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[THROTTLE_READ] = 203 aio_timer_new(new_context, tt->clock_type, SCALE_NS, 204 tt->timer_cb[THROTTLE_READ], tt->timer_opaque); 205 tt->timers[THROTTLE_WRITE] = 206 aio_timer_new(new_context, tt->clock_type, SCALE_NS, 207 tt->timer_cb[THROTTLE_WRITE], tt->timer_opaque); 208 } 209 210 /* 211 * Initialize the ThrottleConfig structure to a valid state 212 * @cfg: the config to initialize 213 */ 214 void throttle_config_init(ThrottleConfig *cfg) 215 { 216 unsigned i; 217 memset(cfg, 0, sizeof(*cfg)); 218 for (i = 0; i < BUCKETS_COUNT; i++) { 219 cfg->buckets[i].burst_length = 1; 220 } 221 } 222 223 /* To be called first on the ThrottleState */ 224 void throttle_init(ThrottleState *ts) 225 { 226 memset(ts, 0, sizeof(ThrottleState)); 227 throttle_config_init(&ts->cfg); 228 } 229 230 /* To be called first on the ThrottleTimers */ 231 void throttle_timers_init(ThrottleTimers *tt, 232 AioContext *aio_context, 233 QEMUClockType clock_type, 234 QEMUTimerCB *read_timer_cb, 235 QEMUTimerCB *write_timer_cb, 236 void *timer_opaque) 237 { 238 memset(tt, 0, sizeof(ThrottleTimers)); 239 240 tt->clock_type = clock_type; 241 tt->timer_cb[THROTTLE_READ] = read_timer_cb; 242 tt->timer_cb[THROTTLE_WRITE] = write_timer_cb; 243 tt->timer_opaque = timer_opaque; 244 throttle_timers_attach_aio_context(tt, aio_context); 245 } 246 247 /* destroy a timer */ 248 static void throttle_timer_destroy(QEMUTimer **timer) 249 { 250 assert(*timer != NULL); 251 252 timer_free(*timer); 253 *timer = NULL; 254 } 255 256 /* Remove timers from event loop */ 257 void throttle_timers_detach_aio_context(ThrottleTimers *tt) 258 { 259 int i; 260 261 for (i = 0; i < THROTTLE_MAX; i++) { 262 throttle_timer_destroy(&tt->timers[i]); 263 } 264 } 265 266 /* To be called last on the ThrottleTimers */ 267 void throttle_timers_destroy(ThrottleTimers *tt) 268 { 269 throttle_timers_detach_aio_context(tt); 270 } 271 272 /* is any throttling timer configured */ 273 bool throttle_timers_are_initialized(ThrottleTimers *tt) 274 { 275 if (tt->timers[0]) { 276 return true; 277 } 278 279 return false; 280 } 281 282 /* Does any throttling must be done 283 * 284 * @cfg: the throttling configuration to inspect 285 * @ret: true if throttling must be done else false 286 */ 287 bool throttle_enabled(ThrottleConfig *cfg) 288 { 289 int i; 290 291 for (i = 0; i < BUCKETS_COUNT; i++) { 292 if (cfg->buckets[i].avg > 0) { 293 return true; 294 } 295 } 296 297 return false; 298 } 299 300 /* check if a throttling configuration is valid 301 * @cfg: the throttling configuration to inspect 302 * @ret: true if valid else false 303 * @errp: error object 304 */ 305 bool throttle_is_valid(ThrottleConfig *cfg, Error **errp) 306 { 307 int i; 308 bool bps_flag, ops_flag; 309 bool bps_max_flag, ops_max_flag; 310 311 bps_flag = cfg->buckets[THROTTLE_BPS_TOTAL].avg && 312 (cfg->buckets[THROTTLE_BPS_READ].avg || 313 cfg->buckets[THROTTLE_BPS_WRITE].avg); 314 315 ops_flag = cfg->buckets[THROTTLE_OPS_TOTAL].avg && 316 (cfg->buckets[THROTTLE_OPS_READ].avg || 317 cfg->buckets[THROTTLE_OPS_WRITE].avg); 318 319 bps_max_flag = cfg->buckets[THROTTLE_BPS_TOTAL].max && 320 (cfg->buckets[THROTTLE_BPS_READ].max || 321 cfg->buckets[THROTTLE_BPS_WRITE].max); 322 323 ops_max_flag = cfg->buckets[THROTTLE_OPS_TOTAL].max && 324 (cfg->buckets[THROTTLE_OPS_READ].max || 325 cfg->buckets[THROTTLE_OPS_WRITE].max); 326 327 if (bps_flag || ops_flag || bps_max_flag || ops_max_flag) { 328 error_setg(errp, "bps/iops/max total values and read/write values" 329 " cannot be used at the same time"); 330 return false; 331 } 332 333 if (cfg->op_size && 334 !cfg->buckets[THROTTLE_OPS_TOTAL].avg && 335 !cfg->buckets[THROTTLE_OPS_READ].avg && 336 !cfg->buckets[THROTTLE_OPS_WRITE].avg) { 337 error_setg(errp, "iops size requires an iops value to be set"); 338 return false; 339 } 340 341 for (i = 0; i < BUCKETS_COUNT; i++) { 342 LeakyBucket *bkt = &cfg->buckets[i]; 343 if (bkt->avg > THROTTLE_VALUE_MAX || bkt->max > THROTTLE_VALUE_MAX) { 344 error_setg(errp, "bps/iops/max values must be within [0, %lld]", 345 THROTTLE_VALUE_MAX); 346 return false; 347 } 348 349 if (!bkt->burst_length) { 350 error_setg(errp, "the burst length cannot be 0"); 351 return false; 352 } 353 354 if (bkt->burst_length > 1 && !bkt->max) { 355 error_setg(errp, "burst length set without burst rate"); 356 return false; 357 } 358 359 if (bkt->max && bkt->burst_length > THROTTLE_VALUE_MAX / bkt->max) { 360 error_setg(errp, "burst length too high for this burst rate"); 361 return false; 362 } 363 364 if (bkt->max && !bkt->avg) { 365 error_setg(errp, "bps_max/iops_max require corresponding" 366 " bps/iops values"); 367 return false; 368 } 369 370 if (bkt->max && bkt->max < bkt->avg) { 371 error_setg(errp, "bps_max/iops_max cannot be lower than bps/iops"); 372 return false; 373 } 374 } 375 376 return true; 377 } 378 379 /* Used to configure the throttle 380 * 381 * @ts: the throttle state we are working on 382 * @clock_type: the group's clock_type 383 * @cfg: the config to set 384 */ 385 void throttle_config(ThrottleState *ts, 386 QEMUClockType clock_type, 387 ThrottleConfig *cfg) 388 { 389 int i; 390 391 ts->cfg = *cfg; 392 393 /* Zero bucket level */ 394 for (i = 0; i < BUCKETS_COUNT; i++) { 395 ts->cfg.buckets[i].level = 0; 396 ts->cfg.buckets[i].burst_level = 0; 397 } 398 399 ts->previous_leak = qemu_clock_get_ns(clock_type); 400 } 401 402 /* used to get config 403 * 404 * @ts: the throttle state we are working on 405 * @cfg: the config to write 406 */ 407 void throttle_get_config(ThrottleState *ts, ThrottleConfig *cfg) 408 { 409 *cfg = ts->cfg; 410 } 411 412 413 /* Schedule the read or write timer if needed 414 * 415 * NOTE: this function is not unit tested due to it's usage of timer_mod 416 * 417 * @tt: the timers structure 418 * @is_write: the type of operation (read/write) 419 * @ret: true if the timer has been scheduled else false 420 */ 421 bool throttle_schedule_timer(ThrottleState *ts, 422 ThrottleTimers *tt, 423 bool is_write) 424 { 425 int64_t now = qemu_clock_get_ns(tt->clock_type); 426 int64_t next_timestamp; 427 bool must_wait; 428 429 must_wait = throttle_compute_timer(ts, 430 is_write, 431 now, 432 &next_timestamp); 433 434 /* request not throttled */ 435 if (!must_wait) { 436 return false; 437 } 438 439 /* request throttled and timer pending -> do nothing */ 440 if (timer_pending(tt->timers[is_write])) { 441 return true; 442 } 443 444 /* request throttled and timer not pending -> arm timer */ 445 timer_mod(tt->timers[is_write], next_timestamp); 446 return true; 447 } 448 449 /* do the accounting for this operation 450 * 451 * @is_write: the type of operation (read/write) 452 * @size: the size of the operation 453 */ 454 void throttle_account(ThrottleState *ts, bool is_write, uint64_t size) 455 { 456 const BucketType bucket_types_size[2][2] = { 457 { THROTTLE_BPS_TOTAL, THROTTLE_BPS_READ }, 458 { THROTTLE_BPS_TOTAL, THROTTLE_BPS_WRITE } 459 }; 460 const BucketType bucket_types_units[2][2] = { 461 { THROTTLE_OPS_TOTAL, THROTTLE_OPS_READ }, 462 { THROTTLE_OPS_TOTAL, THROTTLE_OPS_WRITE } 463 }; 464 double units = 1.0; 465 unsigned i; 466 467 /* if cfg.op_size is defined and smaller than size we compute unit count */ 468 if (ts->cfg.op_size && size > ts->cfg.op_size) { 469 units = (double) size / ts->cfg.op_size; 470 } 471 472 for (i = 0; i < 2; i++) { 473 LeakyBucket *bkt; 474 475 bkt = &ts->cfg.buckets[bucket_types_size[is_write][i]]; 476 bkt->level += size; 477 if (bkt->burst_length > 1) { 478 bkt->burst_level += size; 479 } 480 481 bkt = &ts->cfg.buckets[bucket_types_units[is_write][i]]; 482 bkt->level += units; 483 if (bkt->burst_length > 1) { 484 bkt->burst_level += units; 485 } 486 } 487 } 488 489 /* return a ThrottleConfig based on the options in a ThrottleLimits 490 * 491 * @arg: the ThrottleLimits object to read from 492 * @cfg: the ThrottleConfig to edit 493 * @errp: error object 494 */ 495 void throttle_limits_to_config(ThrottleLimits *arg, ThrottleConfig *cfg, 496 Error **errp) 497 { 498 if (arg->has_bps_total) { 499 cfg->buckets[THROTTLE_BPS_TOTAL].avg = arg->bps_total; 500 } 501 if (arg->has_bps_read) { 502 cfg->buckets[THROTTLE_BPS_READ].avg = arg->bps_read; 503 } 504 if (arg->has_bps_write) { 505 cfg->buckets[THROTTLE_BPS_WRITE].avg = arg->bps_write; 506 } 507 508 if (arg->has_iops_total) { 509 cfg->buckets[THROTTLE_OPS_TOTAL].avg = arg->iops_total; 510 } 511 if (arg->has_iops_read) { 512 cfg->buckets[THROTTLE_OPS_READ].avg = arg->iops_read; 513 } 514 if (arg->has_iops_write) { 515 cfg->buckets[THROTTLE_OPS_WRITE].avg = arg->iops_write; 516 } 517 518 if (arg->has_bps_total_max) { 519 cfg->buckets[THROTTLE_BPS_TOTAL].max = arg->bps_total_max; 520 } 521 if (arg->has_bps_read_max) { 522 cfg->buckets[THROTTLE_BPS_READ].max = arg->bps_read_max; 523 } 524 if (arg->has_bps_write_max) { 525 cfg->buckets[THROTTLE_BPS_WRITE].max = arg->bps_write_max; 526 } 527 if (arg->has_iops_total_max) { 528 cfg->buckets[THROTTLE_OPS_TOTAL].max = arg->iops_total_max; 529 } 530 if (arg->has_iops_read_max) { 531 cfg->buckets[THROTTLE_OPS_READ].max = arg->iops_read_max; 532 } 533 if (arg->has_iops_write_max) { 534 cfg->buckets[THROTTLE_OPS_WRITE].max = arg->iops_write_max; 535 } 536 537 if (arg->has_bps_total_max_length) { 538 if (arg->bps_total_max_length > UINT_MAX) { 539 error_setg(errp, "bps-total-max-length value must be in" 540 " the range [0, %u]", UINT_MAX); 541 return; 542 } 543 cfg->buckets[THROTTLE_BPS_TOTAL].burst_length = arg->bps_total_max_length; 544 } 545 if (arg->has_bps_read_max_length) { 546 if (arg->bps_read_max_length > UINT_MAX) { 547 error_setg(errp, "bps-read-max-length value must be in" 548 " the range [0, %u]", UINT_MAX); 549 return; 550 } 551 cfg->buckets[THROTTLE_BPS_READ].burst_length = arg->bps_read_max_length; 552 } 553 if (arg->has_bps_write_max_length) { 554 if (arg->bps_write_max_length > UINT_MAX) { 555 error_setg(errp, "bps-write-max-length value must be in" 556 " the range [0, %u]", UINT_MAX); 557 return; 558 } 559 cfg->buckets[THROTTLE_BPS_WRITE].burst_length = arg->bps_write_max_length; 560 } 561 if (arg->has_iops_total_max_length) { 562 if (arg->iops_total_max_length > UINT_MAX) { 563 error_setg(errp, "iops-total-max-length value must be in" 564 " the range [0, %u]", UINT_MAX); 565 return; 566 } 567 cfg->buckets[THROTTLE_OPS_TOTAL].burst_length = arg->iops_total_max_length; 568 } 569 if (arg->has_iops_read_max_length) { 570 if (arg->iops_read_max_length > UINT_MAX) { 571 error_setg(errp, "iops-read-max-length value must be in" 572 " the range [0, %u]", UINT_MAX); 573 return; 574 } 575 cfg->buckets[THROTTLE_OPS_READ].burst_length = arg->iops_read_max_length; 576 } 577 if (arg->has_iops_write_max_length) { 578 if (arg->iops_write_max_length > UINT_MAX) { 579 error_setg(errp, "iops-write-max-length value must be in" 580 " the range [0, %u]", UINT_MAX); 581 return; 582 } 583 cfg->buckets[THROTTLE_OPS_WRITE].burst_length = arg->iops_write_max_length; 584 } 585 586 if (arg->has_iops_size) { 587 cfg->op_size = arg->iops_size; 588 } 589 590 throttle_is_valid(cfg, errp); 591 } 592 593 /* write the options of a ThrottleConfig to a ThrottleLimits 594 * 595 * @cfg: the ThrottleConfig to read from 596 * @var: the ThrottleLimits to write to 597 */ 598 void throttle_config_to_limits(ThrottleConfig *cfg, ThrottleLimits *var) 599 { 600 var->bps_total = cfg->buckets[THROTTLE_BPS_TOTAL].avg; 601 var->bps_read = cfg->buckets[THROTTLE_BPS_READ].avg; 602 var->bps_write = cfg->buckets[THROTTLE_BPS_WRITE].avg; 603 var->iops_total = cfg->buckets[THROTTLE_OPS_TOTAL].avg; 604 var->iops_read = cfg->buckets[THROTTLE_OPS_READ].avg; 605 var->iops_write = cfg->buckets[THROTTLE_OPS_WRITE].avg; 606 var->bps_total_max = cfg->buckets[THROTTLE_BPS_TOTAL].max; 607 var->bps_read_max = cfg->buckets[THROTTLE_BPS_READ].max; 608 var->bps_write_max = cfg->buckets[THROTTLE_BPS_WRITE].max; 609 var->iops_total_max = cfg->buckets[THROTTLE_OPS_TOTAL].max; 610 var->iops_read_max = cfg->buckets[THROTTLE_OPS_READ].max; 611 var->iops_write_max = cfg->buckets[THROTTLE_OPS_WRITE].max; 612 var->bps_total_max_length = cfg->buckets[THROTTLE_BPS_TOTAL].burst_length; 613 var->bps_read_max_length = cfg->buckets[THROTTLE_BPS_READ].burst_length; 614 var->bps_write_max_length = cfg->buckets[THROTTLE_BPS_WRITE].burst_length; 615 var->iops_total_max_length = cfg->buckets[THROTTLE_OPS_TOTAL].burst_length; 616 var->iops_read_max_length = cfg->buckets[THROTTLE_OPS_READ].burst_length; 617 var->iops_write_max_length = cfg->buckets[THROTTLE_OPS_WRITE].burst_length; 618 var->iops_size = cfg->op_size; 619 620 var->has_bps_total = true; 621 var->has_bps_read = true; 622 var->has_bps_write = true; 623 var->has_iops_total = true; 624 var->has_iops_read = true; 625 var->has_iops_write = true; 626 var->has_bps_total_max = true; 627 var->has_bps_read_max = true; 628 var->has_bps_write_max = true; 629 var->has_iops_total_max = true; 630 var->has_iops_read_max = true; 631 var->has_iops_write_max = true; 632 var->has_bps_read_max_length = true; 633 var->has_bps_total_max_length = true; 634 var->has_bps_write_max_length = true; 635 var->has_iops_total_max_length = true; 636 var->has_iops_read_max_length = true; 637 var->has_iops_write_max_length = true; 638 var->has_iops_size = true; 639 } 640