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"
44 /* Data structure that identifies a call_rcu thread. */
46 struct call_rcu_data
{
47 struct cds_wfq_queue cbs
;
50 unsigned long qlen
; /* maintained for debugging. */
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) */
119 * per_cpu_call_rcu_data should be constant, but some functions below, used both
120 * for cases where cpu number is available and not available, assume it it not
123 static struct call_rcu_data
**per_cpu_call_rcu_data
= NULL
;
124 static const long maxcpus
= -1;
126 static void alloc_cpu_call_rcu_data(void)
130 static int sched_getcpu(void)
135 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
137 /* Acquire the specified pthread mutex. */
139 static void call_rcu_lock(pthread_mutex_t
*pmp
)
141 if (pthread_mutex_lock(pmp
) != 0) {
142 perror("pthread_mutex_lock");
147 /* Release the specified pthread mutex. */
149 static void call_rcu_unlock(pthread_mutex_t
*pmp
)
151 if (pthread_mutex_unlock(pmp
) != 0) {
152 perror("pthread_mutex_unlock");
157 #if HAVE_SCHED_SETAFFINITY
159 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
163 if (crdp
->cpu_affinity
< 0)
167 CPU_SET(crdp
->cpu_affinity
, &mask
);
168 #if SCHED_SETAFFINITY_ARGS == 2
169 return sched_setaffinity(0, &mask
);
171 return sched_setaffinity(0, sizeof(mask
), &mask
);
176 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
182 static void call_rcu_wait(struct call_rcu_data
*crdp
)
184 /* Read call_rcu list before read futex */
186 if (uatomic_read(&crdp
->futex
) == -1)
187 futex_async(&crdp
->futex
, FUTEX_WAIT
, -1,
191 static void call_rcu_wake_up(struct call_rcu_data
*crdp
)
193 /* Write to call_rcu list before reading/writing futex */
195 if (unlikely(uatomic_read(&crdp
->futex
) == -1)) {
196 uatomic_set(&crdp
->futex
, 0);
197 futex_async(&crdp
->futex
, FUTEX_WAKE
, 1,
202 /* This is the code run by each call_rcu thread. */
204 static void *call_rcu_thread(void *arg
)
206 unsigned long cbcount
;
207 struct cds_wfq_node
*cbs
;
208 struct cds_wfq_node
**cbs_tail
;
209 struct call_rcu_data
*crdp
= (struct call_rcu_data
*)arg
;
210 struct rcu_head
*rhp
;
211 int rt
= !!(uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_RT
);
213 if (set_thread_cpu_affinity(crdp
) != 0) {
214 perror("pthread_setaffinity_np");
219 * If callbacks take a read-side lock, we need to be registered.
221 rcu_register_thread();
223 thread_call_rcu_data
= crdp
;
225 uatomic_dec(&crdp
->futex
);
226 /* Decrement futex before reading call_rcu list */
230 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
231 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
233 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
234 cbs_tail
= (struct cds_wfq_node
**)
235 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
239 while (cbs
->next
== NULL
&&
240 &cbs
->next
!= cbs_tail
)
242 if (cbs
== &crdp
->cbs
.dummy
) {
246 rhp
= (struct rcu_head
*)cbs
;
250 } while (cbs
!= NULL
);
251 uatomic_sub(&crdp
->qlen
, cbcount
);
253 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOP
)
255 rcu_thread_offline();
258 == _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
261 uatomic_dec(&crdp
->futex
);
263 * Decrement futex before reading
277 * Read call_rcu list before write futex.
280 uatomic_set(&crdp
->futex
, 0);
282 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
283 rcu_unregister_thread();
288 * Create both a call_rcu thread and the corresponding call_rcu_data
289 * structure, linking the structure in as specified. Caller must hold
293 static void call_rcu_data_init(struct call_rcu_data
**crdpp
,
297 struct call_rcu_data
*crdp
;
299 crdp
= malloc(sizeof(*crdp
));
301 fprintf(stderr
, "Out of memory.\n");
304 memset(crdp
, '\0', sizeof(*crdp
));
305 cds_wfq_init(&crdp
->cbs
);
309 cds_list_add(&crdp
->list
, &call_rcu_data_list
);
310 crdp
->cpu_affinity
= cpu_affinity
;
311 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
313 if (pthread_create(&crdp
->tid
, NULL
, call_rcu_thread
, crdp
) != 0) {
314 perror("pthread_create");
320 * Return a pointer to the call_rcu_data structure for the specified
321 * CPU, returning NULL if there is none. We cannot automatically
322 * created it because the platform we are running on might not define
326 struct call_rcu_data
*get_cpu_call_rcu_data(int cpu
)
328 static int warned
= 0;
330 if (per_cpu_call_rcu_data
== NULL
)
332 if (!warned
&& maxcpus
> 0 && (cpu
< 0 || maxcpus
<= cpu
)) {
333 fprintf(stderr
, "[error] liburcu: get CPU # out of range\n");
336 if (cpu
< 0 || maxcpus
<= cpu
)
338 return per_cpu_call_rcu_data
[cpu
];
342 * Return the tid corresponding to the call_rcu thread whose
343 * call_rcu_data structure is specified.
346 pthread_t
get_call_rcu_thread(struct call_rcu_data
*crdp
)
352 * Create a call_rcu_data structure (with thread) and return a pointer.
355 static struct call_rcu_data
*__create_call_rcu_data(unsigned long flags
,
358 struct call_rcu_data
*crdp
;
360 call_rcu_data_init(&crdp
, flags
, cpu_affinity
);
364 struct call_rcu_data
*create_call_rcu_data(unsigned long flags
,
367 struct call_rcu_data
*crdp
;
369 call_rcu_lock(&call_rcu_mutex
);
370 crdp
= __create_call_rcu_data(flags
, cpu_affinity
);
371 call_rcu_unlock(&call_rcu_mutex
);
376 * Set the specified CPU to use the specified call_rcu_data structure.
378 * Use NULL to remove a CPU's call_rcu_data structure, but it is
379 * the caller's responsibility to dispose of the removed structure.
380 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
381 * (prior to NULLing it out, of course).
384 int set_cpu_call_rcu_data(int cpu
, struct call_rcu_data
*crdp
)
388 call_rcu_lock(&call_rcu_mutex
);
389 if (cpu
< 0 || maxcpus
<= cpu
) {
391 fprintf(stderr
, "[error] liburcu: set CPU # out of range\n");
394 call_rcu_unlock(&call_rcu_mutex
);
398 alloc_cpu_call_rcu_data();
399 call_rcu_unlock(&call_rcu_mutex
);
400 if (per_cpu_call_rcu_data
== NULL
) {
404 per_cpu_call_rcu_data
[cpu
] = crdp
;
409 * Return a pointer to the default call_rcu_data structure, creating
410 * one if need be. Because we never free call_rcu_data structures,
411 * we don't need to be in an RCU read-side critical section.
414 struct call_rcu_data
*get_default_call_rcu_data(void)
416 if (default_call_rcu_data
!= NULL
)
417 return rcu_dereference(default_call_rcu_data
);
418 call_rcu_lock(&call_rcu_mutex
);
419 if (default_call_rcu_data
!= NULL
) {
420 call_rcu_unlock(&call_rcu_mutex
);
421 return default_call_rcu_data
;
423 call_rcu_data_init(&default_call_rcu_data
, 0, -1);
424 call_rcu_unlock(&call_rcu_mutex
);
425 return default_call_rcu_data
;
429 * Return the call_rcu_data structure that applies to the currently
430 * running thread. Any call_rcu_data structure assigned specifically
431 * to this thread has first priority, followed by any call_rcu_data
432 * structure assigned to the CPU on which the thread is running,
433 * followed by the default call_rcu_data structure. If there is not
434 * yet a default call_rcu_data structure, one will be created.
436 struct call_rcu_data
*get_call_rcu_data(void)
439 static int warned
= 0;
441 if (thread_call_rcu_data
!= NULL
)
442 return thread_call_rcu_data
;
444 return get_default_call_rcu_data();
445 curcpu
= sched_getcpu();
446 if (!warned
&& (curcpu
< 0 || maxcpus
<= curcpu
)) {
447 fprintf(stderr
, "[error] liburcu: gcrd CPU # out of range\n");
450 if (curcpu
>= 0 && maxcpus
> curcpu
&&
451 per_cpu_call_rcu_data
!= NULL
&&
452 per_cpu_call_rcu_data
[curcpu
] != NULL
)
453 return per_cpu_call_rcu_data
[curcpu
];
454 return get_default_call_rcu_data();
458 * Return a pointer to this task's call_rcu_data if there is one.
461 struct call_rcu_data
*get_thread_call_rcu_data(void)
463 return thread_call_rcu_data
;
467 * Set this task's call_rcu_data structure as specified, regardless
468 * of whether or not this task already had one. (This allows switching
469 * to and from real-time call_rcu threads, for example.)
471 * Use NULL to remove a thread's call_rcu_data structure, but it is
472 * the caller's responsibility to dispose of the removed structure.
473 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
474 * (prior to NULLing it out, of course).
477 void set_thread_call_rcu_data(struct call_rcu_data
*crdp
)
479 thread_call_rcu_data
= crdp
;
483 * Create a separate call_rcu thread for each CPU. This does not
484 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
485 * function if you want that behavior. Should be paired with
486 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
490 int create_all_cpu_call_rcu_data(unsigned long flags
)
493 struct call_rcu_data
*crdp
;
496 call_rcu_lock(&call_rcu_mutex
);
497 alloc_cpu_call_rcu_data();
498 call_rcu_unlock(&call_rcu_mutex
);
503 if (per_cpu_call_rcu_data
== NULL
) {
507 for (i
= 0; i
< maxcpus
; i
++) {
508 call_rcu_lock(&call_rcu_mutex
);
509 if (get_cpu_call_rcu_data(i
)) {
510 call_rcu_unlock(&call_rcu_mutex
);
513 crdp
= __create_call_rcu_data(flags
, i
);
515 call_rcu_unlock(&call_rcu_mutex
);
519 call_rcu_unlock(&call_rcu_mutex
);
520 if ((ret
= set_cpu_call_rcu_data(i
, crdp
)) != 0) {
521 /* FIXME: Leaks crdp for now. */
522 return ret
; /* Can happen on race. */
529 * Wake up the call_rcu thread corresponding to the specified
530 * call_rcu_data structure.
532 static void wake_call_rcu_thread(struct call_rcu_data
*crdp
)
534 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RT
))
535 call_rcu_wake_up(crdp
);
539 * Schedule a function to be invoked after a following grace period.
540 * This is the only function that must be called -- the others are
541 * only present to allow applications to tune their use of RCU for
542 * maximum performance.
544 * Note that unless a call_rcu thread has not already been created,
545 * the first invocation of call_rcu() will create one. So, if you
546 * need the first invocation of call_rcu() to be fast, make sure
547 * to create a call_rcu thread first. One way to accomplish this is
548 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
551 void call_rcu(struct rcu_head
*head
,
552 void (*func
)(struct rcu_head
*head
))
554 struct call_rcu_data
*crdp
;
556 cds_wfq_node_init(&head
->next
);
558 crdp
= get_call_rcu_data();
559 cds_wfq_enqueue(&crdp
->cbs
, &head
->next
);
560 uatomic_inc(&crdp
->qlen
);
561 wake_call_rcu_thread(crdp
);
565 * Free up the specified call_rcu_data structure, terminating the
566 * associated call_rcu thread. The caller must have previously
567 * removed the call_rcu_data structure from per-thread or per-CPU
568 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
569 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
570 * per-thread call_rcu_data structures.
572 * We silently refuse to free up the default call_rcu_data structure
573 * because that is where we put any leftover callbacks. Note that
574 * the possibility of self-spawning callbacks makes it impossible
575 * to execute all the callbacks in finite time without putting any
576 * newly spawned callbacks somewhere else. The "somewhere else" of
577 * last resort is the default call_rcu_data structure.
579 * We also silently refuse to free NULL pointers. This simplifies
582 void call_rcu_data_free(struct call_rcu_data
*crdp
)
584 struct cds_wfq_node
*cbs
;
585 struct cds_wfq_node
**cbs_tail
;
586 struct cds_wfq_node
**cbs_endprev
;
588 if (crdp
== NULL
|| crdp
== default_call_rcu_data
) {
591 if ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0) {
592 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOP
);
593 wake_call_rcu_thread(crdp
);
594 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0)
597 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
598 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
600 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
601 cbs_tail
= (struct cds_wfq_node
**)
602 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
603 cbs_endprev
= (struct cds_wfq_node
**)
604 uatomic_xchg(&default_call_rcu_data
, cbs_tail
);
606 uatomic_add(&default_call_rcu_data
->qlen
,
607 uatomic_read(&crdp
->qlen
));
608 cds_list_del(&crdp
->list
);
614 * Clean up all the per-CPU call_rcu threads.
616 void free_all_cpu_call_rcu_data(void)
619 struct call_rcu_data
*crdp
;
623 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
624 crdp
= get_cpu_call_rcu_data(cpu
);
627 set_cpu_call_rcu_data(cpu
, NULL
);
628 call_rcu_data_free(crdp
);
633 * Acquire the call_rcu_mutex in order to ensure that the child sees
634 * all of the call_rcu() data structures in a consistent state.
635 * Suitable for pthread_atfork() and friends.
637 void call_rcu_before_fork(void)
639 call_rcu_lock(&call_rcu_mutex
);
643 * Clean up call_rcu data structures in the parent of a successful fork()
644 * that is not followed by exec() in the child. Suitable for
645 * pthread_atfork() and friends.
647 void call_rcu_after_fork_parent(void)
649 call_rcu_unlock(&call_rcu_mutex
);
653 * Clean up call_rcu data structures in the child of a successful fork()
654 * that is not followed by exec(). Suitable for pthread_atfork() and
657 void call_rcu_after_fork_child(void)
659 struct call_rcu_data
*crdp
;
661 /* Release the mutex. */
662 call_rcu_unlock(&call_rcu_mutex
);
665 * Allocate a new default call_rcu_data structure in order
666 * to get a working call_rcu thread to go with it.
668 default_call_rcu_data
= NULL
;
669 (void)get_default_call_rcu_data();
671 /* Dispose of all of the rest of the call_rcu_data structures. */
672 while (call_rcu_data_list
.next
!= call_rcu_data_list
.prev
) {
673 crdp
= cds_list_entry(call_rcu_data_list
.prev
,
674 struct call_rcu_data
, list
);
675 if (crdp
== default_call_rcu_data
)
676 crdp
= cds_list_entry(crdp
->list
.prev
,
677 struct call_rcu_data
, list
);
678 uatomic_set(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
679 call_rcu_data_free(crdp
);