call_rcu: use futex for wakeup scheme
[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 <string.h>
30 #include <errno.h>
31 #include <poll.h>
32 #include <sys/time.h>
33 #include <syscall.h>
34 #include <unistd.h>
35 #include <sched.h>
36
37 #include "config.h"
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"
43
44 /* Data structure that identifies a call_rcu thread. */
45
46 struct call_rcu_data {
47 struct cds_wfq_queue cbs;
48 unsigned long flags;
49 pthread_mutex_t mtx;
50 int futex;
51 unsigned long qlen;
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
208 if (set_thread_cpu_affinity(crdp) != 0) {
209 perror("pthread_setaffinity_np");
210 exit(-1);
211 }
212
213 thread_call_rcu_data = crdp;
214 for (;;) {
215 if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
216 while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
217 poll(NULL, 0, 1);
218 _CMM_STORE_SHARED(crdp->cbs.head, NULL);
219 cbs_tail = (struct cds_wfq_node **)
220 uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
221 synchronize_rcu();
222 cbcount = 0;
223 do {
224 while (cbs->next == NULL &&
225 &cbs->next != cbs_tail)
226 poll(NULL, 0, 1);
227 if (cbs == &crdp->cbs.dummy) {
228 cbs = cbs->next;
229 continue;
230 }
231 rhp = (struct rcu_head *)cbs;
232 cbs = cbs->next;
233 rhp->func(rhp);
234 cbcount++;
235 } while (cbs != NULL);
236 uatomic_sub(&crdp->qlen, cbcount);
237 }
238 if (crdp->flags & URCU_CALL_RCU_STOP)
239 break;
240 if (crdp->flags & URCU_CALL_RCU_RT)
241 poll(NULL, 0, 10);
242 else {
243 if (&crdp->cbs.head == _CMM_LOAD_SHARED(crdp->cbs.tail))
244 call_rcu_wait(crdp);
245 poll(NULL, 0, 10);
246 }
247 }
248 call_rcu_lock(&crdp->mtx);
249 crdp->flags |= URCU_CALL_RCU_STOPPED;
250 call_rcu_unlock(&crdp->mtx);
251 return NULL;
252 }
253
254 /*
255 * Create both a call_rcu thread and the corresponding call_rcu_data
256 * structure, linking the structure in as specified. Caller must hold
257 * call_rcu_mutex.
258 */
259
260 static void call_rcu_data_init(struct call_rcu_data **crdpp,
261 unsigned long flags,
262 int cpu_affinity)
263 {
264 struct call_rcu_data *crdp;
265
266 crdp = malloc(sizeof(*crdp));
267 if (crdp == NULL) {
268 fprintf(stderr, "Out of memory.\n");
269 exit(-1);
270 }
271 memset(crdp, '\0', sizeof(*crdp));
272 cds_wfq_init(&crdp->cbs);
273 crdp->qlen = 0;
274 if (pthread_mutex_init(&crdp->mtx, NULL) != 0) {
275 perror("pthread_mutex_init");
276 exit(-1);
277 }
278 crdp->futex = 0;
279 crdp->flags = flags;
280 cds_list_add(&crdp->list, &call_rcu_data_list);
281 crdp->cpu_affinity = cpu_affinity;
282 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
283 *crdpp = crdp;
284 if (pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp) != 0) {
285 perror("pthread_create");
286 exit(-1);
287 }
288 }
289
290 /*
291 * Return a pointer to the call_rcu_data structure for the specified
292 * CPU, returning NULL if there is none. We cannot automatically
293 * created it because the platform we are running on might not define
294 * sched_getcpu().
295 */
296
297 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
298 {
299 static int warned = 0;
300
301 if (per_cpu_call_rcu_data == NULL)
302 return NULL;
303 if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
304 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
305 warned = 1;
306 }
307 if (cpu < 0 || maxcpus <= cpu)
308 return NULL;
309 return per_cpu_call_rcu_data[cpu];
310 }
311
312 /*
313 * Return the tid corresponding to the call_rcu thread whose
314 * call_rcu_data structure is specified.
315 */
316
317 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
318 {
319 return crdp->tid;
320 }
321
322 /*
323 * Create a call_rcu_data structure (with thread) and return a pointer.
324 */
325
326 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
327 int cpu_affinity)
328 {
329 struct call_rcu_data *crdp;
330
331 call_rcu_data_init(&crdp, flags, cpu_affinity);
332 return crdp;
333 }
334
335 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
336 int cpu_affinity)
337 {
338 struct call_rcu_data *crdp;
339
340 call_rcu_lock(&call_rcu_mutex);
341 crdp = __create_call_rcu_data(flags, cpu_affinity);
342 call_rcu_unlock(&call_rcu_mutex);
343 return crdp;
344 }
345
346 /*
347 * Set the specified CPU to use the specified call_rcu_data structure.
348 *
349 * Use NULL to remove a CPU's call_rcu_data structure, but it is
350 * the caller's responsibility to dispose of the removed structure.
351 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
352 * (prior to NULLing it out, of course).
353 */
354
355 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
356 {
357 int warned = 0;
358
359 call_rcu_lock(&call_rcu_mutex);
360 if (cpu < 0 || maxcpus <= cpu) {
361 if (!warned) {
362 fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
363 warned = 1;
364 }
365 call_rcu_unlock(&call_rcu_mutex);
366 errno = EINVAL;
367 return -EINVAL;
368 }
369 alloc_cpu_call_rcu_data();
370 call_rcu_unlock(&call_rcu_mutex);
371 if (per_cpu_call_rcu_data == NULL) {
372 errno = ENOMEM;
373 return -ENOMEM;
374 }
375 per_cpu_call_rcu_data[cpu] = crdp;
376 return 0;
377 }
378
379 /*
380 * Return a pointer to the default call_rcu_data structure, creating
381 * one if need be. Because we never free call_rcu_data structures,
382 * we don't need to be in an RCU read-side critical section.
383 */
384
385 struct call_rcu_data *get_default_call_rcu_data(void)
386 {
387 if (default_call_rcu_data != NULL)
388 return rcu_dereference(default_call_rcu_data);
389 call_rcu_lock(&call_rcu_mutex);
390 if (default_call_rcu_data != NULL) {
391 call_rcu_unlock(&call_rcu_mutex);
392 return default_call_rcu_data;
393 }
394 call_rcu_data_init(&default_call_rcu_data, 0, -1);
395 call_rcu_unlock(&call_rcu_mutex);
396 return default_call_rcu_data;
397 }
398
399 /*
400 * Return the call_rcu_data structure that applies to the currently
401 * running thread. Any call_rcu_data structure assigned specifically
402 * to this thread has first priority, followed by any call_rcu_data
403 * structure assigned to the CPU on which the thread is running,
404 * followed by the default call_rcu_data structure. If there is not
405 * yet a default call_rcu_data structure, one will be created.
406 */
407 struct call_rcu_data *get_call_rcu_data(void)
408 {
409 int curcpu;
410 static int warned = 0;
411
412 if (thread_call_rcu_data != NULL)
413 return thread_call_rcu_data;
414 if (maxcpus <= 0)
415 return get_default_call_rcu_data();
416 curcpu = sched_getcpu();
417 if (!warned && (curcpu < 0 || maxcpus <= curcpu)) {
418 fprintf(stderr, "[error] liburcu: gcrd CPU # out of range\n");
419 warned = 1;
420 }
421 if (curcpu >= 0 && maxcpus > curcpu &&
422 per_cpu_call_rcu_data != NULL &&
423 per_cpu_call_rcu_data[curcpu] != NULL)
424 return per_cpu_call_rcu_data[curcpu];
425 return get_default_call_rcu_data();
426 }
427
428 /*
429 * Return a pointer to this task's call_rcu_data if there is one.
430 */
431
432 struct call_rcu_data *get_thread_call_rcu_data(void)
433 {
434 return thread_call_rcu_data;
435 }
436
437 /*
438 * Set this task's call_rcu_data structure as specified, regardless
439 * of whether or not this task already had one. (This allows switching
440 * to and from real-time call_rcu threads, for example.)
441 *
442 * Use NULL to remove a thread's call_rcu_data structure, but it is
443 * the caller's responsibility to dispose of the removed structure.
444 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
445 * (prior to NULLing it out, of course).
446 */
447
448 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
449 {
450 thread_call_rcu_data = crdp;
451 }
452
453 /*
454 * Create a separate call_rcu thread for each CPU. This does not
455 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
456 * function if you want that behavior.
457 */
458
459 int create_all_cpu_call_rcu_data(unsigned long flags)
460 {
461 int i;
462 struct call_rcu_data *crdp;
463 int ret;
464
465 call_rcu_lock(&call_rcu_mutex);
466 alloc_cpu_call_rcu_data();
467 call_rcu_unlock(&call_rcu_mutex);
468 if (maxcpus <= 0) {
469 errno = EINVAL;
470 return -EINVAL;
471 }
472 if (per_cpu_call_rcu_data == NULL) {
473 errno = ENOMEM;
474 return -ENOMEM;
475 }
476 for (i = 0; i < maxcpus; i++) {
477 call_rcu_lock(&call_rcu_mutex);
478 if (get_cpu_call_rcu_data(i)) {
479 call_rcu_unlock(&call_rcu_mutex);
480 continue;
481 }
482 crdp = __create_call_rcu_data(flags, i);
483 if (crdp == NULL) {
484 call_rcu_unlock(&call_rcu_mutex);
485 errno = ENOMEM;
486 return -ENOMEM;
487 }
488 call_rcu_unlock(&call_rcu_mutex);
489 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
490 /* FIXME: Leaks crdp for now. */
491 return ret; /* Can happen on race. */
492 }
493 }
494 return 0;
495 }
496
497 /*
498 * Wake up the call_rcu thread corresponding to the specified
499 * call_rcu_data structure.
500 */
501 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
502 {
503 if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
504 call_rcu_wake_up(crdp);
505 }
506
507 /*
508 * Schedule a function to be invoked after a following grace period.
509 * This is the only function that must be called -- the others are
510 * only present to allow applications to tune their use of RCU for
511 * maximum performance.
512 *
513 * Note that unless a call_rcu thread has not already been created,
514 * the first invocation of call_rcu() will create one. So, if you
515 * need the first invocation of call_rcu() to be fast, make sure
516 * to create a call_rcu thread first. One way to accomplish this is
517 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
518 */
519
520 void call_rcu(struct rcu_head *head,
521 void (*func)(struct rcu_head *head))
522 {
523 struct call_rcu_data *crdp;
524
525 cds_wfq_node_init(&head->next);
526 head->func = func;
527 crdp = get_call_rcu_data();
528 cds_wfq_enqueue(&crdp->cbs, &head->next);
529 uatomic_inc(&crdp->qlen);
530 wake_call_rcu_thread(crdp);
531 }
532
533 /*
534 * Free up the specified call_rcu_data structure, terminating the
535 * associated call_rcu thread. The caller must have previously
536 * removed the call_rcu_data structure from per-thread or per-CPU
537 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
538 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
539 * per-thread call_rcu_data structures.
540 *
541 * We silently refuse to free up the default call_rcu_data structure
542 * because that is where we put any leftover callbacks. Note that
543 * the possibility of self-spawning callbacks makes it impossible
544 * to execute all the callbacks in finite time without putting any
545 * newly spawned callbacks somewhere else. The "somewhere else" of
546 * last resort is the default call_rcu_data structure.
547 *
548 * We also silently refuse to free NULL pointers. This simplifies
549 * the calling code.
550 */
551 void call_rcu_data_free(struct call_rcu_data *crdp)
552 {
553 struct cds_wfq_node *cbs;
554 struct cds_wfq_node **cbs_tail;
555 struct cds_wfq_node **cbs_endprev;
556
557 if (crdp == NULL || crdp == default_call_rcu_data) {
558 return;
559 }
560 if ((crdp->flags & URCU_CALL_RCU_STOPPED) == 0) {
561 call_rcu_lock(&crdp->mtx);
562 crdp->flags |= URCU_CALL_RCU_STOP;
563 call_rcu_unlock(&crdp->mtx);
564 wake_call_rcu_thread(crdp);
565 while ((crdp->flags & URCU_CALL_RCU_STOPPED) == 0)
566 poll(NULL, 0, 1);
567 }
568 if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
569 while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
570 poll(NULL, 0, 1);
571 _CMM_STORE_SHARED(crdp->cbs.head, NULL);
572 cbs_tail = (struct cds_wfq_node **)
573 uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
574 cbs_endprev = (struct cds_wfq_node **)
575 uatomic_xchg(&default_call_rcu_data, cbs_tail);
576 *cbs_endprev = cbs;
577 uatomic_add(&default_call_rcu_data->qlen,
578 uatomic_read(&crdp->qlen));
579 cds_list_del(&crdp->list);
580 free(crdp);
581 }
582 }
583
584 /*
585 * Clean up all the per-CPU call_rcu threads.
586 */
587 void free_all_cpu_call_rcu_data(void)
588 {
589 int cpu;
590 struct call_rcu_data *crdp;
591
592 if (maxcpus <= 0)
593 return;
594 for (cpu = 0; cpu < maxcpus; cpu++) {
595 crdp = get_cpu_call_rcu_data(cpu);
596 if (crdp == NULL)
597 continue;
598 set_cpu_call_rcu_data(cpu, NULL);
599 call_rcu_data_free(crdp);
600 }
601 }
602
603 /*
604 * Acquire the call_rcu_mutex in order to ensure that the child sees
605 * all of the call_rcu() data structures in a consistent state.
606 * Suitable for pthread_atfork() and friends.
607 */
608 void call_rcu_before_fork(void)
609 {
610 call_rcu_lock(&call_rcu_mutex);
611 }
612
613 /*
614 * Clean up call_rcu data structures in the parent of a successful fork()
615 * that is not followed by exec() in the child. Suitable for
616 * pthread_atfork() and friends.
617 */
618 void call_rcu_after_fork_parent(void)
619 {
620 call_rcu_unlock(&call_rcu_mutex);
621 }
622
623 /*
624 * Clean up call_rcu data structures in the child of a successful fork()
625 * that is not followed by exec(). Suitable for pthread_atfork() and
626 * friends.
627 */
628 void call_rcu_after_fork_child(void)
629 {
630 struct call_rcu_data *crdp;
631
632 /* Release the mutex. */
633 call_rcu_unlock(&call_rcu_mutex);
634
635 /*
636 * Allocate a new default call_rcu_data structure in order
637 * to get a working call_rcu thread to go with it.
638 */
639 default_call_rcu_data = NULL;
640 (void)get_default_call_rcu_data();
641
642 /* Dispose of all of the rest of the call_rcu_data structures. */
643 while (call_rcu_data_list.next != call_rcu_data_list.prev) {
644 crdp = cds_list_entry(call_rcu_data_list.prev,
645 struct call_rcu_data, list);
646 if (crdp == default_call_rcu_data)
647 crdp = cds_list_entry(crdp->list.prev,
648 struct call_rcu_data, list);
649 crdp->flags = URCU_CALL_RCU_STOPPED;
650 call_rcu_data_free(crdp);
651 }
652 }
This page took 0.041107 seconds and 4 git commands to generate.