call_rcu: fix futex-based wakeup
[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->flags & URCU_CALL_RCU_RT)) {
216 uatomic_dec(&crdp->futex);
217 /* Decrement futex before reading call_rcu list */
218 cmm_smp_mb();
219 }
220 if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
221 while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
222 poll(NULL, 0, 1);
223 _CMM_STORE_SHARED(crdp->cbs.head, NULL);
224 cbs_tail = (struct cds_wfq_node **)
225 uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
226 synchronize_rcu();
227 cbcount = 0;
228 do {
229 while (cbs->next == NULL &&
230 &cbs->next != cbs_tail)
231 poll(NULL, 0, 1);
232 if (cbs == &crdp->cbs.dummy) {
233 cbs = cbs->next;
234 continue;
235 }
236 rhp = (struct rcu_head *)cbs;
237 cbs = cbs->next;
238 rhp->func(rhp);
239 cbcount++;
240 } while (cbs != NULL);
241 uatomic_sub(&crdp->qlen, cbcount);
242 }
243 if (crdp->flags & URCU_CALL_RCU_STOP) {
244 if (!(crdp->flags & URCU_CALL_RCU_RT)) {
245 /*
246 * Read call_rcu list before write futex.
247 */
248 cmm_smp_mb();
249 uatomic_set(&crdp->futex, 0);
250 }
251 break;
252 }
253 if (!(crdp->flags & URCU_CALL_RCU_RT)) {
254 if (&crdp->cbs.head == _CMM_LOAD_SHARED(crdp->cbs.tail))
255 call_rcu_wait(crdp);
256 }
257 poll(NULL, 0, 10);
258 }
259 call_rcu_lock(&crdp->mtx);
260 crdp->flags |= URCU_CALL_RCU_STOPPED;
261 call_rcu_unlock(&crdp->mtx);
262 return NULL;
263 }
264
265 /*
266 * Create both a call_rcu thread and the corresponding call_rcu_data
267 * structure, linking the structure in as specified. Caller must hold
268 * call_rcu_mutex.
269 */
270
271 static void call_rcu_data_init(struct call_rcu_data **crdpp,
272 unsigned long flags,
273 int cpu_affinity)
274 {
275 struct call_rcu_data *crdp;
276
277 crdp = malloc(sizeof(*crdp));
278 if (crdp == NULL) {
279 fprintf(stderr, "Out of memory.\n");
280 exit(-1);
281 }
282 memset(crdp, '\0', sizeof(*crdp));
283 cds_wfq_init(&crdp->cbs);
284 crdp->qlen = 0;
285 if (pthread_mutex_init(&crdp->mtx, NULL) != 0) {
286 perror("pthread_mutex_init");
287 exit(-1);
288 }
289 crdp->futex = 0;
290 crdp->flags = flags;
291 cds_list_add(&crdp->list, &call_rcu_data_list);
292 crdp->cpu_affinity = cpu_affinity;
293 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
294 *crdpp = crdp;
295 if (pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp) != 0) {
296 perror("pthread_create");
297 exit(-1);
298 }
299 }
300
301 /*
302 * Return a pointer to the call_rcu_data structure for the specified
303 * CPU, returning NULL if there is none. We cannot automatically
304 * created it because the platform we are running on might not define
305 * sched_getcpu().
306 */
307
308 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
309 {
310 static int warned = 0;
311
312 if (per_cpu_call_rcu_data == NULL)
313 return NULL;
314 if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
315 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
316 warned = 1;
317 }
318 if (cpu < 0 || maxcpus <= cpu)
319 return NULL;
320 return per_cpu_call_rcu_data[cpu];
321 }
322
323 /*
324 * Return the tid corresponding to the call_rcu thread whose
325 * call_rcu_data structure is specified.
326 */
327
328 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
329 {
330 return crdp->tid;
331 }
332
333 /*
334 * Create a call_rcu_data structure (with thread) and return a pointer.
335 */
336
337 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
338 int cpu_affinity)
339 {
340 struct call_rcu_data *crdp;
341
342 call_rcu_data_init(&crdp, flags, cpu_affinity);
343 return crdp;
344 }
345
346 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
347 int cpu_affinity)
348 {
349 struct call_rcu_data *crdp;
350
351 call_rcu_lock(&call_rcu_mutex);
352 crdp = __create_call_rcu_data(flags, cpu_affinity);
353 call_rcu_unlock(&call_rcu_mutex);
354 return crdp;
355 }
356
357 /*
358 * Set the specified CPU to use the specified call_rcu_data structure.
359 *
360 * Use NULL to remove a CPU's call_rcu_data structure, but it is
361 * the caller's responsibility to dispose of the removed structure.
362 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
363 * (prior to NULLing it out, of course).
364 */
365
366 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
367 {
368 int warned = 0;
369
370 call_rcu_lock(&call_rcu_mutex);
371 if (cpu < 0 || maxcpus <= cpu) {
372 if (!warned) {
373 fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
374 warned = 1;
375 }
376 call_rcu_unlock(&call_rcu_mutex);
377 errno = EINVAL;
378 return -EINVAL;
379 }
380 alloc_cpu_call_rcu_data();
381 call_rcu_unlock(&call_rcu_mutex);
382 if (per_cpu_call_rcu_data == NULL) {
383 errno = ENOMEM;
384 return -ENOMEM;
385 }
386 per_cpu_call_rcu_data[cpu] = crdp;
387 return 0;
388 }
389
390 /*
391 * Return a pointer to the default call_rcu_data structure, creating
392 * one if need be. Because we never free call_rcu_data structures,
393 * we don't need to be in an RCU read-side critical section.
394 */
395
396 struct call_rcu_data *get_default_call_rcu_data(void)
397 {
398 if (default_call_rcu_data != NULL)
399 return rcu_dereference(default_call_rcu_data);
400 call_rcu_lock(&call_rcu_mutex);
401 if (default_call_rcu_data != NULL) {
402 call_rcu_unlock(&call_rcu_mutex);
403 return default_call_rcu_data;
404 }
405 call_rcu_data_init(&default_call_rcu_data, 0, -1);
406 call_rcu_unlock(&call_rcu_mutex);
407 return default_call_rcu_data;
408 }
409
410 /*
411 * Return the call_rcu_data structure that applies to the currently
412 * running thread. Any call_rcu_data structure assigned specifically
413 * to this thread has first priority, followed by any call_rcu_data
414 * structure assigned to the CPU on which the thread is running,
415 * followed by the default call_rcu_data structure. If there is not
416 * yet a default call_rcu_data structure, one will be created.
417 */
418 struct call_rcu_data *get_call_rcu_data(void)
419 {
420 int curcpu;
421 static int warned = 0;
422
423 if (thread_call_rcu_data != NULL)
424 return thread_call_rcu_data;
425 if (maxcpus <= 0)
426 return get_default_call_rcu_data();
427 curcpu = sched_getcpu();
428 if (!warned && (curcpu < 0 || maxcpus <= curcpu)) {
429 fprintf(stderr, "[error] liburcu: gcrd CPU # out of range\n");
430 warned = 1;
431 }
432 if (curcpu >= 0 && maxcpus > curcpu &&
433 per_cpu_call_rcu_data != NULL &&
434 per_cpu_call_rcu_data[curcpu] != NULL)
435 return per_cpu_call_rcu_data[curcpu];
436 return get_default_call_rcu_data();
437 }
438
439 /*
440 * Return a pointer to this task's call_rcu_data if there is one.
441 */
442
443 struct call_rcu_data *get_thread_call_rcu_data(void)
444 {
445 return thread_call_rcu_data;
446 }
447
448 /*
449 * Set this task's call_rcu_data structure as specified, regardless
450 * of whether or not this task already had one. (This allows switching
451 * to and from real-time call_rcu threads, for example.)
452 *
453 * Use NULL to remove a thread's call_rcu_data structure, but it is
454 * the caller's responsibility to dispose of the removed structure.
455 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
456 * (prior to NULLing it out, of course).
457 */
458
459 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
460 {
461 thread_call_rcu_data = crdp;
462 }
463
464 /*
465 * Create a separate call_rcu thread for each CPU. This does not
466 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
467 * function if you want that behavior.
468 */
469
470 int create_all_cpu_call_rcu_data(unsigned long flags)
471 {
472 int i;
473 struct call_rcu_data *crdp;
474 int ret;
475
476 call_rcu_lock(&call_rcu_mutex);
477 alloc_cpu_call_rcu_data();
478 call_rcu_unlock(&call_rcu_mutex);
479 if (maxcpus <= 0) {
480 errno = EINVAL;
481 return -EINVAL;
482 }
483 if (per_cpu_call_rcu_data == NULL) {
484 errno = ENOMEM;
485 return -ENOMEM;
486 }
487 for (i = 0; i < maxcpus; i++) {
488 call_rcu_lock(&call_rcu_mutex);
489 if (get_cpu_call_rcu_data(i)) {
490 call_rcu_unlock(&call_rcu_mutex);
491 continue;
492 }
493 crdp = __create_call_rcu_data(flags, i);
494 if (crdp == NULL) {
495 call_rcu_unlock(&call_rcu_mutex);
496 errno = ENOMEM;
497 return -ENOMEM;
498 }
499 call_rcu_unlock(&call_rcu_mutex);
500 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
501 /* FIXME: Leaks crdp for now. */
502 return ret; /* Can happen on race. */
503 }
504 }
505 return 0;
506 }
507
508 /*
509 * Wake up the call_rcu thread corresponding to the specified
510 * call_rcu_data structure.
511 */
512 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
513 {
514 if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
515 call_rcu_wake_up(crdp);
516 }
517
518 /*
519 * Schedule a function to be invoked after a following grace period.
520 * This is the only function that must be called -- the others are
521 * only present to allow applications to tune their use of RCU for
522 * maximum performance.
523 *
524 * Note that unless a call_rcu thread has not already been created,
525 * the first invocation of call_rcu() will create one. So, if you
526 * need the first invocation of call_rcu() to be fast, make sure
527 * to create a call_rcu thread first. One way to accomplish this is
528 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
529 */
530
531 void call_rcu(struct rcu_head *head,
532 void (*func)(struct rcu_head *head))
533 {
534 struct call_rcu_data *crdp;
535
536 cds_wfq_node_init(&head->next);
537 head->func = func;
538 crdp = get_call_rcu_data();
539 cds_wfq_enqueue(&crdp->cbs, &head->next);
540 uatomic_inc(&crdp->qlen);
541 wake_call_rcu_thread(crdp);
542 }
543
544 /*
545 * Free up the specified call_rcu_data structure, terminating the
546 * associated call_rcu thread. The caller must have previously
547 * removed the call_rcu_data structure from per-thread or per-CPU
548 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
549 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
550 * per-thread call_rcu_data structures.
551 *
552 * We silently refuse to free up the default call_rcu_data structure
553 * because that is where we put any leftover callbacks. Note that
554 * the possibility of self-spawning callbacks makes it impossible
555 * to execute all the callbacks in finite time without putting any
556 * newly spawned callbacks somewhere else. The "somewhere else" of
557 * last resort is the default call_rcu_data structure.
558 *
559 * We also silently refuse to free NULL pointers. This simplifies
560 * the calling code.
561 */
562 void call_rcu_data_free(struct call_rcu_data *crdp)
563 {
564 struct cds_wfq_node *cbs;
565 struct cds_wfq_node **cbs_tail;
566 struct cds_wfq_node **cbs_endprev;
567
568 if (crdp == NULL || crdp == default_call_rcu_data) {
569 return;
570 }
571 if ((crdp->flags & URCU_CALL_RCU_STOPPED) == 0) {
572 call_rcu_lock(&crdp->mtx);
573 crdp->flags |= URCU_CALL_RCU_STOP;
574 call_rcu_unlock(&crdp->mtx);
575 wake_call_rcu_thread(crdp);
576 while ((crdp->flags & URCU_CALL_RCU_STOPPED) == 0)
577 poll(NULL, 0, 1);
578 }
579 if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
580 while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
581 poll(NULL, 0, 1);
582 _CMM_STORE_SHARED(crdp->cbs.head, NULL);
583 cbs_tail = (struct cds_wfq_node **)
584 uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
585 cbs_endprev = (struct cds_wfq_node **)
586 uatomic_xchg(&default_call_rcu_data, cbs_tail);
587 *cbs_endprev = cbs;
588 uatomic_add(&default_call_rcu_data->qlen,
589 uatomic_read(&crdp->qlen));
590 cds_list_del(&crdp->list);
591 free(crdp);
592 }
593 }
594
595 /*
596 * Clean up all the per-CPU call_rcu threads.
597 */
598 void free_all_cpu_call_rcu_data(void)
599 {
600 int cpu;
601 struct call_rcu_data *crdp;
602
603 if (maxcpus <= 0)
604 return;
605 for (cpu = 0; cpu < maxcpus; cpu++) {
606 crdp = get_cpu_call_rcu_data(cpu);
607 if (crdp == NULL)
608 continue;
609 set_cpu_call_rcu_data(cpu, NULL);
610 call_rcu_data_free(crdp);
611 }
612 }
613
614 /*
615 * Acquire the call_rcu_mutex in order to ensure that the child sees
616 * all of the call_rcu() data structures in a consistent state.
617 * Suitable for pthread_atfork() and friends.
618 */
619 void call_rcu_before_fork(void)
620 {
621 call_rcu_lock(&call_rcu_mutex);
622 }
623
624 /*
625 * Clean up call_rcu data structures in the parent of a successful fork()
626 * that is not followed by exec() in the child. Suitable for
627 * pthread_atfork() and friends.
628 */
629 void call_rcu_after_fork_parent(void)
630 {
631 call_rcu_unlock(&call_rcu_mutex);
632 }
633
634 /*
635 * Clean up call_rcu data structures in the child of a successful fork()
636 * that is not followed by exec(). Suitable for pthread_atfork() and
637 * friends.
638 */
639 void call_rcu_after_fork_child(void)
640 {
641 struct call_rcu_data *crdp;
642
643 /* Release the mutex. */
644 call_rcu_unlock(&call_rcu_mutex);
645
646 /*
647 * Allocate a new default call_rcu_data structure in order
648 * to get a working call_rcu thread to go with it.
649 */
650 default_call_rcu_data = NULL;
651 (void)get_default_call_rcu_data();
652
653 /* Dispose of all of the rest of the call_rcu_data structures. */
654 while (call_rcu_data_list.next != call_rcu_data_list.prev) {
655 crdp = cds_list_entry(call_rcu_data_list.prev,
656 struct call_rcu_data, list);
657 if (crdp == default_call_rcu_data)
658 crdp = cds_list_entry(crdp->list.prev,
659 struct call_rcu_data, list);
660 crdp->flags = URCU_CALL_RCU_STOPPED;
661 call_rcu_data_free(crdp);
662 }
663 }
This page took 0.042182 seconds and 5 git commands to generate.