xref: /openbmc/qemu/util/throttle.c (revision 05caa062)
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     double bucket_size;   /* I/O before throttling to bkt->avg */
99     double burst_bucket_size; /* Before throttling to bkt->max */
100 
101     if (!bkt->avg) {
102         return 0;
103     }
104 
105     if (!bkt->max) {
106         /* If bkt->max is 0 we still want to allow short bursts of I/O
107          * from the guest, otherwise every other request will be throttled
108          * and performance will suffer considerably. */
109         bucket_size = (double) bkt->avg / 10;
110         burst_bucket_size = 0;
111     } else {
112         /* If we have a burst limit then we have to wait until all I/O
113          * at burst rate has finished before throttling to bkt->avg */
114         bucket_size = bkt->max * bkt->burst_length;
115         burst_bucket_size = (double) bkt->max / 10;
116     }
117 
118     /* If the main bucket is full then we have to wait */
119     extra = bkt->level - bucket_size;
120     if (extra > 0) {
121         return throttle_do_compute_wait(bkt->avg, extra);
122     }
123 
124     /* If the main bucket is not full yet we still have to check the
125      * burst bucket in order to enforce the burst limit */
126     if (bkt->burst_length > 1) {
127         assert(bkt->max > 0); /* see throttle_is_valid() */
128         extra = bkt->burst_level - burst_bucket_size;
129         if (extra > 0) {
130             return throttle_do_compute_wait(bkt->max, extra);
131         }
132     }
133 
134     return 0;
135 }
136 
137 /* This function compute the time that must be waited while this IO
138  *
139  * @direction:  throttle direction
140  * @ret:        time to wait
141  */
142 static int64_t throttle_compute_wait_for(ThrottleState *ts,
143                                          ThrottleDirection direction)
144 {
145     static const BucketType to_check[THROTTLE_MAX][4] = {
146                                   {THROTTLE_BPS_TOTAL,
147                                    THROTTLE_OPS_TOTAL,
148                                    THROTTLE_BPS_READ,
149                                    THROTTLE_OPS_READ},
150                                   {THROTTLE_BPS_TOTAL,
151                                    THROTTLE_OPS_TOTAL,
152                                    THROTTLE_BPS_WRITE,
153                                    THROTTLE_OPS_WRITE}, };
154     int64_t wait, max_wait = 0;
155     int i;
156 
157     for (i = 0; i < ARRAY_SIZE(to_check[THROTTLE_READ]); i++) {
158         BucketType index = to_check[direction][i];
159         wait = throttle_compute_wait(&ts->cfg.buckets[index]);
160         if (wait > max_wait) {
161             max_wait = wait;
162         }
163     }
164 
165     return max_wait;
166 }
167 
168 /* compute the timer for this type of operation
169  *
170  * @direction:  throttle direction
171  * @now:        the current clock timestamp
172  * @next_timestamp: the resulting timer
173  * @ret:        true if a timer must be set
174  */
175 static bool throttle_compute_timer(ThrottleState *ts,
176                                    ThrottleDirection direction,
177                                    int64_t now,
178                                    int64_t *next_timestamp)
179 {
180     int64_t wait;
181 
182     /* leak proportionally to the time elapsed */
183     throttle_do_leak(ts, now);
184 
185     /* compute the wait time if any */
186     wait = throttle_compute_wait_for(ts, direction);
187 
188     /* if the code must wait compute when the next timer should fire */
189     if (wait) {
190         *next_timestamp = now + wait;
191         return true;
192     }
193 
194     /* else no need to wait at all */
195     *next_timestamp = now;
196     return false;
197 }
198 
199 /* Add timers to event loop */
200 void throttle_timers_attach_aio_context(ThrottleTimers *tt,
201                                         AioContext *new_context)
202 {
203     ThrottleDirection dir;
204 
205     for (dir = THROTTLE_READ; dir < THROTTLE_MAX; dir++) {
206         if (tt->timer_cb[dir]) {
207             tt->timers[dir] =
208                 aio_timer_new(new_context, tt->clock_type, SCALE_NS,
209                               tt->timer_cb[dir], tt->timer_opaque);
210         }
211     }
212 }
213 
214 /*
215  * Initialize the ThrottleConfig structure to a valid state
216  * @cfg: the config to initialize
217  */
218 void throttle_config_init(ThrottleConfig *cfg)
219 {
220     unsigned i;
221     memset(cfg, 0, sizeof(*cfg));
222     for (i = 0; i < BUCKETS_COUNT; i++) {
223         cfg->buckets[i].burst_length = 1;
224     }
225 }
226 
227 /* To be called first on the ThrottleState */
228 void throttle_init(ThrottleState *ts)
229 {
230     memset(ts, 0, sizeof(ThrottleState));
231     throttle_config_init(&ts->cfg);
232 }
233 
234 /* To be called first on the ThrottleTimers */
235 void throttle_timers_init(ThrottleTimers *tt,
236                           AioContext *aio_context,
237                           QEMUClockType clock_type,
238                           QEMUTimerCB *read_timer_cb,
239                           QEMUTimerCB *write_timer_cb,
240                           void *timer_opaque)
241 {
242     assert(read_timer_cb || write_timer_cb);
243     memset(tt, 0, sizeof(ThrottleTimers));
244 
245     tt->clock_type = clock_type;
246     tt->timer_cb[THROTTLE_READ] = read_timer_cb;
247     tt->timer_cb[THROTTLE_WRITE] = write_timer_cb;
248     tt->timer_opaque = timer_opaque;
249     throttle_timers_attach_aio_context(tt, aio_context);
250 }
251 
252 /* destroy a timer */
253 static void throttle_timer_destroy(QEMUTimer **timer)
254 {
255     if (*timer == NULL) {
256         return;
257     }
258 
259     timer_free(*timer);
260     *timer = NULL;
261 }
262 
263 /* Remove timers from event loop */
264 void throttle_timers_detach_aio_context(ThrottleTimers *tt)
265 {
266     ThrottleDirection dir;
267 
268     for (dir = THROTTLE_READ; dir < THROTTLE_MAX; dir++) {
269         throttle_timer_destroy(&tt->timers[dir]);
270     }
271 }
272 
273 /* To be called last on the ThrottleTimers */
274 void throttle_timers_destroy(ThrottleTimers *tt)
275 {
276     throttle_timers_detach_aio_context(tt);
277 }
278 
279 /* is any throttling timer configured */
280 bool throttle_timers_are_initialized(ThrottleTimers *tt)
281 {
282     ThrottleDirection dir;
283 
284     for (dir = THROTTLE_READ; dir < THROTTLE_MAX; dir++) {
285         if (tt->timers[dir]) {
286             return true;
287         }
288     }
289 
290     return false;
291 }
292 
293 /* Does any throttling must be done
294  *
295  * @cfg: the throttling configuration to inspect
296  * @ret: true if throttling must be done else false
297  */
298 bool throttle_enabled(ThrottleConfig *cfg)
299 {
300     int i;
301 
302     for (i = 0; i < BUCKETS_COUNT; i++) {
303         if (cfg->buckets[i].avg > 0) {
304             return true;
305         }
306     }
307 
308     return false;
309 }
310 
311 /* check if a throttling configuration is valid
312  * @cfg: the throttling configuration to inspect
313  * @ret: true if valid else false
314  * @errp: error object
315  */
316 bool throttle_is_valid(ThrottleConfig *cfg, Error **errp)
317 {
318     int i;
319     bool bps_flag, ops_flag;
320     bool bps_max_flag, ops_max_flag;
321 
322     bps_flag = cfg->buckets[THROTTLE_BPS_TOTAL].avg &&
323                (cfg->buckets[THROTTLE_BPS_READ].avg ||
324                 cfg->buckets[THROTTLE_BPS_WRITE].avg);
325 
326     ops_flag = cfg->buckets[THROTTLE_OPS_TOTAL].avg &&
327                (cfg->buckets[THROTTLE_OPS_READ].avg ||
328                 cfg->buckets[THROTTLE_OPS_WRITE].avg);
329 
330     bps_max_flag = cfg->buckets[THROTTLE_BPS_TOTAL].max &&
331                   (cfg->buckets[THROTTLE_BPS_READ].max  ||
332                    cfg->buckets[THROTTLE_BPS_WRITE].max);
333 
334     ops_max_flag = cfg->buckets[THROTTLE_OPS_TOTAL].max &&
335                    (cfg->buckets[THROTTLE_OPS_READ].max ||
336                    cfg->buckets[THROTTLE_OPS_WRITE].max);
337 
338     if (bps_flag || ops_flag || bps_max_flag || ops_max_flag) {
339         error_setg(errp, "bps/iops/max total values and read/write values"
340                    " cannot be used at the same time");
341         return false;
342     }
343 
344     if (cfg->op_size &&
345         !cfg->buckets[THROTTLE_OPS_TOTAL].avg &&
346         !cfg->buckets[THROTTLE_OPS_READ].avg &&
347         !cfg->buckets[THROTTLE_OPS_WRITE].avg) {
348         error_setg(errp, "iops size requires an iops value to be set");
349         return false;
350     }
351 
352     for (i = 0; i < BUCKETS_COUNT; i++) {
353         LeakyBucket *bkt = &cfg->buckets[i];
354         if (bkt->avg > THROTTLE_VALUE_MAX || bkt->max > THROTTLE_VALUE_MAX) {
355             error_setg(errp, "bps/iops/max values must be within [0, %lld]",
356                        THROTTLE_VALUE_MAX);
357             return false;
358         }
359 
360         if (!bkt->burst_length) {
361             error_setg(errp, "the burst length cannot be 0");
362             return false;
363         }
364 
365         if (bkt->burst_length > 1 && !bkt->max) {
366             error_setg(errp, "burst length set without burst rate");
367             return false;
368         }
369 
370         if (bkt->max && bkt->burst_length > THROTTLE_VALUE_MAX / bkt->max) {
371             error_setg(errp, "burst length too high for this burst rate");
372             return false;
373         }
374 
375         if (bkt->max && !bkt->avg) {
376             error_setg(errp, "bps_max/iops_max require corresponding"
377                        " bps/iops values");
378             return false;
379         }
380 
381         if (bkt->max && bkt->max < bkt->avg) {
382             error_setg(errp, "bps_max/iops_max cannot be lower than bps/iops");
383             return false;
384         }
385     }
386 
387     return true;
388 }
389 
390 /* Used to configure the throttle
391  *
392  * @ts: the throttle state we are working on
393  * @clock_type: the group's clock_type
394  * @cfg: the config to set
395  */
396 void throttle_config(ThrottleState *ts,
397                      QEMUClockType clock_type,
398                      ThrottleConfig *cfg)
399 {
400     int i;
401 
402     ts->cfg = *cfg;
403 
404     /* Zero bucket level */
405     for (i = 0; i < BUCKETS_COUNT; i++) {
406         ts->cfg.buckets[i].level = 0;
407         ts->cfg.buckets[i].burst_level = 0;
408     }
409 
410     ts->previous_leak = qemu_clock_get_ns(clock_type);
411 }
412 
413 /* used to get config
414  *
415  * @ts:  the throttle state we are working on
416  * @cfg: the config to write
417  */
418 void throttle_get_config(ThrottleState *ts, ThrottleConfig *cfg)
419 {
420     *cfg = ts->cfg;
421 }
422 
423 
424 /* Schedule the read or write timer if needed
425  *
426  * NOTE: this function is not unit tested due to it's usage of timer_mod
427  *
428  * @tt:       the timers structure
429  * @direction: throttle direction
430  * @ret:      true if the timer has been scheduled else false
431  */
432 bool throttle_schedule_timer(ThrottleState *ts,
433                              ThrottleTimers *tt,
434                              ThrottleDirection direction)
435 {
436     int64_t now = qemu_clock_get_ns(tt->clock_type);
437     int64_t next_timestamp;
438     QEMUTimer *timer;
439     bool must_wait;
440 
441     assert(direction < THROTTLE_MAX);
442     timer = tt->timers[direction];
443     assert(timer);
444 
445     must_wait = throttle_compute_timer(ts,
446                                        direction,
447                                        now,
448                                        &next_timestamp);
449 
450     /* request not throttled */
451     if (!must_wait) {
452         return false;
453     }
454 
455     /* request throttled and timer pending -> do nothing */
456     if (timer_pending(timer)) {
457         return true;
458     }
459 
460     /* request throttled and timer not pending -> arm timer */
461     timer_mod(timer, next_timestamp);
462     return true;
463 }
464 
465 /* do the accounting for this operation
466  *
467  * @direction: throttle direction
468  * @size:     the size of the operation
469  */
470 void throttle_account(ThrottleState *ts, ThrottleDirection direction,
471                       uint64_t size)
472 {
473     static const BucketType bucket_types_size[THROTTLE_MAX][2] = {
474         { THROTTLE_BPS_TOTAL, THROTTLE_BPS_READ },
475         { THROTTLE_BPS_TOTAL, THROTTLE_BPS_WRITE }
476     };
477     static const BucketType bucket_types_units[THROTTLE_MAX][2] = {
478         { THROTTLE_OPS_TOTAL, THROTTLE_OPS_READ },
479         { THROTTLE_OPS_TOTAL, THROTTLE_OPS_WRITE }
480     };
481     double units = 1.0;
482     unsigned i;
483 
484     assert(direction < THROTTLE_MAX);
485     /* if cfg.op_size is defined and smaller than size we compute unit count */
486     if (ts->cfg.op_size && size > ts->cfg.op_size) {
487         units = (double) size / ts->cfg.op_size;
488     }
489 
490     for (i = 0; i < ARRAY_SIZE(bucket_types_size[THROTTLE_READ]); i++) {
491         LeakyBucket *bkt;
492 
493         bkt = &ts->cfg.buckets[bucket_types_size[direction][i]];
494         bkt->level += size;
495         if (bkt->burst_length > 1) {
496             bkt->burst_level += size;
497         }
498 
499         bkt = &ts->cfg.buckets[bucket_types_units[direction][i]];
500         bkt->level += units;
501         if (bkt->burst_length > 1) {
502             bkt->burst_level += units;
503         }
504     }
505 }
506 
507 /* return a ThrottleConfig based on the options in a ThrottleLimits
508  *
509  * @arg:    the ThrottleLimits object to read from
510  * @cfg:    the ThrottleConfig to edit
511  * @errp:   error object
512  */
513 void throttle_limits_to_config(ThrottleLimits *arg, ThrottleConfig *cfg,
514                                Error **errp)
515 {
516     if (arg->has_bps_total) {
517         cfg->buckets[THROTTLE_BPS_TOTAL].avg = arg->bps_total;
518     }
519     if (arg->has_bps_read) {
520         cfg->buckets[THROTTLE_BPS_READ].avg  = arg->bps_read;
521     }
522     if (arg->has_bps_write) {
523         cfg->buckets[THROTTLE_BPS_WRITE].avg = arg->bps_write;
524     }
525 
526     if (arg->has_iops_total) {
527         cfg->buckets[THROTTLE_OPS_TOTAL].avg = arg->iops_total;
528     }
529     if (arg->has_iops_read) {
530         cfg->buckets[THROTTLE_OPS_READ].avg  = arg->iops_read;
531     }
532     if (arg->has_iops_write) {
533         cfg->buckets[THROTTLE_OPS_WRITE].avg = arg->iops_write;
534     }
535 
536     if (arg->has_bps_total_max) {
537         cfg->buckets[THROTTLE_BPS_TOTAL].max = arg->bps_total_max;
538     }
539     if (arg->has_bps_read_max) {
540         cfg->buckets[THROTTLE_BPS_READ].max = arg->bps_read_max;
541     }
542     if (arg->has_bps_write_max) {
543         cfg->buckets[THROTTLE_BPS_WRITE].max = arg->bps_write_max;
544     }
545     if (arg->has_iops_total_max) {
546         cfg->buckets[THROTTLE_OPS_TOTAL].max = arg->iops_total_max;
547     }
548     if (arg->has_iops_read_max) {
549         cfg->buckets[THROTTLE_OPS_READ].max = arg->iops_read_max;
550     }
551     if (arg->has_iops_write_max) {
552         cfg->buckets[THROTTLE_OPS_WRITE].max = arg->iops_write_max;
553     }
554 
555     if (arg->has_bps_total_max_length) {
556         if (arg->bps_total_max_length > UINT_MAX) {
557             error_setg(errp, "bps-total-max-length value must be in"
558                              " the range [0, %u]", UINT_MAX);
559             return;
560         }
561         cfg->buckets[THROTTLE_BPS_TOTAL].burst_length = arg->bps_total_max_length;
562     }
563     if (arg->has_bps_read_max_length) {
564         if (arg->bps_read_max_length > UINT_MAX) {
565             error_setg(errp, "bps-read-max-length value must be in"
566                              " the range [0, %u]", UINT_MAX);
567             return;
568         }
569         cfg->buckets[THROTTLE_BPS_READ].burst_length = arg->bps_read_max_length;
570     }
571     if (arg->has_bps_write_max_length) {
572         if (arg->bps_write_max_length > UINT_MAX) {
573             error_setg(errp, "bps-write-max-length value must be in"
574                              " the range [0, %u]", UINT_MAX);
575             return;
576         }
577         cfg->buckets[THROTTLE_BPS_WRITE].burst_length = arg->bps_write_max_length;
578     }
579     if (arg->has_iops_total_max_length) {
580         if (arg->iops_total_max_length > UINT_MAX) {
581             error_setg(errp, "iops-total-max-length value must be in"
582                              " the range [0, %u]", UINT_MAX);
583             return;
584         }
585         cfg->buckets[THROTTLE_OPS_TOTAL].burst_length = arg->iops_total_max_length;
586     }
587     if (arg->has_iops_read_max_length) {
588         if (arg->iops_read_max_length > UINT_MAX) {
589             error_setg(errp, "iops-read-max-length value must be in"
590                              " the range [0, %u]", UINT_MAX);
591             return;
592         }
593         cfg->buckets[THROTTLE_OPS_READ].burst_length = arg->iops_read_max_length;
594     }
595     if (arg->has_iops_write_max_length) {
596         if (arg->iops_write_max_length > UINT_MAX) {
597             error_setg(errp, "iops-write-max-length value must be in"
598                              " the range [0, %u]", UINT_MAX);
599             return;
600         }
601         cfg->buckets[THROTTLE_OPS_WRITE].burst_length = arg->iops_write_max_length;
602     }
603 
604     if (arg->has_iops_size) {
605         cfg->op_size = arg->iops_size;
606     }
607 
608     throttle_is_valid(cfg, errp);
609 }
610 
611 /* write the options of a ThrottleConfig to a ThrottleLimits
612  *
613  * @cfg:    the ThrottleConfig to read from
614  * @var:    the ThrottleLimits to write to
615  */
616 void throttle_config_to_limits(ThrottleConfig *cfg, ThrottleLimits *var)
617 {
618     var->bps_total               = cfg->buckets[THROTTLE_BPS_TOTAL].avg;
619     var->bps_read                = cfg->buckets[THROTTLE_BPS_READ].avg;
620     var->bps_write               = cfg->buckets[THROTTLE_BPS_WRITE].avg;
621     var->iops_total              = cfg->buckets[THROTTLE_OPS_TOTAL].avg;
622     var->iops_read               = cfg->buckets[THROTTLE_OPS_READ].avg;
623     var->iops_write              = cfg->buckets[THROTTLE_OPS_WRITE].avg;
624     var->bps_total_max           = cfg->buckets[THROTTLE_BPS_TOTAL].max;
625     var->bps_read_max            = cfg->buckets[THROTTLE_BPS_READ].max;
626     var->bps_write_max           = cfg->buckets[THROTTLE_BPS_WRITE].max;
627     var->iops_total_max          = cfg->buckets[THROTTLE_OPS_TOTAL].max;
628     var->iops_read_max           = cfg->buckets[THROTTLE_OPS_READ].max;
629     var->iops_write_max          = cfg->buckets[THROTTLE_OPS_WRITE].max;
630     var->bps_total_max_length    = cfg->buckets[THROTTLE_BPS_TOTAL].burst_length;
631     var->bps_read_max_length     = cfg->buckets[THROTTLE_BPS_READ].burst_length;
632     var->bps_write_max_length    = cfg->buckets[THROTTLE_BPS_WRITE].burst_length;
633     var->iops_total_max_length   = cfg->buckets[THROTTLE_OPS_TOTAL].burst_length;
634     var->iops_read_max_length    = cfg->buckets[THROTTLE_OPS_READ].burst_length;
635     var->iops_write_max_length   = cfg->buckets[THROTTLE_OPS_WRITE].burst_length;
636     var->iops_size               = cfg->op_size;
637 
638     var->has_bps_total = true;
639     var->has_bps_read = true;
640     var->has_bps_write = true;
641     var->has_iops_total = true;
642     var->has_iops_read = true;
643     var->has_iops_write = true;
644     var->has_bps_total_max = true;
645     var->has_bps_read_max = true;
646     var->has_bps_write_max = true;
647     var->has_iops_total_max = true;
648     var->has_iops_read_max = true;
649     var->has_iops_write_max = true;
650     var->has_bps_read_max_length = true;
651     var->has_bps_total_max_length = true;
652     var->has_bps_write_max_length = true;
653     var->has_iops_total_max_length = true;
654     var->has_iops_read_max_length = true;
655     var->has_iops_write_max_length = true;
656     var->has_iops_size = true;
657 }
658