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 * @is_write: true if the current IO is a write, false if it's a read 140 * @ret: time to wait 141 */ 142 static int64_t throttle_compute_wait_for(ThrottleState *ts, 143 bool is_write) 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[is_write][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 * @is_write: the type of operation 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 bool is_write, 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, is_write); 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 tt->timers[0] = aio_timer_new(new_context, tt->clock_type, SCALE_NS, 203 tt->read_timer_cb, tt->timer_opaque); 204 tt->timers[1] = aio_timer_new(new_context, tt->clock_type, SCALE_NS, 205 tt->write_timer_cb, tt->timer_opaque); 206 } 207 208 /* 209 * Initialize the ThrottleConfig structure to a valid state 210 * @cfg: the config to initialize 211 */ 212 void throttle_config_init(ThrottleConfig *cfg) 213 { 214 unsigned i; 215 memset(cfg, 0, sizeof(*cfg)); 216 for (i = 0; i < BUCKETS_COUNT; i++) { 217 cfg->buckets[i].burst_length = 1; 218 } 219 } 220 221 /* To be called first on the ThrottleState */ 222 void throttle_init(ThrottleState *ts) 223 { 224 memset(ts, 0, sizeof(ThrottleState)); 225 throttle_config_init(&ts->cfg); 226 } 227 228 /* To be called first on the ThrottleTimers */ 229 void throttle_timers_init(ThrottleTimers *tt, 230 AioContext *aio_context, 231 QEMUClockType clock_type, 232 QEMUTimerCB *read_timer_cb, 233 QEMUTimerCB *write_timer_cb, 234 void *timer_opaque) 235 { 236 memset(tt, 0, sizeof(ThrottleTimers)); 237 238 tt->clock_type = clock_type; 239 tt->read_timer_cb = read_timer_cb; 240 tt->write_timer_cb = write_timer_cb; 241 tt->timer_opaque = timer_opaque; 242 throttle_timers_attach_aio_context(tt, aio_context); 243 } 244 245 /* destroy a timer */ 246 static void throttle_timer_destroy(QEMUTimer **timer) 247 { 248 assert(*timer != NULL); 249 250 timer_free(*timer); 251 *timer = NULL; 252 } 253 254 /* Remove timers from event loop */ 255 void throttle_timers_detach_aio_context(ThrottleTimers *tt) 256 { 257 int i; 258 259 for (i = 0; i < 2; i++) { 260 throttle_timer_destroy(&tt->timers[i]); 261 } 262 } 263 264 /* To be called last on the ThrottleTimers */ 265 void throttle_timers_destroy(ThrottleTimers *tt) 266 { 267 throttle_timers_detach_aio_context(tt); 268 } 269 270 /* is any throttling timer configured */ 271 bool throttle_timers_are_initialized(ThrottleTimers *tt) 272 { 273 if (tt->timers[0]) { 274 return true; 275 } 276 277 return false; 278 } 279 280 /* Does any throttling must be done 281 * 282 * @cfg: the throttling configuration to inspect 283 * @ret: true if throttling must be done else false 284 */ 285 bool throttle_enabled(ThrottleConfig *cfg) 286 { 287 int i; 288 289 for (i = 0; i < BUCKETS_COUNT; i++) { 290 if (cfg->buckets[i].avg > 0) { 291 return true; 292 } 293 } 294 295 return false; 296 } 297 298 /* check if a throttling configuration is valid 299 * @cfg: the throttling configuration to inspect 300 * @ret: true if valid else false 301 * @errp: error object 302 */ 303 bool throttle_is_valid(ThrottleConfig *cfg, Error **errp) 304 { 305 int i; 306 bool bps_flag, ops_flag; 307 bool bps_max_flag, ops_max_flag; 308 309 bps_flag = cfg->buckets[THROTTLE_BPS_TOTAL].avg && 310 (cfg->buckets[THROTTLE_BPS_READ].avg || 311 cfg->buckets[THROTTLE_BPS_WRITE].avg); 312 313 ops_flag = cfg->buckets[THROTTLE_OPS_TOTAL].avg && 314 (cfg->buckets[THROTTLE_OPS_READ].avg || 315 cfg->buckets[THROTTLE_OPS_WRITE].avg); 316 317 bps_max_flag = cfg->buckets[THROTTLE_BPS_TOTAL].max && 318 (cfg->buckets[THROTTLE_BPS_READ].max || 319 cfg->buckets[THROTTLE_BPS_WRITE].max); 320 321 ops_max_flag = cfg->buckets[THROTTLE_OPS_TOTAL].max && 322 (cfg->buckets[THROTTLE_OPS_READ].max || 323 cfg->buckets[THROTTLE_OPS_WRITE].max); 324 325 if (bps_flag || ops_flag || bps_max_flag || ops_max_flag) { 326 error_setg(errp, "bps/iops/max total values and read/write values" 327 " cannot be used at the same time"); 328 return false; 329 } 330 331 if (cfg->op_size && 332 !cfg->buckets[THROTTLE_OPS_TOTAL].avg && 333 !cfg->buckets[THROTTLE_OPS_READ].avg && 334 !cfg->buckets[THROTTLE_OPS_WRITE].avg) { 335 error_setg(errp, "iops size requires an iops value to be set"); 336 return false; 337 } 338 339 for (i = 0; i < BUCKETS_COUNT; i++) { 340 LeakyBucket *bkt = &cfg->buckets[i]; 341 if (bkt->avg > THROTTLE_VALUE_MAX || bkt->max > THROTTLE_VALUE_MAX) { 342 error_setg(errp, "bps/iops/max values must be within [0, %lld]", 343 THROTTLE_VALUE_MAX); 344 return false; 345 } 346 347 if (!bkt->burst_length) { 348 error_setg(errp, "the burst length cannot be 0"); 349 return false; 350 } 351 352 if (bkt->burst_length > 1 && !bkt->max) { 353 error_setg(errp, "burst length set without burst rate"); 354 return false; 355 } 356 357 if (bkt->max && bkt->burst_length > THROTTLE_VALUE_MAX / bkt->max) { 358 error_setg(errp, "burst length too high for this burst rate"); 359 return false; 360 } 361 362 if (bkt->max && !bkt->avg) { 363 error_setg(errp, "bps_max/iops_max require corresponding" 364 " bps/iops values"); 365 return false; 366 } 367 368 if (bkt->max && bkt->max < bkt->avg) { 369 error_setg(errp, "bps_max/iops_max cannot be lower than bps/iops"); 370 return false; 371 } 372 } 373 374 return true; 375 } 376 377 /* Used to configure the throttle 378 * 379 * @ts: the throttle state we are working on 380 * @clock_type: the group's clock_type 381 * @cfg: the config to set 382 */ 383 void throttle_config(ThrottleState *ts, 384 QEMUClockType clock_type, 385 ThrottleConfig *cfg) 386 { 387 int i; 388 389 ts->cfg = *cfg; 390 391 /* Zero bucket level */ 392 for (i = 0; i < BUCKETS_COUNT; i++) { 393 ts->cfg.buckets[i].level = 0; 394 ts->cfg.buckets[i].burst_level = 0; 395 } 396 397 ts->previous_leak = qemu_clock_get_ns(clock_type); 398 } 399 400 /* used to get config 401 * 402 * @ts: the throttle state we are working on 403 * @cfg: the config to write 404 */ 405 void throttle_get_config(ThrottleState *ts, ThrottleConfig *cfg) 406 { 407 *cfg = ts->cfg; 408 } 409 410 411 /* Schedule the read or write timer if needed 412 * 413 * NOTE: this function is not unit tested due to it's usage of timer_mod 414 * 415 * @tt: the timers structure 416 * @is_write: the type of operation (read/write) 417 * @ret: true if the timer has been scheduled else false 418 */ 419 bool throttle_schedule_timer(ThrottleState *ts, 420 ThrottleTimers *tt, 421 bool is_write) 422 { 423 int64_t now = qemu_clock_get_ns(tt->clock_type); 424 int64_t next_timestamp; 425 bool must_wait; 426 427 must_wait = throttle_compute_timer(ts, 428 is_write, 429 now, 430 &next_timestamp); 431 432 /* request not throttled */ 433 if (!must_wait) { 434 return false; 435 } 436 437 /* request throttled and timer pending -> do nothing */ 438 if (timer_pending(tt->timers[is_write])) { 439 return true; 440 } 441 442 /* request throttled and timer not pending -> arm timer */ 443 timer_mod(tt->timers[is_write], next_timestamp); 444 return true; 445 } 446 447 /* do the accounting for this operation 448 * 449 * @is_write: the type of operation (read/write) 450 * @size: the size of the operation 451 */ 452 void throttle_account(ThrottleState *ts, bool is_write, uint64_t size) 453 { 454 const BucketType bucket_types_size[2][2] = { 455 { THROTTLE_BPS_TOTAL, THROTTLE_BPS_READ }, 456 { THROTTLE_BPS_TOTAL, THROTTLE_BPS_WRITE } 457 }; 458 const BucketType bucket_types_units[2][2] = { 459 { THROTTLE_OPS_TOTAL, THROTTLE_OPS_READ }, 460 { THROTTLE_OPS_TOTAL, THROTTLE_OPS_WRITE } 461 }; 462 double units = 1.0; 463 unsigned i; 464 465 /* if cfg.op_size is defined and smaller than size we compute unit count */ 466 if (ts->cfg.op_size && size > ts->cfg.op_size) { 467 units = (double) size / ts->cfg.op_size; 468 } 469 470 for (i = 0; i < 2; i++) { 471 LeakyBucket *bkt; 472 473 bkt = &ts->cfg.buckets[bucket_types_size[is_write][i]]; 474 bkt->level += size; 475 if (bkt->burst_length > 1) { 476 bkt->burst_level += size; 477 } 478 479 bkt = &ts->cfg.buckets[bucket_types_units[is_write][i]]; 480 bkt->level += units; 481 if (bkt->burst_length > 1) { 482 bkt->burst_level += units; 483 } 484 } 485 } 486 487 /* return a ThrottleConfig based on the options in a ThrottleLimits 488 * 489 * @arg: the ThrottleLimits object to read from 490 * @cfg: the ThrottleConfig to edit 491 * @errp: error object 492 */ 493 void throttle_limits_to_config(ThrottleLimits *arg, ThrottleConfig *cfg, 494 Error **errp) 495 { 496 if (arg->has_bps_total) { 497 cfg->buckets[THROTTLE_BPS_TOTAL].avg = arg->bps_total; 498 } 499 if (arg->has_bps_read) { 500 cfg->buckets[THROTTLE_BPS_READ].avg = arg->bps_read; 501 } 502 if (arg->has_bps_write) { 503 cfg->buckets[THROTTLE_BPS_WRITE].avg = arg->bps_write; 504 } 505 506 if (arg->has_iops_total) { 507 cfg->buckets[THROTTLE_OPS_TOTAL].avg = arg->iops_total; 508 } 509 if (arg->has_iops_read) { 510 cfg->buckets[THROTTLE_OPS_READ].avg = arg->iops_read; 511 } 512 if (arg->has_iops_write) { 513 cfg->buckets[THROTTLE_OPS_WRITE].avg = arg->iops_write; 514 } 515 516 if (arg->has_bps_total_max) { 517 cfg->buckets[THROTTLE_BPS_TOTAL].max = arg->bps_total_max; 518 } 519 if (arg->has_bps_read_max) { 520 cfg->buckets[THROTTLE_BPS_READ].max = arg->bps_read_max; 521 } 522 if (arg->has_bps_write_max) { 523 cfg->buckets[THROTTLE_BPS_WRITE].max = arg->bps_write_max; 524 } 525 if (arg->has_iops_total_max) { 526 cfg->buckets[THROTTLE_OPS_TOTAL].max = arg->iops_total_max; 527 } 528 if (arg->has_iops_read_max) { 529 cfg->buckets[THROTTLE_OPS_READ].max = arg->iops_read_max; 530 } 531 if (arg->has_iops_write_max) { 532 cfg->buckets[THROTTLE_OPS_WRITE].max = arg->iops_write_max; 533 } 534 535 if (arg->has_bps_total_max_length) { 536 if (arg->bps_total_max_length > UINT_MAX) { 537 error_setg(errp, "bps-total-max-length value must be in" 538 " the range [0, %u]", UINT_MAX); 539 return; 540 } 541 cfg->buckets[THROTTLE_BPS_TOTAL].burst_length = arg->bps_total_max_length; 542 } 543 if (arg->has_bps_read_max_length) { 544 if (arg->bps_read_max_length > UINT_MAX) { 545 error_setg(errp, "bps-read-max-length value must be in" 546 " the range [0, %u]", UINT_MAX); 547 return; 548 } 549 cfg->buckets[THROTTLE_BPS_READ].burst_length = arg->bps_read_max_length; 550 } 551 if (arg->has_bps_write_max_length) { 552 if (arg->bps_write_max_length > UINT_MAX) { 553 error_setg(errp, "bps-write-max-length value must be in" 554 " the range [0, %u]", UINT_MAX); 555 return; 556 } 557 cfg->buckets[THROTTLE_BPS_WRITE].burst_length = arg->bps_write_max_length; 558 } 559 if (arg->has_iops_total_max_length) { 560 if (arg->iops_total_max_length > UINT_MAX) { 561 error_setg(errp, "iops-total-max-length value must be in" 562 " the range [0, %u]", UINT_MAX); 563 return; 564 } 565 cfg->buckets[THROTTLE_OPS_TOTAL].burst_length = arg->iops_total_max_length; 566 } 567 if (arg->has_iops_read_max_length) { 568 if (arg->iops_read_max_length > UINT_MAX) { 569 error_setg(errp, "iops-read-max-length value must be in" 570 " the range [0, %u]", UINT_MAX); 571 return; 572 } 573 cfg->buckets[THROTTLE_OPS_READ].burst_length = arg->iops_read_max_length; 574 } 575 if (arg->has_iops_write_max_length) { 576 if (arg->iops_write_max_length > UINT_MAX) { 577 error_setg(errp, "iops-write-max-length value must be in" 578 " the range [0, %u]", UINT_MAX); 579 return; 580 } 581 cfg->buckets[THROTTLE_OPS_WRITE].burst_length = arg->iops_write_max_length; 582 } 583 584 if (arg->has_iops_size) { 585 cfg->op_size = arg->iops_size; 586 } 587 588 throttle_is_valid(cfg, errp); 589 } 590 591 /* write the options of a ThrottleConfig to a ThrottleLimits 592 * 593 * @cfg: the ThrottleConfig to read from 594 * @var: the ThrottleLimits to write to 595 */ 596 void throttle_config_to_limits(ThrottleConfig *cfg, ThrottleLimits *var) 597 { 598 var->bps_total = cfg->buckets[THROTTLE_BPS_TOTAL].avg; 599 var->bps_read = cfg->buckets[THROTTLE_BPS_READ].avg; 600 var->bps_write = cfg->buckets[THROTTLE_BPS_WRITE].avg; 601 var->iops_total = cfg->buckets[THROTTLE_OPS_TOTAL].avg; 602 var->iops_read = cfg->buckets[THROTTLE_OPS_READ].avg; 603 var->iops_write = cfg->buckets[THROTTLE_OPS_WRITE].avg; 604 var->bps_total_max = cfg->buckets[THROTTLE_BPS_TOTAL].max; 605 var->bps_read_max = cfg->buckets[THROTTLE_BPS_READ].max; 606 var->bps_write_max = cfg->buckets[THROTTLE_BPS_WRITE].max; 607 var->iops_total_max = cfg->buckets[THROTTLE_OPS_TOTAL].max; 608 var->iops_read_max = cfg->buckets[THROTTLE_OPS_READ].max; 609 var->iops_write_max = cfg->buckets[THROTTLE_OPS_WRITE].max; 610 var->bps_total_max_length = cfg->buckets[THROTTLE_BPS_TOTAL].burst_length; 611 var->bps_read_max_length = cfg->buckets[THROTTLE_BPS_READ].burst_length; 612 var->bps_write_max_length = cfg->buckets[THROTTLE_BPS_WRITE].burst_length; 613 var->iops_total_max_length = cfg->buckets[THROTTLE_OPS_TOTAL].burst_length; 614 var->iops_read_max_length = cfg->buckets[THROTTLE_OPS_READ].burst_length; 615 var->iops_write_max_length = cfg->buckets[THROTTLE_OPS_WRITE].burst_length; 616 var->iops_size = cfg->op_size; 617 618 var->has_bps_total = true; 619 var->has_bps_read = true; 620 var->has_bps_write = true; 621 var->has_iops_total = true; 622 var->has_iops_read = true; 623 var->has_iops_write = true; 624 var->has_bps_total_max = true; 625 var->has_bps_read_max = true; 626 var->has_bps_write_max = true; 627 var->has_iops_total_max = true; 628 var->has_iops_read_max = true; 629 var->has_iops_write_max = true; 630 var->has_bps_read_max_length = true; 631 var->has_bps_total_max_length = true; 632 var->has_bps_write_max_length = true; 633 var->has_iops_total_max_length = true; 634 var->has_iops_read_max_length = true; 635 var->has_iops_write_max_length = true; 636 var->has_iops_size = true; 637 } 638