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