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