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 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
87 * Pointer to array of pointers to per-CPU call_rcu_data structures
88 * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
89 * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
90 * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
91 * without mutex. The call_rcu_mutex protects updates.
94 static struct call_rcu_data
**per_cpu_call_rcu_data
;
97 static void maxcpus_reset(void)
102 /* Allocate the array if it has not already been allocated. */
104 static void alloc_cpu_call_rcu_data(void)
106 struct call_rcu_data
**p
;
107 static int warned
= 0;
111 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
115 p
= malloc(maxcpus
* sizeof(*per_cpu_call_rcu_data
));
117 memset(p
, '\0', maxcpus
* sizeof(*per_cpu_call_rcu_data
));
118 rcu_set_pointer(&per_cpu_call_rcu_data
, p
);
121 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
127 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
130 * per_cpu_call_rcu_data should be constant, but some functions below, used both
131 * for cases where cpu number is available and not available, assume it it not
134 static struct call_rcu_data
**per_cpu_call_rcu_data
= NULL
;
135 static const long maxcpus
= -1;
137 static void maxcpus_reset(void)
141 static void alloc_cpu_call_rcu_data(void)
145 static int sched_getcpu(void)
150 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
152 /* Acquire the specified pthread mutex. */
154 static void call_rcu_lock(pthread_mutex_t
*pmp
)
158 ret
= pthread_mutex_lock(pmp
);
163 /* Release the specified pthread mutex. */
165 static void call_rcu_unlock(pthread_mutex_t
*pmp
)
169 ret
= pthread_mutex_unlock(pmp
);
174 #if HAVE_SCHED_SETAFFINITY
176 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
180 if (crdp
->cpu_affinity
< 0)
184 CPU_SET(crdp
->cpu_affinity
, &mask
);
185 #if SCHED_SETAFFINITY_ARGS == 2
186 return sched_setaffinity(0, &mask
);
188 return sched_setaffinity(0, sizeof(mask
), &mask
);
193 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
199 static void call_rcu_wait(struct call_rcu_data
*crdp
)
201 /* Read call_rcu list before read futex */
203 if (uatomic_read(&crdp
->futex
) == -1)
204 futex_async(&crdp
->futex
, FUTEX_WAIT
, -1,
208 static void call_rcu_wake_up(struct call_rcu_data
*crdp
)
210 /* Write to call_rcu list before reading/writing futex */
212 if (caa_unlikely(uatomic_read(&crdp
->futex
) == -1)) {
213 uatomic_set(&crdp
->futex
, 0);
214 futex_async(&crdp
->futex
, FUTEX_WAKE
, 1,
219 /* This is the code run by each call_rcu thread. */
221 static void *call_rcu_thread(void *arg
)
223 unsigned long cbcount
;
224 struct cds_wfq_node
*cbs
;
225 struct cds_wfq_node
**cbs_tail
;
226 struct call_rcu_data
*crdp
= (struct call_rcu_data
*)arg
;
227 struct rcu_head
*rhp
;
228 int rt
= !!(uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_RT
);
231 ret
= set_thread_cpu_affinity(crdp
);
236 * If callbacks take a read-side lock, we need to be registered.
238 rcu_register_thread();
240 URCU_TLS(thread_call_rcu_data
) = crdp
;
242 uatomic_dec(&crdp
->futex
);
243 /* Decrement futex before reading call_rcu list */
247 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSE
) {
249 * Pause requested. Become quiescent: remove
250 * ourself from all global lists, and don't
251 * process any callback. The callback lists may
252 * still be non-empty though.
254 rcu_unregister_thread();
255 cmm_smp_mb__before_uatomic_or();
256 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_PAUSED
);
257 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSE
) != 0)
259 rcu_register_thread();
262 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
263 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
265 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
266 cbs_tail
= (struct cds_wfq_node
**)
267 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
271 while (cbs
->next
== NULL
&&
272 &cbs
->next
!= cbs_tail
)
274 if (cbs
== &crdp
->cbs
.dummy
) {
278 rhp
= (struct rcu_head
*)cbs
;
282 } while (cbs
!= NULL
);
283 uatomic_sub(&crdp
->qlen
, cbcount
);
285 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOP
)
287 rcu_thread_offline();
290 == _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
293 uatomic_dec(&crdp
->futex
);
295 * Decrement futex before reading
309 * Read call_rcu list before write futex.
312 uatomic_set(&crdp
->futex
, 0);
314 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
315 rcu_unregister_thread();
320 * Create both a call_rcu thread and the corresponding call_rcu_data
321 * structure, linking the structure in as specified. Caller must hold
325 static void call_rcu_data_init(struct call_rcu_data
**crdpp
,
329 struct call_rcu_data
*crdp
;
332 crdp
= malloc(sizeof(*crdp
));
335 memset(crdp
, '\0', sizeof(*crdp
));
336 cds_wfq_init(&crdp
->cbs
);
340 cds_list_add(&crdp
->list
, &call_rcu_data_list
);
341 crdp
->cpu_affinity
= cpu_affinity
;
342 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
344 ret
= pthread_create(&crdp
->tid
, NULL
, call_rcu_thread
, crdp
);
350 * Return a pointer to the call_rcu_data structure for the specified
351 * CPU, returning NULL if there is none. We cannot automatically
352 * created it because the platform we are running on might not define
355 * The call to this function and use of the returned call_rcu_data
356 * should be protected by RCU read-side lock.
359 struct call_rcu_data
*get_cpu_call_rcu_data(int cpu
)
361 static int warned
= 0;
362 struct call_rcu_data
**pcpu_crdp
;
364 pcpu_crdp
= rcu_dereference(per_cpu_call_rcu_data
);
365 if (pcpu_crdp
== NULL
)
367 if (!warned
&& maxcpus
> 0 && (cpu
< 0 || maxcpus
<= cpu
)) {
368 fprintf(stderr
, "[error] liburcu: get CPU # out of range\n");
371 if (cpu
< 0 || maxcpus
<= cpu
)
373 return rcu_dereference(pcpu_crdp
[cpu
]);
377 * Return the tid corresponding to the call_rcu thread whose
378 * call_rcu_data structure is specified.
381 pthread_t
get_call_rcu_thread(struct call_rcu_data
*crdp
)
387 * Create a call_rcu_data structure (with thread) and return a pointer.
390 static struct call_rcu_data
*__create_call_rcu_data(unsigned long flags
,
393 struct call_rcu_data
*crdp
;
395 call_rcu_data_init(&crdp
, flags
, cpu_affinity
);
399 struct call_rcu_data
*create_call_rcu_data(unsigned long flags
,
402 struct call_rcu_data
*crdp
;
404 call_rcu_lock(&call_rcu_mutex
);
405 crdp
= __create_call_rcu_data(flags
, cpu_affinity
);
406 call_rcu_unlock(&call_rcu_mutex
);
411 * Set the specified CPU to use the specified call_rcu_data structure.
413 * Use NULL to remove a CPU's call_rcu_data structure, but it is
414 * the caller's responsibility to dispose of the removed structure.
415 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
416 * (prior to NULLing it out, of course).
418 * The caller must wait for a grace-period to pass between return from
419 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
420 * previous call rcu data as argument.
423 int set_cpu_call_rcu_data(int cpu
, struct call_rcu_data
*crdp
)
425 static int warned
= 0;
427 call_rcu_lock(&call_rcu_mutex
);
428 alloc_cpu_call_rcu_data();
429 if (cpu
< 0 || maxcpus
<= cpu
) {
431 fprintf(stderr
, "[error] liburcu: set CPU # out of range\n");
434 call_rcu_unlock(&call_rcu_mutex
);
439 if (per_cpu_call_rcu_data
== NULL
) {
440 call_rcu_unlock(&call_rcu_mutex
);
445 if (per_cpu_call_rcu_data
[cpu
] != NULL
&& crdp
!= NULL
) {
446 call_rcu_unlock(&call_rcu_mutex
);
451 rcu_set_pointer(&per_cpu_call_rcu_data
[cpu
], crdp
);
452 call_rcu_unlock(&call_rcu_mutex
);
457 * Return a pointer to the default call_rcu_data structure, creating
458 * one if need be. Because we never free call_rcu_data structures,
459 * we don't need to be in an RCU read-side critical section.
462 struct call_rcu_data
*get_default_call_rcu_data(void)
464 if (default_call_rcu_data
!= NULL
)
465 return rcu_dereference(default_call_rcu_data
);
466 call_rcu_lock(&call_rcu_mutex
);
467 if (default_call_rcu_data
!= NULL
) {
468 call_rcu_unlock(&call_rcu_mutex
);
469 return default_call_rcu_data
;
471 call_rcu_data_init(&default_call_rcu_data
, 0, -1);
472 call_rcu_unlock(&call_rcu_mutex
);
473 return default_call_rcu_data
;
477 * Return the call_rcu_data structure that applies to the currently
478 * running thread. Any call_rcu_data structure assigned specifically
479 * to this thread has first priority, followed by any call_rcu_data
480 * structure assigned to the CPU on which the thread is running,
481 * followed by the default call_rcu_data structure. If there is not
482 * yet a default call_rcu_data structure, one will be created.
484 * Calls to this function and use of the returned call_rcu_data should
485 * be protected by RCU read-side lock.
487 struct call_rcu_data
*get_call_rcu_data(void)
489 struct call_rcu_data
*crd
;
491 if (URCU_TLS(thread_call_rcu_data
) != NULL
)
492 return URCU_TLS(thread_call_rcu_data
);
495 crd
= get_cpu_call_rcu_data(sched_getcpu());
500 return get_default_call_rcu_data();
504 * Return a pointer to this task's call_rcu_data if there is one.
507 struct call_rcu_data
*get_thread_call_rcu_data(void)
509 return URCU_TLS(thread_call_rcu_data
);
513 * Set this task's call_rcu_data structure as specified, regardless
514 * of whether or not this task already had one. (This allows switching
515 * to and from real-time call_rcu threads, for example.)
517 * Use NULL to remove a thread's call_rcu_data structure, but it is
518 * the caller's responsibility to dispose of the removed structure.
519 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
520 * (prior to NULLing it out, of course).
523 void set_thread_call_rcu_data(struct call_rcu_data
*crdp
)
525 URCU_TLS(thread_call_rcu_data
) = crdp
;
529 * Create a separate call_rcu thread for each CPU. This does not
530 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
531 * function if you want that behavior. Should be paired with
532 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
536 int create_all_cpu_call_rcu_data(unsigned long flags
)
539 struct call_rcu_data
*crdp
;
542 call_rcu_lock(&call_rcu_mutex
);
543 alloc_cpu_call_rcu_data();
544 call_rcu_unlock(&call_rcu_mutex
);
549 if (per_cpu_call_rcu_data
== NULL
) {
553 for (i
= 0; i
< maxcpus
; i
++) {
554 call_rcu_lock(&call_rcu_mutex
);
555 if (get_cpu_call_rcu_data(i
)) {
556 call_rcu_unlock(&call_rcu_mutex
);
559 crdp
= __create_call_rcu_data(flags
, i
);
561 call_rcu_unlock(&call_rcu_mutex
);
565 call_rcu_unlock(&call_rcu_mutex
);
566 if ((ret
= set_cpu_call_rcu_data(i
, crdp
)) != 0) {
567 call_rcu_data_free(crdp
);
569 /* it has been created by other thread */
580 * Wake up the call_rcu thread corresponding to the specified
581 * call_rcu_data structure.
583 static void wake_call_rcu_thread(struct call_rcu_data
*crdp
)
585 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RT
))
586 call_rcu_wake_up(crdp
);
590 * Schedule a function to be invoked after a following grace period.
591 * This is the only function that must be called -- the others are
592 * only present to allow applications to tune their use of RCU for
593 * maximum performance.
595 * Note that unless a call_rcu thread has not already been created,
596 * the first invocation of call_rcu() will create one. So, if you
597 * need the first invocation of call_rcu() to be fast, make sure
598 * to create a call_rcu thread first. One way to accomplish this is
599 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
601 * call_rcu must be called by registered RCU read-side threads.
604 void call_rcu(struct rcu_head
*head
,
605 void (*func
)(struct rcu_head
*head
))
607 struct call_rcu_data
*crdp
;
609 cds_wfq_node_init(&head
->next
);
611 /* Holding rcu read-side lock across use of per-cpu crdp */
613 crdp
= get_call_rcu_data();
614 cds_wfq_enqueue(&crdp
->cbs
, &head
->next
);
615 uatomic_inc(&crdp
->qlen
);
616 wake_call_rcu_thread(crdp
);
621 * Free up the specified call_rcu_data structure, terminating the
622 * associated call_rcu thread. The caller must have previously
623 * removed the call_rcu_data structure from per-thread or per-CPU
624 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
625 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
626 * per-thread call_rcu_data structures.
628 * We silently refuse to free up the default call_rcu_data structure
629 * because that is where we put any leftover callbacks. Note that
630 * the possibility of self-spawning callbacks makes it impossible
631 * to execute all the callbacks in finite time without putting any
632 * newly spawned callbacks somewhere else. The "somewhere else" of
633 * last resort is the default call_rcu_data structure.
635 * We also silently refuse to free NULL pointers. This simplifies
638 * The caller must wait for a grace-period to pass between return from
639 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
640 * previous call rcu data as argument.
642 void call_rcu_data_free(struct call_rcu_data
*crdp
)
644 struct cds_wfq_node
*cbs
;
645 struct cds_wfq_node
**cbs_tail
;
646 struct cds_wfq_node
**cbs_endprev
;
648 if (crdp
== NULL
|| crdp
== default_call_rcu_data
) {
651 if ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0) {
652 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOP
);
653 wake_call_rcu_thread(crdp
);
654 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0)
657 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
658 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
660 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
661 cbs_tail
= (struct cds_wfq_node
**)
662 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
663 /* Create default call rcu data if need be */
664 (void) get_default_call_rcu_data();
665 cbs_endprev
= (struct cds_wfq_node
**)
666 uatomic_xchg(&default_call_rcu_data
->cbs
.tail
,
668 _CMM_STORE_SHARED(*cbs_endprev
, cbs
);
669 uatomic_add(&default_call_rcu_data
->qlen
,
670 uatomic_read(&crdp
->qlen
));
671 wake_call_rcu_thread(default_call_rcu_data
);
674 call_rcu_lock(&call_rcu_mutex
);
675 cds_list_del(&crdp
->list
);
676 call_rcu_unlock(&call_rcu_mutex
);
682 * Clean up all the per-CPU call_rcu threads.
684 void free_all_cpu_call_rcu_data(void)
687 struct call_rcu_data
**crdp
;
688 static int warned
= 0;
693 crdp
= malloc(sizeof(*crdp
) * maxcpus
);
696 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
702 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
703 crdp
[cpu
] = get_cpu_call_rcu_data(cpu
);
704 if (crdp
[cpu
] == NULL
)
706 set_cpu_call_rcu_data(cpu
, NULL
);
709 * Wait for call_rcu sites acting as RCU readers of the
710 * call_rcu_data to become quiescent.
713 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
714 if (crdp
[cpu
] == NULL
)
716 call_rcu_data_free(crdp
[cpu
]);
722 * Acquire the call_rcu_mutex in order to ensure that the child sees
723 * all of the call_rcu() data structures in a consistent state. Ensure
724 * that all call_rcu threads are in a quiescent state across fork.
725 * Suitable for pthread_atfork() and friends.
727 void call_rcu_before_fork(void)
729 struct call_rcu_data
*crdp
;
731 call_rcu_lock(&call_rcu_mutex
);
733 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
734 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_PAUSE
);
735 cmm_smp_mb__after_uatomic_or();
736 wake_call_rcu_thread(crdp
);
738 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
) {
739 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_PAUSED
) == 0)
745 * Clean up call_rcu data structures in the parent of a successful fork()
746 * that is not followed by exec() in the child. Suitable for
747 * pthread_atfork() and friends.
749 void call_rcu_after_fork_parent(void)
751 struct call_rcu_data
*crdp
;
753 cds_list_for_each_entry(crdp
, &call_rcu_data_list
, list
)
754 uatomic_and(&crdp
->flags
, ~URCU_CALL_RCU_PAUSE
);
755 call_rcu_unlock(&call_rcu_mutex
);
759 * Clean up call_rcu data structures in the child of a successful fork()
760 * that is not followed by exec(). Suitable for pthread_atfork() and
763 void call_rcu_after_fork_child(void)
765 struct call_rcu_data
*crdp
, *next
;
767 /* Release the mutex. */
768 call_rcu_unlock(&call_rcu_mutex
);
770 /* Do nothing when call_rcu() has not been used */
771 if (cds_list_empty(&call_rcu_data_list
))
775 * Allocate a new default call_rcu_data structure in order
776 * to get a working call_rcu thread to go with it.
778 default_call_rcu_data
= NULL
;
779 (void)get_default_call_rcu_data();
781 /* Cleanup call_rcu_data pointers before use */
783 free(per_cpu_call_rcu_data
);
784 rcu_set_pointer(&per_cpu_call_rcu_data
, NULL
);
785 URCU_TLS(thread_call_rcu_data
) = NULL
;
788 * Dispose of all of the rest of the call_rcu_data structures.
789 * Leftover call_rcu callbacks will be merged into the new
790 * default call_rcu thread queue.
792 cds_list_for_each_entry_safe(crdp
, next
, &call_rcu_data_list
, list
) {
793 if (crdp
== default_call_rcu_data
)
795 uatomic_set(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
796 call_rcu_data_free(crdp
);