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/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 static void call_rcu_wait(struct call_rcu_data
*crdp
)
93 /* Read call_rcu list before read futex */
95 if (uatomic_read(&crdp
->futex
) == -1)
96 futex_async(&crdp
->futex
, FUTEX_WAIT
, -1,
100 static void call_rcu_wake_up(struct call_rcu_data
*crdp
)
102 /* Write to call_rcu list before reading/writing futex */
104 if (unlikely(uatomic_read(&crdp
->futex
) == -1)) {
105 uatomic_set(&crdp
->futex
, 0);
106 futex_async(&crdp
->futex
, FUTEX_WAKE
, 1,
111 /* Allocate the array if it has not already been allocated. */
113 static void alloc_cpu_call_rcu_data(void)
115 struct call_rcu_data
**p
;
116 static int warned
= 0;
120 maxcpus
= sysconf(_SC_NPROCESSORS_CONF
);
124 p
= malloc(maxcpus
* sizeof(*per_cpu_call_rcu_data
));
126 memset(p
, '\0', maxcpus
* sizeof(*per_cpu_call_rcu_data
));
127 per_cpu_call_rcu_data
= p
;
130 fprintf(stderr
, "[error] liburcu: unable to allocate per-CPU pointer array\n");
136 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
138 static const struct call_rcu_data
**per_cpu_call_rcu_data
= NULL
;
139 static const long maxcpus
= -1;
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
)
156 if (pthread_mutex_lock(pmp
) != 0) {
157 perror("pthread_mutex_lock");
162 /* Release the specified pthread mutex. */
164 static void call_rcu_unlock(pthread_mutex_t
*pmp
)
166 if (pthread_mutex_unlock(pmp
) != 0) {
167 perror("pthread_mutex_unlock");
172 #if HAVE_SCHED_SETAFFINITY
174 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
178 if (crdp
->cpu_affinity
< 0)
182 CPU_SET(crdp
->cpu_affinity
, &mask
);
183 #if SCHED_SETAFFINITY_ARGS == 2
184 return sched_setaffinity(0, &mask
);
186 return sched_setaffinity(0, sizeof(mask
), &mask
);
191 int set_thread_cpu_affinity(struct call_rcu_data
*crdp
)
197 /* This is the code run by each call_rcu thread. */
199 static void *call_rcu_thread(void *arg
)
201 unsigned long cbcount
;
202 struct cds_wfq_node
*cbs
;
203 struct cds_wfq_node
**cbs_tail
;
204 struct call_rcu_data
*crdp
= (struct call_rcu_data
*)arg
;
205 struct rcu_head
*rhp
;
206 int rt
= !!(uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_RT
);
208 if (set_thread_cpu_affinity(crdp
) != 0) {
209 perror("pthread_setaffinity_np");
213 thread_call_rcu_data
= crdp
;
215 uatomic_dec(&crdp
->futex
);
216 /* Decrement futex before reading call_rcu list */
220 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
221 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
223 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
224 cbs_tail
= (struct cds_wfq_node
**)
225 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
229 while (cbs
->next
== NULL
&&
230 &cbs
->next
!= cbs_tail
)
232 if (cbs
== &crdp
->cbs
.dummy
) {
236 rhp
= (struct rcu_head
*)cbs
;
240 } while (cbs
!= NULL
);
241 uatomic_sub(&crdp
->qlen
, cbcount
);
243 if (uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOP
)
247 == _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
250 uatomic_dec(&crdp
->futex
);
252 * Decrement futex before reading
263 * Read call_rcu list before write futex.
266 uatomic_set(&crdp
->futex
, 0);
268 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
273 * Create both a call_rcu thread and the corresponding call_rcu_data
274 * structure, linking the structure in as specified. Caller must hold
278 static void call_rcu_data_init(struct call_rcu_data
**crdpp
,
282 struct call_rcu_data
*crdp
;
284 crdp
= malloc(sizeof(*crdp
));
286 fprintf(stderr
, "Out of memory.\n");
289 memset(crdp
, '\0', sizeof(*crdp
));
290 cds_wfq_init(&crdp
->cbs
);
294 cds_list_add(&crdp
->list
, &call_rcu_data_list
);
295 crdp
->cpu_affinity
= cpu_affinity
;
296 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
298 if (pthread_create(&crdp
->tid
, NULL
, call_rcu_thread
, crdp
) != 0) {
299 perror("pthread_create");
305 * Return a pointer to the call_rcu_data structure for the specified
306 * CPU, returning NULL if there is none. We cannot automatically
307 * created it because the platform we are running on might not define
311 struct call_rcu_data
*get_cpu_call_rcu_data(int cpu
)
313 static int warned
= 0;
315 if (per_cpu_call_rcu_data
== NULL
)
317 if (!warned
&& maxcpus
> 0 && (cpu
< 0 || maxcpus
<= cpu
)) {
318 fprintf(stderr
, "[error] liburcu: get CPU # out of range\n");
321 if (cpu
< 0 || maxcpus
<= cpu
)
323 return per_cpu_call_rcu_data
[cpu
];
327 * Return the tid corresponding to the call_rcu thread whose
328 * call_rcu_data structure is specified.
331 pthread_t
get_call_rcu_thread(struct call_rcu_data
*crdp
)
337 * Create a call_rcu_data structure (with thread) and return a pointer.
340 static struct call_rcu_data
*__create_call_rcu_data(unsigned long flags
,
343 struct call_rcu_data
*crdp
;
345 call_rcu_data_init(&crdp
, flags
, cpu_affinity
);
349 struct call_rcu_data
*create_call_rcu_data(unsigned long flags
,
352 struct call_rcu_data
*crdp
;
354 call_rcu_lock(&call_rcu_mutex
);
355 crdp
= __create_call_rcu_data(flags
, cpu_affinity
);
356 call_rcu_unlock(&call_rcu_mutex
);
361 * Set the specified CPU to use the specified call_rcu_data structure.
363 * Use NULL to remove a CPU's call_rcu_data structure, but it is
364 * the caller's responsibility to dispose of the removed structure.
365 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
366 * (prior to NULLing it out, of course).
369 int set_cpu_call_rcu_data(int cpu
, struct call_rcu_data
*crdp
)
373 call_rcu_lock(&call_rcu_mutex
);
374 if (cpu
< 0 || maxcpus
<= cpu
) {
376 fprintf(stderr
, "[error] liburcu: set CPU # out of range\n");
379 call_rcu_unlock(&call_rcu_mutex
);
383 alloc_cpu_call_rcu_data();
384 call_rcu_unlock(&call_rcu_mutex
);
385 if (per_cpu_call_rcu_data
== NULL
) {
389 per_cpu_call_rcu_data
[cpu
] = crdp
;
394 * Return a pointer to the default call_rcu_data structure, creating
395 * one if need be. Because we never free call_rcu_data structures,
396 * we don't need to be in an RCU read-side critical section.
399 struct call_rcu_data
*get_default_call_rcu_data(void)
401 if (default_call_rcu_data
!= NULL
)
402 return rcu_dereference(default_call_rcu_data
);
403 call_rcu_lock(&call_rcu_mutex
);
404 if (default_call_rcu_data
!= NULL
) {
405 call_rcu_unlock(&call_rcu_mutex
);
406 return default_call_rcu_data
;
408 call_rcu_data_init(&default_call_rcu_data
, 0, -1);
409 call_rcu_unlock(&call_rcu_mutex
);
410 return default_call_rcu_data
;
414 * Return the call_rcu_data structure that applies to the currently
415 * running thread. Any call_rcu_data structure assigned specifically
416 * to this thread has first priority, followed by any call_rcu_data
417 * structure assigned to the CPU on which the thread is running,
418 * followed by the default call_rcu_data structure. If there is not
419 * yet a default call_rcu_data structure, one will be created.
421 struct call_rcu_data
*get_call_rcu_data(void)
424 static int warned
= 0;
426 if (thread_call_rcu_data
!= NULL
)
427 return thread_call_rcu_data
;
429 return get_default_call_rcu_data();
430 curcpu
= sched_getcpu();
431 if (!warned
&& (curcpu
< 0 || maxcpus
<= curcpu
)) {
432 fprintf(stderr
, "[error] liburcu: gcrd CPU # out of range\n");
435 if (curcpu
>= 0 && maxcpus
> curcpu
&&
436 per_cpu_call_rcu_data
!= NULL
&&
437 per_cpu_call_rcu_data
[curcpu
] != NULL
)
438 return per_cpu_call_rcu_data
[curcpu
];
439 return get_default_call_rcu_data();
443 * Return a pointer to this task's call_rcu_data if there is one.
446 struct call_rcu_data
*get_thread_call_rcu_data(void)
448 return thread_call_rcu_data
;
452 * Set this task's call_rcu_data structure as specified, regardless
453 * of whether or not this task already had one. (This allows switching
454 * to and from real-time call_rcu threads, for example.)
456 * Use NULL to remove a thread's call_rcu_data structure, but it is
457 * the caller's responsibility to dispose of the removed structure.
458 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
459 * (prior to NULLing it out, of course).
462 void set_thread_call_rcu_data(struct call_rcu_data
*crdp
)
464 thread_call_rcu_data
= crdp
;
468 * Create a separate call_rcu thread for each CPU. This does not
469 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
470 * function if you want that behavior.
473 int create_all_cpu_call_rcu_data(unsigned long flags
)
476 struct call_rcu_data
*crdp
;
479 call_rcu_lock(&call_rcu_mutex
);
480 alloc_cpu_call_rcu_data();
481 call_rcu_unlock(&call_rcu_mutex
);
486 if (per_cpu_call_rcu_data
== NULL
) {
490 for (i
= 0; i
< maxcpus
; i
++) {
491 call_rcu_lock(&call_rcu_mutex
);
492 if (get_cpu_call_rcu_data(i
)) {
493 call_rcu_unlock(&call_rcu_mutex
);
496 crdp
= __create_call_rcu_data(flags
, i
);
498 call_rcu_unlock(&call_rcu_mutex
);
502 call_rcu_unlock(&call_rcu_mutex
);
503 if ((ret
= set_cpu_call_rcu_data(i
, crdp
)) != 0) {
504 /* FIXME: Leaks crdp for now. */
505 return ret
; /* Can happen on race. */
512 * Wake up the call_rcu thread corresponding to the specified
513 * call_rcu_data structure.
515 static void wake_call_rcu_thread(struct call_rcu_data
*crdp
)
517 if (!(_CMM_LOAD_SHARED(crdp
->flags
) & URCU_CALL_RCU_RT
))
518 call_rcu_wake_up(crdp
);
522 * Schedule a function to be invoked after a following grace period.
523 * This is the only function that must be called -- the others are
524 * only present to allow applications to tune their use of RCU for
525 * maximum performance.
527 * Note that unless a call_rcu thread has not already been created,
528 * the first invocation of call_rcu() will create one. So, if you
529 * need the first invocation of call_rcu() to be fast, make sure
530 * to create a call_rcu thread first. One way to accomplish this is
531 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
534 void call_rcu(struct rcu_head
*head
,
535 void (*func
)(struct rcu_head
*head
))
537 struct call_rcu_data
*crdp
;
539 cds_wfq_node_init(&head
->next
);
541 crdp
= get_call_rcu_data();
542 cds_wfq_enqueue(&crdp
->cbs
, &head
->next
);
543 uatomic_inc(&crdp
->qlen
);
544 wake_call_rcu_thread(crdp
);
548 * Free up the specified call_rcu_data structure, terminating the
549 * associated call_rcu thread. The caller must have previously
550 * removed the call_rcu_data structure from per-thread or per-CPU
551 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
552 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
553 * per-thread call_rcu_data structures.
555 * We silently refuse to free up the default call_rcu_data structure
556 * because that is where we put any leftover callbacks. Note that
557 * the possibility of self-spawning callbacks makes it impossible
558 * to execute all the callbacks in finite time without putting any
559 * newly spawned callbacks somewhere else. The "somewhere else" of
560 * last resort is the default call_rcu_data structure.
562 * We also silently refuse to free NULL pointers. This simplifies
565 void call_rcu_data_free(struct call_rcu_data
*crdp
)
567 struct cds_wfq_node
*cbs
;
568 struct cds_wfq_node
**cbs_tail
;
569 struct cds_wfq_node
**cbs_endprev
;
571 if (crdp
== NULL
|| crdp
== default_call_rcu_data
) {
574 if ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0) {
575 uatomic_or(&crdp
->flags
, URCU_CALL_RCU_STOP
);
576 wake_call_rcu_thread(crdp
);
577 while ((uatomic_read(&crdp
->flags
) & URCU_CALL_RCU_STOPPED
) == 0)
580 if (&crdp
->cbs
.head
!= _CMM_LOAD_SHARED(crdp
->cbs
.tail
)) {
581 while ((cbs
= _CMM_LOAD_SHARED(crdp
->cbs
.head
)) == NULL
)
583 _CMM_STORE_SHARED(crdp
->cbs
.head
, NULL
);
584 cbs_tail
= (struct cds_wfq_node
**)
585 uatomic_xchg(&crdp
->cbs
.tail
, &crdp
->cbs
.head
);
586 cbs_endprev
= (struct cds_wfq_node
**)
587 uatomic_xchg(&default_call_rcu_data
, cbs_tail
);
589 uatomic_add(&default_call_rcu_data
->qlen
,
590 uatomic_read(&crdp
->qlen
));
591 cds_list_del(&crdp
->list
);
597 * Clean up all the per-CPU call_rcu threads.
599 void free_all_cpu_call_rcu_data(void)
602 struct call_rcu_data
*crdp
;
606 for (cpu
= 0; cpu
< maxcpus
; cpu
++) {
607 crdp
= get_cpu_call_rcu_data(cpu
);
610 set_cpu_call_rcu_data(cpu
, NULL
);
611 call_rcu_data_free(crdp
);
616 * Acquire the call_rcu_mutex in order to ensure that the child sees
617 * all of the call_rcu() data structures in a consistent state.
618 * Suitable for pthread_atfork() and friends.
620 void call_rcu_before_fork(void)
622 call_rcu_lock(&call_rcu_mutex
);
626 * Clean up call_rcu data structures in the parent of a successful fork()
627 * that is not followed by exec() in the child. Suitable for
628 * pthread_atfork() and friends.
630 void call_rcu_after_fork_parent(void)
632 call_rcu_unlock(&call_rcu_mutex
);
636 * Clean up call_rcu data structures in the child of a successful fork()
637 * that is not followed by exec(). Suitable for pthread_atfork() and
640 void call_rcu_after_fork_child(void)
642 struct call_rcu_data
*crdp
;
644 /* Release the mutex. */
645 call_rcu_unlock(&call_rcu_mutex
);
648 * Allocate a new default call_rcu_data structure in order
649 * to get a working call_rcu thread to go with it.
651 default_call_rcu_data
= NULL
;
652 (void)get_default_call_rcu_data();
654 /* Dispose of all of the rest of the call_rcu_data structures. */
655 while (call_rcu_data_list
.next
!= call_rcu_data_list
.prev
) {
656 crdp
= cds_list_entry(call_rcu_data_list
.prev
,
657 struct call_rcu_data
, list
);
658 if (crdp
== default_call_rcu_data
)
659 crdp
= cds_list_entry(crdp
->list
.prev
,
660 struct call_rcu_data
, list
);
661 uatomic_set(&crdp
->flags
, URCU_CALL_RCU_STOPPED
);
662 call_rcu_data_free(crdp
);