2 * Copyright © 2016 Advanced Micro Devices, Inc.
5 * Permission is hereby granted, free of charge, to any person obtaining
6 * a copy of this software and associated documentation files (the
7 * "Software"), to deal in the Software without restriction, including
8 * without limitation the rights to use, copy, modify, merge, publish,
9 * distribute, sub license, and/or sell copies of the Software, and to
10 * permit persons to whom the Software is furnished to do so, subject to
11 * the following conditions:
13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND,
14 * EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES
15 * OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND
16 * NON-INFRINGEMENT. IN NO EVENT SHALL THE COPYRIGHT HOLDERS, AUTHORS
17 * AND/OR ITS SUPPLIERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
19 * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE
20 * USE OR OTHER DEALINGS IN THE SOFTWARE.
22 * The above copyright notice and this permission notice (including the
23 * next paragraph) shall be included in all copies or substantial portions
31 #include "util/os_time.h"
32 #include "util/u_string.h"
33 #include "util/u_thread.h"
36 static void util_queue_killall_and_wait(struct util_queue
*queue
);
38 /****************************************************************************
39 * Wait for all queues to assert idle when exit() is called.
41 * Otherwise, C++ static variable destructors can be called while threads
42 * are using the static variables.
45 static once_flag atexit_once_flag
= ONCE_FLAG_INIT
;
46 static struct list_head queue_list
;
47 static mtx_t exit_mutex
= _MTX_INITIALIZER_NP
;
52 struct util_queue
*iter
;
54 mtx_lock(&exit_mutex
);
55 /* Wait for all queues to assert idle. */
56 LIST_FOR_EACH_ENTRY(iter
, &queue_list
, head
) {
57 util_queue_killall_and_wait(iter
);
59 mtx_unlock(&exit_mutex
);
65 LIST_INITHEAD(&queue_list
);
66 atexit(atexit_handler
);
70 add_to_atexit_list(struct util_queue
*queue
)
72 call_once(&atexit_once_flag
, global_init
);
74 mtx_lock(&exit_mutex
);
75 LIST_ADD(&queue
->head
, &queue_list
);
76 mtx_unlock(&exit_mutex
);
80 remove_from_atexit_list(struct util_queue
*queue
)
82 struct util_queue
*iter
, *tmp
;
84 mtx_lock(&exit_mutex
);
85 LIST_FOR_EACH_ENTRY_SAFE(iter
, tmp
, &queue_list
, head
) {
87 LIST_DEL(&iter
->head
);
91 mtx_unlock(&exit_mutex
);
94 /****************************************************************************
98 #ifdef UTIL_QUEUE_FENCE_FUTEX
100 do_futex_fence_wait(struct util_queue_fence
*fence
,
101 bool timeout
, int64_t abs_timeout
)
103 uint32_t v
= fence
->val
;
105 ts
.tv_sec
= abs_timeout
/ (1000*1000*1000);
106 ts
.tv_nsec
= abs_timeout
% (1000*1000*1000);
110 v
= p_atomic_cmpxchg(&fence
->val
, 1, 2);
115 int r
= futex_wait(&fence
->val
, 2, timeout
? &ts
: NULL
);
116 if (timeout
&& r
< 0) {
117 if (errno
== ETIMEDOUT
)
128 _util_queue_fence_wait(struct util_queue_fence
*fence
)
130 do_futex_fence_wait(fence
, false, 0);
134 _util_queue_fence_wait_timeout(struct util_queue_fence
*fence
,
137 return do_futex_fence_wait(fence
, true, abs_timeout
);
142 #ifdef UTIL_QUEUE_FENCE_STANDARD
144 util_queue_fence_signal(struct util_queue_fence
*fence
)
146 mtx_lock(&fence
->mutex
);
147 fence
->signalled
= true;
148 cnd_broadcast(&fence
->cond
);
149 mtx_unlock(&fence
->mutex
);
153 _util_queue_fence_wait(struct util_queue_fence
*fence
)
155 mtx_lock(&fence
->mutex
);
156 while (!fence
->signalled
)
157 cnd_wait(&fence
->cond
, &fence
->mutex
);
158 mtx_unlock(&fence
->mutex
);
162 _util_queue_fence_wait_timeout(struct util_queue_fence
*fence
,
165 /* This terrible hack is made necessary by the fact that we really want an
166 * internal interface consistent with os_time_*, but cnd_timedwait is spec'd
167 * to be relative to the TIME_UTC clock.
169 int64_t rel
= abs_timeout
- os_time_get_nano();
174 timespec_get(&ts
, TIME_UTC
);
176 ts
.tv_sec
+= abs_timeout
/ (1000*1000*1000);
177 ts
.tv_nsec
+= abs_timeout
% (1000*1000*1000);
178 if (ts
.tv_nsec
>= (1000*1000*1000)) {
180 ts
.tv_nsec
-= (1000*1000*1000);
183 mtx_lock(&fence
->mutex
);
184 while (!fence
->signalled
) {
185 if (cnd_timedwait(&fence
->cond
, &fence
->mutex
, &ts
) != thrd_success
)
188 mtx_unlock(&fence
->mutex
);
191 return fence
->signalled
;
195 util_queue_fence_init(struct util_queue_fence
*fence
)
197 memset(fence
, 0, sizeof(*fence
));
198 (void) mtx_init(&fence
->mutex
, mtx_plain
);
199 cnd_init(&fence
->cond
);
200 fence
->signalled
= true;
204 util_queue_fence_destroy(struct util_queue_fence
*fence
)
206 assert(fence
->signalled
);
208 /* Ensure that another thread is not in the middle of
209 * util_queue_fence_signal (having set the fence to signalled but still
210 * holding the fence mutex).
212 * A common contract between threads is that as soon as a fence is signalled
213 * by thread A, thread B is allowed to destroy it. Since
214 * util_queue_fence_is_signalled does not lock the fence mutex (for
215 * performance reasons), we must do so here.
217 mtx_lock(&fence
->mutex
);
218 mtx_unlock(&fence
->mutex
);
220 cnd_destroy(&fence
->cond
);
221 mtx_destroy(&fence
->mutex
);
225 /****************************************************************************
226 * util_queue implementation
229 struct thread_input
{
230 struct util_queue
*queue
;
235 util_queue_thread_func(void *input
)
237 struct util_queue
*queue
= ((struct thread_input
*)input
)->queue
;
238 int thread_index
= ((struct thread_input
*)input
)->thread_index
;
244 util_snprintf(name
, sizeof(name
), "%s%i", queue
->name
, thread_index
);
245 u_thread_setname(name
);
249 struct util_queue_job job
;
251 mtx_lock(&queue
->lock
);
252 assert(queue
->num_queued
>= 0 && queue
->num_queued
<= queue
->max_jobs
);
254 /* wait if the queue is empty */
255 while (!queue
->kill_threads
&& queue
->num_queued
== 0)
256 cnd_wait(&queue
->has_queued_cond
, &queue
->lock
);
258 if (queue
->kill_threads
) {
259 mtx_unlock(&queue
->lock
);
263 job
= queue
->jobs
[queue
->read_idx
];
264 memset(&queue
->jobs
[queue
->read_idx
], 0, sizeof(struct util_queue_job
));
265 queue
->read_idx
= (queue
->read_idx
+ 1) % queue
->max_jobs
;
268 cnd_signal(&queue
->has_space_cond
);
269 mtx_unlock(&queue
->lock
);
272 job
.execute(job
.job
, thread_index
);
273 util_queue_fence_signal(job
.fence
);
275 job
.cleanup(job
.job
, thread_index
);
279 /* signal remaining jobs before terminating */
280 mtx_lock(&queue
->lock
);
281 for (unsigned i
= queue
->read_idx
; i
!= queue
->write_idx
;
282 i
= (i
+ 1) % queue
->max_jobs
) {
283 if (queue
->jobs
[i
].job
) {
284 util_queue_fence_signal(queue
->jobs
[i
].fence
);
285 queue
->jobs
[i
].job
= NULL
;
288 queue
->read_idx
= queue
->write_idx
;
289 queue
->num_queued
= 0;
290 mtx_unlock(&queue
->lock
);
295 util_queue_init(struct util_queue
*queue
,
298 unsigned num_threads
,
303 /* Form the thread name from process_name and name, limited to 13
304 * characters. Characters 14-15 are reserved for the thread number.
305 * Character 16 should be 0. Final form: "process:name12"
307 * If name is too long, it's truncated. If any space is left, the process
310 const char *process_name
= util_get_process_name();
311 int process_len
= process_name
? strlen(process_name
) : 0;
312 int name_len
= strlen(name
);
313 const int max_chars
= sizeof(queue
->name
) - 1;
315 name_len
= MIN2(name_len
, max_chars
);
317 /* See if there is any space left for the process name, reserve 1 for
319 process_len
= MIN2(process_len
, max_chars
- name_len
- 1);
320 process_len
= MAX2(process_len
, 0);
322 memset(queue
, 0, sizeof(*queue
));
325 snprintf(queue
->name
, sizeof(queue
->name
), "%.*s:%s",
326 process_len
, process_name
, name
);
328 snprintf(queue
->name
, sizeof(queue
->name
), "%s", name
);
331 queue
->flags
= flags
;
332 queue
->num_threads
= num_threads
;
333 queue
->max_jobs
= max_jobs
;
335 queue
->jobs
= (struct util_queue_job
*)
336 calloc(max_jobs
, sizeof(struct util_queue_job
));
340 (void) mtx_init(&queue
->lock
, mtx_plain
);
341 (void) mtx_init(&queue
->finish_lock
, mtx_plain
);
343 queue
->num_queued
= 0;
344 cnd_init(&queue
->has_queued_cond
);
345 cnd_init(&queue
->has_space_cond
);
347 queue
->threads
= (thrd_t
*) calloc(num_threads
, sizeof(thrd_t
));
352 for (i
= 0; i
< num_threads
; i
++) {
353 struct thread_input
*input
=
354 (struct thread_input
*) malloc(sizeof(struct thread_input
));
355 input
->queue
= queue
;
356 input
->thread_index
= i
;
358 queue
->threads
[i
] = u_thread_create(util_queue_thread_func
, input
);
360 if (!queue
->threads
[i
]) {
364 /* no threads created, fail */
367 /* at least one thread created, so use it */
368 queue
->num_threads
= i
;
373 if (flags
& UTIL_QUEUE_INIT_USE_MINIMUM_PRIORITY
) {
374 #if defined(__linux__) && defined(SCHED_IDLE)
375 struct sched_param sched_param
= {0};
377 /* The nice() function can only set a maximum of 19.
378 * SCHED_IDLE is the same as nice = 20.
380 * Note that Linux only allows decreasing the priority. The original
381 * priority can't be restored.
383 pthread_setschedparam(queue
->threads
[i
], SCHED_IDLE
, &sched_param
);
388 add_to_atexit_list(queue
);
392 free(queue
->threads
);
395 cnd_destroy(&queue
->has_space_cond
);
396 cnd_destroy(&queue
->has_queued_cond
);
397 mtx_destroy(&queue
->lock
);
400 /* also util_queue_is_initialized can be used to check for success */
401 memset(queue
, 0, sizeof(*queue
));
406 util_queue_killall_and_wait(struct util_queue
*queue
)
410 /* Signal all threads to terminate. */
411 mtx_lock(&queue
->lock
);
412 queue
->kill_threads
= 1;
413 cnd_broadcast(&queue
->has_queued_cond
);
414 mtx_unlock(&queue
->lock
);
416 for (i
= 0; i
< queue
->num_threads
; i
++)
417 thrd_join(queue
->threads
[i
], NULL
);
418 queue
->num_threads
= 0;
422 util_queue_destroy(struct util_queue
*queue
)
424 util_queue_killall_and_wait(queue
);
425 remove_from_atexit_list(queue
);
427 cnd_destroy(&queue
->has_space_cond
);
428 cnd_destroy(&queue
->has_queued_cond
);
429 mtx_destroy(&queue
->finish_lock
);
430 mtx_destroy(&queue
->lock
);
432 free(queue
->threads
);
436 util_queue_add_job(struct util_queue
*queue
,
438 struct util_queue_fence
*fence
,
439 util_queue_execute_func execute
,
440 util_queue_execute_func cleanup
)
442 struct util_queue_job
*ptr
;
444 mtx_lock(&queue
->lock
);
445 if (queue
->kill_threads
) {
446 mtx_unlock(&queue
->lock
);
447 /* well no good option here, but any leaks will be
448 * short-lived as things are shutting down..
453 util_queue_fence_reset(fence
);
455 assert(queue
->num_queued
>= 0 && queue
->num_queued
<= queue
->max_jobs
);
457 if (queue
->num_queued
== queue
->max_jobs
) {
458 if (queue
->flags
& UTIL_QUEUE_INIT_RESIZE_IF_FULL
) {
459 /* If the queue is full, make it larger to avoid waiting for a free
462 unsigned new_max_jobs
= queue
->max_jobs
+ 8;
463 struct util_queue_job
*jobs
=
464 (struct util_queue_job
*)calloc(new_max_jobs
,
465 sizeof(struct util_queue_job
));
468 /* Copy all queued jobs into the new list. */
469 unsigned num_jobs
= 0;
470 unsigned i
= queue
->read_idx
;
473 jobs
[num_jobs
++] = queue
->jobs
[i
];
474 i
= (i
+ 1) % queue
->max_jobs
;
475 } while (i
!= queue
->write_idx
);
477 assert(num_jobs
== queue
->num_queued
);
482 queue
->write_idx
= num_jobs
;
483 queue
->max_jobs
= new_max_jobs
;
485 /* Wait until there is a free slot. */
486 while (queue
->num_queued
== queue
->max_jobs
)
487 cnd_wait(&queue
->has_space_cond
, &queue
->lock
);
491 ptr
= &queue
->jobs
[queue
->write_idx
];
492 assert(ptr
->job
== NULL
);
495 ptr
->execute
= execute
;
496 ptr
->cleanup
= cleanup
;
497 queue
->write_idx
= (queue
->write_idx
+ 1) % queue
->max_jobs
;
500 cnd_signal(&queue
->has_queued_cond
);
501 mtx_unlock(&queue
->lock
);
505 * Remove a queued job. If the job hasn't started execution, it's removed from
506 * the queue. If the job has started execution, the function waits for it to
509 * In all cases, the fence is signalled when the function returns.
511 * The function can be used when destroying an object associated with the job
512 * when you don't care about the job completion state.
515 util_queue_drop_job(struct util_queue
*queue
, struct util_queue_fence
*fence
)
517 bool removed
= false;
519 if (util_queue_fence_is_signalled(fence
))
522 mtx_lock(&queue
->lock
);
523 for (unsigned i
= queue
->read_idx
; i
!= queue
->write_idx
;
524 i
= (i
+ 1) % queue
->max_jobs
) {
525 if (queue
->jobs
[i
].fence
== fence
) {
526 if (queue
->jobs
[i
].cleanup
)
527 queue
->jobs
[i
].cleanup(queue
->jobs
[i
].job
, -1);
529 /* Just clear it. The threads will treat as a no-op job. */
530 memset(&queue
->jobs
[i
], 0, sizeof(queue
->jobs
[i
]));
535 mtx_unlock(&queue
->lock
);
538 util_queue_fence_signal(fence
);
540 util_queue_fence_wait(fence
);
544 util_queue_finish_execute(void *data
, int num_thread
)
546 util_barrier
*barrier
= data
;
547 util_barrier_wait(barrier
);
551 * Wait until all previously added jobs have completed.
554 util_queue_finish(struct util_queue
*queue
)
556 util_barrier barrier
;
557 struct util_queue_fence
*fences
= malloc(queue
->num_threads
* sizeof(*fences
));
559 util_barrier_init(&barrier
, queue
->num_threads
);
561 /* If 2 threads were adding jobs for 2 different barries at the same time,
562 * a deadlock would happen, because 1 barrier requires that all threads
563 * wait for it exclusively.
565 mtx_lock(&queue
->finish_lock
);
567 for (unsigned i
= 0; i
< queue
->num_threads
; ++i
) {
568 util_queue_fence_init(&fences
[i
]);
569 util_queue_add_job(queue
, &barrier
, &fences
[i
], util_queue_finish_execute
, NULL
);
572 for (unsigned i
= 0; i
< queue
->num_threads
; ++i
) {
573 util_queue_fence_wait(&fences
[i
]);
574 util_queue_fence_destroy(&fences
[i
]);
576 mtx_unlock(&queue
->finish_lock
);
578 util_barrier_destroy(&barrier
);
584 util_queue_get_thread_time_nano(struct util_queue
*queue
, unsigned thread_index
)
586 /* Allow some flexibility by not raising an error. */
587 if (thread_index
>= queue
->num_threads
)
590 return u_thread_get_time_nano(queue
->threads
[thread_index
]);