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 int do_dynamic_iosched = 1; 74 TUNABLE_INT("kern.cam.do_dynamic_iosched", &do_dynamic_iosched); 75 SYSCTL_INT(_kern_cam, OID_AUTO, do_dynamic_iosched, CTLFLAG_RD, 76 &do_dynamic_iosched, 1, 77 "Enable Dynamic I/O scheduler optimizations."); 78 79 /* 80 * For an EMA, with an alpha of alpha, we know 81 * alpha = 2 / (N + 1) 82 * or 83 * N = 1 + (2 / alpha) 84 * where N is the number of samples that 86% of the current 85 * EMA is derived from. 86 * 87 * So we invent[*] alpha_bits: 88 * alpha_bits = -log_2(alpha) 89 * alpha = 2^-alpha_bits 90 * So 91 * N = 1 + 2^(alpha_bits + 1) 92 * 93 * The default 9 gives a 1025 lookback for 86% of the data. 94 * For a brief intro: https://en.wikipedia.org/wiki/Moving_average 95 * 96 * [*] Steal from the load average code and many other places. 97 * Note: See computation of EMA and EMVAR for acceptable ranges of alpha. 98 */ 99 static int alpha_bits = 9; 100 TUNABLE_INT("kern.cam.iosched_alpha_bits", &alpha_bits); 101 SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW, 102 &alpha_bits, 1, 103 "Bits in EMA's alpha."); 104 105 struct iop_stats; 106 struct cam_iosched_softc; 107 108 int iosched_debug = 0; 109 110 typedef enum { 111 none = 0, /* No limits */ 112 queue_depth, /* Limit how many ops we queue to SIM */ 113 iops, /* Limit # of IOPS to the drive */ 114 bandwidth, /* Limit bandwidth to the drive */ 115 limiter_max 116 } io_limiter; 117 118 static const char *cam_iosched_limiter_names[] = 119 { "none", "queue_depth", "iops", "bandwidth" }; 120 121 /* 122 * Called to initialize the bits of the iop_stats structure relevant to the 123 * limiter. Called just after the limiter is set. 124 */ 125 typedef int l_init_t(struct iop_stats *); 126 127 /* 128 * Called every tick. 129 */ 130 typedef int l_tick_t(struct iop_stats *); 131 132 /* 133 * Called to see if the limiter thinks this IOP can be allowed to 134 * proceed. If so, the limiter assumes that the IOP proceeded 135 * and makes any accounting of it that's needed. 136 */ 137 typedef int l_iop_t(struct iop_stats *, struct bio *); 138 139 /* 140 * Called when an I/O completes so the limiter can update its 141 * accounting. Pending I/Os may complete in any order (even when 142 * sent to the hardware at the same time), so the limiter may not 143 * make any assumptions other than this I/O has completed. If it 144 * returns 1, then xpt_schedule() needs to be called again. 145 */ 146 typedef int l_iodone_t(struct iop_stats *, struct bio *); 147 148 static l_iop_t cam_iosched_qd_iop; 149 static l_iop_t cam_iosched_qd_caniop; 150 static l_iodone_t cam_iosched_qd_iodone; 151 152 static l_init_t cam_iosched_iops_init; 153 static l_tick_t cam_iosched_iops_tick; 154 static l_iop_t cam_iosched_iops_caniop; 155 static l_iop_t cam_iosched_iops_iop; 156 157 static l_init_t cam_iosched_bw_init; 158 static l_tick_t cam_iosched_bw_tick; 159 static l_iop_t cam_iosched_bw_caniop; 160 static l_iop_t cam_iosched_bw_iop; 161 162 struct limswitch { 163 l_init_t *l_init; 164 l_tick_t *l_tick; 165 l_iop_t *l_iop; 166 l_iop_t *l_caniop; 167 l_iodone_t *l_iodone; 168 } limsw[] = 169 { 170 { /* none */ 171 .l_init = NULL, 172 .l_tick = NULL, 173 .l_iop = NULL, 174 .l_iodone= NULL, 175 }, 176 { /* queue_depth */ 177 .l_init = NULL, 178 .l_tick = NULL, 179 .l_caniop = cam_iosched_qd_caniop, 180 .l_iop = cam_iosched_qd_iop, 181 .l_iodone= cam_iosched_qd_iodone, 182 }, 183 { /* iops */ 184 .l_init = cam_iosched_iops_init, 185 .l_tick = cam_iosched_iops_tick, 186 .l_caniop = cam_iosched_iops_caniop, 187 .l_iop = cam_iosched_iops_iop, 188 .l_iodone= NULL, 189 }, 190 { /* bandwidth */ 191 .l_init = cam_iosched_bw_init, 192 .l_tick = cam_iosched_bw_tick, 193 .l_caniop = cam_iosched_bw_caniop, 194 .l_iop = cam_iosched_bw_iop, 195 .l_iodone= NULL, 196 }, 197 }; 198 199 struct iop_stats { 200 /* 201 * sysctl state for this subnode. 202 */ 203 struct sysctl_ctx_list sysctl_ctx; 204 struct sysctl_oid *sysctl_tree; 205 206 /* 207 * Information about the current rate limiters, if any 208 */ 209 io_limiter limiter; /* How are I/Os being limited */ 210 int min; /* Low range of limit */ 211 int max; /* High range of limit */ 212 int current; /* Current rate limiter */ 213 int l_value1; /* per-limiter scratch value 1. */ 214 int l_value2; /* per-limiter scratch value 2. */ 215 216 /* 217 * Debug information about counts of I/Os that have gone through the 218 * scheduler. 219 */ 220 int pending; /* I/Os pending in the hardware */ 221 int queued; /* number currently in the queue */ 222 int total; /* Total for all time -- wraps */ 223 int in; /* number queued all time -- wraps */ 224 int out; /* number completed all time -- wraps */ 225 int errs; /* Number of I/Os completed with error -- wraps */ 226 227 /* 228 * Statistics on different bits of the process. 229 */ 230 /* Exp Moving Average, see alpha_bits for more details */ 231 sbintime_t ema; 232 sbintime_t emvar; 233 sbintime_t sd; /* Last computed sd */ 234 235 uint32_t state_flags; 236 #define IOP_RATE_LIMITED 1u 237 238 #define LAT_BUCKETS 15 /* < 1ms < 2ms ... < 2^(n-1)ms >= 2^(n-1)ms*/ 239 uint64_t latencies[LAT_BUCKETS]; 240 241 struct cam_iosched_softc *softc; 242 }; 243 244 245 typedef enum { 246 set_max = 0, /* current = max */ 247 read_latency, /* Steer read latency by throttling writes */ 248 cl_max /* Keep last */ 249 } control_type; 250 251 static const char *cam_iosched_control_type_names[] = 252 { "set_max", "read_latency" }; 253 254 struct control_loop { 255 /* 256 * sysctl state for this subnode. 257 */ 258 struct sysctl_ctx_list sysctl_ctx; 259 struct sysctl_oid *sysctl_tree; 260 261 sbintime_t next_steer; /* Time of next steer */ 262 sbintime_t steer_interval; /* How often do we steer? */ 263 sbintime_t lolat; 264 sbintime_t hilat; 265 int alpha; 266 control_type type; /* What type of control? */ 267 int last_count; /* Last I/O count */ 268 269 struct cam_iosched_softc *softc; 270 }; 271 272 #endif 273 274 struct cam_iosched_softc { 275 struct bio_queue_head bio_queue; 276 struct bio_queue_head trim_queue; 277 /* scheduler flags < 16, user flags >= 16 */ 278 uint32_t flags; 279 int sort_io_queue; 280 int trim_goal; /* # of trims to queue before sending */ 281 int trim_ticks; /* Max ticks to hold trims */ 282 int last_trim_tick; /* Last 'tick' time ld a trim */ 283 int queued_trims; /* Number of trims in the queue */ 284 #ifdef CAM_IOSCHED_DYNAMIC 285 int read_bias; /* Read bias setting */ 286 int current_read_bias; /* Current read bias state */ 287 int total_ticks; 288 int load; /* EMA of 'load average' of disk / 2^16 */ 289 290 struct bio_queue_head write_queue; 291 struct iop_stats read_stats, write_stats, trim_stats; 292 struct sysctl_ctx_list sysctl_ctx; 293 struct sysctl_oid *sysctl_tree; 294 295 int quanta; /* Number of quanta per second */ 296 struct callout ticker; /* Callout for our quota system */ 297 struct cam_periph *periph; /* cam periph associated with this device */ 298 uint32_t this_frac; /* Fraction of a second (1024ths) for this tick */ 299 sbintime_t last_time; /* Last time we ticked */ 300 struct control_loop cl; 301 sbintime_t max_lat; /* when != 0, if iop latency > max_lat, call max_lat_fcn */ 302 cam_iosched_latfcn_t latfcn; 303 void *latarg; 304 #endif 305 }; 306 307 #ifdef CAM_IOSCHED_DYNAMIC 308 /* 309 * helper functions to call the limsw functions. 310 */ 311 static int 312 cam_iosched_limiter_init(struct iop_stats *ios) 313 { 314 int lim = ios->limiter; 315 316 /* maybe this should be a kassert */ 317 if (lim < none || lim >= limiter_max) 318 return EINVAL; 319 320 if (limsw[lim].l_init) 321 return limsw[lim].l_init(ios); 322 323 return 0; 324 } 325 326 static int 327 cam_iosched_limiter_tick(struct iop_stats *ios) 328 { 329 int lim = ios->limiter; 330 331 /* maybe this should be a kassert */ 332 if (lim < none || lim >= limiter_max) 333 return EINVAL; 334 335 if (limsw[lim].l_tick) 336 return limsw[lim].l_tick(ios); 337 338 return 0; 339 } 340 341 static int 342 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp) 343 { 344 int lim = ios->limiter; 345 346 /* maybe this should be a kassert */ 347 if (lim < none || lim >= limiter_max) 348 return EINVAL; 349 350 if (limsw[lim].l_iop) 351 return limsw[lim].l_iop(ios, bp); 352 353 return 0; 354 } 355 356 static int 357 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp) 358 { 359 int lim = ios->limiter; 360 361 /* maybe this should be a kassert */ 362 if (lim < none || lim >= limiter_max) 363 return EINVAL; 364 365 if (limsw[lim].l_caniop) 366 return limsw[lim].l_caniop(ios, bp); 367 368 return 0; 369 } 370 371 static int 372 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp) 373 { 374 int lim = ios->limiter; 375 376 /* maybe this should be a kassert */ 377 if (lim < none || lim >= limiter_max) 378 return 0; 379 380 if (limsw[lim].l_iodone) 381 return limsw[lim].l_iodone(ios, bp); 382 383 return 0; 384 } 385 386 /* 387 * Functions to implement the different kinds of limiters 388 */ 389 390 static int 391 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp) 392 { 393 394 if (ios->current <= 0 || ios->pending < ios->current) 395 return 0; 396 397 return EAGAIN; 398 } 399 400 static int 401 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp) 402 { 403 404 if (ios->current <= 0 || ios->pending < ios->current) 405 return 0; 406 407 return EAGAIN; 408 } 409 410 static int 411 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp) 412 { 413 414 if (ios->current <= 0 || ios->pending != ios->current) 415 return 0; 416 417 return 1; 418 } 419 420 static int 421 cam_iosched_iops_init(struct iop_stats *ios) 422 { 423 424 ios->l_value1 = ios->current / ios->softc->quanta; 425 if (ios->l_value1 <= 0) 426 ios->l_value1 = 1; 427 ios->l_value2 = 0; 428 429 return 0; 430 } 431 432 static int 433 cam_iosched_iops_tick(struct iop_stats *ios) 434 { 435 int new_ios; 436 437 /* 438 * Allow at least one IO per tick until all 439 * the IOs for this interval have been spent. 440 */ 441 new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16); 442 if (new_ios < 1 && ios->l_value2 < ios->current) { 443 new_ios = 1; 444 ios->l_value2++; 445 } 446 447 /* 448 * If this a new accounting interval, discard any "unspent" ios 449 * granted in the previous interval. Otherwise add the new ios to 450 * the previously granted ones that haven't been spent yet. 451 */ 452 if ((ios->softc->total_ticks % ios->softc->quanta) == 0) { 453 ios->l_value1 = new_ios; 454 ios->l_value2 = 1; 455 } else { 456 ios->l_value1 += new_ios; 457 } 458 459 460 return 0; 461 } 462 463 static int 464 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp) 465 { 466 467 /* 468 * So if we have any more IOPs left, allow it, 469 * otherwise wait. If current iops is 0, treat that 470 * as unlimited as a failsafe. 471 */ 472 if (ios->current > 0 && ios->l_value1 <= 0) 473 return EAGAIN; 474 return 0; 475 } 476 477 static int 478 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp) 479 { 480 int rv; 481 482 rv = cam_iosched_limiter_caniop(ios, bp); 483 if (rv == 0) 484 ios->l_value1--; 485 486 return rv; 487 } 488 489 static int 490 cam_iosched_bw_init(struct iop_stats *ios) 491 { 492 493 /* ios->current is in kB/s, so scale to bytes */ 494 ios->l_value1 = ios->current * 1000 / ios->softc->quanta; 495 496 return 0; 497 } 498 499 static int 500 cam_iosched_bw_tick(struct iop_stats *ios) 501 { 502 int bw; 503 504 /* 505 * If we're in the hole for available quota from 506 * the last time, then add the quantum for this. 507 * If we have any left over from last quantum, 508 * then too bad, that's lost. Also, ios->current 509 * is in kB/s, so scale. 510 * 511 * We also allow up to 4 quanta of credits to 512 * accumulate to deal with burstiness. 4 is extremely 513 * arbitrary. 514 */ 515 bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16); 516 if (ios->l_value1 < bw * 4) 517 ios->l_value1 += bw; 518 519 return 0; 520 } 521 522 static int 523 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp) 524 { 525 /* 526 * So if we have any more bw quota left, allow it, 527 * otherwise wait. Note, we'll go negative and that's 528 * OK. We'll just get a little less next quota. 529 * 530 * Note on going negative: that allows us to process 531 * requests in order better, since we won't allow 532 * shorter reads to get around the long one that we 533 * don't have the quota to do just yet. It also prevents 534 * starvation by being a little more permissive about 535 * what we let through this quantum (to prevent the 536 * starvation), at the cost of getting a little less 537 * next quantum. 538 * 539 * Also note that if the current limit is <= 0, 540 * we treat it as unlimited as a failsafe. 541 */ 542 if (ios->current > 0 && ios->l_value1 <= 0) 543 return EAGAIN; 544 545 546 return 0; 547 } 548 549 static int 550 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp) 551 { 552 int rv; 553 554 rv = cam_iosched_limiter_caniop(ios, bp); 555 if (rv == 0) 556 ios->l_value1 -= bp->bio_length; 557 558 return rv; 559 } 560 561 static void cam_iosched_cl_maybe_steer(struct control_loop *clp); 562 563 static void 564 cam_iosched_ticker(void *arg) 565 { 566 struct cam_iosched_softc *isc = arg; 567 sbintime_t now, delta; 568 int pending; 569 570 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc); 571 572 now = sbinuptime(); 573 delta = now - isc->last_time; 574 isc->this_frac = (uint32_t)delta >> 16; /* Note: discards seconds -- should be 0 harmless if not */ 575 isc->last_time = now; 576 577 cam_iosched_cl_maybe_steer(&isc->cl); 578 579 cam_iosched_limiter_tick(&isc->read_stats); 580 cam_iosched_limiter_tick(&isc->write_stats); 581 cam_iosched_limiter_tick(&isc->trim_stats); 582 583 cam_iosched_schedule(isc, isc->periph); 584 585 /* 586 * isc->load is an EMA of the pending I/Os at each tick. The number of 587 * pending I/Os is the sum of the I/Os queued to the hardware, and those 588 * in the software queue that could be queued to the hardware if there 589 * were slots. 590 * 591 * ios_stats.pending is a count of requests in the SIM right now for 592 * each of these types of I/O. So the total pending count is the sum of 593 * these I/Os and the sum of the queued I/Os still in the software queue 594 * for those operations that aren't being rate limited at the moment. 595 * 596 * The reason for the rate limiting bit is because those I/Os 597 * aren't part of the software queued load (since we could 598 * give them to hardware, but choose not to). 599 * 600 * Note: due to a bug in counting pending TRIM in the device, we 601 * don't include them in this count. We count each BIO_DELETE in 602 * the pending count, but the periph drivers collapse them down 603 * into one TRIM command. That one trim command gets the completion 604 * so the counts get off. 605 */ 606 pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */; 607 pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued + 608 !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* + 609 !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ; 610 pending <<= 16; 611 pending /= isc->periph->path->device->ccbq.total_openings; 612 613 isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */ 614 615 isc->total_ticks++; 616 } 617 618 619 static void 620 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc) 621 { 622 623 clp->next_steer = sbinuptime(); 624 clp->softc = isc; 625 clp->steer_interval = SBT_1S * 5; /* Let's start out steering every 5s */ 626 clp->lolat = 5 * SBT_1MS; 627 clp->hilat = 15 * SBT_1MS; 628 clp->alpha = 20; /* Alpha == gain. 20 = .2 */ 629 clp->type = set_max; 630 } 631 632 static void 633 cam_iosched_cl_maybe_steer(struct control_loop *clp) 634 { 635 struct cam_iosched_softc *isc; 636 sbintime_t now, lat; 637 int old; 638 639 isc = clp->softc; 640 now = isc->last_time; 641 if (now < clp->next_steer) 642 return; 643 644 clp->next_steer = now + clp->steer_interval; 645 switch (clp->type) { 646 case set_max: 647 if (isc->write_stats.current != isc->write_stats.max) 648 printf("Steering write from %d kBps to %d kBps\n", 649 isc->write_stats.current, isc->write_stats.max); 650 isc->read_stats.current = isc->read_stats.max; 651 isc->write_stats.current = isc->write_stats.max; 652 isc->trim_stats.current = isc->trim_stats.max; 653 break; 654 case read_latency: 655 old = isc->write_stats.current; 656 lat = isc->read_stats.ema; 657 /* 658 * Simple PLL-like engine. Since we're steering to a range for 659 * the SP (set point) that makes things a little more 660 * complicated. In addition, we're not directly controlling our 661 * PV (process variable), the read latency, but instead are 662 * manipulating the write bandwidth limit for our MV 663 * (manipulation variable), analysis of this code gets a bit 664 * messy. Also, the MV is a very noisy control surface for read 665 * latency since it is affected by many hidden processes inside 666 * the device which change how responsive read latency will be 667 * in reaction to changes in write bandwidth. Unlike the classic 668 * boiler control PLL. this may result in over-steering while 669 * the SSD takes its time to react to the new, lower load. This 670 * is why we use a relatively low alpha of between .1 and .25 to 671 * compensate for this effect. At .1, it takes ~22 steering 672 * intervals to back off by a factor of 10. At .2 it only takes 673 * ~10. At .25 it only takes ~8. However some preliminary data 674 * from the SSD drives suggests a reasponse time in 10's of 675 * seconds before latency drops regardless of the new write 676 * rate. Careful observation will be required to tune this 677 * effectively. 678 * 679 * Also, when there's no read traffic, we jack up the write 680 * limit too regardless of the last read latency. 10 is 681 * somewhat arbitrary. 682 */ 683 if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10) 684 isc->write_stats.current = isc->write_stats.current * 685 (100 + clp->alpha) / 100; /* Scale up */ 686 else if (lat > clp->hilat) 687 isc->write_stats.current = isc->write_stats.current * 688 (100 - clp->alpha) / 100; /* Scale down */ 689 clp->last_count = isc->read_stats.total; 690 691 /* 692 * Even if we don't steer, per se, enforce the min/max limits as 693 * those may have changed. 694 */ 695 if (isc->write_stats.current < isc->write_stats.min) 696 isc->write_stats.current = isc->write_stats.min; 697 if (isc->write_stats.current > isc->write_stats.max) 698 isc->write_stats.current = isc->write_stats.max; 699 if (old != isc->write_stats.current && iosched_debug) 700 printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n", 701 old, isc->write_stats.current, 702 (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32); 703 break; 704 case cl_max: 705 break; 706 } 707 } 708 #endif 709 710 /* 711 * Trim or similar currently pending completion. Should only be set for 712 * those drivers wishing only one Trim active at a time. 713 */ 714 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE (1ul << 0) 715 /* Callout active, and needs to be torn down */ 716 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1) 717 718 /* Periph drivers set these flags to indicate work */ 719 #define CAM_IOSCHED_FLAG_WORK_FLAGS ((0xffffu) << 16) 720 721 #ifdef CAM_IOSCHED_DYNAMIC 722 static void 723 cam_iosched_io_metric_update(struct cam_iosched_softc *isc, 724 sbintime_t sim_latency, int cmd, size_t size); 725 #endif 726 727 static inline bool 728 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc) 729 { 730 return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS); 731 } 732 733 static inline bool 734 cam_iosched_has_io(struct cam_iosched_softc *isc) 735 { 736 #ifdef CAM_IOSCHED_DYNAMIC 737 if (do_dynamic_iosched) { 738 struct bio *rbp = bioq_first(&isc->bio_queue); 739 struct bio *wbp = bioq_first(&isc->write_queue); 740 bool can_write = wbp != NULL && 741 cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0; 742 bool can_read = rbp != NULL && 743 cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0; 744 if (iosched_debug > 2) { 745 printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max); 746 printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max); 747 printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued); 748 } 749 return can_read || can_write; 750 } 751 #endif 752 return bioq_first(&isc->bio_queue) != NULL; 753 } 754 755 static inline bool 756 cam_iosched_has_more_trim(struct cam_iosched_softc *isc) 757 { 758 struct bio *bp; 759 760 bp = bioq_first(&isc->trim_queue); 761 #ifdef CAM_IOSCHED_DYNAMIC 762 if (do_dynamic_iosched) { 763 /* 764 * If we're limiting trims, then defer action on trims 765 * for a bit. 766 */ 767 if (bp == NULL || cam_iosched_limiter_caniop(&isc->trim_stats, bp) != 0) 768 return false; 769 } 770 #endif 771 772 /* 773 * If we've set a trim_goal, then if we exceed that allow trims 774 * to be passed back to the driver. If we've also set a tick timeout 775 * allow trims back to the driver. Otherwise, don't allow trims yet. 776 */ 777 if (isc->trim_goal > 0) { 778 if (isc->queued_trims >= isc->trim_goal) 779 return true; 780 if (isc->queued_trims > 0 && 781 isc->trim_ticks > 0 && 782 ticks - isc->last_trim_tick > isc->trim_ticks) 783 return true; 784 return false; 785 } 786 787 /* NB: Should perhaps have a max trim active independent of I/O limiters */ 788 return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) && bp != NULL; 789 } 790 791 #define cam_iosched_sort_queue(isc) ((isc)->sort_io_queue >= 0 ? \ 792 (isc)->sort_io_queue : cam_sort_io_queues) 793 794 795 static inline bool 796 cam_iosched_has_work(struct cam_iosched_softc *isc) 797 { 798 #ifdef CAM_IOSCHED_DYNAMIC 799 if (iosched_debug > 2) 800 printf("has work: %d %d %d\n", cam_iosched_has_io(isc), 801 cam_iosched_has_more_trim(isc), 802 cam_iosched_has_flagged_work(isc)); 803 #endif 804 805 return cam_iosched_has_io(isc) || 806 cam_iosched_has_more_trim(isc) || 807 cam_iosched_has_flagged_work(isc); 808 } 809 810 #ifdef CAM_IOSCHED_DYNAMIC 811 static void 812 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios) 813 { 814 815 ios->limiter = none; 816 ios->in = 0; 817 ios->max = ios->current = 300000; 818 ios->min = 1; 819 ios->out = 0; 820 ios->errs = 0; 821 ios->pending = 0; 822 ios->queued = 0; 823 ios->total = 0; 824 ios->ema = 0; 825 ios->emvar = 0; 826 ios->softc = isc; 827 cam_iosched_limiter_init(ios); 828 } 829 830 static int 831 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS) 832 { 833 char buf[16]; 834 struct iop_stats *ios; 835 struct cam_iosched_softc *isc; 836 int value, i, error; 837 const char *p; 838 839 ios = arg1; 840 isc = ios->softc; 841 value = ios->limiter; 842 if (value < none || value >= limiter_max) 843 p = "UNKNOWN"; 844 else 845 p = cam_iosched_limiter_names[value]; 846 847 strlcpy(buf, p, sizeof(buf)); 848 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 849 if (error != 0 || req->newptr == NULL) 850 return error; 851 852 cam_periph_lock(isc->periph); 853 854 for (i = none; i < limiter_max; i++) { 855 if (strcmp(buf, cam_iosched_limiter_names[i]) != 0) 856 continue; 857 ios->limiter = i; 858 error = cam_iosched_limiter_init(ios); 859 if (error != 0) { 860 ios->limiter = value; 861 cam_periph_unlock(isc->periph); 862 return error; 863 } 864 /* Note: disk load averate requires ticker to be always running */ 865 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc); 866 isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE; 867 868 cam_periph_unlock(isc->periph); 869 return 0; 870 } 871 872 cam_periph_unlock(isc->periph); 873 return EINVAL; 874 } 875 876 static int 877 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS) 878 { 879 char buf[16]; 880 struct control_loop *clp; 881 struct cam_iosched_softc *isc; 882 int value, i, error; 883 const char *p; 884 885 clp = arg1; 886 isc = clp->softc; 887 value = clp->type; 888 if (value < none || value >= cl_max) 889 p = "UNKNOWN"; 890 else 891 p = cam_iosched_control_type_names[value]; 892 893 strlcpy(buf, p, sizeof(buf)); 894 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 895 if (error != 0 || req->newptr == NULL) 896 return error; 897 898 for (i = set_max; i < cl_max; i++) { 899 if (strcmp(buf, cam_iosched_control_type_names[i]) != 0) 900 continue; 901 cam_periph_lock(isc->periph); 902 clp->type = i; 903 cam_periph_unlock(isc->periph); 904 return 0; 905 } 906 907 return EINVAL; 908 } 909 910 static int 911 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS) 912 { 913 char buf[16]; 914 sbintime_t value; 915 int error; 916 uint64_t us; 917 918 value = *(sbintime_t *)arg1; 919 us = (uint64_t)value / SBT_1US; 920 snprintf(buf, sizeof(buf), "%ju", (intmax_t)us); 921 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 922 if (error != 0 || req->newptr == NULL) 923 return error; 924 us = strtoul(buf, NULL, 10); 925 if (us == 0) 926 return EINVAL; 927 *(sbintime_t *)arg1 = us * SBT_1US; 928 return 0; 929 } 930 931 static int 932 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS) 933 { 934 int i, error; 935 struct sbuf sb; 936 uint64_t *latencies; 937 938 latencies = arg1; 939 sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req); 940 941 for (i = 0; i < LAT_BUCKETS - 1; i++) 942 sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]); 943 sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]); 944 error = sbuf_finish(&sb); 945 sbuf_delete(&sb); 946 947 return (error); 948 } 949 950 static int 951 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS) 952 { 953 int *quanta; 954 int error, value; 955 956 quanta = (unsigned *)arg1; 957 value = *quanta; 958 959 error = sysctl_handle_int(oidp, (int *)&value, 0, req); 960 if ((error != 0) || (req->newptr == NULL)) 961 return (error); 962 963 if (value < 1 || value > hz) 964 return (EINVAL); 965 966 *quanta = value; 967 968 return (0); 969 } 970 971 static void 972 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name) 973 { 974 struct sysctl_oid_list *n; 975 struct sysctl_ctx_list *ctx; 976 977 ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx, 978 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name, 979 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, name); 980 n = SYSCTL_CHILDREN(ios->sysctl_tree); 981 ctx = &ios->sysctl_ctx; 982 983 SYSCTL_ADD_UQUAD(ctx, n, 984 OID_AUTO, "ema", CTLFLAG_RD, 985 &ios->ema, 986 "Fast Exponentially Weighted Moving Average"); 987 SYSCTL_ADD_UQUAD(ctx, n, 988 OID_AUTO, "emvar", CTLFLAG_RD, 989 &ios->emvar, 990 "Fast Exponentially Weighted Moving Variance"); 991 992 SYSCTL_ADD_INT(ctx, n, 993 OID_AUTO, "pending", CTLFLAG_RD, 994 &ios->pending, 0, 995 "Instantaneous # of pending transactions"); 996 SYSCTL_ADD_INT(ctx, n, 997 OID_AUTO, "count", CTLFLAG_RD, 998 &ios->total, 0, 999 "# of transactions submitted to hardware"); 1000 SYSCTL_ADD_INT(ctx, n, 1001 OID_AUTO, "queued", CTLFLAG_RD, 1002 &ios->queued, 0, 1003 "# of transactions in the queue"); 1004 SYSCTL_ADD_INT(ctx, n, 1005 OID_AUTO, "in", CTLFLAG_RD, 1006 &ios->in, 0, 1007 "# of transactions queued to driver"); 1008 SYSCTL_ADD_INT(ctx, n, 1009 OID_AUTO, "out", CTLFLAG_RD, 1010 &ios->out, 0, 1011 "# of transactions completed (including with error)"); 1012 SYSCTL_ADD_INT(ctx, n, 1013 OID_AUTO, "errs", CTLFLAG_RD, 1014 &ios->errs, 0, 1015 "# of transactions completed with an error"); 1016 1017 SYSCTL_ADD_PROC(ctx, n, 1018 OID_AUTO, "limiter", 1019 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 1020 ios, 0, cam_iosched_limiter_sysctl, "A", 1021 "Current limiting type."); 1022 SYSCTL_ADD_INT(ctx, n, 1023 OID_AUTO, "min", CTLFLAG_RW, 1024 &ios->min, 0, 1025 "min resource"); 1026 SYSCTL_ADD_INT(ctx, n, 1027 OID_AUTO, "max", CTLFLAG_RW, 1028 &ios->max, 0, 1029 "max resource"); 1030 SYSCTL_ADD_INT(ctx, n, 1031 OID_AUTO, "current", CTLFLAG_RW, 1032 &ios->current, 0, 1033 "current resource"); 1034 1035 SYSCTL_ADD_PROC(ctx, n, 1036 OID_AUTO, "latencies", 1037 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT, 1038 &ios->latencies, 0, 1039 cam_iosched_sysctl_latencies, "A", 1040 "Array of power of 2 latency from 1ms to 1.024s"); 1041 } 1042 1043 static void 1044 cam_iosched_iop_stats_fini(struct iop_stats *ios) 1045 { 1046 if (ios->sysctl_tree) 1047 if (sysctl_ctx_free(&ios->sysctl_ctx) != 0) 1048 printf("can't remove iosched sysctl stats context\n"); 1049 } 1050 1051 static void 1052 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc) 1053 { 1054 struct sysctl_oid_list *n; 1055 struct sysctl_ctx_list *ctx; 1056 struct control_loop *clp; 1057 1058 clp = &isc->cl; 1059 clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx, 1060 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control", 1061 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Control loop info"); 1062 n = SYSCTL_CHILDREN(clp->sysctl_tree); 1063 ctx = &clp->sysctl_ctx; 1064 1065 SYSCTL_ADD_PROC(ctx, n, 1066 OID_AUTO, "type", 1067 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 1068 clp, 0, cam_iosched_control_type_sysctl, "A", 1069 "Control loop algorithm"); 1070 SYSCTL_ADD_PROC(ctx, n, 1071 OID_AUTO, "steer_interval", 1072 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 1073 &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A", 1074 "How often to steer (in us)"); 1075 SYSCTL_ADD_PROC(ctx, n, 1076 OID_AUTO, "lolat", 1077 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 1078 &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A", 1079 "Low water mark for Latency (in us)"); 1080 SYSCTL_ADD_PROC(ctx, n, 1081 OID_AUTO, "hilat", 1082 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 1083 &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A", 1084 "Hi water mark for Latency (in us)"); 1085 SYSCTL_ADD_INT(ctx, n, 1086 OID_AUTO, "alpha", CTLFLAG_RW, 1087 &clp->alpha, 0, 1088 "Alpha for PLL (x100) aka gain"); 1089 } 1090 1091 static void 1092 cam_iosched_cl_sysctl_fini(struct control_loop *clp) 1093 { 1094 if (clp->sysctl_tree) 1095 if (sysctl_ctx_free(&clp->sysctl_ctx) != 0) 1096 printf("can't remove iosched sysctl control loop context\n"); 1097 } 1098 #endif 1099 1100 /* 1101 * Allocate the iosched structure. This also insulates callers from knowing 1102 * sizeof struct cam_iosched_softc. 1103 */ 1104 int 1105 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph) 1106 { 1107 1108 *iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO); 1109 if (*iscp == NULL) 1110 return ENOMEM; 1111 #ifdef CAM_IOSCHED_DYNAMIC 1112 if (iosched_debug) 1113 printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp); 1114 #endif 1115 (*iscp)->sort_io_queue = -1; 1116 bioq_init(&(*iscp)->bio_queue); 1117 bioq_init(&(*iscp)->trim_queue); 1118 #ifdef CAM_IOSCHED_DYNAMIC 1119 if (do_dynamic_iosched) { 1120 bioq_init(&(*iscp)->write_queue); 1121 (*iscp)->read_bias = 100; 1122 (*iscp)->current_read_bias = 100; 1123 (*iscp)->quanta = min(hz, 200); 1124 cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats); 1125 cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats); 1126 cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats); 1127 (*iscp)->trim_stats.max = 1; /* Trims are special: one at a time for now */ 1128 (*iscp)->last_time = sbinuptime(); 1129 callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0); 1130 (*iscp)->periph = periph; 1131 cam_iosched_cl_init(&(*iscp)->cl, *iscp); 1132 callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp); 1133 (*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE; 1134 } 1135 #endif 1136 1137 return 0; 1138 } 1139 1140 /* 1141 * Reclaim all used resources. This assumes that other folks have 1142 * drained the requests in the hardware. Maybe an unwise assumption. 1143 */ 1144 void 1145 cam_iosched_fini(struct cam_iosched_softc *isc) 1146 { 1147 if (isc) { 1148 cam_iosched_flush(isc, NULL, ENXIO); 1149 #ifdef CAM_IOSCHED_DYNAMIC 1150 cam_iosched_iop_stats_fini(&isc->read_stats); 1151 cam_iosched_iop_stats_fini(&isc->write_stats); 1152 cam_iosched_iop_stats_fini(&isc->trim_stats); 1153 cam_iosched_cl_sysctl_fini(&isc->cl); 1154 if (isc->sysctl_tree) 1155 if (sysctl_ctx_free(&isc->sysctl_ctx) != 0) 1156 printf("can't remove iosched sysctl stats context\n"); 1157 if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) { 1158 callout_drain(&isc->ticker); 1159 isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE; 1160 } 1161 #endif 1162 free(isc, M_CAMSCHED); 1163 } 1164 } 1165 1166 /* 1167 * After we're sure we're attaching a device, go ahead and add 1168 * hooks for any sysctl we may wish to honor. 1169 */ 1170 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc, 1171 struct sysctl_ctx_list *ctx, struct sysctl_oid *node) 1172 { 1173 struct sysctl_oid_list *n; 1174 1175 n = SYSCTL_CHILDREN(node); 1176 SYSCTL_ADD_INT(ctx, n, 1177 OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE, 1178 &isc->sort_io_queue, 0, 1179 "Sort IO queue to try and optimise disk access patterns"); 1180 SYSCTL_ADD_INT(ctx, n, 1181 OID_AUTO, "trim_goal", CTLFLAG_RW, 1182 &isc->trim_goal, 0, 1183 "Number of trims to try to accumulate before sending to hardware"); 1184 SYSCTL_ADD_INT(ctx, n, 1185 OID_AUTO, "trim_ticks", CTLFLAG_RW, 1186 &isc->trim_goal, 0, 1187 "IO Schedul qaunta to hold back trims for when accumulating"); 1188 1189 #ifdef CAM_IOSCHED_DYNAMIC 1190 if (!do_dynamic_iosched) 1191 return; 1192 1193 isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx, 1194 SYSCTL_CHILDREN(node), OID_AUTO, "iosched", 1195 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "I/O scheduler statistics"); 1196 n = SYSCTL_CHILDREN(isc->sysctl_tree); 1197 ctx = &isc->sysctl_ctx; 1198 1199 cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read"); 1200 cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write"); 1201 cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim"); 1202 cam_iosched_cl_sysctl_init(isc); 1203 1204 SYSCTL_ADD_INT(ctx, n, 1205 OID_AUTO, "read_bias", CTLFLAG_RW, 1206 &isc->read_bias, 100, 1207 "How biased towards read should we be independent of limits"); 1208 1209 SYSCTL_ADD_PROC(ctx, n, 1210 OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_NEEDGIANT, 1211 &isc->quanta, 0, cam_iosched_quanta_sysctl, "I", 1212 "How many quanta per second do we slice the I/O up into"); 1213 1214 SYSCTL_ADD_INT(ctx, n, 1215 OID_AUTO, "total_ticks", CTLFLAG_RD, 1216 &isc->total_ticks, 0, 1217 "Total number of ticks we've done"); 1218 1219 SYSCTL_ADD_INT(ctx, n, 1220 OID_AUTO, "load", CTLFLAG_RD, 1221 &isc->load, 0, 1222 "scaled load average / 100"); 1223 1224 SYSCTL_ADD_U64(ctx, n, 1225 OID_AUTO, "latency_trigger", CTLFLAG_RW, 1226 &isc->max_lat, 0, 1227 "Latency treshold to trigger callbacks"); 1228 #endif 1229 } 1230 1231 void 1232 cam_iosched_set_latfcn(struct cam_iosched_softc *isc, 1233 cam_iosched_latfcn_t fnp, void *argp) 1234 { 1235 #ifdef CAM_IOSCHED_DYNAMIC 1236 isc->latfcn = fnp; 1237 isc->latarg = argp; 1238 #endif 1239 } 1240 1241 /* 1242 * Client drivers can set two parameters. "goal" is the number of BIO_DELETEs 1243 * that will be queued up before iosched will "release" the trims to the client 1244 * driver to wo with what they will (usually combine as many as possible). If we 1245 * don't get this many, after trim_ticks we'll submit the I/O anyway with 1246 * whatever we have. We do need an I/O of some kind of to clock the deferred 1247 * trims out to disk. Since we will eventually get a write for the super block 1248 * or something before we shutdown, the trims will complete. To be safe, when a 1249 * BIO_FLUSH is presented to the iosched work queue, we set the ticks time far 1250 * enough in the past so we'll present the BIO_DELETEs to the client driver. 1251 * There might be a race if no BIO_DELETESs were queued, a BIO_FLUSH comes in 1252 * and then a BIO_DELETE is sent down. No know client does this, and there's 1253 * already a race between an ordered BIO_FLUSH and any BIO_DELETEs in flight, 1254 * but no client depends on the ordering being honored. 1255 * 1256 * XXX I'm not sure what the interaction between UFS direct BIOs and the BUF 1257 * flushing on shutdown. I think there's bufs that would be dependent on the BIO 1258 * finishing to write out at least metadata, so we'll be fine. To be safe, keep 1259 * the number of ticks low (less than maybe 10s) to avoid shutdown races. 1260 */ 1261 1262 void 1263 cam_iosched_set_trim_goal(struct cam_iosched_softc *isc, int goal) 1264 { 1265 1266 isc->trim_goal = goal; 1267 } 1268 1269 void 1270 cam_iosched_set_trim_ticks(struct cam_iosched_softc *isc, int trim_ticks) 1271 { 1272 1273 isc->trim_ticks = trim_ticks; 1274 } 1275 1276 /* 1277 * Flush outstanding I/O. Consumers of this library don't know all the 1278 * queues we may keep, so this allows all I/O to be flushed in one 1279 * convenient call. 1280 */ 1281 void 1282 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err) 1283 { 1284 bioq_flush(&isc->bio_queue, stp, err); 1285 bioq_flush(&isc->trim_queue, stp, err); 1286 #ifdef CAM_IOSCHED_DYNAMIC 1287 if (do_dynamic_iosched) 1288 bioq_flush(&isc->write_queue, stp, err); 1289 #endif 1290 } 1291 1292 #ifdef CAM_IOSCHED_DYNAMIC 1293 static struct bio * 1294 cam_iosched_get_write(struct cam_iosched_softc *isc) 1295 { 1296 struct bio *bp; 1297 1298 /* 1299 * We control the write rate by controlling how many requests we send 1300 * down to the drive at any one time. Fewer requests limits the 1301 * effects of both starvation when the requests take a while and write 1302 * amplification when each request is causing more than one write to 1303 * the NAND media. Limiting the queue depth like this will also limit 1304 * the write throughput and give and reads that want to compete to 1305 * compete unfairly. 1306 */ 1307 bp = bioq_first(&isc->write_queue); 1308 if (bp == NULL) { 1309 if (iosched_debug > 3) 1310 printf("No writes present in write_queue\n"); 1311 return NULL; 1312 } 1313 1314 /* 1315 * If pending read, prefer that based on current read bias 1316 * setting. 1317 */ 1318 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) { 1319 if (iosched_debug) 1320 printf( 1321 "Reads present and current_read_bias is %d queued " 1322 "writes %d queued reads %d\n", 1323 isc->current_read_bias, isc->write_stats.queued, 1324 isc->read_stats.queued); 1325 isc->current_read_bias--; 1326 /* We're not limiting writes, per se, just doing reads first */ 1327 return NULL; 1328 } 1329 1330 /* 1331 * See if our current limiter allows this I/O. 1332 */ 1333 if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) { 1334 if (iosched_debug) 1335 printf("Can't write because limiter says no.\n"); 1336 isc->write_stats.state_flags |= IOP_RATE_LIMITED; 1337 return NULL; 1338 } 1339 1340 /* 1341 * Let's do this: We've passed all the gates and we're a go 1342 * to schedule the I/O in the SIM. 1343 */ 1344 isc->current_read_bias = isc->read_bias; 1345 bioq_remove(&isc->write_queue, bp); 1346 if (bp->bio_cmd == BIO_WRITE) { 1347 isc->write_stats.queued--; 1348 isc->write_stats.total++; 1349 isc->write_stats.pending++; 1350 } 1351 if (iosched_debug > 9) 1352 printf("HWQ : %p %#x\n", bp, bp->bio_cmd); 1353 isc->write_stats.state_flags &= ~IOP_RATE_LIMITED; 1354 return bp; 1355 } 1356 #endif 1357 1358 /* 1359 * Put back a trim that you weren't able to actually schedule this time. 1360 */ 1361 void 1362 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp) 1363 { 1364 bioq_insert_head(&isc->trim_queue, bp); 1365 if (isc->queued_trims == 0) 1366 isc->last_trim_tick = ticks; 1367 isc->queued_trims++; 1368 #ifdef CAM_IOSCHED_DYNAMIC 1369 isc->trim_stats.queued++; 1370 isc->trim_stats.total--; /* since we put it back, don't double count */ 1371 isc->trim_stats.pending--; 1372 #endif 1373 } 1374 1375 /* 1376 * gets the next trim from the trim queue. 1377 * 1378 * Assumes we're called with the periph lock held. It removes this 1379 * trim from the queue and the device must explicitly reinsert it 1380 * should the need arise. 1381 */ 1382 struct bio * 1383 cam_iosched_next_trim(struct cam_iosched_softc *isc) 1384 { 1385 struct bio *bp; 1386 1387 bp = bioq_first(&isc->trim_queue); 1388 if (bp == NULL) 1389 return NULL; 1390 bioq_remove(&isc->trim_queue, bp); 1391 isc->queued_trims--; 1392 isc->last_trim_tick = ticks; /* Reset the tick timer when we take trims */ 1393 #ifdef CAM_IOSCHED_DYNAMIC 1394 isc->trim_stats.queued--; 1395 isc->trim_stats.total++; 1396 isc->trim_stats.pending++; 1397 #endif 1398 return bp; 1399 } 1400 1401 /* 1402 * gets an available trim from the trim queue, if there's no trim 1403 * already pending. It removes this trim from the queue and the device 1404 * must explicitly reinsert it should the need arise. 1405 * 1406 * Assumes we're called with the periph lock held. 1407 */ 1408 struct bio * 1409 cam_iosched_get_trim(struct cam_iosched_softc *isc) 1410 { 1411 #ifdef CAM_IOSCHED_DYNAMIC 1412 struct bio *bp; 1413 #endif 1414 1415 if (!cam_iosched_has_more_trim(isc)) 1416 return NULL; 1417 #ifdef CAM_IOSCHED_DYNAMIC 1418 bp = bioq_first(&isc->trim_queue); 1419 if (bp == NULL) 1420 return NULL; 1421 1422 /* 1423 * If pending read, prefer that based on current read bias setting. The 1424 * read bias is shared for both writes and TRIMs, but on TRIMs the bias 1425 * is for a combined TRIM not a single TRIM request that's come in. 1426 */ 1427 if (do_dynamic_iosched) { 1428 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) { 1429 if (iosched_debug) 1430 printf("Reads present and current_read_bias is %d" 1431 " queued trims %d queued reads %d\n", 1432 isc->current_read_bias, isc->trim_stats.queued, 1433 isc->read_stats.queued); 1434 isc->current_read_bias--; 1435 /* We're not limiting TRIMS, per se, just doing reads first */ 1436 return NULL; 1437 } 1438 /* 1439 * We're going to do a trim, so reset the bias. 1440 */ 1441 isc->current_read_bias = isc->read_bias; 1442 } 1443 1444 /* 1445 * See if our current limiter allows this I/O. Because we only call this 1446 * here, and not in next_trim, the 'bandwidth' limits for trims won't 1447 * work, while the iops or max queued limits will work. It's tricky 1448 * because we want the limits to be from the perspective of the 1449 * "commands sent to the device." To make iops work, we need to check 1450 * only here (since we want all the ops we combine to count as one). To 1451 * make bw limits work, we'd need to check in next_trim, but that would 1452 * have the effect of limiting the iops as seen from the upper layers. 1453 */ 1454 if (cam_iosched_limiter_iop(&isc->trim_stats, bp) != 0) { 1455 if (iosched_debug) 1456 printf("Can't trim because limiter says no.\n"); 1457 isc->trim_stats.state_flags |= IOP_RATE_LIMITED; 1458 return NULL; 1459 } 1460 isc->current_read_bias = isc->read_bias; 1461 isc->trim_stats.state_flags &= ~IOP_RATE_LIMITED; 1462 /* cam_iosched_next_trim below keeps proper book */ 1463 #endif 1464 return cam_iosched_next_trim(isc); 1465 } 1466 1467 /* 1468 * Determine what the next bit of work to do is for the periph. The 1469 * default implementation looks to see if we have trims to do, but no 1470 * trims outstanding. If so, we do that. Otherwise we see if we have 1471 * other work. If we do, then we do that. Otherwise why were we called? 1472 */ 1473 struct bio * 1474 cam_iosched_next_bio(struct cam_iosched_softc *isc) 1475 { 1476 struct bio *bp; 1477 1478 /* 1479 * See if we have a trim that can be scheduled. We can only send one 1480 * at a time down, so this takes that into account. 1481 * 1482 * XXX newer TRIM commands are queueable. Revisit this when we 1483 * implement them. 1484 */ 1485 if ((bp = cam_iosched_get_trim(isc)) != NULL) 1486 return bp; 1487 1488 #ifdef CAM_IOSCHED_DYNAMIC 1489 /* 1490 * See if we have any pending writes, and room in the queue for them, 1491 * and if so, those are next. 1492 */ 1493 if (do_dynamic_iosched) { 1494 if ((bp = cam_iosched_get_write(isc)) != NULL) 1495 return bp; 1496 } 1497 #endif 1498 1499 /* 1500 * next, see if there's other, normal I/O waiting. If so return that. 1501 */ 1502 if ((bp = bioq_first(&isc->bio_queue)) == NULL) 1503 return NULL; 1504 1505 #ifdef CAM_IOSCHED_DYNAMIC 1506 /* 1507 * For the dynamic scheduler, bio_queue is only for reads, so enforce 1508 * the limits here. Enforce only for reads. 1509 */ 1510 if (do_dynamic_iosched) { 1511 if (bp->bio_cmd == BIO_READ && 1512 cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) { 1513 isc->read_stats.state_flags |= IOP_RATE_LIMITED; 1514 return NULL; 1515 } 1516 } 1517 isc->read_stats.state_flags &= ~IOP_RATE_LIMITED; 1518 #endif 1519 bioq_remove(&isc->bio_queue, bp); 1520 #ifdef CAM_IOSCHED_DYNAMIC 1521 if (do_dynamic_iosched) { 1522 if (bp->bio_cmd == BIO_READ) { 1523 isc->read_stats.queued--; 1524 isc->read_stats.total++; 1525 isc->read_stats.pending++; 1526 } else 1527 printf("Found bio_cmd = %#x\n", bp->bio_cmd); 1528 } 1529 if (iosched_debug > 9) 1530 printf("HWQ : %p %#x\n", bp, bp->bio_cmd); 1531 #endif 1532 return bp; 1533 } 1534 1535 /* 1536 * Driver has been given some work to do by the block layer. Tell the 1537 * scheduler about it and have it queue the work up. The scheduler module 1538 * will then return the currently most useful bit of work later, possibly 1539 * deferring work for various reasons. 1540 */ 1541 void 1542 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp) 1543 { 1544 1545 /* 1546 * A BIO_SPEEDUP from the uppper layers means that they have a block 1547 * shortage. At the present, this is only sent when we're trying to 1548 * allocate blocks, but have a shortage before giving up. bio_length is 1549 * the size of their shortage. We will complete just enough BIO_DELETEs 1550 * in the queue to satisfy the need. If bio_length is 0, we'll complete 1551 * them all. This allows the scheduler to delay BIO_DELETEs to improve 1552 * read/write performance without worrying about the upper layers. When 1553 * it's possibly a problem, we respond by pretending the BIO_DELETEs 1554 * just worked. We can't do anything about the BIO_DELETEs in the 1555 * hardware, though. We have to wait for them to complete. 1556 */ 1557 if (bp->bio_cmd == BIO_SPEEDUP) { 1558 off_t len; 1559 struct bio *nbp; 1560 1561 len = 0; 1562 while (bioq_first(&isc->trim_queue) && 1563 (bp->bio_length == 0 || len < bp->bio_length)) { 1564 nbp = bioq_takefirst(&isc->trim_queue); 1565 len += nbp->bio_length; 1566 nbp->bio_error = 0; 1567 biodone(nbp); 1568 } 1569 if (bp->bio_length > 0) { 1570 if (bp->bio_length > len) 1571 bp->bio_resid = bp->bio_length - len; 1572 else 1573 bp->bio_resid = 0; 1574 } 1575 bp->bio_error = 0; 1576 biodone(bp); 1577 return; 1578 } 1579 1580 /* 1581 * If we get a BIO_FLUSH, and we're doing delayed BIO_DELETEs then we 1582 * set the last tick time to one less than the current ticks minus the 1583 * delay to force the BIO_DELETEs to be presented to the client driver. 1584 */ 1585 if (bp->bio_cmd == BIO_FLUSH && isc->trim_ticks > 0) 1586 isc->last_trim_tick = ticks - isc->trim_ticks - 1; 1587 1588 /* 1589 * Put all trims on the trim queue. Otherwise put the work on the bio 1590 * queue. 1591 */ 1592 if (bp->bio_cmd == BIO_DELETE) { 1593 bioq_insert_tail(&isc->trim_queue, bp); 1594 if (isc->queued_trims == 0) 1595 isc->last_trim_tick = ticks; 1596 isc->queued_trims++; 1597 #ifdef CAM_IOSCHED_DYNAMIC 1598 isc->trim_stats.in++; 1599 isc->trim_stats.queued++; 1600 #endif 1601 } 1602 #ifdef CAM_IOSCHED_DYNAMIC 1603 else if (do_dynamic_iosched && (bp->bio_cmd != BIO_READ)) { 1604 if (cam_iosched_sort_queue(isc)) 1605 bioq_disksort(&isc->write_queue, bp); 1606 else 1607 bioq_insert_tail(&isc->write_queue, bp); 1608 if (iosched_debug > 9) 1609 printf("Qw : %p %#x\n", bp, bp->bio_cmd); 1610 if (bp->bio_cmd == BIO_WRITE) { 1611 isc->write_stats.in++; 1612 isc->write_stats.queued++; 1613 } 1614 } 1615 #endif 1616 else { 1617 if (cam_iosched_sort_queue(isc)) 1618 bioq_disksort(&isc->bio_queue, bp); 1619 else 1620 bioq_insert_tail(&isc->bio_queue, bp); 1621 #ifdef CAM_IOSCHED_DYNAMIC 1622 if (iosched_debug > 9) 1623 printf("Qr : %p %#x\n", bp, bp->bio_cmd); 1624 if (bp->bio_cmd == BIO_READ) { 1625 isc->read_stats.in++; 1626 isc->read_stats.queued++; 1627 } else if (bp->bio_cmd == BIO_WRITE) { 1628 isc->write_stats.in++; 1629 isc->write_stats.queued++; 1630 } 1631 #endif 1632 } 1633 } 1634 1635 /* 1636 * If we have work, get it scheduled. Called with the periph lock held. 1637 */ 1638 void 1639 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph) 1640 { 1641 1642 if (cam_iosched_has_work(isc)) 1643 xpt_schedule(periph, CAM_PRIORITY_NORMAL); 1644 } 1645 1646 /* 1647 * Complete a trim request. Mark that we no longer have one in flight. 1648 */ 1649 void 1650 cam_iosched_trim_done(struct cam_iosched_softc *isc) 1651 { 1652 1653 isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE; 1654 } 1655 1656 /* 1657 * Complete a bio. Called before we release the ccb with xpt_release_ccb so we 1658 * might use notes in the ccb for statistics. 1659 */ 1660 int 1661 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp, 1662 union ccb *done_ccb) 1663 { 1664 int retval = 0; 1665 #ifdef CAM_IOSCHED_DYNAMIC 1666 if (!do_dynamic_iosched) 1667 return retval; 1668 1669 if (iosched_debug > 10) 1670 printf("done: %p %#x\n", bp, bp->bio_cmd); 1671 if (bp->bio_cmd == BIO_WRITE) { 1672 retval = cam_iosched_limiter_iodone(&isc->write_stats, bp); 1673 if ((bp->bio_flags & BIO_ERROR) != 0) 1674 isc->write_stats.errs++; 1675 isc->write_stats.out++; 1676 isc->write_stats.pending--; 1677 } else if (bp->bio_cmd == BIO_READ) { 1678 retval = cam_iosched_limiter_iodone(&isc->read_stats, bp); 1679 if ((bp->bio_flags & BIO_ERROR) != 0) 1680 isc->read_stats.errs++; 1681 isc->read_stats.out++; 1682 isc->read_stats.pending--; 1683 } else if (bp->bio_cmd == BIO_DELETE) { 1684 if ((bp->bio_flags & BIO_ERROR) != 0) 1685 isc->trim_stats.errs++; 1686 isc->trim_stats.out++; 1687 isc->trim_stats.pending--; 1688 } else if (bp->bio_cmd != BIO_FLUSH) { 1689 if (iosched_debug) 1690 printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd); 1691 } 1692 1693 if (!(bp->bio_flags & BIO_ERROR) && done_ccb != NULL) { 1694 sbintime_t sim_latency; 1695 1696 sim_latency = cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data); 1697 1698 cam_iosched_io_metric_update(isc, sim_latency, 1699 bp->bio_cmd, bp->bio_bcount); 1700 /* 1701 * Debugging code: allow callbacks to the periph driver when latency max 1702 * is exceeded. This can be useful for triggering external debugging actions. 1703 */ 1704 if (isc->latfcn && isc->max_lat != 0 && sim_latency > isc->max_lat) 1705 isc->latfcn(isc->latarg, sim_latency, bp); 1706 } 1707 1708 #endif 1709 return retval; 1710 } 1711 1712 /* 1713 * Tell the io scheduler that you've pushed a trim down into the sim. 1714 * This also tells the I/O scheduler not to push any more trims down, so 1715 * some periphs do not call it if they can cope with multiple trims in flight. 1716 */ 1717 void 1718 cam_iosched_submit_trim(struct cam_iosched_softc *isc) 1719 { 1720 1721 isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE; 1722 } 1723 1724 /* 1725 * Change the sorting policy hint for I/O transactions for this device. 1726 */ 1727 void 1728 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val) 1729 { 1730 1731 isc->sort_io_queue = val; 1732 } 1733 1734 int 1735 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags) 1736 { 1737 return isc->flags & flags; 1738 } 1739 1740 void 1741 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags) 1742 { 1743 isc->flags |= flags; 1744 } 1745 1746 void 1747 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags) 1748 { 1749 isc->flags &= ~flags; 1750 } 1751 1752 #ifdef CAM_IOSCHED_DYNAMIC 1753 /* 1754 * After the method presented in Jack Crenshaw's 1998 article "Integer 1755 * Square Roots," reprinted at 1756 * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots 1757 * and well worth the read. Briefly, we find the power of 4 that's the 1758 * largest smaller than val. We then check each smaller power of 4 to 1759 * see if val is still bigger. The right shifts at each step divide 1760 * the result by 2 which after successive application winds up 1761 * accumulating the right answer. It could also have been accumulated 1762 * using a separate root counter, but this code is smaller and faster 1763 * than that method. This method is also integer size invariant. 1764 * It returns floor(sqrt((float)val)), or the largest integer less than 1765 * or equal to the square root. 1766 */ 1767 static uint64_t 1768 isqrt64(uint64_t val) 1769 { 1770 uint64_t res = 0; 1771 uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2); 1772 1773 /* 1774 * Find the largest power of 4 smaller than val. 1775 */ 1776 while (bit > val) 1777 bit >>= 2; 1778 1779 /* 1780 * Accumulate the answer, one bit at a time (we keep moving 1781 * them over since 2 is the square root of 4 and we test 1782 * powers of 4). We accumulate where we find the bit, but 1783 * the successive shifts land the bit in the right place 1784 * by the end. 1785 */ 1786 while (bit != 0) { 1787 if (val >= res + bit) { 1788 val -= res + bit; 1789 res = (res >> 1) + bit; 1790 } else 1791 res >>= 1; 1792 bit >>= 2; 1793 } 1794 1795 return res; 1796 } 1797 1798 static sbintime_t latencies[LAT_BUCKETS - 1] = { 1799 SBT_1MS << 0, 1800 SBT_1MS << 1, 1801 SBT_1MS << 2, 1802 SBT_1MS << 3, 1803 SBT_1MS << 4, 1804 SBT_1MS << 5, 1805 SBT_1MS << 6, 1806 SBT_1MS << 7, 1807 SBT_1MS << 8, 1808 SBT_1MS << 9, 1809 SBT_1MS << 10, 1810 SBT_1MS << 11, 1811 SBT_1MS << 12, 1812 SBT_1MS << 13 /* 8.192s */ 1813 }; 1814 1815 static void 1816 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency) 1817 { 1818 sbintime_t y, deltasq, delta; 1819 int i; 1820 1821 /* 1822 * Keep counts for latency. We do it by power of two buckets. 1823 * This helps us spot outlier behavior obscured by averages. 1824 */ 1825 for (i = 0; i < LAT_BUCKETS - 1; i++) { 1826 if (sim_latency < latencies[i]) { 1827 iop->latencies[i]++; 1828 break; 1829 } 1830 } 1831 if (i == LAT_BUCKETS - 1) 1832 iop->latencies[i]++; /* Put all > 1024ms values into the last bucket. */ 1833 1834 /* 1835 * Classic exponentially decaying average with a tiny alpha 1836 * (2 ^ -alpha_bits). For more info see the NIST statistical 1837 * handbook. 1838 * 1839 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha) [nist] 1840 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1 1841 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1 1842 * alpha = 1 / (1 << alpha_bits) 1843 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1 1844 * = y_t/b - e/b + be/b 1845 * = (y_t - e + be) / b 1846 * = (e + d) / b 1847 * 1848 * Since alpha is a power of two, we can compute this w/o any mult or 1849 * division. 1850 * 1851 * Variance can also be computed. Usually, it would be expressed as follows: 1852 * diff_t = y_t - ema_t-1 1853 * emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha) 1854 * = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2 1855 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2 1856 * = e - e/b + dd/b + dd/bb 1857 * = (bbe - be + bdd + dd) / bb 1858 * = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits)) 1859 */ 1860 /* 1861 * XXX possible numeric issues 1862 * o We assume right shifted integers do the right thing, since that's 1863 * implementation defined. You can change the right shifts to / (1LL << alpha). 1864 * o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits 1865 * for emvar. This puts a ceiling of 13 bits on alpha since we need a 1866 * few tens of seconds of representation. 1867 * o We mitigate alpha issues by never setting it too high. 1868 */ 1869 y = sim_latency; 1870 delta = (y - iop->ema); /* d */ 1871 iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits; 1872 1873 /* 1874 * Were we to naively plow ahead at this point, we wind up with many numerical 1875 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves 1876 * us with microsecond level precision in the input, so the same in the 1877 * output. It means we can't overflow deltasq unless delta > 4k seconds. It 1878 * also means that emvar can be up 46 bits 40 of which are fraction, which 1879 * gives us a way to measure up to ~8s in the SD before the computation goes 1880 * unstable. Even the worst hard disk rarely has > 1s service time in the 1881 * drive. It does mean we have to shift left 12 bits after taking the 1882 * square root to compute the actual standard deviation estimate. This loss of 1883 * precision is preferable to needing int128 types to work. The above numbers 1884 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12, 1885 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases. 1886 */ 1887 delta >>= 12; 1888 deltasq = delta * delta; /* dd */ 1889 iop->emvar = ((iop->emvar << (2 * alpha_bits)) + /* bbe */ 1890 ((deltasq - iop->emvar) << alpha_bits) + /* b(dd-e) */ 1891 deltasq) /* dd */ 1892 >> (2 * alpha_bits); /* div bb */ 1893 iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12; 1894 } 1895 1896 static void 1897 cam_iosched_io_metric_update(struct cam_iosched_softc *isc, 1898 sbintime_t sim_latency, int cmd, size_t size) 1899 { 1900 /* xxx Do we need to scale based on the size of the I/O ? */ 1901 switch (cmd) { 1902 case BIO_READ: 1903 cam_iosched_update(&isc->read_stats, sim_latency); 1904 break; 1905 case BIO_WRITE: 1906 cam_iosched_update(&isc->write_stats, sim_latency); 1907 break; 1908 case BIO_DELETE: 1909 cam_iosched_update(&isc->trim_stats, sim_latency); 1910 break; 1911 default: 1912 break; 1913 } 1914 } 1915 1916 #ifdef DDB 1917 static int biolen(struct bio_queue_head *bq) 1918 { 1919 int i = 0; 1920 struct bio *bp; 1921 1922 TAILQ_FOREACH(bp, &bq->queue, bio_queue) { 1923 i++; 1924 } 1925 return i; 1926 } 1927 1928 /* 1929 * Show the internal state of the I/O scheduler. 1930 */ 1931 DB_SHOW_COMMAND(iosched, cam_iosched_db_show) 1932 { 1933 struct cam_iosched_softc *isc; 1934 1935 if (!have_addr) { 1936 db_printf("Need addr\n"); 1937 return; 1938 } 1939 isc = (struct cam_iosched_softc *)addr; 1940 db_printf("pending_reads: %d\n", isc->read_stats.pending); 1941 db_printf("min_reads: %d\n", isc->read_stats.min); 1942 db_printf("max_reads: %d\n", isc->read_stats.max); 1943 db_printf("reads: %d\n", isc->read_stats.total); 1944 db_printf("in_reads: %d\n", isc->read_stats.in); 1945 db_printf("out_reads: %d\n", isc->read_stats.out); 1946 db_printf("queued_reads: %d\n", isc->read_stats.queued); 1947 db_printf("Read Q len %d\n", biolen(&isc->bio_queue)); 1948 db_printf("pending_writes: %d\n", isc->write_stats.pending); 1949 db_printf("min_writes: %d\n", isc->write_stats.min); 1950 db_printf("max_writes: %d\n", isc->write_stats.max); 1951 db_printf("writes: %d\n", isc->write_stats.total); 1952 db_printf("in_writes: %d\n", isc->write_stats.in); 1953 db_printf("out_writes: %d\n", isc->write_stats.out); 1954 db_printf("queued_writes: %d\n", isc->write_stats.queued); 1955 db_printf("Write Q len %d\n", biolen(&isc->write_queue)); 1956 db_printf("pending_trims: %d\n", isc->trim_stats.pending); 1957 db_printf("min_trims: %d\n", isc->trim_stats.min); 1958 db_printf("max_trims: %d\n", isc->trim_stats.max); 1959 db_printf("trims: %d\n", isc->trim_stats.total); 1960 db_printf("in_trims: %d\n", isc->trim_stats.in); 1961 db_printf("out_trims: %d\n", isc->trim_stats.out); 1962 db_printf("queued_trims: %d\n", isc->trim_stats.queued); 1963 db_printf("Trim Q len %d\n", biolen(&isc->trim_queue)); 1964 db_printf("read_bias: %d\n", isc->read_bias); 1965 db_printf("current_read_bias: %d\n", isc->current_read_bias); 1966 db_printf("Trim active? %s\n", 1967 (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no"); 1968 } 1969 #endif 1970 #endif 1971