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
38 #include "urcu/wfqueue.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 /* Data structure that identifies a call_rcu thread. */
48 struct call_rcu_data
{
49 struct cds_wfq_queue cbs
;
52 unsigned long qlen
; /* maintained for debugging. */
55 struct cds_list_head list
;
56 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
59 * List of all call_rcu_data structures to keep valgrind happy.
60 * Protected by call_rcu_mutex.
63 CDS_LIST_HEAD(call_rcu_data_list
);
65 /* Link a thread using call_rcu() to its call_rcu thread. */
67 static DEFINE_URCU_TLS(struct call_rcu_data
*, thread_call_rcu_data
);
70 * Guard call_rcu thread creation and atfork handlers.
72 static pthread_mutex_t call_rcu_mutex
= PTHREAD_MUTEX_INITIALIZER
;
74 /* If a given thread does not have its own call_rcu thread, this is default. */
76 static struct call_rcu_data
*default_call_rcu_data
;
79 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
80 * available, then we can have call_rcu threads assigned to individual
81 * CPUs rather than only to specific threads.
84 #ifdef HAVE_SCHED_GETCPU
86 static int urcu_sched_getcpu(void)
88 return sched_getcpu();
91 #else /* #ifdef HAVE_SCHED_GETCPU */
93 static int urcu_sched_getcpu(void)
98 #endif /* #else #ifdef HAVE_SCHED_GETCPU */
100 #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU)
103 * Pointer to array of pointers to per-CPU call_rcu_data structures
104 * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
105 * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
106 * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
107 * without mutex. The call_rcu_mutex protects updates.
110 static struct call_rcu_data
**per_cpu_call_rcu_data
;
113 static void maxcpus_reset(void)
118 /* Allocate the array if it has not already been allocated. */
120 static void alloc_cpu_call_rcu_data(void)
122 struct call_rcu_data
**p
;
123 static int warned
= 0;
127 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
131 p
= malloc(maxcpus
* sizeof(*per_cpu_call_rcu_data
));
133 memset(p
, '\0', maxcpus
* sizeof(*per_cpu_call_rcu_data
));
134 rcu_set_pointer(&per_cpu_call_rcu_data
, p
);
137 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
143 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
146 * per_cpu_call_rcu_data should be constant, but some functions below, used both
147 * for cases where cpu number is available and not available, assume it it not
150 static struct call_rcu_data
**per_cpu_call_rcu_data
= NULL
;
151 static const long maxcpus
= -1;
153 static void maxcpus_reset(void)
157 static void alloc_cpu_call_rcu_data(void)
161 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
163 /* Acquire the specified pthread mutex. */
165 static void call_rcu_lock(pthread_mutex_t
*pmp
)
169 ret
= pthread_mutex_lock(pmp
);
174 /* Release the specified pthread mutex. */
176 static void call_rcu_unlock(pthread_mutex_t
*pmp
)
180 ret
= pthread_mutex_unlock(pmp
);
185 #if HAVE_SCHED_SETAFFINITY
187 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
191 if (crdp
->cpu_affinity
< 0)
195 CPU_SET(crdp
->cpu_affinity
, &mask
);
196 #if SCHED_SETAFFINITY_ARGS == 2
197 return sched_setaffinity(0, &mask
);
199 return sched_setaffinity(0, sizeof(mask
), &mask
);
204 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
210 static void call_rcu_wait(struct call_rcu_data
*crdp
)
212 /* Read call_rcu list before read futex */
214 if (uatomic_read(&crdp
->futex
) == -1)
215 futex_async(&crdp
->futex
, FUTEX_WAIT
, -1,
219 static void call_rcu_wake_up(struct call_rcu_data
*crdp
)
221 /* Write to call_rcu list before reading/writing futex */
223 if (caa_unlikely(uatomic_read(&crdp
->futex
) == -1)) {
224 uatomic_set(&crdp
->futex
, 0);
225 futex_async(&crdp
->futex
, FUTEX_WAKE
, 1,
230 /* This is the code run by each call_rcu thread. */
232 static void *call_rcu_thread(void *arg
)
234 unsigned long cbcount
;
235 struct cds_wfq_node
*cbs
;
236 struct cds_wfq_node
**cbs_tail
;
237 struct call_rcu_data
*crdp
= (struct call_rcu_data
*)arg
;
238 struct rcu_head
*rhp
;
239 int rt
= !!(uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_RT
);
242 ret
= set_thread_cpu_affinity(crdp
);
247 * If callbacks take a read-side lock, we need to be registered.
249 rcu_register_thread();
251 URCU_TLS(thread_call_rcu_data
) = crdp
;
253 uatomic_dec(&crdp
->futex
);
254 /* Decrement futex before reading call_rcu list */
258 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSE
) {
260 * Pause requested. Become quiescent: remove
261 * ourself from all global lists, and don't
262 * process any callback. The callback lists may
263 * still be non-empty though.
265 rcu_unregister_thread();
266 cmm_smp_mb__before_uatomic_or();
267 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_PAUSED
);
268 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSE
) != 0)
270 uatomic_and(&crdp
->flags
, ~URCU_CALL_RCU_PAUSED
);
271 cmm_smp_mb__after_uatomic_and();
272 rcu_register_thread();
275 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
276 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
278 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
279 cbs_tail
= (struct cds_wfq_node
**)
280 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
284 while (cbs
->next
== NULL
&&
285 &cbs
->next
!= cbs_tail
)
287 if (cbs
== &crdp
->cbs
.dummy
) {
291 rhp
= (struct rcu_head
*)cbs
;
295 } while (cbs
!= NULL
);
296 uatomic_sub(&crdp
->qlen
, cbcount
);
298 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOP
)
300 rcu_thread_offline();
303 == _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
306 uatomic_dec(&crdp
->futex
);
308 * Decrement futex before reading
322 * Read call_rcu list before write futex.
325 uatomic_set(&crdp
->futex
, 0);
327 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
328 rcu_unregister_thread();
333 * Create both a call_rcu thread and the corresponding call_rcu_data
334 * structure, linking the structure in as specified. Caller must hold
338 static void call_rcu_data_init(struct call_rcu_data
**crdpp
,
342 struct call_rcu_data
*crdp
;
345 crdp
= malloc(sizeof(*crdp
));
348 memset(crdp
, '\0', sizeof(*crdp
));
349 cds_wfq_init(&crdp
->cbs
);
353 cds_list_add(&crdp
->list
, &call_rcu_data_list
);
354 crdp
->cpu_affinity
= cpu_affinity
;
355 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
357 ret
= pthread_create(&crdp
->tid
, NULL
, call_rcu_thread
, crdp
);
363 * Return a pointer to the call_rcu_data structure for the specified
364 * CPU, returning NULL if there is none. We cannot automatically
365 * created it because the platform we are running on might not define
366 * urcu_sched_getcpu().
368 * The call to this function and use of the returned call_rcu_data
369 * should be protected by RCU read-side lock.
372 struct call_rcu_data
*get_cpu_call_rcu_data(int cpu
)
374 static int warned
= 0;
375 struct call_rcu_data
**pcpu_crdp
;
377 pcpu_crdp
= rcu_dereference(per_cpu_call_rcu_data
);
378 if (pcpu_crdp
== NULL
)
380 if (!warned
&& maxcpus
> 0 && (cpu
< 0 || maxcpus
<= cpu
)) {
381 fprintf(stderr
, "[error] liburcu: get CPU # out of range\n");
384 if (cpu
< 0 || maxcpus
<= cpu
)
386 return rcu_dereference(pcpu_crdp
[cpu
]);
390 * Return the tid corresponding to the call_rcu thread whose
391 * call_rcu_data structure is specified.
394 pthread_t
get_call_rcu_thread(struct call_rcu_data
*crdp
)
400 * Create a call_rcu_data structure (with thread) and return a pointer.
403 static struct call_rcu_data
*__create_call_rcu_data(unsigned long flags
,
406 struct call_rcu_data
*crdp
;
408 call_rcu_data_init(&crdp
, flags
, cpu_affinity
);
412 struct call_rcu_data
*create_call_rcu_data(unsigned long flags
,
415 struct call_rcu_data
*crdp
;
417 call_rcu_lock(&call_rcu_mutex
);
418 crdp
= __create_call_rcu_data(flags
, cpu_affinity
);
419 call_rcu_unlock(&call_rcu_mutex
);
424 * Set the specified CPU to use the specified call_rcu_data structure.
426 * Use NULL to remove a CPU's call_rcu_data structure, but it is
427 * the caller's responsibility to dispose of the removed structure.
428 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
429 * (prior to NULLing it out, of course).
431 * The caller must wait for a grace-period to pass between return from
432 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
433 * previous call rcu data as argument.
436 int set_cpu_call_rcu_data(int cpu
, struct call_rcu_data
*crdp
)
438 static int warned
= 0;
440 call_rcu_lock(&call_rcu_mutex
);
441 alloc_cpu_call_rcu_data();
442 if (cpu
< 0 || maxcpus
<= cpu
) {
444 fprintf(stderr
, "[error] liburcu: set CPU # out of range\n");
447 call_rcu_unlock(&call_rcu_mutex
);
452 if (per_cpu_call_rcu_data
== NULL
) {
453 call_rcu_unlock(&call_rcu_mutex
);
458 if (per_cpu_call_rcu_data
[cpu
] != NULL
&& crdp
!= NULL
) {
459 call_rcu_unlock(&call_rcu_mutex
);
464 rcu_set_pointer(&per_cpu_call_rcu_data
[cpu
], crdp
);
465 call_rcu_unlock(&call_rcu_mutex
);
470 * Return a pointer to the default call_rcu_data structure, creating
471 * one if need be. Because we never free call_rcu_data structures,
472 * we don't need to be in an RCU read-side critical section.
475 struct call_rcu_data
*get_default_call_rcu_data(void)
477 if (default_call_rcu_data
!= NULL
)
478 return rcu_dereference(default_call_rcu_data
);
479 call_rcu_lock(&call_rcu_mutex
);
480 if (default_call_rcu_data
!= NULL
) {
481 call_rcu_unlock(&call_rcu_mutex
);
482 return default_call_rcu_data
;
484 call_rcu_data_init(&default_call_rcu_data
, 0, -1);
485 call_rcu_unlock(&call_rcu_mutex
);
486 return default_call_rcu_data
;
490 * Return the call_rcu_data structure that applies to the currently
491 * running thread. Any call_rcu_data structure assigned specifically
492 * to this thread has first priority, followed by any call_rcu_data
493 * structure assigned to the CPU on which the thread is running,
494 * followed by the default call_rcu_data structure. If there is not
495 * yet a default call_rcu_data structure, one will be created.
497 * Calls to this function and use of the returned call_rcu_data should
498 * be protected by RCU read-side lock.
500 struct call_rcu_data
*get_call_rcu_data(void)
502 struct call_rcu_data
*crd
;
504 if (URCU_TLS(thread_call_rcu_data
) != NULL
)
505 return URCU_TLS(thread_call_rcu_data
);
508 crd
= get_cpu_call_rcu_data(urcu_sched_getcpu());
513 return get_default_call_rcu_data();
517 * Return a pointer to this task's call_rcu_data if there is one.
520 struct call_rcu_data
*get_thread_call_rcu_data(void)
522 return URCU_TLS(thread_call_rcu_data
);
526 * Set this task's call_rcu_data structure as specified, regardless
527 * of whether or not this task already had one. (This allows switching
528 * to and from real-time call_rcu threads, for example.)
530 * Use NULL to remove a thread's call_rcu_data structure, but it is
531 * the caller's responsibility to dispose of the removed structure.
532 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
533 * (prior to NULLing it out, of course).
536 void set_thread_call_rcu_data(struct call_rcu_data
*crdp
)
538 URCU_TLS(thread_call_rcu_data
) = crdp
;
542 * Create a separate call_rcu thread for each CPU. This does not
543 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
544 * function if you want that behavior. Should be paired with
545 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
549 int create_all_cpu_call_rcu_data(unsigned long flags
)
552 struct call_rcu_data
*crdp
;
555 call_rcu_lock(&call_rcu_mutex
);
556 alloc_cpu_call_rcu_data();
557 call_rcu_unlock(&call_rcu_mutex
);
562 if (per_cpu_call_rcu_data
== NULL
) {
566 for (i
= 0; i
< maxcpus
; i
++) {
567 call_rcu_lock(&call_rcu_mutex
);
568 if (get_cpu_call_rcu_data(i
)) {
569 call_rcu_unlock(&call_rcu_mutex
);
572 crdp
= __create_call_rcu_data(flags
, i
);
574 call_rcu_unlock(&call_rcu_mutex
);
578 call_rcu_unlock(&call_rcu_mutex
);
579 if ((ret
= set_cpu_call_rcu_data(i
, crdp
)) != 0) {
580 call_rcu_data_free(crdp
);
582 /* it has been created by other thread */
593 * Wake up the call_rcu thread corresponding to the specified
594 * call_rcu_data structure.
596 static void wake_call_rcu_thread(struct call_rcu_data
*crdp
)
598 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RT
))
599 call_rcu_wake_up(crdp
);
603 * Schedule a function to be invoked after a following grace period.
604 * This is the only function that must be called -- the others are
605 * only present to allow applications to tune their use of RCU for
606 * maximum performance.
608 * Note that unless a call_rcu thread has not already been created,
609 * the first invocation of call_rcu() will create one. So, if you
610 * need the first invocation of call_rcu() to be fast, make sure
611 * to create a call_rcu thread first. One way to accomplish this is
612 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
614 * call_rcu must be called by registered RCU read-side threads.
617 void call_rcu(struct rcu_head
*head
,
618 void (*func
)(struct rcu_head
*head
))
620 struct call_rcu_data
*crdp
;
622 cds_wfq_node_init(&head
->next
);
624 /* Holding rcu read-side lock across use of per-cpu crdp */
626 crdp
= get_call_rcu_data();
627 cds_wfq_enqueue(&crdp
->cbs
, &head
->next
);
628 uatomic_inc(&crdp
->qlen
);
629 wake_call_rcu_thread(crdp
);
634 * Free up the specified call_rcu_data structure, terminating the
635 * associated call_rcu thread. The caller must have previously
636 * removed the call_rcu_data structure from per-thread or per-CPU
637 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
638 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
639 * per-thread call_rcu_data structures.
641 * We silently refuse to free up the default call_rcu_data structure
642 * because that is where we put any leftover callbacks. Note that
643 * the possibility of self-spawning callbacks makes it impossible
644 * to execute all the callbacks in finite time without putting any
645 * newly spawned callbacks somewhere else. The "somewhere else" of
646 * last resort is the default call_rcu_data structure.
648 * We also silently refuse to free NULL pointers. This simplifies
651 * The caller must wait for a grace-period to pass between return from
652 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
653 * previous call rcu data as argument.
655 void call_rcu_data_free(struct call_rcu_data
*crdp
)
657 struct cds_wfq_node
*cbs
;
658 struct cds_wfq_node
**cbs_tail
;
659 struct cds_wfq_node
**cbs_endprev
;
661 if (crdp
== NULL
|| crdp
== default_call_rcu_data
) {
664 if ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0) {
665 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOP
);
666 wake_call_rcu_thread(crdp
);
667 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0)
670 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
671 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
673 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
674 cbs_tail
= (struct cds_wfq_node
**)
675 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
676 /* Create default call rcu data if need be */
677 (void) get_default_call_rcu_data();
678 cbs_endprev
= (struct cds_wfq_node
**)
679 uatomic_xchg(&default_call_rcu_data
->cbs
.tail
,
681 _CMM_STORE_SHARED(*cbs_endprev
, cbs
);
682 uatomic_add(&default_call_rcu_data
->qlen
,
683 uatomic_read(&crdp
->qlen
));
684 wake_call_rcu_thread(default_call_rcu_data
);
687 call_rcu_lock(&call_rcu_mutex
);
688 cds_list_del(&crdp
->list
);
689 call_rcu_unlock(&call_rcu_mutex
);
695 * Clean up all the per-CPU call_rcu threads.
697 void free_all_cpu_call_rcu_data(void)
700 struct call_rcu_data
**crdp
;
701 static int warned
= 0;
706 crdp
= malloc(sizeof(*crdp
) * maxcpus
);
709 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
715 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
716 crdp
[cpu
] = get_cpu_call_rcu_data(cpu
);
717 if (crdp
[cpu
] == NULL
)
719 set_cpu_call_rcu_data(cpu
, NULL
);
722 * Wait for call_rcu sites acting as RCU readers of the
723 * call_rcu_data to become quiescent.
726 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
727 if (crdp
[cpu
] == NULL
)
729 call_rcu_data_free(crdp
[cpu
]);
735 * Acquire the call_rcu_mutex in order to ensure that the child sees
736 * all of the call_rcu() data structures in a consistent state. Ensure
737 * that all call_rcu threads are in a quiescent state across fork.
738 * Suitable for pthread_atfork() and friends.
740 void call_rcu_before_fork(void)
742 struct call_rcu_data
*crdp
;
744 call_rcu_lock(&call_rcu_mutex
);
746 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
747 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_PAUSE
);
748 cmm_smp_mb__after_uatomic_or();
749 wake_call_rcu_thread(crdp
);
751 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
752 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSED
) == 0)
758 * Clean up call_rcu data structures in the parent of a successful fork()
759 * that is not followed by exec() in the child. Suitable for
760 * pthread_atfork() and friends.
762 void call_rcu_after_fork_parent(void)
764 struct call_rcu_data
*crdp
;
766 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
)
767 uatomic_and(&crdp
->flags
, ~URCU_CALL_RCU_PAUSE
);
768 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
769 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSED
) != 0)
772 call_rcu_unlock(&call_rcu_mutex
);
776 * Clean up call_rcu data structures in the child of a successful fork()
777 * that is not followed by exec(). Suitable for pthread_atfork() and
780 void call_rcu_after_fork_child(void)
782 struct call_rcu_data
*crdp
, *next
;
784 /* Release the mutex. */
785 call_rcu_unlock(&call_rcu_mutex
);
787 /* Do nothing when call_rcu() has not been used */
788 if (cds_list_empty(&call_rcu_data_list
))
792 * Allocate a new default call_rcu_data structure in order
793 * to get a working call_rcu thread to go with it.
795 default_call_rcu_data
= NULL
;
796 (void)get_default_call_rcu_data();
798 /* Cleanup call_rcu_data pointers before use */
800 free(per_cpu_call_rcu_data
);
801 rcu_set_pointer(&per_cpu_call_rcu_data
, NULL
);
802 URCU_TLS(thread_call_rcu_data
) = NULL
;
805 * Dispose of all of the rest of the call_rcu_data structures.
806 * Leftover call_rcu callbacks will be merged into the new
807 * default call_rcu thread queue.
809 cds_list_for_each_entry_safe(crdp
, next
, &call_rcu_data_list
, list
) {
810 if (crdp
== default_call_rcu_data
)
812 uatomic_set(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
813 call_rcu_data_free(crdp
);