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 * List of all call_rcu_data structures to keep valgrind happy.
86 * Protected by call_rcu_mutex.
89 static CDS_LIST_HEAD(call_rcu_data_list
);
91 /* Link a thread using call_rcu() to its call_rcu thread. */
93 static DEFINE_URCU_TLS(struct call_rcu_data
*, thread_call_rcu_data
);
96 * Guard call_rcu thread creation and atfork handlers.
98 static pthread_mutex_t call_rcu_mutex
= PTHREAD_MUTEX_INITIALIZER
;
100 /* If a given thread does not have its own call_rcu thread, this is default. */
102 static struct call_rcu_data
*default_call_rcu_data
;
104 static struct urcu_atfork
*registered_rculfhash_atfork
;
105 static unsigned long registered_rculfhash_atfork_refcount
;
108 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
109 * available, then we can have call_rcu threads assigned to individual
110 * CPUs rather than only to specific threads.
113 #if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID))
116 * Pointer to array of pointers to per-CPU call_rcu_data structures
117 * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
118 * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
119 * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
120 * without mutex. The call_rcu_mutex protects updates.
123 static struct call_rcu_data
**per_cpu_call_rcu_data
;
124 static long cpus_array_len
;
126 static void cpus_array_len_reset(void)
131 /* Allocate the array if it has not already been allocated. */
133 static void alloc_cpu_call_rcu_data(void)
135 struct call_rcu_data
**p
;
136 static int warned
= 0;
138 if (cpus_array_len
!= 0)
140 cpus_array_len
= get_possible_cpus_array_len();
141 if (cpus_array_len
<= 0) {
144 p
= malloc(cpus_array_len
* sizeof(*per_cpu_call_rcu_data
));
146 memset(p
, '\0', cpus_array_len
* sizeof(*per_cpu_call_rcu_data
));
147 rcu_set_pointer(&per_cpu_call_rcu_data
, p
);
150 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
156 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
159 * per_cpu_call_rcu_data should be constant, but some functions below, used both
160 * for cases where cpu number is available and not available, assume it it not
163 static struct call_rcu_data
**per_cpu_call_rcu_data
= NULL
;
164 static const long cpus_array_len
= -1;
166 static void cpus_array_len_reset(void)
170 static void alloc_cpu_call_rcu_data(void)
174 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
176 /* Acquire the specified pthread mutex. */
178 static void call_rcu_lock(pthread_mutex_t
*pmp
)
182 ret
= pthread_mutex_lock(pmp
);
187 /* Release the specified pthread mutex. */
189 static void call_rcu_unlock(pthread_mutex_t
*pmp
)
193 ret
= pthread_mutex_unlock(pmp
);
199 * Periodically retry setting CPU affinity if we migrate.
200 * Losing affinity can be caused by CPU hotunplug/hotplug, or by
203 #ifdef HAVE_SCHED_SETAFFINITY
205 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
210 if (crdp
->cpu_affinity
< 0)
212 if (++crdp
->gp_count
& SET_AFFINITY_CHECK_PERIOD_MASK
)
214 if (urcu_sched_getcpu() == crdp
->cpu_affinity
)
218 CPU_SET(crdp
->cpu_affinity
, &mask
);
219 ret
= sched_setaffinity(0, sizeof(mask
), &mask
);
222 * EINVAL is fine: can be caused by hotunplugged CPUs, or by
223 * cpuset(7). This is why we should always retry if we detect
226 if (ret
&& errno
== EINVAL
) {
234 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
__attribute__((unused
)))
240 static void call_rcu_wait(struct call_rcu_data
*crdp
)
242 /* Read call_rcu list before read futex */
244 while (uatomic_read(&crdp
->futex
) == -1) {
245 if (!futex_async(&crdp
->futex
, FUTEX_WAIT
, -1, NULL
, NULL
, 0)) {
247 * Prior queued wakeups queued by unrelated code
248 * using the same address can cause futex wait to
249 * return 0 even through the futex value is still
250 * -1 (spurious wakeups). Check the value again
251 * in user-space to validate whether it really
258 /* Value already changed. */
261 /* Retry if interrupted by signal. */
262 break; /* Get out of switch. Check again. */
264 /* Unexpected error. */
270 static void call_rcu_wake_up(struct call_rcu_data
*crdp
)
272 /* Write to call_rcu list before reading/writing futex */
274 if (caa_unlikely(uatomic_read(&crdp
->futex
) == -1)) {
275 uatomic_set(&crdp
->futex
, 0);
276 if (futex_async(&crdp
->futex
, FUTEX_WAKE
, 1,
282 static void call_rcu_completion_wait(struct call_rcu_completion
*completion
)
284 /* Read completion barrier count before read futex */
286 while (uatomic_read(&completion
->futex
) == -1) {
287 if (!futex_async(&completion
->futex
, FUTEX_WAIT
, -1, NULL
, NULL
, 0)) {
289 * Prior queued wakeups queued by unrelated code
290 * using the same address can cause futex wait to
291 * return 0 even through the futex value is still
292 * -1 (spurious wakeups). Check the value again
293 * in user-space to validate whether it really
300 /* Value already changed. */
303 /* Retry if interrupted by signal. */
304 break; /* Get out of switch. Check again. */
306 /* Unexpected error. */
312 static void call_rcu_completion_wake_up(struct call_rcu_completion
*completion
)
314 /* Write to completion barrier count before reading/writing futex */
316 if (caa_unlikely(uatomic_read(&completion
->futex
) == -1)) {
317 uatomic_set(&completion
->futex
, 0);
318 if (futex_async(&completion
->futex
, FUTEX_WAKE
, 1,
324 /* This is the code run by each call_rcu thread. */
326 static void *call_rcu_thread(void *arg
)
328 unsigned long cbcount
;
329 struct call_rcu_data
*crdp
= (struct call_rcu_data
*) arg
;
330 int rt
= !!(uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_RT
);
332 if (set_thread_cpu_affinity(crdp
))
336 * If callbacks take a read-side lock, we need to be registered.
338 rcu_register_thread();
340 URCU_TLS(thread_call_rcu_data
) = crdp
;
342 uatomic_dec(&crdp
->futex
);
343 /* Decrement futex before reading call_rcu list */
347 struct cds_wfcq_head cbs_tmp_head
;
348 struct cds_wfcq_tail cbs_tmp_tail
;
349 struct cds_wfcq_node
*cbs
, *cbs_tmp_n
;
350 enum cds_wfcq_ret splice_ret
;
352 if (set_thread_cpu_affinity(crdp
))
355 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSE
) {
357 * Pause requested. Become quiescent: remove
358 * ourself from all global lists, and don't
359 * process any callback. The callback lists may
360 * still be non-empty though.
362 rcu_unregister_thread();
363 cmm_smp_mb__before_uatomic_or();
364 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_PAUSED
);
365 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSE
) != 0)
366 (void) poll(NULL
, 0, 1);
367 uatomic_and(&crdp
->flags
, ~URCU_CALL_RCU_PAUSED
);
368 cmm_smp_mb__after_uatomic_and();
369 rcu_register_thread();
372 cds_wfcq_init(&cbs_tmp_head
, &cbs_tmp_tail
);
373 splice_ret
= __cds_wfcq_splice_blocking(&cbs_tmp_head
,
374 &cbs_tmp_tail
, &crdp
->cbs_head
, &crdp
->cbs_tail
);
375 urcu_posix_assert(splice_ret
!= CDS_WFCQ_RET_WOULDBLOCK
);
376 urcu_posix_assert(splice_ret
!= CDS_WFCQ_RET_DEST_NON_EMPTY
);
377 if (splice_ret
!= CDS_WFCQ_RET_SRC_EMPTY
) {
380 __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head
,
381 &cbs_tmp_tail
, cbs
, cbs_tmp_n
) {
382 struct rcu_head
*rhp
;
384 rhp
= caa_container_of(cbs
,
385 struct rcu_head
, next
);
389 uatomic_sub(&crdp
->qlen
, cbcount
);
391 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOP
)
393 rcu_thread_offline();
395 if (cds_wfcq_empty(&crdp
->cbs_head
,
398 (void) poll(NULL
, 0, 10);
399 uatomic_dec(&crdp
->futex
);
401 * Decrement futex before reading
406 (void) poll(NULL
, 0, 10);
409 (void) poll(NULL
, 0, 10);
415 * Read call_rcu list before write futex.
418 uatomic_set(&crdp
->futex
, 0);
420 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
421 rcu_unregister_thread();
426 * Create both a call_rcu thread and the corresponding call_rcu_data
427 * structure, linking the structure in as specified. Caller must hold
431 static void call_rcu_data_init(struct call_rcu_data
**crdpp
,
435 struct call_rcu_data
*crdp
;
437 sigset_t newmask
, oldmask
;
439 crdp
= malloc(sizeof(*crdp
));
442 memset(crdp
, '\0', sizeof(*crdp
));
443 cds_wfcq_init(&crdp
->cbs_head
, &crdp
->cbs_tail
);
447 cds_list_add(&crdp
->list
, &call_rcu_data_list
);
448 crdp
->cpu_affinity
= cpu_affinity
;
450 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
453 ret
= sigfillset(&newmask
);
454 urcu_posix_assert(!ret
);
455 ret
= pthread_sigmask(SIG_BLOCK
, &newmask
, &oldmask
);
456 urcu_posix_assert(!ret
);
458 ret
= pthread_create(&crdp
->tid
, NULL
, call_rcu_thread
, crdp
);
462 ret
= pthread_sigmask(SIG_SETMASK
, &oldmask
, NULL
);
463 urcu_posix_assert(!ret
);
467 * Return a pointer to the call_rcu_data structure for the specified
468 * CPU, returning NULL if there is none. We cannot automatically
469 * created it because the platform we are running on might not define
470 * urcu_sched_getcpu().
472 * The call to this function and use of the returned call_rcu_data
473 * should be protected by RCU read-side lock.
476 struct call_rcu_data
*get_cpu_call_rcu_data(int cpu
)
478 static int warned
= 0;
479 struct call_rcu_data
**pcpu_crdp
;
481 pcpu_crdp
= rcu_dereference(per_cpu_call_rcu_data
);
482 if (pcpu_crdp
== NULL
)
484 if (!warned
&& cpus_array_len
> 0 && (cpu
< 0 || cpus_array_len
<= cpu
)) {
485 fprintf(stderr
, "[error] liburcu: get CPU # out of range\n");
488 if (cpu
< 0 || cpus_array_len
<= cpu
)
490 return rcu_dereference(pcpu_crdp
[cpu
]);
494 * Return the tid corresponding to the call_rcu thread whose
495 * call_rcu_data structure is specified.
498 pthread_t
get_call_rcu_thread(struct call_rcu_data
*crdp
)
504 * Create a call_rcu_data structure (with thread) and return a pointer.
507 static struct call_rcu_data
*__create_call_rcu_data(unsigned long flags
,
510 struct call_rcu_data
*crdp
;
512 call_rcu_data_init(&crdp
, flags
, cpu_affinity
);
516 struct call_rcu_data
*create_call_rcu_data(unsigned long flags
,
519 struct call_rcu_data
*crdp
;
521 call_rcu_lock(&call_rcu_mutex
);
522 crdp
= __create_call_rcu_data(flags
, cpu_affinity
);
523 call_rcu_unlock(&call_rcu_mutex
);
528 * Set the specified CPU to use the specified call_rcu_data structure.
530 * Use NULL to remove a CPU's call_rcu_data structure, but it is
531 * the caller's responsibility to dispose of the removed structure.
532 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
533 * (prior to NULLing it out, of course).
535 * The caller must wait for a grace-period to pass between return from
536 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
537 * previous call rcu data as argument.
540 int set_cpu_call_rcu_data(int cpu
, struct call_rcu_data
*crdp
)
542 static int warned
= 0;
544 call_rcu_lock(&call_rcu_mutex
);
545 alloc_cpu_call_rcu_data();
546 if (cpu
< 0 || cpus_array_len
<= cpu
) {
548 fprintf(stderr
, "[error] liburcu: set CPU # out of range\n");
551 call_rcu_unlock(&call_rcu_mutex
);
556 if (per_cpu_call_rcu_data
== NULL
) {
557 call_rcu_unlock(&call_rcu_mutex
);
562 if (per_cpu_call_rcu_data
[cpu
] != NULL
&& crdp
!= NULL
) {
563 call_rcu_unlock(&call_rcu_mutex
);
568 rcu_set_pointer(&per_cpu_call_rcu_data
[cpu
], crdp
);
569 call_rcu_unlock(&call_rcu_mutex
);
574 * Return a pointer to the default call_rcu_data structure, creating
575 * one if need be. Because we never free call_rcu_data structures,
576 * we don't need to be in an RCU read-side critical section.
579 struct call_rcu_data
*get_default_call_rcu_data(void)
581 if (default_call_rcu_data
!= NULL
)
582 return rcu_dereference(default_call_rcu_data
);
583 call_rcu_lock(&call_rcu_mutex
);
584 if (default_call_rcu_data
!= NULL
) {
585 call_rcu_unlock(&call_rcu_mutex
);
586 return default_call_rcu_data
;
588 call_rcu_data_init(&default_call_rcu_data
, 0, -1);
589 call_rcu_unlock(&call_rcu_mutex
);
590 return default_call_rcu_data
;
594 * Return the call_rcu_data structure that applies to the currently
595 * running thread. Any call_rcu_data structure assigned specifically
596 * to this thread has first priority, followed by any call_rcu_data
597 * structure assigned to the CPU on which the thread is running,
598 * followed by the default call_rcu_data structure. If there is not
599 * yet a default call_rcu_data structure, one will be created.
601 * Calls to this function and use of the returned call_rcu_data should
602 * be protected by RCU read-side lock.
604 struct call_rcu_data
*get_call_rcu_data(void)
606 struct call_rcu_data
*crd
;
608 if (URCU_TLS(thread_call_rcu_data
) != NULL
)
609 return URCU_TLS(thread_call_rcu_data
);
611 if (cpus_array_len
> 0) {
612 crd
= get_cpu_call_rcu_data(urcu_sched_getcpu());
617 return get_default_call_rcu_data();
621 * Return a pointer to this task's call_rcu_data if there is one.
624 struct call_rcu_data
*get_thread_call_rcu_data(void)
626 return URCU_TLS(thread_call_rcu_data
);
630 * Set this task's call_rcu_data structure as specified, regardless
631 * of whether or not this task already had one. (This allows switching
632 * to and from real-time call_rcu threads, for example.)
634 * Use NULL to remove a thread's call_rcu_data structure, but it is
635 * the caller's responsibility to dispose of the removed structure.
636 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
637 * (prior to NULLing it out, of course).
640 void set_thread_call_rcu_data(struct call_rcu_data
*crdp
)
642 URCU_TLS(thread_call_rcu_data
) = crdp
;
646 * Create a separate call_rcu thread for each CPU. This does not
647 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
648 * function if you want that behavior. Should be paired with
649 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
653 int create_all_cpu_call_rcu_data(unsigned long flags
)
656 struct call_rcu_data
*crdp
;
659 call_rcu_lock(&call_rcu_mutex
);
660 alloc_cpu_call_rcu_data();
661 call_rcu_unlock(&call_rcu_mutex
);
662 if (cpus_array_len
<= 0) {
666 if (per_cpu_call_rcu_data
== NULL
) {
670 for (i
= 0; i
< cpus_array_len
; i
++) {
671 call_rcu_lock(&call_rcu_mutex
);
672 if (get_cpu_call_rcu_data(i
)) {
673 call_rcu_unlock(&call_rcu_mutex
);
676 crdp
= __create_call_rcu_data(flags
, i
);
678 call_rcu_unlock(&call_rcu_mutex
);
682 call_rcu_unlock(&call_rcu_mutex
);
683 if ((ret
= set_cpu_call_rcu_data(i
, crdp
)) != 0) {
684 call_rcu_data_free(crdp
);
686 /* it has been created by other thread */
697 * Wake up the call_rcu thread corresponding to the specified
698 * call_rcu_data structure.
700 static void wake_call_rcu_thread(struct call_rcu_data
*crdp
)
702 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RT
))
703 call_rcu_wake_up(crdp
);
706 static void _call_rcu(struct rcu_head
*head
,
707 void (*func
)(struct rcu_head
*head
),
708 struct call_rcu_data
*crdp
)
710 cds_wfcq_node_init(&head
->next
);
712 cds_wfcq_enqueue(&crdp
->cbs_head
, &crdp
->cbs_tail
, &head
->next
);
713 uatomic_inc(&crdp
->qlen
);
714 wake_call_rcu_thread(crdp
);
718 * Schedule a function to be invoked after a following grace period.
719 * This is the only function that must be called -- the others are
720 * only present to allow applications to tune their use of RCU for
721 * maximum performance.
723 * Note that unless a call_rcu thread has not already been created,
724 * the first invocation of call_rcu() will create one. So, if you
725 * need the first invocation of call_rcu() to be fast, make sure
726 * to create a call_rcu thread first. One way to accomplish this is
727 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
729 * call_rcu must be called by registered RCU read-side threads.
731 void call_rcu(struct rcu_head
*head
,
732 void (*func
)(struct rcu_head
*head
))
734 struct call_rcu_data
*crdp
;
736 /* Holding rcu read-side lock across use of per-cpu crdp */
738 crdp
= get_call_rcu_data();
739 _call_rcu(head
, func
, crdp
);
744 * Free up the specified call_rcu_data structure, terminating the
745 * associated call_rcu thread. The caller must have previously
746 * removed the call_rcu_data structure from per-thread or per-CPU
747 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
748 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
749 * per-thread call_rcu_data structures.
751 * We silently refuse to free up the default call_rcu_data structure
752 * because that is where we put any leftover callbacks. Note that
753 * the possibility of self-spawning callbacks makes it impossible
754 * to execute all the callbacks in finite time without putting any
755 * newly spawned callbacks somewhere else. The "somewhere else" of
756 * last resort is the default call_rcu_data structure.
758 * We also silently refuse to free NULL pointers. This simplifies
761 * The caller must wait for a grace-period to pass between return from
762 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
763 * previous call rcu data as argument.
765 * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
766 * a list corruption bug in the 0.7.x series. The equivalent fix
767 * appeared in 0.6.8 for the stable-0.6 branch.
769 void call_rcu_data_free(struct call_rcu_data
*crdp
)
771 if (crdp
== NULL
|| crdp
== default_call_rcu_data
) {
774 if ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0) {
775 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOP
);
776 wake_call_rcu_thread(crdp
);
777 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0)
778 (void) poll(NULL
, 0, 1);
780 call_rcu_lock(&call_rcu_mutex
);
781 if (!cds_wfcq_empty(&crdp
->cbs_head
, &crdp
->cbs_tail
)) {
782 call_rcu_unlock(&call_rcu_mutex
);
783 /* Create default call rcu data if need be. */
784 /* CBs queued here will be handed to the default list. */
785 (void) get_default_call_rcu_data();
786 call_rcu_lock(&call_rcu_mutex
);
787 __cds_wfcq_splice_blocking(&default_call_rcu_data
->cbs_head
,
788 &default_call_rcu_data
->cbs_tail
,
789 &crdp
->cbs_head
, &crdp
->cbs_tail
);
790 uatomic_add(&default_call_rcu_data
->qlen
,
791 uatomic_read(&crdp
->qlen
));
792 wake_call_rcu_thread(default_call_rcu_data
);
795 cds_list_del(&crdp
->list
);
796 call_rcu_unlock(&call_rcu_mutex
);
802 * Clean up all the per-CPU call_rcu threads.
804 void free_all_cpu_call_rcu_data(void)
807 struct call_rcu_data
**crdp
;
808 static int warned
= 0;
810 if (cpus_array_len
<= 0)
813 crdp
= malloc(sizeof(*crdp
) * cpus_array_len
);
816 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
822 for (cpu
= 0; cpu
< cpus_array_len
; cpu
++) {
823 crdp
[cpu
] = get_cpu_call_rcu_data(cpu
);
824 if (crdp
[cpu
] == NULL
)
826 set_cpu_call_rcu_data(cpu
, NULL
);
829 * Wait for call_rcu sites acting as RCU readers of the
830 * call_rcu_data to become quiescent.
833 for (cpu
= 0; cpu
< cpus_array_len
; cpu
++) {
834 if (crdp
[cpu
] == NULL
)
836 call_rcu_data_free(crdp
[cpu
]);
842 void free_completion(struct urcu_ref
*ref
)
844 struct call_rcu_completion
*completion
;
846 completion
= caa_container_of(ref
, struct call_rcu_completion
, ref
);
851 void _rcu_barrier_complete(struct rcu_head
*head
)
853 struct call_rcu_completion_work
*work
;
854 struct call_rcu_completion
*completion
;
856 work
= caa_container_of(head
, struct call_rcu_completion_work
, head
);
857 completion
= work
->completion
;
858 if (!uatomic_sub_return(&completion
->barrier_count
, 1))
859 call_rcu_completion_wake_up(completion
);
860 urcu_ref_put(&completion
->ref
, free_completion
);
865 * Wait for all in-flight call_rcu callbacks to complete execution.
867 void rcu_barrier(void)
869 struct call_rcu_data
*crdp
;
870 struct call_rcu_completion
*completion
;
874 /* Put in offline state in QSBR. */
875 was_online
= _rcu_read_ongoing();
877 rcu_thread_offline();
879 * Calling a rcu_barrier() within a RCU read-side critical
880 * section is an error.
882 if (_rcu_read_ongoing()) {
883 static int warned
= 0;
886 fprintf(stderr
, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
892 completion
= calloc(sizeof(*completion
), 1);
896 call_rcu_lock(&call_rcu_mutex
);
897 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
)
900 /* Referenced by rcu_barrier() and each call_rcu thread. */
901 urcu_ref_set(&completion
->ref
, count
+ 1);
902 completion
->barrier_count
= count
;
904 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
905 struct call_rcu_completion_work
*work
;
907 work
= calloc(sizeof(*work
), 1);
910 work
->completion
= completion
;
911 _call_rcu(&work
->head
, _rcu_barrier_complete
, crdp
);
913 call_rcu_unlock(&call_rcu_mutex
);
917 uatomic_dec(&completion
->futex
);
918 /* Decrement futex before reading barrier_count */
920 if (!uatomic_read(&completion
->barrier_count
))
922 call_rcu_completion_wait(completion
);
925 urcu_ref_put(&completion
->ref
, free_completion
);
933 * Acquire the call_rcu_mutex in order to ensure that the child sees
934 * all of the call_rcu() data structures in a consistent state. Ensure
935 * that all call_rcu threads are in a quiescent state across fork.
936 * Suitable for pthread_atfork() and friends.
938 void call_rcu_before_fork(void)
940 struct call_rcu_data
*crdp
;
941 struct urcu_atfork
*atfork
;
943 call_rcu_lock(&call_rcu_mutex
);
945 atfork
= registered_rculfhash_atfork
;
947 atfork
->before_fork(atfork
->priv
);
949 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
950 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_PAUSE
);
951 cmm_smp_mb__after_uatomic_or();
952 wake_call_rcu_thread(crdp
);
954 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
955 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSED
) == 0)
956 (void) poll(NULL
, 0, 1);
961 * Clean up call_rcu data structures in the parent of a successful fork()
962 * that is not followed by exec() in the child. Suitable for
963 * pthread_atfork() and friends.
965 void call_rcu_after_fork_parent(void)
967 struct call_rcu_data
*crdp
;
968 struct urcu_atfork
*atfork
;
970 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
)
971 uatomic_and(&crdp
->flags
, ~URCU_CALL_RCU_PAUSE
);
972 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
973 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSED
) != 0)
974 (void) poll(NULL
, 0, 1);
976 atfork
= registered_rculfhash_atfork
;
978 atfork
->after_fork_parent(atfork
->priv
);
979 call_rcu_unlock(&call_rcu_mutex
);
983 * Clean up call_rcu data structures in the child of a successful fork()
984 * that is not followed by exec(). Suitable for pthread_atfork() and
987 void call_rcu_after_fork_child(void)
989 struct call_rcu_data
*crdp
, *next
;
990 struct urcu_atfork
*atfork
;
992 /* Release the mutex. */
993 call_rcu_unlock(&call_rcu_mutex
);
995 atfork
= registered_rculfhash_atfork
;
997 atfork
->after_fork_child(atfork
->priv
);
999 /* Do nothing when call_rcu() has not been used */
1000 if (cds_list_empty(&call_rcu_data_list
))
1004 * Allocate a new default call_rcu_data structure in order
1005 * to get a working call_rcu thread to go with it.
1007 default_call_rcu_data
= NULL
;
1008 (void)get_default_call_rcu_data();
1010 /* Cleanup call_rcu_data pointers before use */
1011 cpus_array_len_reset();
1012 free(per_cpu_call_rcu_data
);
1013 rcu_set_pointer(&per_cpu_call_rcu_data
, NULL
);
1014 URCU_TLS(thread_call_rcu_data
) = NULL
;
1017 * Dispose of all of the rest of the call_rcu_data structures.
1018 * Leftover call_rcu callbacks will be merged into the new
1019 * default call_rcu thread queue.
1021 cds_list_for_each_entry_safe(crdp
, next
, &call_rcu_data_list
, list
) {
1022 if (crdp
== default_call_rcu_data
)
1024 uatomic_set(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
1025 call_rcu_data_free(crdp
);
1029 void urcu_register_rculfhash_atfork(struct urcu_atfork
*atfork
)
1031 call_rcu_lock(&call_rcu_mutex
);
1032 if (registered_rculfhash_atfork_refcount
++)
1034 registered_rculfhash_atfork
= atfork
;
1036 call_rcu_unlock(&call_rcu_mutex
);
1039 void urcu_unregister_rculfhash_atfork(struct urcu_atfork
*atfork
__attribute__((unused
)))
1041 call_rcu_lock(&call_rcu_mutex
);
1042 if (--registered_rculfhash_atfork_refcount
)
1044 registered_rculfhash_atfork
= NULL
;
1046 call_rcu_unlock(&call_rcu_mutex
);