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
36 #include "compat-getcpu.h"
37 #include <urcu/assert.h>
38 #include <urcu/wfcqueue.h>
39 #include <urcu/call-rcu.h>
40 #include <urcu/pointer.h>
41 #include <urcu/list.h>
42 #include <urcu/futex.h>
43 #include <urcu/tls-compat.h>
46 #include "urcu-utils.h"
47 #include "compat-smp.h"
49 #define SET_AFFINITY_CHECK_PERIOD (1U << 8) /* 256 */
50 #define SET_AFFINITY_CHECK_PERIOD_MASK (SET_AFFINITY_CHECK_PERIOD - 1)
52 /* Data structure that identifies a call_rcu thread. */
54 struct call_rcu_data
{
56 * We do not align head on a different cache-line than tail
57 * mainly because call_rcu callback-invocation threads use
58 * batching ("splice") to get an entire list of callbacks, which
59 * effectively empties the queue, and requires to touch the tail
62 struct cds_wfcq_tail cbs_tail
;
63 struct cds_wfcq_head cbs_head
;
66 unsigned long qlen
; /* maintained for debugging. */
69 unsigned long gp_count
;
70 struct cds_list_head list
;
71 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
73 struct call_rcu_completion
{
79 struct call_rcu_completion_work
{
81 struct call_rcu_completion
*completion
;
85 CRDF_FLAG_JOIN_THREAD
= (1 << 0),
89 * List of all call_rcu_data structures to keep valgrind happy.
90 * Protected by call_rcu_mutex.
93 static CDS_LIST_HEAD(call_rcu_data_list
);
95 /* Link a thread using call_rcu() to its call_rcu thread. */
97 static DEFINE_URCU_TLS(struct call_rcu_data
*, thread_call_rcu_data
);
100 * Guard call_rcu thread creation and atfork handlers.
102 static pthread_mutex_t call_rcu_mutex
= PTHREAD_MUTEX_INITIALIZER
;
104 /* If a given thread does not have its own call_rcu thread, this is default. */
106 static struct call_rcu_data
*default_call_rcu_data
;
108 static struct urcu_atfork
*registered_rculfhash_atfork
;
111 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
112 * available, then we can have call_rcu threads assigned to individual
113 * CPUs rather than only to specific threads.
116 #if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID))
119 * Pointer to array of pointers to per-CPU call_rcu_data structures
120 * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
121 * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
122 * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
123 * without mutex. The call_rcu_mutex protects updates.
126 static struct call_rcu_data
**per_cpu_call_rcu_data
;
127 static long cpus_array_len
;
129 static void cpus_array_len_reset(void)
134 /* Allocate the array if it has not already been allocated. */
136 static void alloc_cpu_call_rcu_data(void)
138 struct call_rcu_data
**p
;
139 static int warned
= 0;
141 if (cpus_array_len
!= 0)
143 cpus_array_len
= get_possible_cpus_array_len();
144 if (cpus_array_len
<= 0) {
147 p
= malloc(cpus_array_len
* sizeof(*per_cpu_call_rcu_data
));
149 memset(p
, '\0', cpus_array_len
* sizeof(*per_cpu_call_rcu_data
));
150 rcu_set_pointer(&per_cpu_call_rcu_data
, p
);
153 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
159 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
162 * per_cpu_call_rcu_data should be constant, but some functions below, used both
163 * for cases where cpu number is available and not available, assume it it not
166 static struct call_rcu_data
**per_cpu_call_rcu_data
= NULL
;
167 static const long cpus_array_len
= -1;
169 static void cpus_array_len_reset(void)
173 static void alloc_cpu_call_rcu_data(void)
177 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
179 /* Acquire the specified pthread mutex. */
181 static void call_rcu_lock(pthread_mutex_t
*pmp
)
185 ret
= pthread_mutex_lock(pmp
);
190 /* Release the specified pthread mutex. */
192 static void call_rcu_unlock(pthread_mutex_t
*pmp
)
196 ret
= pthread_mutex_unlock(pmp
);
202 * Periodically retry setting CPU affinity if we migrate.
203 * Losing affinity can be caused by CPU hotunplug/hotplug, or by
206 #ifdef HAVE_SCHED_SETAFFINITY
208 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
213 if (crdp
->cpu_affinity
< 0)
215 if (++crdp
->gp_count
& SET_AFFINITY_CHECK_PERIOD_MASK
)
217 if (urcu_sched_getcpu() == crdp
->cpu_affinity
)
221 CPU_SET(crdp
->cpu_affinity
, &mask
);
222 ret
= sched_setaffinity(0, sizeof(mask
), &mask
);
225 * EINVAL is fine: can be caused by hotunplugged CPUs, or by
226 * cpuset(7). This is why we should always retry if we detect
229 if (ret
&& errno
== EINVAL
) {
237 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
__attribute__((unused
)))
243 static void call_rcu_wait(struct call_rcu_data
*crdp
)
245 /* Read call_rcu list before read futex */
247 while (uatomic_read(&crdp
->futex
) == -1) {
248 if (!futex_async(&crdp
->futex
, FUTEX_WAIT
, -1, NULL
, NULL
, 0)) {
250 * Prior queued wakeups queued by unrelated code
251 * using the same address can cause futex wait to
252 * return 0 even through the futex value is still
253 * -1 (spurious wakeups). Check the value again
254 * in user-space to validate whether it really
261 /* Value already changed. */
264 /* Retry if interrupted by signal. */
265 break; /* Get out of switch. Check again. */
267 /* Unexpected error. */
273 static void call_rcu_wake_up(struct call_rcu_data
*crdp
)
275 /* Write to call_rcu list before reading/writing futex */
277 if (caa_unlikely(uatomic_read(&crdp
->futex
) == -1)) {
278 uatomic_set(&crdp
->futex
, 0);
279 if (futex_async(&crdp
->futex
, FUTEX_WAKE
, 1,
285 static void call_rcu_completion_wait(struct call_rcu_completion
*completion
)
287 /* Read completion barrier count before read futex */
289 while (uatomic_read(&completion
->futex
) == -1) {
290 if (!futex_async(&completion
->futex
, FUTEX_WAIT
, -1, NULL
, NULL
, 0)) {
292 * Prior queued wakeups queued by unrelated code
293 * using the same address can cause futex wait to
294 * return 0 even through the futex value is still
295 * -1 (spurious wakeups). Check the value again
296 * in user-space to validate whether it really
303 /* Value already changed. */
306 /* Retry if interrupted by signal. */
307 break; /* Get out of switch. Check again. */
309 /* Unexpected error. */
315 static void call_rcu_completion_wake_up(struct call_rcu_completion
*completion
)
317 /* Write to completion barrier count before reading/writing futex */
319 if (caa_unlikely(uatomic_read(&completion
->futex
) == -1)) {
320 uatomic_set(&completion
->futex
, 0);
321 if (futex_async(&completion
->futex
, FUTEX_WAKE
, 1,
327 /* This is the code run by each call_rcu thread. */
329 static void *call_rcu_thread(void *arg
)
331 unsigned long cbcount
;
332 struct call_rcu_data
*crdp
= (struct call_rcu_data
*) arg
;
333 int rt
= !!(uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_RT
);
335 if (set_thread_cpu_affinity(crdp
))
339 * If callbacks take a read-side lock, we need to be registered.
341 rcu_register_thread();
343 URCU_TLS(thread_call_rcu_data
) = crdp
;
345 uatomic_dec(&crdp
->futex
);
346 /* Decrement futex before reading call_rcu list */
350 struct cds_wfcq_head cbs_tmp_head
;
351 struct cds_wfcq_tail cbs_tmp_tail
;
352 struct cds_wfcq_node
*cbs
, *cbs_tmp_n
;
353 enum cds_wfcq_ret splice_ret
;
355 if (set_thread_cpu_affinity(crdp
))
358 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSE
) {
360 * Pause requested. Become quiescent: remove
361 * ourself from all global lists, and don't
362 * process any callback. The callback lists may
363 * still be non-empty though.
365 rcu_unregister_thread();
366 cmm_smp_mb__before_uatomic_or();
367 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_PAUSED
);
368 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSE
) != 0)
369 (void) poll(NULL
, 0, 1);
370 uatomic_and(&crdp
->flags
, ~URCU_CALL_RCU_PAUSED
);
371 cmm_smp_mb__after_uatomic_and();
372 rcu_register_thread();
375 cds_wfcq_init(&cbs_tmp_head
, &cbs_tmp_tail
);
376 splice_ret
= __cds_wfcq_splice_blocking(&cbs_tmp_head
,
377 &cbs_tmp_tail
, &crdp
->cbs_head
, &crdp
->cbs_tail
);
378 urcu_posix_assert(splice_ret
!= CDS_WFCQ_RET_WOULDBLOCK
);
379 urcu_posix_assert(splice_ret
!= CDS_WFCQ_RET_DEST_NON_EMPTY
);
380 if (splice_ret
!= CDS_WFCQ_RET_SRC_EMPTY
) {
383 __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head
,
384 &cbs_tmp_tail
, cbs
, cbs_tmp_n
) {
385 struct rcu_head
*rhp
;
387 rhp
= caa_container_of(cbs
,
388 struct rcu_head
, next
);
392 uatomic_sub(&crdp
->qlen
, cbcount
);
394 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOP
)
396 rcu_thread_offline();
398 if (cds_wfcq_empty(&crdp
->cbs_head
,
401 (void) poll(NULL
, 0, 10);
402 uatomic_dec(&crdp
->futex
);
404 * Decrement futex before reading
409 (void) poll(NULL
, 0, 10);
412 (void) poll(NULL
, 0, 10);
418 * Read call_rcu list before write futex.
421 uatomic_set(&crdp
->futex
, 0);
423 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
424 rcu_unregister_thread();
429 * Create both a call_rcu thread and the corresponding call_rcu_data
430 * structure, linking the structure in as specified. Caller must hold
434 static void call_rcu_data_init(struct call_rcu_data
**crdpp
,
438 struct call_rcu_data
*crdp
;
440 sigset_t newmask
, oldmask
;
442 crdp
= malloc(sizeof(*crdp
));
445 memset(crdp
, '\0', sizeof(*crdp
));
446 cds_wfcq_init(&crdp
->cbs_head
, &crdp
->cbs_tail
);
450 cds_list_add(&crdp
->list
, &call_rcu_data_list
);
451 crdp
->cpu_affinity
= cpu_affinity
;
453 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
456 ret
= sigfillset(&newmask
);
457 urcu_posix_assert(!ret
);
458 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
459 urcu_posix_assert(!ret
);
461 ret
= pthread_create(&crdp
->tid
, NULL
, call_rcu_thread
, crdp
);
465 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
466 urcu_posix_assert(!ret
);
470 * Return a pointer to the call_rcu_data structure for the specified
471 * CPU, returning NULL if there is none. We cannot automatically
472 * created it because the platform we are running on might not define
473 * urcu_sched_getcpu().
475 * The call to this function and use of the returned call_rcu_data
476 * should be protected by RCU read-side lock.
479 struct call_rcu_data
*get_cpu_call_rcu_data(int cpu
)
481 static int warned
= 0;
482 struct call_rcu_data
**pcpu_crdp
;
484 pcpu_crdp
= rcu_dereference(per_cpu_call_rcu_data
);
485 if (pcpu_crdp
== NULL
)
487 if (!warned
&& cpus_array_len
> 0 && (cpu
< 0 || cpus_array_len
<= cpu
)) {
488 fprintf(stderr
, "[error] liburcu: get CPU # out of range\n");
491 if (cpu
< 0 || cpus_array_len
<= cpu
)
493 return rcu_dereference(pcpu_crdp
[cpu
]);
497 * Return the tid corresponding to the call_rcu thread whose
498 * call_rcu_data structure is specified.
501 pthread_t
get_call_rcu_thread(struct call_rcu_data
*crdp
)
507 * Create a call_rcu_data structure (with thread) and return a pointer.
510 static struct call_rcu_data
*__create_call_rcu_data(unsigned long flags
,
513 struct call_rcu_data
*crdp
;
515 call_rcu_data_init(&crdp
, flags
, cpu_affinity
);
519 struct call_rcu_data
*create_call_rcu_data(unsigned long flags
,
522 struct call_rcu_data
*crdp
;
524 call_rcu_lock(&call_rcu_mutex
);
525 crdp
= __create_call_rcu_data(flags
, cpu_affinity
);
526 call_rcu_unlock(&call_rcu_mutex
);
531 * Set the specified CPU to use the specified call_rcu_data structure.
533 * Use NULL to remove a CPU's call_rcu_data structure, but it is
534 * the caller's responsibility to dispose of the removed structure.
535 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
536 * (prior to NULLing it out, of course).
538 * The caller must wait for a grace-period to pass between return from
539 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
540 * previous call rcu data as argument.
543 int set_cpu_call_rcu_data(int cpu
, struct call_rcu_data
*crdp
)
545 static int warned
= 0;
547 call_rcu_lock(&call_rcu_mutex
);
548 alloc_cpu_call_rcu_data();
549 if (cpu
< 0 || cpus_array_len
<= cpu
) {
551 fprintf(stderr
, "[error] liburcu: set CPU # out of range\n");
554 call_rcu_unlock(&call_rcu_mutex
);
559 if (per_cpu_call_rcu_data
== NULL
) {
560 call_rcu_unlock(&call_rcu_mutex
);
565 if (per_cpu_call_rcu_data
[cpu
] != NULL
&& crdp
!= NULL
) {
566 call_rcu_unlock(&call_rcu_mutex
);
571 rcu_set_pointer(&per_cpu_call_rcu_data
[cpu
], crdp
);
572 call_rcu_unlock(&call_rcu_mutex
);
577 * Return a pointer to the default call_rcu_data structure, creating
578 * one if need be. Because we never free call_rcu_data structures,
579 * we don't need to be in an RCU read-side critical section.
582 struct call_rcu_data
*get_default_call_rcu_data(void)
584 if (default_call_rcu_data
!= NULL
)
585 return rcu_dereference(default_call_rcu_data
);
586 call_rcu_lock(&call_rcu_mutex
);
587 if (default_call_rcu_data
!= NULL
) {
588 call_rcu_unlock(&call_rcu_mutex
);
589 return default_call_rcu_data
;
591 call_rcu_data_init(&default_call_rcu_data
, 0, -1);
592 call_rcu_unlock(&call_rcu_mutex
);
593 return default_call_rcu_data
;
597 * Return the call_rcu_data structure that applies to the currently
598 * running thread. Any call_rcu_data structure assigned specifically
599 * to this thread has first priority, followed by any call_rcu_data
600 * structure assigned to the CPU on which the thread is running,
601 * followed by the default call_rcu_data structure. If there is not
602 * yet a default call_rcu_data structure, one will be created.
604 * Calls to this function and use of the returned call_rcu_data should
605 * be protected by RCU read-side lock.
607 struct call_rcu_data
*get_call_rcu_data(void)
609 struct call_rcu_data
*crd
;
611 if (URCU_TLS(thread_call_rcu_data
) != NULL
)
612 return URCU_TLS(thread_call_rcu_data
);
614 if (cpus_array_len
> 0) {
615 crd
= get_cpu_call_rcu_data(urcu_sched_getcpu());
620 return get_default_call_rcu_data();
624 * Return a pointer to this task's call_rcu_data if there is one.
627 struct call_rcu_data
*get_thread_call_rcu_data(void)
629 return URCU_TLS(thread_call_rcu_data
);
633 * Set this task's call_rcu_data structure as specified, regardless
634 * of whether or not this task already had one. (This allows switching
635 * to and from real-time call_rcu threads, for example.)
637 * Use NULL to remove a thread's call_rcu_data structure, but it is
638 * the caller's responsibility to dispose of the removed structure.
639 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
640 * (prior to NULLing it out, of course).
643 void set_thread_call_rcu_data(struct call_rcu_data
*crdp
)
645 URCU_TLS(thread_call_rcu_data
) = crdp
;
649 * Create a separate call_rcu thread for each CPU. This does not
650 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
651 * function if you want that behavior. Should be paired with
652 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
656 int create_all_cpu_call_rcu_data(unsigned long flags
)
659 struct call_rcu_data
*crdp
;
662 call_rcu_lock(&call_rcu_mutex
);
663 alloc_cpu_call_rcu_data();
664 call_rcu_unlock(&call_rcu_mutex
);
665 if (cpus_array_len
<= 0) {
669 if (per_cpu_call_rcu_data
== NULL
) {
673 for (i
= 0; i
< cpus_array_len
; i
++) {
674 call_rcu_lock(&call_rcu_mutex
);
675 if (get_cpu_call_rcu_data(i
)) {
676 call_rcu_unlock(&call_rcu_mutex
);
679 crdp
= __create_call_rcu_data(flags
, i
);
681 call_rcu_unlock(&call_rcu_mutex
);
685 call_rcu_unlock(&call_rcu_mutex
);
686 if ((ret
= set_cpu_call_rcu_data(i
, crdp
)) != 0) {
687 call_rcu_data_free(crdp
);
689 /* it has been created by other thread */
700 * Wake up the call_rcu thread corresponding to the specified
701 * call_rcu_data structure.
703 static void wake_call_rcu_thread(struct call_rcu_data
*crdp
)
705 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RT
))
706 call_rcu_wake_up(crdp
);
709 static void _call_rcu(struct rcu_head
*head
,
710 void (*func
)(struct rcu_head
*head
),
711 struct call_rcu_data
*crdp
)
713 cds_wfcq_node_init(&head
->next
);
715 cds_wfcq_enqueue(&crdp
->cbs_head
, &crdp
->cbs_tail
, &head
->next
);
716 uatomic_inc(&crdp
->qlen
);
717 wake_call_rcu_thread(crdp
);
721 * Schedule a function to be invoked after a following grace period.
722 * This is the only function that must be called -- the others are
723 * only present to allow applications to tune their use of RCU for
724 * maximum performance.
726 * Note that unless a call_rcu thread has not already been created,
727 * the first invocation of call_rcu() will create one. So, if you
728 * need the first invocation of call_rcu() to be fast, make sure
729 * to create a call_rcu thread first. One way to accomplish this is
730 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
732 * call_rcu must be called by registered RCU read-side threads.
734 void call_rcu(struct rcu_head
*head
,
735 void (*func
)(struct rcu_head
*head
))
737 struct call_rcu_data
*crdp
;
739 /* Holding rcu read-side lock across use of per-cpu crdp */
741 crdp
= get_call_rcu_data();
742 _call_rcu(head
, func
, crdp
);
747 * Free up the specified call_rcu_data structure, terminating the
748 * associated call_rcu thread. The caller must have previously
749 * removed the call_rcu_data structure from per-thread or per-CPU
750 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
751 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
752 * per-thread call_rcu_data structures.
754 * We silently refuse to free up the default call_rcu_data structure
755 * because that is where we put any leftover callbacks. Note that
756 * the possibility of self-spawning callbacks makes it impossible
757 * to execute all the callbacks in finite time without putting any
758 * newly spawned callbacks somewhere else. The "somewhere else" of
759 * last resort is the default call_rcu_data structure.
761 * We also silently refuse to free NULL pointers. This simplifies
764 * The caller must wait for a grace-period to pass between return from
765 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
766 * previous call rcu data as argument.
768 * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
769 * a list corruption bug in the 0.7.x series. The equivalent fix
770 * appeared in 0.6.8 for the stable-0.6 branch.
773 void _call_rcu_data_free(struct call_rcu_data
*crdp
, unsigned int flags
)
775 if (crdp
== NULL
|| crdp
== default_call_rcu_data
) {
778 if ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0) {
779 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOP
);
780 wake_call_rcu_thread(crdp
);
781 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0)
782 (void) poll(NULL
, 0, 1);
784 call_rcu_lock(&call_rcu_mutex
);
785 if (!cds_wfcq_empty(&crdp
->cbs_head
, &crdp
->cbs_tail
)) {
786 call_rcu_unlock(&call_rcu_mutex
);
787 /* Create default call rcu data if need be. */
788 /* CBs queued here will be handed to the default list. */
789 (void) get_default_call_rcu_data();
790 call_rcu_lock(&call_rcu_mutex
);
791 __cds_wfcq_splice_blocking(&default_call_rcu_data
->cbs_head
,
792 &default_call_rcu_data
->cbs_tail
,
793 &crdp
->cbs_head
, &crdp
->cbs_tail
);
794 uatomic_add(&default_call_rcu_data
->qlen
,
795 uatomic_read(&crdp
->qlen
));
796 wake_call_rcu_thread(default_call_rcu_data
);
799 cds_list_del(&crdp
->list
);
800 call_rcu_unlock(&call_rcu_mutex
);
802 if (flags
& CRDF_FLAG_JOIN_THREAD
) {
805 ret
= pthread_join(get_call_rcu_thread(crdp
), NULL
);
812 void call_rcu_data_free(struct call_rcu_data
*crdp
)
814 _call_rcu_data_free(crdp
, CRDF_FLAG_JOIN_THREAD
);
818 * Clean up all the per-CPU call_rcu threads.
820 void free_all_cpu_call_rcu_data(void)
823 struct call_rcu_data
**crdp
;
824 static int warned
= 0;
826 if (cpus_array_len
<= 0)
829 crdp
= malloc(sizeof(*crdp
) * cpus_array_len
);
832 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
838 for (cpu
= 0; cpu
< cpus_array_len
; cpu
++) {
839 crdp
[cpu
] = get_cpu_call_rcu_data(cpu
);
840 if (crdp
[cpu
] == NULL
)
842 set_cpu_call_rcu_data(cpu
, NULL
);
845 * Wait for call_rcu sites acting as RCU readers of the
846 * call_rcu_data to become quiescent.
849 for (cpu
= 0; cpu
< cpus_array_len
; cpu
++) {
850 if (crdp
[cpu
] == NULL
)
852 call_rcu_data_free(crdp
[cpu
]);
858 void free_completion(struct urcu_ref
*ref
)
860 struct call_rcu_completion
*completion
;
862 completion
= caa_container_of(ref
, struct call_rcu_completion
, ref
);
867 void _rcu_barrier_complete(struct rcu_head
*head
)
869 struct call_rcu_completion_work
*work
;
870 struct call_rcu_completion
*completion
;
872 work
= caa_container_of(head
, struct call_rcu_completion_work
, head
);
873 completion
= work
->completion
;
874 if (!uatomic_sub_return(&completion
->barrier_count
, 1))
875 call_rcu_completion_wake_up(completion
);
876 urcu_ref_put(&completion
->ref
, free_completion
);
881 * Wait for all in-flight call_rcu callbacks to complete execution.
883 void rcu_barrier(void)
885 struct call_rcu_data
*crdp
;
886 struct call_rcu_completion
*completion
;
890 /* Put in offline state in QSBR. */
891 was_online
= _rcu_read_ongoing();
893 rcu_thread_offline();
895 * Calling a rcu_barrier() within a RCU read-side critical
896 * section is an error.
898 if (_rcu_read_ongoing()) {
899 static int warned
= 0;
902 fprintf(stderr
, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
908 completion
= calloc(sizeof(*completion
), 1);
912 call_rcu_lock(&call_rcu_mutex
);
913 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
)
916 /* Referenced by rcu_barrier() and each call_rcu thread. */
917 urcu_ref_set(&completion
->ref
, count
+ 1);
918 completion
->barrier_count
= count
;
920 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
921 struct call_rcu_completion_work
*work
;
923 work
= calloc(sizeof(*work
), 1);
926 work
->completion
= completion
;
927 _call_rcu(&work
->head
, _rcu_barrier_complete
, crdp
);
929 call_rcu_unlock(&call_rcu_mutex
);
933 uatomic_dec(&completion
->futex
);
934 /* Decrement futex before reading barrier_count */
936 if (!uatomic_read(&completion
->barrier_count
))
938 call_rcu_completion_wait(completion
);
941 urcu_ref_put(&completion
->ref
, free_completion
);
949 * Acquire the call_rcu_mutex in order to ensure that the child sees
950 * all of the call_rcu() data structures in a consistent state. Ensure
951 * that all call_rcu threads are in a quiescent state across fork.
952 * Suitable for pthread_atfork() and friends.
954 void call_rcu_before_fork(void)
956 struct call_rcu_data
*crdp
;
957 struct urcu_atfork
*atfork
;
959 call_rcu_lock(&call_rcu_mutex
);
961 atfork
= registered_rculfhash_atfork
;
963 atfork
->before_fork(atfork
->priv
);
965 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
966 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_PAUSE
);
967 cmm_smp_mb__after_uatomic_or();
968 wake_call_rcu_thread(crdp
);
970 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
971 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSED
) == 0)
972 (void) poll(NULL
, 0, 1);
977 * Clean up call_rcu data structures in the parent of a successful fork()
978 * that is not followed by exec() in the child. Suitable for
979 * pthread_atfork() and friends.
981 void call_rcu_after_fork_parent(void)
983 struct call_rcu_data
*crdp
;
984 struct urcu_atfork
*atfork
;
986 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
)
987 uatomic_and(&crdp
->flags
, ~URCU_CALL_RCU_PAUSE
);
988 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
989 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSED
) != 0)
990 (void) poll(NULL
, 0, 1);
992 atfork
= registered_rculfhash_atfork
;
994 atfork
->after_fork_parent(atfork
->priv
);
995 call_rcu_unlock(&call_rcu_mutex
);
999 * Clean up call_rcu data structures in the child of a successful fork()
1000 * that is not followed by exec(). Suitable for pthread_atfork() and
1003 void call_rcu_after_fork_child(void)
1005 struct call_rcu_data
*crdp
, *next
;
1006 struct urcu_atfork
*atfork
;
1008 /* Release the mutex. */
1009 call_rcu_unlock(&call_rcu_mutex
);
1011 atfork
= registered_rculfhash_atfork
;
1013 atfork
->after_fork_child(atfork
->priv
);
1015 /* Do nothing when call_rcu() has not been used */
1016 if (cds_list_empty(&call_rcu_data_list
))
1020 * Allocate a new default call_rcu_data structure in order
1021 * to get a working call_rcu thread to go with it.
1023 default_call_rcu_data
= NULL
;
1024 (void)get_default_call_rcu_data();
1026 /* Cleanup call_rcu_data pointers before use */
1027 cpus_array_len_reset();
1028 free(per_cpu_call_rcu_data
);
1029 rcu_set_pointer(&per_cpu_call_rcu_data
, NULL
);
1030 URCU_TLS(thread_call_rcu_data
) = NULL
;
1033 * Dispose of all of the rest of the call_rcu_data structures.
1034 * Leftover call_rcu callbacks will be merged into the new
1035 * default call_rcu thread queue.
1037 cds_list_for_each_entry_safe(crdp
, next
, &call_rcu_data_list
, list
) {
1038 if (crdp
== default_call_rcu_data
)
1040 uatomic_set(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
1042 * Do not join the thread because it does not exist in
1045 _call_rcu_data_free(crdp
, 0);
1049 void urcu_register_rculfhash_atfork(struct urcu_atfork
*atfork
)
1051 if (CMM_LOAD_SHARED(registered_rculfhash_atfork
))
1053 call_rcu_lock(&call_rcu_mutex
);
1054 if (!registered_rculfhash_atfork
)
1055 registered_rculfhash_atfork
= atfork
;
1056 call_rcu_unlock(&call_rcu_mutex
);
1060 * This unregistration function is deprecated, meant only for internal
1063 __attribute__((noreturn
))
1064 void urcu_unregister_rculfhash_atfork(struct urcu_atfork
*atfork
__attribute__((unused
)))