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
);
78 /* Guard call_rcu thread creation. */
80 static pthread_mutex_t call_rcu_mutex
= PTHREAD_MUTEX_INITIALIZER
;
82 /* If a given thread does not have its own call_rcu thread, this is default. */
84 static struct call_rcu_data
*default_call_rcu_data
;
87 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
88 * available, then we can have call_rcu threads assigned to individual
89 * CPUs rather than only to specific threads.
92 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
95 * Pointer to array of pointers to per-CPU call_rcu_data structures
96 * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
97 * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
98 * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
99 * without mutex. The call_rcu_mutex protects updates.
102 static struct call_rcu_data
**per_cpu_call_rcu_data
;
105 static void maxcpus_reset(void)
110 /* Allocate the array if it has not already been allocated. */
112 static void alloc_cpu_call_rcu_data(void)
114 struct call_rcu_data
**p
;
115 static int warned
= 0;
119 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
123 p
= malloc(maxcpus
* sizeof(*per_cpu_call_rcu_data
));
125 memset(p
, '\0', maxcpus
* sizeof(*per_cpu_call_rcu_data
));
126 rcu_set_pointer(&per_cpu_call_rcu_data
, p
);
129 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
135 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
138 * per_cpu_call_rcu_data should be constant, but some functions below, used both
139 * for cases where cpu number is available and not available, assume it it not
142 static struct call_rcu_data
**per_cpu_call_rcu_data
= NULL
;
143 static const long maxcpus
= -1;
145 static void maxcpus_reset(void)
149 static void alloc_cpu_call_rcu_data(void)
153 static int sched_getcpu(void)
158 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
160 /* Acquire the specified pthread mutex. */
162 static void call_rcu_lock(pthread_mutex_t
*pmp
)
166 ret
= pthread_mutex_lock(pmp
);
171 /* Release the specified pthread mutex. */
173 static void call_rcu_unlock(pthread_mutex_t
*pmp
)
177 ret
= pthread_mutex_unlock(pmp
);
182 #if HAVE_SCHED_SETAFFINITY
184 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
188 if (crdp
->cpu_affinity
< 0)
192 CPU_SET(crdp
->cpu_affinity
, &mask
);
193 #if SCHED_SETAFFINITY_ARGS == 2
194 return sched_setaffinity(0, &mask
);
196 return sched_setaffinity(0, sizeof(mask
), &mask
);
201 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
207 static void call_rcu_wait(struct call_rcu_data
*crdp
)
209 /* Read call_rcu list before read futex */
211 if (uatomic_read(&crdp
->futex
) == -1)
212 futex_async(&crdp
->futex
, FUTEX_WAIT
, -1,
216 static void call_rcu_wake_up(struct call_rcu_data
*crdp
)
218 /* Write to call_rcu list before reading/writing futex */
220 if (caa_unlikely(uatomic_read(&crdp
->futex
) == -1)) {
221 uatomic_set(&crdp
->futex
, 0);
222 futex_async(&crdp
->futex
, FUTEX_WAKE
, 1,
227 /* This is the code run by each call_rcu thread. */
229 static void *call_rcu_thread(void *arg
)
231 unsigned long cbcount
;
232 struct call_rcu_data
*crdp
= (struct call_rcu_data
*) arg
;
233 int rt
= !!(uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_RT
);
236 ret
= set_thread_cpu_affinity(crdp
);
241 * If callbacks take a read-side lock, we need to be registered.
243 rcu_register_thread();
245 URCU_TLS(thread_call_rcu_data
) = crdp
;
247 uatomic_dec(&crdp
->futex
);
248 /* Decrement futex before reading call_rcu list */
252 struct cds_wfcq_head cbs_tmp_head
;
253 struct cds_wfcq_tail cbs_tmp_tail
;
254 struct cds_wfcq_node
*cbs
, *cbs_tmp_n
;
255 enum cds_wfcq_ret splice_ret
;
257 cds_wfcq_init(&cbs_tmp_head
, &cbs_tmp_tail
);
258 splice_ret
= __cds_wfcq_splice_blocking(&cbs_tmp_head
,
259 &cbs_tmp_tail
, &crdp
->cbs_head
, &crdp
->cbs_tail
);
260 assert(splice_ret
!= CDS_WFCQ_RET_WOULDBLOCK
);
261 assert(splice_ret
!= CDS_WFCQ_RET_DEST_NON_EMPTY
);
262 if (splice_ret
!= CDS_WFCQ_RET_SRC_EMPTY
) {
265 __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head
,
266 &cbs_tmp_tail
, cbs
, cbs_tmp_n
) {
267 struct rcu_head
*rhp
;
269 rhp
= caa_container_of(cbs
,
270 struct rcu_head
, next
);
274 uatomic_sub(&crdp
->qlen
, cbcount
);
276 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOP
)
278 rcu_thread_offline();
280 if (cds_wfcq_empty(&crdp
->cbs_head
,
284 uatomic_dec(&crdp
->futex
);
286 * Decrement futex before reading
300 * Read call_rcu list before write futex.
303 uatomic_set(&crdp
->futex
, 0);
305 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
306 rcu_unregister_thread();
311 * Create both a call_rcu thread and the corresponding call_rcu_data
312 * structure, linking the structure in as specified. Caller must hold
316 static void call_rcu_data_init(struct call_rcu_data
**crdpp
,
320 struct call_rcu_data
*crdp
;
323 crdp
= malloc(sizeof(*crdp
));
326 memset(crdp
, '\0', sizeof(*crdp
));
327 cds_wfcq_init(&crdp
->cbs_head
, &crdp
->cbs_tail
);
331 cds_list_add(&crdp
->list
, &call_rcu_data_list
);
332 crdp
->cpu_affinity
= cpu_affinity
;
333 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
335 ret
= pthread_create(&crdp
->tid
, NULL
, call_rcu_thread
, crdp
);
341 * Return a pointer to the call_rcu_data structure for the specified
342 * CPU, returning NULL if there is none. We cannot automatically
343 * created it because the platform we are running on might not define
346 * The call to this function and use of the returned call_rcu_data
347 * should be protected by RCU read-side lock.
350 struct call_rcu_data
*get_cpu_call_rcu_data(int cpu
)
352 static int warned
= 0;
353 struct call_rcu_data
**pcpu_crdp
;
355 pcpu_crdp
= rcu_dereference(per_cpu_call_rcu_data
);
356 if (pcpu_crdp
== NULL
)
358 if (!warned
&& maxcpus
> 0 && (cpu
< 0 || maxcpus
<= cpu
)) {
359 fprintf(stderr
, "[error] liburcu: get CPU # out of range\n");
362 if (cpu
< 0 || maxcpus
<= cpu
)
364 return rcu_dereference(pcpu_crdp
[cpu
]);
368 * Return the tid corresponding to the call_rcu thread whose
369 * call_rcu_data structure is specified.
372 pthread_t
get_call_rcu_thread(struct call_rcu_data
*crdp
)
378 * Create a call_rcu_data structure (with thread) and return a pointer.
381 static struct call_rcu_data
*__create_call_rcu_data(unsigned long flags
,
384 struct call_rcu_data
*crdp
;
386 call_rcu_data_init(&crdp
, flags
, cpu_affinity
);
390 struct call_rcu_data
*create_call_rcu_data(unsigned long flags
,
393 struct call_rcu_data
*crdp
;
395 call_rcu_lock(&call_rcu_mutex
);
396 crdp
= __create_call_rcu_data(flags
, cpu_affinity
);
397 call_rcu_unlock(&call_rcu_mutex
);
402 * Set the specified CPU to use the specified call_rcu_data structure.
404 * Use NULL to remove a CPU's call_rcu_data structure, but it is
405 * the caller's responsibility to dispose of the removed structure.
406 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
407 * (prior to NULLing it out, of course).
409 * The caller must wait for a grace-period to pass between return from
410 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
411 * previous call rcu data as argument.
414 int set_cpu_call_rcu_data(int cpu
, struct call_rcu_data
*crdp
)
416 static int warned
= 0;
418 call_rcu_lock(&call_rcu_mutex
);
419 alloc_cpu_call_rcu_data();
420 if (cpu
< 0 || maxcpus
<= cpu
) {
422 fprintf(stderr
, "[error] liburcu: set CPU # out of range\n");
425 call_rcu_unlock(&call_rcu_mutex
);
430 if (per_cpu_call_rcu_data
== NULL
) {
431 call_rcu_unlock(&call_rcu_mutex
);
436 if (per_cpu_call_rcu_data
[cpu
] != NULL
&& crdp
!= NULL
) {
437 call_rcu_unlock(&call_rcu_mutex
);
442 rcu_set_pointer(&per_cpu_call_rcu_data
[cpu
], crdp
);
443 call_rcu_unlock(&call_rcu_mutex
);
448 * Return a pointer to the default call_rcu_data structure, creating
449 * one if need be. Because we never free call_rcu_data structures,
450 * we don't need to be in an RCU read-side critical section.
453 struct call_rcu_data
*get_default_call_rcu_data(void)
455 if (default_call_rcu_data
!= NULL
)
456 return rcu_dereference(default_call_rcu_data
);
457 call_rcu_lock(&call_rcu_mutex
);
458 if (default_call_rcu_data
!= NULL
) {
459 call_rcu_unlock(&call_rcu_mutex
);
460 return default_call_rcu_data
;
462 call_rcu_data_init(&default_call_rcu_data
, 0, -1);
463 call_rcu_unlock(&call_rcu_mutex
);
464 return default_call_rcu_data
;
468 * Return the call_rcu_data structure that applies to the currently
469 * running thread. Any call_rcu_data structure assigned specifically
470 * to this thread has first priority, followed by any call_rcu_data
471 * structure assigned to the CPU on which the thread is running,
472 * followed by the default call_rcu_data structure. If there is not
473 * yet a default call_rcu_data structure, one will be created.
475 * Calls to this function and use of the returned call_rcu_data should
476 * be protected by RCU read-side lock.
478 struct call_rcu_data
*get_call_rcu_data(void)
480 struct call_rcu_data
*crd
;
482 if (URCU_TLS(thread_call_rcu_data
) != NULL
)
483 return URCU_TLS(thread_call_rcu_data
);
486 crd
= get_cpu_call_rcu_data(sched_getcpu());
491 return get_default_call_rcu_data();
495 * Return a pointer to this task's call_rcu_data if there is one.
498 struct call_rcu_data
*get_thread_call_rcu_data(void)
500 return URCU_TLS(thread_call_rcu_data
);
504 * Set this task's call_rcu_data structure as specified, regardless
505 * of whether or not this task already had one. (This allows switching
506 * to and from real-time call_rcu threads, for example.)
508 * Use NULL to remove a thread's call_rcu_data structure, but it is
509 * the caller's responsibility to dispose of the removed structure.
510 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
511 * (prior to NULLing it out, of course).
514 void set_thread_call_rcu_data(struct call_rcu_data
*crdp
)
516 URCU_TLS(thread_call_rcu_data
) = crdp
;
520 * Create a separate call_rcu thread for each CPU. This does not
521 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
522 * function if you want that behavior. Should be paired with
523 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
527 int create_all_cpu_call_rcu_data(unsigned long flags
)
530 struct call_rcu_data
*crdp
;
533 call_rcu_lock(&call_rcu_mutex
);
534 alloc_cpu_call_rcu_data();
535 call_rcu_unlock(&call_rcu_mutex
);
540 if (per_cpu_call_rcu_data
== NULL
) {
544 for (i
= 0; i
< maxcpus
; i
++) {
545 call_rcu_lock(&call_rcu_mutex
);
546 if (get_cpu_call_rcu_data(i
)) {
547 call_rcu_unlock(&call_rcu_mutex
);
550 crdp
= __create_call_rcu_data(flags
, i
);
552 call_rcu_unlock(&call_rcu_mutex
);
556 call_rcu_unlock(&call_rcu_mutex
);
557 if ((ret
= set_cpu_call_rcu_data(i
, crdp
)) != 0) {
558 call_rcu_data_free(crdp
);
560 /* it has been created by other thread */
571 * Wake up the call_rcu thread corresponding to the specified
572 * call_rcu_data structure.
574 static void wake_call_rcu_thread(struct call_rcu_data
*crdp
)
576 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RT
))
577 call_rcu_wake_up(crdp
);
581 * Schedule a function to be invoked after a following grace period.
582 * This is the only function that must be called -- the others are
583 * only present to allow applications to tune their use of RCU for
584 * maximum performance.
586 * Note that unless a call_rcu thread has not already been created,
587 * the first invocation of call_rcu() will create one. So, if you
588 * need the first invocation of call_rcu() to be fast, make sure
589 * to create a call_rcu thread first. One way to accomplish this is
590 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
592 * call_rcu must be called by registered RCU read-side threads.
595 void call_rcu(struct rcu_head
*head
,
596 void (*func
)(struct rcu_head
*head
))
598 struct call_rcu_data
*crdp
;
600 cds_wfcq_node_init(&head
->next
);
602 /* Holding rcu read-side lock across use of per-cpu crdp */
604 crdp
= get_call_rcu_data();
605 cds_wfcq_enqueue(&crdp
->cbs_head
, &crdp
->cbs_tail
, &head
->next
);
606 uatomic_inc(&crdp
->qlen
);
607 wake_call_rcu_thread(crdp
);
612 * Free up the specified call_rcu_data structure, terminating the
613 * associated call_rcu thread. The caller must have previously
614 * removed the call_rcu_data structure from per-thread or per-CPU
615 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
616 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
617 * per-thread call_rcu_data structures.
619 * We silently refuse to free up the default call_rcu_data structure
620 * because that is where we put any leftover callbacks. Note that
621 * the possibility of self-spawning callbacks makes it impossible
622 * to execute all the callbacks in finite time without putting any
623 * newly spawned callbacks somewhere else. The "somewhere else" of
624 * last resort is the default call_rcu_data structure.
626 * We also silently refuse to free NULL pointers. This simplifies
629 * The caller must wait for a grace-period to pass between return from
630 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
631 * previous call rcu data as argument.
633 * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
634 * a list corruption bug in the 0.7.x series. The equivalent fix
635 * appeared in 0.6.8 for the stable-0.6 branch.
637 void call_rcu_data_free(struct call_rcu_data
*crdp
)
639 if (crdp
== NULL
|| crdp
== default_call_rcu_data
) {
642 if ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0) {
643 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOP
);
644 wake_call_rcu_thread(crdp
);
645 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0)
648 if (!cds_wfcq_empty(&crdp
->cbs_head
, &crdp
->cbs_tail
)) {
649 /* Create default call rcu data if need be */
650 (void) get_default_call_rcu_data();
651 __cds_wfcq_splice_blocking(&default_call_rcu_data
->cbs_head
,
652 &default_call_rcu_data
->cbs_tail
,
653 &crdp
->cbs_head
, &crdp
->cbs_tail
);
654 uatomic_add(&default_call_rcu_data
->qlen
,
655 uatomic_read(&crdp
->qlen
));
656 wake_call_rcu_thread(default_call_rcu_data
);
659 call_rcu_lock(&call_rcu_mutex
);
660 cds_list_del(&crdp
->list
);
661 call_rcu_unlock(&call_rcu_mutex
);
667 * Clean up all the per-CPU call_rcu threads.
669 void free_all_cpu_call_rcu_data(void)
672 struct call_rcu_data
**crdp
;
673 static int warned
= 0;
678 crdp
= malloc(sizeof(*crdp
) * maxcpus
);
681 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
687 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
688 crdp
[cpu
] = get_cpu_call_rcu_data(cpu
);
689 if (crdp
[cpu
] == NULL
)
691 set_cpu_call_rcu_data(cpu
, NULL
);
694 * Wait for call_rcu sites acting as RCU readers of the
695 * call_rcu_data to become quiescent.
698 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
699 if (crdp
[cpu
] == NULL
)
701 call_rcu_data_free(crdp
[cpu
]);
707 * Acquire the call_rcu_mutex in order to ensure that the child sees
708 * all of the call_rcu() data structures in a consistent state.
709 * Suitable for pthread_atfork() and friends.
711 void call_rcu_before_fork(void)
713 call_rcu_lock(&call_rcu_mutex
);
717 * Clean up call_rcu data structures in the parent of a successful fork()
718 * that is not followed by exec() in the child. Suitable for
719 * pthread_atfork() and friends.
721 void call_rcu_after_fork_parent(void)
723 call_rcu_unlock(&call_rcu_mutex
);
727 * Clean up call_rcu data structures in the child of a successful fork()
728 * that is not followed by exec(). Suitable for pthread_atfork() and
731 void call_rcu_after_fork_child(void)
733 struct call_rcu_data
*crdp
, *next
;
735 /* Release the mutex. */
736 call_rcu_unlock(&call_rcu_mutex
);
738 /* Do nothing when call_rcu() has not been used */
739 if (cds_list_empty(&call_rcu_data_list
))
743 * Allocate a new default call_rcu_data structure in order
744 * to get a working call_rcu thread to go with it.
746 default_call_rcu_data
= NULL
;
747 (void)get_default_call_rcu_data();
749 /* Cleanup call_rcu_data pointers before use */
751 free(per_cpu_call_rcu_data
);
752 rcu_set_pointer(&per_cpu_call_rcu_data
, NULL
);
753 URCU_TLS(thread_call_rcu_data
) = NULL
;
755 /* Dispose of all of the rest of the call_rcu_data structures. */
756 cds_list_for_each_entry_safe(crdp
, next
, &call_rcu_data_list
, list
) {
757 if (crdp
== default_call_rcu_data
)
759 uatomic_set(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
760 call_rcu_data_free(crdp
);