f0751f538b4b66c829d4d988b0d5102192ad889b
[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 "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"
45 #include "urcu-die.h"
46
47 /* Data structure that identifies a call_rcu thread. */
48
49 struct call_rcu_data {
50 /*
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
55 * anyway.
56 */
57 struct cds_wfcq_tail cbs_tail;
58 struct cds_wfcq_head cbs_head;
59 unsigned long flags;
60 int32_t futex;
61 unsigned long qlen; /* maintained for debugging. */
62 pthread_t tid;
63 int cpu_affinity;
64 struct cds_list_head list;
65 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
66
67 struct call_rcu_completion {
68 int barrier_count;
69 int32_t futex;
70 };
71
72 struct call_rcu_completion_work {
73 struct rcu_head head;
74 struct call_rcu_completion *completion;
75 };
76
77 /*
78 * List of all call_rcu_data structures to keep valgrind happy.
79 * Protected by call_rcu_mutex.
80 */
81
82 static CDS_LIST_HEAD(call_rcu_data_list);
83
84 /* Link a thread using call_rcu() to its call_rcu thread. */
85
86 static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data);
87
88 /*
89 * Guard call_rcu thread creation and atfork handlers.
90 */
91 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
92
93 /* If a given thread does not have its own call_rcu thread, this is default. */
94
95 static struct call_rcu_data *default_call_rcu_data;
96
97 /*
98 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
99 * available, then we can have call_rcu threads assigned to individual
100 * CPUs rather than only to specific threads.
101 */
102
103 #ifdef HAVE_SCHED_GETCPU
104
105 static int urcu_sched_getcpu(void)
106 {
107 return sched_getcpu();
108 }
109
110 #else /* #ifdef HAVE_SCHED_GETCPU */
111
112 static int urcu_sched_getcpu(void)
113 {
114 return -1;
115 }
116
117 #endif /* #else #ifdef HAVE_SCHED_GETCPU */
118
119 #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU)
120
121 /*
122 * Pointer to array of pointers to per-CPU call_rcu_data structures
123 * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
124 * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
125 * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
126 * without mutex. The call_rcu_mutex protects updates.
127 */
128
129 static struct call_rcu_data **per_cpu_call_rcu_data;
130 static long maxcpus;
131
132 static void maxcpus_reset(void)
133 {
134 maxcpus = 0;
135 }
136
137 /* Allocate the array if it has not already been allocated. */
138
139 static void alloc_cpu_call_rcu_data(void)
140 {
141 struct call_rcu_data **p;
142 static int warned = 0;
143
144 if (maxcpus != 0)
145 return;
146 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
147 if (maxcpus <= 0) {
148 return;
149 }
150 p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data));
151 if (p != NULL) {
152 memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data));
153 rcu_set_pointer(&per_cpu_call_rcu_data, p);
154 } else {
155 if (!warned) {
156 fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
157 }
158 warned = 1;
159 }
160 }
161
162 #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
163
164 /*
165 * per_cpu_call_rcu_data should be constant, but some functions below, used both
166 * for cases where cpu number is available and not available, assume it it not
167 * constant.
168 */
169 static struct call_rcu_data **per_cpu_call_rcu_data = NULL;
170 static const long maxcpus = -1;
171
172 static void maxcpus_reset(void)
173 {
174 }
175
176 static void alloc_cpu_call_rcu_data(void)
177 {
178 }
179
180 #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
181
182 /* Acquire the specified pthread mutex. */
183
184 static void call_rcu_lock(pthread_mutex_t *pmp)
185 {
186 int ret;
187
188 ret = pthread_mutex_lock(pmp);
189 if (ret)
190 urcu_die(ret);
191 }
192
193 /* Release the specified pthread mutex. */
194
195 static void call_rcu_unlock(pthread_mutex_t *pmp)
196 {
197 int ret;
198
199 ret = pthread_mutex_unlock(pmp);
200 if (ret)
201 urcu_die(ret);
202 }
203
204 #if HAVE_SCHED_SETAFFINITY
205 static
206 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
207 {
208 cpu_set_t mask;
209
210 if (crdp->cpu_affinity < 0)
211 return 0;
212
213 CPU_ZERO(&mask);
214 CPU_SET(crdp->cpu_affinity, &mask);
215 #if SCHED_SETAFFINITY_ARGS == 2
216 return sched_setaffinity(0, &mask);
217 #else
218 return sched_setaffinity(0, sizeof(mask), &mask);
219 #endif
220 }
221 #else
222 static
223 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
224 {
225 return 0;
226 }
227 #endif
228
229 static void call_rcu_wait(struct call_rcu_data *crdp)
230 {
231 /* Read call_rcu list before read futex */
232 cmm_smp_mb();
233 if (uatomic_read(&crdp->futex) == -1)
234 futex_async(&crdp->futex, FUTEX_WAIT, -1,
235 NULL, NULL, 0);
236 }
237
238 static void call_rcu_wake_up(struct call_rcu_data *crdp)
239 {
240 /* Write to call_rcu list before reading/writing futex */
241 cmm_smp_mb();
242 if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
243 uatomic_set(&crdp->futex, 0);
244 futex_async(&crdp->futex, FUTEX_WAKE, 1,
245 NULL, NULL, 0);
246 }
247 }
248
249 static void call_rcu_completion_wait(struct call_rcu_completion *completion)
250 {
251 /* Read completion barrier count before read futex */
252 cmm_smp_mb();
253 if (uatomic_read(&completion->futex) == -1)
254 futex_async(&completion->futex, FUTEX_WAIT, -1,
255 NULL, NULL, 0);
256 }
257
258 static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
259 {
260 /* Write to completion barrier count before reading/writing futex */
261 cmm_smp_mb();
262 if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
263 uatomic_set(&completion->futex, 0);
264 futex_async(&completion->futex, FUTEX_WAKE, 1,
265 NULL, NULL, 0);
266 }
267 }
268
269 /* This is the code run by each call_rcu thread. */
270
271 static void *call_rcu_thread(void *arg)
272 {
273 unsigned long cbcount;
274 struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
275 int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
276 int ret;
277
278 ret = set_thread_cpu_affinity(crdp);
279 if (ret)
280 urcu_die(errno);
281
282 /*
283 * If callbacks take a read-side lock, we need to be registered.
284 */
285 rcu_register_thread();
286
287 URCU_TLS(thread_call_rcu_data) = crdp;
288 if (!rt) {
289 uatomic_dec(&crdp->futex);
290 /* Decrement futex before reading call_rcu list */
291 cmm_smp_mb();
292 }
293 for (;;) {
294 struct cds_wfcq_head cbs_tmp_head;
295 struct cds_wfcq_tail cbs_tmp_tail;
296 struct cds_wfcq_node *cbs, *cbs_tmp_n;
297 enum cds_wfcq_ret splice_ret;
298
299 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
300 /*
301 * Pause requested. Become quiescent: remove
302 * ourself from all global lists, and don't
303 * process any callback. The callback lists may
304 * still be non-empty though.
305 */
306 rcu_unregister_thread();
307 cmm_smp_mb__before_uatomic_or();
308 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
309 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
310 poll(NULL, 0, 1);
311 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
312 cmm_smp_mb__after_uatomic_and();
313 rcu_register_thread();
314 }
315
316 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
317 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
318 &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail);
319 assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
320 assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
321 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
322 synchronize_rcu();
323 cbcount = 0;
324 __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
325 &cbs_tmp_tail, cbs, cbs_tmp_n) {
326 struct rcu_head *rhp;
327
328 rhp = caa_container_of(cbs,
329 struct rcu_head, next);
330 rhp->func(rhp);
331 cbcount++;
332 }
333 uatomic_sub(&crdp->qlen, cbcount);
334 }
335 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
336 break;
337 rcu_thread_offline();
338 if (!rt) {
339 if (cds_wfcq_empty(&crdp->cbs_head,
340 &crdp->cbs_tail)) {
341 call_rcu_wait(crdp);
342 poll(NULL, 0, 10);
343 uatomic_dec(&crdp->futex);
344 /*
345 * Decrement futex before reading
346 * call_rcu list.
347 */
348 cmm_smp_mb();
349 } else {
350 poll(NULL, 0, 10);
351 }
352 } else {
353 poll(NULL, 0, 10);
354 }
355 rcu_thread_online();
356 }
357 if (!rt) {
358 /*
359 * Read call_rcu list before write futex.
360 */
361 cmm_smp_mb();
362 uatomic_set(&crdp->futex, 0);
363 }
364 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
365 rcu_unregister_thread();
366 return NULL;
367 }
368
369 /*
370 * Create both a call_rcu thread and the corresponding call_rcu_data
371 * structure, linking the structure in as specified. Caller must hold
372 * call_rcu_mutex.
373 */
374
375 static void call_rcu_data_init(struct call_rcu_data **crdpp,
376 unsigned long flags,
377 int cpu_affinity)
378 {
379 struct call_rcu_data *crdp;
380 int ret;
381
382 crdp = malloc(sizeof(*crdp));
383 if (crdp == NULL)
384 urcu_die(errno);
385 memset(crdp, '\0', sizeof(*crdp));
386 cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail);
387 crdp->qlen = 0;
388 crdp->futex = 0;
389 crdp->flags = flags;
390 cds_list_add(&crdp->list, &call_rcu_data_list);
391 crdp->cpu_affinity = cpu_affinity;
392 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
393 *crdpp = crdp;
394 ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
395 if (ret)
396 urcu_die(ret);
397 }
398
399 /*
400 * Return a pointer to the call_rcu_data structure for the specified
401 * CPU, returning NULL if there is none. We cannot automatically
402 * created it because the platform we are running on might not define
403 * urcu_sched_getcpu().
404 *
405 * The call to this function and use of the returned call_rcu_data
406 * should be protected by RCU read-side lock.
407 */
408
409 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
410 {
411 static int warned = 0;
412 struct call_rcu_data **pcpu_crdp;
413
414 pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data);
415 if (pcpu_crdp == NULL)
416 return NULL;
417 if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
418 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
419 warned = 1;
420 }
421 if (cpu < 0 || maxcpus <= cpu)
422 return NULL;
423 return rcu_dereference(pcpu_crdp[cpu]);
424 }
425
426 /*
427 * Return the tid corresponding to the call_rcu thread whose
428 * call_rcu_data structure is specified.
429 */
430
431 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
432 {
433 return crdp->tid;
434 }
435
436 /*
437 * Create a call_rcu_data structure (with thread) and return a pointer.
438 */
439
440 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
441 int cpu_affinity)
442 {
443 struct call_rcu_data *crdp;
444
445 call_rcu_data_init(&crdp, flags, cpu_affinity);
446 return crdp;
447 }
448
449 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
450 int cpu_affinity)
451 {
452 struct call_rcu_data *crdp;
453
454 call_rcu_lock(&call_rcu_mutex);
455 crdp = __create_call_rcu_data(flags, cpu_affinity);
456 call_rcu_unlock(&call_rcu_mutex);
457 return crdp;
458 }
459
460 /*
461 * Set the specified CPU to use the specified call_rcu_data structure.
462 *
463 * Use NULL to remove a CPU's call_rcu_data structure, but it is
464 * the caller's responsibility to dispose of the removed structure.
465 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
466 * (prior to NULLing it out, of course).
467 *
468 * The caller must wait for a grace-period to pass between return from
469 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
470 * previous call rcu data as argument.
471 */
472
473 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
474 {
475 static int warned = 0;
476
477 call_rcu_lock(&call_rcu_mutex);
478 alloc_cpu_call_rcu_data();
479 if (cpu < 0 || maxcpus <= cpu) {
480 if (!warned) {
481 fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
482 warned = 1;
483 }
484 call_rcu_unlock(&call_rcu_mutex);
485 errno = EINVAL;
486 return -EINVAL;
487 }
488
489 if (per_cpu_call_rcu_data == NULL) {
490 call_rcu_unlock(&call_rcu_mutex);
491 errno = ENOMEM;
492 return -ENOMEM;
493 }
494
495 if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) {
496 call_rcu_unlock(&call_rcu_mutex);
497 errno = EEXIST;
498 return -EEXIST;
499 }
500
501 rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp);
502 call_rcu_unlock(&call_rcu_mutex);
503 return 0;
504 }
505
506 /*
507 * Return a pointer to the default call_rcu_data structure, creating
508 * one if need be. Because we never free call_rcu_data structures,
509 * we don't need to be in an RCU read-side critical section.
510 */
511
512 struct call_rcu_data *get_default_call_rcu_data(void)
513 {
514 if (default_call_rcu_data != NULL)
515 return rcu_dereference(default_call_rcu_data);
516 call_rcu_lock(&call_rcu_mutex);
517 if (default_call_rcu_data != NULL) {
518 call_rcu_unlock(&call_rcu_mutex);
519 return default_call_rcu_data;
520 }
521 call_rcu_data_init(&default_call_rcu_data, 0, -1);
522 call_rcu_unlock(&call_rcu_mutex);
523 return default_call_rcu_data;
524 }
525
526 /*
527 * Return the call_rcu_data structure that applies to the currently
528 * running thread. Any call_rcu_data structure assigned specifically
529 * to this thread has first priority, followed by any call_rcu_data
530 * structure assigned to the CPU on which the thread is running,
531 * followed by the default call_rcu_data structure. If there is not
532 * yet a default call_rcu_data structure, one will be created.
533 *
534 * Calls to this function and use of the returned call_rcu_data should
535 * be protected by RCU read-side lock.
536 */
537 struct call_rcu_data *get_call_rcu_data(void)
538 {
539 struct call_rcu_data *crd;
540
541 if (URCU_TLS(thread_call_rcu_data) != NULL)
542 return URCU_TLS(thread_call_rcu_data);
543
544 if (maxcpus > 0) {
545 crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
546 if (crd)
547 return crd;
548 }
549
550 return get_default_call_rcu_data();
551 }
552
553 /*
554 * Return a pointer to this task's call_rcu_data if there is one.
555 */
556
557 struct call_rcu_data *get_thread_call_rcu_data(void)
558 {
559 return URCU_TLS(thread_call_rcu_data);
560 }
561
562 /*
563 * Set this task's call_rcu_data structure as specified, regardless
564 * of whether or not this task already had one. (This allows switching
565 * to and from real-time call_rcu threads, for example.)
566 *
567 * Use NULL to remove a thread's call_rcu_data structure, but it is
568 * the caller's responsibility to dispose of the removed structure.
569 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
570 * (prior to NULLing it out, of course).
571 */
572
573 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
574 {
575 URCU_TLS(thread_call_rcu_data) = crdp;
576 }
577
578 /*
579 * Create a separate call_rcu thread for each CPU. This does not
580 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
581 * function if you want that behavior. Should be paired with
582 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
583 * threads.
584 */
585
586 int create_all_cpu_call_rcu_data(unsigned long flags)
587 {
588 int i;
589 struct call_rcu_data *crdp;
590 int ret;
591
592 call_rcu_lock(&call_rcu_mutex);
593 alloc_cpu_call_rcu_data();
594 call_rcu_unlock(&call_rcu_mutex);
595 if (maxcpus <= 0) {
596 errno = EINVAL;
597 return -EINVAL;
598 }
599 if (per_cpu_call_rcu_data == NULL) {
600 errno = ENOMEM;
601 return -ENOMEM;
602 }
603 for (i = 0; i < maxcpus; i++) {
604 call_rcu_lock(&call_rcu_mutex);
605 if (get_cpu_call_rcu_data(i)) {
606 call_rcu_unlock(&call_rcu_mutex);
607 continue;
608 }
609 crdp = __create_call_rcu_data(flags, i);
610 if (crdp == NULL) {
611 call_rcu_unlock(&call_rcu_mutex);
612 errno = ENOMEM;
613 return -ENOMEM;
614 }
615 call_rcu_unlock(&call_rcu_mutex);
616 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
617 call_rcu_data_free(crdp);
618
619 /* it has been created by other thread */
620 if (ret == -EEXIST)
621 continue;
622
623 return ret;
624 }
625 }
626 return 0;
627 }
628
629 /*
630 * Wake up the call_rcu thread corresponding to the specified
631 * call_rcu_data structure.
632 */
633 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
634 {
635 if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
636 call_rcu_wake_up(crdp);
637 }
638
639 static void _call_rcu(struct rcu_head *head,
640 void (*func)(struct rcu_head *head),
641 struct call_rcu_data *crdp)
642 {
643 cds_wfcq_node_init(&head->next);
644 head->func = func;
645 cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
646 uatomic_inc(&crdp->qlen);
647 wake_call_rcu_thread(crdp);
648 }
649
650 /*
651 * Schedule a function to be invoked after a following grace period.
652 * This is the only function that must be called -- the others are
653 * only present to allow applications to tune their use of RCU for
654 * maximum performance.
655 *
656 * Note that unless a call_rcu thread has not already been created,
657 * the first invocation of call_rcu() will create one. So, if you
658 * need the first invocation of call_rcu() to be fast, make sure
659 * to create a call_rcu thread first. One way to accomplish this is
660 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
661 *
662 * call_rcu must be called by registered RCU read-side threads.
663 */
664 void call_rcu(struct rcu_head *head,
665 void (*func)(struct rcu_head *head))
666 {
667 struct call_rcu_data *crdp;
668
669 /* Holding rcu read-side lock across use of per-cpu crdp */
670 rcu_read_lock();
671 crdp = get_call_rcu_data();
672 _call_rcu(head, func, crdp);
673 rcu_read_unlock();
674 }
675
676 /*
677 * Free up the specified call_rcu_data structure, terminating the
678 * associated call_rcu thread. The caller must have previously
679 * removed the call_rcu_data structure from per-thread or per-CPU
680 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
681 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
682 * per-thread call_rcu_data structures.
683 *
684 * We silently refuse to free up the default call_rcu_data structure
685 * because that is where we put any leftover callbacks. Note that
686 * the possibility of self-spawning callbacks makes it impossible
687 * to execute all the callbacks in finite time without putting any
688 * newly spawned callbacks somewhere else. The "somewhere else" of
689 * last resort is the default call_rcu_data structure.
690 *
691 * We also silently refuse to free NULL pointers. This simplifies
692 * the calling code.
693 *
694 * The caller must wait for a grace-period to pass between return from
695 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
696 * previous call rcu data as argument.
697 *
698 * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
699 * a list corruption bug in the 0.7.x series. The equivalent fix
700 * appeared in 0.6.8 for the stable-0.6 branch.
701 */
702 void call_rcu_data_free(struct call_rcu_data *crdp)
703 {
704 if (crdp == NULL || crdp == default_call_rcu_data) {
705 return;
706 }
707 if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
708 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
709 wake_call_rcu_thread(crdp);
710 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
711 poll(NULL, 0, 1);
712 }
713 if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
714 /* Create default call rcu data if need be */
715 (void) get_default_call_rcu_data();
716 __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
717 &default_call_rcu_data->cbs_tail,
718 &crdp->cbs_head, &crdp->cbs_tail);
719 uatomic_add(&default_call_rcu_data->qlen,
720 uatomic_read(&crdp->qlen));
721 wake_call_rcu_thread(default_call_rcu_data);
722 }
723
724 call_rcu_lock(&call_rcu_mutex);
725 cds_list_del(&crdp->list);
726 call_rcu_unlock(&call_rcu_mutex);
727
728 free(crdp);
729 }
730
731 /*
732 * Clean up all the per-CPU call_rcu threads.
733 */
734 void free_all_cpu_call_rcu_data(void)
735 {
736 int cpu;
737 struct call_rcu_data **crdp;
738 static int warned = 0;
739
740 if (maxcpus <= 0)
741 return;
742
743 crdp = malloc(sizeof(*crdp) * maxcpus);
744 if (!crdp) {
745 if (!warned) {
746 fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
747 }
748 warned = 1;
749 return;
750 }
751
752 for (cpu = 0; cpu < maxcpus; cpu++) {
753 crdp[cpu] = get_cpu_call_rcu_data(cpu);
754 if (crdp[cpu] == NULL)
755 continue;
756 set_cpu_call_rcu_data(cpu, NULL);
757 }
758 /*
759 * Wait for call_rcu sites acting as RCU readers of the
760 * call_rcu_data to become quiescent.
761 */
762 synchronize_rcu();
763 for (cpu = 0; cpu < maxcpus; cpu++) {
764 if (crdp[cpu] == NULL)
765 continue;
766 call_rcu_data_free(crdp[cpu]);
767 }
768 free(crdp);
769 }
770
771 static
772 void _rcu_barrier_complete(struct rcu_head *head)
773 {
774 struct call_rcu_completion_work *work;
775 struct call_rcu_completion *completion;
776
777 work = caa_container_of(head, struct call_rcu_completion_work, head);
778 completion = work->completion;
779 uatomic_dec(&completion->barrier_count);
780 call_rcu_completion_wake_up(completion);
781 free(work);
782 }
783
784 /*
785 * Wait for all in-flight call_rcu callbacks to complete execution.
786 */
787 void rcu_barrier(void)
788 {
789 struct call_rcu_data *crdp;
790 struct call_rcu_completion completion;
791 int count = 0;
792 int was_online;
793
794 /* Put in offline state in QSBR. */
795 was_online = rcu_read_ongoing();
796 if (was_online)
797 rcu_thread_offline();
798 /*
799 * Calling a rcu_barrier() within a RCU read-side critical
800 * section is an error.
801 */
802 if (rcu_read_ongoing()) {
803 static int warned = 0;
804
805 if (!warned) {
806 fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
807 }
808 warned = 1;
809 goto online;
810 }
811
812 call_rcu_lock(&call_rcu_mutex);
813 cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
814 count++;
815
816 completion.barrier_count = count;
817 completion.futex = 0;
818
819 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
820 struct call_rcu_completion_work *work;
821
822 work = calloc(sizeof(*work), 1);
823 if (!work)
824 urcu_die(errno);
825 work->completion = &completion;
826 _call_rcu(&work->head, _rcu_barrier_complete, crdp);
827 }
828 call_rcu_unlock(&call_rcu_mutex);
829
830 /* Wait for them */
831 for (;;) {
832 uatomic_dec(&completion.futex);
833 /* Decrement futex before reading barrier_count */
834 cmm_smp_mb();
835 if (!uatomic_read(&completion.barrier_count))
836 break;
837 call_rcu_completion_wait(&completion);
838 }
839 online:
840 if (was_online)
841 rcu_thread_online();
842 }
843
844 /*
845 * Acquire the call_rcu_mutex in order to ensure that the child sees
846 * all of the call_rcu() data structures in a consistent state. Ensure
847 * that all call_rcu threads are in a quiescent state across fork.
848 * Suitable for pthread_atfork() and friends.
849 */
850 void call_rcu_before_fork(void)
851 {
852 struct call_rcu_data *crdp;
853
854 call_rcu_lock(&call_rcu_mutex);
855
856 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
857 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE);
858 cmm_smp_mb__after_uatomic_or();
859 wake_call_rcu_thread(crdp);
860 }
861 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
862 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0)
863 poll(NULL, 0, 1);
864 }
865 }
866
867 /*
868 * Clean up call_rcu data structures in the parent of a successful fork()
869 * that is not followed by exec() in the child. Suitable for
870 * pthread_atfork() and friends.
871 */
872 void call_rcu_after_fork_parent(void)
873 {
874 struct call_rcu_data *crdp;
875
876 cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
877 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
878 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
879 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
880 poll(NULL, 0, 1);
881 }
882 call_rcu_unlock(&call_rcu_mutex);
883 }
884
885 /*
886 * Clean up call_rcu data structures in the child of a successful fork()
887 * that is not followed by exec(). Suitable for pthread_atfork() and
888 * friends.
889 */
890 void call_rcu_after_fork_child(void)
891 {
892 struct call_rcu_data *crdp, *next;
893
894 /* Release the mutex. */
895 call_rcu_unlock(&call_rcu_mutex);
896
897 /* Do nothing when call_rcu() has not been used */
898 if (cds_list_empty(&call_rcu_data_list))
899 return;
900
901 /*
902 * Allocate a new default call_rcu_data structure in order
903 * to get a working call_rcu thread to go with it.
904 */
905 default_call_rcu_data = NULL;
906 (void)get_default_call_rcu_data();
907
908 /* Cleanup call_rcu_data pointers before use */
909 maxcpus_reset();
910 free(per_cpu_call_rcu_data);
911 rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
912 URCU_TLS(thread_call_rcu_data) = NULL;
913
914 /*
915 * Dispose of all of the rest of the call_rcu_data structures.
916 * Leftover call_rcu callbacks will be merged into the new
917 * default call_rcu thread queue.
918 */
919 cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
920 if (crdp == default_call_rcu_data)
921 continue;
922 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
923 call_rcu_data_free(crdp);
924 }
925 }
This page took 0.076 seconds and 3 git commands to generate.