util/throttle.c

   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
  46 /* Calculate the time delta since last leak and make proportionals leaks
  47  *
  48  * @now:      the current timestamp in ns
  49  */
  50 static void throttle_do_leak(ThrottleState *ts, int64_t now)
  51 {
  52     /* compute the time elapsed since the last leak */
  53     int64_t delta_ns = now - ts->previous_leak;
  54     int i;
  55
  56     ts->previous_leak = now;
  57
  58     if (delta_ns <= 0) {
  59         return;
  60     }
  61
  62     /* make each bucket leak */
  63     for (i = 0; i < BUCKETS_COUNT; i++) {
  64         throttle_leak_bucket(&ts->cfg.buckets[i], delta_ns);
  65     }
  66 }
  67
  68 /* do the real job of computing the time to wait
  69  *
  70  * @limit: the throttling limit
  71  * @extra: the number of operation to delay
  72  * @ret:   the time to wait in ns
  73  */
  74 static int64_t throttle_do_compute_wait(double limit, double extra)
  75 {
  76     double wait = extra * NANOSECONDS_PER_SECOND;
  77     wait /= limit;
  78     return wait;
  79 }
  80
  81 /* This function compute the wait time in ns that a leaky bucket should trigger
  82  *
  83  * @bkt: the leaky bucket we operate on
  84  * @ret: the resulting wait time in ns or 0 if the operation can go through
  85  */
  86 int64_t throttle_compute_wait(LeakyBucket *bkt)
  87 {
  88     double extra; /* the number of extra units blocking the io */
  89
  90     if (!bkt->avg) {
  91         return 0;
  92     }
  93
  94     extra = bkt->level - bkt->max;
  95
  96     if (extra <= 0) {
  97         return 0;
  98     }
  99
 100     return throttle_do_compute_wait(bkt->avg, extra);
 101 }
 102
 103 /* This function compute the time that must be waited while this IO
 104  *
 105  * @is_write:   true if the current IO is a write, false if it's a read
 106  * @ret:        time to wait
 107  */
 108 static int64_t throttle_compute_wait_for(ThrottleState *ts,
 109                                          bool is_write)
 110 {
 111     BucketType to_check[2][4] = { {THROTTLE_BPS_TOTAL,
 112                                    THROTTLE_OPS_TOTAL,
 113                                    THROTTLE_BPS_READ,
 114                                    THROTTLE_OPS_READ},
 115                                   {THROTTLE_BPS_TOTAL,
 116                                    THROTTLE_OPS_TOTAL,
 117                                    THROTTLE_BPS_WRITE,
 118                                    THROTTLE_OPS_WRITE}, };
 119     int64_t wait, max_wait = 0;
 120     int i;
 121
 122     for (i = 0; i < 4; i++) {
 123         BucketType index = to_check[is_write][i];
 124         wait = throttle_compute_wait(&ts->cfg.buckets[index]);
 125         if (wait > max_wait) {
 126             max_wait = wait;
 127         }
 128     }
 129
 130     return max_wait;
 131 }
 132
 133 /* compute the timer for this type of operation
 134  *
 135  * @is_write:   the type of operation
 136  * @now:        the current clock timestamp
 137  * @next_timestamp: the resulting timer
 138  * @ret:        true if a timer must be set
 139  */
 140 static bool throttle_compute_timer(ThrottleState *ts,
 141                                    bool is_write,
 142                                    int64_t now,
 143                                    int64_t *next_timestamp)
 144 {
 145     int64_t wait;
 146
 147     /* leak proportionally to the time elapsed */
 148     throttle_do_leak(ts, now);
 149
 150     /* compute the wait time if any */
 151     wait = throttle_compute_wait_for(ts, is_write);
 152
 153     /* if the code must wait compute when the next timer should fire */
 154     if (wait) {
 155         *next_timestamp = now + wait;
 156         return true;
 157     }
 158
 159     /* else no need to wait at all */
 160     *next_timestamp = now;
 161     return false;
 162 }
 163
 164 /* Add timers to event loop */
 165 void throttle_timers_attach_aio_context(ThrottleTimers *tt,
 166                                         AioContext *new_context)
 167 {
 168     tt->timers[0] = aio_timer_new(new_context, tt->clock_type, SCALE_NS,
 169                                   tt->read_timer_cb, tt->timer_opaque);
 170     tt->timers[1] = aio_timer_new(new_context, tt->clock_type, SCALE_NS,
 171                                   tt->write_timer_cb, tt->timer_opaque);
 172 }
 173
 174 /*
 175  * Initialize the ThrottleConfig structure to a valid state
 176  * @cfg: the config to initialize
 177  */
 178 void throttle_config_init(ThrottleConfig *cfg)
 179 {
 180     memset(cfg, 0, sizeof(*cfg));
 181 }
 182
 183 /* To be called first on the ThrottleState */
 184 void throttle_init(ThrottleState *ts)
 185 {
 186     memset(ts, 0, sizeof(ThrottleState));
 187     throttle_config_init(&ts->cfg);
 188 }
 189
 190 /* To be called first on the ThrottleTimers */
 191 void throttle_timers_init(ThrottleTimers *tt,
 192                           AioContext *aio_context,
 193                           QEMUClockType clock_type,
 194                           QEMUTimerCB *read_timer_cb,
 195                           QEMUTimerCB *write_timer_cb,
 196                           void *timer_opaque)
 197 {
 198     memset(tt, 0, sizeof(ThrottleTimers));
 199
 200     tt->clock_type = clock_type;
 201     tt->read_timer_cb = read_timer_cb;
 202     tt->write_timer_cb = write_timer_cb;
 203     tt->timer_opaque = timer_opaque;
 204     throttle_timers_attach_aio_context(tt, aio_context);
 205 }
 206
 207 /* destroy a timer */
 208 static void throttle_timer_destroy(QEMUTimer **timer)
 209 {
 210     assert(*timer != NULL);
 211
 212     timer_del(*timer);
 213     timer_free(*timer);
 214     *timer = NULL;
 215 }
 216
 217 /* Remove timers from event loop */
 218 void throttle_timers_detach_aio_context(ThrottleTimers *tt)
 219 {
 220     int i;
 221
 222     for (i = 0; i < 2; i++) {
 223         throttle_timer_destroy(&tt->timers[i]);
 224     }
 225 }
 226
 227 /* To be called last on the ThrottleTimers */
 228 void throttle_timers_destroy(ThrottleTimers *tt)
 229 {
 230     throttle_timers_detach_aio_context(tt);
 231 }
 232
 233 /* is any throttling timer configured */
 234 bool throttle_timers_are_initialized(ThrottleTimers *tt)
 235 {
 236     if (tt->timers[0]) {
 237         return true;
 238     }
 239
 240     return false;
 241 }
 242
 243 /* Does any throttling must be done
 244  *
 245  * @cfg: the throttling configuration to inspect
 246  * @ret: true if throttling must be done else false
 247  */
 248 bool throttle_enabled(ThrottleConfig *cfg)
 249 {
 250     int i;
 251
 252     for (i = 0; i < BUCKETS_COUNT; i++) {
 253         if (cfg->buckets[i].avg > 0) {
 254             return true;
 255         }
 256     }
 257
 258     return false;
 259 }
 260
 261 /* check if a throttling configuration is valid
 262  * @cfg: the throttling configuration to inspect
 263  * @ret: true if valid else false
 264  * @errp: error object
 265  */
 266 bool throttle_is_valid(ThrottleConfig *cfg, Error **errp)
 267 {
 268     int i;
 269     bool bps_flag, ops_flag;
 270     bool bps_max_flag, ops_max_flag;
 271
 272     bps_flag = cfg->buckets[THROTTLE_BPS_TOTAL].avg &&
 273                (cfg->buckets[THROTTLE_BPS_READ].avg ||
 274                 cfg->buckets[THROTTLE_BPS_WRITE].avg);
 275
 276     ops_flag = cfg->buckets[THROTTLE_OPS_TOTAL].avg &&
 277                (cfg->buckets[THROTTLE_OPS_READ].avg ||
 278                 cfg->buckets[THROTTLE_OPS_WRITE].avg);
 279
 280     bps_max_flag = cfg->buckets[THROTTLE_BPS_TOTAL].max &&
 281                   (cfg->buckets[THROTTLE_BPS_READ].max  ||
 282                    cfg->buckets[THROTTLE_BPS_WRITE].max);
 283
 284     ops_max_flag = cfg->buckets[THROTTLE_OPS_TOTAL].max &&
 285                    (cfg->buckets[THROTTLE_OPS_READ].max ||
 286                    cfg->buckets[THROTTLE_OPS_WRITE].max);
 287
 288     if (bps_flag || ops_flag || bps_max_flag || ops_max_flag) {
 289         error_setg(errp, "bps/iops/max total values and read/write values"
 290                    " cannot be used at the same time");
 291         return false;
 292     }
 293
 294     for (i = 0; i < BUCKETS_COUNT; i++) {
 295         if (cfg->buckets[i].avg < 0 ||
 296             cfg->buckets[i].max < 0 ||
 297             cfg->buckets[i].avg > THROTTLE_VALUE_MAX ||
 298             cfg->buckets[i].max > THROTTLE_VALUE_MAX) {
 299             error_setg(errp, "bps/iops/max values must be within [0, %lld]",
 300                        THROTTLE_VALUE_MAX);
 301             return false;
 302         }
 303
 304         if (cfg->buckets[i].max && !cfg->buckets[i].avg) {
 305             error_setg(errp, "bps_max/iops_max require corresponding"
 306                        " bps/iops values");
 307             return false;
 308         }
 309     }
 310
 311     return true;
 312 }
 313
 314 /* fix bucket parameters */
 315 static void throttle_fix_bucket(LeakyBucket *bkt)
 316 {
 317     double min;
 318
 319     /* zero bucket level */
 320     bkt->level = 0;
 321
 322     /* The following is done to cope with the Linux CFQ block scheduler
 323      * which regroup reads and writes by block of 100ms in the guest.
 324      * When they are two process one making reads and one making writes cfq
 325      * make a pattern looking like the following:
 326      * WWWWWWWWWWWRRRRRRRRRRRRRRWWWWWWWWWWWWWwRRRRRRRRRRRRRRRRR
 327      * Having a max burst value of 100ms of the average will help smooth the
 328      * throttling
 329      */
 330     min = bkt->avg / 10;
 331     if (bkt->avg && !bkt->max) {
 332         bkt->max = min;
 333     }
 334 }
 335
 336 /* take care of canceling a timer */
 337 static void throttle_cancel_timer(QEMUTimer *timer)
 338 {
 339     assert(timer != NULL);
 340
 341     timer_del(timer);
 342 }
 343
 344 /* Used to configure the throttle
 345  *
 346  * @ts: the throttle state we are working on
 347  * @tt: the throttle timers we use in this aio context
 348  * @cfg: the config to set
 349  */
 350 void throttle_config(ThrottleState *ts,
 351                      ThrottleTimers *tt,
 352                      ThrottleConfig *cfg)
 353 {
 354     int i;
 355
 356     ts->cfg = *cfg;
 357
 358     for (i = 0; i < BUCKETS_COUNT; i++) {
 359         throttle_fix_bucket(&ts->cfg.buckets[i]);
 360     }
 361
 362     ts->previous_leak = qemu_clock_get_ns(tt->clock_type);
 363
 364     for (i = 0; i < 2; i++) {
 365         throttle_cancel_timer(tt->timers[i]);
 366     }
 367 }
 368
 369 /* used to get config
 370  *
 371  * @ts:  the throttle state we are working on
 372  * @cfg: the config to write
 373  */
 374 void throttle_get_config(ThrottleState *ts, ThrottleConfig *cfg)
 375 {
 376     *cfg = ts->cfg;
 377 }
 378
 379
 380 /* Schedule the read or write timer if needed
 381  *
 382  * NOTE: this function is not unit tested due to it's usage of timer_mod
 383  *
 384  * @tt:       the timers structure
 385  * @is_write: the type of operation (read/write)
 386  * @ret:      true if the timer has been scheduled else false
 387  */
 388 bool throttle_schedule_timer(ThrottleState *ts,
 389                              ThrottleTimers *tt,
 390                              bool is_write)
 391 {
 392     int64_t now = qemu_clock_get_ns(tt->clock_type);
 393     int64_t next_timestamp;
 394     bool must_wait;
 395
 396     must_wait = throttle_compute_timer(ts,
 397                                        is_write,
 398                                        now,
 399                                        &next_timestamp);
 400
 401     /* request not throttled */
 402     if (!must_wait) {
 403         return false;
 404     }
 405
 406     /* request throttled and timer pending -> do nothing */
 407     if (timer_pending(tt->timers[is_write])) {
 408         return true;
 409     }
 410
 411     /* request throttled and timer not pending -> arm timer */
 412     timer_mod(tt->timers[is_write], next_timestamp);
 413     return true;
 414 }
 415
 416 /* do the accounting for this operation
 417  *
 418  * @is_write: the type of operation (read/write)
 419  * @size:     the size of the operation
 420  */
 421 void throttle_account(ThrottleState *ts, bool is_write, uint64_t size)
 422 {
 423     double units = 1.0;
 424
 425     /* if cfg.op_size is defined and smaller than size we compute unit count */
 426     if (ts->cfg.op_size && size > ts->cfg.op_size) {
 427         units = (double) size / ts->cfg.op_size;
 428     }
 429
 430     ts->cfg.buckets[THROTTLE_BPS_TOTAL].level += size;
 431     ts->cfg.buckets[THROTTLE_OPS_TOTAL].level += units;
 432
 433     if (is_write) {
 434         ts->cfg.buckets[THROTTLE_BPS_WRITE].level += size;
 435         ts->cfg.buckets[THROTTLE_OPS_WRITE].level += units;
 436     } else {
 437         ts->cfg.buckets[THROTTLE_BPS_READ].level += size;
 438         ts->cfg.buckets[THROTTLE_OPS_READ].level += units;
 439     }
 440 }
 441