Port: Add Solaris getcpu support
[urcu.git] / urcu-call-rcu-impl.h
1 /*
2 * urcu-call-rcu.c
3 *
4 * Userspace RCU library - batch memory reclamation with kernel API
5 *
6 * Copyright (c) 2010 Paul E. McKenney <paulmck@linux.vnet.ibm.com>
7 *
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.
12 *
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.
17 *
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
21 */
22
23 #define _GNU_SOURCE
24 #define _LGPL_SOURCE
25 #include <stdio.h>
26 #include <pthread.h>
27 #include <signal.h>
28 #include <assert.h>
29 #include <stdlib.h>
30 #include <stdint.h>
31 #include <string.h>
32 #include <errno.h>
33 #include <poll.h>
34 #include <sys/time.h>
35 #include <unistd.h>
36 #include <sched.h>
37
38 #include "config.h"
39 #include "compat-getcpu.h"
40 #include "urcu/wfcqueue.h"
41 #include "urcu-call-rcu.h"
42 #include "urcu-pointer.h"
43 #include "urcu/list.h"
44 #include "urcu/futex.h"
45 #include "urcu/tls-compat.h"
46 #include "urcu/ref.h"
47 #include "urcu-die.h"
48
49 #define SET_AFFINITY_CHECK_PERIOD (1U << 8) /* 256 */
50 #define SET_AFFINITY_CHECK_PERIOD_MASK (SET_AFFINITY_CHECK_PERIOD - 1)
51
52 /* Data structure that identifies a call_rcu thread. */
53
54 struct call_rcu_data {
55 /*
56 * We do not align head on a different cache-line than tail
57 * mainly because call_rcu callback-invocation threads use
58 * batching ("splice") to get an entire list of callbacks, which
59 * effectively empties the queue, and requires to touch the tail
60 * anyway.
61 */
62 struct cds_wfcq_tail cbs_tail;
63 struct cds_wfcq_head cbs_head;
64 unsigned long flags;
65 int32_t futex;
66 unsigned long qlen; /* maintained for debugging. */
67 pthread_t tid;
68 int cpu_affinity;
69 unsigned long gp_count;
70 struct cds_list_head list;
71 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
72
73 struct call_rcu_completion {
74 int barrier_count;
75 int32_t futex;
76 struct urcu_ref ref;
77 };
78
79 struct call_rcu_completion_work {
80 struct rcu_head head;
81 struct call_rcu_completion *completion;
82 };
83
84 /*
85 * List of all call_rcu_data structures to keep valgrind happy.
86 * Protected by call_rcu_mutex.
87 */
88
89 static CDS_LIST_HEAD(call_rcu_data_list);
90
91 /* Link a thread using call_rcu() to its call_rcu thread. */
92
93 static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data);
94
95 /*
96 * Guard call_rcu thread creation and atfork handlers.
97 */
98 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
99
100 /* If a given thread does not have its own call_rcu thread, this is default. */
101
102 static struct call_rcu_data *default_call_rcu_data;
103
104 /*
105 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
106 * available, then we can have call_rcu threads assigned to individual
107 * CPUs rather than only to specific threads.
108 */
109
110 #if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID))
111
112 /*
113 * Pointer to array of pointers to per-CPU call_rcu_data structures
114 * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
115 * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
116 * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
117 * without mutex. The call_rcu_mutex protects updates.
118 */
119
120 static struct call_rcu_data **per_cpu_call_rcu_data;
121 static long maxcpus;
122
123 static void maxcpus_reset(void)
124 {
125 maxcpus = 0;
126 }
127
128 /* Allocate the array if it has not already been allocated. */
129
130 static void alloc_cpu_call_rcu_data(void)
131 {
132 struct call_rcu_data **p;
133 static int warned = 0;
134
135 if (maxcpus != 0)
136 return;
137 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
138 if (maxcpus <= 0) {
139 return;
140 }
141 p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data));
142 if (p != NULL) {
143 memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data));
144 rcu_set_pointer(&per_cpu_call_rcu_data, p);
145 } else {
146 if (!warned) {
147 fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
148 }
149 warned = 1;
150 }
151 }
152
153 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
154
155 /*
156 * per_cpu_call_rcu_data should be constant, but some functions below, used both
157 * for cases where cpu number is available and not available, assume it it not
158 * constant.
159 */
160 static struct call_rcu_data **per_cpu_call_rcu_data = NULL;
161 static const long maxcpus = -1;
162
163 static void maxcpus_reset(void)
164 {
165 }
166
167 static void alloc_cpu_call_rcu_data(void)
168 {
169 }
170
171 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
172
173 /* Acquire the specified pthread mutex. */
174
175 static void call_rcu_lock(pthread_mutex_t *pmp)
176 {
177 int ret;
178
179 ret = pthread_mutex_lock(pmp);
180 if (ret)
181 urcu_die(ret);
182 }
183
184 /* Release the specified pthread mutex. */
185
186 static void call_rcu_unlock(pthread_mutex_t *pmp)
187 {
188 int ret;
189
190 ret = pthread_mutex_unlock(pmp);
191 if (ret)
192 urcu_die(ret);
193 }
194
195 /*
196 * Periodically retry setting CPU affinity if we migrate.
197 * Losing affinity can be caused by CPU hotunplug/hotplug, or by
198 * cpuset(7).
199 */
200 #if HAVE_SCHED_SETAFFINITY
201 static
202 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
203 {
204 cpu_set_t mask;
205 int ret;
206
207 if (crdp->cpu_affinity < 0)
208 return 0;
209 if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK)
210 return 0;
211 if (urcu_sched_getcpu() == crdp->cpu_affinity)
212 return 0;
213
214 CPU_ZERO(&mask);
215 CPU_SET(crdp->cpu_affinity, &mask);
216 #if SCHED_SETAFFINITY_ARGS == 2
217 ret = sched_setaffinity(0, &mask);
218 #else
219 ret = sched_setaffinity(0, sizeof(mask), &mask);
220 #endif
221 /*
222 * EINVAL is fine: can be caused by hotunplugged CPUs, or by
223 * cpuset(7). This is why we should always retry if we detect
224 * migration.
225 */
226 if (ret && errno == EINVAL) {
227 ret = 0;
228 errno = 0;
229 }
230 return ret;
231 }
232 #else
233 static
234 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
235 {
236 return 0;
237 }
238 #endif
239
240 static void call_rcu_wait(struct call_rcu_data *crdp)
241 {
242 /* Read call_rcu list before read futex */
243 cmm_smp_mb();
244 if (uatomic_read(&crdp->futex) != -1)
245 return;
246 while (futex_async(&crdp->futex, FUTEX_WAIT, -1,
247 NULL, NULL, 0)) {
248 switch (errno) {
249 case EWOULDBLOCK:
250 /* Value already changed. */
251 return;
252 case EINTR:
253 /* Retry if interrupted by signal. */
254 break; /* Get out of switch. */
255 default:
256 /* Unexpected error. */
257 urcu_die(errno);
258 }
259 }
260 }
261
262 static void call_rcu_wake_up(struct call_rcu_data *crdp)
263 {
264 /* Write to call_rcu list before reading/writing futex */
265 cmm_smp_mb();
266 if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
267 uatomic_set(&crdp->futex, 0);
268 if (futex_async(&crdp->futex, FUTEX_WAKE, 1,
269 NULL, NULL, 0) < 0)
270 urcu_die(errno);
271 }
272 }
273
274 static void call_rcu_completion_wait(struct call_rcu_completion *completion)
275 {
276 /* Read completion barrier count before read futex */
277 cmm_smp_mb();
278 if (uatomic_read(&completion->futex) != -1)
279 return;
280 while (futex_async(&completion->futex, FUTEX_WAIT, -1,
281 NULL, NULL, 0)) {
282 switch (errno) {
283 case EWOULDBLOCK:
284 /* Value already changed. */
285 return;
286 case EINTR:
287 /* Retry if interrupted by signal. */
288 break; /* Get out of switch. */
289 default:
290 /* Unexpected error. */
291 urcu_die(errno);
292 }
293 }
294 }
295
296 static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
297 {
298 /* Write to completion barrier count before reading/writing futex */
299 cmm_smp_mb();
300 if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
301 uatomic_set(&completion->futex, 0);
302 if (futex_async(&completion->futex, FUTEX_WAKE, 1,
303 NULL, NULL, 0) < 0)
304 urcu_die(errno);
305 }
306 }
307
308 /* This is the code run by each call_rcu thread. */
309
310 static void *call_rcu_thread(void *arg)
311 {
312 unsigned long cbcount;
313 struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
314 int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
315
316 if (set_thread_cpu_affinity(crdp))
317 urcu_die(errno);
318
319 /*
320 * If callbacks take a read-side lock, we need to be registered.
321 */
322 rcu_register_thread();
323
324 URCU_TLS(thread_call_rcu_data) = crdp;
325 if (!rt) {
326 uatomic_dec(&crdp->futex);
327 /* Decrement futex before reading call_rcu list */
328 cmm_smp_mb();
329 }
330 for (;;) {
331 struct cds_wfcq_head cbs_tmp_head;
332 struct cds_wfcq_tail cbs_tmp_tail;
333 struct cds_wfcq_node *cbs, *cbs_tmp_n;
334 enum cds_wfcq_ret splice_ret;
335
336 if (set_thread_cpu_affinity(crdp))
337 urcu_die(errno);
338
339 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
340 /*
341 * Pause requested. Become quiescent: remove
342 * ourself from all global lists, and don't
343 * process any callback. The callback lists may
344 * still be non-empty though.
345 */
346 rcu_unregister_thread();
347 cmm_smp_mb__before_uatomic_or();
348 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
349 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
350 (void) poll(NULL, 0, 1);
351 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
352 cmm_smp_mb__after_uatomic_and();
353 rcu_register_thread();
354 }
355
356 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
357 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
358 &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail);
359 assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
360 assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
361 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
362 synchronize_rcu();
363 cbcount = 0;
364 __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
365 &cbs_tmp_tail, cbs, cbs_tmp_n) {
366 struct rcu_head *rhp;
367
368 rhp = caa_container_of(cbs,
369 struct rcu_head, next);
370 rhp->func(rhp);
371 cbcount++;
372 }
373 uatomic_sub(&crdp->qlen, cbcount);
374 }
375 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
376 break;
377 rcu_thread_offline();
378 if (!rt) {
379 if (cds_wfcq_empty(&crdp->cbs_head,
380 &crdp->cbs_tail)) {
381 call_rcu_wait(crdp);
382 (void) poll(NULL, 0, 10);
383 uatomic_dec(&crdp->futex);
384 /*
385 * Decrement futex before reading
386 * call_rcu list.
387 */
388 cmm_smp_mb();
389 } else {
390 (void) poll(NULL, 0, 10);
391 }
392 } else {
393 (void) poll(NULL, 0, 10);
394 }
395 rcu_thread_online();
396 }
397 if (!rt) {
398 /*
399 * Read call_rcu list before write futex.
400 */
401 cmm_smp_mb();
402 uatomic_set(&crdp->futex, 0);
403 }
404 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
405 rcu_unregister_thread();
406 return NULL;
407 }
408
409 /*
410 * Create both a call_rcu thread and the corresponding call_rcu_data
411 * structure, linking the structure in as specified. Caller must hold
412 * call_rcu_mutex.
413 */
414
415 static void call_rcu_data_init(struct call_rcu_data **crdpp,
416 unsigned long flags,
417 int cpu_affinity)
418 {
419 struct call_rcu_data *crdp;
420 int ret;
421
422 crdp = malloc(sizeof(*crdp));
423 if (crdp == NULL)
424 urcu_die(errno);
425 memset(crdp, '\0', sizeof(*crdp));
426 cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail);
427 crdp->qlen = 0;
428 crdp->futex = 0;
429 crdp->flags = flags;
430 cds_list_add(&crdp->list, &call_rcu_data_list);
431 crdp->cpu_affinity = cpu_affinity;
432 crdp->gp_count = 0;
433 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
434 *crdpp = crdp;
435 ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
436 if (ret)
437 urcu_die(ret);
438 }
439
440 /*
441 * Return a pointer to the call_rcu_data structure for the specified
442 * CPU, returning NULL if there is none. We cannot automatically
443 * created it because the platform we are running on might not define
444 * urcu_sched_getcpu().
445 *
446 * The call to this function and use of the returned call_rcu_data
447 * should be protected by RCU read-side lock.
448 */
449
450 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
451 {
452 static int warned = 0;
453 struct call_rcu_data **pcpu_crdp;
454
455 pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data);
456 if (pcpu_crdp == NULL)
457 return NULL;
458 if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
459 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
460 warned = 1;
461 }
462 if (cpu < 0 || maxcpus <= cpu)
463 return NULL;
464 return rcu_dereference(pcpu_crdp[cpu]);
465 }
466
467 /*
468 * Return the tid corresponding to the call_rcu thread whose
469 * call_rcu_data structure is specified.
470 */
471
472 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
473 {
474 return crdp->tid;
475 }
476
477 /*
478 * Create a call_rcu_data structure (with thread) and return a pointer.
479 */
480
481 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
482 int cpu_affinity)
483 {
484 struct call_rcu_data *crdp;
485
486 call_rcu_data_init(&crdp, flags, cpu_affinity);
487 return crdp;
488 }
489
490 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
491 int cpu_affinity)
492 {
493 struct call_rcu_data *crdp;
494
495 call_rcu_lock(&call_rcu_mutex);
496 crdp = __create_call_rcu_data(flags, cpu_affinity);
497 call_rcu_unlock(&call_rcu_mutex);
498 return crdp;
499 }
500
501 /*
502 * Set the specified CPU to use the specified call_rcu_data structure.
503 *
504 * Use NULL to remove a CPU's call_rcu_data structure, but it is
505 * the caller's responsibility to dispose of the removed structure.
506 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
507 * (prior to NULLing it out, of course).
508 *
509 * The caller must wait for a grace-period to pass between return from
510 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
511 * previous call rcu data as argument.
512 */
513
514 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
515 {
516 static int warned = 0;
517
518 call_rcu_lock(&call_rcu_mutex);
519 alloc_cpu_call_rcu_data();
520 if (cpu < 0 || maxcpus <= cpu) {
521 if (!warned) {
522 fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
523 warned = 1;
524 }
525 call_rcu_unlock(&call_rcu_mutex);
526 errno = EINVAL;
527 return -EINVAL;
528 }
529
530 if (per_cpu_call_rcu_data == NULL) {
531 call_rcu_unlock(&call_rcu_mutex);
532 errno = ENOMEM;
533 return -ENOMEM;
534 }
535
536 if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) {
537 call_rcu_unlock(&call_rcu_mutex);
538 errno = EEXIST;
539 return -EEXIST;
540 }
541
542 rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp);
543 call_rcu_unlock(&call_rcu_mutex);
544 return 0;
545 }
546
547 /*
548 * Return a pointer to the default call_rcu_data structure, creating
549 * one if need be. Because we never free call_rcu_data structures,
550 * we don't need to be in an RCU read-side critical section.
551 */
552
553 struct call_rcu_data *get_default_call_rcu_data(void)
554 {
555 if (default_call_rcu_data != NULL)
556 return rcu_dereference(default_call_rcu_data);
557 call_rcu_lock(&call_rcu_mutex);
558 if (default_call_rcu_data != NULL) {
559 call_rcu_unlock(&call_rcu_mutex);
560 return default_call_rcu_data;
561 }
562 call_rcu_data_init(&default_call_rcu_data, 0, -1);
563 call_rcu_unlock(&call_rcu_mutex);
564 return default_call_rcu_data;
565 }
566
567 /*
568 * Return the call_rcu_data structure that applies to the currently
569 * running thread. Any call_rcu_data structure assigned specifically
570 * to this thread has first priority, followed by any call_rcu_data
571 * structure assigned to the CPU on which the thread is running,
572 * followed by the default call_rcu_data structure. If there is not
573 * yet a default call_rcu_data structure, one will be created.
574 *
575 * Calls to this function and use of the returned call_rcu_data should
576 * be protected by RCU read-side lock.
577 */
578 struct call_rcu_data *get_call_rcu_data(void)
579 {
580 struct call_rcu_data *crd;
581
582 if (URCU_TLS(thread_call_rcu_data) != NULL)
583 return URCU_TLS(thread_call_rcu_data);
584
585 if (maxcpus > 0) {
586 crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
587 if (crd)
588 return crd;
589 }
590
591 return get_default_call_rcu_data();
592 }
593
594 /*
595 * Return a pointer to this task's call_rcu_data if there is one.
596 */
597
598 struct call_rcu_data *get_thread_call_rcu_data(void)
599 {
600 return URCU_TLS(thread_call_rcu_data);
601 }
602
603 /*
604 * Set this task's call_rcu_data structure as specified, regardless
605 * of whether or not this task already had one. (This allows switching
606 * to and from real-time call_rcu threads, for example.)
607 *
608 * Use NULL to remove a thread's call_rcu_data structure, but it is
609 * the caller's responsibility to dispose of the removed structure.
610 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
611 * (prior to NULLing it out, of course).
612 */
613
614 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
615 {
616 URCU_TLS(thread_call_rcu_data) = crdp;
617 }
618
619 /*
620 * Create a separate call_rcu thread for each CPU. This does not
621 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
622 * function if you want that behavior. Should be paired with
623 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
624 * threads.
625 */
626
627 int create_all_cpu_call_rcu_data(unsigned long flags)
628 {
629 int i;
630 struct call_rcu_data *crdp;
631 int ret;
632
633 call_rcu_lock(&call_rcu_mutex);
634 alloc_cpu_call_rcu_data();
635 call_rcu_unlock(&call_rcu_mutex);
636 if (maxcpus <= 0) {
637 errno = EINVAL;
638 return -EINVAL;
639 }
640 if (per_cpu_call_rcu_data == NULL) {
641 errno = ENOMEM;
642 return -ENOMEM;
643 }
644 for (i = 0; i < maxcpus; i++) {
645 call_rcu_lock(&call_rcu_mutex);
646 if (get_cpu_call_rcu_data(i)) {
647 call_rcu_unlock(&call_rcu_mutex);
648 continue;
649 }
650 crdp = __create_call_rcu_data(flags, i);
651 if (crdp == NULL) {
652 call_rcu_unlock(&call_rcu_mutex);
653 errno = ENOMEM;
654 return -ENOMEM;
655 }
656 call_rcu_unlock(&call_rcu_mutex);
657 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
658 call_rcu_data_free(crdp);
659
660 /* it has been created by other thread */
661 if (ret == -EEXIST)
662 continue;
663
664 return ret;
665 }
666 }
667 return 0;
668 }
669
670 /*
671 * Wake up the call_rcu thread corresponding to the specified
672 * call_rcu_data structure.
673 */
674 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
675 {
676 if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
677 call_rcu_wake_up(crdp);
678 }
679
680 static void _call_rcu(struct rcu_head *head,
681 void (*func)(struct rcu_head *head),
682 struct call_rcu_data *crdp)
683 {
684 cds_wfcq_node_init(&head->next);
685 head->func = func;
686 cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
687 uatomic_inc(&crdp->qlen);
688 wake_call_rcu_thread(crdp);
689 }
690
691 /*
692 * Schedule a function to be invoked after a following grace period.
693 * This is the only function that must be called -- the others are
694 * only present to allow applications to tune their use of RCU for
695 * maximum performance.
696 *
697 * Note that unless a call_rcu thread has not already been created,
698 * the first invocation of call_rcu() will create one. So, if you
699 * need the first invocation of call_rcu() to be fast, make sure
700 * to create a call_rcu thread first. One way to accomplish this is
701 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
702 *
703 * call_rcu must be called by registered RCU read-side threads.
704 */
705 void call_rcu(struct rcu_head *head,
706 void (*func)(struct rcu_head *head))
707 {
708 struct call_rcu_data *crdp;
709
710 /* Holding rcu read-side lock across use of per-cpu crdp */
711 _rcu_read_lock();
712 crdp = get_call_rcu_data();
713 _call_rcu(head, func, crdp);
714 _rcu_read_unlock();
715 }
716
717 /*
718 * Free up the specified call_rcu_data structure, terminating the
719 * associated call_rcu thread. The caller must have previously
720 * removed the call_rcu_data structure from per-thread or per-CPU
721 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
722 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
723 * per-thread call_rcu_data structures.
724 *
725 * We silently refuse to free up the default call_rcu_data structure
726 * because that is where we put any leftover callbacks. Note that
727 * the possibility of self-spawning callbacks makes it impossible
728 * to execute all the callbacks in finite time without putting any
729 * newly spawned callbacks somewhere else. The "somewhere else" of
730 * last resort is the default call_rcu_data structure.
731 *
732 * We also silently refuse to free NULL pointers. This simplifies
733 * the calling code.
734 *
735 * The caller must wait for a grace-period to pass between return from
736 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
737 * previous call rcu data as argument.
738 *
739 * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
740 * a list corruption bug in the 0.7.x series. The equivalent fix
741 * appeared in 0.6.8 for the stable-0.6 branch.
742 */
743 void call_rcu_data_free(struct call_rcu_data *crdp)
744 {
745 if (crdp == NULL || crdp == default_call_rcu_data) {
746 return;
747 }
748 if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
749 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
750 wake_call_rcu_thread(crdp);
751 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
752 (void) poll(NULL, 0, 1);
753 }
754 if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
755 /* Create default call rcu data if need be */
756 (void) get_default_call_rcu_data();
757 __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
758 &default_call_rcu_data->cbs_tail,
759 &crdp->cbs_head, &crdp->cbs_tail);
760 uatomic_add(&default_call_rcu_data->qlen,
761 uatomic_read(&crdp->qlen));
762 wake_call_rcu_thread(default_call_rcu_data);
763 }
764
765 call_rcu_lock(&call_rcu_mutex);
766 cds_list_del(&crdp->list);
767 call_rcu_unlock(&call_rcu_mutex);
768
769 free(crdp);
770 }
771
772 /*
773 * Clean up all the per-CPU call_rcu threads.
774 */
775 void free_all_cpu_call_rcu_data(void)
776 {
777 int cpu;
778 struct call_rcu_data **crdp;
779 static int warned = 0;
780
781 if (maxcpus <= 0)
782 return;
783
784 crdp = malloc(sizeof(*crdp) * maxcpus);
785 if (!crdp) {
786 if (!warned) {
787 fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
788 }
789 warned = 1;
790 return;
791 }
792
793 for (cpu = 0; cpu < maxcpus; cpu++) {
794 crdp[cpu] = get_cpu_call_rcu_data(cpu);
795 if (crdp[cpu] == NULL)
796 continue;
797 set_cpu_call_rcu_data(cpu, NULL);
798 }
799 /*
800 * Wait for call_rcu sites acting as RCU readers of the
801 * call_rcu_data to become quiescent.
802 */
803 synchronize_rcu();
804 for (cpu = 0; cpu < maxcpus; cpu++) {
805 if (crdp[cpu] == NULL)
806 continue;
807 call_rcu_data_free(crdp[cpu]);
808 }
809 free(crdp);
810 }
811
812 static
813 void free_completion(struct urcu_ref *ref)
814 {
815 struct call_rcu_completion *completion;
816
817 completion = caa_container_of(ref, struct call_rcu_completion, ref);
818 free(completion);
819 }
820
821 static
822 void _rcu_barrier_complete(struct rcu_head *head)
823 {
824 struct call_rcu_completion_work *work;
825 struct call_rcu_completion *completion;
826
827 work = caa_container_of(head, struct call_rcu_completion_work, head);
828 completion = work->completion;
829 if (!uatomic_sub_return(&completion->barrier_count, 1))
830 call_rcu_completion_wake_up(completion);
831 urcu_ref_put(&completion->ref, free_completion);
832 free(work);
833 }
834
835 /*
836 * Wait for all in-flight call_rcu callbacks to complete execution.
837 */
838 void rcu_barrier(void)
839 {
840 struct call_rcu_data *crdp;
841 struct call_rcu_completion *completion;
842 int count = 0;
843 int was_online;
844
845 /* Put in offline state in QSBR. */
846 was_online = _rcu_read_ongoing();
847 if (was_online)
848 rcu_thread_offline();
849 /*
850 * Calling a rcu_barrier() within a RCU read-side critical
851 * section is an error.
852 */
853 if (_rcu_read_ongoing()) {
854 static int warned = 0;
855
856 if (!warned) {
857 fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
858 }
859 warned = 1;
860 goto online;
861 }
862
863 completion = calloc(sizeof(*completion), 1);
864 if (!completion)
865 urcu_die(errno);
866
867 call_rcu_lock(&call_rcu_mutex);
868 cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
869 count++;
870
871 /* Referenced by rcu_barrier() and each call_rcu thread. */
872 urcu_ref_set(&completion->ref, count + 1);
873 completion->barrier_count = count;
874
875 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
876 struct call_rcu_completion_work *work;
877
878 work = calloc(sizeof(*work), 1);
879 if (!work)
880 urcu_die(errno);
881 work->completion = completion;
882 _call_rcu(&work->head, _rcu_barrier_complete, crdp);
883 }
884 call_rcu_unlock(&call_rcu_mutex);
885
886 /* Wait for them */
887 for (;;) {
888 uatomic_dec(&completion->futex);
889 /* Decrement futex before reading barrier_count */
890 cmm_smp_mb();
891 if (!uatomic_read(&completion->barrier_count))
892 break;
893 call_rcu_completion_wait(completion);
894 }
895
896 urcu_ref_put(&completion->ref, free_completion);
897
898 online:
899 if (was_online)
900 rcu_thread_online();
901 }
902
903 /*
904 * Acquire the call_rcu_mutex in order to ensure that the child sees
905 * all of the call_rcu() data structures in a consistent state. Ensure
906 * that all call_rcu threads are in a quiescent state across fork.
907 * Suitable for pthread_atfork() and friends.
908 */
909 void call_rcu_before_fork(void)
910 {
911 struct call_rcu_data *crdp;
912
913 call_rcu_lock(&call_rcu_mutex);
914
915 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
916 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE);
917 cmm_smp_mb__after_uatomic_or();
918 wake_call_rcu_thread(crdp);
919 }
920 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
921 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0)
922 (void) poll(NULL, 0, 1);
923 }
924 }
925
926 /*
927 * Clean up call_rcu data structures in the parent of a successful fork()
928 * that is not followed by exec() in the child. Suitable for
929 * pthread_atfork() and friends.
930 */
931 void call_rcu_after_fork_parent(void)
932 {
933 struct call_rcu_data *crdp;
934
935 cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
936 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
937 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
938 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
939 (void) poll(NULL, 0, 1);
940 }
941 call_rcu_unlock(&call_rcu_mutex);
942 }
943
944 /*
945 * Clean up call_rcu data structures in the child of a successful fork()
946 * that is not followed by exec(). Suitable for pthread_atfork() and
947 * friends.
948 */
949 void call_rcu_after_fork_child(void)
950 {
951 struct call_rcu_data *crdp, *next;
952
953 /* Release the mutex. */
954 call_rcu_unlock(&call_rcu_mutex);
955
956 /* Do nothing when call_rcu() has not been used */
957 if (cds_list_empty(&call_rcu_data_list))
958 return;
959
960 /*
961 * Allocate a new default call_rcu_data structure in order
962 * to get a working call_rcu thread to go with it.
963 */
964 default_call_rcu_data = NULL;
965 (void)get_default_call_rcu_data();
966
967 /* Cleanup call_rcu_data pointers before use */
968 maxcpus_reset();
969 free(per_cpu_call_rcu_data);
970 rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
971 URCU_TLS(thread_call_rcu_data) = NULL;
972
973 /*
974 * Dispose of all of the rest of the call_rcu_data structures.
975 * Leftover call_rcu callbacks will be merged into the new
976 * default call_rcu thread queue.
977 */
978 cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
979 if (crdp == default_call_rcu_data)
980 continue;
981 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
982 call_rcu_data_free(crdp);
983 }
984 }
This page took 0.077552 seconds and 4 git commands to generate.