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"
45 /* Data structure that identifies a call_rcu thread. */
47 struct call_rcu_data
{
48 struct cds_wfq_queue cbs
;
51 unsigned long qlen
; /* maintained for debugging. */
54 struct cds_list_head list
;
55 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
58 * List of all call_rcu_data structures to keep valgrind happy.
59 * Protected by call_rcu_mutex.
62 CDS_LIST_HEAD(call_rcu_data_list
);
64 /* Link a thread using call_rcu() to its call_rcu thread. */
66 static DEFINE_URCU_TLS(struct call_rcu_data
*, thread_call_rcu_data
);
68 /* Guard call_rcu thread creation. */
70 static pthread_mutex_t call_rcu_mutex
= PTHREAD_MUTEX_INITIALIZER
;
72 /* If a given thread does not have its own call_rcu thread, this is default. */
74 static struct call_rcu_data
*default_call_rcu_data
;
77 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
78 * available, then we can have call_rcu threads assigned to individual
79 * CPUs rather than only to specific threads.
82 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
85 * Pointer to array of pointers to per-CPU call_rcu_data structures
86 * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
87 * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
88 * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
89 * without mutex. The call_rcu_mutex protects updates.
92 static struct call_rcu_data
**per_cpu_call_rcu_data
;
95 static void maxcpus_reset(void)
100 /* Allocate the array if it has not already been allocated. */
102 static void alloc_cpu_call_rcu_data(void)
104 struct call_rcu_data
**p
;
105 static int warned
= 0;
109 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
113 p
= malloc(maxcpus
* sizeof(*per_cpu_call_rcu_data
));
115 memset(p
, '\0', maxcpus
* sizeof(*per_cpu_call_rcu_data
));
116 rcu_set_pointer(&per_cpu_call_rcu_data
, p
);
119 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
125 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
128 * per_cpu_call_rcu_data should be constant, but some functions below, used both
129 * for cases where cpu number is available and not available, assume it it not
132 static struct call_rcu_data
**per_cpu_call_rcu_data
= NULL
;
133 static const long maxcpus
= -1;
135 static void maxcpus_reset(void)
139 static void alloc_cpu_call_rcu_data(void)
143 static int sched_getcpu(void)
148 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
150 /* Acquire the specified pthread mutex. */
152 static void call_rcu_lock(pthread_mutex_t
*pmp
)
154 if (pthread_mutex_lock(pmp
) != 0) {
155 perror("pthread_mutex_lock");
160 /* Release the specified pthread mutex. */
162 static void call_rcu_unlock(pthread_mutex_t
*pmp
)
164 if (pthread_mutex_unlock(pmp
) != 0) {
165 perror("pthread_mutex_unlock");
170 #if HAVE_SCHED_SETAFFINITY
172 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
176 if (crdp
->cpu_affinity
< 0)
180 CPU_SET(crdp
->cpu_affinity
, &mask
);
181 #if SCHED_SETAFFINITY_ARGS == 2
182 return sched_setaffinity(0, &mask
);
184 return sched_setaffinity(0, sizeof(mask
), &mask
);
189 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
195 static void call_rcu_wait(struct call_rcu_data
*crdp
)
197 /* Read call_rcu list before read futex */
199 if (uatomic_read(&crdp
->futex
) == -1)
200 futex_async(&crdp
->futex
, FUTEX_WAIT
, -1,
204 static void call_rcu_wake_up(struct call_rcu_data
*crdp
)
206 /* Write to call_rcu list before reading/writing futex */
208 if (caa_unlikely(uatomic_read(&crdp
->futex
) == -1)) {
209 uatomic_set(&crdp
->futex
, 0);
210 futex_async(&crdp
->futex
, FUTEX_WAKE
, 1,
215 /* This is the code run by each call_rcu thread. */
217 static void *call_rcu_thread(void *arg
)
219 unsigned long cbcount
;
220 struct cds_wfq_node
*cbs
;
221 struct cds_wfq_node
**cbs_tail
;
222 struct call_rcu_data
*crdp
= (struct call_rcu_data
*)arg
;
223 struct rcu_head
*rhp
;
224 int rt
= !!(uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_RT
);
226 if (set_thread_cpu_affinity(crdp
) != 0) {
227 perror("pthread_setaffinity_np");
232 * If callbacks take a read-side lock, we need to be registered.
234 rcu_register_thread();
236 URCU_TLS(thread_call_rcu_data
) = crdp
;
238 uatomic_dec(&crdp
->futex
);
239 /* Decrement futex before reading call_rcu list */
243 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
244 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
246 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
247 cbs_tail
= (struct cds_wfq_node
**)
248 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
252 while (cbs
->next
== NULL
&&
253 &cbs
->next
!= cbs_tail
)
255 if (cbs
== &crdp
->cbs
.dummy
) {
259 rhp
= (struct rcu_head
*)cbs
;
263 } while (cbs
!= NULL
);
264 uatomic_sub(&crdp
->qlen
, cbcount
);
266 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOP
)
268 rcu_thread_offline();
271 == _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
274 uatomic_dec(&crdp
->futex
);
276 * Decrement futex before reading
290 * Read call_rcu list before write futex.
293 uatomic_set(&crdp
->futex
, 0);
295 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
296 rcu_unregister_thread();
301 * Create both a call_rcu thread and the corresponding call_rcu_data
302 * structure, linking the structure in as specified. Caller must hold
306 static void call_rcu_data_init(struct call_rcu_data
**crdpp
,
310 struct call_rcu_data
*crdp
;
312 crdp
= malloc(sizeof(*crdp
));
314 fprintf(stderr
, "Out of memory.\n");
317 memset(crdp
, '\0', sizeof(*crdp
));
318 cds_wfq_init(&crdp
->cbs
);
322 cds_list_add(&crdp
->list
, &call_rcu_data_list
);
323 crdp
->cpu_affinity
= cpu_affinity
;
324 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
326 if (pthread_create(&crdp
->tid
, NULL
, call_rcu_thread
, crdp
) != 0) {
327 perror("pthread_create");
333 * Return a pointer to the call_rcu_data structure for the specified
334 * CPU, returning NULL if there is none. We cannot automatically
335 * created it because the platform we are running on might not define
338 * The call to this function and use of the returned call_rcu_data
339 * should be protected by RCU read-side lock.
342 struct call_rcu_data
*get_cpu_call_rcu_data(int cpu
)
344 static int warned
= 0;
345 struct call_rcu_data
**pcpu_crdp
;
347 pcpu_crdp
= rcu_dereference(per_cpu_call_rcu_data
);
348 if (pcpu_crdp
== NULL
)
350 if (!warned
&& maxcpus
> 0 && (cpu
< 0 || maxcpus
<= cpu
)) {
351 fprintf(stderr
, "[error] liburcu: get CPU # out of range\n");
354 if (cpu
< 0 || maxcpus
<= cpu
)
356 return rcu_dereference(pcpu_crdp
[cpu
]);
360 * Return the tid corresponding to the call_rcu thread whose
361 * call_rcu_data structure is specified.
364 pthread_t
get_call_rcu_thread(struct call_rcu_data
*crdp
)
370 * Create a call_rcu_data structure (with thread) and return a pointer.
373 static struct call_rcu_data
*__create_call_rcu_data(unsigned long flags
,
376 struct call_rcu_data
*crdp
;
378 call_rcu_data_init(&crdp
, flags
, cpu_affinity
);
382 struct call_rcu_data
*create_call_rcu_data(unsigned long flags
,
385 struct call_rcu_data
*crdp
;
387 call_rcu_lock(&call_rcu_mutex
);
388 crdp
= __create_call_rcu_data(flags
, cpu_affinity
);
389 call_rcu_unlock(&call_rcu_mutex
);
394 * Set the specified CPU to use the specified call_rcu_data structure.
396 * Use NULL to remove a CPU's call_rcu_data structure, but it is
397 * the caller's responsibility to dispose of the removed structure.
398 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
399 * (prior to NULLing it out, of course).
401 * The caller must wait for a grace-period to pass between return from
402 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
403 * previous call rcu data as argument.
406 int set_cpu_call_rcu_data(int cpu
, struct call_rcu_data
*crdp
)
408 static int warned
= 0;
410 call_rcu_lock(&call_rcu_mutex
);
411 alloc_cpu_call_rcu_data();
412 if (cpu
< 0 || maxcpus
<= cpu
) {
414 fprintf(stderr
, "[error] liburcu: set CPU # out of range\n");
417 call_rcu_unlock(&call_rcu_mutex
);
422 if (per_cpu_call_rcu_data
== NULL
) {
423 call_rcu_unlock(&call_rcu_mutex
);
428 if (per_cpu_call_rcu_data
[cpu
] != NULL
&& crdp
!= NULL
) {
429 call_rcu_unlock(&call_rcu_mutex
);
434 rcu_set_pointer(&per_cpu_call_rcu_data
[cpu
], crdp
);
435 call_rcu_unlock(&call_rcu_mutex
);
440 * Return a pointer to the default call_rcu_data structure, creating
441 * one if need be. Because we never free call_rcu_data structures,
442 * we don't need to be in an RCU read-side critical section.
445 struct call_rcu_data
*get_default_call_rcu_data(void)
447 if (default_call_rcu_data
!= NULL
)
448 return rcu_dereference(default_call_rcu_data
);
449 call_rcu_lock(&call_rcu_mutex
);
450 if (default_call_rcu_data
!= NULL
) {
451 call_rcu_unlock(&call_rcu_mutex
);
452 return default_call_rcu_data
;
454 call_rcu_data_init(&default_call_rcu_data
, 0, -1);
455 call_rcu_unlock(&call_rcu_mutex
);
456 return default_call_rcu_data
;
460 * Return the call_rcu_data structure that applies to the currently
461 * running thread. Any call_rcu_data structure assigned specifically
462 * to this thread has first priority, followed by any call_rcu_data
463 * structure assigned to the CPU on which the thread is running,
464 * followed by the default call_rcu_data structure. If there is not
465 * yet a default call_rcu_data structure, one will be created.
467 * Calls to this function and use of the returned call_rcu_data should
468 * be protected by RCU read-side lock.
470 struct call_rcu_data
*get_call_rcu_data(void)
472 struct call_rcu_data
*crd
;
474 if (URCU_TLS(thread_call_rcu_data
) != NULL
)
475 return URCU_TLS(thread_call_rcu_data
);
478 crd
= get_cpu_call_rcu_data(sched_getcpu());
483 return get_default_call_rcu_data();
487 * Return a pointer to this task's call_rcu_data if there is one.
490 struct call_rcu_data
*get_thread_call_rcu_data(void)
492 return URCU_TLS(thread_call_rcu_data
);
496 * Set this task's call_rcu_data structure as specified, regardless
497 * of whether or not this task already had one. (This allows switching
498 * to and from real-time call_rcu threads, for example.)
500 * Use NULL to remove a thread's call_rcu_data structure, but it is
501 * the caller's responsibility to dispose of the removed structure.
502 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
503 * (prior to NULLing it out, of course).
506 void set_thread_call_rcu_data(struct call_rcu_data
*crdp
)
508 URCU_TLS(thread_call_rcu_data
) = crdp
;
512 * Create a separate call_rcu thread for each CPU. This does not
513 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
514 * function if you want that behavior. Should be paired with
515 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
519 int create_all_cpu_call_rcu_data(unsigned long flags
)
522 struct call_rcu_data
*crdp
;
525 call_rcu_lock(&call_rcu_mutex
);
526 alloc_cpu_call_rcu_data();
527 call_rcu_unlock(&call_rcu_mutex
);
532 if (per_cpu_call_rcu_data
== NULL
) {
536 for (i
= 0; i
< maxcpus
; i
++) {
537 call_rcu_lock(&call_rcu_mutex
);
538 if (get_cpu_call_rcu_data(i
)) {
539 call_rcu_unlock(&call_rcu_mutex
);
542 crdp
= __create_call_rcu_data(flags
, i
);
544 call_rcu_unlock(&call_rcu_mutex
);
548 call_rcu_unlock(&call_rcu_mutex
);
549 if ((ret
= set_cpu_call_rcu_data(i
, crdp
)) != 0) {
550 call_rcu_data_free(crdp
);
552 /* it has been created by other thread */
563 * Wake up the call_rcu thread corresponding to the specified
564 * call_rcu_data structure.
566 static void wake_call_rcu_thread(struct call_rcu_data
*crdp
)
568 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RT
))
569 call_rcu_wake_up(crdp
);
573 * Schedule a function to be invoked after a following grace period.
574 * This is the only function that must be called -- the others are
575 * only present to allow applications to tune their use of RCU for
576 * maximum performance.
578 * Note that unless a call_rcu thread has not already been created,
579 * the first invocation of call_rcu() will create one. So, if you
580 * need the first invocation of call_rcu() to be fast, make sure
581 * to create a call_rcu thread first. One way to accomplish this is
582 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
584 * call_rcu must be called by registered RCU read-side threads.
587 void call_rcu(struct rcu_head
*head
,
588 void (*func
)(struct rcu_head
*head
))
590 struct call_rcu_data
*crdp
;
592 cds_wfq_node_init(&head
->next
);
594 /* Holding rcu read-side lock across use of per-cpu crdp */
596 crdp
= get_call_rcu_data();
597 cds_wfq_enqueue(&crdp
->cbs
, &head
->next
);
598 uatomic_inc(&crdp
->qlen
);
599 wake_call_rcu_thread(crdp
);
604 * Free up the specified call_rcu_data structure, terminating the
605 * associated call_rcu thread. The caller must have previously
606 * removed the call_rcu_data structure from per-thread or per-CPU
607 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
608 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
609 * per-thread call_rcu_data structures.
611 * We silently refuse to free up the default call_rcu_data structure
612 * because that is where we put any leftover callbacks. Note that
613 * the possibility of self-spawning callbacks makes it impossible
614 * to execute all the callbacks in finite time without putting any
615 * newly spawned callbacks somewhere else. The "somewhere else" of
616 * last resort is the default call_rcu_data structure.
618 * We also silently refuse to free NULL pointers. This simplifies
621 * The caller must wait for a grace-period to pass between return from
622 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
623 * previous call rcu data as argument.
625 void call_rcu_data_free(struct call_rcu_data
*crdp
)
627 struct cds_wfq_node
*cbs
;
628 struct cds_wfq_node
**cbs_tail
;
629 struct cds_wfq_node
**cbs_endprev
;
631 if (crdp
== NULL
|| crdp
== default_call_rcu_data
) {
634 if ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0) {
635 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOP
);
636 wake_call_rcu_thread(crdp
);
637 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0)
640 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
641 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
643 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
644 cbs_tail
= (struct cds_wfq_node
**)
645 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
646 /* Create default call rcu data if need be */
647 (void) get_default_call_rcu_data();
648 cbs_endprev
= (struct cds_wfq_node
**)
649 uatomic_xchg(&default_call_rcu_data
, cbs_tail
);
651 uatomic_add(&default_call_rcu_data
->qlen
,
652 uatomic_read(&crdp
->qlen
));
653 wake_call_rcu_thread(default_call_rcu_data
);
656 call_rcu_lock(&call_rcu_mutex
);
657 cds_list_del(&crdp
->list
);
658 call_rcu_unlock(&call_rcu_mutex
);
664 * Clean up all the per-CPU call_rcu threads.
666 void free_all_cpu_call_rcu_data(void)
669 struct call_rcu_data
**crdp
;
670 static int warned
= 0;
675 crdp
= malloc(sizeof(*crdp
) * maxcpus
);
678 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
684 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
685 crdp
[cpu
] = get_cpu_call_rcu_data(cpu
);
686 if (crdp
[cpu
] == NULL
)
688 set_cpu_call_rcu_data(cpu
, NULL
);
691 * Wait for call_rcu sites acting as RCU readers of the
692 * call_rcu_data to become quiescent.
695 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
696 if (crdp
[cpu
] == NULL
)
698 call_rcu_data_free(crdp
[cpu
]);
704 * Acquire the call_rcu_mutex in order to ensure that the child sees
705 * all of the call_rcu() data structures in a consistent state.
706 * Suitable for pthread_atfork() and friends.
708 void call_rcu_before_fork(void)
710 call_rcu_lock(&call_rcu_mutex
);
714 * Clean up call_rcu data structures in the parent of a successful fork()
715 * that is not followed by exec() in the child. Suitable for
716 * pthread_atfork() and friends.
718 void call_rcu_after_fork_parent(void)
720 call_rcu_unlock(&call_rcu_mutex
);
724 * Clean up call_rcu data structures in the child of a successful fork()
725 * that is not followed by exec(). Suitable for pthread_atfork() and
728 void call_rcu_after_fork_child(void)
730 struct call_rcu_data
*crdp
, *next
;
732 /* Release the mutex. */
733 call_rcu_unlock(&call_rcu_mutex
);
735 /* Do nothing when call_rcu() has not been used */
736 if (cds_list_empty(&call_rcu_data_list
))
740 * Allocate a new default call_rcu_data structure in order
741 * to get a working call_rcu thread to go with it.
743 default_call_rcu_data
= NULL
;
744 (void)get_default_call_rcu_data();
746 /* Cleanup call_rcu_data pointers before use */
748 free(per_cpu_call_rcu_data
);
749 rcu_set_pointer(&per_cpu_call_rcu_data
, NULL
);
750 URCU_TLS(thread_call_rcu_data
) = NULL
;
752 /* Dispose of all of the rest of the call_rcu_data structures. */
753 cds_list_for_each_entry_safe(crdp
, next
, &call_rcu_data_list
, list
) {
754 if (crdp
== default_call_rcu_data
)
756 uatomic_set(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
757 call_rcu_data_free(crdp
);