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 99 if (!bkt->avg) { 100 return 0; 101 } 102 103 /* If the bucket is full then we have to wait */ 104 extra = bkt->level - bkt->max * bkt->burst_length; 105 if (extra > 0) { 106 return throttle_do_compute_wait(bkt->avg, extra); 107 } 108 109 /* If the bucket is not full yet we have to make sure that we 110 * fulfill the goal of bkt->max units per second. */ 111 if (bkt->burst_length > 1) { 112 /* We use 1/10 of the max value to smooth the throttling. 113 * See throttle_fix_bucket() for more details. */ 114 extra = bkt->burst_level - bkt->max / 10; 115 if (extra > 0) { 116 return throttle_do_compute_wait(bkt->max, extra); 117 } 118 } 119 120 return 0; 121 } 122 123 /* This function compute the time that must be waited while this IO 124 * 125 * @is_write: true if the current IO is a write, false if it's a read 126 * @ret: time to wait 127 */ 128 static int64_t throttle_compute_wait_for(ThrottleState *ts, 129 bool is_write) 130 { 131 BucketType to_check[2][4] = { {THROTTLE_BPS_TOTAL, 132 THROTTLE_OPS_TOTAL, 133 THROTTLE_BPS_READ, 134 THROTTLE_OPS_READ}, 135 {THROTTLE_BPS_TOTAL, 136 THROTTLE_OPS_TOTAL, 137 THROTTLE_BPS_WRITE, 138 THROTTLE_OPS_WRITE}, }; 139 int64_t wait, max_wait = 0; 140 int i; 141 142 for (i = 0; i < 4; i++) { 143 BucketType index = to_check[is_write][i]; 144 wait = throttle_compute_wait(&ts->cfg.buckets[index]); 145 if (wait > max_wait) { 146 max_wait = wait; 147 } 148 } 149 150 return max_wait; 151 } 152 153 /* compute the timer for this type of operation 154 * 155 * @is_write: the type of operation 156 * @now: the current clock timestamp 157 * @next_timestamp: the resulting timer 158 * @ret: true if a timer must be set 159 */ 160 static bool throttle_compute_timer(ThrottleState *ts, 161 bool is_write, 162 int64_t now, 163 int64_t *next_timestamp) 164 { 165 int64_t wait; 166 167 /* leak proportionally to the time elapsed */ 168 throttle_do_leak(ts, now); 169 170 /* compute the wait time if any */ 171 wait = throttle_compute_wait_for(ts, is_write); 172 173 /* if the code must wait compute when the next timer should fire */ 174 if (wait) { 175 *next_timestamp = now + wait; 176 return true; 177 } 178 179 /* else no need to wait at all */ 180 *next_timestamp = now; 181 return false; 182 } 183 184 /* Add timers to event loop */ 185 void throttle_timers_attach_aio_context(ThrottleTimers *tt, 186 AioContext *new_context) 187 { 188 tt->timers[0] = aio_timer_new(new_context, tt->clock_type, SCALE_NS, 189 tt->read_timer_cb, tt->timer_opaque); 190 tt->timers[1] = aio_timer_new(new_context, tt->clock_type, SCALE_NS, 191 tt->write_timer_cb, tt->timer_opaque); 192 } 193 194 /* 195 * Initialize the ThrottleConfig structure to a valid state 196 * @cfg: the config to initialize 197 */ 198 void throttle_config_init(ThrottleConfig *cfg) 199 { 200 unsigned i; 201 memset(cfg, 0, sizeof(*cfg)); 202 for (i = 0; i < BUCKETS_COUNT; i++) { 203 cfg->buckets[i].burst_length = 1; 204 } 205 } 206 207 /* To be called first on the ThrottleState */ 208 void throttle_init(ThrottleState *ts) 209 { 210 memset(ts, 0, sizeof(ThrottleState)); 211 throttle_config_init(&ts->cfg); 212 } 213 214 /* To be called first on the ThrottleTimers */ 215 void throttle_timers_init(ThrottleTimers *tt, 216 AioContext *aio_context, 217 QEMUClockType clock_type, 218 QEMUTimerCB *read_timer_cb, 219 QEMUTimerCB *write_timer_cb, 220 void *timer_opaque) 221 { 222 memset(tt, 0, sizeof(ThrottleTimers)); 223 224 tt->clock_type = clock_type; 225 tt->read_timer_cb = read_timer_cb; 226 tt->write_timer_cb = write_timer_cb; 227 tt->timer_opaque = timer_opaque; 228 throttle_timers_attach_aio_context(tt, aio_context); 229 } 230 231 /* destroy a timer */ 232 static void throttle_timer_destroy(QEMUTimer **timer) 233 { 234 assert(*timer != NULL); 235 236 timer_del(*timer); 237 timer_free(*timer); 238 *timer = NULL; 239 } 240 241 /* Remove timers from event loop */ 242 void throttle_timers_detach_aio_context(ThrottleTimers *tt) 243 { 244 int i; 245 246 for (i = 0; i < 2; i++) { 247 throttle_timer_destroy(&tt->timers[i]); 248 } 249 } 250 251 /* To be called last on the ThrottleTimers */ 252 void throttle_timers_destroy(ThrottleTimers *tt) 253 { 254 throttle_timers_detach_aio_context(tt); 255 } 256 257 /* is any throttling timer configured */ 258 bool throttle_timers_are_initialized(ThrottleTimers *tt) 259 { 260 if (tt->timers[0]) { 261 return true; 262 } 263 264 return false; 265 } 266 267 /* Does any throttling must be done 268 * 269 * @cfg: the throttling configuration to inspect 270 * @ret: true if throttling must be done else false 271 */ 272 bool throttle_enabled(ThrottleConfig *cfg) 273 { 274 int i; 275 276 for (i = 0; i < BUCKETS_COUNT; i++) { 277 if (cfg->buckets[i].avg > 0) { 278 return true; 279 } 280 } 281 282 return false; 283 } 284 285 /* check if a throttling configuration is valid 286 * @cfg: the throttling configuration to inspect 287 * @ret: true if valid else false 288 * @errp: error object 289 */ 290 bool throttle_is_valid(ThrottleConfig *cfg, Error **errp) 291 { 292 int i; 293 bool bps_flag, ops_flag; 294 bool bps_max_flag, ops_max_flag; 295 296 bps_flag = cfg->buckets[THROTTLE_BPS_TOTAL].avg && 297 (cfg->buckets[THROTTLE_BPS_READ].avg || 298 cfg->buckets[THROTTLE_BPS_WRITE].avg); 299 300 ops_flag = cfg->buckets[THROTTLE_OPS_TOTAL].avg && 301 (cfg->buckets[THROTTLE_OPS_READ].avg || 302 cfg->buckets[THROTTLE_OPS_WRITE].avg); 303 304 bps_max_flag = cfg->buckets[THROTTLE_BPS_TOTAL].max && 305 (cfg->buckets[THROTTLE_BPS_READ].max || 306 cfg->buckets[THROTTLE_BPS_WRITE].max); 307 308 ops_max_flag = cfg->buckets[THROTTLE_OPS_TOTAL].max && 309 (cfg->buckets[THROTTLE_OPS_READ].max || 310 cfg->buckets[THROTTLE_OPS_WRITE].max); 311 312 if (bps_flag || ops_flag || bps_max_flag || ops_max_flag) { 313 error_setg(errp, "bps/iops/max total values and read/write values" 314 " cannot be used at the same time"); 315 return false; 316 } 317 318 if (cfg->op_size && 319 !cfg->buckets[THROTTLE_OPS_TOTAL].avg && 320 !cfg->buckets[THROTTLE_OPS_READ].avg && 321 !cfg->buckets[THROTTLE_OPS_WRITE].avg) { 322 error_setg(errp, "iops size requires an iops value to be set"); 323 return false; 324 } 325 326 for (i = 0; i < BUCKETS_COUNT; i++) { 327 if (cfg->buckets[i].avg < 0 || 328 cfg->buckets[i].max < 0 || 329 cfg->buckets[i].avg > THROTTLE_VALUE_MAX || 330 cfg->buckets[i].max > THROTTLE_VALUE_MAX) { 331 error_setg(errp, "bps/iops/max values must be within [0, %lld]", 332 THROTTLE_VALUE_MAX); 333 return false; 334 } 335 336 if (!cfg->buckets[i].burst_length) { 337 error_setg(errp, "the burst length cannot be 0"); 338 return false; 339 } 340 341 if (cfg->buckets[i].burst_length > 1 && !cfg->buckets[i].max) { 342 error_setg(errp, "burst length set without burst rate"); 343 return false; 344 } 345 346 if (cfg->buckets[i].max && !cfg->buckets[i].avg) { 347 error_setg(errp, "bps_max/iops_max require corresponding" 348 " bps/iops values"); 349 return false; 350 } 351 352 if (cfg->buckets[i].max && cfg->buckets[i].max < cfg->buckets[i].avg) { 353 error_setg(errp, "bps_max/iops_max cannot be lower than bps/iops"); 354 return false; 355 } 356 } 357 358 return true; 359 } 360 361 /* fix bucket parameters */ 362 static void throttle_fix_bucket(LeakyBucket *bkt) 363 { 364 double min; 365 366 /* zero bucket level */ 367 bkt->level = bkt->burst_level = 0; 368 369 /* The following is done to cope with the Linux CFQ block scheduler 370 * which regroup reads and writes by block of 100ms in the guest. 371 * When they are two process one making reads and one making writes cfq 372 * make a pattern looking like the following: 373 * WWWWWWWWWWWRRRRRRRRRRRRRRWWWWWWWWWWWWWwRRRRRRRRRRRRRRRRR 374 * Having a max burst value of 100ms of the average will help smooth the 375 * throttling 376 */ 377 min = bkt->avg / 10; 378 if (bkt->avg && !bkt->max) { 379 bkt->max = min; 380 } 381 } 382 383 /* undo internal bucket parameter changes (see throttle_fix_bucket()) */ 384 static void throttle_unfix_bucket(LeakyBucket *bkt) 385 { 386 if (bkt->max < bkt->avg) { 387 bkt->max = 0; 388 } 389 } 390 391 /* take care of canceling a timer */ 392 static void throttle_cancel_timer(QEMUTimer *timer) 393 { 394 assert(timer != NULL); 395 396 timer_del(timer); 397 } 398 399 /* Used to configure the throttle 400 * 401 * @ts: the throttle state we are working on 402 * @tt: the throttle timers we use in this aio context 403 * @cfg: the config to set 404 */ 405 void throttle_config(ThrottleState *ts, 406 ThrottleTimers *tt, 407 ThrottleConfig *cfg) 408 { 409 int i; 410 411 ts->cfg = *cfg; 412 413 for (i = 0; i < BUCKETS_COUNT; i++) { 414 throttle_fix_bucket(&ts->cfg.buckets[i]); 415 } 416 417 ts->previous_leak = qemu_clock_get_ns(tt->clock_type); 418 419 for (i = 0; i < 2; i++) { 420 throttle_cancel_timer(tt->timers[i]); 421 } 422 } 423 424 /* used to get config 425 * 426 * @ts: the throttle state we are working on 427 * @cfg: the config to write 428 */ 429 void throttle_get_config(ThrottleState *ts, ThrottleConfig *cfg) 430 { 431 int i; 432 433 *cfg = ts->cfg; 434 435 for (i = 0; i < BUCKETS_COUNT; i++) { 436 throttle_unfix_bucket(&cfg->buckets[i]); 437 } 438 } 439 440 441 /* Schedule the read or write timer if needed 442 * 443 * NOTE: this function is not unit tested due to it's usage of timer_mod 444 * 445 * @tt: the timers structure 446 * @is_write: the type of operation (read/write) 447 * @ret: true if the timer has been scheduled else false 448 */ 449 bool throttle_schedule_timer(ThrottleState *ts, 450 ThrottleTimers *tt, 451 bool is_write) 452 { 453 int64_t now = qemu_clock_get_ns(tt->clock_type); 454 int64_t next_timestamp; 455 bool must_wait; 456 457 must_wait = throttle_compute_timer(ts, 458 is_write, 459 now, 460 &next_timestamp); 461 462 /* request not throttled */ 463 if (!must_wait) { 464 return false; 465 } 466 467 /* request throttled and timer pending -> do nothing */ 468 if (timer_pending(tt->timers[is_write])) { 469 return true; 470 } 471 472 /* request throttled and timer not pending -> arm timer */ 473 timer_mod(tt->timers[is_write], next_timestamp); 474 return true; 475 } 476 477 /* do the accounting for this operation 478 * 479 * @is_write: the type of operation (read/write) 480 * @size: the size of the operation 481 */ 482 void throttle_account(ThrottleState *ts, bool is_write, uint64_t size) 483 { 484 const BucketType bucket_types_size[2][2] = { 485 { THROTTLE_BPS_TOTAL, THROTTLE_BPS_READ }, 486 { THROTTLE_BPS_TOTAL, THROTTLE_BPS_WRITE } 487 }; 488 const BucketType bucket_types_units[2][2] = { 489 { THROTTLE_OPS_TOTAL, THROTTLE_OPS_READ }, 490 { THROTTLE_OPS_TOTAL, THROTTLE_OPS_WRITE } 491 }; 492 double units = 1.0; 493 unsigned i; 494 495 /* if cfg.op_size is defined and smaller than size we compute unit count */ 496 if (ts->cfg.op_size && size > ts->cfg.op_size) { 497 units = (double) size / ts->cfg.op_size; 498 } 499 500 for (i = 0; i < 2; i++) { 501 LeakyBucket *bkt; 502 503 bkt = &ts->cfg.buckets[bucket_types_size[is_write][i]]; 504 bkt->level += size; 505 if (bkt->burst_length > 1) { 506 bkt->burst_level += size; 507 } 508 509 bkt = &ts->cfg.buckets[bucket_types_units[is_write][i]]; 510 bkt->level += units; 511 if (bkt->burst_length > 1) { 512 bkt->burst_level += units; 513 } 514 } 515 } 516 517