1 /* 2 * QEMU throttling infrastructure 3 * 4 * Copyright (C) Nodalink, EURL. 2013-2014 5 * Copyright (C) Igalia, S.L. 2015 6 * 7 * Authors: 8 * Benoît Canet <benoit.canet@nodalink.com> 9 * Alberto Garcia <berto@igalia.com> 10 * 11 * This program is free software; you can redistribute it and/or 12 * modify it under the terms of the GNU General Public License as 13 * published by the Free Software Foundation; either version 2 or 14 * (at your option) version 3 of the License. 15 * 16 * This program is distributed in the hope that it will be useful, 17 * but WITHOUT ANY WARRANTY; without even the implied warranty of 18 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 19 * GNU General Public License for more details. 20 * 21 * You should have received a copy of the GNU General Public License 22 * along with this program; if not, see <http://www.gnu.org/licenses/>. 23 */ 24 25 #include "qemu/osdep.h" 26 #include "qapi/error.h" 27 #include "qemu/throttle.h" 28 #include "qemu/timer.h" 29 #include "block/aio.h" 30 31 /* This function make a bucket leak 32 * 33 * @bkt: the bucket to make leak 34 * @delta_ns: the time delta 35 */ 36 void throttle_leak_bucket(LeakyBucket *bkt, int64_t delta_ns) 37 { 38 double leak; 39 40 /* compute how much to leak */ 41 leak = (bkt->avg * (double) delta_ns) / NANOSECONDS_PER_SECOND; 42 43 /* make the bucket leak */ 44 bkt->level = MAX(bkt->level - leak, 0); 45 46 /* if we allow bursts for more than one second we also need to 47 * keep track of bkt->burst_level so the bkt->max goal per second 48 * is attained */ 49 if (bkt->burst_length > 1) { 50 leak = (bkt->max * (double) delta_ns) / NANOSECONDS_PER_SECOND; 51 bkt->burst_level = MAX(bkt->burst_level - leak, 0); 52 } 53 } 54 55 /* Calculate the time delta since last leak and make proportionals leaks 56 * 57 * @now: the current timestamp in ns 58 */ 59 static void throttle_do_leak(ThrottleState *ts, int64_t now) 60 { 61 /* compute the time elapsed since the last leak */ 62 int64_t delta_ns = now - ts->previous_leak; 63 int i; 64 65 ts->previous_leak = now; 66 67 if (delta_ns <= 0) { 68 return; 69 } 70 71 /* make each bucket leak */ 72 for (i = 0; i < BUCKETS_COUNT; i++) { 73 throttle_leak_bucket(&ts->cfg.buckets[i], delta_ns); 74 } 75 } 76 77 /* do the real job of computing the time to wait 78 * 79 * @limit: the throttling limit 80 * @extra: the number of operation to delay 81 * @ret: the time to wait in ns 82 */ 83 static int64_t throttle_do_compute_wait(double limit, double extra) 84 { 85 double wait = extra * NANOSECONDS_PER_SECOND; 86 wait /= limit; 87 return wait; 88 } 89 90 /* This function compute the wait time in ns that a leaky bucket should trigger 91 * 92 * @bkt: the leaky bucket we operate on 93 * @ret: the resulting wait time in ns or 0 if the operation can go through 94 */ 95 int64_t throttle_compute_wait(LeakyBucket *bkt) 96 { 97 double extra; /* the number of extra units blocking the io */ 98 double bucket_size; /* I/O before throttling to bkt->avg */ 99 double burst_bucket_size; /* Before throttling to bkt->max */ 100 101 if (!bkt->avg) { 102 return 0; 103 } 104 105 if (!bkt->max) { 106 /* If bkt->max is 0 we still want to allow short bursts of I/O 107 * from the guest, otherwise every other request will be throttled 108 * and performance will suffer considerably. */ 109 bucket_size = (double) bkt->avg / 10; 110 burst_bucket_size = 0; 111 } else { 112 /* If we have a burst limit then we have to wait until all I/O 113 * at burst rate has finished before throttling to bkt->avg */ 114 bucket_size = bkt->max * bkt->burst_length; 115 burst_bucket_size = (double) bkt->max / 10; 116 } 117 118 /* If the main bucket is full then we have to wait */ 119 extra = bkt->level - bucket_size; 120 if (extra > 0) { 121 return throttle_do_compute_wait(bkt->avg, extra); 122 } 123 124 /* If the main bucket is not full yet we still have to check the 125 * burst bucket in order to enforce the burst limit */ 126 if (bkt->burst_length > 1) { 127 assert(bkt->max > 0); /* see throttle_is_valid() */ 128 extra = bkt->burst_level - burst_bucket_size; 129 if (extra > 0) { 130 return throttle_do_compute_wait(bkt->max, extra); 131 } 132 } 133 134 return 0; 135 } 136 137 /* This function compute the time that must be waited while this IO 138 * 139 * @direction: throttle direction 140 * @ret: time to wait 141 */ 142 static int64_t throttle_compute_wait_for(ThrottleState *ts, 143 ThrottleDirection direction) 144 { 145 BucketType to_check[2][4] = { {THROTTLE_BPS_TOTAL, 146 THROTTLE_OPS_TOTAL, 147 THROTTLE_BPS_READ, 148 THROTTLE_OPS_READ}, 149 {THROTTLE_BPS_TOTAL, 150 THROTTLE_OPS_TOTAL, 151 THROTTLE_BPS_WRITE, 152 THROTTLE_OPS_WRITE}, }; 153 int64_t wait, max_wait = 0; 154 int i; 155 156 for (i = 0; i < 4; i++) { 157 BucketType index = to_check[direction][i]; 158 wait = throttle_compute_wait(&ts->cfg.buckets[index]); 159 if (wait > max_wait) { 160 max_wait = wait; 161 } 162 } 163 164 return max_wait; 165 } 166 167 /* compute the timer for this type of operation 168 * 169 * @direction: throttle direction 170 * @now: the current clock timestamp 171 * @next_timestamp: the resulting timer 172 * @ret: true if a timer must be set 173 */ 174 static bool throttle_compute_timer(ThrottleState *ts, 175 ThrottleDirection direction, 176 int64_t now, 177 int64_t *next_timestamp) 178 { 179 int64_t wait; 180 181 /* leak proportionally to the time elapsed */ 182 throttle_do_leak(ts, now); 183 184 /* compute the wait time if any */ 185 wait = throttle_compute_wait_for(ts, direction); 186 187 /* if the code must wait compute when the next timer should fire */ 188 if (wait) { 189 *next_timestamp = now + wait; 190 return true; 191 } 192 193 /* else no need to wait at all */ 194 *next_timestamp = now; 195 return false; 196 } 197 198 /* Add timers to event loop */ 199 void throttle_timers_attach_aio_context(ThrottleTimers *tt, 200 AioContext *new_context) 201 { 202 ThrottleDirection dir; 203 204 for (dir = THROTTLE_READ; dir < THROTTLE_MAX; dir++) { 205 if (tt->timer_cb[dir]) { 206 tt->timers[dir] = 207 aio_timer_new(new_context, tt->clock_type, SCALE_NS, 208 tt->timer_cb[dir], tt->timer_opaque); 209 } 210 } 211 } 212 213 /* 214 * Initialize the ThrottleConfig structure to a valid state 215 * @cfg: the config to initialize 216 */ 217 void throttle_config_init(ThrottleConfig *cfg) 218 { 219 unsigned i; 220 memset(cfg, 0, sizeof(*cfg)); 221 for (i = 0; i < BUCKETS_COUNT; i++) { 222 cfg->buckets[i].burst_length = 1; 223 } 224 } 225 226 /* To be called first on the ThrottleState */ 227 void throttle_init(ThrottleState *ts) 228 { 229 memset(ts, 0, sizeof(ThrottleState)); 230 throttle_config_init(&ts->cfg); 231 } 232 233 /* To be called first on the ThrottleTimers */ 234 void throttle_timers_init(ThrottleTimers *tt, 235 AioContext *aio_context, 236 QEMUClockType clock_type, 237 QEMUTimerCB *read_timer_cb, 238 QEMUTimerCB *write_timer_cb, 239 void *timer_opaque) 240 { 241 assert(read_timer_cb || write_timer_cb); 242 memset(tt, 0, sizeof(ThrottleTimers)); 243 244 tt->clock_type = clock_type; 245 tt->timer_cb[THROTTLE_READ] = read_timer_cb; 246 tt->timer_cb[THROTTLE_WRITE] = write_timer_cb; 247 tt->timer_opaque = timer_opaque; 248 throttle_timers_attach_aio_context(tt, aio_context); 249 } 250 251 /* destroy a timer */ 252 static void throttle_timer_destroy(QEMUTimer **timer) 253 { 254 if (*timer == NULL) { 255 return; 256 } 257 258 timer_free(*timer); 259 *timer = NULL; 260 } 261 262 /* Remove timers from event loop */ 263 void throttle_timers_detach_aio_context(ThrottleTimers *tt) 264 { 265 ThrottleDirection dir; 266 267 for (dir = THROTTLE_READ; dir < THROTTLE_MAX; dir++) { 268 throttle_timer_destroy(&tt->timers[dir]); 269 } 270 } 271 272 /* To be called last on the ThrottleTimers */ 273 void throttle_timers_destroy(ThrottleTimers *tt) 274 { 275 throttle_timers_detach_aio_context(tt); 276 } 277 278 /* is any throttling timer configured */ 279 bool throttle_timers_are_initialized(ThrottleTimers *tt) 280 { 281 ThrottleDirection dir; 282 283 for (dir = THROTTLE_READ; dir < THROTTLE_MAX; dir++) { 284 if (tt->timers[dir]) { 285 return true; 286 } 287 } 288 289 return false; 290 } 291 292 /* Does any throttling must be done 293 * 294 * @cfg: the throttling configuration to inspect 295 * @ret: true if throttling must be done else false 296 */ 297 bool throttle_enabled(ThrottleConfig *cfg) 298 { 299 int i; 300 301 for (i = 0; i < BUCKETS_COUNT; i++) { 302 if (cfg->buckets[i].avg > 0) { 303 return true; 304 } 305 } 306 307 return false; 308 } 309 310 /* check if a throttling configuration is valid 311 * @cfg: the throttling configuration to inspect 312 * @ret: true if valid else false 313 * @errp: error object 314 */ 315 bool throttle_is_valid(ThrottleConfig *cfg, Error **errp) 316 { 317 int i; 318 bool bps_flag, ops_flag; 319 bool bps_max_flag, ops_max_flag; 320 321 bps_flag = cfg->buckets[THROTTLE_BPS_TOTAL].avg && 322 (cfg->buckets[THROTTLE_BPS_READ].avg || 323 cfg->buckets[THROTTLE_BPS_WRITE].avg); 324 325 ops_flag = cfg->buckets[THROTTLE_OPS_TOTAL].avg && 326 (cfg->buckets[THROTTLE_OPS_READ].avg || 327 cfg->buckets[THROTTLE_OPS_WRITE].avg); 328 329 bps_max_flag = cfg->buckets[THROTTLE_BPS_TOTAL].max && 330 (cfg->buckets[THROTTLE_BPS_READ].max || 331 cfg->buckets[THROTTLE_BPS_WRITE].max); 332 333 ops_max_flag = cfg->buckets[THROTTLE_OPS_TOTAL].max && 334 (cfg->buckets[THROTTLE_OPS_READ].max || 335 cfg->buckets[THROTTLE_OPS_WRITE].max); 336 337 if (bps_flag || ops_flag || bps_max_flag || ops_max_flag) { 338 error_setg(errp, "bps/iops/max total values and read/write values" 339 " cannot be used at the same time"); 340 return false; 341 } 342 343 if (cfg->op_size && 344 !cfg->buckets[THROTTLE_OPS_TOTAL].avg && 345 !cfg->buckets[THROTTLE_OPS_READ].avg && 346 !cfg->buckets[THROTTLE_OPS_WRITE].avg) { 347 error_setg(errp, "iops size requires an iops value to be set"); 348 return false; 349 } 350 351 for (i = 0; i < BUCKETS_COUNT; i++) { 352 LeakyBucket *bkt = &cfg->buckets[i]; 353 if (bkt->avg > THROTTLE_VALUE_MAX || bkt->max > THROTTLE_VALUE_MAX) { 354 error_setg(errp, "bps/iops/max values must be within [0, %lld]", 355 THROTTLE_VALUE_MAX); 356 return false; 357 } 358 359 if (!bkt->burst_length) { 360 error_setg(errp, "the burst length cannot be 0"); 361 return false; 362 } 363 364 if (bkt->burst_length > 1 && !bkt->max) { 365 error_setg(errp, "burst length set without burst rate"); 366 return false; 367 } 368 369 if (bkt->max && bkt->burst_length > THROTTLE_VALUE_MAX / bkt->max) { 370 error_setg(errp, "burst length too high for this burst rate"); 371 return false; 372 } 373 374 if (bkt->max && !bkt->avg) { 375 error_setg(errp, "bps_max/iops_max require corresponding" 376 " bps/iops values"); 377 return false; 378 } 379 380 if (bkt->max && bkt->max < bkt->avg) { 381 error_setg(errp, "bps_max/iops_max cannot be lower than bps/iops"); 382 return false; 383 } 384 } 385 386 return true; 387 } 388 389 /* Used to configure the throttle 390 * 391 * @ts: the throttle state we are working on 392 * @clock_type: the group's clock_type 393 * @cfg: the config to set 394 */ 395 void throttle_config(ThrottleState *ts, 396 QEMUClockType clock_type, 397 ThrottleConfig *cfg) 398 { 399 int i; 400 401 ts->cfg = *cfg; 402 403 /* Zero bucket level */ 404 for (i = 0; i < BUCKETS_COUNT; i++) { 405 ts->cfg.buckets[i].level = 0; 406 ts->cfg.buckets[i].burst_level = 0; 407 } 408 409 ts->previous_leak = qemu_clock_get_ns(clock_type); 410 } 411 412 /* used to get config 413 * 414 * @ts: the throttle state we are working on 415 * @cfg: the config to write 416 */ 417 void throttle_get_config(ThrottleState *ts, ThrottleConfig *cfg) 418 { 419 *cfg = ts->cfg; 420 } 421 422 423 /* Schedule the read or write timer if needed 424 * 425 * NOTE: this function is not unit tested due to it's usage of timer_mod 426 * 427 * @tt: the timers structure 428 * @direction: throttle direction 429 * @ret: true if the timer has been scheduled else false 430 */ 431 bool throttle_schedule_timer(ThrottleState *ts, 432 ThrottleTimers *tt, 433 ThrottleDirection direction) 434 { 435 int64_t now = qemu_clock_get_ns(tt->clock_type); 436 int64_t next_timestamp; 437 QEMUTimer *timer; 438 bool must_wait; 439 440 assert(direction < THROTTLE_MAX); 441 timer = tt->timers[direction]; 442 assert(timer); 443 444 must_wait = throttle_compute_timer(ts, 445 direction, 446 now, 447 &next_timestamp); 448 449 /* request not throttled */ 450 if (!must_wait) { 451 return false; 452 } 453 454 /* request throttled and timer pending -> do nothing */ 455 if (timer_pending(timer)) { 456 return true; 457 } 458 459 /* request throttled and timer not pending -> arm timer */ 460 timer_mod(timer, next_timestamp); 461 return true; 462 } 463 464 /* do the accounting for this operation 465 * 466 * @direction: throttle direction 467 * @size: the size of the operation 468 */ 469 void throttle_account(ThrottleState *ts, ThrottleDirection direction, 470 uint64_t size) 471 { 472 const BucketType bucket_types_size[2][2] = { 473 { THROTTLE_BPS_TOTAL, THROTTLE_BPS_READ }, 474 { THROTTLE_BPS_TOTAL, THROTTLE_BPS_WRITE } 475 }; 476 const BucketType bucket_types_units[2][2] = { 477 { THROTTLE_OPS_TOTAL, THROTTLE_OPS_READ }, 478 { THROTTLE_OPS_TOTAL, THROTTLE_OPS_WRITE } 479 }; 480 double units = 1.0; 481 unsigned i; 482 483 assert(direction < THROTTLE_MAX); 484 /* if cfg.op_size is defined and smaller than size we compute unit count */ 485 if (ts->cfg.op_size && size > ts->cfg.op_size) { 486 units = (double) size / ts->cfg.op_size; 487 } 488 489 for (i = 0; i < 2; i++) { 490 LeakyBucket *bkt; 491 492 bkt = &ts->cfg.buckets[bucket_types_size[direction][i]]; 493 bkt->level += size; 494 if (bkt->burst_length > 1) { 495 bkt->burst_level += size; 496 } 497 498 bkt = &ts->cfg.buckets[bucket_types_units[direction][i]]; 499 bkt->level += units; 500 if (bkt->burst_length > 1) { 501 bkt->burst_level += units; 502 } 503 } 504 } 505 506 /* return a ThrottleConfig based on the options in a ThrottleLimits 507 * 508 * @arg: the ThrottleLimits object to read from 509 * @cfg: the ThrottleConfig to edit 510 * @errp: error object 511 */ 512 void throttle_limits_to_config(ThrottleLimits *arg, ThrottleConfig *cfg, 513 Error **errp) 514 { 515 if (arg->has_bps_total) { 516 cfg->buckets[THROTTLE_BPS_TOTAL].avg = arg->bps_total; 517 } 518 if (arg->has_bps_read) { 519 cfg->buckets[THROTTLE_BPS_READ].avg = arg->bps_read; 520 } 521 if (arg->has_bps_write) { 522 cfg->buckets[THROTTLE_BPS_WRITE].avg = arg->bps_write; 523 } 524 525 if (arg->has_iops_total) { 526 cfg->buckets[THROTTLE_OPS_TOTAL].avg = arg->iops_total; 527 } 528 if (arg->has_iops_read) { 529 cfg->buckets[THROTTLE_OPS_READ].avg = arg->iops_read; 530 } 531 if (arg->has_iops_write) { 532 cfg->buckets[THROTTLE_OPS_WRITE].avg = arg->iops_write; 533 } 534 535 if (arg->has_bps_total_max) { 536 cfg->buckets[THROTTLE_BPS_TOTAL].max = arg->bps_total_max; 537 } 538 if (arg->has_bps_read_max) { 539 cfg->buckets[THROTTLE_BPS_READ].max = arg->bps_read_max; 540 } 541 if (arg->has_bps_write_max) { 542 cfg->buckets[THROTTLE_BPS_WRITE].max = arg->bps_write_max; 543 } 544 if (arg->has_iops_total_max) { 545 cfg->buckets[THROTTLE_OPS_TOTAL].max = arg->iops_total_max; 546 } 547 if (arg->has_iops_read_max) { 548 cfg->buckets[THROTTLE_OPS_READ].max = arg->iops_read_max; 549 } 550 if (arg->has_iops_write_max) { 551 cfg->buckets[THROTTLE_OPS_WRITE].max = arg->iops_write_max; 552 } 553 554 if (arg->has_bps_total_max_length) { 555 if (arg->bps_total_max_length > UINT_MAX) { 556 error_setg(errp, "bps-total-max-length value must be in" 557 " the range [0, %u]", UINT_MAX); 558 return; 559 } 560 cfg->buckets[THROTTLE_BPS_TOTAL].burst_length = arg->bps_total_max_length; 561 } 562 if (arg->has_bps_read_max_length) { 563 if (arg->bps_read_max_length > UINT_MAX) { 564 error_setg(errp, "bps-read-max-length value must be in" 565 " the range [0, %u]", UINT_MAX); 566 return; 567 } 568 cfg->buckets[THROTTLE_BPS_READ].burst_length = arg->bps_read_max_length; 569 } 570 if (arg->has_bps_write_max_length) { 571 if (arg->bps_write_max_length > UINT_MAX) { 572 error_setg(errp, "bps-write-max-length value must be in" 573 " the range [0, %u]", UINT_MAX); 574 return; 575 } 576 cfg->buckets[THROTTLE_BPS_WRITE].burst_length = arg->bps_write_max_length; 577 } 578 if (arg->has_iops_total_max_length) { 579 if (arg->iops_total_max_length > UINT_MAX) { 580 error_setg(errp, "iops-total-max-length value must be in" 581 " the range [0, %u]", UINT_MAX); 582 return; 583 } 584 cfg->buckets[THROTTLE_OPS_TOTAL].burst_length = arg->iops_total_max_length; 585 } 586 if (arg->has_iops_read_max_length) { 587 if (arg->iops_read_max_length > UINT_MAX) { 588 error_setg(errp, "iops-read-max-length value must be in" 589 " the range [0, %u]", UINT_MAX); 590 return; 591 } 592 cfg->buckets[THROTTLE_OPS_READ].burst_length = arg->iops_read_max_length; 593 } 594 if (arg->has_iops_write_max_length) { 595 if (arg->iops_write_max_length > UINT_MAX) { 596 error_setg(errp, "iops-write-max-length value must be in" 597 " the range [0, %u]", UINT_MAX); 598 return; 599 } 600 cfg->buckets[THROTTLE_OPS_WRITE].burst_length = arg->iops_write_max_length; 601 } 602 603 if (arg->has_iops_size) { 604 cfg->op_size = arg->iops_size; 605 } 606 607 throttle_is_valid(cfg, errp); 608 } 609 610 /* write the options of a ThrottleConfig to a ThrottleLimits 611 * 612 * @cfg: the ThrottleConfig to read from 613 * @var: the ThrottleLimits to write to 614 */ 615 void throttle_config_to_limits(ThrottleConfig *cfg, ThrottleLimits *var) 616 { 617 var->bps_total = cfg->buckets[THROTTLE_BPS_TOTAL].avg; 618 var->bps_read = cfg->buckets[THROTTLE_BPS_READ].avg; 619 var->bps_write = cfg->buckets[THROTTLE_BPS_WRITE].avg; 620 var->iops_total = cfg->buckets[THROTTLE_OPS_TOTAL].avg; 621 var->iops_read = cfg->buckets[THROTTLE_OPS_READ].avg; 622 var->iops_write = cfg->buckets[THROTTLE_OPS_WRITE].avg; 623 var->bps_total_max = cfg->buckets[THROTTLE_BPS_TOTAL].max; 624 var->bps_read_max = cfg->buckets[THROTTLE_BPS_READ].max; 625 var->bps_write_max = cfg->buckets[THROTTLE_BPS_WRITE].max; 626 var->iops_total_max = cfg->buckets[THROTTLE_OPS_TOTAL].max; 627 var->iops_read_max = cfg->buckets[THROTTLE_OPS_READ].max; 628 var->iops_write_max = cfg->buckets[THROTTLE_OPS_WRITE].max; 629 var->bps_total_max_length = cfg->buckets[THROTTLE_BPS_TOTAL].burst_length; 630 var->bps_read_max_length = cfg->buckets[THROTTLE_BPS_READ].burst_length; 631 var->bps_write_max_length = cfg->buckets[THROTTLE_BPS_WRITE].burst_length; 632 var->iops_total_max_length = cfg->buckets[THROTTLE_OPS_TOTAL].burst_length; 633 var->iops_read_max_length = cfg->buckets[THROTTLE_OPS_READ].burst_length; 634 var->iops_write_max_length = cfg->buckets[THROTTLE_OPS_WRITE].burst_length; 635 var->iops_size = cfg->op_size; 636 637 var->has_bps_total = true; 638 var->has_bps_read = true; 639 var->has_bps_write = true; 640 var->has_iops_total = true; 641 var->has_iops_read = true; 642 var->has_iops_write = true; 643 var->has_bps_total_max = true; 644 var->has_bps_read_max = true; 645 var->has_bps_write_max = true; 646 var->has_iops_total_max = true; 647 var->has_iops_read_max = true; 648 var->has_iops_write_max = true; 649 var->has_bps_read_max_length = true; 650 var->has_bps_total_max_length = true; 651 var->has_bps_write_max_length = true; 652 var->has_iops_total_max_length = true; 653 var->has_iops_read_max_length = true; 654 var->has_iops_write_max_length = true; 655 var->has_iops_size = true; 656 } 657