.current *
684 (100 - clp->alpha) / 100; /* Scale down */
685 clp->last_count = isc->read_stats.total;
688 * Even if we don't steer, per se, enforce the min/max limits as
689 * those may have changed.
691 if (isc->write_stats.current < isc->write_stats.min)
692 isc->write_stats.current = isc->write_stats.min;
693 if (isc->write_stats.current > isc->write_stats.max)
694 isc->write_stats.current = isc->write_stats.max;
695 if (old != isc->write_stats.current && iosched_debug)
696 printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
697 old, isc->write_stats.current,
698 (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
707 * Trim or similar currently pending completion. Should only be set for
708 * those drivers wishing only one Trim active at a time.
710 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE (1ul << 0)
711 /* Callout active, and needs to be torn down */
712 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
714 /* Periph drivers set these flags to indicate work */
715 #define CAM_IOSCHED_FLAG_WORK_FLAGS ((0xffffu) << 16)
717 #ifdef CAM_IOSCHED_DYNAMIC
719 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
720 sbintime_t sim_latency, int cmd, size_t size);
724 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
726 return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
730 cam_iosched_has_io(struct cam_iosched_softc *isc)
732 #ifdef CAM_IOSCHED_DYNAMIC
733 if (do_dynamic_iosched) {
734 struct bio *rbp = bioq_first(&isc->bio_queue);
735 struct bio *wbp = bioq_first(&isc->write_queue);
736 bool can_write = wbp != NULL &&
737 cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
738 bool can_read = rbp != NULL &&
739 cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
740 if (iosched_debug > 2) {
741 printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
742 printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
743 printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
745 return can_read || can_write;
748 return bioq_first(&isc->bio_queue) != NULL;
752 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
756 bp = bioq_first(&isc->trim_queue);
757 #ifdef CAM_IOSCHED_DYNAMIC
758 if (do_dynamic_iosched) {
760 * If we're limiting trims, then defer action on trims
763 if (bp == NULL || cam_iosched_limiter_caniop(&isc->trim_stats, bp) != 0)
769 * If we've set a trim_goal, then if we exceed that allow trims
770 * to be passed back to the driver. If we've also set a tick timeout
771 * allow trims back to the driver. Otherwise, don't allow trims yet.
773 if (isc->trim_goal > 0) {
774 if (isc->queued_trims >= isc->trim_goal)
776 if (isc->queued_trims > 0 &&
777 isc->trim_ticks > 0 &&
778 ticks - isc->last_trim_tick > isc->trim_ticks)
783 /* NB: Should perhaps have a max trim active independent of I/O limiters */
784 return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) && bp != NULL;
787 #define cam_iosched_sort_queue(isc) ((isc)->sort_io_queue >= 0 ? \
788 (isc)->sort_io_queue : cam_sort_io_queues)
791 cam_iosched_has_work(struct cam_iosched_softc *isc)
793 #ifdef CAM_IOSCHED_DYNAMIC
794 if (iosched_debug > 2)
795 printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
796 cam_iosched_has_more_trim(isc),
797 cam_iosched_has_flagged_work(isc));
800 return cam_iosched_has_io(isc) ||
801 cam_iosched_has_more_trim(isc) ||
802 cam_iosched_has_flagged_work(isc);
805 #ifdef CAM_IOSCHED_DYNAMIC
807 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
812 ios->max = ios->current = 300000;
822 cam_iosched_limiter_init(ios);
826 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
829 struct iop_stats *ios;
830 struct cam_iosched_softc *isc;
836 value = ios->limiter;
837 if (value < none || value >= limiter_max)
840 p = cam_iosched_limiter_names[value];
842 strlcpy(buf, p, sizeof(buf));
843 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
844 if (error != 0 || req->newptr == NULL)
847 cam_periph_lock(isc->periph);
849 for (i = none; i < limiter_max; i++) {
850 if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
853 error = cam_iosched_limiter_init(ios);
855 ios->limiter = value;
856 cam_periph_unlock(isc->periph);
859 /* Note: disk load averate requires ticker to be always running */
860 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
861 isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
863 cam_periph_unlock(isc->periph);
867 cam_periph_unlock(isc->periph);
872 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
875 struct control_loop *clp;
876 struct cam_iosched_softc *isc;
883 if (value < none || value >= cl_max)
886 p = cam_iosched_control_type_names[value];
888 strlcpy(buf, p, sizeof(buf));
889 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
890 if (error != 0 || req->newptr == NULL)
893 for (i = set_max; i < cl_max; i++) {
894 if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
896 cam_periph_lock(isc->periph);
898 cam_periph_unlock(isc->periph);
906 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
913 value = *(sbintime_t *)arg1;
914 us = (uint64_t)value / SBT_1US;
915 snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
916 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
917 if (error != 0 || req->newptr == NULL)
919 us = strtoul(buf, NULL, 10);
922 *(sbintime_t *)arg1 = us * SBT_1US;
927 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
934 sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
936 for (i = 0; i < LAT_BUCKETS - 1; i++)
937 sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
938 sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
939 error = sbuf_finish(&sb);
946 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS)
951 quanta = (unsigned *)arg1;
954 error = sysctl_handle_int(oidp, (int *)&value, 0, req);
955 if ((error != 0) || (req->newptr == NULL))
958 if (value < 1 || value > hz)
967 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
969 struct sysctl_oid_list *n;
970 struct sysctl_ctx_list *ctx;
972 ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
973 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
974 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, name);
975 n = SYSCTL_CHILDREN(ios->sysctl_tree);
976 ctx = &ios->sysctl_ctx;
978 SYSCTL_ADD_UQUAD(ctx, n,
979 OID_AUTO, "ema", CTLFLAG_RD,
981 "Fast Exponentially Weighted Moving Average");
982 SYSCTL_ADD_UQUAD(ctx, n,
983 OID_AUTO, "emvar", CTLFLAG_RD,
985 "Fast Exponentially Weighted Moving Variance");
987 SYSCTL_ADD_INT(ctx, n,
988 OID_AUTO, "pending", CTLFLAG_RD,
990 "Instantaneous # of pending transactions");
991 SYSCTL_ADD_INT(ctx, n,
992 OID_AUTO, "count", CTLFLAG_RD,
994 "# of transactions submitted to hardware");
995 SYSCTL_ADD_INT(ctx, n,
996 OID_AUTO, "queued", CTLFLAG_RD,
998 "# of transactions in the queue");
999 SYSCTL_ADD_INT(ctx, n,
1000 OID_AUTO, "in", CTLFLAG_RD,
1002 "# of transactions queued to driver");
1003 SYSCTL_ADD_INT(ctx, n,
1004 OID_AUTO, "out", CTLFLAG_RD,
1006 "# of transactions completed (including with error)");
1007 SYSCTL_ADD_INT(ctx, n,
1008 OID_AUTO, "errs", CTLFLAG_RD,
1010 "# of transactions completed with an error");
1012 SYSCTL_ADD_PROC(ctx, n,
1013 OID_AUTO, "limiter",
1014 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
1015 ios, 0, cam_iosched_limiter_sysctl, "A",
1016 "Current limiting type.");
1017 SYSCTL_ADD_INT(ctx, n,
1018 OID_AUTO, "min", CTLFLAG_RW,
1021 SYSCTL_ADD_INT(ctx, n,
1022 OID_AUTO, "max", CTLFLAG_RW,
1025 SYSCTL_ADD_INT(ctx, n,
1026 OID_AUTO, "current", CTLFLAG_RW,
1028 "current resource");
1030 SYSCTL_ADD_PROC(ctx, n,
1031 OID_AUTO, "latencies",
1032 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
1034 cam_iosched_sysctl_latencies, "A",
1035 "Array of power of 2 latency from 1ms to 1.024s");
1039 cam_iosched_iop_stats_fini(struct iop_stats *ios)
1041 if (ios->sysctl_tree)
1042 if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
1043 printf("can't remove iosched sysctl stats context\n");
1047 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
1049 struct sysctl_oid_list *n;
1050 struct sysctl_ctx_list *ctx;
1051 struct control_loop *clp;
1054 clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1055 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
1056 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Control loop info");
1057 n = SYSCTL_CHILDREN(clp->sysctl_tree);
1058 ctx = &clp->sysctl_ctx;
1060 SYSCTL_ADD_PROC(ctx, n,
1062 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
1063 clp, 0, cam_iosched_control_type_sysctl, "A",
1064 "Control loop algorithm");
1065 SYSCTL_ADD_PROC(ctx, n,
1066 OID_AUTO, "steer_interval",
1067 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
1068 &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
1069 "How often to steer (in us)");
1070 SYSCTL_ADD_PROC(ctx, n,
1072 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
1073 &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
1074 "Low water mark for Latency (in us)");
1075 SYSCTL_ADD_PROC(ctx, n,
1077 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
1078 &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
1079 "Hi water mark for Latency (in us)");
1080 SYSCTL_ADD_INT(ctx, n,
1081 OID_AUTO, "alpha", CTLFLAG_RW,
1083 "Alpha for PLL (x100) aka gain");
1087 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1089 if (clp->sysctl_tree)
1090 if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1091 printf("can't remove iosched sysctl control loop context\n");
1096 * Allocate the iosched structure. This also insulates callers from knowing
1097 * sizeof struct cam_iosched_softc.
1100 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1103 *iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1106 #ifdef CAM_IOSCHED_DYNAMIC
1108 printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1110 (*iscp)->sort_io_queue = -1;
1111 bioq_init(&(*iscp)->bio_queue);
1112 bioq_init(&(*iscp)->trim_queue);
1113 #ifdef CAM_IOSCHED_DYNAMIC
1114 if (do_dynamic_iosched) {
1115 bioq_init(&(*iscp)->write_queue);
1116 (*iscp)->read_bias = 100;
1117 (*iscp)->current_read_bias = 100;
1118 (*iscp)->quanta = min(hz, 200);
1119 cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1120 cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1121 cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1122 (*iscp)->trim_stats.max = 1; /* Trims are special: one at a time for now */
1123 (*iscp)->last_time = sbinuptime();
1124 callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1125 (*iscp)->periph = periph;
1126 cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1127 callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp);
1128 (*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1136 * Reclaim all used resources. This assumes that other folks have
1137 * drained the requests in the hardware. Maybe an unwise assumption.
1140 cam_iosched_fini(struct cam_iosched_softc *isc)
1143 cam_iosched_flush(isc, NULL, ENXIO);
1144 #ifdef CAM_IOSCHED_DYNAMIC
1145 cam_iosched_iop_stats_fini(&isc->read_stats);
1146 cam_iosched_iop_stats_fini(&isc->write_stats);
1147 cam_iosched_iop_stats_fini(&isc->trim_stats);
1148 cam_iosched_cl_sysctl_fini(&isc->cl);
1149 if (isc->sysctl_tree)
1150 if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1151 printf("can't remove iosched sysctl stats context\n");
1152 if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1153 callout_drain(&isc->ticker);
1154 isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1157 free(isc, M_CAMSCHED);
1162 * After we're sure we're attaching a device, go ahead and add
1163 * hooks for any sysctl we may wish to honor.
1165 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1166 struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1168 struct sysctl_oid_list *n;
1170 n = SYSCTL_CHILDREN(node);
1171 SYSCTL_ADD_INT(ctx, n,
1172 OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1173 &isc->sort_io_queue, 0,
1174 "Sort IO queue to try and optimise disk access patterns");
1175 SYSCTL_ADD_INT(ctx, n,
1176 OID_AUTO, "trim_goal", CTLFLAG_RW,
1178 "Number of trims to try to accumulate before sending to hardware");
1179 SYSCTL_ADD_INT(ctx, n,
1180 OID_AUTO, "trim_ticks", CTLFLAG_RW,
1182 "IO Schedul qaunta to hold back trims for when accumulating");
1184 #ifdef CAM_IOSCHED_DYNAMIC
1185 if (!do_dynamic_iosched)
1188 isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1189 SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1190 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "I/O scheduler statistics");
1191 n = SYSCTL_CHILDREN(isc->sysctl_tree);
1192 ctx = &isc->sysctl_ctx;
1194 cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1195 cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1196 cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1197 cam_iosched_cl_sysctl_init(isc);
1199 SYSCTL_ADD_INT(ctx, n,
1200 OID_AUTO, "read_bias", CTLFLAG_RW,
1201 &isc->read_bias, 100,
1202 "How biased towards read should we be independent of limits");
1204 SYSCTL_ADD_PROC(ctx, n,
1205 OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
1206 &isc->quanta, 0, cam_iosched_quanta_sysctl, "I",
1207 "How many quanta per second do we slice the I/O up into");
1209 SYSCTL_ADD_INT(ctx, n,
1210 OID_AUTO, "total_ticks", CTLFLAG_RD,
1211 &isc->total_ticks, 0,
1212 "Total number of ticks we've done");
1214 SYSCTL_ADD_INT(ctx, n,
1215 OID_AUTO, "load", CTLFLAG_RD,
1217 "scaled load average / 100");
1219 SYSCTL_ADD_U64(ctx, n,
1220 OID_AUTO, "latency_trigger", CTLFLAG_RW,
1222 "Latency treshold to trigger callbacks");
1227 cam_iosched_set_latfcn(struct cam_iosched_softc *isc,
1228 cam_iosched_latfcn_t fnp, void *argp)
1230 #ifdef CAM_IOSCHED_DYNAMIC
1237 * Client drivers can set two parameters. "goal" is the number of BIO_DELETEs
1238 * that will be queued up before iosched will "release" the trims to the client
1239 * driver to wo with what they will (usually combine as many as possible). If we
1240 * don't get this many, after trim_ticks we'll submit the I/O anyway with
1241 * whatever we have. We do need an I/O of some kind of to clock the deferred
1242 * trims out to disk. Since we will eventually get a write for the super block
1243 * or something before we shutdown, the trims will complete. To be safe, when a
1244 * BIO_FLUSH is presented to the iosched work queue, we set the ticks time far
1245 * enough in the past so we'll present the BIO_DELETEs to the client driver.
1246 * There might be a race if no BIO_DELETESs were queued, a BIO_FLUSH comes in
1247 * and then a BIO_DELETE is sent down. No know client does this, and there's
1248 * already a race between an ordered BIO_FLUSH and any BIO_DELETEs in flight,
1249 * but no client depends on the ordering being honored.
1251 * XXX I'm not sure what the interaction between UFS direct BIOs and the BUF
1252 * flushing on shutdown. I think there's bufs that would be dependent on the BIO
1253 * finishing to write out at least metadata, so we'll be fine. To be safe, keep
1254 * the number of ticks low (less than maybe 10s) to avoid shutdown races.
1258 cam_iosched_set_trim_goal(struct cam_iosched_softc *isc, int goal)
1261 isc->trim_goal = goal;
1265 cam_iosched_set_trim_ticks(struct cam_iosched_softc *isc, int trim_ticks)
1268 isc->trim_ticks = trim_ticks;
1272 * Flush outstanding I/O. Consumers of this library don't know all the
1273 * queues we may keep, so this allows all I/O to be flushed in one
1277 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1279 bioq_flush(&isc->bio_queue, stp, err);
1280 bioq_flush(&isc->trim_queue, stp, err);
1281 #ifdef CAM_IOSCHED_DYNAMIC
1282 if (do_dynamic_iosched)
1283 bioq_flush(&isc->write_queue, stp, err);
1287 #ifdef CAM_IOSCHED_DYNAMIC
1289 cam_iosched_get_write(struct cam_iosched_softc *isc)
1294 * We control the write rate by controlling how many requests we send
1295 * down to the drive at any one time. Fewer requests limits the
1296 * effects of both starvation when the requests take a while and write
1297 * amplification when each request is causing more than one write to
1298 * the NAND media. Limiting the queue depth like this will also limit
1299 * the write throughput and give and reads that want to compete to
1302 bp = bioq_first(&isc->write_queue);
1304 if (iosched_debug > 3)
1305 printf("No writes present in write_queue\n");
1310 * If pending read, prefer that based on current read bias
1313 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1316 "Reads present and current_read_bias is %d queued "
1317 "writes %d queued reads %d\n",
1318 isc->current_read_bias, isc->write_stats.queued,
1319 isc->read_stats.queued);
1320 isc->current_read_bias--;
1321 /* We're not limiting writes, per se, just doing reads first */
1326 * See if our current limiter allows this I/O.
1328 if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1330 printf("Can't write because limiter says no.\n");
1331 isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1336 * Let's do this: We've passed all the gates and we're a go
1337 * to schedule the I/O in the SIM.
1339 isc->current_read_bias = isc->read_bias;
1340 bioq_remove(&isc->write_queue, bp);
1341 if (bp->bio_cmd == BIO_WRITE) {
1342 isc->write_stats.queued--;
1343 isc->write_stats.total++;
1344 isc->write_stats.pending++;
1346 if (iosched_debug > 9)
1347 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1348 isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1354 * Put back a trim that you weren't able to actually schedule this time.
1357 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1359 bioq_insert_head(&isc->trim_queue, bp);
1360 if (isc->queued_trims == 0)
1361 isc->last_trim_tick = ticks;
1362 isc->queued_trims++;
1363 #ifdef CAM_IOSCHED_DYNAMIC
1364 isc->trim_stats.queued++;
1365 isc->trim_stats.total--; /* since we put it back, don't double count */
1366 isc->trim_stats.pending--;
1371 * gets the next trim from the trim queue.
1373 * Assumes we're called with the periph lock held. It removes this
1374 * trim from the queue and the device must explicitly reinsert it
1375 * should the need arise.
1378 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1382 bp = bioq_first(&isc->trim_queue);
1385 bioq_remove(&isc->trim_queue, bp);
1386 isc->queued_trims--;
1387 isc->last_trim_tick = ticks; /* Reset the tick timer when we take trims */
1388 #ifdef CAM_IOSCHED_DYNAMIC
1389 isc->trim_stats.queued--;
1390 isc->trim_stats.total++;
1391 isc->trim_stats.pending++;
1397 * gets an available trim from the trim queue, if there's no trim
1398 * already pending. It removes this trim from the queue and the device
1399 * must explicitly reinsert it should the need arise.
1401 * Assumes we're called with the periph lock held.
1404 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1406 #ifdef CAM_IOSCHED_DYNAMIC
1410 if (!cam_iosched_has_more_trim(isc))
1412 #ifdef CAM_IOSCHED_DYNAMIC
1413 bp = bioq_first(&isc->trim_queue);
1418 * If pending read, prefer that based on current read bias setting. The
1419 * read bias is shared for both writes and TRIMs, but on TRIMs the bias
1420 * is for a combined TRIM not a single TRIM request that's come in.
1422 if (do_dynamic_iosched) {
1423 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1425 printf("Reads present and current_read_bias is %d"
1426 " queued trims %d queued reads %d\n",
1427 isc->current_read_bias, isc->trim_stats.queued,
1428 isc->read_stats.queued);
1429 isc->current_read_bias--;
1430 /* We're not limiting TRIMS, per se, just doing reads first */
1434 * We're going to do a trim, so reset the bias.
1436 isc->current_read_bias = isc->read_bias;
1440 * See if our current limiter allows this I/O. Because we only call this
1441 * here, and not in next_trim, the 'bandwidth' limits for trims won't
1442 * work, while the iops or max queued limits will work. It's tricky
1443 * because we want the limits to be from the perspective of the
1444 * "commands sent to the device." To make iops work, we need to check
1445 * only here (since we want all the ops we combine to count as one). To
1446 * make bw limits work, we'd need to check in next_trim, but that would
1447 * have the effect of limiting the iops as seen from the upper layers.
1449 if (cam_iosched_limiter_iop(&isc->trim_stats, bp) != 0) {
1451 printf("Can't trim because limiter says no.\n");
1452 isc->trim_stats.state_flags |= IOP_RATE_LIMITED;
1455 isc->current_read_bias = isc->read_bias;
1456 isc->trim_stats.state_flags &= ~IOP_RATE_LIMITED;
1457 /* cam_iosched_next_trim below keeps proper book */
1459 return cam_iosched_next_trim(isc);
1463 * Determine what the next bit of work to do is for the periph. The
1464 * default implementation looks to see if we have trims to do, but no
1465 * trims outstanding. If so, we do that. Otherwise we see if we have
1466 * other work. If we do, then we do that. Otherwise why were we called?
1469 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1474 * See if we have a trim that can be scheduled. We can only send one
1475 * at a time down, so this takes that into account.
1477 * XXX newer TRIM commands are queueable. Revisit this when we
1480 if ((bp = cam_iosched_get_trim(isc)) != NULL)
1483 #ifdef CAM_IOSCHED_DYNAMIC
1485 * See if we have any pending writes, and room in the queue for them,
1486 * and if so, those are next.
1488 if (do_dynamic_iosched) {
1489 if ((bp = cam_iosched_get_write(isc)) != NULL)
1495 * next, see if there's other, normal I/O waiting. If so return that.
1497 if ((bp = bioq_first(&isc->bio_queue)) == NULL)
1500 #ifdef CAM_IOSCHED_DYNAMIC
1502 * For the dynamic scheduler, bio_queue is only for reads, so enforce
1503 * the limits here. Enforce only for reads.
1505 if (do_dynamic_iosched) {
1506 if (bp->bio_cmd == BIO_READ &&
1507 cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1508 isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1512 isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1514 bioq_remove(&isc->bio_queue, bp);
1515 #ifdef CAM_IOSCHED_DYNAMIC
1516 if (do_dynamic_iosched) {
1517 if (bp->bio_cmd == BIO_READ) {
1518 isc->read_stats.queued--;
1519 isc->read_stats.total++;
1520 isc->read_stats.pending++;
1522 printf("Found bio_cmd = %#x\n", bp->bio_cmd);
1524 if (iosched_debug > 9)
1525 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1531 * Driver has been given some work to do by the block layer. Tell the
1532 * scheduler about it and have it queue the work up. The scheduler module
1533 * will then return the currently most useful bit of work later, possibly
1534 * deferring work for various reasons.
1537 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1541 * A BIO_SPEEDUP from the uppper layers means that they have a block
1542 * shortage. At the present, this is only sent when we're trying to
1543 * allocate blocks, but have a shortage before giving up. bio_length is
1544 * the size of their shortage. We will complete just enough BIO_DELETEs
1545 * in the queue to satisfy the need. If bio_length is 0, we'll complete
1546 * them all. This allows the scheduler to delay BIO_DELETEs to improve
1547 * read/write performance without worrying about the upper layers. When
1548 * it's possibly a problem, we respond by pretending the BIO_DELETEs
1549 * just worked. We can't do anything about the BIO_DELETEs in the
1550 * hardware, though. We have to wait for them to complete.
1552 if (bp->bio_cmd == BIO_SPEEDUP) {
1557 while (bioq_first(&isc->trim_queue) &&
1558 (bp->bio_length == 0 || len < bp->bio_length)) {
1559 nbp = bioq_takefirst(&isc->trim_queue);
1560 len += nbp->bio_length;
1564 if (bp->bio_length > 0) {
1565 if (bp->bio_length > len)
1566 bp->bio_resid = bp->bio_length - len;
1576 * If we get a BIO_FLUSH, and we're doing delayed BIO_DELETEs then we
1577 * set the last tick time to one less than the current ticks minus the
1578 * delay to force the BIO_DELETEs to be presented to the client driver.
1580 if (bp->bio_cmd == BIO_FLUSH && isc->trim_ticks > 0)
1581 isc->last_trim_tick = ticks - isc->trim_ticks - 1;
1584 * Put all trims on the trim queue. Otherwise put the work on the bio
1587 if (bp->bio_cmd == BIO_DELETE) {
1588 bioq_insert_tail(&isc->trim_queue, bp);
1589 if (isc->queued_trims == 0)
1590 isc->last_trim_tick = ticks;
1591 isc->queued_trims++;
1592 #ifdef CAM_IOSCHED_DYNAMIC
1593 isc->trim_stats.in++;
1594 isc->trim_stats.queued++;
1597 #ifdef CAM_IOSCHED_DYNAMIC
1598 else if (do_dynamic_iosched && (bp->bio_cmd != BIO_READ)) {
1599 if (cam_iosched_sort_queue(isc))
1600 bioq_disksort(&isc->write_queue, bp);
1602 bioq_insert_tail(&isc->write_queue, bp);
1603 if (iosched_debug > 9)
1604 printf("Qw : %p %#x\n", bp, bp->bio_cmd);
1605 if (bp->bio_cmd == BIO_WRITE) {
1606 isc->write_stats.in++;
1607 isc->write_stats.queued++;
1612 if (cam_iosched_sort_queue(isc))
1613 bioq_disksort(&isc->bio_queue, bp);
1615 bioq_insert_tail(&isc->bio_queue, bp);
1616 #ifdef CAM_IOSCHED_DYNAMIC
1617 if (iosched_debug > 9)
1618 printf("Qr : %p %#x\n", bp, bp->bio_cmd);
1619 if (bp->bio_cmd == BIO_READ) {
1620 isc->read_stats.in++;
1621 isc->read_stats.queued++;
1622 } else if (bp->bio_cmd == BIO_WRITE) {
1623 isc->write_stats.in++;
1624 isc->write_stats.queued++;
1631 * If we have work, get it scheduled. Called with the periph lock held.
1634 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1637 if (cam_iosched_has_work(isc))
1638 xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1642 * Complete a trim request. Mark that we no longer have one in flight.
1645 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1648 isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1652 * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1653 * might use notes in the ccb for statistics.
1656 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1657 union ccb *done_ccb)
1660 #ifdef CAM_IOSCHED_DYNAMIC
1661 if (!do_dynamic_iosched)
1664 if (iosched_debug > 10)
1665 printf("done: %p %#x\n", bp, bp->bio_cmd);
1666 if (bp->bio_cmd == BIO_WRITE) {
1667 retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1668 if ((bp->bio_flags & BIO_ERROR) != 0)
1669 isc->write_stats.errs++;
1670 isc->write_stats.out++;
1671 isc->write_stats.pending--;
1672 } else if (bp->bio_cmd == BIO_READ) {
1673 retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1674 if ((bp->bio_flags & BIO_ERROR) != 0)
1675 isc->read_stats.errs++;
1676 isc->read_stats.out++;
1677 isc->read_stats.pending--;
1678 } else if (bp->bio_cmd == BIO_DELETE) {
1679 if ((bp->bio_flags & BIO_ERROR) != 0)
1680 isc->trim_stats.errs++;
1681 isc->trim_stats.out++;
1682 isc->trim_stats.pending--;
1683 } else if (bp->bio_cmd != BIO_FLUSH) {
1685 printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1688 if (!(bp->bio_flags & BIO_ERROR) && done_ccb != NULL) {
1689 sbintime_t sim_latency;
1691 sim_latency = cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data);
1693 cam_iosched_io_metric_update(isc, sim_latency,
1694 bp->bio_cmd, bp->bio_bcount);
1696 * Debugging code: allow callbacks to the periph driver when latency max
1697 * is exceeded. This can be useful for triggering external debugging actions.
1699 if (isc->latfcn && isc->max_lat != 0 && sim_latency > isc->max_lat)
1700 isc->latfcn(isc->latarg, sim_latency, bp);
1708 * Tell the io scheduler that you've pushed a trim down into the sim.
1709 * This also tells the I/O scheduler not to push any more trims down, so
1710 * some periphs do not call it if they can cope with multiple trims in flight.
1713 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1716 isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1720 * Change the sorting policy hint for I/O transactions for this device.
1723 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1726 isc->sort_io_queue = val;
1730 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1732 return isc->flags & flags;
1736 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1738 isc->flags |= flags;
1742 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1744 isc->flags &= ~flags;
1747 #ifdef CAM_IOSCHED_DYNAMIC
1749 * After the method presented in Jack Crenshaw's 1998 article "Integer
1750 * Square Roots," reprinted at
1751 * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1752 * and well worth the read. Briefly, we find the power of 4 that's the
1753 * largest smaller than val. We then check each smaller power of 4 to
1754 * see if val is still bigger. The right shifts at each step divide
1755 * the result by 2 which after successive application winds up
1756 * accumulating the right answer. It could also have been accumulated
1757 * using a separate root counter, but this code is smaller and faster
1758 * than that method. This method is also integer size invariant.
1759 * It returns floor(sqrt((float)val)), or the largest integer less than
1760 * or equal to the square root.
1763 isqrt64(uint64_t val)
1766 uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1769 * Find the largest power of 4 smaller than val.
1775 * Accumulate the answer, one bit at a time (we keep moving
1776 * them over since 2 is the square root of 4 and we test
1777 * powers of 4). We accumulate where we find the bit, but
1778 * the successive shifts land the bit in the right place
1782 if (val >= res + bit) {
1784 res = (res >> 1) + bit;
1793 static sbintime_t latencies[LAT_BUCKETS - 1] = {
1807 SBT_1MS << 13 /* 8.192s */
1811 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1813 sbintime_t y, deltasq, delta;
1817 * Keep counts for latency. We do it by power of two buckets.
1818 * This helps us spot outlier behavior obscured by averages.
1820 for (i = 0; i < LAT_BUCKETS - 1; i++) {
1821 if (sim_latency < latencies[i]) {
1822 iop->latencies[i]++;
1826 if (i == LAT_BUCKETS - 1)
1827 iop->latencies[i]++; /* Put all > 1024ms values into the last bucket. */
1830 * Classic exponentially decaying average with a tiny alpha
1831 * (2 ^ -alpha_bits). For more info see the NIST statistical
1834 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha) [nist]
1835 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
1836 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1837 * alpha = 1 / (1 << alpha_bits)
1838 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
1839 * = y_t/b - e/b + be/b
1840 * = (y_t - e + be) / b
1843 * Since alpha is a power of two, we can compute this w/o any mult or
1846 * Variance can also be computed. Usually, it would be expressed as follows:
1847 * diff_t = y_t - ema_t-1
1848 * emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
1849 * = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
1850 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
1851 * = e - e/b + dd/b + dd/bb
1852 * = (bbe - be + bdd + dd) / bb
1853 * = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
1856 * XXX possible numeric issues
1857 * o We assume right shifted integers do the right thing, since that's
1858 * implementation defined. You can change the right shifts to / (1LL << alpha).
1859 * o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
1860 * for emvar. This puts a ceiling of 13 bits on alpha since we need a
1861 * few tens of seconds of representation.
1862 * o We mitigate alpha issues by never setting it too high.
1865 delta = (y - iop->ema); /* d */
1866 iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1869 * Were we to naively plow ahead at this point, we wind up with many numerical
1870 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
1871 * us with microsecond level precision in the input, so the same in the
1872 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
1873 * also means that emvar can be up 46 bits 40 of which are fraction, which
1874 * gives us a way to measure up to ~8s in the SD before the computation goes
1875 * unstable. Even the worst hard disk rarely has > 1s service time in the
1876 * drive. It does mean we have to shift left 12 bits after taking the
1877 * square root to compute the actual standard deviation estimate. This loss of
1878 * precision is preferable to needing int128 types to work. The above numbers
1879 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
1880 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1883 deltasq = delta * delta; /* dd */
1884 iop->emvar = ((iop->emvar << (2 * alpha_bits)) + /* bbe */
1885 ((deltasq - iop->emvar) << alpha_bits) + /* b(dd-e) */
1887 >> (2 * alpha_bits); /* div bb */
1888 iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1892 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1893 sbintime_t sim_latency, int cmd, size_t size)
1895 /* xxx Do we need to scale based on the size of the I/O ? */
1898 cam_iosched_update(&isc->read_stats, sim_latency);
1901 cam_iosched_update(&isc->write_stats, sim_latency);
1904 cam_iosched_update(&isc->trim_stats, sim_latency);
1912 static int biolen(struct bio_queue_head *bq)
1917 TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
1924 * Show the internal state of the I/O scheduler.
1926 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
1928 struct cam_iosched_softc *isc;
1931 db_printf("Need addr\n");
1934 isc = (struct cam_iosched_softc *)addr;
1935 db_printf("pending_reads: %d\n", isc->read_stats.pending);
1936 db_printf("min_reads: %d\n", isc->read_stats.min);
1937 db_printf("max_reads: %d\n", isc->read_stats.max);
1938 db_printf("reads: %d\n", isc->read_stats.total);
1939 db_printf("in_reads: %d\n", isc->read_stats.in);
1940 db_printf("out_reads: %d\n", isc->read_stats.out);
1941 db_printf("queued_reads: %d\n", isc->read_stats.queued);
1942 db_printf("Read Q len %d\n", biolen(&isc->bio_queue));
1943 db_printf("pending_writes: %d\n", isc->write_stats.pending);
1944 db_printf("min_writes: %d\n", isc->write_stats.min);
1945 db_printf("max_writes: %d\n", isc->write_stats.max);
1946 db_printf("writes: %d\n", isc->write_stats.total);
1947 db_printf("in_writes: %d\n", isc->write_stats.in);
1948 db_printf("out_writes: %d\n", isc->write_stats.out);
1949 db_printf("queued_writes: %d\n", isc->write_stats.queued);
1950 db_printf("Write Q len %d\n", biolen(&isc->write_queue));
1951 db_printf("pending_trims: %d\n", isc->trim_stats.pending);
1952 db_printf("min_trims: %d\n", isc->trim_stats.min);
1953 db_printf("max_trims: %d\n", isc->trim_stats.max);
1954 db_printf("trims: %d\n", isc->trim_stats.total);
1955 db_printf("in_trims: %d\n", isc->trim_stats.in);
1956 db_printf("out_trims: %d\n", isc->trim_stats.out);
1957 db_printf("queued_trims: %d\n", isc->trim_stats.queued);
1958 db_printf("Trim Q len %d\n", biolen(&isc->trim_queue));
1959 db_printf("read_bias: %d\n", isc->read_bias);
1960 db_printf("current_read_bias: %d\n", isc->current_read_bias);
1961 db_printf("Trim active? %s\n",
1962 (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");