4 * Userspace RCU library - batch memory reclamation with kernel API
6 * Copyright (c) 2010 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
8 * This library is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU Lesser General Public
10 * License as published by the Free Software Foundation; either
11 * version 2.1 of the License, or (at your option) any later version.
13 * This library is distributed in the hope that it will be useful,
14 * but WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * Lesser General Public License for more details.
18 * You should have received a copy of the GNU Lesser General Public
19 * License along with this library; if not, write to the Free Software
20 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
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 #define SET_AFFINITY_CHECK_PERIOD (1U << 8) /* 256 */
49 #define SET_AFFINITY_CHECK_PERIOD_MASK (SET_AFFINITY_CHECK_PERIOD - 1)
51 /* Data structure that identifies a call_rcu thread. */
53 struct call_rcu_data
{
55 * We do not align head on a different cache-line than tail
56 * mainly because call_rcu callback-invocation threads use
57 * batching ("splice") to get an entire list of callbacks, which
58 * effectively empties the queue, and requires to touch the tail
61 struct cds_wfcq_tail cbs_tail
;
62 struct cds_wfcq_head cbs_head
;
65 unsigned long qlen
; /* maintained for debugging. */
68 unsigned long gp_count
;
69 struct cds_list_head list
;
70 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
72 struct call_rcu_completion
{
78 struct call_rcu_completion_work
{
80 struct call_rcu_completion
*completion
;
84 * List of all call_rcu_data structures to keep valgrind happy.
85 * Protected by call_rcu_mutex.
88 static CDS_LIST_HEAD(call_rcu_data_list
);
90 /* Link a thread using call_rcu() to its call_rcu thread. */
92 static DEFINE_URCU_TLS(struct call_rcu_data
*, thread_call_rcu_data
);
95 * Guard call_rcu thread creation and atfork handlers.
97 static pthread_mutex_t call_rcu_mutex
= PTHREAD_MUTEX_INITIALIZER
;
99 /* If a given thread does not have its own call_rcu thread, this is default. */
101 static struct call_rcu_data
*default_call_rcu_data
;
104 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
105 * available, then we can have call_rcu threads assigned to individual
106 * CPUs rather than only to specific threads.
109 #ifdef HAVE_SCHED_GETCPU
111 static int urcu_sched_getcpu(void)
113 return sched_getcpu();
116 #else /* #ifdef HAVE_SCHED_GETCPU */
118 static int urcu_sched_getcpu(void)
123 #endif /* #else #ifdef HAVE_SCHED_GETCPU */
125 #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU)
128 * Pointer to array of pointers to per-CPU call_rcu_data structures
129 * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
130 * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
131 * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
132 * without mutex. The call_rcu_mutex protects updates.
135 static struct call_rcu_data
**per_cpu_call_rcu_data
;
138 static void maxcpus_reset(void)
143 /* Allocate the array if it has not already been allocated. */
145 static void alloc_cpu_call_rcu_data(void)
147 struct call_rcu_data
**p
;
148 static int warned
= 0;
152 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
156 p
= malloc(maxcpus
* sizeof(*per_cpu_call_rcu_data
));
158 memset(p
, '\0', maxcpus
* sizeof(*per_cpu_call_rcu_data
));
159 rcu_set_pointer(&per_cpu_call_rcu_data
, p
);
162 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
168 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
171 * per_cpu_call_rcu_data should be constant, but some functions below, used both
172 * for cases where cpu number is available and not available, assume it it not
175 static struct call_rcu_data
**per_cpu_call_rcu_data
= NULL
;
176 static const long maxcpus
= -1;
178 static void maxcpus_reset(void)
182 static void alloc_cpu_call_rcu_data(void)
186 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
188 /* Acquire the specified pthread mutex. */
190 static void call_rcu_lock(pthread_mutex_t
*pmp
)
194 ret
= pthread_mutex_lock(pmp
);
199 /* Release the specified pthread mutex. */
201 static void call_rcu_unlock(pthread_mutex_t
*pmp
)
205 ret
= pthread_mutex_unlock(pmp
);
211 * Periodically retry setting CPU affinity if we migrate.
212 * Losing affinity can be caused by CPU hotunplug/hotplug, or by
215 #if HAVE_SCHED_SETAFFINITY
217 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
222 if (crdp
->cpu_affinity
< 0)
224 if (++crdp
->gp_count
& SET_AFFINITY_CHECK_PERIOD_MASK
)
226 if (urcu_sched_getcpu() == crdp
->cpu_affinity
)
230 CPU_SET(crdp
->cpu_affinity
, &mask
);
231 #if SCHED_SETAFFINITY_ARGS == 2
232 ret
= sched_setaffinity(0, &mask
);
234 ret
= sched_setaffinity(0, sizeof(mask
), &mask
);
237 * EINVAL is fine: can be caused by hotunplugged CPUs, or by
238 * cpuset(7). This is why we should always retry if we detect
241 if (ret
&& errno
== EINVAL
) {
249 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
255 static void call_rcu_wait(struct call_rcu_data
*crdp
)
257 /* Read call_rcu list before read futex */
259 if (uatomic_read(&crdp
->futex
) == -1)
260 futex_async(&crdp
->futex
, FUTEX_WAIT
, -1,
264 static void call_rcu_wake_up(struct call_rcu_data
*crdp
)
266 /* Write to call_rcu list before reading/writing futex */
268 if (caa_unlikely(uatomic_read(&crdp
->futex
) == -1)) {
269 uatomic_set(&crdp
->futex
, 0);
270 futex_async(&crdp
->futex
, FUTEX_WAKE
, 1,
275 static void call_rcu_completion_wait(struct call_rcu_completion
*completion
)
277 /* Read completion barrier count before read futex */
279 if (uatomic_read(&completion
->futex
) == -1)
280 futex_async(&completion
->futex
, FUTEX_WAIT
, -1,
284 static void call_rcu_completion_wake_up(struct call_rcu_completion
*completion
)
286 /* Write to completion barrier count before reading/writing futex */
288 if (caa_unlikely(uatomic_read(&completion
->futex
) == -1)) {
289 uatomic_set(&completion
->futex
, 0);
290 futex_async(&completion
->futex
, FUTEX_WAKE
, 1,
295 /* This is the code run by each call_rcu thread. */
297 static void *call_rcu_thread(void *arg
)
299 unsigned long cbcount
;
300 struct call_rcu_data
*crdp
= (struct call_rcu_data
*) arg
;
301 int rt
= !!(uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_RT
);
303 if (set_thread_cpu_affinity(crdp
))
307 * If callbacks take a read-side lock, we need to be registered.
309 rcu_register_thread();
311 URCU_TLS(thread_call_rcu_data
) = crdp
;
313 uatomic_dec(&crdp
->futex
);
314 /* Decrement futex before reading call_rcu list */
318 struct cds_wfcq_head cbs_tmp_head
;
319 struct cds_wfcq_tail cbs_tmp_tail
;
320 struct cds_wfcq_node
*cbs
, *cbs_tmp_n
;
321 enum cds_wfcq_ret splice_ret
;
323 if (set_thread_cpu_affinity(crdp
))
326 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSE
) {
328 * Pause requested. Become quiescent: remove
329 * ourself from all global lists, and don't
330 * process any callback. The callback lists may
331 * still be non-empty though.
333 rcu_unregister_thread();
334 cmm_smp_mb__before_uatomic_or();
335 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_PAUSED
);
336 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSE
) != 0)
337 (void) poll(NULL
, 0, 1);
338 uatomic_and(&crdp
->flags
, ~URCU_CALL_RCU_PAUSED
);
339 cmm_smp_mb__after_uatomic_and();
340 rcu_register_thread();
343 cds_wfcq_init(&cbs_tmp_head
, &cbs_tmp_tail
);
344 splice_ret
= __cds_wfcq_splice_blocking(&cbs_tmp_head
,
345 &cbs_tmp_tail
, &crdp
->cbs_head
, &crdp
->cbs_tail
);
346 assert(splice_ret
!= CDS_WFCQ_RET_WOULDBLOCK
);
347 assert(splice_ret
!= CDS_WFCQ_RET_DEST_NON_EMPTY
);
348 if (splice_ret
!= CDS_WFCQ_RET_SRC_EMPTY
) {
351 __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head
,
352 &cbs_tmp_tail
, cbs
, cbs_tmp_n
) {
353 struct rcu_head
*rhp
;
355 rhp
= caa_container_of(cbs
,
356 struct rcu_head
, next
);
360 uatomic_sub(&crdp
->qlen
, cbcount
);
362 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOP
)
364 rcu_thread_offline();
366 if (cds_wfcq_empty(&crdp
->cbs_head
,
369 (void) poll(NULL
, 0, 10);
370 uatomic_dec(&crdp
->futex
);
372 * Decrement futex before reading
377 (void) poll(NULL
, 0, 10);
380 (void) poll(NULL
, 0, 10);
386 * Read call_rcu list before write futex.
389 uatomic_set(&crdp
->futex
, 0);
391 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
392 rcu_unregister_thread();
397 * Create both a call_rcu thread and the corresponding call_rcu_data
398 * structure, linking the structure in as specified. Caller must hold
402 static void call_rcu_data_init(struct call_rcu_data
**crdpp
,
406 struct call_rcu_data
*crdp
;
409 crdp
= malloc(sizeof(*crdp
));
412 memset(crdp
, '\0', sizeof(*crdp
));
413 cds_wfcq_init(&crdp
->cbs_head
, &crdp
->cbs_tail
);
417 cds_list_add(&crdp
->list
, &call_rcu_data_list
);
418 crdp
->cpu_affinity
= cpu_affinity
;
420 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
422 ret
= pthread_create(&crdp
->tid
, NULL
, call_rcu_thread
, crdp
);
428 * Return a pointer to the call_rcu_data structure for the specified
429 * CPU, returning NULL if there is none. We cannot automatically
430 * created it because the platform we are running on might not define
431 * urcu_sched_getcpu().
433 * The call to this function and use of the returned call_rcu_data
434 * should be protected by RCU read-side lock.
437 struct call_rcu_data
*get_cpu_call_rcu_data(int cpu
)
439 static int warned
= 0;
440 struct call_rcu_data
**pcpu_crdp
;
442 pcpu_crdp
= rcu_dereference(per_cpu_call_rcu_data
);
443 if (pcpu_crdp
== NULL
)
445 if (!warned
&& maxcpus
> 0 && (cpu
< 0 || maxcpus
<= cpu
)) {
446 fprintf(stderr
, "[error] liburcu: get CPU # out of range\n");
449 if (cpu
< 0 || maxcpus
<= cpu
)
451 return rcu_dereference(pcpu_crdp
[cpu
]);
455 * Return the tid corresponding to the call_rcu thread whose
456 * call_rcu_data structure is specified.
459 pthread_t
get_call_rcu_thread(struct call_rcu_data
*crdp
)
465 * Create a call_rcu_data structure (with thread) and return a pointer.
468 static struct call_rcu_data
*__create_call_rcu_data(unsigned long flags
,
471 struct call_rcu_data
*crdp
;
473 call_rcu_data_init(&crdp
, flags
, cpu_affinity
);
477 struct call_rcu_data
*create_call_rcu_data(unsigned long flags
,
480 struct call_rcu_data
*crdp
;
482 call_rcu_lock(&call_rcu_mutex
);
483 crdp
= __create_call_rcu_data(flags
, cpu_affinity
);
484 call_rcu_unlock(&call_rcu_mutex
);
489 * Set the specified CPU to use the specified call_rcu_data structure.
491 * Use NULL to remove a CPU's call_rcu_data structure, but it is
492 * the caller's responsibility to dispose of the removed structure.
493 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
494 * (prior to NULLing it out, of course).
496 * The caller must wait for a grace-period to pass between return from
497 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
498 * previous call rcu data as argument.
501 int set_cpu_call_rcu_data(int cpu
, struct call_rcu_data
*crdp
)
503 static int warned
= 0;
505 call_rcu_lock(&call_rcu_mutex
);
506 alloc_cpu_call_rcu_data();
507 if (cpu
< 0 || maxcpus
<= cpu
) {
509 fprintf(stderr
, "[error] liburcu: set CPU # out of range\n");
512 call_rcu_unlock(&call_rcu_mutex
);
517 if (per_cpu_call_rcu_data
== NULL
) {
518 call_rcu_unlock(&call_rcu_mutex
);
523 if (per_cpu_call_rcu_data
[cpu
] != NULL
&& crdp
!= NULL
) {
524 call_rcu_unlock(&call_rcu_mutex
);
529 rcu_set_pointer(&per_cpu_call_rcu_data
[cpu
], crdp
);
530 call_rcu_unlock(&call_rcu_mutex
);
535 * Return a pointer to the default call_rcu_data structure, creating
536 * one if need be. Because we never free call_rcu_data structures,
537 * we don't need to be in an RCU read-side critical section.
540 struct call_rcu_data
*get_default_call_rcu_data(void)
542 if (default_call_rcu_data
!= NULL
)
543 return rcu_dereference(default_call_rcu_data
);
544 call_rcu_lock(&call_rcu_mutex
);
545 if (default_call_rcu_data
!= NULL
) {
546 call_rcu_unlock(&call_rcu_mutex
);
547 return default_call_rcu_data
;
549 call_rcu_data_init(&default_call_rcu_data
, 0, -1);
550 call_rcu_unlock(&call_rcu_mutex
);
551 return default_call_rcu_data
;
555 * Return the call_rcu_data structure that applies to the currently
556 * running thread. Any call_rcu_data structure assigned specifically
557 * to this thread has first priority, followed by any call_rcu_data
558 * structure assigned to the CPU on which the thread is running,
559 * followed by the default call_rcu_data structure. If there is not
560 * yet a default call_rcu_data structure, one will be created.
562 * Calls to this function and use of the returned call_rcu_data should
563 * be protected by RCU read-side lock.
565 struct call_rcu_data
*get_call_rcu_data(void)
567 struct call_rcu_data
*crd
;
569 if (URCU_TLS(thread_call_rcu_data
) != NULL
)
570 return URCU_TLS(thread_call_rcu_data
);
573 crd
= get_cpu_call_rcu_data(urcu_sched_getcpu());
578 return get_default_call_rcu_data();
582 * Return a pointer to this task's call_rcu_data if there is one.
585 struct call_rcu_data
*get_thread_call_rcu_data(void)
587 return URCU_TLS(thread_call_rcu_data
);
591 * Set this task's call_rcu_data structure as specified, regardless
592 * of whether or not this task already had one. (This allows switching
593 * to and from real-time call_rcu threads, for example.)
595 * Use NULL to remove a thread's call_rcu_data structure, but it is
596 * the caller's responsibility to dispose of the removed structure.
597 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
598 * (prior to NULLing it out, of course).
601 void set_thread_call_rcu_data(struct call_rcu_data
*crdp
)
603 URCU_TLS(thread_call_rcu_data
) = crdp
;
607 * Create a separate call_rcu thread for each CPU. This does not
608 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
609 * function if you want that behavior. Should be paired with
610 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
614 int create_all_cpu_call_rcu_data(unsigned long flags
)
617 struct call_rcu_data
*crdp
;
620 call_rcu_lock(&call_rcu_mutex
);
621 alloc_cpu_call_rcu_data();
622 call_rcu_unlock(&call_rcu_mutex
);
627 if (per_cpu_call_rcu_data
== NULL
) {
631 for (i
= 0; i
< maxcpus
; i
++) {
632 call_rcu_lock(&call_rcu_mutex
);
633 if (get_cpu_call_rcu_data(i
)) {
634 call_rcu_unlock(&call_rcu_mutex
);
637 crdp
= __create_call_rcu_data(flags
, i
);
639 call_rcu_unlock(&call_rcu_mutex
);
643 call_rcu_unlock(&call_rcu_mutex
);
644 if ((ret
= set_cpu_call_rcu_data(i
, crdp
)) != 0) {
645 call_rcu_data_free(crdp
);
647 /* it has been created by other thread */
658 * Wake up the call_rcu thread corresponding to the specified
659 * call_rcu_data structure.
661 static void wake_call_rcu_thread(struct call_rcu_data
*crdp
)
663 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RT
))
664 call_rcu_wake_up(crdp
);
667 static void _call_rcu(struct rcu_head
*head
,
668 void (*func
)(struct rcu_head
*head
),
669 struct call_rcu_data
*crdp
)
671 cds_wfcq_node_init(&head
->next
);
673 cds_wfcq_enqueue(&crdp
->cbs_head
, &crdp
->cbs_tail
, &head
->next
);
674 uatomic_inc(&crdp
->qlen
);
675 wake_call_rcu_thread(crdp
);
679 * Schedule a function to be invoked after a following grace period.
680 * This is the only function that must be called -- the others are
681 * only present to allow applications to tune their use of RCU for
682 * maximum performance.
684 * Note that unless a call_rcu thread has not already been created,
685 * the first invocation of call_rcu() will create one. So, if you
686 * need the first invocation of call_rcu() to be fast, make sure
687 * to create a call_rcu thread first. One way to accomplish this is
688 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
690 * call_rcu must be called by registered RCU read-side threads.
692 void call_rcu(struct rcu_head
*head
,
693 void (*func
)(struct rcu_head
*head
))
695 struct call_rcu_data
*crdp
;
697 /* Holding rcu read-side lock across use of per-cpu crdp */
699 crdp
= get_call_rcu_data();
700 _call_rcu(head
, func
, crdp
);
705 * Free up the specified call_rcu_data structure, terminating the
706 * associated call_rcu thread. The caller must have previously
707 * removed the call_rcu_data structure from per-thread or per-CPU
708 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
709 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
710 * per-thread call_rcu_data structures.
712 * We silently refuse to free up the default call_rcu_data structure
713 * because that is where we put any leftover callbacks. Note that
714 * the possibility of self-spawning callbacks makes it impossible
715 * to execute all the callbacks in finite time without putting any
716 * newly spawned callbacks somewhere else. The "somewhere else" of
717 * last resort is the default call_rcu_data structure.
719 * We also silently refuse to free NULL pointers. This simplifies
722 * The caller must wait for a grace-period to pass between return from
723 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
724 * previous call rcu data as argument.
726 * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
727 * a list corruption bug in the 0.7.x series. The equivalent fix
728 * appeared in 0.6.8 for the stable-0.6 branch.
730 void call_rcu_data_free(struct call_rcu_data
*crdp
)
732 if (crdp
== NULL
|| crdp
== default_call_rcu_data
) {
735 if ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0) {
736 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOP
);
737 wake_call_rcu_thread(crdp
);
738 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0)
739 (void) poll(NULL
, 0, 1);
741 if (!cds_wfcq_empty(&crdp
->cbs_head
, &crdp
->cbs_tail
)) {
742 /* Create default call rcu data if need be */
743 (void) get_default_call_rcu_data();
744 __cds_wfcq_splice_blocking(&default_call_rcu_data
->cbs_head
,
745 &default_call_rcu_data
->cbs_tail
,
746 &crdp
->cbs_head
, &crdp
->cbs_tail
);
747 uatomic_add(&default_call_rcu_data
->qlen
,
748 uatomic_read(&crdp
->qlen
));
749 wake_call_rcu_thread(default_call_rcu_data
);
752 call_rcu_lock(&call_rcu_mutex
);
753 cds_list_del(&crdp
->list
);
754 call_rcu_unlock(&call_rcu_mutex
);
760 * Clean up all the per-CPU call_rcu threads.
762 void free_all_cpu_call_rcu_data(void)
765 struct call_rcu_data
**crdp
;
766 static int warned
= 0;
771 crdp
= malloc(sizeof(*crdp
) * maxcpus
);
774 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
780 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
781 crdp
[cpu
] = get_cpu_call_rcu_data(cpu
);
782 if (crdp
[cpu
] == NULL
)
784 set_cpu_call_rcu_data(cpu
, NULL
);
787 * Wait for call_rcu sites acting as RCU readers of the
788 * call_rcu_data to become quiescent.
791 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
792 if (crdp
[cpu
] == NULL
)
794 call_rcu_data_free(crdp
[cpu
]);
800 void free_completion(struct urcu_ref
*ref
)
802 struct call_rcu_completion
*completion
;
804 completion
= caa_container_of(ref
, struct call_rcu_completion
, ref
);
809 void _rcu_barrier_complete(struct rcu_head
*head
)
811 struct call_rcu_completion_work
*work
;
812 struct call_rcu_completion
*completion
;
814 work
= caa_container_of(head
, struct call_rcu_completion_work
, head
);
815 completion
= work
->completion
;
816 if (!uatomic_sub_return(&completion
->barrier_count
, 1))
817 call_rcu_completion_wake_up(completion
);
818 urcu_ref_put(&completion
->ref
, free_completion
);
823 * Wait for all in-flight call_rcu callbacks to complete execution.
825 void rcu_barrier(void)
827 struct call_rcu_data
*crdp
;
828 struct call_rcu_completion
*completion
;
832 /* Put in offline state in QSBR. */
833 was_online
= _rcu_read_ongoing();
835 rcu_thread_offline();
837 * Calling a rcu_barrier() within a RCU read-side critical
838 * section is an error.
840 if (_rcu_read_ongoing()) {
841 static int warned
= 0;
844 fprintf(stderr
, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
850 completion
= calloc(sizeof(*completion
), 1);
854 call_rcu_lock(&call_rcu_mutex
);
855 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
)
858 /* Referenced by rcu_barrier() and each call_rcu thread. */
859 urcu_ref_set(&completion
->ref
, count
+ 1);
860 completion
->barrier_count
= count
;
862 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
863 struct call_rcu_completion_work
*work
;
865 work
= calloc(sizeof(*work
), 1);
868 work
->completion
= completion
;
869 _call_rcu(&work
->head
, _rcu_barrier_complete
, crdp
);
871 call_rcu_unlock(&call_rcu_mutex
);
875 uatomic_dec(&completion
->futex
);
876 /* Decrement futex before reading barrier_count */
878 if (!uatomic_read(&completion
->barrier_count
))
880 call_rcu_completion_wait(completion
);
883 urcu_ref_put(&completion
->ref
, free_completion
);
891 * Acquire the call_rcu_mutex in order to ensure that the child sees
892 * all of the call_rcu() data structures in a consistent state. Ensure
893 * that all call_rcu threads are in a quiescent state across fork.
894 * Suitable for pthread_atfork() and friends.
896 void call_rcu_before_fork(void)
898 struct call_rcu_data
*crdp
;
900 call_rcu_lock(&call_rcu_mutex
);
902 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
903 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_PAUSE
);
904 cmm_smp_mb__after_uatomic_or();
905 wake_call_rcu_thread(crdp
);
907 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
908 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSED
) == 0)
909 (void) poll(NULL
, 0, 1);
914 * Clean up call_rcu data structures in the parent of a successful fork()
915 * that is not followed by exec() in the child. Suitable for
916 * pthread_atfork() and friends.
918 void call_rcu_after_fork_parent(void)
920 struct call_rcu_data
*crdp
;
922 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
)
923 uatomic_and(&crdp
->flags
, ~URCU_CALL_RCU_PAUSE
);
924 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
925 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSED
) != 0)
926 (void) poll(NULL
, 0, 1);
928 call_rcu_unlock(&call_rcu_mutex
);
932 * Clean up call_rcu data structures in the child of a successful fork()
933 * that is not followed by exec(). Suitable for pthread_atfork() and
936 void call_rcu_after_fork_child(void)
938 struct call_rcu_data
*crdp
, *next
;
940 /* Release the mutex. */
941 call_rcu_unlock(&call_rcu_mutex
);
943 /* Do nothing when call_rcu() has not been used */
944 if (cds_list_empty(&call_rcu_data_list
))
948 * Allocate a new default call_rcu_data structure in order
949 * to get a working call_rcu thread to go with it.
951 default_call_rcu_data
= NULL
;
952 (void)get_default_call_rcu_data();
954 /* Cleanup call_rcu_data pointers before use */
956 free(per_cpu_call_rcu_data
);
957 rcu_set_pointer(&per_cpu_call_rcu_data
, NULL
);
958 URCU_TLS(thread_call_rcu_data
) = NULL
;
961 * Dispose of all of the rest of the call_rcu_data structures.
962 * Leftover call_rcu callbacks will be merged into the new
963 * default call_rcu thread queue.
965 cds_list_for_each_entry_safe(crdp
, next
, &call_rcu_data_list
, list
) {
966 if (crdp
== default_call_rcu_data
)
968 uatomic_set(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
969 call_rcu_data_free(crdp
);