Futex: turn "int" into "int32_t" for portability
[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 #include <stdio.h>
25 #include <pthread.h>
26 #include <signal.h>
27 #include <assert.h>
28 #include <stdlib.h>
29 #include <stdint.h>
30 #include <string.h>
31 #include <errno.h>
32 #include <poll.h>
33 #include <sys/time.h>
34 #include <syscall.h>
35 #include <unistd.h>
36 #include <sched.h>
37
38 #include "config.h"
39 #include "urcu/wfqueue.h"
40 #include "urcu-call-rcu.h"
41 #include "urcu-pointer.h"
42 #include "urcu/list.h"
43 #include "urcu/futex.h"
44
45 /* Data structure that identifies a call_rcu thread. */
46
47 struct call_rcu_data {
48 struct cds_wfq_queue cbs;
49 unsigned long flags;
50 int32_t futex;
51 unsigned long qlen; /* maintained for debugging. */
52 pthread_t tid;
53 int cpu_affinity;
54 struct cds_list_head list;
55 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
56
57 /*
58 * List of all call_rcu_data structures to keep valgrind happy.
59 * Protected by call_rcu_mutex.
60 */
61
62 CDS_LIST_HEAD(call_rcu_data_list);
63
64 /* Link a thread using call_rcu() to its call_rcu thread. */
65
66 static __thread struct call_rcu_data *thread_call_rcu_data;
67
68 /* Guard call_rcu thread creation. */
69
70 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
71
72 /* If a given thread does not have its own call_rcu thread, this is default. */
73
74 static struct call_rcu_data *default_call_rcu_data;
75
76 /*
77 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
78 * available, then we can have call_rcu threads assigned to individual
79 * CPUs rather than only to specific threads.
80 */
81
82 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
83
84 /*
85 * Pointer to array of pointers to per-CPU call_rcu_data structures
86 * and # CPUs.
87 */
88
89 static struct call_rcu_data **per_cpu_call_rcu_data;
90 static long maxcpus;
91
92 static void call_rcu_wait(struct call_rcu_data *crdp)
93 {
94 /* Read call_rcu list before read futex */
95 cmm_smp_mb();
96 if (uatomic_read(&crdp->futex) == -1)
97 futex_async(&crdp->futex, FUTEX_WAIT, -1,
98 NULL, NULL, 0);
99 }
100
101 static void call_rcu_wake_up(struct call_rcu_data *crdp)
102 {
103 /* Write to call_rcu list before reading/writing futex */
104 cmm_smp_mb();
105 if (unlikely(uatomic_read(&crdp->futex) == -1)) {
106 uatomic_set(&crdp->futex, 0);
107 futex_async(&crdp->futex, FUTEX_WAKE, 1,
108 NULL, NULL, 0);
109 }
110 }
111
112 /* Allocate the array if it has not already been allocated. */
113
114 static void alloc_cpu_call_rcu_data(void)
115 {
116 struct call_rcu_data **p;
117 static int warned = 0;
118
119 if (maxcpus != 0)
120 return;
121 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
122 if (maxcpus <= 0) {
123 return;
124 }
125 p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data));
126 if (p != NULL) {
127 memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data));
128 per_cpu_call_rcu_data = p;
129 } else {
130 if (!warned) {
131 fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
132 }
133 warned = 1;
134 }
135 }
136
137 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
138
139 static const struct call_rcu_data **per_cpu_call_rcu_data = NULL;
140 static const long maxcpus = -1;
141
142 static void alloc_cpu_call_rcu_data(void)
143 {
144 }
145
146 static int sched_getcpu(void)
147 {
148 return -1;
149 }
150
151 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
152
153 /* Acquire the specified pthread mutex. */
154
155 static void call_rcu_lock(pthread_mutex_t *pmp)
156 {
157 if (pthread_mutex_lock(pmp) != 0) {
158 perror("pthread_mutex_lock");
159 exit(-1);
160 }
161 }
162
163 /* Release the specified pthread mutex. */
164
165 static void call_rcu_unlock(pthread_mutex_t *pmp)
166 {
167 if (pthread_mutex_unlock(pmp) != 0) {
168 perror("pthread_mutex_unlock");
169 exit(-1);
170 }
171 }
172
173 #if HAVE_SCHED_SETAFFINITY
174 static
175 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
176 {
177 cpu_set_t mask;
178
179 if (crdp->cpu_affinity < 0)
180 return 0;
181
182 CPU_ZERO(&mask);
183 CPU_SET(crdp->cpu_affinity, &mask);
184 #if SCHED_SETAFFINITY_ARGS == 2
185 return sched_setaffinity(0, &mask);
186 #else
187 return sched_setaffinity(0, sizeof(mask), &mask);
188 #endif
189 }
190 #else
191 static
192 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
193 {
194 return 0;
195 }
196 #endif
197
198 /* This is the code run by each call_rcu thread. */
199
200 static void *call_rcu_thread(void *arg)
201 {
202 unsigned long cbcount;
203 struct cds_wfq_node *cbs;
204 struct cds_wfq_node **cbs_tail;
205 struct call_rcu_data *crdp = (struct call_rcu_data *)arg;
206 struct rcu_head *rhp;
207 int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
208
209 if (set_thread_cpu_affinity(crdp) != 0) {
210 perror("pthread_setaffinity_np");
211 exit(-1);
212 }
213
214 thread_call_rcu_data = crdp;
215 if (!rt) {
216 uatomic_dec(&crdp->futex);
217 /* Decrement futex before reading call_rcu list */
218 cmm_smp_mb();
219 }
220 for (;;) {
221 if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
222 while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
223 poll(NULL, 0, 1);
224 _CMM_STORE_SHARED(crdp->cbs.head, NULL);
225 cbs_tail = (struct cds_wfq_node **)
226 uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
227 synchronize_rcu();
228 cbcount = 0;
229 do {
230 while (cbs->next == NULL &&
231 &cbs->next != cbs_tail)
232 poll(NULL, 0, 1);
233 if (cbs == &crdp->cbs.dummy) {
234 cbs = cbs->next;
235 continue;
236 }
237 rhp = (struct rcu_head *)cbs;
238 cbs = cbs->next;
239 rhp->func(rhp);
240 cbcount++;
241 } while (cbs != NULL);
242 uatomic_sub(&crdp->qlen, cbcount);
243 }
244 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
245 break;
246 if (!rt) {
247 if (&crdp->cbs.head
248 == _CMM_LOAD_SHARED(crdp->cbs.tail)) {
249 call_rcu_wait(crdp);
250 poll(NULL, 0, 10);
251 uatomic_dec(&crdp->futex);
252 /*
253 * Decrement futex before reading
254 * call_rcu list.
255 */
256 cmm_smp_mb();
257 } else {
258 poll(NULL, 0, 10);
259 }
260 } else {
261 poll(NULL, 0, 10);
262 }
263 }
264 if (!rt) {
265 /*
266 * Read call_rcu list before write futex.
267 */
268 cmm_smp_mb();
269 uatomic_set(&crdp->futex, 0);
270 }
271 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
272 return NULL;
273 }
274
275 /*
276 * Create both a call_rcu thread and the corresponding call_rcu_data
277 * structure, linking the structure in as specified. Caller must hold
278 * call_rcu_mutex.
279 */
280
281 static void call_rcu_data_init(struct call_rcu_data **crdpp,
282 unsigned long flags,
283 int cpu_affinity)
284 {
285 struct call_rcu_data *crdp;
286
287 crdp = malloc(sizeof(*crdp));
288 if (crdp == NULL) {
289 fprintf(stderr, "Out of memory.\n");
290 exit(-1);
291 }
292 memset(crdp, '\0', sizeof(*crdp));
293 cds_wfq_init(&crdp->cbs);
294 crdp->qlen = 0;
295 crdp->futex = 0;
296 crdp->flags = flags;
297 cds_list_add(&crdp->list, &call_rcu_data_list);
298 crdp->cpu_affinity = cpu_affinity;
299 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
300 *crdpp = crdp;
301 if (pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp) != 0) {
302 perror("pthread_create");
303 exit(-1);
304 }
305 }
306
307 /*
308 * Return a pointer to the call_rcu_data structure for the specified
309 * CPU, returning NULL if there is none. We cannot automatically
310 * created it because the platform we are running on might not define
311 * sched_getcpu().
312 */
313
314 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
315 {
316 static int warned = 0;
317
318 if (per_cpu_call_rcu_data == NULL)
319 return NULL;
320 if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
321 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
322 warned = 1;
323 }
324 if (cpu < 0 || maxcpus <= cpu)
325 return NULL;
326 return per_cpu_call_rcu_data[cpu];
327 }
328
329 /*
330 * Return the tid corresponding to the call_rcu thread whose
331 * call_rcu_data structure is specified.
332 */
333
334 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
335 {
336 return crdp->tid;
337 }
338
339 /*
340 * Create a call_rcu_data structure (with thread) and return a pointer.
341 */
342
343 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
344 int cpu_affinity)
345 {
346 struct call_rcu_data *crdp;
347
348 call_rcu_data_init(&crdp, flags, cpu_affinity);
349 return crdp;
350 }
351
352 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
353 int cpu_affinity)
354 {
355 struct call_rcu_data *crdp;
356
357 call_rcu_lock(&call_rcu_mutex);
358 crdp = __create_call_rcu_data(flags, cpu_affinity);
359 call_rcu_unlock(&call_rcu_mutex);
360 return crdp;
361 }
362
363 /*
364 * Set the specified CPU to use the specified call_rcu_data structure.
365 *
366 * Use NULL to remove a CPU's call_rcu_data structure, but it is
367 * the caller's responsibility to dispose of the removed structure.
368 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
369 * (prior to NULLing it out, of course).
370 */
371
372 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
373 {
374 int warned = 0;
375
376 call_rcu_lock(&call_rcu_mutex);
377 if (cpu < 0 || maxcpus <= cpu) {
378 if (!warned) {
379 fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
380 warned = 1;
381 }
382 call_rcu_unlock(&call_rcu_mutex);
383 errno = EINVAL;
384 return -EINVAL;
385 }
386 alloc_cpu_call_rcu_data();
387 call_rcu_unlock(&call_rcu_mutex);
388 if (per_cpu_call_rcu_data == NULL) {
389 errno = ENOMEM;
390 return -ENOMEM;
391 }
392 per_cpu_call_rcu_data[cpu] = crdp;
393 return 0;
394 }
395
396 /*
397 * Return a pointer to the default call_rcu_data structure, creating
398 * one if need be. Because we never free call_rcu_data structures,
399 * we don't need to be in an RCU read-side critical section.
400 */
401
402 struct call_rcu_data *get_default_call_rcu_data(void)
403 {
404 if (default_call_rcu_data != NULL)
405 return rcu_dereference(default_call_rcu_data);
406 call_rcu_lock(&call_rcu_mutex);
407 if (default_call_rcu_data != NULL) {
408 call_rcu_unlock(&call_rcu_mutex);
409 return default_call_rcu_data;
410 }
411 call_rcu_data_init(&default_call_rcu_data, 0, -1);
412 call_rcu_unlock(&call_rcu_mutex);
413 return default_call_rcu_data;
414 }
415
416 /*
417 * Return the call_rcu_data structure that applies to the currently
418 * running thread. Any call_rcu_data structure assigned specifically
419 * to this thread has first priority, followed by any call_rcu_data
420 * structure assigned to the CPU on which the thread is running,
421 * followed by the default call_rcu_data structure. If there is not
422 * yet a default call_rcu_data structure, one will be created.
423 */
424 struct call_rcu_data *get_call_rcu_data(void)
425 {
426 int curcpu;
427 static int warned = 0;
428
429 if (thread_call_rcu_data != NULL)
430 return thread_call_rcu_data;
431 if (maxcpus <= 0)
432 return get_default_call_rcu_data();
433 curcpu = sched_getcpu();
434 if (!warned && (curcpu < 0 || maxcpus <= curcpu)) {
435 fprintf(stderr, "[error] liburcu: gcrd CPU # out of range\n");
436 warned = 1;
437 }
438 if (curcpu >= 0 && maxcpus > curcpu &&
439 per_cpu_call_rcu_data != NULL &&
440 per_cpu_call_rcu_data[curcpu] != NULL)
441 return per_cpu_call_rcu_data[curcpu];
442 return get_default_call_rcu_data();
443 }
444
445 /*
446 * Return a pointer to this task's call_rcu_data if there is one.
447 */
448
449 struct call_rcu_data *get_thread_call_rcu_data(void)
450 {
451 return thread_call_rcu_data;
452 }
453
454 /*
455 * Set this task's call_rcu_data structure as specified, regardless
456 * of whether or not this task already had one. (This allows switching
457 * to and from real-time call_rcu threads, for example.)
458 *
459 * Use NULL to remove a thread's call_rcu_data structure, but it is
460 * the caller's responsibility to dispose of the removed structure.
461 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
462 * (prior to NULLing it out, of course).
463 */
464
465 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
466 {
467 thread_call_rcu_data = crdp;
468 }
469
470 /*
471 * Create a separate call_rcu thread for each CPU. This does not
472 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
473 * function if you want that behavior.
474 */
475
476 int create_all_cpu_call_rcu_data(unsigned long flags)
477 {
478 int i;
479 struct call_rcu_data *crdp;
480 int ret;
481
482 call_rcu_lock(&call_rcu_mutex);
483 alloc_cpu_call_rcu_data();
484 call_rcu_unlock(&call_rcu_mutex);
485 if (maxcpus <= 0) {
486 errno = EINVAL;
487 return -EINVAL;
488 }
489 if (per_cpu_call_rcu_data == NULL) {
490 errno = ENOMEM;
491 return -ENOMEM;
492 }
493 for (i = 0; i < maxcpus; i++) {
494 call_rcu_lock(&call_rcu_mutex);
495 if (get_cpu_call_rcu_data(i)) {
496 call_rcu_unlock(&call_rcu_mutex);
497 continue;
498 }
499 crdp = __create_call_rcu_data(flags, i);
500 if (crdp == NULL) {
501 call_rcu_unlock(&call_rcu_mutex);
502 errno = ENOMEM;
503 return -ENOMEM;
504 }
505 call_rcu_unlock(&call_rcu_mutex);
506 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
507 /* FIXME: Leaks crdp for now. */
508 return ret; /* Can happen on race. */
509 }
510 }
511 return 0;
512 }
513
514 /*
515 * Wake up the call_rcu thread corresponding to the specified
516 * call_rcu_data structure.
517 */
518 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
519 {
520 if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
521 call_rcu_wake_up(crdp);
522 }
523
524 /*
525 * Schedule a function to be invoked after a following grace period.
526 * This is the only function that must be called -- the others are
527 * only present to allow applications to tune their use of RCU for
528 * maximum performance.
529 *
530 * Note that unless a call_rcu thread has not already been created,
531 * the first invocation of call_rcu() will create one. So, if you
532 * need the first invocation of call_rcu() to be fast, make sure
533 * to create a call_rcu thread first. One way to accomplish this is
534 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
535 */
536
537 void call_rcu(struct rcu_head *head,
538 void (*func)(struct rcu_head *head))
539 {
540 struct call_rcu_data *crdp;
541
542 cds_wfq_node_init(&head->next);
543 head->func = func;
544 crdp = get_call_rcu_data();
545 cds_wfq_enqueue(&crdp->cbs, &head->next);
546 uatomic_inc(&crdp->qlen);
547 wake_call_rcu_thread(crdp);
548 }
549
550 /*
551 * Free up the specified call_rcu_data structure, terminating the
552 * associated call_rcu thread. The caller must have previously
553 * removed the call_rcu_data structure from per-thread or per-CPU
554 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
555 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
556 * per-thread call_rcu_data structures.
557 *
558 * We silently refuse to free up the default call_rcu_data structure
559 * because that is where we put any leftover callbacks. Note that
560 * the possibility of self-spawning callbacks makes it impossible
561 * to execute all the callbacks in finite time without putting any
562 * newly spawned callbacks somewhere else. The "somewhere else" of
563 * last resort is the default call_rcu_data structure.
564 *
565 * We also silently refuse to free NULL pointers. This simplifies
566 * the calling code.
567 */
568 void call_rcu_data_free(struct call_rcu_data *crdp)
569 {
570 struct cds_wfq_node *cbs;
571 struct cds_wfq_node **cbs_tail;
572 struct cds_wfq_node **cbs_endprev;
573
574 if (crdp == NULL || crdp == default_call_rcu_data) {
575 return;
576 }
577 if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
578 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
579 wake_call_rcu_thread(crdp);
580 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
581 poll(NULL, 0, 1);
582 }
583 if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
584 while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
585 poll(NULL, 0, 1);
586 _CMM_STORE_SHARED(crdp->cbs.head, NULL);
587 cbs_tail = (struct cds_wfq_node **)
588 uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
589 cbs_endprev = (struct cds_wfq_node **)
590 uatomic_xchg(&default_call_rcu_data, cbs_tail);
591 *cbs_endprev = cbs;
592 uatomic_add(&default_call_rcu_data->qlen,
593 uatomic_read(&crdp->qlen));
594 cds_list_del(&crdp->list);
595 free(crdp);
596 }
597 }
598
599 /*
600 * Clean up all the per-CPU call_rcu threads.
601 */
602 void free_all_cpu_call_rcu_data(void)
603 {
604 int cpu;
605 struct call_rcu_data *crdp;
606
607 if (maxcpus <= 0)
608 return;
609 for (cpu = 0; cpu < maxcpus; cpu++) {
610 crdp = get_cpu_call_rcu_data(cpu);
611 if (crdp == NULL)
612 continue;
613 set_cpu_call_rcu_data(cpu, NULL);
614 call_rcu_data_free(crdp);
615 }
616 }
617
618 /*
619 * Acquire the call_rcu_mutex in order to ensure that the child sees
620 * all of the call_rcu() data structures in a consistent state.
621 * Suitable for pthread_atfork() and friends.
622 */
623 void call_rcu_before_fork(void)
624 {
625 call_rcu_lock(&call_rcu_mutex);
626 }
627
628 /*
629 * Clean up call_rcu data structures in the parent of a successful fork()
630 * that is not followed by exec() in the child. Suitable for
631 * pthread_atfork() and friends.
632 */
633 void call_rcu_after_fork_parent(void)
634 {
635 call_rcu_unlock(&call_rcu_mutex);
636 }
637
638 /*
639 * Clean up call_rcu data structures in the child of a successful fork()
640 * that is not followed by exec(). Suitable for pthread_atfork() and
641 * friends.
642 */
643 void call_rcu_after_fork_child(void)
644 {
645 struct call_rcu_data *crdp;
646
647 /* Release the mutex. */
648 call_rcu_unlock(&call_rcu_mutex);
649
650 /*
651 * Allocate a new default call_rcu_data structure in order
652 * to get a working call_rcu thread to go with it.
653 */
654 default_call_rcu_data = NULL;
655 (void)get_default_call_rcu_data();
656
657 /* Dispose of all of the rest of the call_rcu_data structures. */
658 while (call_rcu_data_list.next != call_rcu_data_list.prev) {
659 crdp = cds_list_entry(call_rcu_data_list.prev,
660 struct call_rcu_data, list);
661 if (crdp == default_call_rcu_data)
662 crdp = cds_list_entry(crdp->list.prev,
663 struct call_rcu_data, list);
664 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
665 call_rcu_data_free(crdp);
666 }
667 }
This page took 0.041796 seconds and 5 git commands to generate.