1 /*- 2 * CAM IO Scheduler Interface 3 * 4 * SPDX-License-Identifier: BSD-2-Clause 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 * 2. Redistributions in binary form must reproduce the above copyright 14 * notice, this list of conditions and the following disclaimer in the 15 * documentation and/or other materials provided with the distribution. 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 21 * FOR 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 30 #include "opt_ddb.h" 31 32 #include <sys/param.h> 33 #include <sys/systm.h> 34 #include <sys/kernel.h> 35 #include <sys/bio.h> 36 #include <sys/lock.h> 37 #include <sys/malloc.h> 38 #include <sys/mutex.h> 39 #include <sys/sbuf.h> 40 #include <sys/sysctl.h> 41 42 #include <cam/cam.h> 43 #include <cam/cam_ccb.h> 44 #include <cam/cam_periph.h> 45 #include <cam/cam_xpt_periph.h> 46 #include <cam/cam_xpt_internal.h> 47 #include <cam/cam_iosched.h> 48 49 #include <ddb/ddb.h> 50 51 #include <geom/geom_disk.h> 52 53 static MALLOC_DEFINE(M_CAMSCHED, "CAM I/O Scheduler", 54 "CAM I/O Scheduler buffers"); 55 56 static SYSCTL_NODE(_kern_cam, OID_AUTO, iosched, CTLFLAG_RD | CTLFLAG_MPSAFE, 0, 57 "CAM I/O Scheduler parameters"); 58 59 /* 60 * Default I/O scheduler for FreeBSD. This implementation is just a thin-vineer 61 * over the bioq_* interface, with notions of separate calls for normal I/O and 62 * for trims. 63 * 64 * When CAM_IOSCHED_DYNAMIC is defined, the scheduler is enhanced to dynamically 65 * steer the rate of one type of traffic to help other types of traffic (eg 66 * limit writes when read latency deteriorates on SSDs). 67 */ 68 69 #ifdef CAM_IOSCHED_DYNAMIC 70 71 static bool do_dynamic_iosched = true; 72 SYSCTL_BOOL(_kern_cam_iosched, OID_AUTO, dynamic, CTLFLAG_RDTUN, 73 &do_dynamic_iosched, 1, 74 "Enable Dynamic I/O scheduler optimizations."); 75 76 /* 77 * For an EMA, with an alpha of alpha, we know 78 * alpha = 2 / (N + 1) 79 * or 80 * N = 1 + (2 / alpha) 81 * where N is the number of samples that 86% of the current 82 * EMA is derived from. 83 * 84 * So we invent[*] alpha_bits: 85 * alpha_bits = -log_2(alpha) 86 * alpha = 2^-alpha_bits 87 * So 88 * N = 1 + 2^(alpha_bits + 1) 89 * 90 * The default 9 gives a 1025 lookback for 86% of the data. 91 * For a brief intro: https://en.wikipedia.org/wiki/Moving_average 92 * 93 * [*] Steal from the load average code and many other places. 94 * Note: See computation of EMA and EMVAR for acceptable ranges of alpha. 95 */ 96 static int alpha_bits = 9; 97 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, alpha_bits, CTLFLAG_RWTUN, 98 &alpha_bits, 1, 99 "Bits in EMA's alpha."); 100 101 /* 102 * Different parameters for the buckets of latency we keep track of. These are all 103 * published read-only since at present they are compile time constants. 104 * 105 * Bucket base is the upper bounds of the first latency bucket. It's currently 20us. 106 * With 20 buckets (see below), that leads to a geometric progression with a max size 107 * of 5.2s which is safeily larger than 1s to help diagnose extreme outliers better. 108 */ 109 #ifndef BUCKET_BASE 110 #define BUCKET_BASE ((SBT_1S / 50000) + 1) /* 20us */ 111 #endif 112 static sbintime_t bucket_base = BUCKET_BASE; 113 SYSCTL_SBINTIME_USEC(_kern_cam_iosched, OID_AUTO, bucket_base_us, CTLFLAG_RD, 114 &bucket_base, 115 "Size of the smallest latency bucket"); 116 117 /* 118 * Bucket ratio is the geometric progression for the bucket. For a bucket b_n 119 * the size of bucket b_n+1 is b_n * bucket_ratio / 100. 120 */ 121 static int bucket_ratio = 200; /* Rather hard coded at the moment */ 122 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, bucket_ratio, CTLFLAG_RD, 123 &bucket_ratio, 200, 124 "Latency Bucket Ratio for geometric progression."); 125 126 /* 127 * Number of total buckets. Starting at BUCKET_BASE, each one is a power of 2. 128 */ 129 #ifndef LAT_BUCKETS 130 #define LAT_BUCKETS 20 /* < 20us < 40us ... < 2^(n-1)*20us >= 2^(n-1)*20us */ 131 #endif 132 static int lat_buckets = LAT_BUCKETS; 133 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, buckets, CTLFLAG_RD, 134 &lat_buckets, LAT_BUCKETS, 135 "Total number of latency buckets published"); 136 137 /* 138 * Read bias: how many reads do we favor before scheduling a write 139 * when we have a choice. 140 */ 141 static int default_read_bias = 0; 142 SYSCTL_INT(_kern_cam_iosched, OID_AUTO, read_bias, CTLFLAG_RWTUN, 143 &default_read_bias, 0, 144 "Default read bias for new devices."); 145 146 struct iop_stats; 147 struct cam_iosched_softc; 148 149 int iosched_debug = 0; 150 151 typedef enum { 152 none = 0, /* No limits */ 153 queue_depth, /* Limit how many ops we queue to SIM */ 154 iops, /* Limit # of IOPS to the drive */ 155 bandwidth, /* Limit bandwidth to the drive */ 156 limiter_max 157 } io_limiter; 158 159 static const char *cam_iosched_limiter_names[] = 160 { "none", "queue_depth", "iops", "bandwidth" }; 161 162 /* 163 * Called to initialize the bits of the iop_stats structure relevant to the 164 * limiter. Called just after the limiter is set. 165 */ 166 typedef int l_init_t(struct iop_stats *); 167 168 /* 169 * Called every tick. 170 */ 171 typedef int l_tick_t(struct iop_stats *); 172 173 /* 174 * Called to see if the limiter thinks this IOP can be allowed to 175 * proceed. If so, the limiter assumes that the IOP proceeded 176 * and makes any accounting of it that's needed. 177 */ 178 typedef int l_iop_t(struct iop_stats *, struct bio *); 179 180 /* 181 * Called when an I/O completes so the limiter can update its 182 * accounting. Pending I/Os may complete in any order (even when 183 * sent to the hardware at the same time), so the limiter may not 184 * make any assumptions other than this I/O has completed. If it 185 * returns 1, then xpt_schedule() needs to be called again. 186 */ 187 typedef int l_iodone_t(struct iop_stats *, struct bio *); 188 189 static l_iop_t cam_iosched_qd_iop; 190 static l_iop_t cam_iosched_qd_caniop; 191 static l_iodone_t cam_iosched_qd_iodone; 192 193 static l_init_t cam_iosched_iops_init; 194 static l_tick_t cam_iosched_iops_tick; 195 static l_iop_t cam_iosched_iops_caniop; 196 static l_iop_t cam_iosched_iops_iop; 197 198 static l_init_t cam_iosched_bw_init; 199 static l_tick_t cam_iosched_bw_tick; 200 static l_iop_t cam_iosched_bw_caniop; 201 static l_iop_t cam_iosched_bw_iop; 202 203 struct limswitch { 204 l_init_t *l_init; 205 l_tick_t *l_tick; 206 l_iop_t *l_iop; 207 l_iop_t *l_caniop; 208 l_iodone_t *l_iodone; 209 } limsw[] = 210 { 211 { /* none */ 212 .l_init = NULL, 213 .l_tick = NULL, 214 .l_iop = NULL, 215 .l_iodone= NULL, 216 }, 217 { /* queue_depth */ 218 .l_init = NULL, 219 .l_tick = NULL, 220 .l_caniop = cam_iosched_qd_caniop, 221 .l_iop = cam_iosched_qd_iop, 222 .l_iodone= cam_iosched_qd_iodone, 223 }, 224 { /* iops */ 225 .l_init = cam_iosched_iops_init, 226 .l_tick = cam_iosched_iops_tick, 227 .l_caniop = cam_iosched_iops_caniop, 228 .l_iop = cam_iosched_iops_iop, 229 .l_iodone= NULL, 230 }, 231 { /* bandwidth */ 232 .l_init = cam_iosched_bw_init, 233 .l_tick = cam_iosched_bw_tick, 234 .l_caniop = cam_iosched_bw_caniop, 235 .l_iop = cam_iosched_bw_iop, 236 .l_iodone= NULL, 237 }, 238 }; 239 240 struct iop_stats { 241 /* 242 * sysctl state for this subnode. 243 */ 244 struct sysctl_ctx_list sysctl_ctx; 245 struct sysctl_oid *sysctl_tree; 246 247 /* 248 * Information about the current rate limiters, if any 249 */ 250 io_limiter limiter; /* How are I/Os being limited */ 251 int min; /* Low range of limit */ 252 int max; /* High range of limit */ 253 int current; /* Current rate limiter */ 254 int l_value1; /* per-limiter scratch value 1. */ 255 int l_value2; /* per-limiter scratch value 2. */ 256 257 /* 258 * Debug information about counts of I/Os that have gone through the 259 * scheduler. 260 */ 261 int pending; /* I/Os pending in the hardware */ 262 int queued; /* number currently in the queue */ 263 int total; /* Total for all time -- wraps */ 264 int in; /* number queued all time -- wraps */ 265 int out; /* number completed all time -- wraps */ 266 int errs; /* Number of I/Os completed with error -- wraps */ 267 268 /* 269 * Statistics on different bits of the process. 270 */ 271 /* Exp Moving Average, see alpha_bits for more details */ 272 sbintime_t ema; 273 sbintime_t emvar; 274 sbintime_t sd; /* Last computed sd */ 275 276 uint64_t too_long; /* Number of I/Os greater than bad lat threshold */ 277 sbintime_t bad_latency; /* Latency threshold */ 278 279 uint32_t state_flags; 280 #define IOP_RATE_LIMITED 1u 281 282 uint64_t latencies[LAT_BUCKETS]; 283 284 struct cam_iosched_softc *softc; 285 }; 286 287 typedef enum { 288 set_max = 0, /* current = max */ 289 read_latency, /* Steer read latency by throttling writes */ 290 cl_max /* Keep last */ 291 } control_type; 292 293 static const char *cam_iosched_control_type_names[] = 294 { "set_max", "read_latency" }; 295 296 struct control_loop { 297 /* 298 * sysctl state for this subnode. 299 */ 300 struct sysctl_ctx_list sysctl_ctx; 301 struct sysctl_oid *sysctl_tree; 302 303 sbintime_t next_steer; /* Time of next steer */ 304 sbintime_t steer_interval; /* How often do we steer? */ 305 sbintime_t lolat; 306 sbintime_t hilat; 307 int alpha; 308 control_type type; /* What type of control? */ 309 int last_count; /* Last I/O count */ 310 311 struct cam_iosched_softc *softc; 312 }; 313 314 #endif 315 316 struct cam_iosched_softc { 317 struct bio_queue_head bio_queue; 318 struct bio_queue_head trim_queue; 319 const struct disk *disk; 320 cam_iosched_schedule_t schedfnc; 321 /* scheduler flags < 16, user flags >= 16 */ 322 uint32_t flags; 323 int sort_io_queue; 324 int trim_goal; /* # of trims to queue before sending */ 325 int trim_ticks; /* Max ticks to hold trims */ 326 int last_trim_tick; /* Last 'tick' time ld a trim */ 327 int queued_trims; /* Number of trims in the queue */ 328 #ifdef CAM_IOSCHED_DYNAMIC 329 int read_bias; /* Read bias setting */ 330 int current_read_bias; /* Current read bias state */ 331 int total_ticks; 332 int load; /* EMA of 'load average' of disk / 2^16 */ 333 334 struct bio_queue_head write_queue; 335 struct iop_stats read_stats, write_stats, trim_stats; 336 struct sysctl_ctx_list sysctl_ctx; 337 struct sysctl_oid *sysctl_tree; 338 339 int quanta; /* Number of quanta per second */ 340 struct callout ticker; /* Callout for our quota system */ 341 struct cam_periph *periph; /* cam periph associated with this device */ 342 uint32_t this_frac; /* Fraction of a second (1024ths) for this tick */ 343 sbintime_t last_time; /* Last time we ticked */ 344 struct control_loop cl; 345 sbintime_t max_lat; /* when != 0, if iop latency > max_lat, call max_lat_fcn */ 346 cam_iosched_latfcn_t latfcn; 347 void *latarg; 348 #endif 349 }; 350 351 #ifdef CAM_IOSCHED_DYNAMIC 352 /* 353 * helper functions to call the limsw functions. 354 */ 355 static int 356 cam_iosched_limiter_init(struct iop_stats *ios) 357 { 358 int lim = ios->limiter; 359 360 /* maybe this should be a kassert */ 361 if (lim < none || lim >= limiter_max) 362 return EINVAL; 363 364 if (limsw[lim].l_init) 365 return limsw[lim].l_init(ios); 366 367 return 0; 368 } 369 370 static int 371 cam_iosched_limiter_tick(struct iop_stats *ios) 372 { 373 int lim = ios->limiter; 374 375 /* maybe this should be a kassert */ 376 if (lim < none || lim >= limiter_max) 377 return EINVAL; 378 379 if (limsw[lim].l_tick) 380 return limsw[lim].l_tick(ios); 381 382 return 0; 383 } 384 385 static int 386 cam_iosched_limiter_iop(struct iop_stats *ios, struct bio *bp) 387 { 388 int lim = ios->limiter; 389 390 /* maybe this should be a kassert */ 391 if (lim < none || lim >= limiter_max) 392 return EINVAL; 393 394 if (limsw[lim].l_iop) 395 return limsw[lim].l_iop(ios, bp); 396 397 return 0; 398 } 399 400 static int 401 cam_iosched_limiter_caniop(struct iop_stats *ios, struct bio *bp) 402 { 403 int lim = ios->limiter; 404 405 /* maybe this should be a kassert */ 406 if (lim < none || lim >= limiter_max) 407 return EINVAL; 408 409 if (limsw[lim].l_caniop) 410 return limsw[lim].l_caniop(ios, bp); 411 412 return 0; 413 } 414 415 static int 416 cam_iosched_limiter_iodone(struct iop_stats *ios, struct bio *bp) 417 { 418 int lim = ios->limiter; 419 420 /* maybe this should be a kassert */ 421 if (lim < none || lim >= limiter_max) 422 return 0; 423 424 if (limsw[lim].l_iodone) 425 return limsw[lim].l_iodone(ios, bp); 426 427 return 0; 428 } 429 430 /* 431 * Functions to implement the different kinds of limiters 432 */ 433 434 static int 435 cam_iosched_qd_iop(struct iop_stats *ios, struct bio *bp) 436 { 437 438 if (ios->current <= 0 || ios->pending < ios->current) 439 return 0; 440 441 return EAGAIN; 442 } 443 444 static int 445 cam_iosched_qd_caniop(struct iop_stats *ios, struct bio *bp) 446 { 447 448 if (ios->current <= 0 || ios->pending < ios->current) 449 return 0; 450 451 return EAGAIN; 452 } 453 454 static int 455 cam_iosched_qd_iodone(struct iop_stats *ios, struct bio *bp) 456 { 457 458 if (ios->current <= 0 || ios->pending != ios->current) 459 return 0; 460 461 return 1; 462 } 463 464 static int 465 cam_iosched_iops_init(struct iop_stats *ios) 466 { 467 468 ios->l_value1 = ios->current / ios->softc->quanta; 469 if (ios->l_value1 <= 0) 470 ios->l_value1 = 1; 471 ios->l_value2 = 0; 472 473 return 0; 474 } 475 476 static int 477 cam_iosched_iops_tick(struct iop_stats *ios) 478 { 479 int new_ios; 480 481 /* 482 * Allow at least one IO per tick until all 483 * the IOs for this interval have been spent. 484 */ 485 new_ios = (int)((ios->current * (uint64_t)ios->softc->this_frac) >> 16); 486 if (new_ios < 1 && ios->l_value2 < ios->current) { 487 new_ios = 1; 488 ios->l_value2++; 489 } 490 491 /* 492 * If this a new accounting interval, discard any "unspent" ios 493 * granted in the previous interval. Otherwise add the new ios to 494 * the previously granted ones that haven't been spent yet. 495 */ 496 if ((ios->softc->total_ticks % ios->softc->quanta) == 0) { 497 ios->l_value1 = new_ios; 498 ios->l_value2 = 1; 499 } else { 500 ios->l_value1 += new_ios; 501 } 502 503 return 0; 504 } 505 506 static int 507 cam_iosched_iops_caniop(struct iop_stats *ios, struct bio *bp) 508 { 509 510 /* 511 * So if we have any more IOPs left, allow it, 512 * otherwise wait. If current iops is 0, treat that 513 * as unlimited as a failsafe. 514 */ 515 if (ios->current > 0 && ios->l_value1 <= 0) 516 return EAGAIN; 517 return 0; 518 } 519 520 static int 521 cam_iosched_iops_iop(struct iop_stats *ios, struct bio *bp) 522 { 523 int rv; 524 525 rv = cam_iosched_limiter_caniop(ios, bp); 526 if (rv == 0) 527 ios->l_value1--; 528 529 return rv; 530 } 531 532 static int 533 cam_iosched_bw_init(struct iop_stats *ios) 534 { 535 536 /* ios->current is in kB/s, so scale to bytes */ 537 ios->l_value1 = ios->current * 1000 / ios->softc->quanta; 538 539 return 0; 540 } 541 542 static int 543 cam_iosched_bw_tick(struct iop_stats *ios) 544 { 545 int bw; 546 547 /* 548 * If we're in the hole for available quota from 549 * the last time, then add the quantum for this. 550 * If we have any left over from last quantum, 551 * then too bad, that's lost. Also, ios->current 552 * is in kB/s, so scale. 553 * 554 * We also allow up to 4 quanta of credits to 555 * accumulate to deal with burstiness. 4 is extremely 556 * arbitrary. 557 */ 558 bw = (int)((ios->current * 1000ull * (uint64_t)ios->softc->this_frac) >> 16); 559 if (ios->l_value1 < bw * 4) 560 ios->l_value1 += bw; 561 562 return 0; 563 } 564 565 static int 566 cam_iosched_bw_caniop(struct iop_stats *ios, struct bio *bp) 567 { 568 /* 569 * So if we have any more bw quota left, allow it, 570 * otherwise wait. Note, we'll go negative and that's 571 * OK. We'll just get a little less next quota. 572 * 573 * Note on going negative: that allows us to process 574 * requests in order better, since we won't allow 575 * shorter reads to get around the long one that we 576 * don't have the quota to do just yet. It also prevents 577 * starvation by being a little more permissive about 578 * what we let through this quantum (to prevent the 579 * starvation), at the cost of getting a little less 580 * next quantum. 581 * 582 * Also note that if the current limit is <= 0, 583 * we treat it as unlimited as a failsafe. 584 */ 585 if (ios->current > 0 && ios->l_value1 <= 0) 586 return EAGAIN; 587 588 return 0; 589 } 590 591 static int 592 cam_iosched_bw_iop(struct iop_stats *ios, struct bio *bp) 593 { 594 int rv; 595 596 rv = cam_iosched_limiter_caniop(ios, bp); 597 if (rv == 0) 598 ios->l_value1 -= bp->bio_length; 599 600 return rv; 601 } 602 603 static void cam_iosched_cl_maybe_steer(struct control_loop *clp); 604 605 static void 606 cam_iosched_ticker(void *arg) 607 { 608 struct cam_iosched_softc *isc = arg; 609 sbintime_t now, delta; 610 int pending; 611 612 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc); 613 614 now = sbinuptime(); 615 delta = now - isc->last_time; 616 isc->this_frac = (uint32_t)delta >> 16; /* Note: discards seconds -- should be 0 harmless if not */ 617 isc->last_time = now; 618 619 cam_iosched_cl_maybe_steer(&isc->cl); 620 621 cam_iosched_limiter_tick(&isc->read_stats); 622 cam_iosched_limiter_tick(&isc->write_stats); 623 cam_iosched_limiter_tick(&isc->trim_stats); 624 625 isc->schedfnc(isc->periph); 626 627 /* 628 * isc->load is an EMA of the pending I/Os at each tick. The number of 629 * pending I/Os is the sum of the I/Os queued to the hardware, and those 630 * in the software queue that could be queued to the hardware if there 631 * were slots. 632 * 633 * ios_stats.pending is a count of requests in the SIM right now for 634 * each of these types of I/O. So the total pending count is the sum of 635 * these I/Os and the sum of the queued I/Os still in the software queue 636 * for those operations that aren't being rate limited at the moment. 637 * 638 * The reason for the rate limiting bit is because those I/Os 639 * aren't part of the software queued load (since we could 640 * give them to hardware, but choose not to). 641 * 642 * Note: due to a bug in counting pending TRIM in the device, we 643 * don't include them in this count. We count each BIO_DELETE in 644 * the pending count, but the periph drivers collapse them down 645 * into one TRIM command. That one trim command gets the completion 646 * so the counts get off. 647 */ 648 pending = isc->read_stats.pending + isc->write_stats.pending /* + isc->trim_stats.pending */; 649 pending += !!(isc->read_stats.state_flags & IOP_RATE_LIMITED) * isc->read_stats.queued + 650 !!(isc->write_stats.state_flags & IOP_RATE_LIMITED) * isc->write_stats.queued /* + 651 !!(isc->trim_stats.state_flags & IOP_RATE_LIMITED) * isc->trim_stats.queued */ ; 652 pending <<= 16; 653 pending /= isc->periph->path->device->ccbq.total_openings; 654 655 isc->load = (pending + (isc->load << 13) - isc->load) >> 13; /* see above: 13 -> 16139 / 200/s = ~81s ~1 minute */ 656 657 isc->total_ticks++; 658 } 659 660 static void 661 cam_iosched_cl_init(struct control_loop *clp, struct cam_iosched_softc *isc) 662 { 663 664 clp->next_steer = sbinuptime(); 665 clp->softc = isc; 666 clp->steer_interval = SBT_1S * 5; /* Let's start out steering every 5s */ 667 clp->lolat = 5 * SBT_1MS; 668 clp->hilat = 15 * SBT_1MS; 669 clp->alpha = 20; /* Alpha == gain. 20 = .2 */ 670 clp->type = set_max; 671 } 672 673 static void 674 cam_iosched_cl_maybe_steer(struct control_loop *clp) 675 { 676 struct cam_iosched_softc *isc; 677 sbintime_t now, lat; 678 int old; 679 680 isc = clp->softc; 681 now = isc->last_time; 682 if (now < clp->next_steer) 683 return; 684 685 clp->next_steer = now + clp->steer_interval; 686 switch (clp->type) { 687 case set_max: 688 if (isc->write_stats.current != isc->write_stats.max) 689 printf("Steering write from %d kBps to %d kBps\n", 690 isc->write_stats.current, isc->write_stats.max); 691 isc->read_stats.current = isc->read_stats.max; 692 isc->write_stats.current = isc->write_stats.max; 693 isc->trim_stats.current = isc->trim_stats.max; 694 break; 695 case read_latency: 696 old = isc->write_stats.current; 697 lat = isc->read_stats.ema; 698 /* 699 * Simple PLL-like engine. Since we're steering to a range for 700 * the SP (set point) that makes things a little more 701 * complicated. In addition, we're not directly controlling our 702 * PV (process variable), the read latency, but instead are 703 * manipulating the write bandwidth limit for our MV 704 * (manipulation variable), analysis of this code gets a bit 705 * messy. Also, the MV is a very noisy control surface for read 706 * latency since it is affected by many hidden processes inside 707 * the device which change how responsive read latency will be 708 * in reaction to changes in write bandwidth. Unlike the classic 709 * boiler control PLL. this may result in over-steering while 710 * the SSD takes its time to react to the new, lower load. This 711 * is why we use a relatively low alpha of between .1 and .25 to 712 * compensate for this effect. At .1, it takes ~22 steering 713 * intervals to back off by a factor of 10. At .2 it only takes 714 * ~10. At .25 it only takes ~8. However some preliminary data 715 * from the SSD drives suggests a reasponse time in 10's of 716 * seconds before latency drops regardless of the new write 717 * rate. Careful observation will be required to tune this 718 * effectively. 719 * 720 * Also, when there's no read traffic, we jack up the write 721 * limit too regardless of the last read latency. 10 is 722 * somewhat arbitrary. 723 */ 724 if (lat < clp->lolat || isc->read_stats.total - clp->last_count < 10) 725 isc->write_stats.current = isc->write_stats.current * 726 (100 + clp->alpha) / 100; /* Scale up */ 727 else if (lat > clp->hilat) 728 isc->write_stats.current = isc->write_stats.current * 729 (100 - clp->alpha) / 100; /* Scale down */ 730 clp->last_count = isc->read_stats.total; 731 732 /* 733 * Even if we don't steer, per se, enforce the min/max limits as 734 * those may have changed. 735 */ 736 if (isc->write_stats.current < isc->write_stats.min) 737 isc->write_stats.current = isc->write_stats.min; 738 if (isc->write_stats.current > isc->write_stats.max) 739 isc->write_stats.current = isc->write_stats.max; 740 if (old != isc->write_stats.current && iosched_debug) 741 printf("Steering write from %d kBps to %d kBps due to latency of %jdus\n", 742 old, isc->write_stats.current, 743 (uintmax_t)((uint64_t)1000000 * (uint32_t)lat) >> 32); 744 break; 745 case cl_max: 746 break; 747 } 748 } 749 #endif 750 751 /* 752 * Trim or similar currently pending completion. Should only be set for 753 * those drivers wishing only one Trim active at a time. 754 */ 755 #define CAM_IOSCHED_FLAG_TRIM_ACTIVE (1ul << 0) 756 /* Callout active, and needs to be torn down */ 757 #define CAM_IOSCHED_FLAG_CALLOUT_ACTIVE (1ul << 1) 758 759 /* Periph drivers set these flags to indicate work */ 760 #define CAM_IOSCHED_FLAG_WORK_FLAGS ((0xffffu) << 16) 761 762 #ifdef CAM_IOSCHED_DYNAMIC 763 static void 764 cam_iosched_io_metric_update(struct cam_iosched_softc *isc, 765 sbintime_t sim_latency, const struct bio *bp); 766 #endif 767 768 static inline bool 769 cam_iosched_has_flagged_work(struct cam_iosched_softc *isc) 770 { 771 return !!(isc->flags & CAM_IOSCHED_FLAG_WORK_FLAGS); 772 } 773 774 static inline bool 775 cam_iosched_has_io(struct cam_iosched_softc *isc) 776 { 777 #ifdef CAM_IOSCHED_DYNAMIC 778 if (do_dynamic_iosched) { 779 struct bio *rbp = bioq_first(&isc->bio_queue); 780 struct bio *wbp = bioq_first(&isc->write_queue); 781 bool can_write = wbp != NULL && 782 cam_iosched_limiter_caniop(&isc->write_stats, wbp) == 0; 783 bool can_read = rbp != NULL && 784 cam_iosched_limiter_caniop(&isc->read_stats, rbp) == 0; 785 if (iosched_debug > 2) { 786 printf("can write %d: pending_writes %d max_writes %d\n", can_write, isc->write_stats.pending, isc->write_stats.max); 787 printf("can read %d: read_stats.pending %d max_reads %d\n", can_read, isc->read_stats.pending, isc->read_stats.max); 788 printf("Queued reads %d writes %d\n", isc->read_stats.queued, isc->write_stats.queued); 789 } 790 return can_read || can_write; 791 } 792 #endif 793 return bioq_first(&isc->bio_queue) != NULL; 794 } 795 796 static inline bool 797 cam_iosched_has_more_trim(struct cam_iosched_softc *isc) 798 { 799 struct bio *bp; 800 801 bp = bioq_first(&isc->trim_queue); 802 #ifdef CAM_IOSCHED_DYNAMIC 803 if (do_dynamic_iosched) { 804 /* 805 * If we're limiting trims, then defer action on trims 806 * for a bit. 807 */ 808 if (bp == NULL || cam_iosched_limiter_caniop(&isc->trim_stats, bp) != 0) 809 return false; 810 } 811 #endif 812 813 /* 814 * If we've set a trim_goal, then if we exceed that allow trims 815 * to be passed back to the driver. If we've also set a tick timeout 816 * allow trims back to the driver. Otherwise, don't allow trims yet. 817 */ 818 if (isc->trim_goal > 0) { 819 if (isc->queued_trims >= isc->trim_goal) 820 return true; 821 if (isc->queued_trims > 0 && 822 isc->trim_ticks > 0 && 823 ticks - isc->last_trim_tick > isc->trim_ticks) 824 return true; 825 return false; 826 } 827 828 /* NB: Should perhaps have a max trim active independent of I/O limiters */ 829 return !(isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) && bp != NULL; 830 } 831 832 #define cam_iosched_sort_queue(isc) ((isc)->sort_io_queue >= 0 ? \ 833 (isc)->sort_io_queue : cam_sort_io_queues) 834 835 static inline bool 836 cam_iosched_has_work(struct cam_iosched_softc *isc) 837 { 838 #ifdef CAM_IOSCHED_DYNAMIC 839 if (iosched_debug > 2) 840 printf("has work: %d %d %d\n", cam_iosched_has_io(isc), 841 cam_iosched_has_more_trim(isc), 842 cam_iosched_has_flagged_work(isc)); 843 #endif 844 845 return cam_iosched_has_io(isc) || 846 cam_iosched_has_more_trim(isc) || 847 cam_iosched_has_flagged_work(isc); 848 } 849 850 #ifdef CAM_IOSCHED_DYNAMIC 851 static void 852 cam_iosched_iop_stats_init(struct cam_iosched_softc *isc, struct iop_stats *ios) 853 { 854 855 ios->limiter = none; 856 ios->in = 0; 857 ios->max = ios->current = 300000; 858 ios->min = 1; 859 ios->out = 0; 860 ios->errs = 0; 861 ios->pending = 0; 862 ios->queued = 0; 863 ios->total = 0; 864 ios->ema = 0; 865 ios->emvar = 0; 866 ios->bad_latency = SBT_1S / 2; /* Default to 500ms */ 867 ios->softc = isc; 868 cam_iosched_limiter_init(ios); 869 } 870 871 static int 872 cam_iosched_limiter_sysctl(SYSCTL_HANDLER_ARGS) 873 { 874 char buf[16]; 875 struct iop_stats *ios; 876 struct cam_iosched_softc *isc; 877 int value, i, error; 878 const char *p; 879 880 ios = arg1; 881 isc = ios->softc; 882 value = ios->limiter; 883 if (value < none || value >= limiter_max) 884 p = "UNKNOWN"; 885 else 886 p = cam_iosched_limiter_names[value]; 887 888 strlcpy(buf, p, sizeof(buf)); 889 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 890 if (error != 0 || req->newptr == NULL) 891 return error; 892 893 cam_periph_lock(isc->periph); 894 895 for (i = none; i < limiter_max; i++) { 896 if (strcmp(buf, cam_iosched_limiter_names[i]) != 0) 897 continue; 898 ios->limiter = i; 899 error = cam_iosched_limiter_init(ios); 900 if (error != 0) { 901 ios->limiter = value; 902 cam_periph_unlock(isc->periph); 903 return error; 904 } 905 /* Note: disk load averate requires ticker to be always running */ 906 callout_reset(&isc->ticker, hz / isc->quanta, cam_iosched_ticker, isc); 907 isc->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE; 908 909 cam_periph_unlock(isc->periph); 910 return 0; 911 } 912 913 cam_periph_unlock(isc->periph); 914 return EINVAL; 915 } 916 917 static int 918 cam_iosched_control_type_sysctl(SYSCTL_HANDLER_ARGS) 919 { 920 char buf[16]; 921 struct control_loop *clp; 922 struct cam_iosched_softc *isc; 923 int value, i, error; 924 const char *p; 925 926 clp = arg1; 927 isc = clp->softc; 928 value = clp->type; 929 if (value < none || value >= cl_max) 930 p = "UNKNOWN"; 931 else 932 p = cam_iosched_control_type_names[value]; 933 934 strlcpy(buf, p, sizeof(buf)); 935 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 936 if (error != 0 || req->newptr == NULL) 937 return error; 938 939 for (i = set_max; i < cl_max; i++) { 940 if (strcmp(buf, cam_iosched_control_type_names[i]) != 0) 941 continue; 942 cam_periph_lock(isc->periph); 943 clp->type = i; 944 cam_periph_unlock(isc->periph); 945 return 0; 946 } 947 948 return EINVAL; 949 } 950 951 static int 952 cam_iosched_sbintime_sysctl(SYSCTL_HANDLER_ARGS) 953 { 954 char buf[16]; 955 sbintime_t value; 956 int error; 957 uint64_t us; 958 959 value = *(sbintime_t *)arg1; 960 us = (uint64_t)value / SBT_1US; 961 snprintf(buf, sizeof(buf), "%ju", (intmax_t)us); 962 error = sysctl_handle_string(oidp, buf, sizeof(buf), req); 963 if (error != 0 || req->newptr == NULL) 964 return error; 965 us = strtoul(buf, NULL, 10); 966 if (us == 0) 967 return EINVAL; 968 *(sbintime_t *)arg1 = us * SBT_1US; 969 return 0; 970 } 971 972 static int 973 cam_iosched_sysctl_latencies(SYSCTL_HANDLER_ARGS) 974 { 975 int i, error; 976 struct sbuf sb; 977 uint64_t *latencies; 978 979 latencies = arg1; 980 sbuf_new_for_sysctl(&sb, NULL, LAT_BUCKETS * 16, req); 981 982 for (i = 0; i < LAT_BUCKETS - 1; i++) 983 sbuf_printf(&sb, "%jd,", (intmax_t)latencies[i]); 984 sbuf_printf(&sb, "%jd", (intmax_t)latencies[LAT_BUCKETS - 1]); 985 error = sbuf_finish(&sb); 986 sbuf_delete(&sb); 987 988 return (error); 989 } 990 991 static int 992 cam_iosched_quanta_sysctl(SYSCTL_HANDLER_ARGS) 993 { 994 int *quanta; 995 int error, value; 996 997 quanta = (unsigned *)arg1; 998 value = *quanta; 999 1000 error = sysctl_handle_int(oidp, (int *)&value, 0, req); 1001 if ((error != 0) || (req->newptr == NULL)) 1002 return (error); 1003 1004 if (value < 1 || value > hz) 1005 return (EINVAL); 1006 1007 *quanta = value; 1008 1009 return (0); 1010 } 1011 1012 static void 1013 cam_iosched_iop_stats_sysctl_init(struct cam_iosched_softc *isc, struct iop_stats *ios, char *name) 1014 { 1015 struct sysctl_oid_list *n; 1016 struct sysctl_ctx_list *ctx; 1017 1018 ios->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx, 1019 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, name, 1020 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, name); 1021 n = SYSCTL_CHILDREN(ios->sysctl_tree); 1022 ctx = &ios->sysctl_ctx; 1023 1024 SYSCTL_ADD_UQUAD(ctx, n, 1025 OID_AUTO, "ema", CTLFLAG_RD, 1026 &ios->ema, 1027 "Fast Exponentially Weighted Moving Average"); 1028 SYSCTL_ADD_UQUAD(ctx, n, 1029 OID_AUTO, "emvar", CTLFLAG_RD, 1030 &ios->emvar, 1031 "Fast Exponentially Weighted Moving Variance"); 1032 1033 SYSCTL_ADD_INT(ctx, n, 1034 OID_AUTO, "pending", CTLFLAG_RD, 1035 &ios->pending, 0, 1036 "Instantaneous # of pending transactions"); 1037 SYSCTL_ADD_INT(ctx, n, 1038 OID_AUTO, "count", CTLFLAG_RD, 1039 &ios->total, 0, 1040 "# of transactions submitted to hardware"); 1041 SYSCTL_ADD_INT(ctx, n, 1042 OID_AUTO, "queued", CTLFLAG_RD, 1043 &ios->queued, 0, 1044 "# of transactions in the queue"); 1045 SYSCTL_ADD_INT(ctx, n, 1046 OID_AUTO, "in", CTLFLAG_RD, 1047 &ios->in, 0, 1048 "# of transactions queued to driver"); 1049 SYSCTL_ADD_INT(ctx, n, 1050 OID_AUTO, "out", CTLFLAG_RD, 1051 &ios->out, 0, 1052 "# of transactions completed (including with error)"); 1053 SYSCTL_ADD_INT(ctx, n, 1054 OID_AUTO, "errs", CTLFLAG_RD, 1055 &ios->errs, 0, 1056 "# of transactions completed with an error"); 1057 SYSCTL_ADD_U64(ctx, n, 1058 OID_AUTO, "too_long", CTLFLAG_RD, 1059 &ios->too_long, 0, 1060 "# of transactions completed took too long"); 1061 SYSCTL_ADD_PROC(ctx, n, 1062 OID_AUTO, "bad_latency", 1063 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, 1064 &ios->bad_latency, 0, cam_iosched_sbintime_sysctl, "A", 1065 "Threshold for counting transactions that took too long (in us)"); 1066 1067 SYSCTL_ADD_PROC(ctx, n, 1068 OID_AUTO, "limiter", 1069 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, 1070 ios, 0, cam_iosched_limiter_sysctl, "A", 1071 "Current limiting type."); 1072 SYSCTL_ADD_INT(ctx, n, 1073 OID_AUTO, "min", CTLFLAG_RW, 1074 &ios->min, 0, 1075 "min resource"); 1076 SYSCTL_ADD_INT(ctx, n, 1077 OID_AUTO, "max", CTLFLAG_RW, 1078 &ios->max, 0, 1079 "max resource"); 1080 SYSCTL_ADD_INT(ctx, n, 1081 OID_AUTO, "current", CTLFLAG_RW, 1082 &ios->current, 0, 1083 "current resource"); 1084 1085 SYSCTL_ADD_PROC(ctx, n, 1086 OID_AUTO, "latencies", 1087 CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, 1088 &ios->latencies, 0, 1089 cam_iosched_sysctl_latencies, "A", 1090 "Array of latencies, a geometric progresson from\n" 1091 "kern.cam.iosched.bucket_base_us with a ratio of\n" 1092 "kern.cam.iosched.bucket_ration / 100 from one to\n" 1093 "the next. By default 20 steps from 20us to 10.485s\n" 1094 "by doubling."); 1095 1096 } 1097 1098 static void 1099 cam_iosched_iop_stats_fini(struct iop_stats *ios) 1100 { 1101 if (ios->sysctl_tree) 1102 if (sysctl_ctx_free(&ios->sysctl_ctx) != 0) 1103 printf("can't remove iosched sysctl stats context\n"); 1104 } 1105 1106 static void 1107 cam_iosched_cl_sysctl_init(struct cam_iosched_softc *isc) 1108 { 1109 struct sysctl_oid_list *n; 1110 struct sysctl_ctx_list *ctx; 1111 struct control_loop *clp; 1112 1113 clp = &isc->cl; 1114 clp->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx, 1115 SYSCTL_CHILDREN(isc->sysctl_tree), OID_AUTO, "control", 1116 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "Control loop info"); 1117 n = SYSCTL_CHILDREN(clp->sysctl_tree); 1118 ctx = &clp->sysctl_ctx; 1119 1120 SYSCTL_ADD_PROC(ctx, n, 1121 OID_AUTO, "type", 1122 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, 1123 clp, 0, cam_iosched_control_type_sysctl, "A", 1124 "Control loop algorithm"); 1125 SYSCTL_ADD_PROC(ctx, n, 1126 OID_AUTO, "steer_interval", 1127 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, 1128 &clp->steer_interval, 0, cam_iosched_sbintime_sysctl, "A", 1129 "How often to steer (in us)"); 1130 SYSCTL_ADD_PROC(ctx, n, 1131 OID_AUTO, "lolat", 1132 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, 1133 &clp->lolat, 0, cam_iosched_sbintime_sysctl, "A", 1134 "Low water mark for Latency (in us)"); 1135 SYSCTL_ADD_PROC(ctx, n, 1136 OID_AUTO, "hilat", 1137 CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE, 1138 &clp->hilat, 0, cam_iosched_sbintime_sysctl, "A", 1139 "Hi water mark for Latency (in us)"); 1140 SYSCTL_ADD_INT(ctx, n, 1141 OID_AUTO, "alpha", CTLFLAG_RW, 1142 &clp->alpha, 0, 1143 "Alpha for PLL (x100) aka gain"); 1144 } 1145 1146 static void 1147 cam_iosched_cl_sysctl_fini(struct control_loop *clp) 1148 { 1149 if (clp->sysctl_tree) 1150 if (sysctl_ctx_free(&clp->sysctl_ctx) != 0) 1151 printf("can't remove iosched sysctl control loop context\n"); 1152 } 1153 #endif 1154 1155 /* 1156 * Allocate the iosched structure. This also insulates callers from knowing 1157 * sizeof struct cam_iosched_softc. 1158 */ 1159 int 1160 cam_iosched_init(struct cam_iosched_softc **iscp, struct cam_periph *periph, 1161 const struct disk *dp, cam_iosched_schedule_t schedfnc) 1162 { 1163 1164 *iscp = malloc(sizeof(**iscp), M_CAMSCHED, M_NOWAIT | M_ZERO); 1165 if (*iscp == NULL) 1166 return ENOMEM; 1167 (*iscp)->disk = dp; 1168 (*iscp)->schedfnc = schedfnc; 1169 #ifdef CAM_IOSCHED_DYNAMIC 1170 if (iosched_debug) 1171 printf("CAM IOSCHEDULER Allocating entry at %p\n", *iscp); 1172 #endif 1173 (*iscp)->sort_io_queue = -1; 1174 bioq_init(&(*iscp)->bio_queue); 1175 bioq_init(&(*iscp)->trim_queue); 1176 #ifdef CAM_IOSCHED_DYNAMIC 1177 if (do_dynamic_iosched) { 1178 bioq_init(&(*iscp)->write_queue); 1179 (*iscp)->read_bias = default_read_bias; 1180 (*iscp)->current_read_bias = 0; 1181 (*iscp)->quanta = min(hz, 200); 1182 cam_iosched_iop_stats_init(*iscp, &(*iscp)->read_stats); 1183 cam_iosched_iop_stats_init(*iscp, &(*iscp)->write_stats); 1184 cam_iosched_iop_stats_init(*iscp, &(*iscp)->trim_stats); 1185 (*iscp)->trim_stats.max = 1; /* Trims are special: one at a time for now */ 1186 (*iscp)->last_time = sbinuptime(); 1187 callout_init_mtx(&(*iscp)->ticker, cam_periph_mtx(periph), 0); 1188 (*iscp)->periph = periph; 1189 cam_iosched_cl_init(&(*iscp)->cl, *iscp); 1190 callout_reset(&(*iscp)->ticker, hz / (*iscp)->quanta, cam_iosched_ticker, *iscp); 1191 (*iscp)->flags |= CAM_IOSCHED_FLAG_CALLOUT_ACTIVE; 1192 } 1193 #endif 1194 1195 return 0; 1196 } 1197 1198 /* 1199 * Reclaim all used resources. This assumes that other folks have 1200 * drained the requests in the hardware. Maybe an unwise assumption. 1201 */ 1202 void 1203 cam_iosched_fini(struct cam_iosched_softc *isc) 1204 { 1205 if (isc) { 1206 cam_iosched_flush(isc, NULL, ENXIO); 1207 #ifdef CAM_IOSCHED_DYNAMIC 1208 cam_iosched_iop_stats_fini(&isc->read_stats); 1209 cam_iosched_iop_stats_fini(&isc->write_stats); 1210 cam_iosched_iop_stats_fini(&isc->trim_stats); 1211 cam_iosched_cl_sysctl_fini(&isc->cl); 1212 if (isc->sysctl_tree) 1213 if (sysctl_ctx_free(&isc->sysctl_ctx) != 0) 1214 printf("can't remove iosched sysctl stats context\n"); 1215 if (isc->flags & CAM_IOSCHED_FLAG_CALLOUT_ACTIVE) { 1216 callout_drain(&isc->ticker); 1217 isc->flags &= ~ CAM_IOSCHED_FLAG_CALLOUT_ACTIVE; 1218 } 1219 #endif 1220 free(isc, M_CAMSCHED); 1221 } 1222 } 1223 1224 /* 1225 * After we're sure we're attaching a device, go ahead and add 1226 * hooks for any sysctl we may wish to honor. 1227 */ 1228 void cam_iosched_sysctl_init(struct cam_iosched_softc *isc, 1229 struct sysctl_ctx_list *ctx, struct sysctl_oid *node) 1230 { 1231 struct sysctl_oid_list *n; 1232 1233 n = SYSCTL_CHILDREN(node); 1234 SYSCTL_ADD_INT(ctx, n, 1235 OID_AUTO, "sort_io_queue", CTLFLAG_RW | CTLFLAG_MPSAFE, 1236 &isc->sort_io_queue, 0, 1237 "Sort IO queue to try and optimise disk access patterns"); 1238 SYSCTL_ADD_INT(ctx, n, 1239 OID_AUTO, "trim_goal", CTLFLAG_RW, 1240 &isc->trim_goal, 0, 1241 "Number of trims to try to accumulate before sending to hardware"); 1242 SYSCTL_ADD_INT(ctx, n, 1243 OID_AUTO, "trim_ticks", CTLFLAG_RW, 1244 &isc->trim_goal, 0, 1245 "IO Schedul qaunta to hold back trims for when accumulating"); 1246 1247 #ifdef CAM_IOSCHED_DYNAMIC 1248 if (!do_dynamic_iosched) 1249 return; 1250 1251 isc->sysctl_tree = SYSCTL_ADD_NODE(&isc->sysctl_ctx, 1252 SYSCTL_CHILDREN(node), OID_AUTO, "iosched", 1253 CTLFLAG_RD | CTLFLAG_MPSAFE, 0, "I/O scheduler statistics"); 1254 n = SYSCTL_CHILDREN(isc->sysctl_tree); 1255 ctx = &isc->sysctl_ctx; 1256 1257 cam_iosched_iop_stats_sysctl_init(isc, &isc->read_stats, "read"); 1258 cam_iosched_iop_stats_sysctl_init(isc, &isc->write_stats, "write"); 1259 cam_iosched_iop_stats_sysctl_init(isc, &isc->trim_stats, "trim"); 1260 cam_iosched_cl_sysctl_init(isc); 1261 1262 SYSCTL_ADD_INT(ctx, n, 1263 OID_AUTO, "read_bias", CTLFLAG_RW, 1264 &isc->read_bias, default_read_bias, 1265 "How biased towards read should we be independent of limits"); 1266 1267 SYSCTL_ADD_PROC(ctx, n, 1268 OID_AUTO, "quanta", CTLTYPE_UINT | CTLFLAG_RW | CTLFLAG_MPSAFE, 1269 &isc->quanta, 0, cam_iosched_quanta_sysctl, "I", 1270 "How many quanta per second do we slice the I/O up into"); 1271 1272 SYSCTL_ADD_INT(ctx, n, 1273 OID_AUTO, "total_ticks", CTLFLAG_RD, 1274 &isc->total_ticks, 0, 1275 "Total number of ticks we've done"); 1276 1277 SYSCTL_ADD_INT(ctx, n, 1278 OID_AUTO, "load", CTLFLAG_RD, 1279 &isc->load, 0, 1280 "scaled load average / 100"); 1281 1282 SYSCTL_ADD_U64(ctx, n, 1283 OID_AUTO, "latency_trigger", CTLFLAG_RW, 1284 &isc->max_lat, 0, 1285 "Latency treshold to trigger callbacks"); 1286 #endif 1287 } 1288 1289 void 1290 cam_iosched_set_latfcn(struct cam_iosched_softc *isc, 1291 cam_iosched_latfcn_t fnp, void *argp) 1292 { 1293 #ifdef CAM_IOSCHED_DYNAMIC 1294 isc->latfcn = fnp; 1295 isc->latarg = argp; 1296 #endif 1297 } 1298 1299 /* 1300 * Client drivers can set two parameters. "goal" is the number of BIO_DELETEs 1301 * that will be queued up before iosched will "release" the trims to the client 1302 * driver to wo with what they will (usually combine as many as possible). If we 1303 * don't get this many, after trim_ticks we'll submit the I/O anyway with 1304 * whatever we have. We do need an I/O of some kind of to clock the deferred 1305 * trims out to disk. Since we will eventually get a write for the super block 1306 * or something before we shutdown, the trims will complete. To be safe, when a 1307 * BIO_FLUSH is presented to the iosched work queue, we set the ticks time far 1308 * enough in the past so we'll present the BIO_DELETEs to the client driver. 1309 * There might be a race if no BIO_DELETESs were queued, a BIO_FLUSH comes in 1310 * and then a BIO_DELETE is sent down. No know client does this, and there's 1311 * already a race between an ordered BIO_FLUSH and any BIO_DELETEs in flight, 1312 * but no client depends on the ordering being honored. 1313 * 1314 * XXX I'm not sure what the interaction between UFS direct BIOs and the BUF 1315 * flushing on shutdown. I think there's bufs that would be dependent on the BIO 1316 * finishing to write out at least metadata, so we'll be fine. To be safe, keep 1317 * the number of ticks low (less than maybe 10s) to avoid shutdown races. 1318 */ 1319 1320 void 1321 cam_iosched_set_trim_goal(struct cam_iosched_softc *isc, int goal) 1322 { 1323 1324 isc->trim_goal = goal; 1325 } 1326 1327 void 1328 cam_iosched_set_trim_ticks(struct cam_iosched_softc *isc, int trim_ticks) 1329 { 1330 1331 isc->trim_ticks = trim_ticks; 1332 } 1333 1334 /* 1335 * Flush outstanding I/O. Consumers of this library don't know all the 1336 * queues we may keep, so this allows all I/O to be flushed in one 1337 * convenient call. 1338 */ 1339 void 1340 cam_iosched_flush(struct cam_iosched_softc *isc, struct devstat *stp, int err) 1341 { 1342 bioq_flush(&isc->bio_queue, stp, err); 1343 bioq_flush(&isc->trim_queue, stp, err); 1344 #ifdef CAM_IOSCHED_DYNAMIC 1345 if (do_dynamic_iosched) 1346 bioq_flush(&isc->write_queue, stp, err); 1347 #endif 1348 } 1349 1350 #ifdef CAM_IOSCHED_DYNAMIC 1351 static struct bio * 1352 cam_iosched_get_write(struct cam_iosched_softc *isc) 1353 { 1354 struct bio *bp; 1355 1356 /* 1357 * We control the write rate by controlling how many requests we send 1358 * down to the drive at any one time. Fewer requests limits the 1359 * effects of both starvation when the requests take a while and write 1360 * amplification when each request is causing more than one write to 1361 * the NAND media. Limiting the queue depth like this will also limit 1362 * the write throughput and give and reads that want to compete to 1363 * compete unfairly. 1364 */ 1365 bp = bioq_first(&isc->write_queue); 1366 if (bp == NULL) { 1367 if (iosched_debug > 3) 1368 printf("No writes present in write_queue\n"); 1369 return NULL; 1370 } 1371 1372 /* 1373 * If pending read, prefer that based on current read bias 1374 * setting. 1375 */ 1376 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) { 1377 if (iosched_debug) 1378 printf( 1379 "Reads present and current_read_bias is %d queued " 1380 "writes %d queued reads %d\n", 1381 isc->current_read_bias, isc->write_stats.queued, 1382 isc->read_stats.queued); 1383 isc->current_read_bias--; 1384 /* We're not limiting writes, per se, just doing reads first */ 1385 return NULL; 1386 } 1387 1388 /* 1389 * See if our current limiter allows this I/O. 1390 */ 1391 if (cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) { 1392 if (iosched_debug) 1393 printf("Can't write because limiter says no.\n"); 1394 isc->write_stats.state_flags |= IOP_RATE_LIMITED; 1395 return NULL; 1396 } 1397 1398 /* 1399 * Let's do this: We've passed all the gates and we're a go 1400 * to schedule the I/O in the SIM. 1401 */ 1402 isc->current_read_bias = isc->read_bias; 1403 bioq_remove(&isc->write_queue, bp); 1404 if (bp->bio_cmd == BIO_WRITE) { 1405 isc->write_stats.queued--; 1406 isc->write_stats.total++; 1407 isc->write_stats.pending++; 1408 } 1409 if (iosched_debug > 9) 1410 printf("HWQ : %p %#x\n", bp, bp->bio_cmd); 1411 isc->write_stats.state_flags &= ~IOP_RATE_LIMITED; 1412 return bp; 1413 } 1414 #endif 1415 1416 /* 1417 * Put back a trim that you weren't able to actually schedule this time. 1418 */ 1419 void 1420 cam_iosched_put_back_trim(struct cam_iosched_softc *isc, struct bio *bp) 1421 { 1422 bioq_insert_head(&isc->trim_queue, bp); 1423 if (isc->queued_trims == 0) 1424 isc->last_trim_tick = ticks; 1425 isc->queued_trims++; 1426 #ifdef CAM_IOSCHED_DYNAMIC 1427 isc->trim_stats.queued++; 1428 isc->trim_stats.total--; /* since we put it back, don't double count */ 1429 isc->trim_stats.pending--; 1430 #endif 1431 } 1432 1433 /* 1434 * gets the next trim from the trim queue. 1435 * 1436 * Assumes we're called with the periph lock held. It removes this 1437 * trim from the queue and the device must explicitly reinsert it 1438 * should the need arise. 1439 */ 1440 struct bio * 1441 cam_iosched_next_trim(struct cam_iosched_softc *isc) 1442 { 1443 struct bio *bp; 1444 1445 bp = bioq_first(&isc->trim_queue); 1446 if (bp == NULL) 1447 return NULL; 1448 bioq_remove(&isc->trim_queue, bp); 1449 isc->queued_trims--; 1450 isc->last_trim_tick = ticks; /* Reset the tick timer when we take trims */ 1451 #ifdef CAM_IOSCHED_DYNAMIC 1452 isc->trim_stats.queued--; 1453 isc->trim_stats.total++; 1454 isc->trim_stats.pending++; 1455 #endif 1456 return bp; 1457 } 1458 1459 /* 1460 * gets an available trim from the trim queue, if there's no trim 1461 * already pending. It removes this trim from the queue and the device 1462 * must explicitly reinsert it should the need arise. 1463 * 1464 * Assumes we're called with the periph lock held. 1465 */ 1466 struct bio * 1467 cam_iosched_get_trim(struct cam_iosched_softc *isc) 1468 { 1469 #ifdef CAM_IOSCHED_DYNAMIC 1470 struct bio *bp; 1471 #endif 1472 1473 if (!cam_iosched_has_more_trim(isc)) 1474 return NULL; 1475 #ifdef CAM_IOSCHED_DYNAMIC 1476 bp = bioq_first(&isc->trim_queue); 1477 if (bp == NULL) 1478 return NULL; 1479 1480 /* 1481 * If pending read, prefer that based on current read bias setting. The 1482 * read bias is shared for both writes and TRIMs, but on TRIMs the bias 1483 * is for a combined TRIM not a single TRIM request that's come in. 1484 */ 1485 if (do_dynamic_iosched) { 1486 if (bioq_first(&isc->bio_queue) && isc->current_read_bias) { 1487 if (iosched_debug) 1488 printf("Reads present and current_read_bias is %d" 1489 " queued trims %d queued reads %d\n", 1490 isc->current_read_bias, isc->trim_stats.queued, 1491 isc->read_stats.queued); 1492 isc->current_read_bias--; 1493 /* We're not limiting TRIMS, per se, just doing reads first */ 1494 return NULL; 1495 } 1496 /* 1497 * We're going to do a trim, so reset the bias. 1498 */ 1499 isc->current_read_bias = isc->read_bias; 1500 } 1501 1502 /* 1503 * See if our current limiter allows this I/O. Because we only call this 1504 * here, and not in next_trim, the 'bandwidth' limits for trims won't 1505 * work, while the iops or max queued limits will work. It's tricky 1506 * because we want the limits to be from the perspective of the 1507 * "commands sent to the device." To make iops work, we need to check 1508 * only here (since we want all the ops we combine to count as one). To 1509 * make bw limits work, we'd need to check in next_trim, but that would 1510 * have the effect of limiting the iops as seen from the upper layers. 1511 */ 1512 if (cam_iosched_limiter_iop(&isc->trim_stats, bp) != 0) { 1513 if (iosched_debug) 1514 printf("Can't trim because limiter says no.\n"); 1515 isc->trim_stats.state_flags |= IOP_RATE_LIMITED; 1516 return NULL; 1517 } 1518 isc->current_read_bias = isc->read_bias; 1519 isc->trim_stats.state_flags &= ~IOP_RATE_LIMITED; 1520 /* cam_iosched_next_trim below keeps proper book */ 1521 #endif 1522 return cam_iosched_next_trim(isc); 1523 } 1524 1525 1526 #ifdef CAM_IOSCHED_DYNAMIC 1527 static struct bio * 1528 bio_next(struct bio *bp) 1529 { 1530 bp = TAILQ_NEXT(bp, bio_queue); 1531 /* 1532 * After the first commands, the ordered bit terminates 1533 * our search because BIO_ORDERED acts like a barrier. 1534 */ 1535 if (bp == NULL || bp->bio_flags & BIO_ORDERED) 1536 return NULL; 1537 return bp; 1538 } 1539 1540 static bool 1541 cam_iosched_rate_limited(struct iop_stats *ios) 1542 { 1543 return ios->state_flags & IOP_RATE_LIMITED; 1544 } 1545 #endif 1546 1547 /* 1548 * Determine what the next bit of work to do is for the periph. The 1549 * default implementation looks to see if we have trims to do, but no 1550 * trims outstanding. If so, we do that. Otherwise we see if we have 1551 * other work. If we do, then we do that. Otherwise why were we called? 1552 */ 1553 struct bio * 1554 cam_iosched_next_bio(struct cam_iosched_softc *isc) 1555 { 1556 struct bio *bp; 1557 1558 /* 1559 * See if we have a trim that can be scheduled. We can only send one 1560 * at a time down, so this takes that into account. 1561 * 1562 * XXX newer TRIM commands are queueable. Revisit this when we 1563 * implement them. 1564 */ 1565 if ((bp = cam_iosched_get_trim(isc)) != NULL) 1566 return bp; 1567 1568 #ifdef CAM_IOSCHED_DYNAMIC 1569 /* 1570 * See if we have any pending writes, room in the queue for them, 1571 * and no pending reads (unless we've scheduled too many). 1572 * if so, those are next. 1573 */ 1574 if (do_dynamic_iosched) { 1575 if ((bp = cam_iosched_get_write(isc)) != NULL) 1576 return bp; 1577 } 1578 #endif 1579 /* 1580 * next, see if there's other, normal I/O waiting. If so return that. 1581 */ 1582 #ifdef CAM_IOSCHED_DYNAMIC 1583 if (do_dynamic_iosched) { 1584 for (bp = bioq_first(&isc->bio_queue); bp != NULL; 1585 bp = bio_next(bp)) { 1586 /* 1587 * For the dynamic scheduler with a read bias, bio_queue 1588 * is only for reads. However, without one, all 1589 * operations are queued. Enforce limits here for any 1590 * operation we find here. 1591 */ 1592 if (bp->bio_cmd == BIO_READ) { 1593 if (cam_iosched_rate_limited(&isc->read_stats) || 1594 cam_iosched_limiter_iop(&isc->read_stats, bp) != 0) { 1595 isc->read_stats.state_flags |= IOP_RATE_LIMITED; 1596 continue; 1597 } 1598 isc->read_stats.state_flags &= ~IOP_RATE_LIMITED; 1599 } 1600 /* 1601 * There can only be write requests on the queue when 1602 * the read bias is 0, but we need to process them 1603 * here. We do not assert for read bias == 0, however, 1604 * since it is dynamic and we can have WRITE operations 1605 * in the queue after we transition from 0 to non-zero. 1606 */ 1607 if (bp->bio_cmd == BIO_WRITE) { 1608 if (cam_iosched_rate_limited(&isc->write_stats) || 1609 cam_iosched_limiter_iop(&isc->write_stats, bp) != 0) { 1610 isc->write_stats.state_flags |= IOP_RATE_LIMITED; 1611 continue; 1612 } 1613 isc->write_stats.state_flags &= ~IOP_RATE_LIMITED; 1614 } 1615 /* 1616 * here we know we have a bp that's != NULL, that's not rate limited 1617 * and can be the next I/O. 1618 */ 1619 break; 1620 } 1621 } else 1622 #endif 1623 bp = bioq_first(&isc->bio_queue); 1624 1625 if (bp == NULL) 1626 return (NULL); 1627 bioq_remove(&isc->bio_queue, bp); 1628 #ifdef CAM_IOSCHED_DYNAMIC 1629 if (do_dynamic_iosched) { 1630 if (bp->bio_cmd == BIO_READ) { 1631 isc->read_stats.queued--; 1632 isc->read_stats.total++; 1633 isc->read_stats.pending++; 1634 } else if (bp->bio_cmd == BIO_WRITE) { 1635 isc->write_stats.queued--; 1636 isc->write_stats.total++; 1637 isc->write_stats.pending++; 1638 } 1639 } 1640 if (iosched_debug > 9) 1641 printf("HWQ : %p %#x\n", bp, bp->bio_cmd); 1642 #endif 1643 return bp; 1644 } 1645 1646 /* 1647 * Driver has been given some work to do by the block layer. Tell the 1648 * scheduler about it and have it queue the work up. The scheduler module 1649 * will then return the currently most useful bit of work later, possibly 1650 * deferring work for various reasons. 1651 */ 1652 void 1653 cam_iosched_queue_work(struct cam_iosched_softc *isc, struct bio *bp) 1654 { 1655 1656 /* 1657 * A BIO_SPEEDUP from the upper layers means that they have a block 1658 * shortage. At the present, this is only sent when we're trying to 1659 * allocate blocks, but have a shortage before giving up. bio_length is 1660 * the size of their shortage. We will complete just enough BIO_DELETEs 1661 * in the queue to satisfy the need. If bio_length is 0, we'll complete 1662 * them all. This allows the scheduler to delay BIO_DELETEs to improve 1663 * read/write performance without worrying about the upper layers. When 1664 * it's possibly a problem, we respond by pretending the BIO_DELETEs 1665 * just worked. We can't do anything about the BIO_DELETEs in the 1666 * hardware, though. We have to wait for them to complete. 1667 */ 1668 if (bp->bio_cmd == BIO_SPEEDUP) { 1669 off_t len; 1670 struct bio *nbp; 1671 1672 len = 0; 1673 while (bioq_first(&isc->trim_queue) && 1674 (bp->bio_length == 0 || len < bp->bio_length)) { 1675 nbp = bioq_takefirst(&isc->trim_queue); 1676 len += nbp->bio_length; 1677 nbp->bio_error = 0; 1678 biodone(nbp); 1679 } 1680 if (bp->bio_length > 0) { 1681 if (bp->bio_length > len) 1682 bp->bio_resid = bp->bio_length - len; 1683 else 1684 bp->bio_resid = 0; 1685 } 1686 bp->bio_error = 0; 1687 biodone(bp); 1688 return; 1689 } 1690 1691 /* 1692 * If we get a BIO_FLUSH, and we're doing delayed BIO_DELETEs then we 1693 * set the last tick time to one less than the current ticks minus the 1694 * delay to force the BIO_DELETEs to be presented to the client driver. 1695 */ 1696 if (bp->bio_cmd == BIO_FLUSH && isc->trim_ticks > 0) 1697 isc->last_trim_tick = ticks - isc->trim_ticks - 1; 1698 1699 /* 1700 * Put all trims on the trim queue. Otherwise put the work on the bio 1701 * queue. 1702 */ 1703 if (bp->bio_cmd == BIO_DELETE) { 1704 bioq_insert_tail(&isc->trim_queue, bp); 1705 if (isc->queued_trims == 0) 1706 isc->last_trim_tick = ticks; 1707 isc->queued_trims++; 1708 #ifdef CAM_IOSCHED_DYNAMIC 1709 isc->trim_stats.in++; 1710 isc->trim_stats.queued++; 1711 #endif 1712 } 1713 #ifdef CAM_IOSCHED_DYNAMIC 1714 else if (do_dynamic_iosched && isc->read_bias != 0 && 1715 (bp->bio_cmd != BIO_READ)) { 1716 if (cam_iosched_sort_queue(isc)) 1717 bioq_disksort(&isc->write_queue, bp); 1718 else 1719 bioq_insert_tail(&isc->write_queue, bp); 1720 if (iosched_debug > 9) 1721 printf("Qw : %p %#x\n", bp, bp->bio_cmd); 1722 if (bp->bio_cmd == BIO_WRITE) { 1723 isc->write_stats.in++; 1724 isc->write_stats.queued++; 1725 } 1726 } 1727 #endif 1728 else { 1729 if (cam_iosched_sort_queue(isc)) 1730 bioq_disksort(&isc->bio_queue, bp); 1731 else 1732 bioq_insert_tail(&isc->bio_queue, bp); 1733 #ifdef CAM_IOSCHED_DYNAMIC 1734 if (iosched_debug > 9) 1735 printf("Qr : %p %#x\n", bp, bp->bio_cmd); 1736 if (bp->bio_cmd == BIO_READ) { 1737 isc->read_stats.in++; 1738 isc->read_stats.queued++; 1739 } else if (bp->bio_cmd == BIO_WRITE) { 1740 isc->write_stats.in++; 1741 isc->write_stats.queued++; 1742 } 1743 #endif 1744 } 1745 } 1746 1747 /* 1748 * If we have work, get it scheduled. Called with the periph lock held. 1749 */ 1750 void 1751 cam_iosched_schedule(struct cam_iosched_softc *isc, struct cam_periph *periph) 1752 { 1753 1754 if (cam_iosched_has_work(isc)) 1755 xpt_schedule(periph, CAM_PRIORITY_NORMAL); 1756 } 1757 1758 /* 1759 * Complete a trim request. Mark that we no longer have one in flight. 1760 */ 1761 void 1762 cam_iosched_trim_done(struct cam_iosched_softc *isc) 1763 { 1764 1765 isc->flags &= ~CAM_IOSCHED_FLAG_TRIM_ACTIVE; 1766 } 1767 1768 /* 1769 * Complete a bio. Called before we release the ccb with xpt_release_ccb so we 1770 * might use notes in the ccb for statistics. 1771 */ 1772 int 1773 cam_iosched_bio_complete(struct cam_iosched_softc *isc, struct bio *bp, 1774 union ccb *done_ccb) 1775 { 1776 int retval = 0; 1777 #ifdef CAM_IOSCHED_DYNAMIC 1778 if (!do_dynamic_iosched) 1779 return retval; 1780 1781 if (iosched_debug > 10) 1782 printf("done: %p %#x\n", bp, bp->bio_cmd); 1783 if (bp->bio_cmd == BIO_WRITE) { 1784 retval = cam_iosched_limiter_iodone(&isc->write_stats, bp); 1785 if ((bp->bio_flags & BIO_ERROR) != 0) 1786 isc->write_stats.errs++; 1787 isc->write_stats.out++; 1788 isc->write_stats.pending--; 1789 } else if (bp->bio_cmd == BIO_READ) { 1790 retval = cam_iosched_limiter_iodone(&isc->read_stats, bp); 1791 if ((bp->bio_flags & BIO_ERROR) != 0) 1792 isc->read_stats.errs++; 1793 isc->read_stats.out++; 1794 isc->read_stats.pending--; 1795 } else if (bp->bio_cmd == BIO_DELETE) { 1796 if ((bp->bio_flags & BIO_ERROR) != 0) 1797 isc->trim_stats.errs++; 1798 isc->trim_stats.out++; 1799 isc->trim_stats.pending--; 1800 } else if (bp->bio_cmd != BIO_FLUSH) { 1801 if (iosched_debug) 1802 printf("Completing command with bio_cmd == %#x\n", bp->bio_cmd); 1803 } 1804 1805 if ((bp->bio_flags & BIO_ERROR) == 0 && done_ccb != NULL && 1806 (done_ccb->ccb_h.status & CAM_QOS_VALID) != 0) { 1807 sbintime_t sim_latency; 1808 1809 sim_latency = cam_iosched_sbintime_t(done_ccb->ccb_h.qos.periph_data); 1810 1811 cam_iosched_io_metric_update(isc, sim_latency, bp); 1812 1813 /* 1814 * Debugging code: allow callbacks to the periph driver when latency max 1815 * is exceeded. This can be useful for triggering external debugging actions. 1816 */ 1817 if (isc->latfcn && isc->max_lat != 0 && sim_latency > isc->max_lat) 1818 isc->latfcn(isc->latarg, sim_latency, bp); 1819 } 1820 #endif 1821 return retval; 1822 } 1823 1824 /* 1825 * Tell the io scheduler that you've pushed a trim down into the sim. 1826 * This also tells the I/O scheduler not to push any more trims down, so 1827 * some periphs do not call it if they can cope with multiple trims in flight. 1828 */ 1829 void 1830 cam_iosched_submit_trim(struct cam_iosched_softc *isc) 1831 { 1832 1833 isc->flags |= CAM_IOSCHED_FLAG_TRIM_ACTIVE; 1834 } 1835 1836 /* 1837 * Change the sorting policy hint for I/O transactions for this device. 1838 */ 1839 void 1840 cam_iosched_set_sort_queue(struct cam_iosched_softc *isc, int val) 1841 { 1842 1843 isc->sort_io_queue = val; 1844 } 1845 1846 int 1847 cam_iosched_has_work_flags(struct cam_iosched_softc *isc, uint32_t flags) 1848 { 1849 return isc->flags & flags; 1850 } 1851 1852 void 1853 cam_iosched_set_work_flags(struct cam_iosched_softc *isc, uint32_t flags) 1854 { 1855 isc->flags |= flags; 1856 } 1857 1858 void 1859 cam_iosched_clr_work_flags(struct cam_iosched_softc *isc, uint32_t flags) 1860 { 1861 isc->flags &= ~flags; 1862 } 1863 1864 #ifdef CAM_IOSCHED_DYNAMIC 1865 /* 1866 * After the method presented in Jack Crenshaw's 1998 article "Integer 1867 * Square Roots," reprinted at 1868 * http://www.embedded.com/electronics-blogs/programmer-s-toolbox/4219659/Integer-Square-Roots 1869 * and well worth the read. Briefly, we find the power of 4 that's the 1870 * largest smaller than val. We then check each smaller power of 4 to 1871 * see if val is still bigger. The right shifts at each step divide 1872 * the result by 2 which after successive application winds up 1873 * accumulating the right answer. It could also have been accumulated 1874 * using a separate root counter, but this code is smaller and faster 1875 * than that method. This method is also integer size invariant. 1876 * It returns floor(sqrt((float)val)), or the largest integer less than 1877 * or equal to the square root. 1878 */ 1879 static uint64_t 1880 isqrt64(uint64_t val) 1881 { 1882 uint64_t res = 0; 1883 uint64_t bit = 1ULL << (sizeof(uint64_t) * NBBY - 2); 1884 1885 /* 1886 * Find the largest power of 4 smaller than val. 1887 */ 1888 while (bit > val) 1889 bit >>= 2; 1890 1891 /* 1892 * Accumulate the answer, one bit at a time (we keep moving 1893 * them over since 2 is the square root of 4 and we test 1894 * powers of 4). We accumulate where we find the bit, but 1895 * the successive shifts land the bit in the right place 1896 * by the end. 1897 */ 1898 while (bit != 0) { 1899 if (val >= res + bit) { 1900 val -= res + bit; 1901 res = (res >> 1) + bit; 1902 } else 1903 res >>= 1; 1904 bit >>= 2; 1905 } 1906 1907 return res; 1908 } 1909 1910 static sbintime_t latencies[LAT_BUCKETS - 1] = { 1911 BUCKET_BASE << 0, /* 20us */ 1912 BUCKET_BASE << 1, 1913 BUCKET_BASE << 2, 1914 BUCKET_BASE << 3, 1915 BUCKET_BASE << 4, 1916 BUCKET_BASE << 5, 1917 BUCKET_BASE << 6, 1918 BUCKET_BASE << 7, 1919 BUCKET_BASE << 8, 1920 BUCKET_BASE << 9, 1921 BUCKET_BASE << 10, 1922 BUCKET_BASE << 11, 1923 BUCKET_BASE << 12, 1924 BUCKET_BASE << 13, 1925 BUCKET_BASE << 14, 1926 BUCKET_BASE << 15, 1927 BUCKET_BASE << 16, 1928 BUCKET_BASE << 17, 1929 BUCKET_BASE << 18 /* 5,242,880us */ 1930 }; 1931 1932 #define CAM_IOSCHED_DEVD_MSG_SIZE 256 1933 1934 static void 1935 cam_iosched_devctl_outlier(struct iop_stats *iop, sbintime_t sim_latency, 1936 const struct bio *bp) 1937 { 1938 daddr_t lba = bp->bio_pblkno; 1939 daddr_t cnt = bp->bio_bcount / iop->softc->disk->d_sectorsize; 1940 char *sbmsg; 1941 struct sbuf sb; 1942 1943 sbmsg = malloc(CAM_IOSCHED_DEVD_MSG_SIZE, M_CAMSCHED, M_NOWAIT); 1944 if (sbmsg == NULL) 1945 return; 1946 sbuf_new(&sb, sbmsg, CAM_IOSCHED_DEVD_MSG_SIZE, SBUF_FIXEDLEN); 1947 1948 sbuf_printf(&sb, "device=%s%d lba=%jd blocks=%jd latency=%jd", 1949 iop->softc->periph->periph_name, 1950 iop->softc->periph->unit_number, 1951 lba, cnt, sbttons(sim_latency)); 1952 if (sbuf_finish(&sb) == 0) 1953 devctl_notify("CAM", "iosched", "latency", sbuf_data(&sb)); 1954 sbuf_delete(&sb); 1955 free(sbmsg, M_CAMSCHED); 1956 } 1957 1958 static void 1959 cam_iosched_update(struct iop_stats *iop, sbintime_t sim_latency, 1960 const struct bio *bp) 1961 { 1962 sbintime_t y, deltasq, delta; 1963 int i; 1964 1965 /* 1966 * Simple threshold: count the number of events that excede the 1967 * configured threshold. 1968 */ 1969 if (sim_latency > iop->bad_latency) { 1970 cam_iosched_devctl_outlier(iop, sim_latency, bp); 1971 iop->too_long++; 1972 } 1973 1974 /* 1975 * Keep counts for latency. We do it by power of two buckets. 1976 * This helps us spot outlier behavior obscured by averages. 1977 */ 1978 for (i = 0; i < LAT_BUCKETS - 1; i++) { 1979 if (sim_latency < latencies[i]) { 1980 iop->latencies[i]++; 1981 break; 1982 } 1983 } 1984 if (i == LAT_BUCKETS - 1) 1985 iop->latencies[i]++; /* Put all > 8192ms values into the last bucket. */ 1986 1987 /* 1988 * Classic exponentially decaying average with a tiny alpha 1989 * (2 ^ -alpha_bits). For more info see the NIST statistical 1990 * handbook. 1991 * 1992 * ema_t = y_t * alpha + ema_t-1 * (1 - alpha) [nist] 1993 * ema_t = y_t * alpha + ema_t-1 - alpha * ema_t-1 1994 * ema_t = alpha * y_t - alpha * ema_t-1 + ema_t-1 1995 * alpha = 1 / (1 << alpha_bits) 1996 * sub e == ema_t-1, b == 1/alpha (== 1 << alpha_bits), d == y_t - ema_t-1 1997 * = y_t/b - e/b + be/b 1998 * = (y_t - e + be) / b 1999 * = (e + d) / b 2000 * 2001 * Since alpha is a power of two, we can compute this w/o any mult or 2002 * division. 2003 * 2004 * Variance can also be computed. Usually, it would be expressed as follows: 2005 * diff_t = y_t - ema_t-1 2006 * emvar_t = (1 - alpha) * (emavar_t-1 + diff_t^2 * alpha) 2007 * = emavar_t-1 - alpha * emavar_t-1 + delta_t^2 * alpha - (delta_t * alpha)^2 2008 * sub b == 1/alpha (== 1 << alpha_bits), e == emavar_t-1, d = delta_t^2 2009 * = e - e/b + dd/b + dd/bb 2010 * = (bbe - be + bdd + dd) / bb 2011 * = (bbe + b(dd-e) + dd) / bb (which is expanded below bb = 1<<(2*alpha_bits)) 2012 */ 2013 /* 2014 * XXX possible numeric issues 2015 * o We assume right shifted integers do the right thing, since that's 2016 * implementation defined. You can change the right shifts to / (1LL << alpha). 2017 * o alpha_bits = 9 gives ema ceiling of 23 bits of seconds for ema and 14 bits 2018 * for emvar. This puts a ceiling of 13 bits on alpha since we need a 2019 * few tens of seconds of representation. 2020 * o We mitigate alpha issues by never setting it too high. 2021 */ 2022 y = sim_latency; 2023 delta = (y - iop->ema); /* d */ 2024 iop->ema = ((iop->ema << alpha_bits) + delta) >> alpha_bits; 2025 2026 /* 2027 * Were we to naively plow ahead at this point, we wind up with many numerical 2028 * issues making any SD > ~3ms unreliable. So, we shift right by 12. This leaves 2029 * us with microsecond level precision in the input, so the same in the 2030 * output. It means we can't overflow deltasq unless delta > 4k seconds. It 2031 * also means that emvar can be up 46 bits 40 of which are fraction, which 2032 * gives us a way to measure up to ~8s in the SD before the computation goes 2033 * unstable. Even the worst hard disk rarely has > 1s service time in the 2034 * drive. It does mean we have to shift left 12 bits after taking the 2035 * square root to compute the actual standard deviation estimate. This loss of 2036 * precision is preferable to needing int128 types to work. The above numbers 2037 * assume alpha=9. 10 or 11 are ok, but we start to run into issues at 12, 2038 * so 12 or 13 is OK for EMA, EMVAR and SD will be wrong in those cases. 2039 */ 2040 delta >>= 12; 2041 deltasq = delta * delta; /* dd */ 2042 iop->emvar = ((iop->emvar << (2 * alpha_bits)) + /* bbe */ 2043 ((deltasq - iop->emvar) << alpha_bits) + /* b(dd-e) */ 2044 deltasq) /* dd */ 2045 >> (2 * alpha_bits); /* div bb */ 2046 iop->sd = (sbintime_t)isqrt64((uint64_t)iop->emvar) << 12; 2047 } 2048 2049 static void 2050 cam_iosched_io_metric_update(struct cam_iosched_softc *isc, 2051 sbintime_t sim_latency, const struct bio *bp) 2052 { 2053 switch (bp->bio_cmd) { 2054 case BIO_READ: 2055 cam_iosched_update(&isc->read_stats, sim_latency, bp); 2056 break; 2057 case BIO_WRITE: 2058 cam_iosched_update(&isc->write_stats, sim_latency, bp); 2059 break; 2060 case BIO_DELETE: 2061 cam_iosched_update(&isc->trim_stats, sim_latency, bp); 2062 break; 2063 default: 2064 break; 2065 } 2066 } 2067 2068 #ifdef DDB 2069 static int biolen(struct bio_queue_head *bq) 2070 { 2071 int i = 0; 2072 struct bio *bp; 2073 2074 TAILQ_FOREACH(bp, &bq->queue, bio_queue) { 2075 i++; 2076 } 2077 return i; 2078 } 2079 2080 /* 2081 * Show the internal state of the I/O scheduler. 2082 */ 2083 DB_SHOW_COMMAND(iosched, cam_iosched_db_show) 2084 { 2085 struct cam_iosched_softc *isc; 2086 2087 if (!have_addr) { 2088 db_printf("Need addr\n"); 2089 return; 2090 } 2091 isc = (struct cam_iosched_softc *)addr; 2092 db_printf("pending_reads: %d\n", isc->read_stats.pending); 2093 db_printf("min_reads: %d\n", isc->read_stats.min); 2094 db_printf("max_reads: %d\n", isc->read_stats.max); 2095 db_printf("reads: %d\n", isc->read_stats.total); 2096 db_printf("in_reads: %d\n", isc->read_stats.in); 2097 db_printf("out_reads: %d\n", isc->read_stats.out); 2098 db_printf("queued_reads: %d\n", isc->read_stats.queued); 2099 db_printf("Read Q len %d\n", biolen(&isc->bio_queue)); 2100 db_printf("pending_writes: %d\n", isc->write_stats.pending); 2101 db_printf("min_writes: %d\n", isc->write_stats.min); 2102 db_printf("max_writes: %d\n", isc->write_stats.max); 2103 db_printf("writes: %d\n", isc->write_stats.total); 2104 db_printf("in_writes: %d\n", isc->write_stats.in); 2105 db_printf("out_writes: %d\n", isc->write_stats.out); 2106 db_printf("queued_writes: %d\n", isc->write_stats.queued); 2107 db_printf("Write Q len %d\n", biolen(&isc->write_queue)); 2108 db_printf("pending_trims: %d\n", isc->trim_stats.pending); 2109 db_printf("min_trims: %d\n", isc->trim_stats.min); 2110 db_printf("max_trims: %d\n", isc->trim_stats.max); 2111 db_printf("trims: %d\n", isc->trim_stats.total); 2112 db_printf("in_trims: %d\n", isc->trim_stats.in); 2113 db_printf("out_trims: %d\n", isc->trim_stats.out); 2114 db_printf("queued_trims: %d\n", isc->trim_stats.queued); 2115 db_printf("Trim Q len %d\n", biolen(&isc->trim_queue)); 2116 db_printf("read_bias: %d\n", isc->read_bias); 2117 db_printf("current_read_bias: %d\n", isc->current_read_bias); 2118 db_printf("Trim active? %s\n", 2119 (isc->flags & CAM_IOSCHED_FLAG_TRIM_ACTIVE) ? "yes" : "no"); 2120 } 2121 #endif 2122 #endif 2123