2 * QEMU throttling infrastructure
4 * Copyright (C) Nodalink, EURL. 2013-2014
5 * Copyright (C) Igalia, S.L. 2015
8 * BenoƮt Canet <benoit.canet@nodalink.com>
9 * Alberto Garcia <berto@igalia.com>
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.
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.
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/>.
25 #include "qemu/throttle.h"
26 #include "qemu/timer.h"
27 #include "block/aio.h"
29 /* This function make a bucket leak
31 * @bkt: the bucket to make leak
32 * @delta_ns: the time delta
34 void throttle_leak_bucket(LeakyBucket *bkt, int64_t delta_ns)
38 /* compute how much to leak */
39 leak = (bkt->avg * (double) delta_ns) / NANOSECONDS_PER_SECOND;
41 /* make the bucket leak */
42 bkt->level = MAX(bkt->level - leak, 0);
45 /* Calculate the time delta since last leak and make proportionals leaks
47 * @now: the current timestamp in ns
49 static void throttle_do_leak(ThrottleState *ts, int64_t now)
51 /* compute the time elapsed since the last leak */
52 int64_t delta_ns = now - ts->previous_leak;
55 ts->previous_leak = now;
61 /* make each bucket leak */
62 for (i = 0; i < BUCKETS_COUNT; i++) {
63 throttle_leak_bucket(&ts->cfg.buckets[i], delta_ns);
67 /* do the real job of computing the time to wait
69 * @limit: the throttling limit
70 * @extra: the number of operation to delay
71 * @ret: the time to wait in ns
73 static int64_t throttle_do_compute_wait(double limit, double extra)
75 double wait = extra * NANOSECONDS_PER_SECOND;
80 /* This function compute the wait time in ns that a leaky bucket should trigger
82 * @bkt: the leaky bucket we operate on
83 * @ret: the resulting wait time in ns or 0 if the operation can go through
85 int64_t throttle_compute_wait(LeakyBucket *bkt)
87 double extra; /* the number of extra units blocking the io */
93 extra = bkt->level - bkt->max;
99 return throttle_do_compute_wait(bkt->avg, extra);
102 /* This function compute the time that must be waited while this IO
104 * @is_write: true if the current IO is a write, false if it's a read
107 static int64_t throttle_compute_wait_for(ThrottleState *ts,
110 BucketType to_check[2][4] = { {THROTTLE_BPS_TOTAL,
117 THROTTLE_OPS_WRITE}, };
118 int64_t wait, max_wait = 0;
121 for (i = 0; i < 4; i++) {
122 BucketType index = to_check[is_write][i];
123 wait = throttle_compute_wait(&ts->cfg.buckets[index]);
124 if (wait > max_wait) {
132 /* compute the timer for this type of operation
134 * @is_write: the type of operation
135 * @now: the current clock timestamp
136 * @next_timestamp: the resulting timer
137 * @ret: true if a timer must be set
139 bool throttle_compute_timer(ThrottleState *ts,
142 int64_t *next_timestamp)
146 /* leak proportionally to the time elapsed */
147 throttle_do_leak(ts, now);
149 /* compute the wait time if any */
150 wait = throttle_compute_wait_for(ts, is_write);
152 /* if the code must wait compute when the next timer should fire */
154 *next_timestamp = now + wait;
158 /* else no need to wait at all */
159 *next_timestamp = now;
163 /* Add timers to event loop */
164 void throttle_timers_attach_aio_context(ThrottleTimers *tt,
165 AioContext *new_context)
167 tt->timers[0] = aio_timer_new(new_context, tt->clock_type, SCALE_NS,
168 tt->read_timer_cb, tt->timer_opaque);
169 tt->timers[1] = aio_timer_new(new_context, tt->clock_type, SCALE_NS,
170 tt->write_timer_cb, tt->timer_opaque);
173 /* To be called first on the ThrottleState */
174 void throttle_init(ThrottleState *ts)
176 memset(ts, 0, sizeof(ThrottleState));
179 /* To be called first on the ThrottleTimers */
180 void throttle_timers_init(ThrottleTimers *tt,
181 AioContext *aio_context,
182 QEMUClockType clock_type,
183 QEMUTimerCB *read_timer_cb,
184 QEMUTimerCB *write_timer_cb,
187 memset(tt, 0, sizeof(ThrottleTimers));
189 tt->clock_type = clock_type;
190 tt->read_timer_cb = read_timer_cb;
191 tt->write_timer_cb = write_timer_cb;
192 tt->timer_opaque = timer_opaque;
193 throttle_timers_attach_aio_context(tt, aio_context);
196 /* destroy a timer */
197 static void throttle_timer_destroy(QEMUTimer **timer)
199 assert(*timer != NULL);
206 /* Remove timers from event loop */
207 void throttle_timers_detach_aio_context(ThrottleTimers *tt)
211 for (i = 0; i < 2; i++) {
212 throttle_timer_destroy(&tt->timers[i]);
216 /* To be called last on the ThrottleTimers */
217 void throttle_timers_destroy(ThrottleTimers *tt)
219 throttle_timers_detach_aio_context(tt);
222 /* is any throttling timer configured */
223 bool throttle_timers_are_initialized(ThrottleTimers *tt)
232 /* Does any throttling must be done
234 * @cfg: the throttling configuration to inspect
235 * @ret: true if throttling must be done else false
237 bool throttle_enabled(ThrottleConfig *cfg)
241 for (i = 0; i < BUCKETS_COUNT; i++) {
242 if (cfg->buckets[i].avg > 0) {
250 /* return true if any two throttling parameters conflicts
252 * @cfg: the throttling configuration to inspect
253 * @ret: true if any conflict detected else false
255 bool throttle_conflicting(ThrottleConfig *cfg)
257 bool bps_flag, ops_flag;
258 bool bps_max_flag, ops_max_flag;
260 bps_flag = cfg->buckets[THROTTLE_BPS_TOTAL].avg &&
261 (cfg->buckets[THROTTLE_BPS_READ].avg ||
262 cfg->buckets[THROTTLE_BPS_WRITE].avg);
264 ops_flag = cfg->buckets[THROTTLE_OPS_TOTAL].avg &&
265 (cfg->buckets[THROTTLE_OPS_READ].avg ||
266 cfg->buckets[THROTTLE_OPS_WRITE].avg);
268 bps_max_flag = cfg->buckets[THROTTLE_BPS_TOTAL].max &&
269 (cfg->buckets[THROTTLE_BPS_READ].max ||
270 cfg->buckets[THROTTLE_BPS_WRITE].max);
272 ops_max_flag = cfg->buckets[THROTTLE_OPS_TOTAL].max &&
273 (cfg->buckets[THROTTLE_OPS_READ].max ||
274 cfg->buckets[THROTTLE_OPS_WRITE].max);
276 return bps_flag || ops_flag || bps_max_flag || ops_max_flag;
279 /* check if a throttling configuration is valid
280 * @cfg: the throttling configuration to inspect
281 * @ret: true if valid else false
283 bool throttle_is_valid(ThrottleConfig *cfg)
285 bool invalid = false;
288 for (i = 0; i < BUCKETS_COUNT; i++) {
289 if (cfg->buckets[i].avg < 0) {
294 for (i = 0; i < BUCKETS_COUNT; i++) {
295 if (cfg->buckets[i].max < 0) {
303 /* fix bucket parameters */
304 static void throttle_fix_bucket(LeakyBucket *bkt)
308 /* zero bucket level */
311 /* The following is done to cope with the Linux CFQ block scheduler
312 * which regroup reads and writes by block of 100ms in the guest.
313 * When they are two process one making reads and one making writes cfq
314 * make a pattern looking like the following:
315 * WWWWWWWWWWWRRRRRRRRRRRRRRWWWWWWWWWWWWWwRRRRRRRRRRRRRRRRR
316 * Having a max burst value of 100ms of the average will help smooth the
320 if (bkt->avg && !bkt->max) {
325 /* take care of canceling a timer */
326 static void throttle_cancel_timer(QEMUTimer *timer)
328 assert(timer != NULL);
333 /* Used to configure the throttle
335 * @ts: the throttle state we are working on
336 * @tt: the throttle timers we use in this aio context
337 * @cfg: the config to set
339 void throttle_config(ThrottleState *ts,
347 for (i = 0; i < BUCKETS_COUNT; i++) {
348 throttle_fix_bucket(&ts->cfg.buckets[i]);
351 ts->previous_leak = qemu_clock_get_ns(tt->clock_type);
353 for (i = 0; i < 2; i++) {
354 throttle_cancel_timer(tt->timers[i]);
358 /* used to get config
360 * @ts: the throttle state we are working on
361 * @cfg: the config to write
363 void throttle_get_config(ThrottleState *ts, ThrottleConfig *cfg)
369 /* Schedule the read or write timer if needed
371 * NOTE: this function is not unit tested due to it's usage of timer_mod
373 * @tt: the timers structure
374 * @is_write: the type of operation (read/write)
375 * @ret: true if the timer has been scheduled else false
377 bool throttle_schedule_timer(ThrottleState *ts,
381 int64_t now = qemu_clock_get_ns(tt->clock_type);
382 int64_t next_timestamp;
385 must_wait = throttle_compute_timer(ts,
390 /* request not throttled */
395 /* request throttled and timer pending -> do nothing */
396 if (timer_pending(tt->timers[is_write])) {
400 /* request throttled and timer not pending -> arm timer */
401 timer_mod(tt->timers[is_write], next_timestamp);
405 /* do the accounting for this operation
407 * @is_write: the type of operation (read/write)
408 * @size: the size of the operation
410 void throttle_account(ThrottleState *ts, bool is_write, uint64_t size)
414 /* if cfg.op_size is defined and smaller than size we compute unit count */
415 if (ts->cfg.op_size && size > ts->cfg.op_size) {
416 units = (double) size / ts->cfg.op_size;
419 ts->cfg.buckets[THROTTLE_BPS_TOTAL].level += size;
420 ts->cfg.buckets[THROTTLE_OPS_TOTAL].level += units;
423 ts->cfg.buckets[THROTTLE_BPS_WRITE].level += size;
424 ts->cfg.buckets[THROTTLE_OPS_WRITE].level += units;
426 ts->cfg.buckets[THROTTLE_BPS_READ].level += size;
427 ts->cfg.buckets[THROTTLE_OPS_READ].level += units;