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"
47 /* Data structure that identifies a call_rcu thread. */
49 struct call_rcu_data
{
51 * We do not align head on a different cache-line than tail
52 * mainly because call_rcu callback-invocation threads use
53 * batching ("splice") to get an entire list of callbacks, which
54 * effectively empties the queue, and requires to touch the tail
57 struct cds_wfcq_tail cbs_tail
;
58 struct cds_wfcq_head cbs_head
;
61 unsigned long qlen
; /* maintained for debugging. */
64 struct cds_list_head list
;
65 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
68 * List of all call_rcu_data structures to keep valgrind happy.
69 * Protected by call_rcu_mutex.
72 static CDS_LIST_HEAD(call_rcu_data_list
);
74 /* Link a thread using call_rcu() to its call_rcu thread. */
76 static DEFINE_URCU_TLS(struct call_rcu_data
*, thread_call_rcu_data
);
79 * Guard call_rcu thread creation and atfork handlers.
81 static pthread_mutex_t call_rcu_mutex
= PTHREAD_MUTEX_INITIALIZER
;
83 /* If a given thread does not have its own call_rcu thread, this is default. */
85 static struct call_rcu_data
*default_call_rcu_data
;
88 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
89 * available, then we can have call_rcu threads assigned to individual
90 * CPUs rather than only to specific threads.
93 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
96 * Pointer to array of pointers to per-CPU call_rcu_data structures
97 * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
98 * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
99 * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
100 * without mutex. The call_rcu_mutex protects updates.
103 static struct call_rcu_data
**per_cpu_call_rcu_data
;
106 static void maxcpus_reset(void)
111 /* Allocate the array if it has not already been allocated. */
113 static void alloc_cpu_call_rcu_data(void)
115 struct call_rcu_data
**p
;
116 static int warned
= 0;
120 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
124 p
= malloc(maxcpus
* sizeof(*per_cpu_call_rcu_data
));
126 memset(p
, '\0', maxcpus
* sizeof(*per_cpu_call_rcu_data
));
127 rcu_set_pointer(&per_cpu_call_rcu_data
, p
);
130 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
136 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
139 * per_cpu_call_rcu_data should be constant, but some functions below, used both
140 * for cases where cpu number is available and not available, assume it it not
143 static struct call_rcu_data
**per_cpu_call_rcu_data
= NULL
;
144 static const long maxcpus
= -1;
146 static void maxcpus_reset(void)
150 static void alloc_cpu_call_rcu_data(void)
154 static int sched_getcpu(void)
159 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
161 /* Acquire the specified pthread mutex. */
163 static void call_rcu_lock(pthread_mutex_t
*pmp
)
167 ret
= pthread_mutex_lock(pmp
);
172 /* Release the specified pthread mutex. */
174 static void call_rcu_unlock(pthread_mutex_t
*pmp
)
178 ret
= pthread_mutex_unlock(pmp
);
183 #if HAVE_SCHED_SETAFFINITY
185 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
189 if (crdp
->cpu_affinity
< 0)
193 CPU_SET(crdp
->cpu_affinity
, &mask
);
194 #if SCHED_SETAFFINITY_ARGS == 2
195 return sched_setaffinity(0, &mask
);
197 return sched_setaffinity(0, sizeof(mask
), &mask
);
202 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
208 static void call_rcu_wait(struct call_rcu_data
*crdp
)
210 /* Read call_rcu list before read futex */
212 if (uatomic_read(&crdp
->futex
) == -1)
213 futex_async(&crdp
->futex
, FUTEX_WAIT
, -1,
217 static void call_rcu_wake_up(struct call_rcu_data
*crdp
)
219 /* Write to call_rcu list before reading/writing futex */
221 if (caa_unlikely(uatomic_read(&crdp
->futex
) == -1)) {
222 uatomic_set(&crdp
->futex
, 0);
223 futex_async(&crdp
->futex
, FUTEX_WAKE
, 1,
228 /* This is the code run by each call_rcu thread. */
230 static void *call_rcu_thread(void *arg
)
232 unsigned long cbcount
;
233 struct call_rcu_data
*crdp
= (struct call_rcu_data
*) arg
;
234 int rt
= !!(uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_RT
);
237 ret
= set_thread_cpu_affinity(crdp
);
242 * If callbacks take a read-side lock, we need to be registered.
244 rcu_register_thread();
246 URCU_TLS(thread_call_rcu_data
) = crdp
;
248 uatomic_dec(&crdp
->futex
);
249 /* Decrement futex before reading call_rcu list */
253 struct cds_wfcq_head cbs_tmp_head
;
254 struct cds_wfcq_tail cbs_tmp_tail
;
255 struct cds_wfcq_node
*cbs
, *cbs_tmp_n
;
256 enum cds_wfcq_ret splice_ret
;
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 rcu_register_thread();
273 cds_wfcq_init(&cbs_tmp_head
, &cbs_tmp_tail
);
274 splice_ret
= __cds_wfcq_splice_blocking(&cbs_tmp_head
,
275 &cbs_tmp_tail
, &crdp
->cbs_head
, &crdp
->cbs_tail
);
276 assert(splice_ret
!= CDS_WFCQ_RET_WOULDBLOCK
);
277 assert(splice_ret
!= CDS_WFCQ_RET_DEST_NON_EMPTY
);
278 if (splice_ret
!= CDS_WFCQ_RET_SRC_EMPTY
) {
281 __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head
,
282 &cbs_tmp_tail
, cbs
, cbs_tmp_n
) {
283 struct rcu_head
*rhp
;
285 rhp
= caa_container_of(cbs
,
286 struct rcu_head
, next
);
290 uatomic_sub(&crdp
->qlen
, cbcount
);
292 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOP
)
294 rcu_thread_offline();
296 if (cds_wfcq_empty(&crdp
->cbs_head
,
300 uatomic_dec(&crdp
->futex
);
302 * Decrement futex before reading
316 * Read call_rcu list before write futex.
319 uatomic_set(&crdp
->futex
, 0);
321 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
322 rcu_unregister_thread();
327 * Create both a call_rcu thread and the corresponding call_rcu_data
328 * structure, linking the structure in as specified. Caller must hold
332 static void call_rcu_data_init(struct call_rcu_data
**crdpp
,
336 struct call_rcu_data
*crdp
;
339 crdp
= malloc(sizeof(*crdp
));
342 memset(crdp
, '\0', sizeof(*crdp
));
343 cds_wfcq_init(&crdp
->cbs_head
, &crdp
->cbs_tail
);
347 cds_list_add(&crdp
->list
, &call_rcu_data_list
);
348 crdp
->cpu_affinity
= cpu_affinity
;
349 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
351 ret
= pthread_create(&crdp
->tid
, NULL
, call_rcu_thread
, crdp
);
357 * Return a pointer to the call_rcu_data structure for the specified
358 * CPU, returning NULL if there is none. We cannot automatically
359 * created it because the platform we are running on might not define
362 * The call to this function and use of the returned call_rcu_data
363 * should be protected by RCU read-side lock.
366 struct call_rcu_data
*get_cpu_call_rcu_data(int cpu
)
368 static int warned
= 0;
369 struct call_rcu_data
**pcpu_crdp
;
371 pcpu_crdp
= rcu_dereference(per_cpu_call_rcu_data
);
372 if (pcpu_crdp
== NULL
)
374 if (!warned
&& maxcpus
> 0 && (cpu
< 0 || maxcpus
<= cpu
)) {
375 fprintf(stderr
, "[error] liburcu: get CPU # out of range\n");
378 if (cpu
< 0 || maxcpus
<= cpu
)
380 return rcu_dereference(pcpu_crdp
[cpu
]);
384 * Return the tid corresponding to the call_rcu thread whose
385 * call_rcu_data structure is specified.
388 pthread_t
get_call_rcu_thread(struct call_rcu_data
*crdp
)
394 * Create a call_rcu_data structure (with thread) and return a pointer.
397 static struct call_rcu_data
*__create_call_rcu_data(unsigned long flags
,
400 struct call_rcu_data
*crdp
;
402 call_rcu_data_init(&crdp
, flags
, cpu_affinity
);
406 struct call_rcu_data
*create_call_rcu_data(unsigned long flags
,
409 struct call_rcu_data
*crdp
;
411 call_rcu_lock(&call_rcu_mutex
);
412 crdp
= __create_call_rcu_data(flags
, cpu_affinity
);
413 call_rcu_unlock(&call_rcu_mutex
);
418 * Set the specified CPU to use the specified call_rcu_data structure.
420 * Use NULL to remove a CPU's call_rcu_data structure, but it is
421 * the caller's responsibility to dispose of the removed structure.
422 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
423 * (prior to NULLing it out, of course).
425 * The caller must wait for a grace-period to pass between return from
426 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
427 * previous call rcu data as argument.
430 int set_cpu_call_rcu_data(int cpu
, struct call_rcu_data
*crdp
)
432 static int warned
= 0;
434 call_rcu_lock(&call_rcu_mutex
);
435 alloc_cpu_call_rcu_data();
436 if (cpu
< 0 || maxcpus
<= cpu
) {
438 fprintf(stderr
, "[error] liburcu: set CPU # out of range\n");
441 call_rcu_unlock(&call_rcu_mutex
);
446 if (per_cpu_call_rcu_data
== NULL
) {
447 call_rcu_unlock(&call_rcu_mutex
);
452 if (per_cpu_call_rcu_data
[cpu
] != NULL
&& crdp
!= NULL
) {
453 call_rcu_unlock(&call_rcu_mutex
);
458 rcu_set_pointer(&per_cpu_call_rcu_data
[cpu
], crdp
);
459 call_rcu_unlock(&call_rcu_mutex
);
464 * Return a pointer to the default call_rcu_data structure, creating
465 * one if need be. Because we never free call_rcu_data structures,
466 * we don't need to be in an RCU read-side critical section.
469 struct call_rcu_data
*get_default_call_rcu_data(void)
471 if (default_call_rcu_data
!= NULL
)
472 return rcu_dereference(default_call_rcu_data
);
473 call_rcu_lock(&call_rcu_mutex
);
474 if (default_call_rcu_data
!= NULL
) {
475 call_rcu_unlock(&call_rcu_mutex
);
476 return default_call_rcu_data
;
478 call_rcu_data_init(&default_call_rcu_data
, 0, -1);
479 call_rcu_unlock(&call_rcu_mutex
);
480 return default_call_rcu_data
;
484 * Return the call_rcu_data structure that applies to the currently
485 * running thread. Any call_rcu_data structure assigned specifically
486 * to this thread has first priority, followed by any call_rcu_data
487 * structure assigned to the CPU on which the thread is running,
488 * followed by the default call_rcu_data structure. If there is not
489 * yet a default call_rcu_data structure, one will be created.
491 * Calls to this function and use of the returned call_rcu_data should
492 * be protected by RCU read-side lock.
494 struct call_rcu_data
*get_call_rcu_data(void)
496 struct call_rcu_data
*crd
;
498 if (URCU_TLS(thread_call_rcu_data
) != NULL
)
499 return URCU_TLS(thread_call_rcu_data
);
502 crd
= get_cpu_call_rcu_data(sched_getcpu());
507 return get_default_call_rcu_data();
511 * Return a pointer to this task's call_rcu_data if there is one.
514 struct call_rcu_data
*get_thread_call_rcu_data(void)
516 return URCU_TLS(thread_call_rcu_data
);
520 * Set this task's call_rcu_data structure as specified, regardless
521 * of whether or not this task already had one. (This allows switching
522 * to and from real-time call_rcu threads, for example.)
524 * Use NULL to remove a thread's call_rcu_data structure, but it is
525 * the caller's responsibility to dispose of the removed structure.
526 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
527 * (prior to NULLing it out, of course).
530 void set_thread_call_rcu_data(struct call_rcu_data
*crdp
)
532 URCU_TLS(thread_call_rcu_data
) = crdp
;
536 * Create a separate call_rcu thread for each CPU. This does not
537 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
538 * function if you want that behavior. Should be paired with
539 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
543 int create_all_cpu_call_rcu_data(unsigned long flags
)
546 struct call_rcu_data
*crdp
;
549 call_rcu_lock(&call_rcu_mutex
);
550 alloc_cpu_call_rcu_data();
551 call_rcu_unlock(&call_rcu_mutex
);
556 if (per_cpu_call_rcu_data
== NULL
) {
560 for (i
= 0; i
< maxcpus
; i
++) {
561 call_rcu_lock(&call_rcu_mutex
);
562 if (get_cpu_call_rcu_data(i
)) {
563 call_rcu_unlock(&call_rcu_mutex
);
566 crdp
= __create_call_rcu_data(flags
, i
);
568 call_rcu_unlock(&call_rcu_mutex
);
572 call_rcu_unlock(&call_rcu_mutex
);
573 if ((ret
= set_cpu_call_rcu_data(i
, crdp
)) != 0) {
574 call_rcu_data_free(crdp
);
576 /* it has been created by other thread */
587 * Wake up the call_rcu thread corresponding to the specified
588 * call_rcu_data structure.
590 static void wake_call_rcu_thread(struct call_rcu_data
*crdp
)
592 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RT
))
593 call_rcu_wake_up(crdp
);
597 * Schedule a function to be invoked after a following grace period.
598 * This is the only function that must be called -- the others are
599 * only present to allow applications to tune their use of RCU for
600 * maximum performance.
602 * Note that unless a call_rcu thread has not already been created,
603 * the first invocation of call_rcu() will create one. So, if you
604 * need the first invocation of call_rcu() to be fast, make sure
605 * to create a call_rcu thread first. One way to accomplish this is
606 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
608 * call_rcu must be called by registered RCU read-side threads.
611 void call_rcu(struct rcu_head
*head
,
612 void (*func
)(struct rcu_head
*head
))
614 struct call_rcu_data
*crdp
;
616 cds_wfcq_node_init(&head
->next
);
618 /* Holding rcu read-side lock across use of per-cpu crdp */
620 crdp
= get_call_rcu_data();
621 cds_wfcq_enqueue(&crdp
->cbs_head
, &crdp
->cbs_tail
, &head
->next
);
622 uatomic_inc(&crdp
->qlen
);
623 wake_call_rcu_thread(crdp
);
628 * Free up the specified call_rcu_data structure, terminating the
629 * associated call_rcu thread. The caller must have previously
630 * removed the call_rcu_data structure from per-thread or per-CPU
631 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
632 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
633 * per-thread call_rcu_data structures.
635 * We silently refuse to free up the default call_rcu_data structure
636 * because that is where we put any leftover callbacks. Note that
637 * the possibility of self-spawning callbacks makes it impossible
638 * to execute all the callbacks in finite time without putting any
639 * newly spawned callbacks somewhere else. The "somewhere else" of
640 * last resort is the default call_rcu_data structure.
642 * We also silently refuse to free NULL pointers. This simplifies
645 * The caller must wait for a grace-period to pass between return from
646 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
647 * previous call rcu data as argument.
649 * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
650 * a list corruption bug in the 0.7.x series. The equivalent fix
651 * appeared in 0.6.8 for the stable-0.6 branch.
653 void call_rcu_data_free(struct call_rcu_data
*crdp
)
655 if (crdp
== NULL
|| crdp
== default_call_rcu_data
) {
658 if ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0) {
659 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOP
);
660 wake_call_rcu_thread(crdp
);
661 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0)
664 if (!cds_wfcq_empty(&crdp
->cbs_head
, &crdp
->cbs_tail
)) {
665 /* Create default call rcu data if need be */
666 (void) get_default_call_rcu_data();
667 __cds_wfcq_splice_blocking(&default_call_rcu_data
->cbs_head
,
668 &default_call_rcu_data
->cbs_tail
,
669 &crdp
->cbs_head
, &crdp
->cbs_tail
);
670 uatomic_add(&default_call_rcu_data
->qlen
,
671 uatomic_read(&crdp
->qlen
));
672 wake_call_rcu_thread(default_call_rcu_data
);
675 call_rcu_lock(&call_rcu_mutex
);
676 cds_list_del(&crdp
->list
);
677 call_rcu_unlock(&call_rcu_mutex
);
683 * Clean up all the per-CPU call_rcu threads.
685 void free_all_cpu_call_rcu_data(void)
688 struct call_rcu_data
**crdp
;
689 static int warned
= 0;
694 crdp
= malloc(sizeof(*crdp
) * maxcpus
);
697 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
703 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
704 crdp
[cpu
] = get_cpu_call_rcu_data(cpu
);
705 if (crdp
[cpu
] == NULL
)
707 set_cpu_call_rcu_data(cpu
, NULL
);
710 * Wait for call_rcu sites acting as RCU readers of the
711 * call_rcu_data to become quiescent.
714 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
715 if (crdp
[cpu
] == NULL
)
717 call_rcu_data_free(crdp
[cpu
]);
723 * Acquire the call_rcu_mutex in order to ensure that the child sees
724 * all of the call_rcu() data structures in a consistent state. Ensure
725 * that all call_rcu threads are in a quiescent state across fork.
726 * Suitable for pthread_atfork() and friends.
728 void call_rcu_before_fork(void)
730 struct call_rcu_data
*crdp
;
732 call_rcu_lock(&call_rcu_mutex
);
734 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
735 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_PAUSE
);
736 cmm_smp_mb__after_uatomic_or();
737 wake_call_rcu_thread(crdp
);
739 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
740 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSED
) == 0)
746 * Clean up call_rcu data structures in the parent of a successful fork()
747 * that is not followed by exec() in the child. Suitable for
748 * pthread_atfork() and friends.
750 void call_rcu_after_fork_parent(void)
752 struct call_rcu_data
*crdp
;
754 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
)
755 uatomic_and(&crdp
->flags
, ~URCU_CALL_RCU_PAUSE
);
756 call_rcu_unlock(&call_rcu_mutex
);
760 * Clean up call_rcu data structures in the child of a successful fork()
761 * that is not followed by exec(). Suitable for pthread_atfork() and
764 void call_rcu_after_fork_child(void)
766 struct call_rcu_data
*crdp
, *next
;
768 /* Release the mutex. */
769 call_rcu_unlock(&call_rcu_mutex
);
771 /* Do nothing when call_rcu() has not been used */
772 if (cds_list_empty(&call_rcu_data_list
))
776 * Allocate a new default call_rcu_data structure in order
777 * to get a working call_rcu thread to go with it.
779 default_call_rcu_data
= NULL
;
780 (void)get_default_call_rcu_data();
782 /* Cleanup call_rcu_data pointers before use */
784 free(per_cpu_call_rcu_data
);
785 rcu_set_pointer(&per_cpu_call_rcu_data
, NULL
);
786 URCU_TLS(thread_call_rcu_data
) = NULL
;
789 * Dispose of all of the rest of the call_rcu_data structures.
790 * Leftover call_rcu callbacks will be merged into the new
791 * default call_rcu thread queue.
793 cds_list_for_each_entry_safe(crdp
, next
, &call_rcu_data_list
, list
) {
794 if (crdp
== default_call_rcu_data
)
796 uatomic_set(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
797 call_rcu_data_free(crdp
);