4 * Userspace RCU library - Userspace workqeues
6 * Copyright (c) 2010 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
7 * Copyright (c) 2017 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
9 * This library is free software; you can redistribute it and/or
10 * modify it under the terms of the GNU Lesser General Public
11 * License as published by the Free Software Foundation; either
12 * version 2.1 of the License, or (at your option) any later version.
14 * This library is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
17 * Lesser General Public License for more details.
19 * You should have received a copy of the GNU Lesser General Public
20 * License along with this library; if not, write to the Free Software
21 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
38 #include "compat-getcpu.h"
39 #include "urcu/wfcqueue.h"
40 #include "urcu-call-rcu.h"
41 #include "urcu-pointer.h"
42 #include "urcu/list.h"
43 #include "urcu/futex.h"
44 #include "urcu/tls-compat.h"
48 #include "workqueue.h"
50 #define SET_AFFINITY_CHECK_PERIOD (1U << 8) /* 256 */
51 #define SET_AFFINITY_CHECK_PERIOD_MASK (SET_AFFINITY_CHECK_PERIOD - 1)
53 /* Data structure that identifies a workqueue. */
55 struct urcu_workqueue
{
57 * We do not align head on a different cache-line than tail
58 * mainly because call_rcu callback-invocation threads use
59 * batching ("splice") to get an entire list of callbacks, which
60 * effectively empties the queue, and requires to touch the tail
63 struct cds_wfcq_tail cbs_tail
;
64 struct cds_wfcq_head cbs_head
;
67 unsigned long qlen
; /* maintained for debugging. */
70 unsigned long loop_count
;
72 void (*grace_period_fct
)(struct urcu_workqueue
*workqueue
, void *priv
);
73 void (*initialize_worker_fct
)(struct urcu_workqueue
*workqueue
, void *priv
);
74 void (*finalize_worker_fct
)(struct urcu_workqueue
*workqueue
, void *priv
);
75 void (*worker_before_pause_fct
)(struct urcu_workqueue
*workqueue
, void *priv
);
76 void (*worker_after_resume_fct
)(struct urcu_workqueue
*workqueue
, void *priv
);
77 void (*worker_before_wait_fct
)(struct urcu_workqueue
*workqueue
, void *priv
);
78 void (*worker_after_wake_up_fct
)(struct urcu_workqueue
*workqueue
, void *priv
);
79 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
81 struct urcu_workqueue_completion
{
87 struct urcu_workqueue_completion_work
{
88 struct urcu_work work
;
89 struct urcu_workqueue_completion
*completion
;
93 * Periodically retry setting CPU affinity if we migrate.
94 * Losing affinity can be caused by CPU hotunplug/hotplug, or by
97 #if HAVE_SCHED_SETAFFINITY
98 static int set_thread_cpu_affinity(struct urcu_workqueue
*workqueue
)
103 if (workqueue
->cpu_affinity
< 0)
105 if (++workqueue
->loop_count
& SET_AFFINITY_CHECK_PERIOD_MASK
)
107 if (urcu_sched_getcpu() == workqueue
->cpu_affinity
)
111 CPU_SET(workqueue
->cpu_affinity
, &mask
);
112 #if SCHED_SETAFFINITY_ARGS == 2
113 ret
= sched_setaffinity(0, &mask
);
115 ret
= sched_setaffinity(0, sizeof(mask
), &mask
);
118 * EINVAL is fine: can be caused by hotunplugged CPUs, or by
119 * cpuset(7). This is why we should always retry if we detect
122 if (ret
&& errno
== EINVAL
) {
129 static int set_thread_cpu_affinity(struct urcu_workqueue
*workqueue
)
135 static void futex_wait(int32_t *futex
)
137 /* Read condition before read futex */
139 if (uatomic_read(futex
) != -1)
141 while (futex_async(futex
, FUTEX_WAIT
, -1, NULL
, NULL
, 0)) {
144 /* Value already changed. */
147 /* Retry if interrupted by signal. */
148 break; /* Get out of switch. */
150 /* Unexpected error. */
156 static void futex_wake_up(int32_t *futex
)
158 /* Write to condition before reading/writing futex */
160 if (caa_unlikely(uatomic_read(futex
) == -1)) {
161 uatomic_set(futex
, 0);
162 if (futex_async(futex
, FUTEX_WAKE
, 1,
168 /* This is the code run by each worker thread. */
170 static void *workqueue_thread(void *arg
)
172 unsigned long cbcount
;
173 struct urcu_workqueue
*workqueue
= (struct urcu_workqueue
*) arg
;
174 int rt
= !!(uatomic_read(&workqueue
->flags
) & URCU_WORKQUEUE_RT
);
176 if (set_thread_cpu_affinity(workqueue
))
179 if (workqueue
->initialize_worker_fct
)
180 workqueue
->initialize_worker_fct(workqueue
, workqueue
->priv
);
183 uatomic_dec(&workqueue
->futex
);
184 /* Decrement futex before reading workqueue */
188 struct cds_wfcq_head cbs_tmp_head
;
189 struct cds_wfcq_tail cbs_tmp_tail
;
190 struct cds_wfcq_node
*cbs
, *cbs_tmp_n
;
191 enum cds_wfcq_ret splice_ret
;
193 if (set_thread_cpu_affinity(workqueue
))
196 if (uatomic_read(&workqueue
->flags
) & URCU_WORKQUEUE_PAUSE
) {
198 * Pause requested. Become quiescent: remove
199 * ourself from all global lists, and don't
200 * process any callback. The callback lists may
201 * still be non-empty though.
203 if (workqueue
->worker_before_pause_fct
)
204 workqueue
->worker_before_pause_fct(workqueue
, workqueue
->priv
);
205 cmm_smp_mb__before_uatomic_or();
206 uatomic_or(&workqueue
->flags
, URCU_WORKQUEUE_PAUSED
);
207 while ((uatomic_read(&workqueue
->flags
) & URCU_WORKQUEUE_PAUSE
) != 0)
208 (void) poll(NULL
, 0, 1);
209 uatomic_and(&workqueue
->flags
, ~URCU_WORKQUEUE_PAUSED
);
210 cmm_smp_mb__after_uatomic_and();
211 if (workqueue
->worker_after_resume_fct
)
212 workqueue
->worker_after_resume_fct(workqueue
, workqueue
->priv
);
215 cds_wfcq_init(&cbs_tmp_head
, &cbs_tmp_tail
);
216 splice_ret
= __cds_wfcq_splice_blocking(&cbs_tmp_head
,
217 &cbs_tmp_tail
, &workqueue
->cbs_head
, &workqueue
->cbs_tail
);
218 assert(splice_ret
!= CDS_WFCQ_RET_WOULDBLOCK
);
219 assert(splice_ret
!= CDS_WFCQ_RET_DEST_NON_EMPTY
);
220 if (splice_ret
!= CDS_WFCQ_RET_SRC_EMPTY
) {
221 if (workqueue
->grace_period_fct
)
222 workqueue
->grace_period_fct(workqueue
, workqueue
->priv
);
224 __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head
,
225 &cbs_tmp_tail
, cbs
, cbs_tmp_n
) {
226 struct rcu_head
*rhp
;
228 rhp
= caa_container_of(cbs
,
229 struct rcu_head
, next
);
233 uatomic_sub(&workqueue
->qlen
, cbcount
);
235 if (uatomic_read(&workqueue
->flags
) & URCU_WORKQUEUE_STOP
)
237 if (workqueue
->worker_before_wait_fct
)
238 workqueue
->worker_before_wait_fct(workqueue
, workqueue
->priv
);
240 if (cds_wfcq_empty(&workqueue
->cbs_head
,
241 &workqueue
->cbs_tail
)) {
242 futex_wait(&workqueue
->futex
);
243 (void) poll(NULL
, 0, 10);
244 uatomic_dec(&workqueue
->futex
);
246 * Decrement futex before reading
251 (void) poll(NULL
, 0, 10);
254 (void) poll(NULL
, 0, 10);
256 if (workqueue
->worker_after_wake_up_fct
)
257 workqueue
->worker_after_wake_up_fct(workqueue
, workqueue
->priv
);
261 * Read call_rcu list before write futex.
264 uatomic_set(&workqueue
->futex
, 0);
266 if (workqueue
->finalize_worker_fct
)
267 workqueue
->finalize_worker_fct(workqueue
, workqueue
->priv
);
271 struct urcu_workqueue
*urcu_workqueue_create(unsigned long flags
,
272 int cpu_affinity
, void *priv
,
273 void (*grace_period_fct
)(struct urcu_workqueue
*workqueue
, void *priv
),
274 void (*initialize_worker_fct
)(struct urcu_workqueue
*workqueue
, void *priv
),
275 void (*finalize_worker_fct
)(struct urcu_workqueue
*workqueue
, void *priv
),
276 void (*worker_before_wait_fct
)(struct urcu_workqueue
*workqueue
, void *priv
),
277 void (*worker_after_wake_up_fct
)(struct urcu_workqueue
*workqueue
, void *priv
),
278 void (*worker_before_pause_fct
)(struct urcu_workqueue
*workqueue
, void *priv
),
279 void (*worker_after_resume_fct
)(struct urcu_workqueue
*workqueue
, void *priv
))
281 struct urcu_workqueue
*workqueue
;
284 workqueue
= malloc(sizeof(*workqueue
));
285 if (workqueue
== NULL
)
287 memset(workqueue
, '\0', sizeof(*workqueue
));
288 cds_wfcq_init(&workqueue
->cbs_head
, &workqueue
->cbs_tail
);
290 workqueue
->futex
= 0;
291 workqueue
->flags
= flags
;
292 workqueue
->priv
= priv
;
293 workqueue
->grace_period_fct
= grace_period_fct
;
294 workqueue
->initialize_worker_fct
= initialize_worker_fct
;
295 workqueue
->finalize_worker_fct
= finalize_worker_fct
;
296 workqueue
->worker_before_wait_fct
= worker_before_wait_fct
;
297 workqueue
->worker_after_wake_up_fct
= worker_after_wake_up_fct
;
298 workqueue
->worker_before_pause_fct
= worker_before_pause_fct
;
299 workqueue
->worker_after_resume_fct
= worker_after_resume_fct
;
300 workqueue
->cpu_affinity
= cpu_affinity
;
301 workqueue
->loop_count
= 0;
302 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
303 ret
= pthread_create(&workqueue
->tid
, NULL
, workqueue_thread
, workqueue
);
310 static void wake_worker_thread(struct urcu_workqueue
*workqueue
)
312 if (!(_CMM_LOAD_SHARED(workqueue
->flags
) & URCU_CALL_RCU_RT
))
313 futex_wake_up(&workqueue
->futex
);
316 static int urcu_workqueue_destroy_worker(struct urcu_workqueue
*workqueue
)
321 uatomic_or(&workqueue
->flags
, URCU_WORKQUEUE_STOP
);
322 wake_worker_thread(workqueue
);
324 ret
= pthread_join(workqueue
->tid
, &retval
);
328 if (retval
!= NULL
) {
331 workqueue
->flags
&= ~URCU_WORKQUEUE_STOP
;
336 void urcu_workqueue_destroy(struct urcu_workqueue
*workqueue
)
338 if (workqueue
== NULL
) {
341 if (urcu_workqueue_destroy_worker(workqueue
)) {
344 assert(cds_wfcq_empty(&workqueue
->cbs_head
, &workqueue
->cbs_tail
));
348 void urcu_workqueue_queue_work(struct urcu_workqueue
*workqueue
,
349 struct urcu_work
*work
,
350 void (*func
)(struct urcu_work
*work
))
352 cds_wfcq_node_init(&work
->next
);
354 cds_wfcq_enqueue(&workqueue
->cbs_head
, &workqueue
->cbs_tail
, &work
->next
);
355 uatomic_inc(&workqueue
->qlen
);
356 wake_worker_thread(workqueue
);
360 void free_completion(struct urcu_ref
*ref
)
362 struct urcu_workqueue_completion
*completion
;
364 completion
= caa_container_of(ref
, struct urcu_workqueue_completion
, ref
);
369 void _urcu_workqueue_wait_complete(struct urcu_work
*work
)
371 struct urcu_workqueue_completion_work
*completion_work
;
372 struct urcu_workqueue_completion
*completion
;
374 completion_work
= caa_container_of(work
, struct urcu_workqueue_completion_work
, work
);
375 completion
= completion_work
->completion
;
376 if (!uatomic_sub_return(&completion
->barrier_count
, 1))
377 futex_wake_up(&completion
->futex
);
378 urcu_ref_put(&completion
->ref
, free_completion
);
379 free(completion_work
);
382 struct urcu_workqueue_completion
*urcu_workqueue_create_completion(void)
384 struct urcu_workqueue_completion
*completion
;
386 completion
= calloc(sizeof(*completion
), 1);
389 urcu_ref_set(&completion
->ref
, 1);
390 completion
->barrier_count
= 0;
394 void urcu_workqueue_destroy_completion(struct urcu_workqueue_completion
*completion
)
396 urcu_ref_put(&completion
->ref
, free_completion
);
399 void urcu_workqueue_wait_completion(struct urcu_workqueue_completion
*completion
)
403 uatomic_dec(&completion
->futex
);
404 /* Decrement futex before reading barrier_count */
406 if (!uatomic_read(&completion
->barrier_count
))
408 futex_wait(&completion
->futex
);
412 void urcu_workqueue_queue_completion(struct urcu_workqueue
*workqueue
,
413 struct urcu_workqueue_completion
*completion
)
415 struct urcu_workqueue_completion_work
*work
;
417 work
= calloc(sizeof(*work
), 1);
420 work
->completion
= completion
;
421 urcu_ref_get(&completion
->ref
);
422 uatomic_inc(&completion
->barrier_count
);
423 urcu_workqueue_queue_work(workqueue
, &work
->work
, _urcu_workqueue_wait_complete
);
427 * Wait for all in-flight work to complete execution.
429 void urcu_workqueue_flush_queued_work(struct urcu_workqueue
*workqueue
)
431 struct urcu_workqueue_completion
*completion
;
433 completion
= urcu_workqueue_create_completion();
436 urcu_workqueue_queue_completion(workqueue
, completion
);
437 urcu_workqueue_wait_completion(completion
);
438 urcu_workqueue_destroy_completion(completion
);
441 /* To be used in before fork handler. */
442 void urcu_workqueue_pause_worker(struct urcu_workqueue
*workqueue
)
444 uatomic_or(&workqueue
->flags
, URCU_WORKQUEUE_PAUSE
);
445 cmm_smp_mb__after_uatomic_or();
446 wake_worker_thread(workqueue
);
448 while ((uatomic_read(&workqueue
->flags
) & URCU_WORKQUEUE_PAUSED
) == 0)
449 (void) poll(NULL
, 0, 1);
452 /* To be used in after fork parent handler. */
453 void urcu_workqueue_resume_worker(struct urcu_workqueue
*workqueue
)
455 uatomic_and(&workqueue
->flags
, ~URCU_WORKQUEUE_PAUSE
);
456 while ((uatomic_read(&workqueue
->flags
) & URCU_WORKQUEUE_PAUSED
) != 0)
457 (void) poll(NULL
, 0, 1);
460 void urcu_workqueue_create_worker(struct urcu_workqueue
*workqueue
)
464 /* Clear workqueue state from parent. */
465 workqueue
->flags
&= ~URCU_WORKQUEUE_PAUSED
;
466 workqueue
->flags
&= ~URCU_WORKQUEUE_PAUSE
;
468 ret
= pthread_create(&workqueue
->tid
, NULL
, workqueue_thread
, workqueue
);