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 "qemu/throttle.h" 27 #include "qemu/timer.h" 28 #include "block/aio.h" 29 30 /* This function make a bucket leak 31 * 32 * @bkt: the bucket to make leak 33 * @delta_ns: the time delta 34 */ 35 void throttle_leak_bucket(LeakyBucket *bkt, int64_t delta_ns) 36 { 37 double leak; 38 39 /* compute how much to leak */ 40 leak = (bkt->avg * (double) delta_ns) / NANOSECONDS_PER_SECOND; 41 42 /* make the bucket leak */ 43 bkt->level = MAX(bkt->level - leak, 0); 44 45 /* if we allow bursts for more than one second we also need to 46 * keep track of bkt->burst_level so the bkt->max goal per second 47 * is attained */ 48 if (bkt->burst_length > 1) { 49 leak = (bkt->max * (double) delta_ns) / NANOSECONDS_PER_SECOND; 50 bkt->burst_level = MAX(bkt->burst_level - leak, 0); 51 } 52 } 53 54 /* Calculate the time delta since last leak and make proportionals leaks 55 * 56 * @now: the current timestamp in ns 57 */ 58 static void throttle_do_leak(ThrottleState *ts, int64_t now) 59 { 60 /* compute the time elapsed since the last leak */ 61 int64_t delta_ns = now - ts->previous_leak; 62 int i; 63 64 ts->previous_leak = now; 65 66 if (delta_ns <= 0) { 67 return; 68 } 69 70 /* make each bucket leak */ 71 for (i = 0; i < BUCKETS_COUNT; i++) { 72 throttle_leak_bucket(&ts->cfg.buckets[i], delta_ns); 73 } 74 } 75 76 /* do the real job of computing the time to wait 77 * 78 * @limit: the throttling limit 79 * @extra: the number of operation to delay 80 * @ret: the time to wait in ns 81 */ 82 static int64_t throttle_do_compute_wait(double limit, double extra) 83 { 84 double wait = extra * NANOSECONDS_PER_SECOND; 85 wait /= limit; 86 return wait; 87 } 88 89 /* This function compute the wait time in ns that a leaky bucket should trigger 90 * 91 * @bkt: the leaky bucket we operate on 92 * @ret: the resulting wait time in ns or 0 if the operation can go through 93 */ 94 int64_t throttle_compute_wait(LeakyBucket *bkt) 95 { 96 double extra; /* the number of extra units blocking the io */ 97 98 if (!bkt->avg) { 99 return 0; 100 } 101 102 /* If the bucket is full then we have to wait */ 103 extra = bkt->level - bkt->max * bkt->burst_length; 104 if (extra > 0) { 105 return throttle_do_compute_wait(bkt->avg, extra); 106 } 107 108 /* If the bucket is not full yet we have to make sure that we 109 * fulfill the goal of bkt->max units per second. */ 110 if (bkt->burst_length > 1) { 111 /* We use 1/10 of the max value to smooth the throttling. 112 * See throttle_fix_bucket() for more details. */ 113 extra = bkt->burst_level - bkt->max / 10; 114 if (extra > 0) { 115 return throttle_do_compute_wait(bkt->max, extra); 116 } 117 } 118 119 return 0; 120 } 121 122 /* This function compute the time that must be waited while this IO 123 * 124 * @is_write: true if the current IO is a write, false if it's a read 125 * @ret: time to wait 126 */ 127 static int64_t throttle_compute_wait_for(ThrottleState *ts, 128 bool is_write) 129 { 130 BucketType to_check[2][4] = { {THROTTLE_BPS_TOTAL, 131 THROTTLE_OPS_TOTAL, 132 THROTTLE_BPS_READ, 133 THROTTLE_OPS_READ}, 134 {THROTTLE_BPS_TOTAL, 135 THROTTLE_OPS_TOTAL, 136 THROTTLE_BPS_WRITE, 137 THROTTLE_OPS_WRITE}, }; 138 int64_t wait, max_wait = 0; 139 int i; 140 141 for (i = 0; i < 4; i++) { 142 BucketType index = to_check[is_write][i]; 143 wait = throttle_compute_wait(&ts->cfg.buckets[index]); 144 if (wait > max_wait) { 145 max_wait = wait; 146 } 147 } 148 149 return max_wait; 150 } 151 152 /* compute the timer for this type of operation 153 * 154 * @is_write: the type of operation 155 * @now: the current clock timestamp 156 * @next_timestamp: the resulting timer 157 * @ret: true if a timer must be set 158 */ 159 static bool throttle_compute_timer(ThrottleState *ts, 160 bool is_write, 161 int64_t now, 162 int64_t *next_timestamp) 163 { 164 int64_t wait; 165 166 /* leak proportionally to the time elapsed */ 167 throttle_do_leak(ts, now); 168 169 /* compute the wait time if any */ 170 wait = throttle_compute_wait_for(ts, is_write); 171 172 /* if the code must wait compute when the next timer should fire */ 173 if (wait) { 174 *next_timestamp = now + wait; 175 return true; 176 } 177 178 /* else no need to wait at all */ 179 *next_timestamp = now; 180 return false; 181 } 182 183 /* Add timers to event loop */ 184 void throttle_timers_attach_aio_context(ThrottleTimers *tt, 185 AioContext *new_context) 186 { 187 tt->timers[0] = aio_timer_new(new_context, tt->clock_type, SCALE_NS, 188 tt->read_timer_cb, tt->timer_opaque); 189 tt->timers[1] = aio_timer_new(new_context, tt->clock_type, SCALE_NS, 190 tt->write_timer_cb, tt->timer_opaque); 191 } 192 193 /* 194 * Initialize the ThrottleConfig structure to a valid state 195 * @cfg: the config to initialize 196 */ 197 void throttle_config_init(ThrottleConfig *cfg) 198 { 199 unsigned i; 200 memset(cfg, 0, sizeof(*cfg)); 201 for (i = 0; i < BUCKETS_COUNT; i++) { 202 cfg->buckets[i].burst_length = 1; 203 } 204 } 205 206 /* To be called first on the ThrottleState */ 207 void throttle_init(ThrottleState *ts) 208 { 209 memset(ts, 0, sizeof(ThrottleState)); 210 throttle_config_init(&ts->cfg); 211 } 212 213 /* To be called first on the ThrottleTimers */ 214 void throttle_timers_init(ThrottleTimers *tt, 215 AioContext *aio_context, 216 QEMUClockType clock_type, 217 QEMUTimerCB *read_timer_cb, 218 QEMUTimerCB *write_timer_cb, 219 void *timer_opaque) 220 { 221 memset(tt, 0, sizeof(ThrottleTimers)); 222 223 tt->clock_type = clock_type; 224 tt->read_timer_cb = read_timer_cb; 225 tt->write_timer_cb = write_timer_cb; 226 tt->timer_opaque = timer_opaque; 227 throttle_timers_attach_aio_context(tt, aio_context); 228 } 229 230 /* destroy a timer */ 231 static void throttle_timer_destroy(QEMUTimer **timer) 232 { 233 assert(*timer != NULL); 234 235 timer_del(*timer); 236 timer_free(*timer); 237 *timer = NULL; 238 } 239 240 /* Remove timers from event loop */ 241 void throttle_timers_detach_aio_context(ThrottleTimers *tt) 242 { 243 int i; 244 245 for (i = 0; i < 2; i++) { 246 throttle_timer_destroy(&tt->timers[i]); 247 } 248 } 249 250 /* To be called last on the ThrottleTimers */ 251 void throttle_timers_destroy(ThrottleTimers *tt) 252 { 253 throttle_timers_detach_aio_context(tt); 254 } 255 256 /* is any throttling timer configured */ 257 bool throttle_timers_are_initialized(ThrottleTimers *tt) 258 { 259 if (tt->timers[0]) { 260 return true; 261 } 262 263 return false; 264 } 265 266 /* Does any throttling must be done 267 * 268 * @cfg: the throttling configuration to inspect 269 * @ret: true if throttling must be done else false 270 */ 271 bool throttle_enabled(ThrottleConfig *cfg) 272 { 273 int i; 274 275 for (i = 0; i < BUCKETS_COUNT; i++) { 276 if (cfg->buckets[i].avg > 0) { 277 return true; 278 } 279 } 280 281 return false; 282 } 283 284 /* check if a throttling configuration is valid 285 * @cfg: the throttling configuration to inspect 286 * @ret: true if valid else false 287 * @errp: error object 288 */ 289 bool throttle_is_valid(ThrottleConfig *cfg, Error **errp) 290 { 291 int i; 292 bool bps_flag, ops_flag; 293 bool bps_max_flag, ops_max_flag; 294 295 bps_flag = cfg->buckets[THROTTLE_BPS_TOTAL].avg && 296 (cfg->buckets[THROTTLE_BPS_READ].avg || 297 cfg->buckets[THROTTLE_BPS_WRITE].avg); 298 299 ops_flag = cfg->buckets[THROTTLE_OPS_TOTAL].avg && 300 (cfg->buckets[THROTTLE_OPS_READ].avg || 301 cfg->buckets[THROTTLE_OPS_WRITE].avg); 302 303 bps_max_flag = cfg->buckets[THROTTLE_BPS_TOTAL].max && 304 (cfg->buckets[THROTTLE_BPS_READ].max || 305 cfg->buckets[THROTTLE_BPS_WRITE].max); 306 307 ops_max_flag = cfg->buckets[THROTTLE_OPS_TOTAL].max && 308 (cfg->buckets[THROTTLE_OPS_READ].max || 309 cfg->buckets[THROTTLE_OPS_WRITE].max); 310 311 if (bps_flag || ops_flag || bps_max_flag || ops_max_flag) { 312 error_setg(errp, "bps/iops/max total values and read/write values" 313 " cannot be used at the same time"); 314 return false; 315 } 316 317 for (i = 0; i < BUCKETS_COUNT; i++) { 318 if (cfg->buckets[i].avg < 0 || 319 cfg->buckets[i].max < 0 || 320 cfg->buckets[i].avg > THROTTLE_VALUE_MAX || 321 cfg->buckets[i].max > THROTTLE_VALUE_MAX) { 322 error_setg(errp, "bps/iops/max values must be within [0, %lld]", 323 THROTTLE_VALUE_MAX); 324 return false; 325 } 326 327 if (!cfg->buckets[i].burst_length) { 328 error_setg(errp, "the burst length cannot be 0"); 329 return false; 330 } 331 332 if (cfg->buckets[i].burst_length > 1 && !cfg->buckets[i].max) { 333 error_setg(errp, "burst length set without burst rate"); 334 return false; 335 } 336 337 if (cfg->buckets[i].max && !cfg->buckets[i].avg) { 338 error_setg(errp, "bps_max/iops_max require corresponding" 339 " bps/iops values"); 340 return false; 341 } 342 } 343 344 return true; 345 } 346 347 /* fix bucket parameters */ 348 static void throttle_fix_bucket(LeakyBucket *bkt) 349 { 350 double min; 351 352 /* zero bucket level */ 353 bkt->level = bkt->burst_level = 0; 354 355 /* The following is done to cope with the Linux CFQ block scheduler 356 * which regroup reads and writes by block of 100ms in the guest. 357 * When they are two process one making reads and one making writes cfq 358 * make a pattern looking like the following: 359 * WWWWWWWWWWWRRRRRRRRRRRRRRWWWWWWWWWWWWWwRRRRRRRRRRRRRRRRR 360 * Having a max burst value of 100ms of the average will help smooth the 361 * throttling 362 */ 363 min = bkt->avg / 10; 364 if (bkt->avg && !bkt->max) { 365 bkt->max = min; 366 } 367 } 368 369 /* take care of canceling a timer */ 370 static void throttle_cancel_timer(QEMUTimer *timer) 371 { 372 assert(timer != NULL); 373 374 timer_del(timer); 375 } 376 377 /* Used to configure the throttle 378 * 379 * @ts: the throttle state we are working on 380 * @tt: the throttle timers we use in this aio context 381 * @cfg: the config to set 382 */ 383 void throttle_config(ThrottleState *ts, 384 ThrottleTimers *tt, 385 ThrottleConfig *cfg) 386 { 387 int i; 388 389 ts->cfg = *cfg; 390 391 for (i = 0; i < BUCKETS_COUNT; i++) { 392 throttle_fix_bucket(&ts->cfg.buckets[i]); 393 } 394 395 ts->previous_leak = qemu_clock_get_ns(tt->clock_type); 396 397 for (i = 0; i < 2; i++) { 398 throttle_cancel_timer(tt->timers[i]); 399 } 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