block: throttle-groups: Use BlockBackend pointers internally
[qemu.git] / block / throttle-groups.c
1 /*
2 * QEMU block throttling group infrastructure
3 *
4 * Copyright (C) Nodalink, EURL. 2014
5 * Copyright (C) Igalia, S.L. 2015
6 *
7 * Authors:
8 * BenoƮt 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 "sysemu/block-backend.h"
27 #include "block/throttle-groups.h"
28 #include "qemu/queue.h"
29 #include "qemu/thread.h"
30 #include "sysemu/qtest.h"
31
32 /* The ThrottleGroup structure (with its ThrottleState) is shared
33 * among different BlockDriverState and it's independent from
34 * AioContext, so in order to use it from different threads it needs
35 * its own locking.
36 *
37 * This locking is however handled internally in this file, so it's
38 * transparent to outside users.
39 *
40 * The whole ThrottleGroup structure is private and invisible to
41 * outside users, that only use it through its ThrottleState.
42 *
43 * In addition to the ThrottleGroup structure, BlockDriverState has
44 * fields that need to be accessed by other members of the group and
45 * therefore also need to be protected by this lock. Once a BDS is
46 * registered in a group those fields can be accessed by other threads
47 * any time.
48 *
49 * Again, all this is handled internally and is mostly transparent to
50 * the outside. The 'throttle_timers' field however has an additional
51 * constraint because it may be temporarily invalid (see for example
52 * bdrv_set_aio_context()). Therefore in this file a thread will
53 * access some other BDS's timers only after verifying that that BDS
54 * has throttled requests in the queue.
55 */
56 typedef struct ThrottleGroup {
57 char *name; /* This is constant during the lifetime of the group */
58
59 QemuMutex lock; /* This lock protects the following four fields */
60 ThrottleState ts;
61 QLIST_HEAD(, BlockBackendPublic) head;
62 BlockBackend *tokens[2];
63 bool any_timer_armed[2];
64
65 /* These two are protected by the global throttle_groups_lock */
66 unsigned refcount;
67 QTAILQ_ENTRY(ThrottleGroup) list;
68 } ThrottleGroup;
69
70 static QemuMutex throttle_groups_lock;
71 static QTAILQ_HEAD(, ThrottleGroup) throttle_groups =
72 QTAILQ_HEAD_INITIALIZER(throttle_groups);
73
74 /* Increments the reference count of a ThrottleGroup given its name.
75 *
76 * If no ThrottleGroup is found with the given name a new one is
77 * created.
78 *
79 * @name: the name of the ThrottleGroup
80 * @ret: the ThrottleState member of the ThrottleGroup
81 */
82 ThrottleState *throttle_group_incref(const char *name)
83 {
84 ThrottleGroup *tg = NULL;
85 ThrottleGroup *iter;
86
87 qemu_mutex_lock(&throttle_groups_lock);
88
89 /* Look for an existing group with that name */
90 QTAILQ_FOREACH(iter, &throttle_groups, list) {
91 if (!strcmp(name, iter->name)) {
92 tg = iter;
93 break;
94 }
95 }
96
97 /* Create a new one if not found */
98 if (!tg) {
99 tg = g_new0(ThrottleGroup, 1);
100 tg->name = g_strdup(name);
101 qemu_mutex_init(&tg->lock);
102 throttle_init(&tg->ts);
103 QLIST_INIT(&tg->head);
104
105 QTAILQ_INSERT_TAIL(&throttle_groups, tg, list);
106 }
107
108 tg->refcount++;
109
110 qemu_mutex_unlock(&throttle_groups_lock);
111
112 return &tg->ts;
113 }
114
115 /* Decrease the reference count of a ThrottleGroup.
116 *
117 * When the reference count reaches zero the ThrottleGroup is
118 * destroyed.
119 *
120 * @ts: The ThrottleGroup to unref, given by its ThrottleState member
121 */
122 void throttle_group_unref(ThrottleState *ts)
123 {
124 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
125
126 qemu_mutex_lock(&throttle_groups_lock);
127 if (--tg->refcount == 0) {
128 QTAILQ_REMOVE(&throttle_groups, tg, list);
129 qemu_mutex_destroy(&tg->lock);
130 g_free(tg->name);
131 g_free(tg);
132 }
133 qemu_mutex_unlock(&throttle_groups_lock);
134 }
135
136 /* Get the name from a BlockDriverState's ThrottleGroup. The name (and
137 * the pointer) is guaranteed to remain constant during the lifetime
138 * of the group.
139 *
140 * @bs: a BlockDriverState that is member of a throttling group
141 * @ret: the name of the group.
142 */
143 const char *throttle_group_get_name(BlockDriverState *bs)
144 {
145 ThrottleGroup *tg = container_of(bs->throttle_state, ThrottleGroup, ts);
146 return tg->name;
147 }
148
149 /* Return the next BlockBackend in the round-robin sequence, simulating a
150 * circular list.
151 *
152 * This assumes that tg->lock is held.
153 *
154 * @blk: the current BlockBackend
155 * @ret: the next BlockBackend in the sequence
156 */
157 static BlockBackend *throttle_group_next_blk(BlockBackend *blk)
158 {
159 BlockDriverState *bs = blk_bs(blk);
160 ThrottleState *ts = bs->throttle_state;
161 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
162 BlockBackendPublic *next = QLIST_NEXT(blk_get_public(blk), round_robin);
163
164 if (!next) {
165 next = QLIST_FIRST(&tg->head);
166 }
167
168 return blk_by_public(next);
169 }
170
171 /* Return the next BlockBackend in the round-robin sequence with pending I/O
172 * requests.
173 *
174 * This assumes that tg->lock is held.
175 *
176 * @blk: the current BlockBackend
177 * @is_write: the type of operation (read/write)
178 * @ret: the next BlockBackend with pending requests, or blk if there is
179 * none.
180 */
181 static BlockBackend *next_throttle_token(BlockBackend *blk, bool is_write)
182 {
183 ThrottleGroup *tg = container_of(blk_bs(blk)->throttle_state,
184 ThrottleGroup, ts);
185 BlockBackend *token, *start;
186
187 start = token = tg->tokens[is_write];
188
189 /* get next bs round in round robin style */
190 token = throttle_group_next_blk(token);
191 while (token != start && !blk_bs(token)->pending_reqs[is_write]) {
192 token = throttle_group_next_blk(token);
193 }
194
195 /* If no IO are queued for scheduling on the next round robin token
196 * then decide the token is the current bs because chances are
197 * the current bs get the current request queued.
198 */
199 if (token == start && !blk_bs(token)->pending_reqs[is_write]) {
200 token = blk;
201 }
202
203 return token;
204 }
205
206 /* Check if the next I/O request for a BlockBackend needs to be throttled or
207 * not. If there's no timer set in this group, set one and update the token
208 * accordingly.
209 *
210 * This assumes that tg->lock is held.
211 *
212 * @blk: the current BlockBackend
213 * @is_write: the type of operation (read/write)
214 * @ret: whether the I/O request needs to be throttled or not
215 */
216 static bool throttle_group_schedule_timer(BlockBackend *blk, bool is_write)
217 {
218 ThrottleState *ts = blk_bs(blk)->throttle_state;
219 ThrottleTimers *tt = &blk_bs(blk)->throttle_timers;
220 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
221 bool must_wait;
222
223 if (blk_bs(blk)->io_limits_disabled) {
224 return false;
225 }
226
227 /* Check if any of the timers in this group is already armed */
228 if (tg->any_timer_armed[is_write]) {
229 return true;
230 }
231
232 must_wait = throttle_schedule_timer(ts, tt, is_write);
233
234 /* If a timer just got armed, set blk as the current token */
235 if (must_wait) {
236 tg->tokens[is_write] = blk;
237 tg->any_timer_armed[is_write] = true;
238 }
239
240 return must_wait;
241 }
242
243 /* Look for the next pending I/O request and schedule it.
244 *
245 * This assumes that tg->lock is held.
246 *
247 * @blk: the current BlockBackend
248 * @is_write: the type of operation (read/write)
249 */
250 static void schedule_next_request(BlockBackend *blk, bool is_write)
251 {
252 BlockDriverState *bs = blk_bs(blk);
253 ThrottleGroup *tg = container_of(bs->throttle_state, ThrottleGroup, ts);
254 bool must_wait;
255 BlockBackend *token;
256
257 /* Check if there's any pending request to schedule next */
258 token = next_throttle_token(blk, is_write);
259 if (!blk_bs(token)->pending_reqs[is_write]) {
260 return;
261 }
262
263 /* Set a timer for the request if it needs to be throttled */
264 must_wait = throttle_group_schedule_timer(token, is_write);
265
266 /* If it doesn't have to wait, queue it for immediate execution */
267 if (!must_wait) {
268 /* Give preference to requests from the current bs */
269 if (qemu_in_coroutine() &&
270 qemu_co_queue_next(&bs->throttled_reqs[is_write])) {
271 token = blk;
272 } else {
273 ThrottleTimers *tt = &blk_bs(token)->throttle_timers;
274 int64_t now = qemu_clock_get_ns(tt->clock_type);
275 timer_mod(tt->timers[is_write], now + 1);
276 tg->any_timer_armed[is_write] = true;
277 }
278 tg->tokens[is_write] = token;
279 }
280 }
281
282 /* Check if an I/O request needs to be throttled, wait and set a timer
283 * if necessary, and schedule the next request using a round robin
284 * algorithm.
285 *
286 * @bs: the current BlockDriverState
287 * @bytes: the number of bytes for this I/O
288 * @is_write: the type of operation (read/write)
289 */
290 void coroutine_fn throttle_group_co_io_limits_intercept(BlockDriverState *bs,
291 unsigned int bytes,
292 bool is_write)
293 {
294 bool must_wait;
295 BlockBackend *token;
296
297 ThrottleGroup *tg = container_of(bs->throttle_state, ThrottleGroup, ts);
298 qemu_mutex_lock(&tg->lock);
299
300 /* First we check if this I/O has to be throttled. */
301 token = next_throttle_token(bs->blk, is_write);
302 must_wait = throttle_group_schedule_timer(token, is_write);
303
304 /* Wait if there's a timer set or queued requests of this type */
305 if (must_wait || bs->pending_reqs[is_write]) {
306 bs->pending_reqs[is_write]++;
307 qemu_mutex_unlock(&tg->lock);
308 qemu_co_queue_wait(&bs->throttled_reqs[is_write]);
309 qemu_mutex_lock(&tg->lock);
310 bs->pending_reqs[is_write]--;
311 }
312
313 /* The I/O will be executed, so do the accounting */
314 throttle_account(bs->throttle_state, is_write, bytes);
315
316 /* Schedule the next request */
317 schedule_next_request(bs->blk, is_write);
318
319 qemu_mutex_unlock(&tg->lock);
320 }
321
322 void throttle_group_restart_bs(BlockDriverState *bs)
323 {
324 int i;
325
326 for (i = 0; i < 2; i++) {
327 while (qemu_co_enter_next(&bs->throttled_reqs[i])) {
328 ;
329 }
330 }
331 }
332
333 /* Update the throttle configuration for a particular group. Similar
334 * to throttle_config(), but guarantees atomicity within the
335 * throttling group.
336 *
337 * @bs: a BlockDriverState that is member of the group
338 * @cfg: the configuration to set
339 */
340 void throttle_group_config(BlockDriverState *bs, ThrottleConfig *cfg)
341 {
342 ThrottleTimers *tt = &bs->throttle_timers;
343 ThrottleState *ts = bs->throttle_state;
344 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
345 qemu_mutex_lock(&tg->lock);
346 /* throttle_config() cancels the timers */
347 if (timer_pending(tt->timers[0])) {
348 tg->any_timer_armed[0] = false;
349 }
350 if (timer_pending(tt->timers[1])) {
351 tg->any_timer_armed[1] = false;
352 }
353 throttle_config(ts, tt, cfg);
354 qemu_mutex_unlock(&tg->lock);
355
356 qemu_co_enter_next(&bs->throttled_reqs[0]);
357 qemu_co_enter_next(&bs->throttled_reqs[1]);
358 }
359
360 /* Get the throttle configuration from a particular group. Similar to
361 * throttle_get_config(), but guarantees atomicity within the
362 * throttling group.
363 *
364 * @bs: a BlockDriverState that is member of the group
365 * @cfg: the configuration will be written here
366 */
367 void throttle_group_get_config(BlockDriverState *bs, ThrottleConfig *cfg)
368 {
369 ThrottleState *ts = bs->throttle_state;
370 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
371 qemu_mutex_lock(&tg->lock);
372 throttle_get_config(ts, cfg);
373 qemu_mutex_unlock(&tg->lock);
374 }
375
376 /* ThrottleTimers callback. This wakes up a request that was waiting
377 * because it had been throttled.
378 *
379 * @bs: the BlockDriverState whose request had been throttled
380 * @is_write: the type of operation (read/write)
381 */
382 static void timer_cb(BlockDriverState *bs, bool is_write)
383 {
384 ThrottleState *ts = bs->throttle_state;
385 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
386 bool empty_queue;
387
388 /* The timer has just been fired, so we can update the flag */
389 qemu_mutex_lock(&tg->lock);
390 tg->any_timer_armed[is_write] = false;
391 qemu_mutex_unlock(&tg->lock);
392
393 /* Run the request that was waiting for this timer */
394 empty_queue = !qemu_co_enter_next(&bs->throttled_reqs[is_write]);
395
396 /* If the request queue was empty then we have to take care of
397 * scheduling the next one */
398 if (empty_queue) {
399 qemu_mutex_lock(&tg->lock);
400 schedule_next_request(bs->blk, is_write);
401 qemu_mutex_unlock(&tg->lock);
402 }
403 }
404
405 static void read_timer_cb(void *opaque)
406 {
407 timer_cb(opaque, false);
408 }
409
410 static void write_timer_cb(void *opaque)
411 {
412 timer_cb(opaque, true);
413 }
414
415 /* Register a BlockBackend in the throttling group, also initializing its
416 * timers and updating its throttle_state pointer to point to it. If a
417 * throttling group with that name does not exist yet, it will be created.
418 *
419 * @blk: the BlockBackend to insert
420 * @groupname: the name of the group
421 */
422 void throttle_group_register_blk(BlockBackend *blk, const char *groupname)
423 {
424 int i;
425 BlockDriverState *bs = blk_bs(blk);
426 ThrottleState *ts = throttle_group_incref(groupname);
427 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
428 int clock_type = QEMU_CLOCK_REALTIME;
429
430 if (qtest_enabled()) {
431 /* For testing block IO throttling only */
432 clock_type = QEMU_CLOCK_VIRTUAL;
433 }
434
435 bs->throttle_state = ts;
436
437 qemu_mutex_lock(&tg->lock);
438 /* If the ThrottleGroup is new set this BlockBackend as the token */
439 for (i = 0; i < 2; i++) {
440 if (!tg->tokens[i]) {
441 tg->tokens[i] = blk;
442 }
443 }
444
445 QLIST_INSERT_HEAD(&tg->head, blk_get_public(blk), round_robin);
446
447 throttle_timers_init(&bs->throttle_timers,
448 bdrv_get_aio_context(bs),
449 clock_type,
450 read_timer_cb,
451 write_timer_cb,
452 bs);
453
454 qemu_mutex_unlock(&tg->lock);
455 }
456
457 /* Unregister a BlockBackend from its group, removing it from the list,
458 * destroying the timers and setting the throttle_state pointer to NULL.
459 *
460 * The BlockBackend must not have pending throttled requests, so the caller has
461 * to drain them first.
462 *
463 * The group will be destroyed if it's empty after this operation.
464 *
465 * @blk: the BlockBackend to remove
466 */
467 void throttle_group_unregister_blk(BlockBackend *blk)
468 {
469 BlockDriverState *bs = blk_bs(blk);
470 ThrottleGroup *tg = container_of(bs->throttle_state, ThrottleGroup, ts);
471 int i;
472
473 assert(bs->pending_reqs[0] == 0 && bs->pending_reqs[1] == 0);
474 assert(qemu_co_queue_empty(&bs->throttled_reqs[0]));
475 assert(qemu_co_queue_empty(&bs->throttled_reqs[1]));
476
477 qemu_mutex_lock(&tg->lock);
478 for (i = 0; i < 2; i++) {
479 if (tg->tokens[i] == blk) {
480 BlockBackend *token = throttle_group_next_blk(blk);
481 /* Take care of the case where this is the last bs in the group */
482 if (token == blk) {
483 token = NULL;
484 }
485 tg->tokens[i] = token;
486 }
487 }
488
489 /* remove the current bs from the list */
490 QLIST_REMOVE(blk_get_public(blk), round_robin);
491 throttle_timers_destroy(&bs->throttle_timers);
492 qemu_mutex_unlock(&tg->lock);
493
494 throttle_group_unref(&tg->ts);
495 bs->throttle_state = NULL;
496 }
497
498 static void throttle_groups_init(void)
499 {
500 qemu_mutex_init(&throttle_groups_lock);
501 }
502
503 block_init(throttle_groups_init);