Benoît Canet | 5ddfffb | 2013-09-02 14:14:37 +0200 | [diff] [blame] | 1 | /* |
| 2 | * QEMU throttling infrastructure |
| 3 | * |
| 4 | * Copyright (C) Nodalink, SARL. 2013 |
| 5 | * |
| 6 | * Author: |
| 7 | * Benoît Canet <benoit.canet@irqsave.net> |
| 8 | * |
| 9 | * This program is free software; you can redistribute it and/or |
| 10 | * modify it under the terms of the GNU General Public License as |
| 11 | * published by the Free Software Foundation; either version 2 or |
| 12 | * (at your option) version 3 of the License. |
| 13 | * |
| 14 | * This program is distributed in the hope that it will be useful, |
| 15 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 16 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 17 | * GNU General Public License for more details. |
| 18 | * |
| 19 | * You should have received a copy of the GNU General Public License |
| 20 | * along with this program; if not, see <http://www.gnu.org/licenses/>. |
| 21 | */ |
| 22 | |
| 23 | #include "qemu/throttle.h" |
| 24 | #include "qemu/timer.h" |
| 25 | |
| 26 | /* This function make a bucket leak |
| 27 | * |
| 28 | * @bkt: the bucket to make leak |
| 29 | * @delta_ns: the time delta |
| 30 | */ |
| 31 | void throttle_leak_bucket(LeakyBucket *bkt, int64_t delta_ns) |
| 32 | { |
| 33 | double leak; |
| 34 | |
| 35 | /* compute how much to leak */ |
| 36 | leak = (bkt->avg * (double) delta_ns) / NANOSECONDS_PER_SECOND; |
| 37 | |
| 38 | /* make the bucket leak */ |
| 39 | bkt->level = MAX(bkt->level - leak, 0); |
| 40 | } |
| 41 | |
| 42 | /* Calculate the time delta since last leak and make proportionals leaks |
| 43 | * |
| 44 | * @now: the current timestamp in ns |
| 45 | */ |
| 46 | static void throttle_do_leak(ThrottleState *ts, int64_t now) |
| 47 | { |
| 48 | /* compute the time elapsed since the last leak */ |
| 49 | int64_t delta_ns = now - ts->previous_leak; |
| 50 | int i; |
| 51 | |
| 52 | ts->previous_leak = now; |
| 53 | |
| 54 | if (delta_ns <= 0) { |
| 55 | return; |
| 56 | } |
| 57 | |
| 58 | /* make each bucket leak */ |
| 59 | for (i = 0; i < BUCKETS_COUNT; i++) { |
| 60 | throttle_leak_bucket(&ts->cfg.buckets[i], delta_ns); |
| 61 | } |
| 62 | } |
| 63 | |
| 64 | /* do the real job of computing the time to wait |
| 65 | * |
| 66 | * @limit: the throttling limit |
| 67 | * @extra: the number of operation to delay |
| 68 | * @ret: the time to wait in ns |
| 69 | */ |
| 70 | static int64_t throttle_do_compute_wait(double limit, double extra) |
| 71 | { |
| 72 | double wait = extra * NANOSECONDS_PER_SECOND; |
| 73 | wait /= limit; |
| 74 | return wait; |
| 75 | } |
| 76 | |
| 77 | /* This function compute the wait time in ns that a leaky bucket should trigger |
| 78 | * |
| 79 | * @bkt: the leaky bucket we operate on |
| 80 | * @ret: the resulting wait time in ns or 0 if the operation can go through |
| 81 | */ |
| 82 | int64_t throttle_compute_wait(LeakyBucket *bkt) |
| 83 | { |
| 84 | double extra; /* the number of extra units blocking the io */ |
| 85 | |
| 86 | if (!bkt->avg) { |
| 87 | return 0; |
| 88 | } |
| 89 | |
| 90 | extra = bkt->level - bkt->max; |
| 91 | |
| 92 | if (extra <= 0) { |
| 93 | return 0; |
| 94 | } |
| 95 | |
| 96 | return throttle_do_compute_wait(bkt->avg, extra); |
| 97 | } |
| 98 | |
| 99 | /* This function compute the time that must be waited while this IO |
| 100 | * |
| 101 | * @is_write: true if the current IO is a write, false if it's a read |
| 102 | * @ret: time to wait |
| 103 | */ |
| 104 | static int64_t throttle_compute_wait_for(ThrottleState *ts, |
| 105 | bool is_write) |
| 106 | { |
| 107 | BucketType to_check[2][4] = { {THROTTLE_BPS_TOTAL, |
| 108 | THROTTLE_OPS_TOTAL, |
| 109 | THROTTLE_BPS_READ, |
| 110 | THROTTLE_OPS_READ}, |
| 111 | {THROTTLE_BPS_TOTAL, |
| 112 | THROTTLE_OPS_TOTAL, |
| 113 | THROTTLE_BPS_WRITE, |
| 114 | THROTTLE_OPS_WRITE}, }; |
| 115 | int64_t wait, max_wait = 0; |
| 116 | int i; |
| 117 | |
| 118 | for (i = 0; i < 4; i++) { |
| 119 | BucketType index = to_check[is_write][i]; |
| 120 | wait = throttle_compute_wait(&ts->cfg.buckets[index]); |
| 121 | if (wait > max_wait) { |
| 122 | max_wait = wait; |
| 123 | } |
| 124 | } |
| 125 | |
| 126 | return max_wait; |
| 127 | } |
| 128 | |
| 129 | /* compute the timer for this type of operation |
| 130 | * |
| 131 | * @is_write: the type of operation |
| 132 | * @now: the current clock timestamp |
| 133 | * @next_timestamp: the resulting timer |
| 134 | * @ret: true if a timer must be set |
| 135 | */ |
| 136 | bool throttle_compute_timer(ThrottleState *ts, |
| 137 | bool is_write, |
| 138 | int64_t now, |
| 139 | int64_t *next_timestamp) |
| 140 | { |
| 141 | int64_t wait; |
| 142 | |
| 143 | /* leak proportionally to the time elapsed */ |
| 144 | throttle_do_leak(ts, now); |
| 145 | |
| 146 | /* compute the wait time if any */ |
| 147 | wait = throttle_compute_wait_for(ts, is_write); |
| 148 | |
| 149 | /* if the code must wait compute when the next timer should fire */ |
| 150 | if (wait) { |
| 151 | *next_timestamp = now + wait; |
| 152 | return true; |
| 153 | } |
| 154 | |
| 155 | /* else no need to wait at all */ |
| 156 | *next_timestamp = now; |
| 157 | return false; |
| 158 | } |
| 159 | |
| 160 | /* To be called first on the ThrottleState */ |
| 161 | void throttle_init(ThrottleState *ts, |
| 162 | QEMUClockType clock_type, |
| 163 | QEMUTimerCB *read_timer_cb, |
| 164 | QEMUTimerCB *write_timer_cb, |
| 165 | void *timer_opaque) |
| 166 | { |
| 167 | memset(ts, 0, sizeof(ThrottleState)); |
| 168 | |
| 169 | ts->clock_type = clock_type; |
| 170 | ts->timers[0] = timer_new_ns(clock_type, read_timer_cb, timer_opaque); |
| 171 | ts->timers[1] = timer_new_ns(clock_type, write_timer_cb, timer_opaque); |
| 172 | } |
| 173 | |
| 174 | /* destroy a timer */ |
| 175 | static void throttle_timer_destroy(QEMUTimer **timer) |
| 176 | { |
| 177 | assert(*timer != NULL); |
| 178 | |
| 179 | timer_del(*timer); |
| 180 | timer_free(*timer); |
| 181 | *timer = NULL; |
| 182 | } |
| 183 | |
| 184 | /* To be called last on the ThrottleState */ |
| 185 | void throttle_destroy(ThrottleState *ts) |
| 186 | { |
| 187 | int i; |
| 188 | |
| 189 | for (i = 0; i < 2; i++) { |
| 190 | throttle_timer_destroy(&ts->timers[i]); |
| 191 | } |
| 192 | } |
| 193 | |
| 194 | /* is any throttling timer configured */ |
| 195 | bool throttle_have_timer(ThrottleState *ts) |
| 196 | { |
| 197 | if (ts->timers[0]) { |
| 198 | return true; |
| 199 | } |
| 200 | |
| 201 | return false; |
| 202 | } |
| 203 | |
| 204 | /* Does any throttling must be done |
| 205 | * |
| 206 | * @cfg: the throttling configuration to inspect |
| 207 | * @ret: true if throttling must be done else false |
| 208 | */ |
| 209 | bool throttle_enabled(ThrottleConfig *cfg) |
| 210 | { |
| 211 | int i; |
| 212 | |
| 213 | for (i = 0; i < BUCKETS_COUNT; i++) { |
| 214 | if (cfg->buckets[i].avg > 0) { |
| 215 | return true; |
| 216 | } |
| 217 | } |
| 218 | |
| 219 | return false; |
| 220 | } |
| 221 | |
| 222 | /* return true if any two throttling parameters conflicts |
| 223 | * |
| 224 | * @cfg: the throttling configuration to inspect |
| 225 | * @ret: true if any conflict detected else false |
| 226 | */ |
| 227 | bool throttle_conflicting(ThrottleConfig *cfg) |
| 228 | { |
| 229 | bool bps_flag, ops_flag; |
| 230 | bool bps_max_flag, ops_max_flag; |
| 231 | |
| 232 | bps_flag = cfg->buckets[THROTTLE_BPS_TOTAL].avg && |
| 233 | (cfg->buckets[THROTTLE_BPS_READ].avg || |
| 234 | cfg->buckets[THROTTLE_BPS_WRITE].avg); |
| 235 | |
| 236 | ops_flag = cfg->buckets[THROTTLE_OPS_TOTAL].avg && |
| 237 | (cfg->buckets[THROTTLE_OPS_READ].avg || |
| 238 | cfg->buckets[THROTTLE_OPS_WRITE].avg); |
| 239 | |
| 240 | bps_max_flag = cfg->buckets[THROTTLE_BPS_TOTAL].max && |
| 241 | (cfg->buckets[THROTTLE_BPS_READ].max || |
| 242 | cfg->buckets[THROTTLE_BPS_WRITE].max); |
| 243 | |
| 244 | ops_max_flag = cfg->buckets[THROTTLE_OPS_TOTAL].max && |
| 245 | (cfg->buckets[THROTTLE_OPS_READ].max || |
| 246 | cfg->buckets[THROTTLE_OPS_WRITE].max); |
| 247 | |
| 248 | return bps_flag || ops_flag || bps_max_flag || ops_max_flag; |
| 249 | } |
| 250 | |
| 251 | /* check if a throttling configuration is valid |
| 252 | * @cfg: the throttling configuration to inspect |
| 253 | * @ret: true if valid else false |
| 254 | */ |
| 255 | bool throttle_is_valid(ThrottleConfig *cfg) |
| 256 | { |
| 257 | bool invalid = false; |
| 258 | int i; |
| 259 | |
| 260 | for (i = 0; i < BUCKETS_COUNT; i++) { |
| 261 | if (cfg->buckets[i].avg < 0) { |
| 262 | invalid = true; |
| 263 | } |
| 264 | } |
| 265 | |
| 266 | for (i = 0; i < BUCKETS_COUNT; i++) { |
| 267 | if (cfg->buckets[i].max < 0) { |
| 268 | invalid = true; |
| 269 | } |
| 270 | } |
| 271 | |
| 272 | return !invalid; |
| 273 | } |
| 274 | |
| 275 | /* fix bucket parameters */ |
| 276 | static void throttle_fix_bucket(LeakyBucket *bkt) |
| 277 | { |
| 278 | double min; |
| 279 | |
| 280 | /* zero bucket level */ |
| 281 | bkt->level = 0; |
| 282 | |
| 283 | /* The following is done to cope with the Linux CFQ block scheduler |
| 284 | * which regroup reads and writes by block of 100ms in the guest. |
| 285 | * When they are two process one making reads and one making writes cfq |
| 286 | * make a pattern looking like the following: |
| 287 | * WWWWWWWWWWWRRRRRRRRRRRRRRWWWWWWWWWWWWWwRRRRRRRRRRRRRRRRR |
| 288 | * Having a max burst value of 100ms of the average will help smooth the |
| 289 | * throttling |
| 290 | */ |
| 291 | min = bkt->avg / 10; |
| 292 | if (bkt->avg && !bkt->max) { |
| 293 | bkt->max = min; |
| 294 | } |
| 295 | } |
| 296 | |
| 297 | /* take care of canceling a timer */ |
| 298 | static void throttle_cancel_timer(QEMUTimer *timer) |
| 299 | { |
| 300 | assert(timer != NULL); |
| 301 | |
| 302 | timer_del(timer); |
| 303 | } |
| 304 | |
| 305 | /* Used to configure the throttle |
| 306 | * |
| 307 | * @ts: the throttle state we are working on |
| 308 | * @cfg: the config to set |
| 309 | */ |
| 310 | void throttle_config(ThrottleState *ts, ThrottleConfig *cfg) |
| 311 | { |
| 312 | int i; |
| 313 | |
| 314 | ts->cfg = *cfg; |
| 315 | |
| 316 | for (i = 0; i < BUCKETS_COUNT; i++) { |
| 317 | throttle_fix_bucket(&ts->cfg.buckets[i]); |
| 318 | } |
| 319 | |
| 320 | ts->previous_leak = qemu_clock_get_ns(ts->clock_type); |
| 321 | |
| 322 | for (i = 0; i < 2; i++) { |
| 323 | throttle_cancel_timer(ts->timers[i]); |
| 324 | } |
| 325 | } |
| 326 | |
| 327 | /* used to get config |
| 328 | * |
| 329 | * @ts: the throttle state we are working on |
| 330 | * @cfg: the config to write |
| 331 | */ |
| 332 | void throttle_get_config(ThrottleState *ts, ThrottleConfig *cfg) |
| 333 | { |
| 334 | *cfg = ts->cfg; |
| 335 | } |
| 336 | |
| 337 | |
| 338 | /* Schedule the read or write timer if needed |
| 339 | * |
| 340 | * NOTE: this function is not unit tested due to it's usage of timer_mod |
| 341 | * |
| 342 | * @is_write: the type of operation (read/write) |
| 343 | * @ret: true if the timer has been scheduled else false |
| 344 | */ |
| 345 | bool throttle_schedule_timer(ThrottleState *ts, bool is_write) |
| 346 | { |
| 347 | int64_t now = qemu_clock_get_ns(ts->clock_type); |
| 348 | int64_t next_timestamp; |
| 349 | bool must_wait; |
| 350 | |
| 351 | must_wait = throttle_compute_timer(ts, |
| 352 | is_write, |
| 353 | now, |
| 354 | &next_timestamp); |
| 355 | |
| 356 | /* request not throttled */ |
| 357 | if (!must_wait) { |
| 358 | return false; |
| 359 | } |
| 360 | |
| 361 | /* request throttled and timer pending -> do nothing */ |
| 362 | if (timer_pending(ts->timers[is_write])) { |
| 363 | return true; |
| 364 | } |
| 365 | |
| 366 | /* request throttled and timer not pending -> arm timer */ |
| 367 | timer_mod(ts->timers[is_write], next_timestamp); |
| 368 | return true; |
| 369 | } |
| 370 | |
| 371 | /* do the accounting for this operation |
| 372 | * |
| 373 | * @is_write: the type of operation (read/write) |
| 374 | * @size: the size of the operation |
| 375 | */ |
| 376 | void throttle_account(ThrottleState *ts, bool is_write, uint64_t size) |
| 377 | { |
| 378 | double units = 1.0; |
| 379 | |
| 380 | /* if cfg.op_size is defined and smaller than size we compute unit count */ |
| 381 | if (ts->cfg.op_size && size > ts->cfg.op_size) { |
| 382 | units = (double) size / ts->cfg.op_size; |
| 383 | } |
| 384 | |
| 385 | ts->cfg.buckets[THROTTLE_BPS_TOTAL].level += size; |
| 386 | ts->cfg.buckets[THROTTLE_OPS_TOTAL].level += units; |
| 387 | |
| 388 | if (is_write) { |
| 389 | ts->cfg.buckets[THROTTLE_BPS_WRITE].level += size; |
| 390 | ts->cfg.buckets[THROTTLE_OPS_WRITE].level += units; |
| 391 | } else { |
| 392 | ts->cfg.buckets[THROTTLE_BPS_READ].level += size; |
| 393 | ts->cfg.buckets[THROTTLE_OPS_READ].level += units; |
| 394 | } |
| 395 | } |
| 396 | |