2 * CAM IO Scheduler Interface
4 * SPDX-License-Identifier: BSD-2-Clause
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 * 2. Redistributions in binary form must reproduce the above copyright
14 * notice, this list of conditions and the following disclaimer in the
15 * documentation and/or other materials provided with the distribution.
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
21 * FOR 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
32 #include <sys/param.h>
33 #include <sys/systm.h>
34 #include <sys/kernel.h>
37 #include <sys/malloc.h>
38 #include <sys/mutex.h>
40 #include <sys/sysctl.h>
43 #include <cam/cam_ccb.h>
44 #include <cam/cam_periph.h>
45 #include <cam/cam_xpt_periph.h>
46 #include <cam/cam_xpt_internal.h>
47 #include <cam/cam_iosched.h>
51 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler",
52 "CAM I/O Scheduler buffers");
54 static SYSCTL_NODE(_kern_cam, OID_AUTO, iosched, CTLFLAG_RD | CTLFLAG_MPSAFE, 0,
55 "CAM I/O Scheduler parameters");
58 * Default I/O scheduler for FreeBSD. This implementation is just a thin-vineer
59 * over the bioq_* interface, with notions of separate calls for normal I/O and
62 * When CAM_IOSCHED_DYNAMIC is defined, the scheduler is enhanced to dynamically
63 * steer the rate of one type of traffic to help other types of traffic (eg
64 * limit writes when read latency deteriorates on SSDs).
67 #ifdef CAM_IOSCHED_DYNAMIC
69 static bool do_dynamic_iosched = true;
70 SYSCTL_BOOL(_kern_cam_iosched, OID_AUTO, dynamic, CTLFLAG_RDTUN,
71 &do_dynamic_iosched, 1,
72 "Enable Dynamic I/O scheduler optimizations.");
75 * For an EMA, with an alpha of alpha, we know
79 * where N is the number of samples that 86% of the current
80 * EMA is derived from.
82 * So we invent[*] alpha_bits:
83 * alpha_bits = -log_2(alpha)
84 * alpha = 2^-alpha_bits
86 * N = 1 + 2^(alpha_bits + 1)
88 * The default 9 gives a 1025 lookback for 86% of the data.
89 * For a brief intro: https://en.wikipedia.org/wiki/Moving_average
91 * [*] Steal from the load average code and many other places.
92 * Note: See computation of EMA and EMVAR for acceptable ranges of alpha.
94 static int alpha_bits = 9;
95 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, alpha_bits, CTLFLAG_RWTUN,
97 "Bits in EMA's alpha.");
100 * Different parameters for the buckets of latency we keep track of. These are all
101 * published read-only since at present they are compile time constants.
103 * Bucket base is the upper bounds of the first latency bucket. It's currently 20us.
104 * With 20 buckets (see below), that leads to a geometric progression with a max size
105 * of 5.2s which is safeily larger than 1s to help diagnose extreme outliers better.
108 #define BUCKET_BASE ((SBT_1S / 50000) + 1) /* 20us */
110 static sbintime_t bucket_base = BUCKET_BASE;
111 SYSCTL_SBINTIME_USEC(_kern_cam_iosched, OID_AUTO, bucket_base_us, CTLFLAG_RD,
113 "Size of the smallest latency bucket");
116 * Bucket ratio is the geometric progression for the bucket. For a bucket b_n
117 * the size of bucket b_n+1 is b_n * bucket_ratio / 100.
119 static int bucket_ratio = 200; /* Rather hard coded at the moment */
120 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, bucket_ratio, CTLFLAG_RD,
122 "Latency Bucket Ratio for geometric progression.");
125 * Number of total buckets. Starting at BUCKET_BASE, each one is a power of 2.
128 #define LAT_BUCKETS 20 /* < 20us < 40us ... < 2^(n-1)*20us >= 2^(n-1)*20us */
130 static int lat_buckets = LAT_BUCKETS;
131 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, buckets, CTLFLAG_RD,
132 &lat_buckets, LAT_BUCKETS,
133 "Total number of latency buckets published");
136 * Read bias: how many reads do we favor before scheduling a write
137 * when we have a choice.
139 static int default_read_bias = 0;
140 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, read_bias, CTLFLAG_RWTUN,
141 &default_read_bias, 0,
142 "Default read bias for new devices.");
145 struct cam_iosched_softc;
147 int iosched_debug = 0;
150 none = 0, /* No limits */
151 queue_depth, /* Limit how many ops we queue to SIM */
152 iops, /* Limit # of IOPS to the drive */
153 bandwidth, /* Limit bandwidth to the drive */
157 static const char *cam_iosched_limiter_names[] =
158 { "none", "queue_depth", "iops", "bandwidth" };
161 * Called to initialize the bits of the iop_stats structure relevant to the
162 * limiter. Called just after the limiter is set.
164 typedef int l_init_t(struct iop_stats *);
169 typedef int l_tick_t(struct iop_stats *);
172 * Called to see if the limiter thinks this IOP can be allowed to
173 * proceed. If so, the limiter assumes that the IOP proceeded
174 * and makes any accounting of it that's needed.
176 typedef int l_iop_t(struct iop_stats *, struct bio *);
179 * Called when an I/O completes so the limiter can update its
180 * accounting. Pending I/Os may complete in any order (even when
181 * sent to the hardware at the same time), so the limiter may not
182 * make any assumptions other than this I/O has completed. If it
183 * returns 1, then xpt_schedule() needs to be called again.
185 typedef int l_iodone_t(struct iop_stats *, struct bio *);
187 static l_iop_t cam_iosched_qd_iop;
188 static l_iop_t cam_iosched_qd_caniop;
189 static l_iodone_t cam_iosched_qd_iodone;
191 static l_init_t cam_iosched_iops_init;
192 static l_tick_t cam_iosched_iops_tick;
193 static l_iop_t cam_iosched_iops_caniop;
194 static l_iop_t cam_iosched_iops_iop;
196 static l_init_t cam_iosched_bw_init;
197 static l_tick_t cam_iosched_bw_tick;
198 static l_iop_t cam_iosched_bw_caniop;
199 static l_iop_t cam_iosched_bw_iop;
206 l_iodone_t *l_iodone;
218 .l_caniop = cam_iosched_qd_caniop,
219 .l_iop = cam_iosched_qd_iop,
220 .l_iodone= cam_iosched_qd_iodone,
223 .l_init = cam_iosched_iops_init,
224 .l_tick = cam_iosched_iops_tick,
225 .l_caniop = cam_iosched_iops_caniop,
226 .l_iop = cam_iosched_iops_iop,
230 .l_init = cam_iosched_bw_init,
231 .l_tick = cam_iosched_bw_tick,
232 .l_caniop = cam_iosched_bw_caniop,
233 .l_iop = cam_iosched_bw_iop,
240 * sysctl state for this subnode.
242 struct sysctl_ctx_list sysctl_ctx;
243 struct sysctl_oid *sysctl_tree;
246 * Information about the current rate limiters, if any
248 io_limiter limiter; /* How are I/Os being limited */
249 int min; /* Low range of limit */
250 int max; /* High range of limit */
251 int current; /* Current rate limiter */
252 int l_value1; /* per-limiter scratch value 1. */
253 int l_value2; /* per-limiter scratch value 2. */
256 * Debug information about counts of I/Os that have gone through the
259 int pending; /* I/Os pending in the hardware */
260 int queued; /* number currently in the queue */
261 int total; /* Total for all time -- wraps */
262 int in; /* number queued all time -- wraps */
263 int out; /* number completed all time -- wraps */
264 int errs; /* Number of I/Os completed with error -- wraps */
267 * Statistics on different bits of the process.
269 /* Exp Moving Average, see alpha_bits for more details */
272 sbintime_t sd; /* Last computed sd */
274 uint32_t state_flags;
275 #define IOP_RATE_LIMITED 1u
277 uint64_t latencies[LAT_BUCKETS];
279 struct cam_iosched_softc *softc;
283 set_max = 0, /* current = max */
284 read_latency, /* Steer read latency by throttling writes */
285 cl_max /* Keep last */
288 static const char *cam_iosched_control_type_names[] =
289 { "set_max", "read_latency" };
291 struct control_loop {
293 * sysctl state for this subnode.
295 struct sysctl_ctx_list sysctl_ctx;
296 struct sysctl_oid *sysctl_tree;
298 sbintime_t next_steer; /* Time of next steer */
299 sbintime_t steer_interval; /* How often do we steer? */
303 control_type type; /* What type of control? */
304 int last_count; /* Last I/O count */
306 struct cam_iosched_softc *softc;
311 struct cam_iosched_softc {
312 struct bio_queue_head bio_queue;
313 struct bio_queue_head trim_queue;
314 /* scheduler flags < 16, user flags >= 16 */
317 int trim_goal; /* # of trims to queue before sending */
318 int trim_ticks; /* Max ticks to hold trims */
319 int last_trim_tick; /* Last 'tick' time ld a trim */
320 int queued_trims; /* Number of trims in the queue */
321 #ifdef CAM_IOSCHED_DYNAMIC
322 int read_bias; /* Read bias setting */
323 int current_read_bias; /* Current read bias state */
325 int load; /* EMA of 'load average' of disk / 2^16 */
327 struct bio_queue_head write_queue;
328 struct iop_stats read_stats, write_stats, trim_stats;
329 struct sysctl_ctx_list sysctl_ctx;
330 struct sysctl_oid *sysctl_tree;
332 int quanta; /* Number of quanta per second */
333 struct callout ticker; /* Callout for our quota system */
334 struct cam_periph *periph; /* cam periph associated with this device */
335 uint32_t this_frac; /* Fraction of a second (1024ths) for this tick */
336 sbintime_t last_time; /* Last time we ticked */
337 struct control_loop cl;
338 sbintime_t max_lat; /* when != 0, if iop latency > max_lat, call max_lat_fcn */
339 cam_iosched_latfcn_t latfcn;
344 #ifdef CAM_IOSCHED_DYNAMIC
346 * helper functions to call the limsw functions.
349 cam_iosched_limiter_init(struct iop_stats *ios)
351 int lim = ios->limiter;
353 /* maybe this should be a kassert */
354 if (lim < none || lim >= limiter_max)
357 if (limsw[lim].l_init)
358 return limsw[lim].l_init(ios);
364 cam_iosched_limiter_tick(struct iop_stats *ios)
366 int lim = ios->limiter;
368 /* maybe this should be a kassert */
369 if (lim < none || lim >= limiter_max)
372 if (limsw[lim].l_tick)
373 return limsw[lim].l_tick(ios);
379 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
381 int lim = ios->limiter;
383 /* maybe this should be a kassert */
384 if (lim < none || lim >= limiter_max)
387 if (limsw[lim].l_iop)
388 return limsw[lim].l_iop(ios, bp);
394 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
396 int lim = ios->limiter;
398 /* maybe this should be a kassert */
399 if (lim < none || lim >= limiter_max)
402 if (limsw[lim].l_caniop)
403 return limsw[lim].l_caniop(ios, bp);
409 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
411 int lim = ios->limiter;
413 /* maybe this should be a kassert */
414 if (lim < none || lim >= limiter_max)
417 if (limsw[lim].l_iodone)
418 return limsw[lim].l_iodone(ios, bp);
424 * Functions to implement the different kinds of limiters
428 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
431 if (ios->current <= 0 || ios->pending < ios->current)
438 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
441 if (ios->current <= 0 || ios->pending < ios->current)
448 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
451 if (ios->current <= 0 || ios->pending != ios->current)
458 cam_iosched_iops_init(struct iop_stats *ios)
461 ios->l_value1 = ios->current / ios->softc->quanta;
462 if (ios->l_value1 <= 0)
470 cam_iosched_iops_tick(struct iop_stats *ios)
475 * Allow at least one IO per tick until all
476 * the IOs for this interval have been spent.
478 new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
479 if (new_ios < 1 && ios->l_value2 < ios->current) {
485 * If this a new accounting interval, discard any "unspent" ios
486 * granted in the previous interval. Otherwise add the new ios to
487 * the previously granted ones that haven't been spent yet.
489 if ((ios->softc->total_ticks % ios->softc->quanta) == 0) {
490 ios->l_value1 = new_ios;
493 ios->l_value1 += new_ios;
500 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
504 * So if we have any more IOPs left, allow it,
505 * otherwise wait. If current iops is 0, treat that
506 * as unlimited as a failsafe.
508 if (ios->current > 0 && ios->l_value1 <= 0)
514 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
518 rv = cam_iosched_limiter_caniop(ios, bp);
526 cam_iosched_bw_init(struct iop_stats *ios)
529 /* ios->current is in kB/s, so scale to bytes */
530 ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
536 cam_iosched_bw_tick(struct iop_stats *ios)
541 * If we're in the hole for available quota from
542 * the last time, then add the quantum for this.
543 * If we have any left over from last quantum,
544 * then too bad, that's lost. Also, ios->current
545 * is in kB/s, so scale.
547 * We also allow up to 4 quanta of credits to
548 * accumulate to deal with burstiness. 4 is extremely
551 bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
552 if (ios->l_value1 < bw * 4)
559 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
562 * So if we have any more bw quota left, allow it,
563 * otherwise wait. Note, we'll go negative and that's
564 * OK. We'll just get a little less next quota.
566 * Note on going negative: that allows us to process
567 * requests in order better, since we won't allow
568 * shorter reads to get around the long one that we
569 * don't have the quota to do just yet. It also prevents
570 * starvation by being a little more permissive about
571 * what we let through this quantum (to prevent the
572 * starvation), at the cost of getting a little less
575 * Also note that if the current limit is <= 0,
576 * we treat it as unlimited as a failsafe.
578 if (ios->current > 0 && ios->l_value1 <= 0)
585 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
589 rv = cam_iosched_limiter_caniop(ios, bp);
591 ios->l_value1 -= bp->bio_length;
596 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
599 cam_iosched_ticker(void *arg)
601 struct cam_iosched_softc *isc = arg;
602 sbintime_t now, delta;
605 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
608 delta = now - isc->last_time;
609 isc->this_frac = (uint32_t)delta >> 16; /* Note: discards seconds -- should be 0 harmless if not */
610 isc->last_time = now;
612 cam_iosched_cl_maybe_steer(&isc->cl);
614 cam_iosched_limiter_tick(&isc->read_stats);
615 cam_iosched_limiter_tick(&isc->write_stats);
616 cam_iosched_limiter_tick(&isc->trim_stats);
618 cam_iosched_schedule(isc, isc->periph);
621 * isc->load is an EMA of the pending I/Os at each tick. The number of
622 * pending I/Os is the sum of the I/Os queued to the hardware, and those
623 * in the software queue that could be queued to the hardware if there
626 * ios_stats.pending is a count of requests in the SIM right now for
627 * each of these types of I/O. So the total pending count is the sum of
628 * these I/Os and the sum of the queued I/Os still in the software queue
629 * for those operations that aren't being rate limited at the moment.
631 * The reason for the rate limiting bit is because those I/Os
632 * aren't part of the software queued load (since we could
633 * give them to hardware, but choose not to).
635 * Note: due to a bug in counting pending TRIM in the device, we
636 * don't include them in this count. We count each BIO_DELETE in
637 * the pending count, but the periph drivers collapse them down
638 * into one TRIM command. That one trim command gets the completion
639 * so the counts get off.
641 pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
642 pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
643 !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
644 !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
646 pending /= isc->periph->path->device->ccbq.total_openings;
648 isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
654 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
657 clp->next_steer = sbinuptime();
659 clp->steer_interval = SBT_1S * 5; /* Let's start out steering every 5s */
660 clp->lolat = 5 * SBT_1MS;
661 clp->hilat = 15 * SBT_1MS;
662 clp->alpha = 20; /* Alpha == gain. 20 = .2 */
667 cam_iosched_cl_maybe_steer(struct control_loop *clp)
669 struct cam_iosched_softc *isc;
674 now = isc->last_time;
675 if (now < clp->next_steer)
678 clp->next_steer = now + clp->steer_interval;
681 if (isc->write_stats.current != isc->write_stats.max)
682 printf("Steering write from %d kBps to %d kBps\n",
683 isc->write_stats.current, isc->write_stats.max);
684 isc->read_stats.current = isc->read_stats.max;
685 isc->write_stats.current = isc->write_stats.max;
686 isc->trim_stats.current = isc->trim_stats.max;
689 old = isc->write_stats.current;
690 lat = isc->read_stats.ema;
692 * Simple PLL-like engine. Since we're steering to a range for
693 * the SP (set point) that makes things a little more
694 * complicated. In addition, we're not directly controlling our
695 * PV (process variable), the read latency, but instead are
696 * manipulating the write bandwidth limit for our MV
697 * (manipulation variable), analysis of this code gets a bit
698 * messy. Also, the MV is a very noisy control surface for read
699 * latency since it is affected by many hidden processes inside
700 * the device which change how responsive read latency will be
701 * in reaction to changes in write bandwidth. Unlike the classic
702 * boiler control PLL. this may result in over-steering while
703 * the SSD takes its time to react to the new, lower load. This
704 * is why we use a relatively low alpha of between .1 and .25 to
705 * compensate for this effect. At .1, it takes ~22 steering
706 * intervals to back off by a factor of 10. At .2 it only takes
707 * ~10. At .25 it only takes ~8. However some preliminary data
708 * from the SSD drives suggests a reasponse time in 10's of
709 * seconds before latency drops regardless of the new write
710 * rate. Careful observation will be required to tune this
713 * Also, when there's no read traffic, we jack up the write
714 * limit too regardless of the last read latency. 10 is
715 * somewhat arbitrary.
717 if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
718 isc->write_stats.current = isc->write_stats.current *
719 (100 + clp->alpha) / 100; /* Scale up */
720 else if (lat > clp->hilat)
721 isc->write_stats.current = isc->write_stats.current *
722 (100 - clp->alpha) / 100; /* Scale down */
723 clp->last_count = isc->read_stats.total;
726 * Even if we don't steer, per se, enforce the min/max limits as
727 * those may have changed.
729 if (isc->write_stats.current < isc->write_stats.min)
730 isc->write_stats.current = isc->write_stats.min;
731 if (isc->write_stats.current > isc->write_stats.max)
732 isc->write_stats.current = isc->write_stats.max;
733 if (old != isc->write_stats.current && iosched_debug)
734 printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
735 old, isc->write_stats.current,
736 (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
745 * Trim or similar currently pending completion. Should only be set for
746 * those drivers wishing only one Trim active at a time.
748 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE (1ul << 0)
749 /* Callout active, and needs to be torn down */
750 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
752 /* Periph drivers set these flags to indicate work */
753 #define CAM_IOSCHED_FLAG_WORK_FLAGS ((0xffffu) << 16)
755 #ifdef CAM_IOSCHED_DYNAMIC
757 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
758 sbintime_t sim_latency, int cmd, size_t size);
762 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
764 return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
768 cam_iosched_has_io(struct cam_iosched_softc *isc)
770 #ifdef CAM_IOSCHED_DYNAMIC
771 if (do_dynamic_iosched) {
772 struct bio *rbp = bioq_first(&isc->bio_queue);
773 struct bio *wbp = bioq_first(&isc->write_queue);
774 bool can_write = wbp != NULL &&
775 cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
776 bool can_read = rbp != NULL &&
777 cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
778 if (iosched_debug > 2) {
779 printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
780 printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
781 printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
783 return can_read || can_write;
786 return bioq_first(&isc->bio_queue) != NULL;
790 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
794 bp = bioq_first(&isc->trim_queue);
795 #ifdef CAM_IOSCHED_DYNAMIC
796 if (do_dynamic_iosched) {
798 * If we're limiting trims, then defer action on trims
801 if (bp == NULL || cam_iosched_limiter_caniop(&isc->trim_stats, bp) != 0)
807 * If we've set a trim_goal, then if we exceed that allow trims
808 * to be passed back to the driver. If we've also set a tick timeout
809 * allow trims back to the driver. Otherwise, don't allow trims yet.
811 if (isc->trim_goal > 0) {
812 if (isc->queued_trims >= isc->trim_goal)
814 if (isc->queued_trims > 0 &&
815 isc->trim_ticks > 0 &&
816 ticks - isc->last_trim_tick > isc->trim_ticks)
821 /* NB: Should perhaps have a max trim active independent of I/O limiters */
822 return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) && bp != NULL;
825 #define cam_iosched_sort_queue(isc) ((isc)->sort_io_queue >= 0 ? \
826 (isc)->sort_io_queue : cam_sort_io_queues)
829 cam_iosched_has_work(struct cam_iosched_softc *isc)
831 #ifdef CAM_IOSCHED_DYNAMIC
832 if (iosched_debug > 2)
833 printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
834 cam_iosched_has_more_trim(isc),
835 cam_iosched_has_flagged_work(isc));
838 return cam_iosched_has_io(isc) ||
839 cam_iosched_has_more_trim(isc) ||
840 cam_iosched_has_flagged_work(isc);
843 #ifdef CAM_IOSCHED_DYNAMIC
845 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
850 ios->max = ios->current = 300000;
860 cam_iosched_limiter_init(ios);
864 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
867 struct iop_stats *ios;
868 struct cam_iosched_softc *isc;
874 value = ios->limiter;
875 if (value < none || value >= limiter_max)
878 p = cam_iosched_limiter_names[value];
880 strlcpy(buf, p, sizeof(buf));
881 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
882 if (error != 0 || req->newptr == NULL)
885 cam_periph_lock(isc->periph);
887 for (i = none; i < limiter_max; i++) {
888 if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
891 error = cam_iosched_limiter_init(ios);
893 ios->limiter = value;
894 cam_periph_unlock(isc->periph);
897 /* Note: disk load averate requires ticker to be always running */
898 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
899 isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
901 cam_periph_unlock(isc->periph);
905 cam_periph_unlock(isc->periph);
910 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
913 struct control_loop *clp;
914 struct cam_iosched_softc *isc;
921 if (value < none || value >= cl_max)
924 p = cam_iosched_control_type_names[value];
926 strlcpy(buf, p, sizeof(buf));
927 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
928 if (error != 0 || req->newptr == NULL)
931 for (i = set_max; i < cl_max; i++) {
932 if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
934 cam_periph_lock(isc->periph);
936 cam_periph_unlock(isc->periph);
944 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
951 value = *(sbintime_t *)arg1;
952 us = (uint64_t)value / SBT_1US;
953 snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
954 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
955 if (error != 0 || req->newptr == NULL)
957 us = strtoul(buf, NULL, 10);
960 *(sbintime_t *)arg1 = us * SBT_1US;
965 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
972 sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
974 for (i = 0; i < LAT_BUCKETS - 1; i++)
975 sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
976 sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
977 error = sbuf_finish(&sb);
984 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS)
989 quanta = (unsigned *)arg1;
992 error = sysctl_handle_int(oidp, (int *)&value, 0, req);
993 if ((error != 0) || (req->newptr == NULL))
996 if (value < 1 || value > hz)
1005 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
1007 struct sysctl_oid_list *n;
1008 struct sysctl_ctx_list *ctx;
1010 ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1011 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
1012 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, name);
1013 n = SYSCTL_CHILDREN(ios->sysctl_tree);
1014 ctx = &ios->sysctl_ctx;
1016 SYSCTL_ADD_UQUAD(ctx, n,
1017 OID_AUTO, "ema", CTLFLAG_RD,
1019 "Fast Exponentially Weighted Moving Average");
1020 SYSCTL_ADD_UQUAD(ctx, n,
1021 OID_AUTO, "emvar", CTLFLAG_RD,
1023 "Fast Exponentially Weighted Moving Variance");
1025 SYSCTL_ADD_INT(ctx, n,
1026 OID_AUTO, "pending", CTLFLAG_RD,
1028 "Instantaneous # of pending transactions");
1029 SYSCTL_ADD_INT(ctx, n,
1030 OID_AUTO, "count", CTLFLAG_RD,
1032 "# of transactions submitted to hardware");
1033 SYSCTL_ADD_INT(ctx, n,
1034 OID_AUTO, "queued", CTLFLAG_RD,
1036 "# of transactions in the queue");
1037 SYSCTL_ADD_INT(ctx, n,
1038 OID_AUTO, "in", CTLFLAG_RD,
1040 "# of transactions queued to driver");
1041 SYSCTL_ADD_INT(ctx, n,
1042 OID_AUTO, "out", CTLFLAG_RD,
1044 "# of transactions completed (including with error)");
1045 SYSCTL_ADD_INT(ctx, n,
1046 OID_AUTO, "errs", CTLFLAG_RD,
1048 "# of transactions completed with an error");
1050 SYSCTL_ADD_PROC(ctx, n,
1051 OID_AUTO, "limiter",
1052 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1053 ios, 0, cam_iosched_limiter_sysctl, "A",
1054 "Current limiting type.");
1055 SYSCTL_ADD_INT(ctx, n,
1056 OID_AUTO, "min", CTLFLAG_RW,
1059 SYSCTL_ADD_INT(ctx, n,
1060 OID_AUTO, "max", CTLFLAG_RW,
1063 SYSCTL_ADD_INT(ctx, n,
1064 OID_AUTO, "current", CTLFLAG_RW,
1066 "current resource");
1068 SYSCTL_ADD_PROC(ctx, n,
1069 OID_AUTO, "latencies",
1070 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
1072 cam_iosched_sysctl_latencies, "A",
1073 "Array of power of 2 latency from 1ms to 1.024s");
1077 cam_iosched_iop_stats_fini(struct iop_stats *ios)
1079 if (ios->sysctl_tree)
1080 if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
1081 printf("can't remove iosched sysctl stats context\n");
1085 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
1087 struct sysctl_oid_list *n;
1088 struct sysctl_ctx_list *ctx;
1089 struct control_loop *clp;
1092 clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1093 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
1094 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Control loop info");
1095 n = SYSCTL_CHILDREN(clp->sysctl_tree);
1096 ctx = &clp->sysctl_ctx;
1098 SYSCTL_ADD_PROC(ctx, n,
1100 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1101 clp, 0, cam_iosched_control_type_sysctl, "A",
1102 "Control loop algorithm");
1103 SYSCTL_ADD_PROC(ctx, n,
1104 OID_AUTO, "steer_interval",
1105 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1106 &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
1107 "How often to steer (in us)");
1108 SYSCTL_ADD_PROC(ctx, n,
1110 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1111 &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
1112 "Low water mark for Latency (in us)");
1113 SYSCTL_ADD_PROC(ctx, n,
1115 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
1116 &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
1117 "Hi water mark for Latency (in us)");
1118 SYSCTL_ADD_INT(ctx, n,
1119 OID_AUTO, "alpha", CTLFLAG_RW,
1121 "Alpha for PLL (x100) aka gain");
1125 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1127 if (clp->sysctl_tree)
1128 if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1129 printf("can't remove iosched sysctl control loop context\n");
1134 * Allocate the iosched structure. This also insulates callers from knowing
1135 * sizeof struct cam_iosched_softc.
1138 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1141 *iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1144 #ifdef CAM_IOSCHED_DYNAMIC
1146 printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1148 (*iscp)->sort_io_queue = -1;
1149 bioq_init(&(*iscp)->bio_queue);
1150 bioq_init(&(*iscp)->trim_queue);
1151 #ifdef CAM_IOSCHED_DYNAMIC
1152 if (do_dynamic_iosched) {
1153 bioq_init(&(*iscp)->write_queue);
1154 (*iscp)->read_bias = default_read_bias;
1155 (*iscp)->current_read_bias = 0;
1156 (*iscp)->quanta = min(hz, 200);
1157 cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1158 cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1159 cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1160 (*iscp)->trim_stats.max = 1; /* Trims are special: one at a time for now */
1161 (*iscp)->last_time = sbinuptime();
1162 callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1163 (*iscp)->periph = periph;
1164 cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1165 callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp);
1166 (*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1174 * Reclaim all used resources. This assumes that other folks have
1175 * drained the requests in the hardware. Maybe an unwise assumption.
1178 cam_iosched_fini(struct cam_iosched_softc *isc)
1181 cam_iosched_flush(isc, NULL, ENXIO);
1182 #ifdef CAM_IOSCHED_DYNAMIC
1183 cam_iosched_iop_stats_fini(&isc->read_stats);
1184 cam_iosched_iop_stats_fini(&isc->write_stats);
1185 cam_iosched_iop_stats_fini(&isc->trim_stats);
1186 cam_iosched_cl_sysctl_fini(&isc->cl);
1187 if (isc->sysctl_tree)
1188 if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1189 printf("can't remove iosched sysctl stats context\n");
1190 if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1191 callout_drain(&isc->ticker);
1192 isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1195 free(isc, M_CAMSCHED);
1200 * After we're sure we're attaching a device, go ahead and add
1201 * hooks for any sysctl we may wish to honor.
1203 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1204 struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1206 struct sysctl_oid_list *n;
1208 n = SYSCTL_CHILDREN(node);
1209 SYSCTL_ADD_INT(ctx, n,
1210 OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1211 &isc->sort_io_queue, 0,
1212 "Sort IO queue to try and optimise disk access patterns");
1213 SYSCTL_ADD_INT(ctx, n,
1214 OID_AUTO, "trim_goal", CTLFLAG_RW,
1216 "Number of trims to try to accumulate before sending to hardware");
1217 SYSCTL_ADD_INT(ctx, n,
1218 OID_AUTO, "trim_ticks", CTLFLAG_RW,
1220 "IO Schedul qaunta to hold back trims for when accumulating");
1222 #ifdef CAM_IOSCHED_DYNAMIC
1223 if (!do_dynamic_iosched)
1226 isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1227 SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1228 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "I/O scheduler statistics");
1229 n = SYSCTL_CHILDREN(isc->sysctl_tree);
1230 ctx = &isc->sysctl_ctx;
1232 cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1233 cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1234 cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1235 cam_iosched_cl_sysctl_init(isc);
1237 SYSCTL_ADD_INT(ctx, n,
1238 OID_AUTO, "read_bias", CTLFLAG_RW,
1239 &isc->read_bias, default_read_bias,
1240 "How biased towards read should we be independent of limits");
1242 SYSCTL_ADD_PROC(ctx, n,
1243 OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE,
1244 &isc->quanta, 0, cam_iosched_quanta_sysctl, "I",
1245 "How many quanta per second do we slice the I/O up into");
1247 SYSCTL_ADD_INT(ctx, n,
1248 OID_AUTO, "total_ticks", CTLFLAG_RD,
1249 &isc->total_ticks, 0,
1250 "Total number of ticks we've done");
1252 SYSCTL_ADD_INT(ctx, n,
1253 OID_AUTO, "load", CTLFLAG_RD,
1255 "scaled load average / 100");
1257 SYSCTL_ADD_U64(ctx, n,
1258 OID_AUTO, "latency_trigger", CTLFLAG_RW,
1260 "Latency treshold to trigger callbacks");
1265 cam_iosched_set_latfcn(struct cam_iosched_softc *isc,
1266 cam_iosched_latfcn_t fnp, void *argp)
1268 #ifdef CAM_IOSCHED_DYNAMIC
1275 * Client drivers can set two parameters. "goal" is the number of BIO_DELETEs
1276 * that will be queued up before iosched will "release" the trims to the client
1277 * driver to wo with what they will (usually combine as many as possible). If we
1278 * don't get this many, after trim_ticks we'll submit the I/O anyway with
1279 * whatever we have. We do need an I/O of some kind of to clock the deferred
1280 * trims out to disk. Since we will eventually get a write for the super block
1281 * or something before we shutdown, the trims will complete. To be safe, when a
1282 * BIO_FLUSH is presented to the iosched work queue, we set the ticks time far
1283 * enough in the past so we'll present the BIO_DELETEs to the client driver.
1284 * There might be a race if no BIO_DELETESs were queued, a BIO_FLUSH comes in
1285 * and then a BIO_DELETE is sent down. No know client does this, and there's
1286 * already a race between an ordered BIO_FLUSH and any BIO_DELETEs in flight,
1287 * but no client depends on the ordering being honored.
1289 * XXX I'm not sure what the interaction between UFS direct BIOs and the BUF
1290 * flushing on shutdown. I think there's bufs that would be dependent on the BIO
1291 * finishing to write out at least metadata, so we'll be fine. To be safe, keep
1292 * the number of ticks low (less than maybe 10s) to avoid shutdown races.
1296 cam_iosched_set_trim_goal(struct cam_iosched_softc *isc, int goal)
1299 isc->trim_goal = goal;
1303 cam_iosched_set_trim_ticks(struct cam_iosched_softc *isc, int trim_ticks)
1306 isc->trim_ticks = trim_ticks;
1310 * Flush outstanding I/O. Consumers of this library don't know all the
1311 * queues we may keep, so this allows all I/O to be flushed in one
1315 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1317 bioq_flush(&isc->bio_queue, stp, err);
1318 bioq_flush(&isc->trim_queue, stp, err);
1319 #ifdef CAM_IOSCHED_DYNAMIC
1320 if (do_dynamic_iosched)
1321 bioq_flush(&isc->write_queue, stp, err);
1325 #ifdef CAM_IOSCHED_DYNAMIC
1327 cam_iosched_get_write(struct cam_iosched_softc *isc)
1332 * We control the write rate by controlling how many requests we send
1333 * down to the drive at any one time. Fewer requests limits the
1334 * effects of both starvation when the requests take a while and write
1335 * amplification when each request is causing more than one write to
1336 * the NAND media. Limiting the queue depth like this will also limit
1337 * the write throughput and give and reads that want to compete to
1340 bp = bioq_first(&isc->write_queue);
1342 if (iosched_debug > 3)
1343 printf("No writes present in write_queue\n");
1348 * If pending read, prefer that based on current read bias
1351 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1354 "Reads present and current_read_bias is %d queued "
1355 "writes %d queued reads %d\n",
1356 isc->current_read_bias, isc->write_stats.queued,
1357 isc->read_stats.queued);
1358 isc->current_read_bias--;
1359 /* We're not limiting writes, per se, just doing reads first */
1364 * See if our current limiter allows this I/O.
1366 if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1368 printf("Can't write because limiter says no.\n");
1369 isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1374 * Let's do this: We've passed all the gates and we're a go
1375 * to schedule the I/O in the SIM.
1377 isc->current_read_bias = isc->read_bias;
1378 bioq_remove(&isc->write_queue, bp);
1379 if (bp->bio_cmd == BIO_WRITE) {
1380 isc->write_stats.queued--;
1381 isc->write_stats.total++;
1382 isc->write_stats.pending++;
1384 if (iosched_debug > 9)
1385 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1386 isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1392 * Put back a trim that you weren't able to actually schedule this time.
1395 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1397 bioq_insert_head(&isc->trim_queue, bp);
1398 if (isc->queued_trims == 0)
1399 isc->last_trim_tick = ticks;
1400 isc->queued_trims++;
1401 #ifdef CAM_IOSCHED_DYNAMIC
1402 isc->trim_stats.queued++;
1403 isc->trim_stats.total--; /* since we put it back, don't double count */
1404 isc->trim_stats.pending--;
1409 * gets the next trim from the trim queue.
1411 * Assumes we're called with the periph lock held. It removes this
1412 * trim from the queue and the device must explicitly reinsert it
1413 * should the need arise.
1416 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1420 bp = bioq_first(&isc->trim_queue);
1423 bioq_remove(&isc->trim_queue, bp);
1424 isc->queued_trims--;
1425 isc->last_trim_tick = ticks; /* Reset the tick timer when we take trims */
1426 #ifdef CAM_IOSCHED_DYNAMIC
1427 isc->trim_stats.queued--;
1428 isc->trim_stats.total++;
1429 isc->trim_stats.pending++;
1435 * gets an available trim from the trim queue, if there's no trim
1436 * already pending. It removes this trim from the queue and the device
1437 * must explicitly reinsert it should the need arise.
1439 * Assumes we're called with the periph lock held.
1442 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1444 #ifdef CAM_IOSCHED_DYNAMIC
1448 if (!cam_iosched_has_more_trim(isc))
1450 #ifdef CAM_IOSCHED_DYNAMIC
1451 bp = bioq_first(&isc->trim_queue);
1456 * If pending read, prefer that based on current read bias setting. The
1457 * read bias is shared for both writes and TRIMs, but on TRIMs the bias
1458 * is for a combined TRIM not a single TRIM request that's come in.
1460 if (do_dynamic_iosched) {
1461 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1463 printf("Reads present and current_read_bias is %d"
1464 " queued trims %d queued reads %d\n",
1465 isc->current_read_bias, isc->trim_stats.queued,
1466 isc->read_stats.queued);
1467 isc->current_read_bias--;
1468 /* We're not limiting TRIMS, per se, just doing reads first */
1472 * We're going to do a trim, so reset the bias.
1474 isc->current_read_bias = isc->read_bias;
1478 * See if our current limiter allows this I/O. Because we only call this
1479 * here, and not in next_trim, the 'bandwidth' limits for trims won't
1480 * work, while the iops or max queued limits will work. It's tricky
1481 * because we want the limits to be from the perspective of the
1482 * "commands sent to the device." To make iops work, we need to check
1483 * only here (since we want all the ops we combine to count as one). To
1484 * make bw limits work, we'd need to check in next_trim, but that would
1485 * have the effect of limiting the iops as seen from the upper layers.
1487 if (cam_iosched_limiter_iop(&isc->trim_stats, bp) != 0) {
1489 printf("Can't trim because limiter says no.\n");
1490 isc->trim_stats.state_flags |= IOP_RATE_LIMITED;
1493 isc->current_read_bias = isc->read_bias;
1494 isc->trim_stats.state_flags &= ~IOP_RATE_LIMITED;
1495 /* cam_iosched_next_trim below keeps proper book */
1497 return cam_iosched_next_trim(isc);
1501 #ifdef CAM_IOSCHED_DYNAMIC
1503 bio_next(struct bio *bp)
1505 bp = TAILQ_NEXT(bp, bio_queue);
1507 * After the first commands, the ordered bit terminates
1508 * our search because BIO_ORDERED acts like a barrier.
1510 if (bp == NULL || bp->bio_flags & BIO_ORDERED)
1516 cam_iosched_rate_limited(struct iop_stats *ios)
1518 return ios->state_flags & IOP_RATE_LIMITED;
1523 * Determine what the next bit of work to do is for the periph. The
1524 * default implementation looks to see if we have trims to do, but no
1525 * trims outstanding. If so, we do that. Otherwise we see if we have
1526 * other work. If we do, then we do that. Otherwise why were we called?
1529 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1534 * See if we have a trim that can be scheduled. We can only send one
1535 * at a time down, so this takes that into account.
1537 * XXX newer TRIM commands are queueable. Revisit this when we
1540 if ((bp = cam_iosched_get_trim(isc)) != NULL)
1543 #ifdef CAM_IOSCHED_DYNAMIC
1545 * See if we have any pending writes, room in the queue for them,
1546 * and no pending reads (unless we've scheduled too many).
1547 * if so, those are next.
1549 if (do_dynamic_iosched) {
1550 if ((bp = cam_iosched_get_write(isc)) != NULL)
1555 * next, see if there's other, normal I/O waiting. If so return that.
1557 #ifdef CAM_IOSCHED_DYNAMIC
1558 if (do_dynamic_iosched) {
1559 for (bp = bioq_first(&isc->bio_queue); bp != NULL;
1560 bp = bio_next(bp)) {
1562 * For the dynamic scheduler with a read bias, bio_queue
1563 * is only for reads. However, without one, all
1564 * operations are queued. Enforce limits here for any
1565 * operation we find here.
1567 if (bp->bio_cmd == BIO_READ) {
1568 if (cam_iosched_rate_limited(&isc->read_stats) ||
1569 cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1570 isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1573 isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1576 * There can only be write requests on the queue when
1577 * the read bias is 0, but we need to process them
1578 * here. We do not assert for read bias == 0, however,
1579 * since it is dynamic and we can have WRITE operations
1580 * in the queue after we transition from 0 to non-zero.
1582 if (bp->bio_cmd == BIO_WRITE) {
1583 if (cam_iosched_rate_limited(&isc->write_stats) ||
1584 cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1585 isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1588 isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1591 * here we know we have a bp that's != NULL, that's not rate limited
1592 * and can be the next I/O.
1598 bp = bioq_first(&isc->bio_queue);
1602 bioq_remove(&isc->bio_queue, bp);
1603 #ifdef CAM_IOSCHED_DYNAMIC
1604 if (do_dynamic_iosched) {
1605 if (bp->bio_cmd == BIO_READ) {
1606 isc->read_stats.queued--;
1607 isc->read_stats.total++;
1608 isc->read_stats.pending++;
1609 } else if (bp->bio_cmd == BIO_WRITE) {
1610 isc->write_stats.queued--;
1611 isc->write_stats.total++;
1612 isc->write_stats.pending++;
1615 if (iosched_debug > 9)
1616 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1622 * Driver has been given some work to do by the block layer. Tell the
1623 * scheduler about it and have it queue the work up. The scheduler module
1624 * will then return the currently most useful bit of work later, possibly
1625 * deferring work for various reasons.
1628 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1632 * A BIO_SPEEDUP from the upper layers means that they have a block
1633 * shortage. At the present, this is only sent when we're trying to
1634 * allocate blocks, but have a shortage before giving up. bio_length is
1635 * the size of their shortage. We will complete just enough BIO_DELETEs
1636 * in the queue to satisfy the need. If bio_length is 0, we'll complete
1637 * them all. This allows the scheduler to delay BIO_DELETEs to improve
1638 * read/write performance without worrying about the upper layers. When
1639 * it's possibly a problem, we respond by pretending the BIO_DELETEs
1640 * just worked. We can't do anything about the BIO_DELETEs in the
1641 * hardware, though. We have to wait for them to complete.
1643 if (bp->bio_cmd == BIO_SPEEDUP) {
1648 while (bioq_first(&isc->trim_queue) &&
1649 (bp->bio_length == 0 || len < bp->bio_length)) {
1650 nbp = bioq_takefirst(&isc->trim_queue);
1651 len += nbp->bio_length;
1655 if (bp->bio_length > 0) {
1656 if (bp->bio_length > len)
1657 bp->bio_resid = bp->bio_length - len;
1667 * If we get a BIO_FLUSH, and we're doing delayed BIO_DELETEs then we
1668 * set the last tick time to one less than the current ticks minus the
1669 * delay to force the BIO_DELETEs to be presented to the client driver.
1671 if (bp->bio_cmd == BIO_FLUSH && isc->trim_ticks > 0)
1672 isc->last_trim_tick = ticks - isc->trim_ticks - 1;
1675 * Put all trims on the trim queue. Otherwise put the work on the bio
1678 if (bp->bio_cmd == BIO_DELETE) {
1679 bioq_insert_tail(&isc->trim_queue, bp);
1680 if (isc->queued_trims == 0)
1681 isc->last_trim_tick = ticks;
1682 isc->queued_trims++;
1683 #ifdef CAM_IOSCHED_DYNAMIC
1684 isc->trim_stats.in++;
1685 isc->trim_stats.queued++;
1688 #ifdef CAM_IOSCHED_DYNAMIC
1689 else if (do_dynamic_iosched && isc->read_bias != 0 &&
1690 (bp->bio_cmd != BIO_READ)) {
1691 if (cam_iosched_sort_queue(isc))
1692 bioq_disksort(&isc->write_queue, bp);
1694 bioq_insert_tail(&isc->write_queue, bp);
1695 if (iosched_debug > 9)
1696 printf("Qw : %p %#x\n", bp, bp->bio_cmd);
1697 if (bp->bio_cmd == BIO_WRITE) {
1698 isc->write_stats.in++;
1699 isc->write_stats.queued++;
1704 if (cam_iosched_sort_queue(isc))
1705 bioq_disksort(&isc->bio_queue, bp);
1707 bioq_insert_tail(&isc->bio_queue, bp);
1708 #ifdef CAM_IOSCHED_DYNAMIC
1709 if (iosched_debug > 9)
1710 printf("Qr : %p %#x\n", bp, bp->bio_cmd);
1711 if (bp->bio_cmd == BIO_READ) {
1712 isc->read_stats.in++;
1713 isc->read_stats.queued++;
1714 } else if (bp->bio_cmd == BIO_WRITE) {
1715 isc->write_stats.in++;
1716 isc->write_stats.queued++;
1723 * If we have work, get it scheduled. Called with the periph lock held.
1726 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1729 if (cam_iosched_has_work(isc))
1730 xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1734 * Complete a trim request. Mark that we no longer have one in flight.
1737 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1740 isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1744 * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1745 * might use notes in the ccb for statistics.
1748 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1749 union ccb *done_ccb)
1752 #ifdef CAM_IOSCHED_DYNAMIC
1753 if (!do_dynamic_iosched)
1756 if (iosched_debug > 10)
1757 printf("done: %p %#x\n", bp, bp->bio_cmd);
1758 if (bp->bio_cmd == BIO_WRITE) {
1759 retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1760 if ((bp->bio_flags & BIO_ERROR) != 0)
1761 isc->write_stats.errs++;
1762 isc->write_stats.out++;
1763 isc->write_stats.pending--;
1764 } else if (bp->bio_cmd == BIO_READ) {
1765 retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1766 if ((bp->bio_flags & BIO_ERROR) != 0)
1767 isc->read_stats.errs++;
1768 isc->read_stats.out++;
1769 isc->read_stats.pending--;
1770 } else if (bp->bio_cmd == BIO_DELETE) {
1771 if ((bp->bio_flags & BIO_ERROR) != 0)
1772 isc->trim_stats.errs++;
1773 isc->trim_stats.out++;
1774 isc->trim_stats.pending--;
1775 } else if (bp->bio_cmd != BIO_FLUSH) {
1777 printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1780 if ((bp->bio_flags & BIO_ERROR) == 0 && done_ccb != NULL &&
1781 (done_ccb->ccb_h.status & CAM_QOS_VALID) != 0) {
1782 sbintime_t sim_latency;
1784 sim_latency = cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data);
1786 cam_iosched_io_metric_update(isc, sim_latency,
1787 bp->bio_cmd, bp->bio_bcount);
1789 * Debugging code: allow callbacks to the periph driver when latency max
1790 * is exceeded. This can be useful for triggering external debugging actions.
1792 if (isc->latfcn && isc->max_lat != 0 && sim_latency > isc->max_lat)
1793 isc->latfcn(isc->latarg, sim_latency, bp);
1801 * Tell the io scheduler that you've pushed a trim down into the sim.
1802 * This also tells the I/O scheduler not to push any more trims down, so
1803 * some periphs do not call it if they can cope with multiple trims in flight.
1806 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1809 isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1813 * Change the sorting policy hint for I/O transactions for this device.
1816 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1819 isc->sort_io_queue = val;
1823 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1825 return isc->flags & flags;
1829 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1831 isc->flags |= flags;
1835 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1837 isc->flags &= ~flags;
1840 #ifdef CAM_IOSCHED_DYNAMIC
1842 * After the method presented in Jack Crenshaw's 1998 article "Integer
1843 * Square Roots," reprinted at
1844 * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1845 * and well worth the read. Briefly, we find the power of 4 that's the
1846 * largest smaller than val. We then check each smaller power of 4 to
1847 * see if val is still bigger. The right shifts at each step divide
1848 * the result by 2 which after successive application winds up
1849 * accumulating the right answer. It could also have been accumulated
1850 * using a separate root counter, but this code is smaller and faster
1851 * than that method. This method is also integer size invariant.
1852 * It returns floor(sqrt((float)val)), or the largest integer less than
1853 * or equal to the square root.
1856 isqrt64(uint64_t val)
1859 uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1862 * Find the largest power of 4 smaller than val.
1868 * Accumulate the answer, one bit at a time (we keep moving
1869 * them over since 2 is the square root of 4 and we test
1870 * powers of 4). We accumulate where we find the bit, but
1871 * the successive shifts land the bit in the right place
1875 if (val >= res + bit) {
1877 res = (res >> 1) + bit;
1886 static sbintime_t latencies[LAT_BUCKETS - 1] = {
1887 BUCKET_BASE << 0, /* 20us */
1905 BUCKET_BASE << 18 /* 5,242,880us */
1909 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1911 sbintime_t y, deltasq, delta;
1915 * Keep counts for latency. We do it by power of two buckets.
1916 * This helps us spot outlier behavior obscured by averages.
1918 for (i = 0; i < LAT_BUCKETS - 1; i++) {
1919 if (sim_latency < latencies[i]) {
1920 iop->latencies[i]++;
1924 if (i == LAT_BUCKETS - 1)
1925 iop->latencies[i]++; /* Put all > 8192ms values into the last bucket. */
1928 * Classic exponentially decaying average with a tiny alpha
1929 * (2 ^ -alpha_bits). For more info see the NIST statistical
1932 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha) [nist]
1933 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
1934 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1935 * alpha = 1 / (1 << alpha_bits)
1936 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
1937 * = y_t/b - e/b + be/b
1938 * = (y_t - e + be) / b
1941 * Since alpha is a power of two, we can compute this w/o any mult or
1944 * Variance can also be computed. Usually, it would be expressed as follows:
1945 * diff_t = y_t - ema_t-1
1946 * emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
1947 * = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
1948 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
1949 * = e - e/b + dd/b + dd/bb
1950 * = (bbe - be + bdd + dd) / bb
1951 * = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
1954 * XXX possible numeric issues
1955 * o We assume right shifted integers do the right thing, since that's
1956 * implementation defined. You can change the right shifts to / (1LL << alpha).
1957 * o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
1958 * for emvar. This puts a ceiling of 13 bits on alpha since we need a
1959 * few tens of seconds of representation.
1960 * o We mitigate alpha issues by never setting it too high.
1963 delta = (y - iop->ema); /* d */
1964 iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1967 * Were we to naively plow ahead at this point, we wind up with many numerical
1968 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
1969 * us with microsecond level precision in the input, so the same in the
1970 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
1971 * also means that emvar can be up 46 bits 40 of which are fraction, which
1972 * gives us a way to measure up to ~8s in the SD before the computation goes
1973 * unstable. Even the worst hard disk rarely has > 1s service time in the
1974 * drive. It does mean we have to shift left 12 bits after taking the
1975 * square root to compute the actual standard deviation estimate. This loss of
1976 * precision is preferable to needing int128 types to work. The above numbers
1977 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
1978 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1981 deltasq = delta * delta; /* dd */
1982 iop->emvar = ((iop->emvar << (2 * alpha_bits)) + /* bbe */
1983 ((deltasq - iop->emvar) << alpha_bits) + /* b(dd-e) */
1985 >> (2 * alpha_bits); /* div bb */
1986 iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1990 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1991 sbintime_t sim_latency, int cmd, size_t size)
1993 /* xxx Do we need to scale based on the size of the I/O ? */
1996 cam_iosched_update(&isc->read_stats, sim_latency);
1999 cam_iosched_update(&isc->write_stats, sim_latency);
2002 cam_iosched_update(&isc->trim_stats, sim_latency);
2010 static int biolen(struct bio_queue_head *bq)
2015 TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
2022 * Show the internal state of the I/O scheduler.
2024 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
2026 struct cam_iosched_softc *isc;
2029 db_printf("Need addr\n");
2032 isc = (struct cam_iosched_softc *)addr;
2033 db_printf("pending_reads: %d\n", isc->read_stats.pending);
2034 db_printf("min_reads: %d\n", isc->read_stats.min);
2035 db_printf("max_reads: %d\n", isc->read_stats.max);
2036 db_printf("reads: %d\n", isc->read_stats.total);
2037 db_printf("in_reads: %d\n", isc->read_stats.in);
2038 db_printf("out_reads: %d\n", isc->read_stats.out);
2039 db_printf("queued_reads: %d\n", isc->read_stats.queued);
2040 db_printf("Read Q len %d\n", biolen(&isc->bio_queue));
2041 db_printf("pending_writes: %d\n", isc->write_stats.pending);
2042 db_printf("min_writes: %d\n", isc->write_stats.min);
2043 db_printf("max_writes: %d\n", isc->write_stats.max);
2044 db_printf("writes: %d\n", isc->write_stats.total);
2045 db_printf("in_writes: %d\n", isc->write_stats.in);
2046 db_printf("out_writes: %d\n", isc->write_stats.out);
2047 db_printf("queued_writes: %d\n", isc->write_stats.queued);
2048 db_printf("Write Q len %d\n", biolen(&isc->write_queue));
2049 db_printf("pending_trims: %d\n", isc->trim_stats.pending);
2050 db_printf("min_trims: %d\n", isc->trim_stats.min);
2051 db_printf("max_trims: %d\n", isc->trim_stats.max);
2052 db_printf("trims: %d\n", isc->trim_stats.total);
2053 db_printf("in_trims: %d\n", isc->trim_stats.in);
2054 db_printf("out_trims: %d\n", isc->trim_stats.out);
2055 db_printf("queued_trims: %d\n", isc->trim_stats.queued);
2056 db_printf("Trim Q len %d\n", biolen(&isc->trim_queue));
2057 db_printf("read_bias: %d\n", isc->read_bias);
2058 db_printf("current_read_bias: %d\n", isc->current_read_bias);
2059 db_printf("Trim active? %s\n",
2060 (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");