1 /*- 2 * CAM IO Scheduler Interface 3 * 4 * Copyright (c) 2015 Netflix, Inc. 5 * All rights reserved. 6 * 7 * Redistribution and use in source and binary forms, with or without 8 * modification, are permitted provided that the following conditions 9 * are met: 10 * 1. Redistributions of source code must retain the above copyright 11 * notice, this list of conditions, and the following disclaimer, 12 * without modification, immediately at the beginning of the file. 13 * 2. The name of the author may not be used to endorse or promote products 14 * derived from this software without specific prior written permission. 15 * 16 * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND 17 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE 18 * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE 19 * ARE DISCLAIMED. IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE FOR 20 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL 21 * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS 22 * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) 23 * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT 24 * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY 25 * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF 26 * SUCH DAMAGE. 27 * 28 * $FreeBSD$ 29 */ 30 31 #include "opt_cam.h" 32 #include "opt_ddb.h" 33 34 #include <sys/cdefs.h> 35 __FBSDID("$FreeBSD$"); 36 37 #include <sys/param.h> 38 39 #include <sys/systm.h> 40 #include <sys/kernel.h> 41 #include <sys/bio.h> 42 #include <sys/lock.h> 43 #include <sys/malloc.h> 44 #include <sys/mutex.h> 45 #include <sys/sysctl.h> 46 47 #include <cam/cam.h> 48 #include <cam/cam_ccb.h> 49 #include <cam/cam_periph.h> 50 #include <cam/cam_xpt_periph.h> 51 #include <cam/cam_iosched.h> 52 53 #include <ddb/ddb.h> 54 55 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler", 56 "CAM I/O Scheduler buffers"); 57 58 /* 59 * Default I/O scheduler for FreeBSD. This implementation is just a thin-vineer 60 * over the bioq_* interface, with notions of separate calls for normal I/O and 61 * for trims. 62 */ 63 64 #ifdef CAM_NETFLIX_IOSCHED 65 66 static int do_netflix_iosched = 1; 67 TUNABLE_INT("kern.cam.do_netflix_iosched", &do_netflix_iosched); 68 SYSCTL_INT(_kern_cam, OID_AUTO, do_netflix_iosched, CTLFLAG_RD, 69 &do_netflix_iosched, 1, 70 "Enable Netflix I/O scheduler optimizations."); 71 72 static int alpha_bits = 9; 73 TUNABLE_INT("kern.cam.iosched_alpha_bits", &alpha_bits); 74 SYSCTL_INT(_kern_cam, OID_AUTO, iosched_alpha_bits, CTLFLAG_RW, 75 &alpha_bits, 1, 76 "Bits in EMA's alpha."); 77 78 79 80 struct iop_stats; 81 struct cam_iosched_softc; 82 83 int iosched_debug = 0; 84 85 typedef enum { 86 none = 0, /* No limits */ 87 queue_depth, /* Limit how many ops we queue to SIM */ 88 iops, /* Limit # of IOPS to the drive */ 89 bandwidth, /* Limit bandwidth to the drive */ 90 limiter_max 91 } io_limiter; 92 93 static const char *cam_iosched_limiter_names[] = 94 { "none", "queue_depth", "iops", "bandwidth" }; 95 96 /* 97 * Called to initialize the bits of the iop_stats structure relevant to the 98 * limiter. Called just after the limiter is set. 99 */ 100 typedef int l_init_t(struct iop_stats *); 101 102 /* 103 * Called every tick. 104 */ 105 typedef int l_tick_t(struct iop_stats *); 106 107 /* 108 * Called to see if the limiter thinks this IOP can be allowed to 109 * proceed. If so, the limiter assumes that the while IOP proceeded 110 * and makes any accounting of it that's needed. 111 */ 112 typedef int l_iop_t(struct iop_stats *, struct bio *); 113 114 /* 115 * Called when an I/O completes so the limiter can updates its 116 * accounting. Pending I/Os may complete in any order (even when 117 * sent to the hardware at the same time), so the limiter may not 118 * make any assumptions other than this I/O has completed. If it 119 * returns 1, then xpt_schedule() needs to be called again. 120 */ 121 typedef int l_iodone_t(struct iop_stats *, struct bio *); 122 123 static l_iop_t cam_iosched_qd_iop; 124 static l_iop_t cam_iosched_qd_caniop; 125 static l_iodone_t cam_iosched_qd_iodone; 126 127 static l_init_t cam_iosched_iops_init; 128 static l_tick_t cam_iosched_iops_tick; 129 static l_iop_t cam_iosched_iops_caniop; 130 static l_iop_t cam_iosched_iops_iop; 131 132 static l_init_t cam_iosched_bw_init; 133 static l_tick_t cam_iosched_bw_tick; 134 static l_iop_t cam_iosched_bw_caniop; 135 static l_iop_t cam_iosched_bw_iop; 136 137 struct limswitch 138 { 139 l_init_t *l_init; 140 l_tick_t *l_tick; 141 l_iop_t *l_iop; 142 l_iop_t *l_caniop; 143 l_iodone_t *l_iodone; 144 } limsw[] = 145 { 146 { /* none */ 147 .l_init = NULL, 148 .l_tick = NULL, 149 .l_iop = NULL, 150 .l_iodone= NULL, 151 }, 152 { /* queue_depth */ 153 .l_init = NULL, 154 .l_tick = NULL, 155 .l_caniop = cam_iosched_qd_caniop, 156 .l_iop = cam_iosched_qd_iop, 157 .l_iodone= cam_iosched_qd_iodone, 158 }, 159 { /* iops */ 160 .l_init = cam_iosched_iops_init, 161 .l_tick = cam_iosched_iops_tick, 162 .l_caniop = cam_iosched_iops_caniop, 163 .l_iop = cam_iosched_iops_iop, 164 .l_iodone= NULL, 165 }, 166 { /* bandwidth */ 167 .l_init = cam_iosched_bw_init, 168 .l_tick = cam_iosched_bw_tick, 169 .l_caniop = cam_iosched_bw_caniop, 170 .l_iop = cam_iosched_bw_iop, 171 .l_iodone= NULL, 172 }, 173 }; 174 175 struct iop_stats 176 { 177 /* 178 * sysctl state for this subnode. 179 */ 180 struct sysctl_ctx_list sysctl_ctx; 181 struct sysctl_oid *sysctl_tree; 182 183 /* 184 * Information about the current rate limiters, if any 185 */ 186 io_limiter limiter; /* How are I/Os being limited */ 187 int min; /* Low range of limit */ 188 int max; /* High range of limit */ 189 int current; /* Current rate limiter */ 190 int l_value1; /* per-limiter scratch value 1. */ 191 int l_value2; /* per-limiter scratch value 2. */ 192 193 194 /* 195 * Debug information about counts of I/Os that have gone through the 196 * scheduler. 197 */ 198 int pending; /* I/Os pending in the hardware */ 199 int queued; /* number currently in the queue */ 200 int total; /* Total for all time -- wraps */ 201 int in; /* number queued all time -- wraps */ 202 int out; /* number completed all time -- wraps */ 203 204 /* 205 * Statistics on different bits of the process. 206 */ 207 /* Exp Moving Average, alpha = 1 / (1 << alpha_bits) */ 208 sbintime_t ema; 209 sbintime_t emss; /* Exp Moving sum of the squares */ 210 sbintime_t sd; /* Last computed sd */ 211 212 struct cam_iosched_softc *softc; 213 }; 214 215 216 typedef enum { 217 set_max = 0, /* current = max */ 218 read_latency, /* Steer read latency by throttling writes */ 219 cl_max /* Keep last */ 220 } control_type; 221 222 static const char *cam_iosched_control_type_names[] = 223 { "set_max", "read_latency" }; 224 225 struct control_loop 226 { 227 /* 228 * sysctl state for this subnode. 229 */ 230 struct sysctl_ctx_list sysctl_ctx; 231 struct sysctl_oid *sysctl_tree; 232 233 sbintime_t next_steer; /* Time of next steer */ 234 sbintime_t steer_interval; /* How often do we steer? */ 235 sbintime_t lolat; 236 sbintime_t hilat; 237 int alpha; 238 control_type type; /* What type of control? */ 239 int last_count; /* Last I/O count */ 240 241 struct cam_iosched_softc *softc; 242 }; 243 244 #endif 245 246 struct cam_iosched_softc 247 { 248 struct bio_queue_head bio_queue; 249 struct bio_queue_head trim_queue; 250 /* scheduler flags < 16, user flags >= 16 */ 251 uint32_t flags; 252 int sort_io_queue; 253 #ifdef CAM_NETFLIX_IOSCHED 254 int read_bias; /* Read bias setting */ 255 int current_read_bias; /* Current read bias state */ 256 int total_ticks; 257 258 struct bio_queue_head write_queue; 259 struct iop_stats read_stats, write_stats, trim_stats; 260 struct sysctl_ctx_list sysctl_ctx; 261 struct sysctl_oid *sysctl_tree; 262 263 int quanta; /* Number of quanta per second */ 264 struct callout ticker; /* Callout for our quota system */ 265 struct cam_periph *periph; /* cam periph associated with this device */ 266 uint32_t this_frac; /* Fraction of a second (1024ths) for this tick */ 267 sbintime_t last_time; /* Last time we ticked */ 268 struct control_loop cl; 269 #endif 270 }; 271 272 #ifdef CAM_NETFLIX_IOSCHED 273 /* 274 * helper functions to call the limsw functions. 275 */ 276 static int 277 cam_iosched_limiter_init(struct iop_stats *ios) 278 { 279 int lim = ios->limiter; 280 281 /* maybe this should be a kassert */ 282 if (lim < none || lim >= limiter_max) 283 return EINVAL; 284 285 if (limsw[lim].l_init) 286 return limsw[lim].l_init(ios); 287 288 return 0; 289 } 290 291 static int 292 cam_iosched_limiter_tick(struct iop_stats *ios) 293 { 294 int lim = ios->limiter; 295 296 /* maybe this should be a kassert */ 297 if (lim < none || lim >= limiter_max) 298 return EINVAL; 299 300 if (limsw[lim].l_tick) 301 return limsw[lim].l_tick(ios); 302 303 return 0; 304 } 305 306 static int 307 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp) 308 { 309 int lim = ios->limiter; 310 311 /* maybe this should be a kassert */ 312 if (lim < none || lim >= limiter_max) 313 return EINVAL; 314 315 if (limsw[lim].l_iop) 316 return limsw[lim].l_iop(ios, bp); 317 318 return 0; 319 } 320 321 static int 322 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp) 323 { 324 int lim = ios->limiter; 325 326 /* maybe this should be a kassert */ 327 if (lim < none || lim >= limiter_max) 328 return EINVAL; 329 330 if (limsw[lim].l_caniop) 331 return limsw[lim].l_caniop(ios, bp); 332 333 return 0; 334 } 335 336 static int 337 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp) 338 { 339 int lim = ios->limiter; 340 341 /* maybe this should be a kassert */ 342 if (lim < none || lim >= limiter_max) 343 return 0; 344 345 if (limsw[lim].l_iodone) 346 return limsw[lim].l_iodone(ios, bp); 347 348 return 0; 349 } 350 351 /* 352 * Functions to implement the different kinds of limiters 353 */ 354 355 static int 356 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp) 357 { 358 359 if (ios->current <= 0 || ios->pending < ios->current) 360 return 0; 361 362 return EAGAIN; 363 } 364 365 static int 366 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp) 367 { 368 369 if (ios->current <= 0 || ios->pending < ios->current) 370 return 0; 371 372 return EAGAIN; 373 } 374 375 static int 376 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp) 377 { 378 379 if (ios->current <= 0 || ios->pending != ios->current) 380 return 0; 381 382 return 1; 383 } 384 385 static int 386 cam_iosched_iops_init(struct iop_stats *ios) 387 { 388 389 ios->l_value1 = ios->current / ios->softc->quanta; 390 if (ios->l_value1 <= 0) 391 ios->l_value1 = 1; 392 393 return 0; 394 } 395 396 static int 397 cam_iosched_iops_tick(struct iop_stats *ios) 398 { 399 400 ios->l_value1 = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16); 401 if (ios->l_value1 <= 0) 402 ios->l_value1 = 1; 403 404 return 0; 405 } 406 407 static int 408 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp) 409 { 410 411 /* 412 * So if we have any more IOPs left, allow it, 413 * otherwise wait. 414 */ 415 if (ios->l_value1 <= 0) 416 return EAGAIN; 417 return 0; 418 } 419 420 static int 421 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp) 422 { 423 int rv; 424 425 rv = cam_iosched_limiter_caniop(ios, bp); 426 if (rv == 0) 427 ios->l_value1--; 428 429 return rv; 430 } 431 432 static int 433 cam_iosched_bw_init(struct iop_stats *ios) 434 { 435 436 /* ios->current is in kB/s, so scale to bytes */ 437 ios->l_value1 = ios->current * 1000 / ios->softc->quanta; 438 439 return 0; 440 } 441 442 static int 443 cam_iosched_bw_tick(struct iop_stats *ios) 444 { 445 int bw; 446 447 /* 448 * If we're in the hole for available quota from 449 * the last time, then add the quantum for this. 450 * If we have any left over from last quantum, 451 * then too bad, that's lost. Also, ios->current 452 * is in kB/s, so scale. 453 * 454 * We also allow up to 4 quanta of credits to 455 * accumulate to deal with burstiness. 4 is extremely 456 * arbitrary. 457 */ 458 bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16); 459 if (ios->l_value1 < bw * 4) 460 ios->l_value1 += bw; 461 462 return 0; 463 } 464 465 static int 466 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp) 467 { 468 /* 469 * So if we have any more bw quota left, allow it, 470 * otherwise wait. Not, we'll go negative and that's 471 * OK. We'll just get a lettle less next quota. 472 * 473 * Note on going negative: that allows us to process 474 * requests in order better, since we won't allow 475 * shorter reads to get around the long one that we 476 * don't have the quota to do just yet. It also prevents 477 * starvation by being a little more permissive about 478 * what we let through this quantum (to prevent the 479 * starvation), at the cost of getting a little less 480 * next quantum. 481 */ 482 if (ios->l_value1 <= 0) 483 return EAGAIN; 484 485 486 return 0; 487 } 488 489 static int 490 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp) 491 { 492 int rv; 493 494 rv = cam_iosched_limiter_caniop(ios, bp); 495 if (rv == 0) 496 ios->l_value1 -= bp->bio_length; 497 498 return rv; 499 } 500 501 static void cam_iosched_cl_maybe_steer(struct control_loop *clp); 502 503 static void 504 cam_iosched_ticker(void *arg) 505 { 506 struct cam_iosched_softc *isc = arg; 507 sbintime_t now, delta; 508 509 callout_reset(&isc->ticker, hz / isc->quanta - 1, cam_iosched_ticker, isc); 510 511 now = sbinuptime(); 512 delta = now - isc->last_time; 513 isc->this_frac = (uint32_t)delta >> 16; /* Note: discards seconds -- should be 0 harmless if not */ 514 isc->last_time = now; 515 516 cam_iosched_cl_maybe_steer(&isc->cl); 517 518 cam_iosched_limiter_tick(&isc->read_stats); 519 cam_iosched_limiter_tick(&isc->write_stats); 520 cam_iosched_limiter_tick(&isc->trim_stats); 521 522 cam_iosched_schedule(isc, isc->periph); 523 524 isc->total_ticks++; 525 } 526 527 528 static void 529 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc) 530 { 531 532 clp->next_steer = sbinuptime(); 533 clp->softc = isc; 534 clp->steer_interval = SBT_1S * 5; /* Let's start out steering every 5s */ 535 clp->lolat = 5 * SBT_1MS; 536 clp->hilat = 15 * SBT_1MS; 537 clp->alpha = 20; /* Alpha == gain. 20 = .2 */ 538 clp->type = set_max; 539 } 540 541 static void 542 cam_iosched_cl_maybe_steer(struct control_loop *clp) 543 { 544 struct cam_iosched_softc *isc; 545 sbintime_t now, lat; 546 int old; 547 548 isc = clp->softc; 549 now = isc->last_time; 550 if (now < clp->next_steer) 551 return; 552 553 clp->next_steer = now + clp->steer_interval; 554 switch (clp->type) { 555 case set_max: 556 if (isc->write_stats.current != isc->write_stats.max) 557 printf("Steering write from %d kBps to %d kBps\n", 558 isc->write_stats.current, isc->write_stats.max); 559 isc->read_stats.current = isc->read_stats.max; 560 isc->write_stats.current = isc->write_stats.max; 561 isc->trim_stats.current = isc->trim_stats.max; 562 break; 563 case read_latency: 564 old = isc->write_stats.current; 565 lat = isc->read_stats.ema; 566 /* 567 * Simple PLL-like engine. Since we're steering to a range for 568 * the SP (set point) that makes things a little more 569 * complicated. In addition, we're not directly controlling our 570 * PV (process variable), the read latency, but instead are 571 * manipulating the write bandwidth limit for our MV 572 * (manipulation variable), analysis of this code gets a bit 573 * messy. Also, the MV is a very noisy control surface for read 574 * latency since it is affected by many hidden processes inside 575 * the device which change how responsive read latency will be 576 * in reaction to changes in write bandwidth. Unlike the classic 577 * boiler control PLL. this may result in over-steering while 578 * the SSD takes its time to react to the new, lower load. This 579 * is why we use a relatively low alpha of between .1 and .25 to 580 * compensate for this effect. At .1, it takes ~22 steering 581 * intervals to back off by a factor of 10. At .2 it only takes 582 * ~10. At .25 it only takes ~8. However some preliminary data 583 * from the SSD drives suggests a reasponse time in 10's of 584 * seconds before latency drops regardless of the new write 585 * rate. Careful observation will be reqiured to tune this 586 * effectively. 587 * 588 * Also, when there's no read traffic, we jack up the write 589 * limit too regardless of the last read latency. 10 is 590 * somewhat arbitrary. 591 */ 592 if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10) 593 isc->write_stats.current = isc->write_stats.current * 594 (100 + clp->alpha) / 100; /* Scale up */ 595 else if (lat > clp->hilat) 596 isc->write_stats.current = isc->write_stats.current * 597 (100 - clp->alpha) / 100; /* Scale down */ 598 clp->last_count = isc->read_stats.total; 599 600 /* 601 * Even if we don't steer, per se, enforce the min/max limits as 602 * those may have changed. 603 */ 604 if (isc->write_stats.current < isc->write_stats.min) 605 isc->write_stats.current = isc->write_stats.min; 606 if (isc->write_stats.current > isc->write_stats.max) 607 isc->write_stats.current = isc->write_stats.max; 608 if (old != isc->write_stats.current && iosched_debug) 609 printf("Steering write from %d kBps to %d kBps due to latency of %jdms\n", 610 old, isc->write_stats.current, 611 (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32); 612 break; 613 case cl_max: 614 break; 615 } 616 } 617 #endif 618 619 /* Trim or similar currently pending completion */ 620 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE (1ul << 0) 621 /* Callout active, and needs to be torn down */ 622 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1) 623 624 /* Periph drivers set these flags to indicate work */ 625 #define CAM_IOSCHED_FLAG_WORK_FLAGS ((0xffffu) << 16) 626 627 #ifdef CAM_NETFLIX_IOSCHED 628 static void 629 cam_iosched_io_metric_update(struct cam_iosched_softc *isc, 630 sbintime_t sim_latency, int cmd, size_t size); 631 #endif 632 633 static inline int 634 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc) 635 { 636 return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS); 637 } 638 639 static inline int 640 cam_iosched_has_io(struct cam_iosched_softc *isc) 641 { 642 #ifdef CAM_NETFLIX_IOSCHED 643 if (do_netflix_iosched) { 644 struct bio *rbp = bioq_first(&isc->bio_queue); 645 struct bio *wbp = bioq_first(&isc->write_queue); 646 int can_write = wbp != NULL && 647 cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0; 648 int can_read = rbp != NULL && 649 cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0; 650 if (iosched_debug > 2) { 651 printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max); 652 printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max); 653 printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued); 654 } 655 return can_read || can_write; 656 } 657 #endif 658 return bioq_first(&isc->bio_queue) != NULL; 659 } 660 661 static inline int 662 cam_iosched_has_more_trim(struct cam_iosched_softc *isc) 663 { 664 return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) && 665 bioq_first(&isc->trim_queue); 666 } 667 668 #define cam_iosched_sort_queue(isc) ((isc)->sort_io_queue >= 0 ? \ 669 (isc)->sort_io_queue : cam_sort_io_queues) 670 671 672 static inline int 673 cam_iosched_has_work(struct cam_iosched_softc *isc) 674 { 675 #ifdef CAM_NETFLIX_IOSCHED 676 if (iosched_debug > 2) 677 printf("has work: %d %d %d\n", cam_iosched_has_io(isc), 678 cam_iosched_has_more_trim(isc), 679 cam_iosched_has_flagged_work(isc)); 680 #endif 681 682 return cam_iosched_has_io(isc) || 683 cam_iosched_has_more_trim(isc) || 684 cam_iosched_has_flagged_work(isc); 685 } 686 687 #ifdef CAM_NETFLIX_IOSCHED 688 static void 689 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios) 690 { 691 692 ios->limiter = none; 693 cam_iosched_limiter_init(ios); 694 ios->in = 0; 695 ios->max = 300000; 696 ios->min = 1; 697 ios->out = 0; 698 ios->pending = 0; 699 ios->queued = 0; 700 ios->total = 0; 701 ios->ema = 0; 702 ios->emss = 0; 703 ios->sd = 0; 704 ios->softc = isc; 705 } 706 707 static int 708 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS) 709 { 710 char buf[16]; 711 struct iop_stats *ios; 712 struct cam_iosched_softc *isc; 713 int value, i, error, cantick; 714 const char *p; 715 716 ios = arg1; 717 isc = ios->softc; 718 value = ios->limiter; 719 if (value < none || value >= limiter_max) 720 p = "UNKNOWN"; 721 else 722 p = cam_iosched_limiter_names[value]; 723 724 strlcpy(buf, p, sizeof(buf)); 725 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 726 if (error != 0 || req->newptr == NULL) 727 return error; 728 729 cam_periph_lock(isc->periph); 730 731 for (i = none; i < limiter_max; i++) { 732 if (strcmp(buf, cam_iosched_limiter_names[i]) != 0) 733 continue; 734 ios->limiter = i; 735 error = cam_iosched_limiter_init(ios); 736 if (error != 0) { 737 ios->limiter = value; 738 cam_periph_unlock(isc->periph); 739 return error; 740 } 741 cantick = !!limsw[isc->read_stats.limiter].l_tick + 742 !!limsw[isc->write_stats.limiter].l_tick + 743 !!limsw[isc->trim_stats.limiter].l_tick + 744 1; /* Control loop requires it */ 745 if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) { 746 if (cantick == 0) { 747 callout_stop(&isc->ticker); 748 isc->flags &= ~CAM_IOSCHED_FLAG_CALLOUT_ACTIVE; 749 } 750 } else { 751 if (cantick != 0) { 752 callout_reset(&isc->ticker, hz / isc->quanta - 1, cam_iosched_ticker, isc); 753 isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE; 754 } 755 } 756 757 cam_periph_unlock(isc->periph); 758 return 0; 759 } 760 761 cam_periph_unlock(isc->periph); 762 return EINVAL; 763 } 764 765 static int 766 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS) 767 { 768 char buf[16]; 769 struct control_loop *clp; 770 struct cam_iosched_softc *isc; 771 int value, i, error; 772 const char *p; 773 774 clp = arg1; 775 isc = clp->softc; 776 value = clp->type; 777 if (value < none || value >= cl_max) 778 p = "UNKNOWN"; 779 else 780 p = cam_iosched_control_type_names[value]; 781 782 strlcpy(buf, p, sizeof(buf)); 783 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 784 if (error != 0 || req->newptr == NULL) 785 return error; 786 787 for (i = set_max; i < cl_max; i++) { 788 if (strcmp(buf, cam_iosched_control_type_names[i]) != 0) 789 continue; 790 cam_periph_lock(isc->periph); 791 clp->type = i; 792 cam_periph_unlock(isc->periph); 793 return 0; 794 } 795 796 return EINVAL; 797 } 798 799 static int 800 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS) 801 { 802 char buf[16]; 803 sbintime_t value; 804 int error; 805 uint64_t us; 806 807 value = *(sbintime_t *)arg1; 808 us = (uint64_t)value / SBT_1US; 809 snprintf(buf, sizeof(buf), "%ju", (intmax_t)us); 810 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 811 if (error != 0 || req->newptr == NULL) 812 return error; 813 us = strtoul(buf, NULL, 10); 814 if (us == 0) 815 return EINVAL; 816 *(sbintime_t *)arg1 = us * SBT_1US; 817 return 0; 818 } 819 820 static void 821 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name) 822 { 823 struct sysctl_oid_list *n; 824 struct sysctl_ctx_list *ctx; 825 826 ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx, 827 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name, 828 CTLFLAG_RD, 0, name); 829 n = SYSCTL_CHILDREN(ios->sysctl_tree); 830 ctx = &ios->sysctl_ctx; 831 832 SYSCTL_ADD_UQUAD(ctx, n, 833 OID_AUTO, "ema", CTLFLAG_RD, 834 &ios->ema, 835 "Fast Exponentially Weighted Moving Average"); 836 SYSCTL_ADD_UQUAD(ctx, n, 837 OID_AUTO, "emss", CTLFLAG_RD, 838 &ios->emss, 839 "Fast Exponentially Weighted Moving Sum of Squares (maybe wrong)"); 840 SYSCTL_ADD_UQUAD(ctx, n, 841 OID_AUTO, "sd", CTLFLAG_RD, 842 &ios->sd, 843 "Estimated SD for fast ema (may be wrong)"); 844 845 SYSCTL_ADD_INT(ctx, n, 846 OID_AUTO, "pending", CTLFLAG_RD, 847 &ios->pending, 0, 848 "Instantaneous # of pending transactions"); 849 SYSCTL_ADD_INT(ctx, n, 850 OID_AUTO, "count", CTLFLAG_RD, 851 &ios->total, 0, 852 "# of transactions submitted to hardware"); 853 SYSCTL_ADD_INT(ctx, n, 854 OID_AUTO, "queued", CTLFLAG_RD, 855 &ios->queued, 0, 856 "# of transactions in the queue"); 857 SYSCTL_ADD_INT(ctx, n, 858 OID_AUTO, "in", CTLFLAG_RD, 859 &ios->in, 0, 860 "# of transactions queued to driver"); 861 SYSCTL_ADD_INT(ctx, n, 862 OID_AUTO, "out", CTLFLAG_RD, 863 &ios->out, 0, 864 "# of transactions completed"); 865 866 SYSCTL_ADD_PROC(ctx, n, 867 OID_AUTO, "limiter", CTLTYPE_STRING | CTLFLAG_RW, 868 ios, 0, cam_iosched_limiter_sysctl, "A", 869 "Current limiting type."); 870 SYSCTL_ADD_INT(ctx, n, 871 OID_AUTO, "min", CTLFLAG_RW, 872 &ios->min, 0, 873 "min resource"); 874 SYSCTL_ADD_INT(ctx, n, 875 OID_AUTO, "max", CTLFLAG_RW, 876 &ios->max, 0, 877 "max resource"); 878 SYSCTL_ADD_INT(ctx, n, 879 OID_AUTO, "current", CTLFLAG_RW, 880 &ios->current, 0, 881 "current resource"); 882 883 } 884 885 static void 886 cam_iosched_iop_stats_fini(struct iop_stats *ios) 887 { 888 if (ios->sysctl_tree) 889 if (sysctl_ctx_free(&ios->sysctl_ctx) != 0) 890 printf("can't remove iosched sysctl stats context\n"); 891 } 892 893 static void 894 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc) 895 { 896 struct sysctl_oid_list *n; 897 struct sysctl_ctx_list *ctx; 898 struct control_loop *clp; 899 900 clp = &isc->cl; 901 clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx, 902 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control", 903 CTLFLAG_RD, 0, "Control loop info"); 904 n = SYSCTL_CHILDREN(clp->sysctl_tree); 905 ctx = &clp->sysctl_ctx; 906 907 SYSCTL_ADD_PROC(ctx, n, 908 OID_AUTO, "type", CTLTYPE_STRING | CTLFLAG_RW, 909 clp, 0, cam_iosched_control_type_sysctl, "A", 910 "Control loop algorithm"); 911 SYSCTL_ADD_PROC(ctx, n, 912 OID_AUTO, "steer_interval", CTLTYPE_STRING | CTLFLAG_RW, 913 &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A", 914 "How often to steer (in us)"); 915 SYSCTL_ADD_PROC(ctx, n, 916 OID_AUTO, "lolat", CTLTYPE_STRING | CTLFLAG_RW, 917 &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A", 918 "Low water mark for Latency (in us)"); 919 SYSCTL_ADD_PROC(ctx, n, 920 OID_AUTO, "hilat", CTLTYPE_STRING | CTLFLAG_RW, 921 &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A", 922 "Hi water mark for Latency (in us)"); 923 SYSCTL_ADD_INT(ctx, n, 924 OID_AUTO, "alpha", CTLFLAG_RW, 925 &clp->alpha, 0, 926 "Alpha for PLL (x100) aka gain"); 927 } 928 929 static void 930 cam_iosched_cl_sysctl_fini(struct control_loop *clp) 931 { 932 if (clp->sysctl_tree) 933 if (sysctl_ctx_free(&clp->sysctl_ctx) != 0) 934 printf("can't remove iosched sysctl control loop context\n"); 935 } 936 #endif 937 938 /* 939 * Allocate the iosched structure. This also insulates callers from knowing 940 * sizeof struct cam_iosched_softc. 941 */ 942 int 943 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph) 944 { 945 946 *iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO); 947 if (*iscp == NULL) 948 return ENOMEM; 949 #ifdef CAM_NETFLIX_IOSCHED 950 if (iosched_debug) 951 printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp); 952 #endif 953 (*iscp)->sort_io_queue = -1; 954 bioq_init(&(*iscp)->bio_queue); 955 bioq_init(&(*iscp)->trim_queue); 956 #ifdef CAM_NETFLIX_IOSCHED 957 if (do_netflix_iosched) { 958 bioq_init(&(*iscp)->write_queue); 959 (*iscp)->read_bias = 100; 960 (*iscp)->current_read_bias = 100; 961 (*iscp)->quanta = 200; 962 cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats); 963 cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats); 964 cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats); 965 (*iscp)->trim_stats.max = 1; /* Trims are special: one at a time for now */ 966 (*iscp)->last_time = sbinuptime(); 967 callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0); 968 (*iscp)->periph = periph; 969 cam_iosched_cl_init(&(*iscp)->cl, *iscp); 970 callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta - 1, cam_iosched_ticker, *iscp); 971 (*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE; 972 } 973 #endif 974 975 return 0; 976 } 977 978 /* 979 * Reclaim all used resources. This assumes that other folks have 980 * drained the requests in the hardware. Maybe an unwise assumption. 981 */ 982 void 983 cam_iosched_fini(struct cam_iosched_softc *isc) 984 { 985 if (isc) { 986 cam_iosched_flush(isc, NULL, ENXIO); 987 #ifdef CAM_NETFLIX_IOSCHED 988 cam_iosched_iop_stats_fini(&isc->read_stats); 989 cam_iosched_iop_stats_fini(&isc->write_stats); 990 cam_iosched_iop_stats_fini(&isc->trim_stats); 991 cam_iosched_cl_sysctl_fini(&isc->cl); 992 if (isc->sysctl_tree) 993 if (sysctl_ctx_free(&isc->sysctl_ctx) != 0) 994 printf("can't remove iosched sysctl stats context\n"); 995 if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) { 996 callout_drain(&isc->ticker); 997 isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE; 998 } 999 1000 #endif 1001 free(isc, M_CAMSCHED); 1002 } 1003 } 1004 1005 /* 1006 * After we're sure we're attaching a device, go ahead and add 1007 * hooks for any sysctl we may wish to honor. 1008 */ 1009 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc, 1010 struct sysctl_ctx_list *ctx, struct sysctl_oid *node) 1011 { 1012 #ifdef CAM_NETFLIX_IOSCHED 1013 struct sysctl_oid_list *n; 1014 #endif 1015 1016 SYSCTL_ADD_INT(ctx, SYSCTL_CHILDREN(node), 1017 OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE, 1018 &isc->sort_io_queue, 0, 1019 "Sort IO queue to try and optimise disk access patterns"); 1020 1021 #ifdef CAM_NETFLIX_IOSCHED 1022 if (!do_netflix_iosched) 1023 return; 1024 1025 isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx, 1026 SYSCTL_CHILDREN(node), OID_AUTO, "iosched", 1027 CTLFLAG_RD, 0, "I/O scheduler statistics"); 1028 n = SYSCTL_CHILDREN(isc->sysctl_tree); 1029 ctx = &isc->sysctl_ctx; 1030 1031 cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read"); 1032 cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write"); 1033 cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim"); 1034 cam_iosched_cl_sysctl_init(isc); 1035 1036 SYSCTL_ADD_INT(ctx, n, 1037 OID_AUTO, "read_bias", CTLFLAG_RW, 1038 &isc->read_bias, 100, 1039 "How biased towards read should we be independent of limits"); 1040 1041 SYSCTL_ADD_INT(ctx, n, 1042 OID_AUTO, "quanta", CTLFLAG_RW, 1043 &isc->quanta, 200, 1044 "How many quanta per second do we slice the I/O up into"); 1045 1046 SYSCTL_ADD_INT(ctx, n, 1047 OID_AUTO, "total_ticks", CTLFLAG_RD, 1048 &isc->total_ticks, 0, 1049 "Total number of ticks we've done"); 1050 #endif 1051 } 1052 1053 /* 1054 * Flush outstanding I/O. Consumers of this library don't know all the 1055 * queues we may keep, so this allows all I/O to be flushed in one 1056 * convenient call. 1057 */ 1058 void 1059 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err) 1060 { 1061 bioq_flush(&isc->bio_queue, stp, err); 1062 bioq_flush(&isc->trim_queue, stp, err); 1063 #ifdef CAM_NETFLIX_IOSCHED 1064 if (do_netflix_iosched) 1065 bioq_flush(&isc->write_queue, stp, err); 1066 #endif 1067 } 1068 1069 #ifdef CAM_NETFLIX_IOSCHED 1070 static struct bio * 1071 cam_iosched_get_write(struct cam_iosched_softc *isc) 1072 { 1073 struct bio *bp; 1074 1075 /* 1076 * We control the write rate by controlling how many requests we send 1077 * down to the drive at any one time. Fewer requests limits the 1078 * effects of both starvation when the requests take a while and write 1079 * amplification when each request is causing more than one write to 1080 * the NAND media. Limiting the queue depth like this will also limit 1081 * the write throughput and give and reads that want to compete to 1082 * compete unfairly. 1083 */ 1084 bp = bioq_first(&isc->write_queue); 1085 if (bp == NULL) { 1086 if (iosched_debug > 3) 1087 printf("No writes present in write_queue\n"); 1088 return NULL; 1089 } 1090 1091 /* 1092 * If pending read, prefer that based on current read bias 1093 * setting. 1094 */ 1095 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) { 1096 if (iosched_debug) 1097 printf("Reads present and current_read_bias is %d queued writes %d queued reads %d\n", isc->current_read_bias, isc->write_stats.queued, isc->read_stats.queued); 1098 isc->current_read_bias--; 1099 return NULL; 1100 } 1101 1102 /* 1103 * See if our current limiter allows this I/O. 1104 */ 1105 if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) { 1106 if (iosched_debug) 1107 printf("Can't write because limiter says no.\n"); 1108 return NULL; 1109 } 1110 1111 /* 1112 * Let's do this: We've passed all the gates and we're a go 1113 * to schedule the I/O in the SIM. 1114 */ 1115 isc->current_read_bias = isc->read_bias; 1116 bioq_remove(&isc->write_queue, bp); 1117 if (bp->bio_cmd == BIO_WRITE) { 1118 isc->write_stats.queued--; 1119 isc->write_stats.total++; 1120 isc->write_stats.pending++; 1121 } 1122 if (iosched_debug > 9) 1123 printf("HWQ : %p %#x\n", bp, bp->bio_cmd); 1124 return bp; 1125 } 1126 #endif 1127 1128 /* 1129 * Put back a trim that you weren't able to actually schedule this time. 1130 */ 1131 void 1132 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp) 1133 { 1134 bioq_insert_head(&isc->trim_queue, bp); 1135 #ifdef CAM_NETFLIX_IOSCHED 1136 isc->trim_stats.queued++; 1137 isc->trim_stats.total--; /* since we put it back, don't double count */ 1138 isc->trim_stats.pending--; 1139 #endif 1140 } 1141 1142 /* 1143 * gets the next trim from the trim queue. 1144 * 1145 * Assumes we're called with the periph lock held. It removes this 1146 * trim from the queue and the device must explicitly reinstert it 1147 * should the need arise. 1148 */ 1149 struct bio * 1150 cam_iosched_next_trim(struct cam_iosched_softc *isc) 1151 { 1152 struct bio *bp; 1153 1154 bp = bioq_first(&isc->trim_queue); 1155 if (bp == NULL) 1156 return NULL; 1157 bioq_remove(&isc->trim_queue, bp); 1158 #ifdef CAM_NETFLIX_IOSCHED 1159 isc->trim_stats.queued--; 1160 isc->trim_stats.total++; 1161 isc->trim_stats.pending++; 1162 #endif 1163 return bp; 1164 } 1165 1166 /* 1167 * gets the an available trim from the trim queue, if there's no trim 1168 * already pending. It removes this trim from the queue and the device 1169 * must explicitly reinstert it should the need arise. 1170 * 1171 * Assumes we're called with the periph lock held. 1172 */ 1173 struct bio * 1174 cam_iosched_get_trim(struct cam_iosched_softc *isc) 1175 { 1176 1177 if (!cam_iosched_has_more_trim(isc)) 1178 return NULL; 1179 1180 return cam_iosched_next_trim(isc); 1181 } 1182 1183 /* 1184 * Determine what the next bit of work to do is for the periph. The 1185 * default implementation looks to see if we have trims to do, but no 1186 * trims outstanding. If so, we do that. Otherwise we see if we have 1187 * other work. If we do, then we do that. Otherwise why were we called? 1188 */ 1189 struct bio * 1190 cam_iosched_next_bio(struct cam_iosched_softc *isc) 1191 { 1192 struct bio *bp; 1193 1194 /* 1195 * See if we have a trim that can be scheduled. We can only send one 1196 * at a time down, so this takes that into account. 1197 * 1198 * XXX newer TRIM commands are queueable. Revisit this when we 1199 * implement them. 1200 */ 1201 if ((bp = cam_iosched_get_trim(isc)) != NULL) 1202 return bp; 1203 1204 #ifdef CAM_NETFLIX_IOSCHED 1205 /* 1206 * See if we have any pending writes, and room in the queue for them, 1207 * and if so, those are next. 1208 */ 1209 if (do_netflix_iosched) { 1210 if ((bp = cam_iosched_get_write(isc)) != NULL) 1211 return bp; 1212 } 1213 #endif 1214 1215 /* 1216 * next, see if there's other, normal I/O waiting. If so return that. 1217 */ 1218 if ((bp = bioq_first(&isc->bio_queue)) == NULL) 1219 return NULL; 1220 1221 #ifdef CAM_NETFLIX_IOSCHED 1222 /* 1223 * For the netflix scheduler, bio_queue is only for reads, so enforce 1224 * the limits here. Enforce only for reads. 1225 */ 1226 if (do_netflix_iosched) { 1227 if (bp->bio_cmd == BIO_READ && 1228 cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) 1229 return NULL; 1230 } 1231 #endif 1232 bioq_remove(&isc->bio_queue, bp); 1233 #ifdef CAM_NETFLIX_IOSCHED 1234 if (do_netflix_iosched) { 1235 if (bp->bio_cmd == BIO_READ) { 1236 isc->read_stats.queued--; 1237 isc->read_stats.total++; 1238 isc->read_stats.pending++; 1239 } else 1240 printf("Found bio_cmd = %#x\n", bp->bio_cmd); 1241 } 1242 if (iosched_debug > 9) 1243 printf("HWQ : %p %#x\n", bp, bp->bio_cmd); 1244 #endif 1245 return bp; 1246 } 1247 1248 /* 1249 * Driver has been given some work to do by the block layer. Tell the 1250 * scheduler about it and have it queue the work up. The scheduler module 1251 * will then return the currently most useful bit of work later, possibly 1252 * deferring work for various reasons. 1253 */ 1254 void 1255 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp) 1256 { 1257 1258 /* 1259 * Put all trims on the trim queue sorted, since we know 1260 * that the collapsing code requires this. Otherwise put 1261 * the work on the bio queue. 1262 */ 1263 if (bp->bio_cmd == BIO_DELETE) { 1264 bioq_disksort(&isc->trim_queue, bp); 1265 #ifdef CAM_NETFLIX_IOSCHED 1266 isc->trim_stats.in++; 1267 isc->trim_stats.queued++; 1268 #endif 1269 } 1270 #ifdef CAM_NETFLIX_IOSCHED 1271 else if (do_netflix_iosched && 1272 (bp->bio_cmd == BIO_WRITE || bp->bio_cmd == BIO_FLUSH)) { 1273 if (cam_iosched_sort_queue(isc)) 1274 bioq_disksort(&isc->write_queue, bp); 1275 else 1276 bioq_insert_tail(&isc->write_queue, bp); 1277 if (iosched_debug > 9) 1278 printf("Qw : %p %#x\n", bp, bp->bio_cmd); 1279 if (bp->bio_cmd == BIO_WRITE) { 1280 isc->write_stats.in++; 1281 isc->write_stats.queued++; 1282 } 1283 } 1284 #endif 1285 else { 1286 if (cam_iosched_sort_queue(isc)) 1287 bioq_disksort(&isc->bio_queue, bp); 1288 else 1289 bioq_insert_tail(&isc->bio_queue, bp); 1290 #ifdef CAM_NETFLIX_IOSCHED 1291 if (iosched_debug > 9) 1292 printf("Qr : %p %#x\n", bp, bp->bio_cmd); 1293 if (bp->bio_cmd == BIO_READ) { 1294 isc->read_stats.in++; 1295 isc->read_stats.queued++; 1296 } else if (bp->bio_cmd == BIO_WRITE) { 1297 isc->write_stats.in++; 1298 isc->write_stats.queued++; 1299 } 1300 #endif 1301 } 1302 } 1303 1304 /* 1305 * If we have work, get it scheduled. Called with the periph lock held. 1306 */ 1307 void 1308 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph) 1309 { 1310 1311 if (cam_iosched_has_work(isc)) 1312 xpt_schedule(periph, CAM_PRIORITY_NORMAL); 1313 } 1314 1315 /* 1316 * Complete a trim request 1317 */ 1318 void 1319 cam_iosched_trim_done(struct cam_iosched_softc *isc) 1320 { 1321 1322 isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE; 1323 } 1324 1325 /* 1326 * Complete a bio. Called before we release the ccb with xpt_release_ccb so we 1327 * might use notes in the ccb for statistics. 1328 */ 1329 int 1330 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp, 1331 union ccb *done_ccb) 1332 { 1333 int retval = 0; 1334 #ifdef CAM_NETFLIX_IOSCHED 1335 if (!do_netflix_iosched) 1336 return retval; 1337 1338 if (iosched_debug > 10) 1339 printf("done: %p %#x\n", bp, bp->bio_cmd); 1340 if (bp->bio_cmd == BIO_WRITE) { 1341 retval = cam_iosched_limiter_iodone(&isc->write_stats, bp); 1342 isc->write_stats.out++; 1343 isc->write_stats.pending--; 1344 } else if (bp->bio_cmd == BIO_READ) { 1345 retval = cam_iosched_limiter_iodone(&isc->read_stats, bp); 1346 isc->read_stats.out++; 1347 isc->read_stats.pending--; 1348 } else if (bp->bio_cmd == BIO_DELETE) { 1349 isc->trim_stats.out++; 1350 isc->trim_stats.pending--; 1351 } else if (bp->bio_cmd != BIO_FLUSH) { 1352 if (iosched_debug) 1353 printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd); 1354 } 1355 1356 if (!(bp->bio_flags & BIO_ERROR)) 1357 cam_iosched_io_metric_update(isc, done_ccb->ccb_h.qos.sim_data, 1358 bp->bio_cmd, bp->bio_bcount); 1359 #endif 1360 return retval; 1361 } 1362 1363 /* 1364 * Tell the io scheduler that you've pushed a trim down into the sim. 1365 * xxx better place for this? 1366 */ 1367 void 1368 cam_iosched_submit_trim(struct cam_iosched_softc *isc) 1369 { 1370 1371 isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE; 1372 } 1373 1374 /* 1375 * Change the sorting policy hint for I/O transactions for this device. 1376 */ 1377 void 1378 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val) 1379 { 1380 1381 isc->sort_io_queue = val; 1382 } 1383 1384 int 1385 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags) 1386 { 1387 return isc->flags & flags; 1388 } 1389 1390 void 1391 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags) 1392 { 1393 isc->flags |= flags; 1394 } 1395 1396 void 1397 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags) 1398 { 1399 isc->flags &= ~flags; 1400 } 1401 1402 #ifdef CAM_NETFLIX_IOSCHED 1403 /* 1404 * After the method presented in Jack Crenshaw's 1998 article "Integer 1405 * Suqare Roots," reprinted at 1406 * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots 1407 * and well worth the read. Briefly, we find the power of 4 that's the 1408 * largest smaller than val. We then check each smaller power of 4 to 1409 * see if val is still bigger. The right shifts at each step divide 1410 * the result by 2 which after successive application winds up 1411 * accumulating the right answer. It could also have been accumulated 1412 * using a separate root counter, but this code is smaller and faster 1413 * than that method. This method is also integer size invariant. 1414 * It returns floor(sqrt((float)val)), or the larget integer less than 1415 * or equal to the square root. 1416 */ 1417 static uint64_t 1418 isqrt64(uint64_t val) 1419 { 1420 uint64_t res = 0; 1421 uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2); 1422 1423 /* 1424 * Find the largest power of 4 smaller than val. 1425 */ 1426 while (bit > val) 1427 bit >>= 2; 1428 1429 /* 1430 * Accumulate the answer, one bit at a time (we keep moving 1431 * them over since 2 is the square root of 4 and we test 1432 * powers of 4). We accumulate where we find the bit, but 1433 * the successive shifts land the bit in the right place 1434 * by the end. 1435 */ 1436 while (bit != 0) { 1437 if (val >= res + bit) { 1438 val -= res + bit; 1439 res = (res >> 1) + bit; 1440 } else 1441 res >>= 1; 1442 bit >>= 2; 1443 } 1444 1445 return res; 1446 } 1447 1448 /* 1449 * a and b are 32.32 fixed point stored in a 64-bit word. 1450 * Let al and bl be the .32 part of a and b. 1451 * Let ah and bh be the 32 part of a and b. 1452 * R is the radix and is 1 << 32 1453 * 1454 * a * b 1455 * (ah + al / R) * (bh + bl / R) 1456 * ah * bh + (al * bh + ah * bl) / R + al * bl / R^2 1457 * 1458 * After multiplicaiton, we have to renormalize by multiply by 1459 * R, so we wind up with 1460 * ah * bh * R + al * bh + ah * bl + al * bl / R 1461 * which turns out to be a very nice way to compute this value 1462 * so long as ah and bh are < 65536 there's no loss of high bits 1463 * and the low order bits are below the threshold of caring for 1464 * this application. 1465 */ 1466 static uint64_t 1467 mul(uint64_t a, uint64_t b) 1468 { 1469 uint64_t al, ah, bl, bh; 1470 al = a & 0xffffffff; 1471 ah = a >> 32; 1472 bl = b & 0xffffffff; 1473 bh = b >> 32; 1474 return ((ah * bh) << 32) + al * bh + ah * bl + ((al * bl) >> 32); 1475 } 1476 1477 static void 1478 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency) 1479 { 1480 sbintime_t y, yy; 1481 uint64_t var; 1482 1483 /* 1484 * Classic expoentially decaying average with a tiny alpha 1485 * (2 ^ -alpha_bits). For more info see the NIST statistical 1486 * handbook. 1487 * 1488 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha) 1489 * alpha = 1 / (1 << alpha_bits) 1490 * 1491 * Since alpha is a power of two, we can compute this w/o any mult or 1492 * division. 1493 */ 1494 y = sim_latency; 1495 iop->ema = (y + (iop->ema << alpha_bits) - iop->ema) >> alpha_bits; 1496 1497 yy = mul(y, y); 1498 iop->emss = (yy + (iop->emss << alpha_bits) - iop->emss) >> alpha_bits; 1499 1500 /* 1501 * s_1 = sum of data 1502 * s_2 = sum of data * data 1503 * ema ~ mean (or s_1 / N) 1504 * emss ~ s_2 / N 1505 * 1506 * sd = sqrt((N * s_2 - s_1 ^ 2) / (N * (N - 1))) 1507 * sd = sqrt((N * s_2 / N * (N - 1)) - (s_1 ^ 2 / (N * (N - 1)))) 1508 * 1509 * N ~ 2 / alpha - 1 1510 * alpha < 1 / 16 (typically much less) 1511 * N > 31 --> N large so N * (N - 1) is approx N * N 1512 * 1513 * substituting and rearranging: 1514 * sd ~ sqrt(s_2 / N - (s_1 / N) ^ 2) 1515 * ~ sqrt(emss - ema ^ 2); 1516 * which is the formula used here to get a decent estimate of sd which 1517 * we use to detect outliers. Note that when first starting up, it 1518 * takes a while for emss sum of squares estimator to converge on a 1519 * good value. during this time, it can be less than ema^2. We 1520 * compute a sd of 0 in that case, and ignore outliers. 1521 */ 1522 var = iop->emss - mul(iop->ema, iop->ema); 1523 iop->sd = (int64_t)var < 0 ? 0 : isqrt64(var); 1524 } 1525 1526 #ifdef CAM_NETFLIX_IOSCHED 1527 static void 1528 cam_iosched_io_metric_update(struct cam_iosched_softc *isc, 1529 sbintime_t sim_latency, int cmd, size_t size) 1530 { 1531 /* xxx Do we need to scale based on the size of the I/O ? */ 1532 switch (cmd) { 1533 case BIO_READ: 1534 cam_iosched_update(&isc->read_stats, sim_latency); 1535 break; 1536 case BIO_WRITE: 1537 cam_iosched_update(&isc->write_stats, sim_latency); 1538 break; 1539 case BIO_DELETE: 1540 cam_iosched_update(&isc->trim_stats, sim_latency); 1541 break; 1542 default: 1543 break; 1544 } 1545 } 1546 #endif 1547 1548 #ifdef DDB 1549 static int biolen(struct bio_queue_head *bq) 1550 { 1551 int i = 0; 1552 struct bio *bp; 1553 1554 TAILQ_FOREACH(bp, &bq->queue, bio_queue) { 1555 i++; 1556 } 1557 return i; 1558 } 1559 1560 /* 1561 * Show the internal state of the I/O scheduler. 1562 */ 1563 DB_SHOW_COMMAND(iosched, cam_iosched_db_show) 1564 { 1565 struct cam_iosched_softc *isc; 1566 1567 if (!have_addr) { 1568 db_printf("Need addr\n"); 1569 return; 1570 } 1571 isc = (struct cam_iosched_softc *)addr; 1572 db_printf("pending_reads: %d\n", isc->read_stats.pending); 1573 db_printf("min_reads: %d\n", isc->read_stats.min); 1574 db_printf("max_reads: %d\n", isc->read_stats.max); 1575 db_printf("reads: %d\n", isc->read_stats.total); 1576 db_printf("in_reads: %d\n", isc->read_stats.in); 1577 db_printf("out_reads: %d\n", isc->read_stats.out); 1578 db_printf("queued_reads: %d\n", isc->read_stats.queued); 1579 db_printf("Current Q len %d\n", biolen(&isc->bio_queue)); 1580 db_printf("pending_writes: %d\n", isc->write_stats.pending); 1581 db_printf("min_writes: %d\n", isc->write_stats.min); 1582 db_printf("max_writes: %d\n", isc->write_stats.max); 1583 db_printf("writes: %d\n", isc->write_stats.total); 1584 db_printf("in_writes: %d\n", isc->write_stats.in); 1585 db_printf("out_writes: %d\n", isc->write_stats.out); 1586 db_printf("queued_writes: %d\n", isc->write_stats.queued); 1587 db_printf("Current Q len %d\n", biolen(&isc->write_queue)); 1588 db_printf("pending_trims: %d\n", isc->trim_stats.pending); 1589 db_printf("min_trims: %d\n", isc->trim_stats.min); 1590 db_printf("max_trims: %d\n", isc->trim_stats.max); 1591 db_printf("trims: %d\n", isc->trim_stats.total); 1592 db_printf("in_trims: %d\n", isc->trim_stats.in); 1593 db_printf("out_trims: %d\n", isc->trim_stats.out); 1594 db_printf("queued_trims: %d\n", isc->trim_stats.queued); 1595 db_printf("Current Q len %d\n", biolen(&isc->trim_queue)); 1596 db_printf("read_bias: %d\n", isc->read_bias); 1597 db_printf("current_read_bias: %d\n", isc->current_read_bias); 1598 db_printf("Trim active? %s\n", 1599 (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no"); 1600 } 1601 #endif 1602 #endif 1603