FreeBSD/Linux Kernel Cross Reference
sys/cam/cam_iosched.c
1 /*-
2 * CAM IO Scheduler Interface
3 *
4 * SPDX-License-Identifier: BSD-2-Clause-FreeBSD
5 *
6 * Copyright (c) 2015 Netflix, Inc.
7 *
8 * Redistribution and use in source and binary forms, with or without
9 * modification, are permitted provided that the following conditions
10 * are met:
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.
16 *
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
27 * SUCH DAMAGE.
28 *
29 * $FreeBSD$
30 */
31
32 #include "opt_cam.h"
33 #include "opt_ddb.h"
34
35 #include <sys/cdefs.h>
36 __FBSDID("$FreeBSD$");
37
38 #include <sys/param.h>
39
40 #include <sys/systm.h>
41 #include <sys/kernel.h>
42 #include <sys/bio.h>
43 #include <sys/lock.h>
44 #include <sys/malloc.h>
45 #include <sys/mutex.h>
46 #include <sys/sbuf.h>
47 #include <sys/sysctl.h>
48
49 #include <cam/cam.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>
55
56 #include <ddb/ddb.h>
57
58 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler",
59 "CAM I/O Scheduler buffers");
60
61 /*
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
64 * for trims.
65 *
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).
69 */
70
71 #ifdef CAM_IOSCHED_DYNAMIC
72
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.");
77
78 /*
79 * For an EMA, with an alpha of alpha, we know
80 * alpha = 2 / (N + 1)
81 * or
82 * N = 1 + (2 / alpha)
83 * where N is the number of samples that 86% of the current
84 * EMA is derived from.
85 *
86 * So we invent[*] alpha_bits:
87 * alpha_bits = -log_2(alpha)
88 * alpha = 2^-alpha_bits
89 * So
90 * N = 1 + 2^(alpha_bits + 1)
91 *
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
94 *
95 * [*] Steal from the load average code and many other places.
96 * Note: See computation of EMA and EMVAR for acceptable ranges of alpha.
97 */
98 static int alpha_bits = 9;
99 SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW | CTLFLAG_TUN,
100 &alpha_bits, 1,
101 "Bits in EMA's alpha.");
102
103 struct iop_stats;
104 struct cam_iosched_softc;
105
106 int iosched_debug = 0;
107
108 typedef enum {
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 */
113 limiter_max
114 } io_limiter;
115
116 static const char *cam_iosched_limiter_names[] =
117 { "none", "queue_depth", "iops", "bandwidth" };
118
119 /*
120 * Called to initialize the bits of the iop_stats structure relevant to the
121 * limiter. Called just after the limiter is set.
122 */
123 typedef int l_init_t(struct iop_stats *);
124
125 /*
126 * Called every tick.
127 */
128 typedef int l_tick_t(struct iop_stats *);
129
130 /*
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.
134 */
135 typedef int l_iop_t(struct iop_stats *, struct bio *);
136
137 /*
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.
143 */
144 typedef int l_iodone_t(struct iop_stats *, struct bio *);
145
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;
149
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;
154
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;
159
160 struct limswitch {
161 l_init_t *l_init;
162 l_tick_t *l_tick;
163 l_iop_t *l_iop;
164 l_iop_t *l_caniop;
165 l_iodone_t *l_iodone;
166 } limsw[] =
167 {
168 { /* none */
169 .l_init = NULL,
170 .l_tick = NULL,
171 .l_iop = NULL,
172 .l_iodone= NULL,
173 },
174 { /* queue_depth */
175 .l_init = NULL,
176 .l_tick = NULL,
177 .l_caniop = cam_iosched_qd_caniop,
178 .l_iop = cam_iosched_qd_iop,
179 .l_iodone= cam_iosched_qd_iodone,
180 },
181 { /* iops */
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,
186 .l_iodone= NULL,
187 },
188 { /* bandwidth */
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,
193 .l_iodone= NULL,
194 },
195 };
196
197 struct iop_stats {
198 /*
199 * sysctl state for this subnode.
200 */
201 struct sysctl_ctx_list sysctl_ctx;
202 struct sysctl_oid *sysctl_tree;
203
204 /*
205 * Information about the current rate limiters, if any
206 */
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. */
213
214 /*
215 * Debug information about counts of I/Os that have gone through the
216 * scheduler.
217 */
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 */
224
225 /*
226 * Statistics on different bits of the process.
227 */
228 /* Exp Moving Average, see alpha_bits for more details */
229 sbintime_t ema;
230 sbintime_t emvar;
231 sbintime_t sd; /* Last computed sd */
232
233 uint32_t state_flags;
234 #define IOP_RATE_LIMITED 1u
235
236 #define LAT_BUCKETS 15 /* < 1ms < 2ms ... < 2^(n-1)ms >= 2^(n-1)ms*/
237 uint64_t latencies[LAT_BUCKETS];
238
239 struct cam_iosched_softc *softc;
240 };
241
242
243 typedef enum {
244 set_max = 0, /* current = max */
245 read_latency, /* Steer read latency by throttling writes */
246 cl_max /* Keep last */
247 } control_type;
248
249 static const char *cam_iosched_control_type_names[] =
250 { "set_max", "read_latency" };
251
252 struct control_loop {
253 /*
254 * sysctl state for this subnode.
255 */
256 struct sysctl_ctx_list sysctl_ctx;
257 struct sysctl_oid *sysctl_tree;
258
259 sbintime_t next_steer; /* Time of next steer */
260 sbintime_t steer_interval; /* How often do we steer? */
261 sbintime_t lolat;
262 sbintime_t hilat;
263 int alpha;
264 control_type type; /* What type of control? */
265 int last_count; /* Last I/O count */
266
267 struct cam_iosched_softc *softc;
268 };
269
270 #endif
271
272 struct cam_iosched_softc {
273 struct bio_queue_head bio_queue;
274 struct bio_queue_head trim_queue;
275 /* scheduler flags < 16, user flags >= 16 */
276 uint32_t flags;
277 int sort_io_queue;
278 #ifdef CAM_IOSCHED_DYNAMIC
279 int read_bias; /* Read bias setting */
280 int current_read_bias; /* Current read bias state */
281 int total_ticks;
282 int load; /* EMA of 'load average' of disk / 2^16 */
283
284 struct bio_queue_head write_queue;
285 struct iop_stats read_stats, write_stats, trim_stats;
286 struct sysctl_ctx_list sysctl_ctx;
287 struct sysctl_oid *sysctl_tree;
288
289 int quanta; /* Number of quanta per second */
290 struct callout ticker; /* Callout for our quota system */
291 struct cam_periph *periph; /* cam periph associated with this device */
292 uint32_t this_frac; /* Fraction of a second (1024ths) for this tick */
293 sbintime_t last_time; /* Last time we ticked */
294 struct control_loop cl;
295 #endif
296 };
297
298 #ifdef CAM_IOSCHED_DYNAMIC
299 /*
300 * helper functions to call the limsw functions.
301 */
302 static int
303 cam_iosched_limiter_init(struct iop_stats *ios)
304 {
305 int lim = ios->limiter;
306
307 /* maybe this should be a kassert */
308 if (lim < none || lim >= limiter_max)
309 return EINVAL;
310
311 if (limsw[lim].l_init)
312 return limsw[lim].l_init(ios);
313
314 return 0;
315 }
316
317 static int
318 cam_iosched_limiter_tick(struct iop_stats *ios)
319 {
320 int lim = ios->limiter;
321
322 /* maybe this should be a kassert */
323 if (lim < none || lim >= limiter_max)
324 return EINVAL;
325
326 if (limsw[lim].l_tick)
327 return limsw[lim].l_tick(ios);
328
329 return 0;
330 }
331
332 static int
333 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp)
334 {
335 int lim = ios->limiter;
336
337 /* maybe this should be a kassert */
338 if (lim < none || lim >= limiter_max)
339 return EINVAL;
340
341 if (limsw[lim].l_iop)
342 return limsw[lim].l_iop(ios, bp);
343
344 return 0;
345 }
346
347 static int
348 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp)
349 {
350 int lim = ios->limiter;
351
352 /* maybe this should be a kassert */
353 if (lim < none || lim >= limiter_max)
354 return EINVAL;
355
356 if (limsw[lim].l_caniop)
357 return limsw[lim].l_caniop(ios, bp);
358
359 return 0;
360 }
361
362 static int
363 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp)
364 {
365 int lim = ios->limiter;
366
367 /* maybe this should be a kassert */
368 if (lim < none || lim >= limiter_max)
369 return 0;
370
371 if (limsw[lim].l_iodone)
372 return limsw[lim].l_iodone(ios, bp);
373
374 return 0;
375 }
376
377 /*
378 * Functions to implement the different kinds of limiters
379 */
380
381 static int
382 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp)
383 {
384
385 if (ios->current <= 0 || ios->pending < ios->current)
386 return 0;
387
388 return EAGAIN;
389 }
390
391 static int
392 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp)
393 {
394
395 if (ios->current <= 0 || ios->pending < ios->current)
396 return 0;
397
398 return EAGAIN;
399 }
400
401 static int
402 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp)
403 {
404
405 if (ios->current <= 0 || ios->pending != ios->current)
406 return 0;
407
408 return 1;
409 }
410
411 static int
412 cam_iosched_iops_init(struct iop_stats *ios)
413 {
414
415 ios->l_value1 = ios->current / ios->softc->quanta;
416 if (ios->l_value1 <= 0)
417 ios->l_value1 = 1;
418 ios->l_value2 = 0;
419
420 return 0;
421 }
422
423 static int
424 cam_iosched_iops_tick(struct iop_stats *ios)
425 {
426 int new_ios;
427
428 /*
429 * Allow at least one IO per tick until all
430 * the IOs for this interval have been spent.
431 */
432 new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16);
433 if (new_ios < 1 && ios->l_value2 < ios->current) {
434 new_ios = 1;
435 ios->l_value2++;
436 }
437
438 /*
439 * If this a new accounting interval, discard any "unspent" ios
440 * granted in the previous interval. Otherwise add the new ios to
441 * the previously granted ones that haven't been spent yet.
442 */
443 if ((ios->softc->total_ticks % ios->softc->quanta) == 0) {
444 ios->l_value1 = new_ios;
445 ios->l_value2 = 1;
446 } else {
447 ios->l_value1 += new_ios;
448 }
449
450
451 return 0;
452 }
453
454 static int
455 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp)
456 {
457
458 /*
459 * So if we have any more IOPs left, allow it,
460 * otherwise wait. If current iops is 0, treat that
461 * as unlimited as a failsafe.
462 */
463 if (ios->current > 0 && ios->l_value1 <= 0)
464 return EAGAIN;
465 return 0;
466 }
467
468 static int
469 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp)
470 {
471 int rv;
472
473 rv = cam_iosched_limiter_caniop(ios, bp);
474 if (rv == 0)
475 ios->l_value1--;
476
477 return rv;
478 }
479
480 static int
481 cam_iosched_bw_init(struct iop_stats *ios)
482 {
483
484 /* ios->current is in kB/s, so scale to bytes */
485 ios->l_value1 = ios->current * 1000 / ios->softc->quanta;
486
487 return 0;
488 }
489
490 static int
491 cam_iosched_bw_tick(struct iop_stats *ios)
492 {
493 int bw;
494
495 /*
496 * If we're in the hole for available quota from
497 * the last time, then add the quantum for this.
498 * If we have any left over from last quantum,
499 * then too bad, that's lost. Also, ios->current
500 * is in kB/s, so scale.
501 *
502 * We also allow up to 4 quanta of credits to
503 * accumulate to deal with burstiness. 4 is extremely
504 * arbitrary.
505 */
506 bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16);
507 if (ios->l_value1 < bw * 4)
508 ios->l_value1 += bw;
509
510 return 0;
511 }
512
513 static int
514 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp)
515 {
516 /*
517 * So if we have any more bw quota left, allow it,
518 * otherwise wait. Note, we'll go negative and that's
519 * OK. We'll just get a little less next quota.
520 *
521 * Note on going negative: that allows us to process
522 * requests in order better, since we won't allow
523 * shorter reads to get around the long one that we
524 * don't have the quota to do just yet. It also prevents
525 * starvation by being a little more permissive about
526 * what we let through this quantum (to prevent the
527 * starvation), at the cost of getting a little less
528 * next quantum.
529 *
530 * Also note that if the current limit is <= 0,
531 * we treat it as unlimited as a failsafe.
532 */
533 if (ios->current > 0 && ios->l_value1 <= 0)
534 return EAGAIN;
535
536
537 return 0;
538 }
539
540 static int
541 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp)
542 {
543 int rv;
544
545 rv = cam_iosched_limiter_caniop(ios, bp);
546 if (rv == 0)
547 ios->l_value1 -= bp->bio_length;
548
549 return rv;
550 }
551
552 static void cam_iosched_cl_maybe_steer(struct control_loop *clp);
553
554 static void
555 cam_iosched_ticker(void *arg)
556 {
557 struct cam_iosched_softc *isc = arg;
558 sbintime_t now, delta;
559 int pending;
560
561 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
562
563 now = sbinuptime();
564 delta = now - isc->last_time;
565 isc->this_frac = (uint32_t)delta >> 16; /* Note: discards seconds -- should be 0 harmless if not */
566 isc->last_time = now;
567
568 cam_iosched_cl_maybe_steer(&isc->cl);
569
570 cam_iosched_limiter_tick(&isc->read_stats);
571 cam_iosched_limiter_tick(&isc->write_stats);
572 cam_iosched_limiter_tick(&isc->trim_stats);
573
574 cam_iosched_schedule(isc, isc->periph);
575
576 /*
577 * isc->load is an EMA of the pending I/Os at each tick. The number of
578 * pending I/Os is the sum of the I/Os queued to the hardware, and those
579 * in the software queue that could be queued to the hardware if there
580 * were slots.
581 *
582 * ios_stats.pending is a count of requests in the SIM right now for
583 * each of these types of I/O. So the total pending count is the sum of
584 * these I/Os and the sum of the queued I/Os still in the software queue
585 * for those operations that aren't being rate limited at the moment.
586 *
587 * The reason for the rate limiting bit is because those I/Os
588 * aren't part of the software queued load (since we could
589 * give them to hardware, but choose not to).
590 *
591 * Note: due to a bug in counting pending TRIM in the device, we
592 * don't include them in this count. We count each BIO_DELETE in
593 * the pending count, but the periph drivers collapse them down
594 * into one TRIM command. That one trim command gets the completion
595 * so the counts get off.
596 */
597 pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */;
598 pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued +
599 !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* +
600 !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ;
601 pending <<= 16;
602 pending /= isc->periph->path->device->ccbq.total_openings;
603
604 isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */
605
606 isc->total_ticks++;
607 }
608
609
610 static void
611 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc)
612 {
613
614 clp->next_steer = sbinuptime();
615 clp->softc = isc;
616 clp->steer_interval = SBT_1S * 5; /* Let's start out steering every 5s */
617 clp->lolat = 5 * SBT_1MS;
618 clp->hilat = 15 * SBT_1MS;
619 clp->alpha = 20; /* Alpha == gain. 20 = .2 */
620 clp->type = set_max;
621 }
622
623 static void
624 cam_iosched_cl_maybe_steer(struct control_loop *clp)
625 {
626 struct cam_iosched_softc *isc;
627 sbintime_t now, lat;
628 int old;
629
630 isc = clp->softc;
631 now = isc->last_time;
632 if (now < clp->next_steer)
633 return;
634
635 clp->next_steer = now + clp->steer_interval;
636 switch (clp->type) {
637 case set_max:
638 if (isc->write_stats.current != isc->write_stats.max)
639 printf("Steering write from %d kBps to %d kBps\n",
640 isc->write_stats.current, isc->write_stats.max);
641 isc->read_stats.current = isc->read_stats.max;
642 isc->write_stats.current = isc->write_stats.max;
643 isc->trim_stats.current = isc->trim_stats.max;
644 break;
645 case read_latency:
646 old = isc->write_stats.current;
647 lat = isc->read_stats.ema;
648 /*
649 * Simple PLL-like engine. Since we're steering to a range for
650 * the SP (set point) that makes things a little more
651 * complicated. In addition, we're not directly controlling our
652 * PV (process variable), the read latency, but instead are
653 * manipulating the write bandwidth limit for our MV
654 * (manipulation variable), analysis of this code gets a bit
655 * messy. Also, the MV is a very noisy control surface for read
656 * latency since it is affected by many hidden processes inside
657 * the device which change how responsive read latency will be
658 * in reaction to changes in write bandwidth. Unlike the classic
659 * boiler control PLL. this may result in over-steering while
660 * the SSD takes its time to react to the new, lower load. This
661 * is why we use a relatively low alpha of between .1 and .25 to
662 * compensate for this effect. At .1, it takes ~22 steering
663 * intervals to back off by a factor of 10. At .2 it only takes
664 * ~10. At .25 it only takes ~8. However some preliminary data
665 * from the SSD drives suggests a reasponse time in 10's of
666 * seconds before latency drops regardless of the new write
667 * rate. Careful observation will be required to tune this
668 * effectively.
669 *
670 * Also, when there's no read traffic, we jack up the write
671 * limit too regardless of the last read latency. 10 is
672 * somewhat arbitrary.
673 */
674 if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10)
675 isc->write_stats.current = isc->write_stats.current *
676 (100 + clp->alpha) / 100; /* Scale up */
677 else if (lat > clp->hilat)
678 isc->write_stats.current = isc->write_stats.current *
679 (100 - clp->alpha) / 100; /* Scale down */
680 clp->last_count = isc->read_stats.total;
681
682 /*
683 * Even if we don't steer, per se, enforce the min/max limits as
684 * those may have changed.
685 */
686 if (isc->write_stats.current < isc->write_stats.min)
687 isc->write_stats.current = isc->write_stats.min;
688 if (isc->write_stats.current > isc->write_stats.max)
689 isc->write_stats.current = isc->write_stats.max;
690 if (old != isc->write_stats.current && iosched_debug)
691 printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n",
692 old, isc->write_stats.current,
693 (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32);
694 break;
695 case cl_max:
696 break;
697 }
698 }
699 #endif
700
701 /*
702 * Trim or similar currently pending completion. Should only be set for
703 * those drivers wishing only one Trim active at a time.
704 */
705 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE (1ul << 0)
706 /* Callout active, and needs to be torn down */
707 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1)
708
709 /* Periph drivers set these flags to indicate work */
710 #define CAM_IOSCHED_FLAG_WORK_FLAGS ((0xffffu) << 16)
711
712 #ifdef CAM_IOSCHED_DYNAMIC
713 static void
714 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
715 sbintime_t sim_latency, int cmd, size_t size);
716 #endif
717
718 static inline bool
719 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc)
720 {
721 return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS);
722 }
723
724 static inline bool
725 cam_iosched_has_io(struct cam_iosched_softc *isc)
726 {
727 #ifdef CAM_IOSCHED_DYNAMIC
728 if (do_dynamic_iosched) {
729 struct bio *rbp = bioq_first(&isc->bio_queue);
730 struct bio *wbp = bioq_first(&isc->write_queue);
731 bool can_write = wbp != NULL &&
732 cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0;
733 bool can_read = rbp != NULL &&
734 cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0;
735 if (iosched_debug > 2) {
736 printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max);
737 printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max);
738 printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued);
739 }
740 return can_read || can_write;
741 }
742 #endif
743 return bioq_first(&isc->bio_queue) != NULL;
744 }
745
746 static inline bool
747 cam_iosched_has_more_trim(struct cam_iosched_softc *isc)
748 {
749 return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) &&
750 bioq_first(&isc->trim_queue);
751 }
752
753 #define cam_iosched_sort_queue(isc) ((isc)->sort_io_queue >= 0 ? \
754 (isc)->sort_io_queue : cam_sort_io_queues)
755
756
757 static inline bool
758 cam_iosched_has_work(struct cam_iosched_softc *isc)
759 {
760 #ifdef CAM_IOSCHED_DYNAMIC
761 if (iosched_debug > 2)
762 printf("has work: %d %d %d\n", cam_iosched_has_io(isc),
763 cam_iosched_has_more_trim(isc),
764 cam_iosched_has_flagged_work(isc));
765 #endif
766
767 return cam_iosched_has_io(isc) ||
768 cam_iosched_has_more_trim(isc) ||
769 cam_iosched_has_flagged_work(isc);
770 }
771
772 #ifdef CAM_IOSCHED_DYNAMIC
773 static void
774 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios)
775 {
776
777 ios->limiter = none;
778 ios->in = 0;
779 ios->max = ios->current = 300000;
780 ios->min = 1;
781 ios->out = 0;
782 ios->errs = 0;
783 ios->pending = 0;
784 ios->queued = 0;
785 ios->total = 0;
786 ios->ema = 0;
787 ios->emvar = 0;
788 ios->softc = isc;
789 cam_iosched_limiter_init(ios);
790 }
791
792 static int
793 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS)
794 {
795 char buf[16];
796 struct iop_stats *ios;
797 struct cam_iosched_softc *isc;
798 int value, i, error;
799 const char *p;
800
801 ios = arg1;
802 isc = ios->softc;
803 value = ios->limiter;
804 if (value < none || value >= limiter_max)
805 p = "UNKNOWN";
806 else
807 p = cam_iosched_limiter_names[value];
808
809 strlcpy(buf, p, sizeof(buf));
810 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
811 if (error != 0 || req->newptr == NULL)
812 return error;
813
814 cam_periph_lock(isc->periph);
815
816 for (i = none; i < limiter_max; i++) {
817 if (strcmp(buf, cam_iosched_limiter_names[i]) != 0)
818 continue;
819 ios->limiter = i;
820 error = cam_iosched_limiter_init(ios);
821 if (error != 0) {
822 ios->limiter = value;
823 cam_periph_unlock(isc->periph);
824 return error;
825 }
826 /* Note: disk load averate requires ticker to be always running */
827 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc);
828 isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
829
830 cam_periph_unlock(isc->periph);
831 return 0;
832 }
833
834 cam_periph_unlock(isc->periph);
835 return EINVAL;
836 }
837
838 static int
839 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS)
840 {
841 char buf[16];
842 struct control_loop *clp;
843 struct cam_iosched_softc *isc;
844 int value, i, error;
845 const char *p;
846
847 clp = arg1;
848 isc = clp->softc;
849 value = clp->type;
850 if (value < none || value >= cl_max)
851 p = "UNKNOWN";
852 else
853 p = cam_iosched_control_type_names[value];
854
855 strlcpy(buf, p, sizeof(buf));
856 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
857 if (error != 0 || req->newptr == NULL)
858 return error;
859
860 for (i = set_max; i < cl_max; i++) {
861 if (strcmp(buf, cam_iosched_control_type_names[i]) != 0)
862 continue;
863 cam_periph_lock(isc->periph);
864 clp->type = i;
865 cam_periph_unlock(isc->periph);
866 return 0;
867 }
868
869 return EINVAL;
870 }
871
872 static int
873 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS)
874 {
875 char buf[16];
876 sbintime_t value;
877 int error;
878 uint64_t us;
879
880 value = *(sbintime_t *)arg1;
881 us = (uint64_t)value / SBT_1US;
882 snprintf(buf, sizeof(buf), "%ju", (intmax_t)us);
883 error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
884 if (error != 0 || req->newptr == NULL)
885 return error;
886 us = strtoul(buf, NULL, 10);
887 if (us == 0)
888 return EINVAL;
889 *(sbintime_t *)arg1 = us * SBT_1US;
890 return 0;
891 }
892
893 static int
894 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS)
895 {
896 int i, error;
897 struct sbuf sb;
898 uint64_t *latencies;
899
900 latencies = arg1;
901 sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req);
902
903 for (i = 0; i < LAT_BUCKETS - 1; i++)
904 sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]);
905 sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]);
906 error = sbuf_finish(&sb);
907 sbuf_delete(&sb);
908
909 return (error);
910 }
911
912 static int
913 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS)
914 {
915 int *quanta;
916 int error, value;
917
918 quanta = (unsigned *)arg1;
919 value = *quanta;
920
921 error = sysctl_handle_int(oidp, (int *)&value, 0, req);
922 if ((error != 0) || (req->newptr == NULL))
923 return (error);
924
925 if (value < 1 || value > hz)
926 return (EINVAL);
927
928 *quanta = value;
929
930 return (0);
931 }
932
933 static void
934 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name)
935 {
936 struct sysctl_oid_list *n;
937 struct sysctl_ctx_list *ctx;
938
939 ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
940 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name,
941 CTLFLAG_RD, 0, name);
942 n = SYSCTL_CHILDREN(ios->sysctl_tree);
943 ctx = &ios->sysctl_ctx;
944
945 SYSCTL_ADD_UQUAD(ctx, n,
946 OID_AUTO, "ema", CTLFLAG_RD,
947 &ios->ema,
948 "Fast Exponentially Weighted Moving Average");
949 SYSCTL_ADD_UQUAD(ctx, n,
950 OID_AUTO, "emvar", CTLFLAG_RD,
951 &ios->emvar,
952 "Fast Exponentially Weighted Moving Variance");
953
954 SYSCTL_ADD_INT(ctx, n,
955 OID_AUTO, "pending", CTLFLAG_RD,
956 &ios->pending, 0,
957 "Instantaneous # of pending transactions");
958 SYSCTL_ADD_INT(ctx, n,
959 OID_AUTO, "count", CTLFLAG_RD,
960 &ios->total, 0,
961 "# of transactions submitted to hardware");
962 SYSCTL_ADD_INT(ctx, n,
963 OID_AUTO, "queued", CTLFLAG_RD,
964 &ios->queued, 0,
965 "# of transactions in the queue");
966 SYSCTL_ADD_INT(ctx, n,
967 OID_AUTO, "in", CTLFLAG_RD,
968 &ios->in, 0,
969 "# of transactions queued to driver");
970 SYSCTL_ADD_INT(ctx, n,
971 OID_AUTO, "out", CTLFLAG_RD,
972 &ios->out, 0,
973 "# of transactions completed (including with error)");
974 SYSCTL_ADD_INT(ctx, n,
975 OID_AUTO, "errs", CTLFLAG_RD,
976 &ios->errs, 0,
977 "# of transactions completed with an error");
978
979 SYSCTL_ADD_PROC(ctx, n,
980 OID_AUTO, "limiter", CTLTYPE_STRING | CTLFLAG_RW,
981 ios, 0, cam_iosched_limiter_sysctl, "A",
982 "Current limiting type.");
983 SYSCTL_ADD_INT(ctx, n,
984 OID_AUTO, "min", CTLFLAG_RW,
985 &ios->min, 0,
986 "min resource");
987 SYSCTL_ADD_INT(ctx, n,
988 OID_AUTO, "max", CTLFLAG_RW,
989 &ios->max, 0,
990 "max resource");
991 SYSCTL_ADD_INT(ctx, n,
992 OID_AUTO, "current", CTLFLAG_RW,
993 &ios->current, 0,
994 "current resource");
995
996 SYSCTL_ADD_PROC(ctx, n,
997 OID_AUTO, "latencies", CTLTYPE_STRING | CTLFLAG_RD,
998 &ios->latencies, 0,
999 cam_iosched_sysctl_latencies, "A",
1000 "Array of power of 2 latency from 1ms to 1.024s");
1001 }
1002
1003 static void
1004 cam_iosched_iop_stats_fini(struct iop_stats *ios)
1005 {
1006 if (ios->sysctl_tree)
1007 if (sysctl_ctx_free(&ios->sysctl_ctx) != 0)
1008 printf("can't remove iosched sysctl stats context\n");
1009 }
1010
1011 static void
1012 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc)
1013 {
1014 struct sysctl_oid_list *n;
1015 struct sysctl_ctx_list *ctx;
1016 struct control_loop *clp;
1017
1018 clp = &isc->cl;
1019 clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1020 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control",
1021 CTLFLAG_RD, 0, "Control loop info");
1022 n = SYSCTL_CHILDREN(clp->sysctl_tree);
1023 ctx = &clp->sysctl_ctx;
1024
1025 SYSCTL_ADD_PROC(ctx, n,
1026 OID_AUTO, "type", CTLTYPE_STRING | CTLFLAG_RW,
1027 clp, 0, cam_iosched_control_type_sysctl, "A",
1028 "Control loop algorithm");
1029 SYSCTL_ADD_PROC(ctx, n,
1030 OID_AUTO, "steer_interval", CTLTYPE_STRING | CTLFLAG_RW,
1031 &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A",
1032 "How often to steer (in us)");
1033 SYSCTL_ADD_PROC(ctx, n,
1034 OID_AUTO, "lolat", CTLTYPE_STRING | CTLFLAG_RW,
1035 &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A",
1036 "Low water mark for Latency (in us)");
1037 SYSCTL_ADD_PROC(ctx, n,
1038 OID_AUTO, "hilat", CTLTYPE_STRING | CTLFLAG_RW,
1039 &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A",
1040 "Hi water mark for Latency (in us)");
1041 SYSCTL_ADD_INT(ctx, n,
1042 OID_AUTO, "alpha", CTLFLAG_RW,
1043 &clp->alpha, 0,
1044 "Alpha for PLL (x100) aka gain");
1045 }
1046
1047 static void
1048 cam_iosched_cl_sysctl_fini(struct control_loop *clp)
1049 {
1050 if (clp->sysctl_tree)
1051 if (sysctl_ctx_free(&clp->sysctl_ctx) != 0)
1052 printf("can't remove iosched sysctl control loop context\n");
1053 }
1054 #endif
1055
1056 /*
1057 * Allocate the iosched structure. This also insulates callers from knowing
1058 * sizeof struct cam_iosched_softc.
1059 */
1060 int
1061 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph)
1062 {
1063
1064 *iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO);
1065 if (*iscp == NULL)
1066 return ENOMEM;
1067 #ifdef CAM_IOSCHED_DYNAMIC
1068 if (iosched_debug)
1069 printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp);
1070 #endif
1071 (*iscp)->sort_io_queue = -1;
1072 bioq_init(&(*iscp)->bio_queue);
1073 bioq_init(&(*iscp)->trim_queue);
1074 #ifdef CAM_IOSCHED_DYNAMIC
1075 if (do_dynamic_iosched) {
1076 bioq_init(&(*iscp)->write_queue);
1077 (*iscp)->read_bias = 100;
1078 (*iscp)->current_read_bias = 100;
1079 (*iscp)->quanta = min(hz, 200);
1080 cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats);
1081 cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats);
1082 cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats);
1083 (*iscp)->trim_stats.max = 1; /* Trims are special: one at a time for now */
1084 (*iscp)->last_time = sbinuptime();
1085 callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0);
1086 (*iscp)->periph = periph;
1087 cam_iosched_cl_init(&(*iscp)->cl, *iscp);
1088 callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp);
1089 (*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1090 }
1091 #endif
1092
1093 return 0;
1094 }
1095
1096 /*
1097 * Reclaim all used resources. This assumes that other folks have
1098 * drained the requests in the hardware. Maybe an unwise assumption.
1099 */
1100 void
1101 cam_iosched_fini(struct cam_iosched_softc *isc)
1102 {
1103 if (isc) {
1104 cam_iosched_flush(isc, NULL, ENXIO);
1105 #ifdef CAM_IOSCHED_DYNAMIC
1106 cam_iosched_iop_stats_fini(&isc->read_stats);
1107 cam_iosched_iop_stats_fini(&isc->write_stats);
1108 cam_iosched_iop_stats_fini(&isc->trim_stats);
1109 cam_iosched_cl_sysctl_fini(&isc->cl);
1110 if (isc->sysctl_tree)
1111 if (sysctl_ctx_free(&isc->sysctl_ctx) != 0)
1112 printf("can't remove iosched sysctl stats context\n");
1113 if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) {
1114 callout_drain(&isc->ticker);
1115 isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE;
1116 }
1117 #endif
1118 free(isc, M_CAMSCHED);
1119 }
1120 }
1121
1122 /*
1123 * After we're sure we're attaching a device, go ahead and add
1124 * hooks for any sysctl we may wish to honor.
1125 */
1126 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc,
1127 struct sysctl_ctx_list *ctx, struct sysctl_oid *node)
1128 {
1129 #ifdef CAM_IOSCHED_DYNAMIC
1130 struct sysctl_oid_list *n;
1131 #endif
1132
1133 SYSCTL_ADD_INT(ctx, SYSCTL_CHILDREN(node),
1134 OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE,
1135 &isc->sort_io_queue, 0,
1136 "Sort IO queue to try and optimise disk access patterns");
1137
1138 #ifdef CAM_IOSCHED_DYNAMIC
1139 if (!do_dynamic_iosched)
1140 return;
1141
1142 isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx,
1143 SYSCTL_CHILDREN(node), OID_AUTO, "iosched",
1144 CTLFLAG_RD, 0, "I/O scheduler statistics");
1145 n = SYSCTL_CHILDREN(isc->sysctl_tree);
1146 ctx = &isc->sysctl_ctx;
1147
1148 cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read");
1149 cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write");
1150 cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim");
1151 cam_iosched_cl_sysctl_init(isc);
1152
1153 SYSCTL_ADD_INT(ctx, n,
1154 OID_AUTO, "read_bias", CTLFLAG_RW,
1155 &isc->read_bias, 100,
1156 "How biased towards read should we be independent of limits");
1157
1158 SYSCTL_ADD_PROC(ctx, n,
1159 OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW,
1160 &isc->quanta, 0, cam_iosched_quanta_sysctl, "I",
1161 "How many quanta per second do we slice the I/O up into");
1162
1163 SYSCTL_ADD_INT(ctx, n,
1164 OID_AUTO, "total_ticks", CTLFLAG_RD,
1165 &isc->total_ticks, 0,
1166 "Total number of ticks we've done");
1167
1168 SYSCTL_ADD_INT(ctx, n,
1169 OID_AUTO, "load", CTLFLAG_RD,
1170 &isc->load, 0,
1171 "scaled load average / 100");
1172 #endif
1173 }
1174
1175 /*
1176 * Flush outstanding I/O. Consumers of this library don't know all the
1177 * queues we may keep, so this allows all I/O to be flushed in one
1178 * convenient call.
1179 */
1180 void
1181 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err)
1182 {
1183 bioq_flush(&isc->bio_queue, stp, err);
1184 bioq_flush(&isc->trim_queue, stp, err);
1185 #ifdef CAM_IOSCHED_DYNAMIC
1186 if (do_dynamic_iosched)
1187 bioq_flush(&isc->write_queue, stp, err);
1188 #endif
1189 }
1190
1191 #ifdef CAM_IOSCHED_DYNAMIC
1192 static struct bio *
1193 cam_iosched_get_write(struct cam_iosched_softc *isc)
1194 {
1195 struct bio *bp;
1196
1197 /*
1198 * We control the write rate by controlling how many requests we send
1199 * down to the drive at any one time. Fewer requests limits the
1200 * effects of both starvation when the requests take a while and write
1201 * amplification when each request is causing more than one write to
1202 * the NAND media. Limiting the queue depth like this will also limit
1203 * the write throughput and give and reads that want to compete to
1204 * compete unfairly.
1205 */
1206 bp = bioq_first(&isc->write_queue);
1207 if (bp == NULL) {
1208 if (iosched_debug > 3)
1209 printf("No writes present in write_queue\n");
1210 return NULL;
1211 }
1212
1213 /*
1214 * If pending read, prefer that based on current read bias
1215 * setting.
1216 */
1217 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1218 if (iosched_debug)
1219 printf(
1220 "Reads present and current_read_bias is %d queued "
1221 "writes %d queued reads %d\n",
1222 isc->current_read_bias, isc->write_stats.queued,
1223 isc->read_stats.queued);
1224 isc->current_read_bias--;
1225 /* We're not limiting writes, per se, just doing reads first */
1226 return NULL;
1227 }
1228
1229 /*
1230 * See if our current limiter allows this I/O.
1231 */
1232 if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) {
1233 if (iosched_debug)
1234 printf("Can't write because limiter says no.\n");
1235 isc->write_stats.state_flags |= IOP_RATE_LIMITED;
1236 return NULL;
1237 }
1238
1239 /*
1240 * Let's do this: We've passed all the gates and we're a go
1241 * to schedule the I/O in the SIM.
1242 */
1243 isc->current_read_bias = isc->read_bias;
1244 bioq_remove(&isc->write_queue, bp);
1245 if (bp->bio_cmd == BIO_WRITE) {
1246 isc->write_stats.queued--;
1247 isc->write_stats.total++;
1248 isc->write_stats.pending++;
1249 }
1250 if (iosched_debug > 9)
1251 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1252 isc->write_stats.state_flags &= ~IOP_RATE_LIMITED;
1253 return bp;
1254 }
1255 #endif
1256
1257 /*
1258 * Put back a trim that you weren't able to actually schedule this time.
1259 */
1260 void
1261 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp)
1262 {
1263 bioq_insert_head(&isc->trim_queue, bp);
1264 #ifdef CAM_IOSCHED_DYNAMIC
1265 isc->trim_stats.queued++;
1266 isc->trim_stats.total--; /* since we put it back, don't double count */
1267 isc->trim_stats.pending--;
1268 #endif
1269 }
1270
1271 /*
1272 * gets the next trim from the trim queue.
1273 *
1274 * Assumes we're called with the periph lock held. It removes this
1275 * trim from the queue and the device must explicitly reinsert it
1276 * should the need arise.
1277 */
1278 struct bio *
1279 cam_iosched_next_trim(struct cam_iosched_softc *isc)
1280 {
1281 struct bio *bp;
1282
1283 bp = bioq_first(&isc->trim_queue);
1284 if (bp == NULL)
1285 return NULL;
1286 bioq_remove(&isc->trim_queue, bp);
1287 #ifdef CAM_IOSCHED_DYNAMIC
1288 isc->trim_stats.queued--;
1289 isc->trim_stats.total++;
1290 isc->trim_stats.pending++;
1291 #endif
1292 return bp;
1293 }
1294
1295 /*
1296 * gets an available trim from the trim queue, if there's no trim
1297 * already pending. It removes this trim from the queue and the device
1298 * must explicitly reinsert it should the need arise.
1299 *
1300 * Assumes we're called with the periph lock held.
1301 */
1302 struct bio *
1303 cam_iosched_get_trim(struct cam_iosched_softc *isc)
1304 {
1305
1306 if (!cam_iosched_has_more_trim(isc))
1307 return NULL;
1308 #ifdef CAM_IOSCHED_DYNAMIC
1309 if (do_dynamic_iosched) {
1310 /*
1311 * If pending read, prefer that based on current read bias
1312 * setting. The read bias is shared for both writes and
1313 * TRIMs, but on TRIMs the bias is for a combined TRIM
1314 * not a single TRIM request that's come in.
1315 */
1316 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) {
1317 isc->current_read_bias--;
1318 /* We're not limiting TRIMS, per se, just doing reads first */
1319 return NULL;
1320 }
1321 /*
1322 * We're going to do a trim, so reset the bias.
1323 */
1324 isc->current_read_bias = isc->read_bias;
1325 }
1326 #endif
1327 return cam_iosched_next_trim(isc);
1328 }
1329
1330 /*
1331 * Determine what the next bit of work to do is for the periph. The
1332 * default implementation looks to see if we have trims to do, but no
1333 * trims outstanding. If so, we do that. Otherwise we see if we have
1334 * other work. If we do, then we do that. Otherwise why were we called?
1335 */
1336 struct bio *
1337 cam_iosched_next_bio(struct cam_iosched_softc *isc)
1338 {
1339 struct bio *bp;
1340
1341 /*
1342 * See if we have a trim that can be scheduled. We can only send one
1343 * at a time down, so this takes that into account.
1344 *
1345 * XXX newer TRIM commands are queueable. Revisit this when we
1346 * implement them.
1347 */
1348 if ((bp = cam_iosched_get_trim(isc)) != NULL)
1349 return bp;
1350
1351 #ifdef CAM_IOSCHED_DYNAMIC
1352 /*
1353 * See if we have any pending writes, and room in the queue for them,
1354 * and if so, those are next.
1355 */
1356 if (do_dynamic_iosched) {
1357 if ((bp = cam_iosched_get_write(isc)) != NULL)
1358 return bp;
1359 }
1360 #endif
1361
1362 /*
1363 * next, see if there's other, normal I/O waiting. If so return that.
1364 */
1365 if ((bp = bioq_first(&isc->bio_queue)) == NULL)
1366 return NULL;
1367
1368 #ifdef CAM_IOSCHED_DYNAMIC
1369 /*
1370 * For the dynamic scheduler, bio_queue is only for reads, so enforce
1371 * the limits here. Enforce only for reads.
1372 */
1373 if (do_dynamic_iosched) {
1374 if (bp->bio_cmd == BIO_READ &&
1375 cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) {
1376 isc->read_stats.state_flags |= IOP_RATE_LIMITED;
1377 return NULL;
1378 }
1379 }
1380 isc->read_stats.state_flags &= ~IOP_RATE_LIMITED;
1381 #endif
1382 bioq_remove(&isc->bio_queue, bp);
1383 #ifdef CAM_IOSCHED_DYNAMIC
1384 if (do_dynamic_iosched) {
1385 if (bp->bio_cmd == BIO_READ) {
1386 isc->read_stats.queued--;
1387 isc->read_stats.total++;
1388 isc->read_stats.pending++;
1389 } else
1390 printf("Found bio_cmd = %#x\n", bp->bio_cmd);
1391 }
1392 if (iosched_debug > 9)
1393 printf("HWQ : %p %#x\n", bp, bp->bio_cmd);
1394 #endif
1395 return bp;
1396 }
1397
1398 /*
1399 * Driver has been given some work to do by the block layer. Tell the
1400 * scheduler about it and have it queue the work up. The scheduler module
1401 * will then return the currently most useful bit of work later, possibly
1402 * deferring work for various reasons.
1403 */
1404 void
1405 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp)
1406 {
1407
1408 /*
1409 * Put all trims on the trim queue sorted, since we know
1410 * that the collapsing code requires this. Otherwise put
1411 * the work on the bio queue.
1412 */
1413 if (bp->bio_cmd == BIO_DELETE) {
1414 bioq_insert_tail(&isc->trim_queue, bp);
1415 #ifdef CAM_IOSCHED_DYNAMIC
1416 isc->trim_stats.in++;
1417 isc->trim_stats.queued++;
1418 #endif
1419 }
1420 #ifdef CAM_IOSCHED_DYNAMIC
1421 else if (do_dynamic_iosched && (bp->bio_cmd != BIO_READ)) {
1422 if (cam_iosched_sort_queue(isc))
1423 bioq_disksort(&isc->write_queue, bp);
1424 else
1425 bioq_insert_tail(&isc->write_queue, bp);
1426 if (iosched_debug > 9)
1427 printf("Qw : %p %#x\n", bp, bp->bio_cmd);
1428 if (bp->bio_cmd == BIO_WRITE) {
1429 isc->write_stats.in++;
1430 isc->write_stats.queued++;
1431 }
1432 }
1433 #endif
1434 else {
1435 if (cam_iosched_sort_queue(isc))
1436 bioq_disksort(&isc->bio_queue, bp);
1437 else
1438 bioq_insert_tail(&isc->bio_queue, bp);
1439 #ifdef CAM_IOSCHED_DYNAMIC
1440 if (iosched_debug > 9)
1441 printf("Qr : %p %#x\n", bp, bp->bio_cmd);
1442 if (bp->bio_cmd == BIO_READ) {
1443 isc->read_stats.in++;
1444 isc->read_stats.queued++;
1445 } else if (bp->bio_cmd == BIO_WRITE) {
1446 isc->write_stats.in++;
1447 isc->write_stats.queued++;
1448 }
1449 #endif
1450 }
1451 }
1452
1453 /*
1454 * If we have work, get it scheduled. Called with the periph lock held.
1455 */
1456 void
1457 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph)
1458 {
1459
1460 if (cam_iosched_has_work(isc))
1461 xpt_schedule(periph, CAM_PRIORITY_NORMAL);
1462 }
1463
1464 /*
1465 * Complete a trim request. Mark that we no longer have one in flight.
1466 */
1467 void
1468 cam_iosched_trim_done(struct cam_iosched_softc *isc)
1469 {
1470
1471 isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1472 }
1473
1474 /*
1475 * Complete a bio. Called before we release the ccb with xpt_release_ccb so we
1476 * might use notes in the ccb for statistics.
1477 */
1478 int
1479 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp,
1480 union ccb *done_ccb)
1481 {
1482 int retval = 0;
1483 #ifdef CAM_IOSCHED_DYNAMIC
1484 if (!do_dynamic_iosched)
1485 return retval;
1486
1487 if (iosched_debug > 10)
1488 printf("done: %p %#x\n", bp, bp->bio_cmd);
1489 if (bp->bio_cmd == BIO_WRITE) {
1490 retval = cam_iosched_limiter_iodone(&isc->write_stats, bp);
1491 if ((bp->bio_flags & BIO_ERROR) != 0)
1492 isc->write_stats.errs++;
1493 isc->write_stats.out++;
1494 isc->write_stats.pending--;
1495 } else if (bp->bio_cmd == BIO_READ) {
1496 retval = cam_iosched_limiter_iodone(&isc->read_stats, bp);
1497 if ((bp->bio_flags & BIO_ERROR) != 0)
1498 isc->read_stats.errs++;
1499 isc->read_stats.out++;
1500 isc->read_stats.pending--;
1501 } else if (bp->bio_cmd == BIO_DELETE) {
1502 if ((bp->bio_flags & BIO_ERROR) != 0)
1503 isc->trim_stats.errs++;
1504 isc->trim_stats.out++;
1505 isc->trim_stats.pending--;
1506 } else if (bp->bio_cmd != BIO_FLUSH) {
1507 if (iosched_debug)
1508 printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd);
1509 }
1510
1511 if (!(bp->bio_flags & BIO_ERROR) && done_ccb != NULL)
1512 cam_iosched_io_metric_update(isc,
1513 cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data),
1514 bp->bio_cmd, bp->bio_bcount);
1515 #endif
1516 return retval;
1517 }
1518
1519 /*
1520 * Tell the io scheduler that you've pushed a trim down into the sim.
1521 * This also tells the I/O scheduler not to push any more trims down, so
1522 * some periphs do not call it if they can cope with multiple trims in flight.
1523 */
1524 void
1525 cam_iosched_submit_trim(struct cam_iosched_softc *isc)
1526 {
1527
1528 isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE;
1529 }
1530
1531 /*
1532 * Change the sorting policy hint for I/O transactions for this device.
1533 */
1534 void
1535 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val)
1536 {
1537
1538 isc->sort_io_queue = val;
1539 }
1540
1541 int
1542 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1543 {
1544 return isc->flags & flags;
1545 }
1546
1547 void
1548 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1549 {
1550 isc->flags |= flags;
1551 }
1552
1553 void
1554 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags)
1555 {
1556 isc->flags &= ~flags;
1557 }
1558
1559 #ifdef CAM_IOSCHED_DYNAMIC
1560 /*
1561 * After the method presented in Jack Crenshaw's 1998 article "Integer
1562 * Square Roots," reprinted at
1563 * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots
1564 * and well worth the read. Briefly, we find the power of 4 that's the
1565 * largest smaller than val. We then check each smaller power of 4 to
1566 * see if val is still bigger. The right shifts at each step divide
1567 * the result by 2 which after successive application winds up
1568 * accumulating the right answer. It could also have been accumulated
1569 * using a separate root counter, but this code is smaller and faster
1570 * than that method. This method is also integer size invariant.
1571 * It returns floor(sqrt((float)val)), or the largest integer less than
1572 * or equal to the square root.
1573 */
1574 static uint64_t
1575 isqrt64(uint64_t val)
1576 {
1577 uint64_t res = 0;
1578 uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2);
1579
1580 /*
1581 * Find the largest power of 4 smaller than val.
1582 */
1583 while (bit > val)
1584 bit >>= 2;
1585
1586 /*
1587 * Accumulate the answer, one bit at a time (we keep moving
1588 * them over since 2 is the square root of 4 and we test
1589 * powers of 4). We accumulate where we find the bit, but
1590 * the successive shifts land the bit in the right place
1591 * by the end.
1592 */
1593 while (bit != 0) {
1594 if (val >= res + bit) {
1595 val -= res + bit;
1596 res = (res >> 1) + bit;
1597 } else
1598 res >>= 1;
1599 bit >>= 2;
1600 }
1601
1602 return res;
1603 }
1604
1605 static sbintime_t latencies[LAT_BUCKETS - 1] = {
1606 SBT_1MS << 0,
1607 SBT_1MS << 1,
1608 SBT_1MS << 2,
1609 SBT_1MS << 3,
1610 SBT_1MS << 4,
1611 SBT_1MS << 5,
1612 SBT_1MS << 6,
1613 SBT_1MS << 7,
1614 SBT_1MS << 8,
1615 SBT_1MS << 9,
1616 SBT_1MS << 10,
1617 SBT_1MS << 11,
1618 SBT_1MS << 12,
1619 SBT_1MS << 13 /* 8.192s */
1620 };
1621
1622 static void
1623 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency)
1624 {
1625 sbintime_t y, deltasq, delta;
1626 int i;
1627
1628 /*
1629 * Keep counts for latency. We do it by power of two buckets.
1630 * This helps us spot outlier behavior obscured by averages.
1631 */
1632 for (i = 0; i < LAT_BUCKETS - 1; i++) {
1633 if (sim_latency < latencies[i]) {
1634 iop->latencies[i]++;
1635 break;
1636 }
1637 }
1638 if (i == LAT_BUCKETS - 1)
1639 iop->latencies[i]++; /* Put all > 1024ms values into the last bucket. */
1640
1641 /*
1642 * Classic exponentially decaying average with a tiny alpha
1643 * (2 ^ -alpha_bits). For more info see the NIST statistical
1644 * handbook.
1645 *
1646 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha) [nist]
1647 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1
1648 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1
1649 * alpha = 1 / (1 << alpha_bits)
1650 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1
1651 * = y_t/b - e/b + be/b
1652 * = (y_t - e + be) / b
1653 * = (e + d) / b
1654 *
1655 * Since alpha is a power of two, we can compute this w/o any mult or
1656 * division.
1657 *
1658 * Variance can also be computed. Usually, it would be expressed as follows:
1659 * diff_t = y_t - ema_t-1
1660 * emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha)
1661 * = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2
1662 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2
1663 * = e - e/b + dd/b + dd/bb
1664 * = (bbe - be + bdd + dd) / bb
1665 * = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits))
1666 */
1667 /*
1668 * XXX possible numeric issues
1669 * o We assume right shifted integers do the right thing, since that's
1670 * implementation defined. You can change the right shifts to / (1LL << alpha).
1671 * o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits
1672 * for emvar. This puts a ceiling of 13 bits on alpha since we need a
1673 * few tens of seconds of representation.
1674 * o We mitigate alpha issues by never setting it too high.
1675 */
1676 y = sim_latency;
1677 delta = (y - iop->ema); /* d */
1678 iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits;
1679
1680 /*
1681 * Were we to naively plow ahead at this point, we wind up with many numerical
1682 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves
1683 * us with microsecond level precision in the input, so the same in the
1684 * output. It means we can't overflow deltasq unless delta > 4k seconds. It
1685 * also means that emvar can be up 46 bits 40 of which are fraction, which
1686 * gives us a way to measure up to ~8s in the SD before the computation goes
1687 * unstable. Even the worst hard disk rarely has > 1s service time in the
1688 * drive. It does mean we have to shift left 12 bits after taking the
1689 * square root to compute the actual standard deviation estimate. This loss of
1690 * precision is preferable to needing int128 types to work. The above numbers
1691 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12,
1692 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases.
1693 */
1694 delta >>= 12;
1695 deltasq = delta * delta; /* dd */
1696 iop->emvar = ((iop->emvar << (2 * alpha_bits)) + /* bbe */
1697 ((deltasq - iop->emvar) << alpha_bits) + /* b(dd-e) */
1698 deltasq) /* dd */
1699 >> (2 * alpha_bits); /* div bb */
1700 iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12;
1701 }
1702
1703 static void
1704 cam_iosched_io_metric_update(struct cam_iosched_softc *isc,
1705 sbintime_t sim_latency, int cmd, size_t size)
1706 {
1707 /* xxx Do we need to scale based on the size of the I/O ? */
1708 switch (cmd) {
1709 case BIO_READ:
1710 cam_iosched_update(&isc->read_stats, sim_latency);
1711 break;
1712 case BIO_WRITE:
1713 cam_iosched_update(&isc->write_stats, sim_latency);
1714 break;
1715 case BIO_DELETE:
1716 cam_iosched_update(&isc->trim_stats, sim_latency);
1717 break;
1718 default:
1719 break;
1720 }
1721 }
1722
1723 #ifdef DDB
1724 static int biolen(struct bio_queue_head *bq)
1725 {
1726 int i = 0;
1727 struct bio *bp;
1728
1729 TAILQ_FOREACH(bp, &bq->queue, bio_queue) {
1730 i++;
1731 }
1732 return i;
1733 }
1734
1735 /*
1736 * Show the internal state of the I/O scheduler.
1737 */
1738 DB_SHOW_COMMAND(iosched, cam_iosched_db_show)
1739 {
1740 struct cam_iosched_softc *isc;
1741
1742 if (!have_addr) {
1743 db_printf("Need addr\n");
1744 return;
1745 }
1746 isc = (struct cam_iosched_softc *)addr;
1747 db_printf("pending_reads: %d\n", isc->read_stats.pending);
1748 db_printf("min_reads: %d\n", isc->read_stats.min);
1749 db_printf("max_reads: %d\n", isc->read_stats.max);
1750 db_printf("reads: %d\n", isc->read_stats.total);
1751 db_printf("in_reads: %d\n", isc->read_stats.in);
1752 db_printf("out_reads: %d\n", isc->read_stats.out);
1753 db_printf("queued_reads: %d\n", isc->read_stats.queued);
1754 db_printf("Current Q len %d\n", biolen(&isc->bio_queue));
1755 db_printf("pending_writes: %d\n", isc->write_stats.pending);
1756 db_printf("min_writes: %d\n", isc->write_stats.min);
1757 db_printf("max_writes: %d\n", isc->write_stats.max);
1758 db_printf("writes: %d\n", isc->write_stats.total);
1759 db_printf("in_writes: %d\n", isc->write_stats.in);
1760 db_printf("out_writes: %d\n", isc->write_stats.out);
1761 db_printf("queued_writes: %d\n", isc->write_stats.queued);
1762 db_printf("Current Q len %d\n", biolen(&isc->write_queue));
1763 db_printf("pending_trims: %d\n", isc->trim_stats.pending);
1764 db_printf("min_trims: %d\n", isc->trim_stats.min);
1765 db_printf("max_trims: %d\n", isc->trim_stats.max);
1766 db_printf("trims: %d\n", isc->trim_stats.total);
1767 db_printf("in_trims: %d\n", isc->trim_stats.in);
1768 db_printf("out_trims: %d\n", isc->trim_stats.out);
1769 db_printf("queued_trims: %d\n", isc->trim_stats.queued);
1770 db_printf("Current Q len %d\n", biolen(&isc->trim_queue));
1771 db_printf("read_bias: %d\n", isc->read_bias);
1772 db_printf("current_read_bias: %d\n", isc->current_read_bias);
1773 db_printf("Trim active? %s\n",
1774 (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no");
1775 }
1776 #endif
1777 #endif
Cache object: 4fab059fa66f2cc645efad39d1b59c43
|