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"
43 /* Data structure that identifies a call_rcu thread. */
45 struct call_rcu_data
{
46 struct cds_wfq_queue cbs
;
53 struct cds_list_head list
;
54 } __attribute__((aligned(CAA_CACHE_LINE_SIZE
)));
57 * List of all call_rcu_data structures to keep valgrind happy.
58 * Protected by call_rcu_mutex.
61 CDS_LIST_HEAD(call_rcu_data_list
);
63 /* Link a thread using call_rcu() to its call_rcu thread. */
65 static __thread
struct call_rcu_data
*thread_call_rcu_data
;
67 /* Guard call_rcu thread creation. */
69 static pthread_mutex_t call_rcu_mutex
= PTHREAD_MUTEX_INITIALIZER
;
71 /* If a given thread does not have its own call_rcu thread, this is default. */
73 static struct call_rcu_data
*default_call_rcu_data
;
76 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
77 * available, then we can have call_rcu threads assigned to individual
78 * CPUs rather than only to specific threads.
81 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
84 * Pointer to array of pointers to per-CPU call_rcu_data structures
88 static struct call_rcu_data
**per_cpu_call_rcu_data
;
91 /* Allocate the array if it has not already been allocated. */
93 static void alloc_cpu_call_rcu_data(void)
95 struct call_rcu_data
**p
;
96 static int warned
= 0;
100 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
104 p
= malloc(maxcpus
* sizeof(*per_cpu_call_rcu_data
));
106 memset(p
, '\0', maxcpus
* sizeof(*per_cpu_call_rcu_data
));
107 per_cpu_call_rcu_data
= p
;
110 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
116 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
118 static const struct call_rcu_data
**per_cpu_call_rcu_data
= NULL
;
119 static const long maxcpus
= -1;
121 static void alloc_cpu_call_rcu_data(void)
125 static int sched_getcpu(void)
130 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
132 /* Acquire the specified pthread mutex. */
134 static void call_rcu_lock(pthread_mutex_t
*pmp
)
136 if (pthread_mutex_lock(pmp
) != 0) {
137 perror("pthread_mutex_lock");
142 /* Release the specified pthread mutex. */
144 static void call_rcu_unlock(pthread_mutex_t
*pmp
)
146 if (pthread_mutex_unlock(pmp
) != 0) {
147 perror("pthread_mutex_unlock");
152 #if HAVE_SCHED_SETAFFINITY
154 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
158 if (crdp
->cpu_affinity
< 0)
162 CPU_SET(crdp
->cpu_affinity
, &mask
);
163 #if SCHED_SETAFFINITY_ARGS == 2
164 return sched_setaffinity(0, &mask
);
166 return sched_setaffinity(0, sizeof(mask
), &mask
);
171 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
177 /* This is the code run by each call_rcu thread. */
179 static void *call_rcu_thread(void *arg
)
181 unsigned long cbcount
;
182 struct cds_wfq_node
*cbs
;
183 struct cds_wfq_node
**cbs_tail
;
184 struct call_rcu_data
*crdp
= (struct call_rcu_data
*)arg
;
185 struct rcu_head
*rhp
;
187 if (set_thread_cpu_affinity(crdp
) != 0) {
188 perror("pthread_setaffinity_np");
192 thread_call_rcu_data
= crdp
;
194 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
195 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
197 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
198 cbs_tail
= (struct cds_wfq_node
**)
199 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
203 while (cbs
->next
== NULL
&&
204 &cbs
->next
!= cbs_tail
)
206 if (cbs
== &crdp
->cbs
.dummy
) {
210 rhp
= (struct rcu_head
*)cbs
;
214 } while (cbs
!= NULL
);
215 uatomic_sub(&crdp
->qlen
, cbcount
);
217 if (crdp
->flags
& URCU_CALL_RCU_STOP
)
219 if (crdp
->flags
& URCU_CALL_RCU_RT
)
222 call_rcu_lock(&crdp
->mtx
);
223 _CMM_STORE_SHARED(crdp
->flags
,
224 crdp
->flags
& ~URCU_CALL_RCU_RUNNING
);
225 if (&crdp
->cbs
.head
==
226 _CMM_LOAD_SHARED(crdp
->cbs
.tail
) &&
227 pthread_cond_wait(&crdp
->cond
, &crdp
->mtx
) != 0) {
228 perror("pthread_cond_wait");
231 _CMM_STORE_SHARED(crdp
->flags
,
232 crdp
->flags
| URCU_CALL_RCU_RUNNING
);
234 call_rcu_unlock(&crdp
->mtx
);
237 call_rcu_lock(&crdp
->mtx
);
238 crdp
->flags
|= URCU_CALL_RCU_STOPPED
;
239 call_rcu_unlock(&crdp
->mtx
);
244 * Create both a call_rcu thread and the corresponding call_rcu_data
245 * structure, linking the structure in as specified. Caller must hold
249 static void call_rcu_data_init(struct call_rcu_data
**crdpp
,
253 struct call_rcu_data
*crdp
;
255 crdp
= malloc(sizeof(*crdp
));
257 fprintf(stderr
, "Out of memory.\n");
260 memset(crdp
, '\0', sizeof(*crdp
));
261 cds_wfq_init(&crdp
->cbs
);
263 if (pthread_mutex_init(&crdp
->mtx
, NULL
) != 0) {
264 perror("pthread_mutex_init");
267 if (pthread_cond_init(&crdp
->cond
, NULL
) != 0) {
268 perror("pthread_cond_init");
271 crdp
->flags
= flags
| URCU_CALL_RCU_RUNNING
;
272 cds_list_add(&crdp
->list
, &call_rcu_data_list
);
273 crdp
->cpu_affinity
= cpu_affinity
;
274 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
276 if (pthread_create(&crdp
->tid
, NULL
, call_rcu_thread
, crdp
) != 0) {
277 perror("pthread_create");
283 * Return a pointer to the call_rcu_data structure for the specified
284 * CPU, returning NULL if there is none. We cannot automatically
285 * created it because the platform we are running on might not define
289 struct call_rcu_data
*get_cpu_call_rcu_data(int cpu
)
291 static int warned
= 0;
293 if (per_cpu_call_rcu_data
== NULL
)
295 if (!warned
&& maxcpus
> 0 && (cpu
< 0 || maxcpus
<= cpu
)) {
296 fprintf(stderr
, "[error] liburcu: get CPU # out of range\n");
299 if (cpu
< 0 || maxcpus
<= cpu
)
301 return per_cpu_call_rcu_data
[cpu
];
305 * Return the tid corresponding to the call_rcu thread whose
306 * call_rcu_data structure is specified.
309 pthread_t
get_call_rcu_thread(struct call_rcu_data
*crdp
)
315 * Create a call_rcu_data structure (with thread) and return a pointer.
318 static struct call_rcu_data
*__create_call_rcu_data(unsigned long flags
,
321 struct call_rcu_data
*crdp
;
323 call_rcu_data_init(&crdp
, flags
, cpu_affinity
);
327 struct call_rcu_data
*create_call_rcu_data(unsigned long flags
,
330 struct call_rcu_data
*crdp
;
332 call_rcu_lock(&call_rcu_mutex
);
333 crdp
= __create_call_rcu_data(flags
, cpu_affinity
);
334 call_rcu_unlock(&call_rcu_mutex
);
339 * Set the specified CPU to use the specified call_rcu_data structure.
341 * Use NULL to remove a CPU's call_rcu_data structure, but it is
342 * the caller's responsibility to dispose of the removed structure.
343 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
344 * (prior to NULLing it out, of course).
347 int set_cpu_call_rcu_data(int cpu
, struct call_rcu_data
*crdp
)
351 call_rcu_lock(&call_rcu_mutex
);
352 if (cpu
< 0 || maxcpus
<= cpu
) {
354 fprintf(stderr
, "[error] liburcu: set CPU # out of range\n");
357 call_rcu_unlock(&call_rcu_mutex
);
361 alloc_cpu_call_rcu_data();
362 call_rcu_unlock(&call_rcu_mutex
);
363 if (per_cpu_call_rcu_data
== NULL
) {
367 per_cpu_call_rcu_data
[cpu
] = crdp
;
372 * Return a pointer to the default call_rcu_data structure, creating
373 * one if need be. Because we never free call_rcu_data structures,
374 * we don't need to be in an RCU read-side critical section.
377 struct call_rcu_data
*get_default_call_rcu_data(void)
379 if (default_call_rcu_data
!= NULL
)
380 return rcu_dereference(default_call_rcu_data
);
381 call_rcu_lock(&call_rcu_mutex
);
382 if (default_call_rcu_data
!= NULL
) {
383 call_rcu_unlock(&call_rcu_mutex
);
384 return default_call_rcu_data
;
386 call_rcu_data_init(&default_call_rcu_data
, 0, -1);
387 call_rcu_unlock(&call_rcu_mutex
);
388 return default_call_rcu_data
;
392 * Return the call_rcu_data structure that applies to the currently
393 * running thread. Any call_rcu_data structure assigned specifically
394 * to this thread has first priority, followed by any call_rcu_data
395 * structure assigned to the CPU on which the thread is running,
396 * followed by the default call_rcu_data structure. If there is not
397 * yet a default call_rcu_data structure, one will be created.
399 struct call_rcu_data
*get_call_rcu_data(void)
402 static int warned
= 0;
404 if (thread_call_rcu_data
!= NULL
)
405 return thread_call_rcu_data
;
407 return get_default_call_rcu_data();
408 curcpu
= sched_getcpu();
409 if (!warned
&& (curcpu
< 0 || maxcpus
<= curcpu
)) {
410 fprintf(stderr
, "[error] liburcu: gcrd CPU # out of range\n");
413 if (curcpu
>= 0 && maxcpus
> curcpu
&&
414 per_cpu_call_rcu_data
!= NULL
&&
415 per_cpu_call_rcu_data
[curcpu
] != NULL
)
416 return per_cpu_call_rcu_data
[curcpu
];
417 return get_default_call_rcu_data();
421 * Return a pointer to this task's call_rcu_data if there is one.
424 struct call_rcu_data
*get_thread_call_rcu_data(void)
426 return thread_call_rcu_data
;
430 * Set this task's call_rcu_data structure as specified, regardless
431 * of whether or not this task already had one. (This allows switching
432 * to and from real-time call_rcu threads, for example.)
434 * Use NULL to remove a thread's call_rcu_data structure, but it is
435 * the caller's responsibility to dispose of the removed structure.
436 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
437 * (prior to NULLing it out, of course).
440 void set_thread_call_rcu_data(struct call_rcu_data
*crdp
)
442 thread_call_rcu_data
= crdp
;
446 * Create a separate call_rcu thread for each CPU. This does not
447 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
448 * function if you want that behavior.
451 int create_all_cpu_call_rcu_data(unsigned long flags
)
454 struct call_rcu_data
*crdp
;
457 call_rcu_lock(&call_rcu_mutex
);
458 alloc_cpu_call_rcu_data();
459 call_rcu_unlock(&call_rcu_mutex
);
464 if (per_cpu_call_rcu_data
== NULL
) {
468 for (i
= 0; i
< maxcpus
; i
++) {
469 call_rcu_lock(&call_rcu_mutex
);
470 if (get_cpu_call_rcu_data(i
)) {
471 call_rcu_unlock(&call_rcu_mutex
);
474 crdp
= __create_call_rcu_data(flags
, i
);
476 call_rcu_unlock(&call_rcu_mutex
);
480 call_rcu_unlock(&call_rcu_mutex
);
481 if ((ret
= set_cpu_call_rcu_data(i
, crdp
)) != 0) {
482 /* FIXME: Leaks crdp for now. */
483 return ret
; /* Can happen on race. */
490 * Wake up the call_rcu thread corresponding to the specified
491 * call_rcu_data structure.
493 static void wake_call_rcu_thread(struct call_rcu_data
*crdp
)
495 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RT
)) {
496 call_rcu_lock(&crdp
->mtx
);
497 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RUNNING
)) {
498 if (pthread_cond_signal(&crdp
->cond
) != 0) {
499 perror("pthread_cond_signal");
503 call_rcu_unlock(&crdp
->mtx
);
508 * Schedule a function to be invoked after a following grace period.
509 * This is the only function that must be called -- the others are
510 * only present to allow applications to tune their use of RCU for
511 * maximum performance.
513 * Note that unless a call_rcu thread has not already been created,
514 * the first invocation of call_rcu() will create one. So, if you
515 * need the first invocation of call_rcu() to be fast, make sure
516 * to create a call_rcu thread first. One way to accomplish this is
517 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
520 void call_rcu(struct rcu_head
*head
,
521 void (*func
)(struct rcu_head
*head
))
523 struct call_rcu_data
*crdp
;
525 cds_wfq_node_init(&head
->next
);
527 crdp
= get_call_rcu_data();
528 cds_wfq_enqueue(&crdp
->cbs
, &head
->next
);
529 uatomic_inc(&crdp
->qlen
);
530 wake_call_rcu_thread(crdp
);
534 * Free up the specified call_rcu_data structure, terminating the
535 * associated call_rcu thread. The caller must have previously
536 * removed the call_rcu_data structure from per-thread or per-CPU
537 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
538 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
539 * per-thread call_rcu_data structures.
541 * We silently refuse to free up the default call_rcu_data structure
542 * because that is where we put any leftover callbacks. Note that
543 * the possibility of self-spawning callbacks makes it impossible
544 * to execute all the callbacks in finite time without putting any
545 * newly spawned callbacks somewhere else. The "somewhere else" of
546 * last resort is the default call_rcu_data structure.
548 * We also silently refuse to free NULL pointers. This simplifies
551 void call_rcu_data_free(struct call_rcu_data
*crdp
)
553 struct cds_wfq_node
*cbs
;
554 struct cds_wfq_node
**cbs_tail
;
555 struct cds_wfq_node
**cbs_endprev
;
557 if (crdp
== NULL
|| crdp
== default_call_rcu_data
) {
560 if ((crdp
->flags
& URCU_CALL_RCU_STOPPED
) == 0) {
561 call_rcu_lock(&crdp
->mtx
);
562 crdp
->flags
|= URCU_CALL_RCU_STOP
;
563 call_rcu_unlock(&crdp
->mtx
);
564 wake_call_rcu_thread(crdp
);
565 while ((crdp
->flags
& URCU_CALL_RCU_STOPPED
) == 0)
568 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
569 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
571 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
572 cbs_tail
= (struct cds_wfq_node
**)
573 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
574 cbs_endprev
= (struct cds_wfq_node
**)
575 uatomic_xchg(&default_call_rcu_data
, cbs_tail
);
577 uatomic_add(&default_call_rcu_data
->qlen
,
578 uatomic_read(&crdp
->qlen
));
579 cds_list_del(&crdp
->list
);
585 * Clean up all the per-CPU call_rcu threads.
587 void free_all_cpu_call_rcu_data(void)
590 struct call_rcu_data
*crdp
;
594 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
595 crdp
= get_cpu_call_rcu_data(cpu
);
598 set_cpu_call_rcu_data(cpu
, NULL
);
599 call_rcu_data_free(crdp
);
604 * Acquire the call_rcu_mutex in order to ensure that the child sees
605 * all of the call_rcu() data structures in a consistent state.
606 * Suitable for pthread_atfork() and friends.
608 void call_rcu_before_fork(void)
610 call_rcu_lock(&call_rcu_mutex
);
614 * Clean up call_rcu data structures in the parent of a successful fork()
615 * that is not followed by exec() in the child. Suitable for
616 * pthread_atfork() and friends.
618 void call_rcu_after_fork_parent(void)
620 call_rcu_unlock(&call_rcu_mutex
);
624 * Clean up call_rcu data structures in the child of a successful fork()
625 * that is not followed by exec(). Suitable for pthread_atfork() and
628 void call_rcu_after_fork_child(void)
630 struct call_rcu_data
*crdp
;
632 /* Release the mutex. */
633 call_rcu_unlock(&call_rcu_mutex
);
636 * Allocate a new default call_rcu_data structure in order
637 * to get a working call_rcu thread to go with it.
639 default_call_rcu_data
= NULL
;
640 (void)get_default_call_rcu_data();
642 /* Dispose of all of the rest of the call_rcu_data structures. */
643 while (call_rcu_data_list
.next
!= call_rcu_data_list
.prev
) {
644 crdp
= cds_list_entry(call_rcu_data_list
.prev
,
645 struct call_rcu_data
, list
);
646 if (crdp
== default_call_rcu_data
)
647 crdp
= cds_list_entry(crdp
->list
.prev
,
648 struct call_rcu_data
, list
);
649 crdp
->flags
= URCU_CALL_RCU_STOPPED
;
650 call_rcu_data_free(crdp
);