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 bool do_dynamic_iosched = 1;
74 SYSCTL_BOOL(_kern_cam, OID_AUTO, do_dynamic_iosched, CTLFLAG_RD | CTLFLAG_TUN,
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 SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW | CTLFLAG_TUN,
101 "Bits in EMA's alpha.");
104 struct cam_iosched_softc;
106 int iosched_debug = 0;
109 none = 0, /* No limits */
110 queue_depth, /* Limit how many ops we queue to SIM */
111 iops, /* Limit # of IOPS to the drive */
112 bandwidth, /* Limit bandwidth to the drive */
116 static const char *cam_iosched_limiter_names[] =
117 { "none", "queue_depth", "iops", "bandwidth" };
120 * Called to initialize the bits of the iop_stats structure relevant to the
121 * limiter. Called just after the limiter is set.
123 typedef int l_init_t(struct iop_stats *);
128 typedef int l_tick_t(struct iop_stats *);
131 * Called to see if the limiter thinks this IOP can be allowed to
132 * proceed. If so, the limiter assumes that the IOP proceeded
133 * and makes any accounting of it that's needed.
135 typedef int l_iop_t(struct iop_stats *, struct bio *);
138 * Called when an I/O completes so the limiter can update its
139 * accounting. Pending I/Os may complete in any order (even when
140 * sent to the hardware at the same time), so the limiter may not
141 * make any assumptions other than this I/O has completed. If it
142 * returns 1, then xpt_schedule() needs to be called again.
144 typedef int l_iodone_t(struct iop_stats *, struct bio *);
146 static l_iop_t cam_iosched_qd_iop;
147 static l_iop_t cam_iosched_qd_caniop;
148 static l_iodone_t cam_iosched_qd_iodone;
150 static l_init_t cam_iosched_iops_init;
151 static l_tick_t cam_iosched_iops_tick;
152 static l_iop_t cam_iosched_iops_caniop;
153 static l_iop_t cam_iosched_iops_iop;
155 static l_init_t cam_iosched_bw_init;
156 static l_tick_t cam_iosched_bw_tick;
157 static l_iop_t cam_iosched_bw_caniop;
158 static l_iop_t cam_iosched_bw_iop;
165 l_iodone_t *l_iodone;
177 .l_caniop = cam_iosched_qd_caniop,
178 .l_iop = cam_iosched_qd_iop,
179 .l_iodone= cam_iosched_qd_iodone,
182 .l_init = cam_iosched_iops_init,
183 .l_tick = cam_iosched_iops_tick,
184 .l_caniop = cam_iosched_iops_caniop,
185 .l_iop = cam_iosched_iops_iop,
189 .l_init = cam_iosched_bw_init,
190 .l_tick = cam_iosched_bw_tick,
191 .l_caniop = cam_iosched_bw_caniop,
192 .l_iop = cam_iosched_bw_iop,
199 * sysctl state for this subnode.
201 struct sysctl_ctx_list sysctl_ctx;
202 struct sysctl_oid *sysctl_tree;
205 * Information about the current rate limiters, if any
207 io_limiter limiter; /* How are I/Os being limited */
208 int min; /* Low range of limit */
209 int max; /* High range of limit */
210 int current; /* Current rate limiter */
211 int l_value1; /* per-limiter scratch value 1. */
212 int l_value2; /* per-limiter scratch value 2. */
215 * Debug information about counts of I/Os that have gone through the
218 int pending; /* I/Os pending in the hardware */
219 int queued; /* number currently in the queue */
220 int total; /* Total for all time -- wraps */
221 int in; /* number queued all time -- wraps */
222 int out; /* number completed all time -- wraps */
223 int errs; /* Number of I/Os completed with error -- 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;
243 set_max = 0, /* current = max */
244 read_latency, /* Steer read latency by throttling writes */
245 cl_max /* Keep last */
248 static const char *cam_iosched_control_type_names[] =
249 { "set_max", "read_latency" };
251 struct control_loop {
253 * sysctl state for this subnode.
255 struct sysctl_ctx_list sysctl_ctx;
256 struct sysctl_oid *sysctl_tree;
258 sbintime_t next_steer; /* Time of next steer */
259 sbintime_t steer_interval; /* How often do we steer? */
263 control_type type; /* What type of control? */
264 int last_count; /* Last I/O count */
266 struct cam_iosched_softc *softc;
271 struct cam_iosched_softc {
272 struct bio_queue_head bio_queue;
273 struct bio_queue_head trim_queue;
274 /* scheduler flags < 16, user flags >= 16 */
277 int trim_goal; /* # of trims to queue before sending */
278 int trim_ticks; /* Max ticks to hold trims */
279 int last_trim_tick; /* Last 'tick' time ld a trim */
280 int queued_trims; /* Number of trims in the queue */
281 #ifdef CAM_IOSCHED_DYNAMIC
282 int read_bias; /* Read bias setting */
283 int current_read_bias; /* Current read bias state */
285 int load; /* EMA of 'load average' of disk / 2^16 */
287 struct bio_queue_head write_queue;
288 struct iop_stats read_stats, write_stats, trim_stats;
289 struct sysctl_ctx_list sysctl_ctx;
290 struct sysctl_oid *sysctl_tree;
292 int quanta; /* Number of quanta per second */
293 struct callout ticker; /* Callout for our quota system */
294 struct cam_periph *periph; /* cam periph associated with this device */
295 uint32_t this_frac; /* Fraction of a second (1024ths) for this tick */
296 sbintime_t last_time; /* Last time we ticked */
297 struct control_loop cl;
298 sbintime_t max_lat; /* when != 0, if iop latency > max_lat, call max_lat_fcn */
299 cam_iosched_latfcn_t latfcn;
304 #ifdef CAM_IOSCHED_DYNAMIC
306 * helper functions to call the limsw functions.
309 cam_iosched_limiter_init(struct iop_stats *ios)
311 int lim = ios->limiter;
313 /* maybe this should be a kassert */
314 if (lim < none || lim >= limiter_max)
317 if (limsw[lim].l_init)
318 return limsw[lim].l_init(ios);
324 cam_iosched_limiter_tick(struct iop_stats *ios)
326 int lim = ios->limiter;
328 /* maybe this should be a kassert */
329 if (lim < none || lim >= limiter_max)
332 if (limsw[lim].l_tick)
333 return limsw[lim].l_tick(ios);
339 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
341 int lim = ios->limiter;
343 /* maybe this should be a kassert */
344 if (lim < none || lim >= limiter_max)
347 if (limsw[lim].l_iop)
348 return limsw[lim].l_iop(ios, bp);
354 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
356 int lim = ios->limiter;
358 /* maybe this should be a kassert */
359 if (lim < none || lim >= limiter_max)
362 if (limsw[lim].l_caniop)
363 return limsw[lim].l_caniop(ios, bp);
369 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
371 int lim = ios->limiter;
373 /* maybe this should be a kassert */
374 if (lim < none || lim >= limiter_max)
377 if (limsw[lim].l_iodone)
378 return limsw[lim].l_iodone(ios, bp);
384 * Functions to implement the different kinds of limiters
388 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
391 if (ios->current <= 0 || ios->pending < ios->current)
398 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
401 if (ios->current <= 0 || ios->pending < ios->current)
408 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
411 if (ios->current <= 0 || ios->pending != ios->current)
418 cam_iosched_iops_init(struct iop_stats *ios)
421 ios->l_value1 = ios->current / ios->softc->quanta;
422 if (ios->l_value1 <= 0)
430 cam_iosched_iops_tick(struct iop_stats *ios)
435 * Allow at least one IO per tick until all
436 * the IOs for this interval have been spent.
438 new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
439 if (new_ios < 1 && ios->l_value2 < ios->current) {
445 * If this a new accounting interval, discard any "unspent" ios
446 * granted in the previous interval. Otherwise add the new ios to
447 * the previously granted ones that haven't been spent yet.
449 if ((ios->softc->total_ticks % ios->softc->quanta) == 0) {
450 ios->l_value1 = new_ios;
453 ios->l_value1 += new_ios;
460 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
464 * So if we have any more IOPs left, allow it,
465 * otherwise wait. If current iops is 0, treat that
466 * as unlimited as a failsafe.
468 if (ios->current > 0 && ios->l_value1 <= 0)
474 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
478 rv = cam_iosched_limiter_caniop(ios, bp);
486 cam_iosched_bw_init(struct iop_stats *ios)
489 /* ios->current is in kB/s, so scale to bytes */
490 ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
496 cam_iosched_bw_tick(struct iop_stats *ios)
501 * If we're in the hole for available quota from
502 * the last time, then add the quantum for this.
503 * If we have any left over from last quantum,
504 * then too bad, that's lost. Also, ios->current
505 * is in kB/s, so scale.
507 * We also allow up to 4 quanta of credits to
508 * accumulate to deal with burstiness. 4 is extremely
511 bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
512 if (ios->l_value1 < bw * 4)
519 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
522 * So if we have any more bw quota left, allow it,
523 * otherwise wait. Note, we'll go negative and that's
524 * OK. We'll just get a little less next quota.
526 * Note on going negative: that allows us to process
527 * requests in order better, since we won't allow
528 * shorter reads to get around the long one that we
529 * don't have the quota to do just yet. It also prevents
530 * starvation by being a little more permissive about
531 * what we let through this quantum (to prevent the
532 * starvation), at the cost of getting a little less
535 * Also note that if the current limit is <= 0,
536 * we treat it as unlimited as a failsafe.
538 if (ios->current > 0 && ios->l_value1 <= 0)
545 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
549 rv = cam_iosched_limiter_caniop(ios, bp);
551 ios->l_value1 -= bp->bio_length;
556 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
559 cam_iosched_ticker(void *arg)
561 struct cam_iosched_softc *isc = arg;
562 sbintime_t now, delta;
565 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
568 delta = now - isc->last_time;
569 isc->this_frac = (uint32_t)delta >> 16; /* Note: discards seconds -- should be 0 harmless if not */
570 isc->last_time = now;
572 cam_iosched_cl_maybe_steer(&isc->cl);
574 cam_iosched_limiter_tick(&isc->read_stats);
575 cam_iosched_limiter_tick(&isc->write_stats);
576 cam_iosched_limiter_tick(&isc->trim_stats);
578 cam_iosched_schedule(isc, isc->periph);
581 * isc->load is an EMA of the pending I/Os at each tick. The number of
582 * pending I/Os is the sum of the I/Os queued to the hardware, and those
583 * in the software queue that could be queued to the hardware if there
586 * ios_stats.pending is a count of requests in the SIM right now for
587 * each of these types of I/O. So the total pending count is the sum of
588 * these I/Os and the sum of the queued I/Os still in the software queue
589 * for those operations that aren't being rate limited at the moment.
591 * The reason for the rate limiting bit is because those I/Os
592 * aren't part of the software queued load (since we could
593 * give them to hardware, but choose not to).
595 * Note: due to a bug in counting pending TRIM in the device, we
596 * don't include them in this count. We count each BIO_DELETE in
597 * the pending count, but the periph drivers collapse them down
598 * into one TRIM command. That one trim command gets the completion
599 * so the counts get off.
601 pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
602 pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
603 !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
604 !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
606 pending /= isc->periph->path->device->ccbq.total_openings;
608 isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
614 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
617 clp->next_steer = sbinuptime();
619 clp->steer_interval = SBT_1S * 5; /* Let's start out steering every 5s */
620 clp->lolat = 5 * SBT_1MS;
621 clp->hilat = 15 * SBT_1MS;
622 clp->alpha = 20; /* Alpha == gain. 20 = .2 */
627 cam_iosched_cl_maybe_steer(struct control_loop *clp)
629 struct cam_iosched_softc *isc;
634 now = isc->last_time;
635 if (now < clp->next_steer)
638 clp->next_steer = now + clp->steer_interval;
641 if (isc->write_stats.current != isc->write_stats.max)
642 printf("Steering write from %d kBps to %d kBps\n",
643 isc->write_stats.current, isc->write_stats.max);
644 isc->read_stats.current = isc->read_stats.max;
645 isc->write_stats.current = isc->write_stats.max;
646 isc->trim_stats.current = isc->trim_stats.max;
649 old = isc->write_stats.current;
650 lat = isc->read_stats.ema;
652 * Simple PLL-like engine. Since we're steering to a range for
653 * the SP (set point) that makes things a little more
654 * complicated. In addition, we're not directly controlling our
655 * PV (process variable), the read latency, but instead are
656 * manipulating the write bandwidth limit for our MV
657 * (manipulation variable), analysis of this code gets a bit
658 * messy. Also, the MV is a very noisy control surface for read
659 * latency since it is affected by many hidden processes inside
660 * the device which change how responsive read latency will be
661 * in reaction to changes in write bandwidth. Unlike the classic
662 * boiler control PLL. this may result in over-steering while
663 * the SSD takes its time to react to the new, lower load. This
664 * is why we use a relatively low alpha of between .1 and .25 to
665 * compensate for this effect. At .1, it takes ~22 steering
666 * intervals to back off by a factor of 10. At .2 it only takes
667 * ~10. At .25 it only takes ~8. However some preliminary data
668 * from the SSD drives suggests a reasponse time in 10's of
669 * seconds before latency drops regardless of the new write
670 * rate. Careful observation will be required to tune this
673 * Also, when there's no read traffic, we jack up the write
674 * limit too regardless of the last read latency. 10 is
675 * somewhat arbitrary.
677 if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
678 isc->write_stats.current = isc->write_stats.current *
679 (100 + clp->alpha) / 100; /* Scale up */
680 else if (lat > clp->hilat)
681 isc->write_stats.current = isc->write_stats.current *
682 (100 - clp->alpha) / 100; /* Scale down */
683 clp->last_count = isc->read_stats.total;
686 * Even if we don't steer, per se, enforce the min/max limits as
687 * those may have changed.
689 if (isc->write_stats.current < isc->write_stats.min)
690 isc->write_stats.current = isc->write_stats.min;
691 if (isc->write_stats.current > isc->write_stats.max)
692 isc->write_stats.current = isc->write_stats.max;
693 if (old != isc->write_stats.current && iosched_debug)
694 printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
695 old, isc->write_stats.current,
696 (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
705 * Trim or similar currently pending completion. Should only be set for
706 * those drivers wishing only one Trim active at a time.
708 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE (1ul << 0)
709 /* Callout active, and needs to be torn down */
710 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
712 /* Periph drivers set these flags to indicate work */
713 #define CAM_IOSCHED_FLAG_WORK_FLAGS ((0xffffu) << 16)
715 #ifdef CAM_IOSCHED_DYNAMIC
717 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
718 sbintime_t sim_latency, int cmd, size_t size);
722 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
724 return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
728 cam_iosched_has_io(struct cam_iosched_softc *isc)
730 #ifdef CAM_IOSCHED_DYNAMIC
731 if (do_dynamic_iosched) {
732 struct bio *rbp = bioq_first(&isc->bio_queue);
733 struct bio *wbp = bioq_first(&isc->write_queue);
734 bool can_write = wbp != NULL &&
735 cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
736 bool can_read = rbp != NULL &&
737 cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
738 if (iosched_debug > 2) {
739 printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
740 printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
741 printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
743 return can_read || can_write;
746 return bioq_first(&isc->bio_queue) != NULL;
750 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
754 bp = bioq_first(&isc->trim_queue);
755 #ifdef CAM_IOSCHED_DYNAMIC
756 if (do_dynamic_iosched) {
758 * If we're limiting trims, then defer action on trims
761 if (bp == NULL || cam_iosched_limiter_caniop(&isc->trim_stats, bp) != 0)
767 * If we've set a trim_goal, then if we exceed that allow trims
768 * to be passed back to the driver. If we've also set a tick timeout
769 * allow trims back to the driver. Otherwise, don't allow trims yet.
771 if (isc->trim_goal > 0) {
772 if (isc->queued_trims >= isc->trim_goal)
774 if (isc->queued_trims > 0 &&
775 isc->trim_ticks > 0 &&
776 ticks - isc->last_trim_tick > isc->trim_ticks)
781 /* NB: Should perhaps have a max trim active independent of I/O limiters */
782 return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) && bp != NULL;
785 #define cam_iosched_sort_queue(isc) ((isc)->sort_io_queue >= 0 ? \
786 (isc)->sort_io_queue : cam_sort_io_queues)
789 cam_iosched_has_work(struct cam_iosched_softc *isc)
791 #ifdef CAM_IOSCHED_DYNAMIC
792 if (iosched_debug > 2)
793 printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
794 cam_iosched_has_more_trim(isc),
795 cam_iosched_has_flagged_work(isc));
798 return cam_iosched_has_io(isc) ||
799 cam_iosched_has_more_trim(isc) ||
800 cam_iosched_has_flagged_work(isc);
803 #ifdef CAM_IOSCHED_DYNAMIC
805 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
810 ios->max = ios->current = 300000;
820 cam_iosched_limiter_init(ios);
824 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
827 struct iop_stats *ios;
828 struct cam_iosched_softc *isc;
834 value = ios->limiter;
835 if (value < none || value >= limiter_max)
838 p = cam_iosched_limiter_names[value];
840 strlcpy(buf, p, sizeof(buf));
841 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
842 if (error != 0 || req->newptr == NULL)
845 cam_periph_lock(isc->periph);
847 for (i = none; i < limiter_max; i++) {
848 if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
851 error = cam_iosched_limiter_init(ios);
853 ios->limiter = value;
854 cam_periph_unlock(isc->periph);
857 /* Note: disk load averate requires ticker to be always running */
858 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
859 isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
861 cam_periph_unlock(isc->periph);
865 cam_periph_unlock(isc->periph);
870 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
873 struct control_loop *clp;
874 struct cam_iosched_softc *isc;
881 if (value < none || value >= cl_max)
884 p = cam_iosched_control_type_names[value];
886 strlcpy(buf, p, sizeof(buf));
887 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
888 if (error != 0 || req->newptr == NULL)
891 for (i = set_max; i < cl_max; i++) {
892 if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
894 cam_periph_lock(isc->periph);
896 cam_periph_unlock(isc->periph);
904 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
911 value = *(sbintime_t *)arg1;
912 us = (uint64_t)value / SBT_1US;
913 snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
914 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
915 if (error != 0 || req->newptr == NULL)
917 us = strtoul(buf, NULL, 10);
920 *(sbintime_t *)arg1 = us * SBT_1US;
925 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
932 sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
934 for (i = 0; i < LAT_BUCKETS - 1; i++)
935 sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
936 sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
937 error = sbuf_finish(&sb);
944 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS)
949 quanta = (unsigned *)arg1;
952 error = sysctl_handle_int(oidp, (int *)&value, 0, req);
953 if ((error != 0) || (req->newptr == NULL))
956 if (value < 1 || value > hz)
965 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
967 struct sysctl_oid_list *n;
968 struct sysctl_ctx_list *ctx;
970 ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
971 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
972 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, name);
973 n = SYSCTL_CHILDREN(ios->sysctl_tree);
974 ctx = &ios->sysctl_ctx;
976 SYSCTL_ADD_UQUAD(ctx, n,
977 OID_AUTO, "ema", CTLFLAG_RD,
979 "Fast Exponentially Weighted Moving Average");
980 SYSCTL_ADD_UQUAD(ctx, n,
981 OID_AUTO, "emvar", CTLFLAG_RD,
983 "Fast Exponentially Weighted Moving Variance");
985 SYSCTL_ADD_INT(ctx, n,
986 OID_AUTO, "pending", CTLFLAG_RD,
988 "Instantaneous # of pending transactions");
989 SYSCTL_ADD_INT(ctx, n,
990 OID_AUTO, "count", CTLFLAG_RD,
992 "# of transactions submitted to hardware");
993 SYSCTL_ADD_INT(ctx, n,
994 OID_AUTO, "queued", CTLFLAG_RD,
996 "# of transactions in the queue");
997 SYSCTL_ADD_INT(ctx, n,
998 OID_AUTO, "in", CTLFLAG_RD,
1000 "# of transactions queued to driver");
1001 SYSCTL_ADD_INT(ctx, n,
1002 OID_AUTO, "out", CTLFLAG_RD,
1004 "# of transactions completed (including with error)");
1005 SYSCTL_ADD_INT(ctx, n,
1006 OID_AUTO, "errs", CTLFLAG_RD,
1008 "# of transactions completed with an error");
1010 SYSCTL_ADD_PROC(ctx, n,
1011 OID_AUTO, "limiter",
1012 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1013 ios, 0, cam_iosched_limiter_sysctl, "A",
1014 "Current limiting type.");
1015 SYSCTL_ADD_INT(ctx, n,
1016 OID_AUTO, "min", CTLFLAG_RW,
1019 SYSCTL_ADD_INT(ctx, n,
1020 OID_AUTO, "max", CTLFLAG_RW,
1023 SYSCTL_ADD_INT(ctx, n,
1024 OID_AUTO, "current", CTLFLAG_RW,
1026 "current resource");
1028 SYSCTL_ADD_PROC(ctx, n,
1029 OID_AUTO, "latencies",
1030 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
1032 cam_iosched_sysctl_latencies, "A",
1033 "Array of power of 2 latency from 1ms to 1.024s");
1037 cam_iosched_iop_stats_fini(struct iop_stats *ios)
1039 if (ios->sysctl_tree)
1040 if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
1041 printf("can't remove iosched sysctl stats context\n");
1045 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
1047 struct sysctl_oid_list *n;
1048 struct sysctl_ctx_list *ctx;
1049 struct control_loop *clp;
1052 clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1053 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
1054 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Control loop info");
1055 n = SYSCTL_CHILDREN(clp->sysctl_tree);
1056 ctx = &clp->sysctl_ctx;
1058 SYSCTL_ADD_PROC(ctx, n,
1060 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1061 clp, 0, cam_iosched_control_type_sysctl, "A",
1062 "Control loop algorithm");
1063 SYSCTL_ADD_PROC(ctx, n,
1064 OID_AUTO, "steer_interval",
1065 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1066 &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
1067 "How often to steer (in us)");
1068 SYSCTL_ADD_PROC(ctx, n,
1070 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1071 &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
1072 "Low water mark for Latency (in us)");
1073 SYSCTL_ADD_PROC(ctx, n,
1075 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1076 &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
1077 "Hi water mark for Latency (in us)");
1078 SYSCTL_ADD_INT(ctx, n,
1079 OID_AUTO, "alpha", CTLFLAG_RW,
1081 "Alpha for PLL (x100) aka gain");
1085 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1087 if (clp->sysctl_tree)
1088 if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1089 printf("can't remove iosched sysctl control loop context\n");
1094 * Allocate the iosched structure. This also insulates callers from knowing
1095 * sizeof struct cam_iosched_softc.
1098 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1101 *iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1104 #ifdef CAM_IOSCHED_DYNAMIC
1106 printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1108 (*iscp)->sort_io_queue = -1;
1109 bioq_init(&(*iscp)->bio_queue);
1110 bioq_init(&(*iscp)->trim_queue);
1111 #ifdef CAM_IOSCHED_DYNAMIC
1112 if (do_dynamic_iosched) {
1113 bioq_init(&(*iscp)->write_queue);
1114 (*iscp)->read_bias = 100;
1115 (*iscp)->current_read_bias = 100;
1116 (*iscp)->quanta = min(hz, 200);
1117 cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1118 cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1119 cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1120 (*iscp)->trim_stats.max = 1; /* Trims are special: one at a time for now */
1121 (*iscp)->last_time = sbinuptime();
1122 callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1123 (*iscp)->periph = periph;
1124 cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1125 callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp);
1126 (*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1134 * Reclaim all used resources. This assumes that other folks have
1135 * drained the requests in the hardware. Maybe an unwise assumption.
1138 cam_iosched_fini(struct cam_iosched_softc *isc)
1141 cam_iosched_flush(isc, NULL, ENXIO);
1142 #ifdef CAM_IOSCHED_DYNAMIC
1143 cam_iosched_iop_stats_fini(&isc->read_stats);
1144 cam_iosched_iop_stats_fini(&isc->write_stats);
1145 cam_iosched_iop_stats_fini(&isc->trim_stats);
1146 cam_iosched_cl_sysctl_fini(&isc->cl);
1147 if (isc->sysctl_tree)
1148 if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1149 printf("can't remove iosched sysctl stats context\n");
1150 if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1151 callout_drain(&isc->ticker);
1152 isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1155 free(isc, M_CAMSCHED);
1160 * After we're sure we're attaching a device, go ahead and add
1161 * hooks for any sysctl we may wish to honor.
1163 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1164 struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1166 struct sysctl_oid_list *n;
1168 n = SYSCTL_CHILDREN(node);
1169 SYSCTL_ADD_INT(ctx, n,
1170 OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1171 &isc->sort_io_queue, 0,
1172 "Sort IO queue to try and optimise disk access patterns");
1173 SYSCTL_ADD_INT(ctx, n,
1174 OID_AUTO, "trim_goal", CTLFLAG_RW,
1176 "Number of trims to try to accumulate before sending to hardware");
1177 SYSCTL_ADD_INT(ctx, n,
1178 OID_AUTO, "trim_ticks", CTLFLAG_RW,
1180 "IO Schedul qaunta to hold back trims for when accumulating");
1182 #ifdef CAM_IOSCHED_DYNAMIC
1183 if (!do_dynamic_iosched)
1186 isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1187 SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1188 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "I/O scheduler statistics");
1189 n = SYSCTL_CHILDREN(isc->sysctl_tree);
1190 ctx = &isc->sysctl_ctx;
1192 cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1193 cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1194 cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1195 cam_iosched_cl_sysctl_init(isc);
1197 SYSCTL_ADD_INT(ctx, n,
1198 OID_AUTO, "read_bias", CTLFLAG_RW,
1199 &isc->read_bias, 100,
1200 "How biased towards read should we be independent of limits");
1202 SYSCTL_ADD_PROC(ctx, n,
1203 OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1204 &isc->quanta, 0, cam_iosched_quanta_sysctl, "I",
1205 "How many quanta per second do we slice the I/O up into");
1207 SYSCTL_ADD_INT(ctx, n,
1208 OID_AUTO, "total_ticks", CTLFLAG_RD,
1209 &isc->total_ticks, 0,
1210 "Total number of ticks we've done");
1212 SYSCTL_ADD_INT(ctx, n,
1213 OID_AUTO, "load", CTLFLAG_RD,
1215 "scaled load average / 100");
1217 SYSCTL_ADD_U64(ctx, n,
1218 OID_AUTO, "latency_trigger", CTLFLAG_RW,
1220 "Latency treshold to trigger callbacks");
1225 cam_iosched_set_latfcn(struct cam_iosched_softc *isc,
1226 cam_iosched_latfcn_t fnp, void *argp)
1228 #ifdef CAM_IOSCHED_DYNAMIC
1235 * Client drivers can set two parameters. "goal" is the number of BIO_DELETEs
1236 * that will be queued up before iosched will "release" the trims to the client
1237 * driver to wo with what they will (usually combine as many as possible). If we
1238 * don't get this many, after trim_ticks we'll submit the I/O anyway with
1239 * whatever we have. We do need an I/O of some kind of to clock the deferred
1240 * trims out to disk. Since we will eventually get a write for the super block
1241 * or something before we shutdown, the trims will complete. To be safe, when a
1242 * BIO_FLUSH is presented to the iosched work queue, we set the ticks time far
1243 * enough in the past so we'll present the BIO_DELETEs to the client driver.
1244 * There might be a race if no BIO_DELETESs were queued, a BIO_FLUSH comes in
1245 * and then a BIO_DELETE is sent down. No know client does this, and there's
1246 * already a race between an ordered BIO_FLUSH and any BIO_DELETEs in flight,
1247 * but no client depends on the ordering being honored.
1249 * XXX I'm not sure what the interaction between UFS direct BIOs and the BUF
1250 * flushing on shutdown. I think there's bufs that would be dependent on the BIO
1251 * finishing to write out at least metadata, so we'll be fine. To be safe, keep
1252 * the number of ticks low (less than maybe 10s) to avoid shutdown races.
1256 cam_iosched_set_trim_goal(struct cam_iosched_softc *isc, int goal)
1259 isc->trim_goal = goal;
1263 cam_iosched_set_trim_ticks(struct cam_iosched_softc *isc, int trim_ticks)
1266 isc->trim_ticks = trim_ticks;
1270 * Flush outstanding I/O. Consumers of this library don't know all the
1271 * queues we may keep, so this allows all I/O to be flushed in one
1275 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1277 bioq_flush(&isc->bio_queue, stp, err);
1278 bioq_flush(&isc->trim_queue, stp, err);
1279 #ifdef CAM_IOSCHED_DYNAMIC
1280 if (do_dynamic_iosched)
1281 bioq_flush(&isc->write_queue, stp, err);
1285 #ifdef CAM_IOSCHED_DYNAMIC
1287 cam_iosched_get_write(struct cam_iosched_softc *isc)
1292 * We control the write rate by controlling how many requests we send
1293 * down to the drive at any one time. Fewer requests limits the
1294 * effects of both starvation when the requests take a while and write
1295 * amplification when each request is causing more than one write to
1296 * the NAND media. Limiting the queue depth like this will also limit
1297 * the write throughput and give and reads that want to compete to
1300 bp = bioq_first(&isc->write_queue);
1302 if (iosched_debug > 3)
1303 printf("No writes present in write_queue\n");
1308 * If pending read, prefer that based on current read bias
1311 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1314 "Reads present and current_read_bias is %d queued "
1315 "writes %d queued reads %d\n",
1316 isc->current_read_bias, isc->write_stats.queued,
1317 isc->read_stats.queued);
1318 isc->current_read_bias--;
1319 /* We're not limiting writes, per se, just doing reads first */
1324 * See if our current limiter allows this I/O.
1326 if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1328 printf("Can't write because limiter says no.\n");
1329 isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1334 * Let's do this: We've passed all the gates and we're a go
1335 * to schedule the I/O in the SIM.
1337 isc->current_read_bias = isc->read_bias;
1338 bioq_remove(&isc->write_queue, bp);
1339 if (bp->bio_cmd == BIO_WRITE) {
1340 isc->write_stats.queued--;
1341 isc->write_stats.total++;
1342 isc->write_stats.pending++;
1344 if (iosched_debug > 9)
1345 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1346 isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1352 * Put back a trim that you weren't able to actually schedule this time.
1355 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1357 bioq_insert_head(&isc->trim_queue, bp);
1358 if (isc->queued_trims == 0)
1359 isc->last_trim_tick = ticks;
1360 isc->queued_trims++;
1361 #ifdef CAM_IOSCHED_DYNAMIC
1362 isc->trim_stats.queued++;
1363 isc->trim_stats.total--; /* since we put it back, don't double count */
1364 isc->trim_stats.pending--;
1369 * gets the next trim from the trim queue.
1371 * Assumes we're called with the periph lock held. It removes this
1372 * trim from the queue and the device must explicitly reinsert it
1373 * should the need arise.
1376 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1380 bp = bioq_first(&isc->trim_queue);
1383 bioq_remove(&isc->trim_queue, bp);
1384 isc->queued_trims--;
1385 isc->last_trim_tick = ticks; /* Reset the tick timer when we take trims */
1386 #ifdef CAM_IOSCHED_DYNAMIC
1387 isc->trim_stats.queued--;
1388 isc->trim_stats.total++;
1389 isc->trim_stats.pending++;
1395 * gets an available trim from the trim queue, if there's no trim
1396 * already pending. It removes this trim from the queue and the device
1397 * must explicitly reinsert it should the need arise.
1399 * Assumes we're called with the periph lock held.
1402 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1404 #ifdef CAM_IOSCHED_DYNAMIC
1408 if (!cam_iosched_has_more_trim(isc))
1410 #ifdef CAM_IOSCHED_DYNAMIC
1411 bp = bioq_first(&isc->trim_queue);
1416 * If pending read, prefer that based on current read bias setting. The
1417 * read bias is shared for both writes and TRIMs, but on TRIMs the bias
1418 * is for a combined TRIM not a single TRIM request that's come in.
1420 if (do_dynamic_iosched) {
1421 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1423 printf("Reads present and current_read_bias is %d"
1424 " queued trims %d queued reads %d\n",
1425 isc->current_read_bias, isc->trim_stats.queued,
1426 isc->read_stats.queued);
1427 isc->current_read_bias--;
1428 /* We're not limiting TRIMS, per se, just doing reads first */
1432 * We're going to do a trim, so reset the bias.
1434 isc->current_read_bias = isc->read_bias;
1438 * See if our current limiter allows this I/O. Because we only call this
1439 * here, and not in next_trim, the 'bandwidth' limits for trims won't
1440 * work, while the iops or max queued limits will work. It's tricky
1441 * because we want the limits to be from the perspective of the
1442 * "commands sent to the device." To make iops work, we need to check
1443 * only here (since we want all the ops we combine to count as one). To
1444 * make bw limits work, we'd need to check in next_trim, but that would
1445 * have the effect of limiting the iops as seen from the upper layers.
1447 if (cam_iosched_limiter_iop(&isc->trim_stats, bp) != 0) {
1449 printf("Can't trim because limiter says no.\n");
1450 isc->trim_stats.state_flags |= IOP_RATE_LIMITED;
1453 isc->current_read_bias = isc->read_bias;
1454 isc->trim_stats.state_flags &= ~IOP_RATE_LIMITED;
1455 /* cam_iosched_next_trim below keeps proper book */
1457 return cam_iosched_next_trim(isc);
1461 * Determine what the next bit of work to do is for the periph. The
1462 * default implementation looks to see if we have trims to do, but no
1463 * trims outstanding. If so, we do that. Otherwise we see if we have
1464 * other work. If we do, then we do that. Otherwise why were we called?
1467 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1472 * See if we have a trim that can be scheduled. We can only send one
1473 * at a time down, so this takes that into account.
1475 * XXX newer TRIM commands are queueable. Revisit this when we
1478 if ((bp = cam_iosched_get_trim(isc)) != NULL)
1481 #ifdef CAM_IOSCHED_DYNAMIC
1483 * See if we have any pending writes, and room in the queue for them,
1484 * and if so, those are next.
1486 if (do_dynamic_iosched) {
1487 if ((bp = cam_iosched_get_write(isc)) != NULL)
1493 * next, see if there's other, normal I/O waiting. If so return that.
1495 if ((bp = bioq_first(&isc->bio_queue)) == NULL)
1498 #ifdef CAM_IOSCHED_DYNAMIC
1500 * For the dynamic scheduler, bio_queue is only for reads, so enforce
1501 * the limits here. Enforce only for reads.
1503 if (do_dynamic_iosched) {
1504 if (bp->bio_cmd == BIO_READ &&
1505 cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1506 isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1510 isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1512 bioq_remove(&isc->bio_queue, bp);
1513 #ifdef CAM_IOSCHED_DYNAMIC
1514 if (do_dynamic_iosched) {
1515 if (bp->bio_cmd == BIO_READ) {
1516 isc->read_stats.queued--;
1517 isc->read_stats.total++;
1518 isc->read_stats.pending++;
1520 printf("Found bio_cmd = %#x\n", bp->bio_cmd);
1522 if (iosched_debug > 9)
1523 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1529 * Driver has been given some work to do by the block layer. Tell the
1530 * scheduler about it and have it queue the work up. The scheduler module
1531 * will then return the currently most useful bit of work later, possibly
1532 * deferring work for various reasons.
1535 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1539 * A BIO_SPEEDUP from the uppper layers means that they have a block
1540 * shortage. At the present, this is only sent when we're trying to
1541 * allocate blocks, but have a shortage before giving up. bio_length is
1542 * the size of their shortage. We will complete just enough BIO_DELETEs
1543 * in the queue to satisfy the need. If bio_length is 0, we'll complete
1544 * them all. This allows the scheduler to delay BIO_DELETEs to improve
1545 * read/write performance without worrying about the upper layers. When
1546 * it's possibly a problem, we respond by pretending the BIO_DELETEs
1547 * just worked. We can't do anything about the BIO_DELETEs in the
1548 * hardware, though. We have to wait for them to complete.
1550 if (bp->bio_cmd == BIO_SPEEDUP) {
1555 while (bioq_first(&isc->trim_queue) &&
1556 (bp->bio_length == 0 || len < bp->bio_length)) {
1557 nbp = bioq_takefirst(&isc->trim_queue);
1558 len += nbp->bio_length;
1562 if (bp->bio_length > 0) {
1563 if (bp->bio_length > len)
1564 bp->bio_resid = bp->bio_length - len;
1574 * If we get a BIO_FLUSH, and we're doing delayed BIO_DELETEs then we
1575 * set the last tick time to one less than the current ticks minus the
1576 * delay to force the BIO_DELETEs to be presented to the client driver.
1578 if (bp->bio_cmd == BIO_FLUSH && isc->trim_ticks > 0)
1579 isc->last_trim_tick = ticks - isc->trim_ticks - 1;
1582 * Put all trims on the trim queue. Otherwise put the work on the bio
1585 if (bp->bio_cmd == BIO_DELETE) {
1586 bioq_insert_tail(&isc->trim_queue, bp);
1587 if (isc->queued_trims == 0)
1588 isc->last_trim_tick = ticks;
1589 isc->queued_trims++;
1590 #ifdef CAM_IOSCHED_DYNAMIC
1591 isc->trim_stats.in++;
1592 isc->trim_stats.queued++;
1595 #ifdef CAM_IOSCHED_DYNAMIC
1596 else if (do_dynamic_iosched && (bp->bio_cmd != BIO_READ)) {
1597 if (cam_iosched_sort_queue(isc))
1598 bioq_disksort(&isc->write_queue, bp);
1600 bioq_insert_tail(&isc->write_queue, bp);
1601 if (iosched_debug > 9)
1602 printf("Qw : %p %#x\n", bp, bp->bio_cmd);
1603 if (bp->bio_cmd == BIO_WRITE) {
1604 isc->write_stats.in++;
1605 isc->write_stats.queued++;
1610 if (cam_iosched_sort_queue(isc))
1611 bioq_disksort(&isc->bio_queue, bp);
1613 bioq_insert_tail(&isc->bio_queue, bp);
1614 #ifdef CAM_IOSCHED_DYNAMIC
1615 if (iosched_debug > 9)
1616 printf("Qr : %p %#x\n", bp, bp->bio_cmd);
1617 if (bp->bio_cmd == BIO_READ) {
1618 isc->read_stats.in++;
1619 isc->read_stats.queued++;
1620 } else if (bp->bio_cmd == BIO_WRITE) {
1621 isc->write_stats.in++;
1622 isc->write_stats.queued++;
1629 * If we have work, get it scheduled. Called with the periph lock held.
1632 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1635 if (cam_iosched_has_work(isc))
1636 xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1640 * Complete a trim request. Mark that we no longer have one in flight.
1643 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1646 isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1650 * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1651 * might use notes in the ccb for statistics.
1654 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1655 union ccb *done_ccb)
1658 #ifdef CAM_IOSCHED_DYNAMIC
1659 if (!do_dynamic_iosched)
1662 if (iosched_debug > 10)
1663 printf("done: %p %#x\n", bp, bp->bio_cmd);
1664 if (bp->bio_cmd == BIO_WRITE) {
1665 retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1666 if ((bp->bio_flags & BIO_ERROR) != 0)
1667 isc->write_stats.errs++;
1668 isc->write_stats.out++;
1669 isc->write_stats.pending--;
1670 } else if (bp->bio_cmd == BIO_READ) {
1671 retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1672 if ((bp->bio_flags & BIO_ERROR) != 0)
1673 isc->read_stats.errs++;
1674 isc->read_stats.out++;
1675 isc->read_stats.pending--;
1676 } else if (bp->bio_cmd == BIO_DELETE) {
1677 if ((bp->bio_flags & BIO_ERROR) != 0)
1678 isc->trim_stats.errs++;
1679 isc->trim_stats.out++;
1680 isc->trim_stats.pending--;
1681 } else if (bp->bio_cmd != BIO_FLUSH) {
1683 printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1686 if (!(bp->bio_flags & BIO_ERROR) && done_ccb != NULL) {
1687 sbintime_t sim_latency;
1689 sim_latency = cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data);
1691 cam_iosched_io_metric_update(isc, sim_latency,
1692 bp->bio_cmd, bp->bio_bcount);
1694 * Debugging code: allow callbacks to the periph driver when latency max
1695 * is exceeded. This can be useful for triggering external debugging actions.
1697 if (isc->latfcn && isc->max_lat != 0 && sim_latency > isc->max_lat)
1698 isc->latfcn(isc->latarg, sim_latency, bp);
1706 * Tell the io scheduler that you've pushed a trim down into the sim.
1707 * This also tells the I/O scheduler not to push any more trims down, so
1708 * some periphs do not call it if they can cope with multiple trims in flight.
1711 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1714 isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1718 * Change the sorting policy hint for I/O transactions for this device.
1721 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1724 isc->sort_io_queue = val;
1728 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1730 return isc->flags & flags;
1734 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1736 isc->flags |= flags;
1740 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1742 isc->flags &= ~flags;
1745 #ifdef CAM_IOSCHED_DYNAMIC
1747 * After the method presented in Jack Crenshaw's 1998 article "Integer
1748 * Square Roots," reprinted at
1749 * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1750 * and well worth the read. Briefly, we find the power of 4 that's the
1751 * largest smaller than val. We then check each smaller power of 4 to
1752 * see if val is still bigger. The right shifts at each step divide
1753 * the result by 2 which after successive application winds up
1754 * accumulating the right answer. It could also have been accumulated
1755 * using a separate root counter, but this code is smaller and faster
1756 * than that method. This method is also integer size invariant.
1757 * It returns floor(sqrt((float)val)), or the largest integer less than
1758 * or equal to the square root.
1761 isqrt64(uint64_t val)
1764 uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1767 * Find the largest power of 4 smaller than val.
1773 * Accumulate the answer, one bit at a time (we keep moving
1774 * them over since 2 is the square root of 4 and we test
1775 * powers of 4). We accumulate where we find the bit, but
1776 * the successive shifts land the bit in the right place
1780 if (val >= res + bit) {
1782 res = (res >> 1) + bit;
1791 static sbintime_t latencies[LAT_BUCKETS - 1] = {
1805 SBT_1MS << 13 /* 8.192s */
1809 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1811 sbintime_t y, deltasq, delta;
1815 * Keep counts for latency. We do it by power of two buckets.
1816 * This helps us spot outlier behavior obscured by averages.
1818 for (i = 0; i < LAT_BUCKETS - 1; i++) {
1819 if (sim_latency < latencies[i]) {
1820 iop->latencies[i]++;
1824 if (i == LAT_BUCKETS - 1)
1825 iop->latencies[i]++; /* Put all > 1024ms values into the last bucket. */
1828 * Classic exponentially decaying average with a tiny alpha
1829 * (2 ^ -alpha_bits). For more info see the NIST statistical
1832 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha) [nist]
1833 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
1834 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1835 * alpha = 1 / (1 << alpha_bits)
1836 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
1837 * = y_t/b - e/b + be/b
1838 * = (y_t - e + be) / b
1841 * Since alpha is a power of two, we can compute this w/o any mult or
1844 * Variance can also be computed. Usually, it would be expressed as follows:
1845 * diff_t = y_t - ema_t-1
1846 * emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
1847 * = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
1848 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
1849 * = e - e/b + dd/b + dd/bb
1850 * = (bbe - be + bdd + dd) / bb
1851 * = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
1854 * XXX possible numeric issues
1855 * o We assume right shifted integers do the right thing, since that's
1856 * implementation defined. You can change the right shifts to / (1LL << alpha).
1857 * o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
1858 * for emvar. This puts a ceiling of 13 bits on alpha since we need a
1859 * few tens of seconds of representation.
1860 * o We mitigate alpha issues by never setting it too high.
1863 delta = (y - iop->ema); /* d */
1864 iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1867 * Were we to naively plow ahead at this point, we wind up with many numerical
1868 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
1869 * us with microsecond level precision in the input, so the same in the
1870 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
1871 * also means that emvar can be up 46 bits 40 of which are fraction, which
1872 * gives us a way to measure up to ~8s in the SD before the computation goes
1873 * unstable. Even the worst hard disk rarely has > 1s service time in the
1874 * drive. It does mean we have to shift left 12 bits after taking the
1875 * square root to compute the actual standard deviation estimate. This loss of
1876 * precision is preferable to needing int128 types to work. The above numbers
1877 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
1878 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1881 deltasq = delta * delta; /* dd */
1882 iop->emvar = ((iop->emvar << (2 * alpha_bits)) + /* bbe */
1883 ((deltasq - iop->emvar) << alpha_bits) + /* b(dd-e) */
1885 >> (2 * alpha_bits); /* div bb */
1886 iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1890 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1891 sbintime_t sim_latency, int cmd, size_t size)
1893 /* xxx Do we need to scale based on the size of the I/O ? */
1896 cam_iosched_update(&isc->read_stats, sim_latency);
1899 cam_iosched_update(&isc->write_stats, sim_latency);
1902 cam_iosched_update(&isc->trim_stats, sim_latency);
1910 static int biolen(struct bio_queue_head *bq)
1915 TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
1922 * Show the internal state of the I/O scheduler.
1924 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
1926 struct cam_iosched_softc *isc;
1929 db_printf("Need addr\n");
1932 isc = (struct cam_iosched_softc *)addr;
1933 db_printf("pending_reads: %d\n", isc->read_stats.pending);
1934 db_printf("min_reads: %d\n", isc->read_stats.min);
1935 db_printf("max_reads: %d\n", isc->read_stats.max);
1936 db_printf("reads: %d\n", isc->read_stats.total);
1937 db_printf("in_reads: %d\n", isc->read_stats.in);
1938 db_printf("out_reads: %d\n", isc->read_stats.out);
1939 db_printf("queued_reads: %d\n", isc->read_stats.queued);
1940 db_printf("Read Q len %d\n", biolen(&isc->bio_queue));
1941 db_printf("pending_writes: %d\n", isc->write_stats.pending);
1942 db_printf("min_writes: %d\n", isc->write_stats.min);
1943 db_printf("max_writes: %d\n", isc->write_stats.max);
1944 db_printf("writes: %d\n", isc->write_stats.total);
1945 db_printf("in_writes: %d\n", isc->write_stats.in);
1946 db_printf("out_writes: %d\n", isc->write_stats.out);
1947 db_printf("queued_writes: %d\n", isc->write_stats.queued);
1948 db_printf("Write Q len %d\n", biolen(&isc->write_queue));
1949 db_printf("pending_trims: %d\n", isc->trim_stats.pending);
1950 db_printf("min_trims: %d\n", isc->trim_stats.min);
1951 db_printf("max_trims: %d\n", isc->trim_stats.max);
1952 db_printf("trims: %d\n", isc->trim_stats.total);
1953 db_printf("in_trims: %d\n", isc->trim_stats.in);
1954 db_printf("out_trims: %d\n", isc->trim_stats.out);
1955 db_printf("queued_trims: %d\n", isc->trim_stats.queued);
1956 db_printf("Trim Q len %d\n", biolen(&isc->trim_queue));
1957 db_printf("read_bias: %d\n", isc->read_bias);
1958 db_printf("current_read_bias: %d\n", isc->current_read_bias);
1959 db_printf("Trim active? %s\n",
1960 (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");