tcg: add CONFIG_TCG guards in headers
[qemu/ar7.git] / block / throttle-groups.c
blobda2b490c381abb08e4be2e24b8df1a4e93ff651a
1 /*
2 * QEMU block throttling group infrastructure
4 * Copyright (C) Nodalink, EURL. 2014
5 * Copyright (C) Igalia, S.L. 2015
7 * Authors:
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/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"
32 /* The ThrottleGroup structure (with its ThrottleState) is shared
33 * among different BlockBackends and it's independent from
34 * AioContext, so in order to use it from different threads it needs
35 * its own locking.
37 * This locking is however handled internally in this file, so it's
38 * transparent to outside users.
40 * The whole ThrottleGroup structure is private and invisible to
41 * outside users, that only use it through its ThrottleState.
43 * In addition to the ThrottleGroup structure, BlockBackendPublic 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
46 * BlockBackend is registered in a group those fields can be accessed
47 * by other threads any time.
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 * blk_set_aio_context()). Therefore in this file a thread will
53 * access some other BlockBackend's timers only after verifying that
54 * that BlockBackend has throttled requests in the queue.
56 typedef struct ThrottleGroup {
57 char *name; /* This is constant during the lifetime of the group */
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];
65 /* These two are protected by the global throttle_groups_lock */
66 unsigned refcount;
67 QTAILQ_ENTRY(ThrottleGroup) list;
68 } ThrottleGroup;
70 static QemuMutex throttle_groups_lock;
71 static QTAILQ_HEAD(, ThrottleGroup) throttle_groups =
72 QTAILQ_HEAD_INITIALIZER(throttle_groups);
74 /* Increments the reference count of a ThrottleGroup given its name.
76 * If no ThrottleGroup is found with the given name a new one is
77 * created.
79 * @name: the name of the ThrottleGroup
80 * @ret: the ThrottleState member of the ThrottleGroup
82 ThrottleState *throttle_group_incref(const char *name)
84 ThrottleGroup *tg = NULL;
85 ThrottleGroup *iter;
87 qemu_mutex_lock(&throttle_groups_lock);
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;
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);
105 QTAILQ_INSERT_TAIL(&throttle_groups, tg, list);
108 tg->refcount++;
110 qemu_mutex_unlock(&throttle_groups_lock);
112 return &tg->ts;
115 /* Decrease the reference count of a ThrottleGroup.
117 * When the reference count reaches zero the ThrottleGroup is
118 * destroyed.
120 * @ts: The ThrottleGroup to unref, given by its ThrottleState member
122 void throttle_group_unref(ThrottleState *ts)
124 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
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);
133 qemu_mutex_unlock(&throttle_groups_lock);
136 /* Get the name from a BlockBackend's ThrottleGroup. The name (and the pointer)
137 * is guaranteed to remain constant during the lifetime of the group.
139 * @blk: a BlockBackend that is member of a throttling group
140 * @ret: the name of the group.
142 const char *throttle_group_get_name(BlockBackend *blk)
144 BlockBackendPublic *blkp = blk_get_public(blk);
145 ThrottleGroup *tg = container_of(blkp->throttle_state, ThrottleGroup, ts);
146 return tg->name;
149 /* Return the next BlockBackend in the round-robin sequence, simulating a
150 * circular list.
152 * This assumes that tg->lock is held.
154 * @blk: the current BlockBackend
155 * @ret: the next BlockBackend in the sequence
157 static BlockBackend *throttle_group_next_blk(BlockBackend *blk)
159 BlockBackendPublic *blkp = blk_get_public(blk);
160 ThrottleState *ts = blkp->throttle_state;
161 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
162 BlockBackendPublic *next = QLIST_NEXT(blkp, round_robin);
164 if (!next) {
165 next = QLIST_FIRST(&tg->head);
168 return blk_by_public(next);
172 * Return whether a BlockBackend has pending requests.
174 * This assumes that tg->lock is held.
176 * @blk: the BlockBackend
177 * @is_write: the type of operation (read/write)
178 * @ret: whether the BlockBackend has pending requests.
180 static inline bool blk_has_pending_reqs(BlockBackend *blk,
181 bool is_write)
183 const BlockBackendPublic *blkp = blk_get_public(blk);
184 return blkp->pending_reqs[is_write];
187 /* Return the next BlockBackend in the round-robin sequence with pending I/O
188 * requests.
190 * This assumes that tg->lock is held.
192 * @blk: the current BlockBackend
193 * @is_write: the type of operation (read/write)
194 * @ret: the next BlockBackend with pending requests, or blk if there is
195 * none.
197 static BlockBackend *next_throttle_token(BlockBackend *blk, bool is_write)
199 BlockBackendPublic *blkp = blk_get_public(blk);
200 ThrottleGroup *tg = container_of(blkp->throttle_state, ThrottleGroup, ts);
201 BlockBackend *token, *start;
203 start = token = tg->tokens[is_write];
205 /* get next bs round in round robin style */
206 token = throttle_group_next_blk(token);
207 while (token != start && !blk_has_pending_reqs(token, is_write)) {
208 token = throttle_group_next_blk(token);
211 /* If no IO are queued for scheduling on the next round robin token
212 * then decide the token is the current bs because chances are
213 * the current bs get the current request queued.
215 if (token == start && !blk_has_pending_reqs(token, is_write)) {
216 token = blk;
219 /* Either we return the original BB, or one with pending requests */
220 assert(token == blk || blk_has_pending_reqs(token, is_write));
222 return token;
225 /* Check if the next I/O request for a BlockBackend needs to be throttled or
226 * not. If there's no timer set in this group, set one and update the token
227 * accordingly.
229 * This assumes that tg->lock is held.
231 * @blk: the current BlockBackend
232 * @is_write: the type of operation (read/write)
233 * @ret: whether the I/O request needs to be throttled or not
235 static bool throttle_group_schedule_timer(BlockBackend *blk, bool is_write)
237 BlockBackendPublic *blkp = blk_get_public(blk);
238 ThrottleState *ts = blkp->throttle_state;
239 ThrottleTimers *tt = &blkp->throttle_timers;
240 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
241 bool must_wait;
243 if (atomic_read(&blkp->io_limits_disabled)) {
244 return false;
247 /* Check if any of the timers in this group is already armed */
248 if (tg->any_timer_armed[is_write]) {
249 return true;
252 must_wait = throttle_schedule_timer(ts, tt, is_write);
254 /* If a timer just got armed, set blk as the current token */
255 if (must_wait) {
256 tg->tokens[is_write] = blk;
257 tg->any_timer_armed[is_write] = true;
260 return must_wait;
263 /* Start the next pending I/O request for a BlockBackend. Return whether
264 * any request was actually pending.
266 * @blk: the current BlockBackend
267 * @is_write: the type of operation (read/write)
269 static bool coroutine_fn throttle_group_co_restart_queue(BlockBackend *blk,
270 bool is_write)
272 BlockBackendPublic *blkp = blk_get_public(blk);
273 bool ret;
275 qemu_co_mutex_lock(&blkp->throttled_reqs_lock);
276 ret = qemu_co_queue_next(&blkp->throttled_reqs[is_write]);
277 qemu_co_mutex_unlock(&blkp->throttled_reqs_lock);
279 return ret;
282 /* Look for the next pending I/O request and schedule it.
284 * This assumes that tg->lock is held.
286 * @blk: the current BlockBackend
287 * @is_write: the type of operation (read/write)
289 static void schedule_next_request(BlockBackend *blk, bool is_write)
291 BlockBackendPublic *blkp = blk_get_public(blk);
292 ThrottleGroup *tg = container_of(blkp->throttle_state, ThrottleGroup, ts);
293 bool must_wait;
294 BlockBackend *token;
296 /* Check if there's any pending request to schedule next */
297 token = next_throttle_token(blk, is_write);
298 if (!blk_has_pending_reqs(token, is_write)) {
299 return;
302 /* Set a timer for the request if it needs to be throttled */
303 must_wait = throttle_group_schedule_timer(token, is_write);
305 /* If it doesn't have to wait, queue it for immediate execution */
306 if (!must_wait) {
307 /* Give preference to requests from the current blk */
308 if (qemu_in_coroutine() &&
309 throttle_group_co_restart_queue(blk, is_write)) {
310 token = blk;
311 } else {
312 ThrottleTimers *tt = &blk_get_public(token)->throttle_timers;
313 int64_t now = qemu_clock_get_ns(tt->clock_type);
314 timer_mod(tt->timers[is_write], now);
315 tg->any_timer_armed[is_write] = true;
317 tg->tokens[is_write] = token;
321 /* Check if an I/O request needs to be throttled, wait and set a timer
322 * if necessary, and schedule the next request using a round robin
323 * algorithm.
325 * @blk: the current BlockBackend
326 * @bytes: the number of bytes for this I/O
327 * @is_write: the type of operation (read/write)
329 void coroutine_fn throttle_group_co_io_limits_intercept(BlockBackend *blk,
330 unsigned int bytes,
331 bool is_write)
333 bool must_wait;
334 BlockBackend *token;
336 BlockBackendPublic *blkp = blk_get_public(blk);
337 ThrottleGroup *tg = container_of(blkp->throttle_state, ThrottleGroup, ts);
338 qemu_mutex_lock(&tg->lock);
340 /* First we check if this I/O has to be throttled. */
341 token = next_throttle_token(blk, is_write);
342 must_wait = throttle_group_schedule_timer(token, is_write);
344 /* Wait if there's a timer set or queued requests of this type */
345 if (must_wait || blkp->pending_reqs[is_write]) {
346 blkp->pending_reqs[is_write]++;
347 qemu_mutex_unlock(&tg->lock);
348 qemu_co_mutex_lock(&blkp->throttled_reqs_lock);
349 qemu_co_queue_wait(&blkp->throttled_reqs[is_write],
350 &blkp->throttled_reqs_lock);
351 qemu_co_mutex_unlock(&blkp->throttled_reqs_lock);
352 qemu_mutex_lock(&tg->lock);
353 blkp->pending_reqs[is_write]--;
356 /* The I/O will be executed, so do the accounting */
357 throttle_account(blkp->throttle_state, is_write, bytes);
359 /* Schedule the next request */
360 schedule_next_request(blk, is_write);
362 qemu_mutex_unlock(&tg->lock);
365 typedef struct {
366 BlockBackend *blk;
367 bool is_write;
368 } RestartData;
370 static void coroutine_fn throttle_group_restart_queue_entry(void *opaque)
372 RestartData *data = opaque;
373 BlockBackend *blk = data->blk;
374 bool is_write = data->is_write;
375 BlockBackendPublic *blkp = blk_get_public(blk);
376 ThrottleGroup *tg = container_of(blkp->throttle_state, ThrottleGroup, ts);
377 bool empty_queue;
379 empty_queue = !throttle_group_co_restart_queue(blk, is_write);
381 /* If the request queue was empty then we have to take care of
382 * scheduling the next one */
383 if (empty_queue) {
384 qemu_mutex_lock(&tg->lock);
385 schedule_next_request(blk, is_write);
386 qemu_mutex_unlock(&tg->lock);
390 static void throttle_group_restart_queue(BlockBackend *blk, bool is_write)
392 Coroutine *co;
393 RestartData rd = {
394 .blk = blk,
395 .is_write = is_write
398 co = qemu_coroutine_create(throttle_group_restart_queue_entry, &rd);
399 aio_co_enter(blk_get_aio_context(blk), co);
402 void throttle_group_restart_blk(BlockBackend *blk)
404 BlockBackendPublic *blkp = blk_get_public(blk);
406 if (blkp->throttle_state) {
407 throttle_group_restart_queue(blk, 0);
408 throttle_group_restart_queue(blk, 1);
412 /* Update the throttle configuration for a particular group. Similar
413 * to throttle_config(), but guarantees atomicity within the
414 * throttling group.
416 * @blk: a BlockBackend that is a member of the group
417 * @cfg: the configuration to set
419 void throttle_group_config(BlockBackend *blk, ThrottleConfig *cfg)
421 BlockBackendPublic *blkp = blk_get_public(blk);
422 ThrottleTimers *tt = &blkp->throttle_timers;
423 ThrottleState *ts = blkp->throttle_state;
424 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
425 qemu_mutex_lock(&tg->lock);
426 /* throttle_config() cancels the timers */
427 if (timer_pending(tt->timers[0])) {
428 tg->any_timer_armed[0] = false;
430 if (timer_pending(tt->timers[1])) {
431 tg->any_timer_armed[1] = false;
433 throttle_config(ts, tt, cfg);
434 qemu_mutex_unlock(&tg->lock);
436 throttle_group_restart_blk(blk);
439 /* Get the throttle configuration from a particular group. Similar to
440 * throttle_get_config(), but guarantees atomicity within the
441 * throttling group.
443 * @blk: a BlockBackend that is a member of the group
444 * @cfg: the configuration will be written here
446 void throttle_group_get_config(BlockBackend *blk, ThrottleConfig *cfg)
448 BlockBackendPublic *blkp = blk_get_public(blk);
449 ThrottleState *ts = blkp->throttle_state;
450 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
451 qemu_mutex_lock(&tg->lock);
452 throttle_get_config(ts, cfg);
453 qemu_mutex_unlock(&tg->lock);
456 /* ThrottleTimers callback. This wakes up a request that was waiting
457 * because it had been throttled.
459 * @blk: the BlockBackend whose request had been throttled
460 * @is_write: the type of operation (read/write)
462 static void timer_cb(BlockBackend *blk, bool is_write)
464 BlockBackendPublic *blkp = blk_get_public(blk);
465 ThrottleState *ts = blkp->throttle_state;
466 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
468 /* The timer has just been fired, so we can update the flag */
469 qemu_mutex_lock(&tg->lock);
470 tg->any_timer_armed[is_write] = false;
471 qemu_mutex_unlock(&tg->lock);
473 /* Run the request that was waiting for this timer */
474 throttle_group_restart_queue(blk, is_write);
477 static void read_timer_cb(void *opaque)
479 timer_cb(opaque, false);
482 static void write_timer_cb(void *opaque)
484 timer_cb(opaque, true);
487 /* Register a BlockBackend in the throttling group, also initializing its
488 * timers and updating its throttle_state pointer to point to it. If a
489 * throttling group with that name does not exist yet, it will be created.
491 * @blk: the BlockBackend to insert
492 * @groupname: the name of the group
494 void throttle_group_register_blk(BlockBackend *blk, const char *groupname)
496 int i;
497 BlockBackendPublic *blkp = blk_get_public(blk);
498 ThrottleState *ts = throttle_group_incref(groupname);
499 ThrottleGroup *tg = container_of(ts, ThrottleGroup, ts);
500 int clock_type = QEMU_CLOCK_REALTIME;
502 if (qtest_enabled()) {
503 /* For testing block IO throttling only */
504 clock_type = QEMU_CLOCK_VIRTUAL;
507 blkp->throttle_state = ts;
509 qemu_mutex_lock(&tg->lock);
510 /* If the ThrottleGroup is new set this BlockBackend as the token */
511 for (i = 0; i < 2; i++) {
512 if (!tg->tokens[i]) {
513 tg->tokens[i] = blk;
517 QLIST_INSERT_HEAD(&tg->head, blkp, round_robin);
519 throttle_timers_init(&blkp->throttle_timers,
520 blk_get_aio_context(blk),
521 clock_type,
522 read_timer_cb,
523 write_timer_cb,
524 blk);
526 qemu_mutex_unlock(&tg->lock);
529 /* Unregister a BlockBackend from its group, removing it from the list,
530 * destroying the timers and setting the throttle_state pointer to NULL.
532 * The BlockBackend must not have pending throttled requests, so the caller has
533 * to drain them first.
535 * The group will be destroyed if it's empty after this operation.
537 * @blk: the BlockBackend to remove
539 void throttle_group_unregister_blk(BlockBackend *blk)
541 BlockBackendPublic *blkp = blk_get_public(blk);
542 ThrottleGroup *tg = container_of(blkp->throttle_state, ThrottleGroup, ts);
543 int i;
545 assert(blkp->pending_reqs[0] == 0 && blkp->pending_reqs[1] == 0);
546 assert(qemu_co_queue_empty(&blkp->throttled_reqs[0]));
547 assert(qemu_co_queue_empty(&blkp->throttled_reqs[1]));
549 qemu_mutex_lock(&tg->lock);
550 for (i = 0; i < 2; i++) {
551 if (tg->tokens[i] == blk) {
552 BlockBackend *token = throttle_group_next_blk(blk);
553 /* Take care of the case where this is the last blk in the group */
554 if (token == blk) {
555 token = NULL;
557 tg->tokens[i] = token;
561 /* remove the current blk from the list */
562 QLIST_REMOVE(blkp, round_robin);
563 throttle_timers_destroy(&blkp->throttle_timers);
564 qemu_mutex_unlock(&tg->lock);
566 throttle_group_unref(&tg->ts);
567 blkp->throttle_state = NULL;
570 static void throttle_groups_init(void)
572 qemu_mutex_init(&throttle_groups_lock);
575 block_init(throttle_groups_init);