Merge remote branch 'origin/urcu/ht' into urcu/ht
[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 <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/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 int32_t futex;
50 unsigned long qlen; /* maintained for debugging. */
51 pthread_t tid;
52 int cpu_affinity;
53 struct cds_list_head list;
54 } __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
55
56 /*
57 * List of all call_rcu_data structures to keep valgrind happy.
58 * Protected by call_rcu_mutex.
59 */
60
61 CDS_LIST_HEAD(call_rcu_data_list);
62
63 /* Link a thread using call_rcu() to its call_rcu thread. */
64
65 static __thread struct call_rcu_data *thread_call_rcu_data;
66
67 /* Guard call_rcu thread creation. */
68
69 static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
70
71 /* If a given thread does not have its own call_rcu thread, this is default. */
72
73 static struct call_rcu_data *default_call_rcu_data;
74
75 /*
76 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
77 * available, then we can have call_rcu threads assigned to individual
78 * CPUs rather than only to specific threads.
79 */
80
81 #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF)
82
83 /*
84 * Pointer to array of pointers to per-CPU call_rcu_data structures
85 * and # CPUs.
86 */
87
88 static struct call_rcu_data **per_cpu_call_rcu_data;
89 static long maxcpus;
90
91 /* Allocate the array if it has not already been allocated. */
92
93 static void alloc_cpu_call_rcu_data(void)
94 {
95 struct call_rcu_data **p;
96 static int warned = 0;
97
98 if (maxcpus != 0)
99 return;
100 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
101 if (maxcpus <= 0) {
102 return;
103 }
104 p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data));
105 if (p != NULL) {
106 memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data));
107 per_cpu_call_rcu_data = p;
108 } else {
109 if (!warned) {
110 fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
111 }
112 warned = 1;
113 }
114 }
115
116 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
117
118 /*
119 * per_cpu_call_rcu_data should be constant, but some functions below, used both
120 * for cases where cpu number is available and not available, assume it it not
121 * constant.
122 */
123 static struct call_rcu_data **per_cpu_call_rcu_data = NULL;
124 static const long maxcpus = -1;
125
126 static void alloc_cpu_call_rcu_data(void)
127 {
128 }
129
130 static int sched_getcpu(void)
131 {
132 return -1;
133 }
134
135 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
136
137 /* Acquire the specified pthread mutex. */
138
139 static void call_rcu_lock(pthread_mutex_t *pmp)
140 {
141 if (pthread_mutex_lock(pmp) != 0) {
142 perror("pthread_mutex_lock");
143 exit(-1);
144 }
145 }
146
147 /* Release the specified pthread mutex. */
148
149 static void call_rcu_unlock(pthread_mutex_t *pmp)
150 {
151 if (pthread_mutex_unlock(pmp) != 0) {
152 perror("pthread_mutex_unlock");
153 exit(-1);
154 }
155 }
156
157 #if HAVE_SCHED_SETAFFINITY
158 static
159 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
160 {
161 cpu_set_t mask;
162
163 if (crdp->cpu_affinity < 0)
164 return 0;
165
166 CPU_ZERO(&mask);
167 CPU_SET(crdp->cpu_affinity, &mask);
168 #if SCHED_SETAFFINITY_ARGS == 2
169 return sched_setaffinity(0, &mask);
170 #else
171 return sched_setaffinity(0, sizeof(mask), &mask);
172 #endif
173 }
174 #else
175 static
176 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
177 {
178 return 0;
179 }
180 #endif
181
182 static void call_rcu_wait(struct call_rcu_data *crdp)
183 {
184 /* Read call_rcu list before read futex */
185 cmm_smp_mb();
186 if (uatomic_read(&crdp->futex) == -1)
187 futex_async(&crdp->futex, FUTEX_WAIT, -1,
188 NULL, NULL, 0);
189 }
190
191 static void call_rcu_wake_up(struct call_rcu_data *crdp)
192 {
193 /* Write to call_rcu list before reading/writing futex */
194 cmm_smp_mb();
195 if (unlikely(uatomic_read(&crdp->futex) == -1)) {
196 uatomic_set(&crdp->futex, 0);
197 futex_async(&crdp->futex, FUTEX_WAKE, 1,
198 NULL, NULL, 0);
199 }
200 }
201
202 /* This is the code run by each call_rcu thread. */
203
204 static void *call_rcu_thread(void *arg)
205 {
206 unsigned long cbcount;
207 struct cds_wfq_node *cbs;
208 struct cds_wfq_node **cbs_tail;
209 struct call_rcu_data *crdp = (struct call_rcu_data *)arg;
210 struct rcu_head *rhp;
211 int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
212
213 if (set_thread_cpu_affinity(crdp) != 0) {
214 perror("pthread_setaffinity_np");
215 exit(-1);
216 }
217
218 thread_call_rcu_data = crdp;
219 if (!rt) {
220 uatomic_dec(&crdp->futex);
221 /* Decrement futex before reading call_rcu list */
222 cmm_smp_mb();
223 }
224 for (;;) {
225 if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
226 while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
227 poll(NULL, 0, 1);
228 _CMM_STORE_SHARED(crdp->cbs.head, NULL);
229 cbs_tail = (struct cds_wfq_node **)
230 uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
231 synchronize_rcu();
232 cbcount = 0;
233 do {
234 while (cbs->next == NULL &&
235 &cbs->next != cbs_tail)
236 poll(NULL, 0, 1);
237 if (cbs == &crdp->cbs.dummy) {
238 cbs = cbs->next;
239 continue;
240 }
241 rhp = (struct rcu_head *)cbs;
242 cbs = cbs->next;
243 rhp->func(rhp);
244 cbcount++;
245 } while (cbs != NULL);
246 uatomic_sub(&crdp->qlen, cbcount);
247 }
248 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
249 break;
250 if (!rt) {
251 if (&crdp->cbs.head
252 == _CMM_LOAD_SHARED(crdp->cbs.tail)) {
253 call_rcu_wait(crdp);
254 poll(NULL, 0, 10);
255 uatomic_dec(&crdp->futex);
256 /*
257 * Decrement futex before reading
258 * call_rcu list.
259 */
260 cmm_smp_mb();
261 } else {
262 poll(NULL, 0, 10);
263 }
264 } else {
265 poll(NULL, 0, 10);
266 }
267 }
268 if (!rt) {
269 /*
270 * Read call_rcu list before write futex.
271 */
272 cmm_smp_mb();
273 uatomic_set(&crdp->futex, 0);
274 }
275 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
276 return NULL;
277 }
278
279 /*
280 * Create both a call_rcu thread and the corresponding call_rcu_data
281 * structure, linking the structure in as specified. Caller must hold
282 * call_rcu_mutex.
283 */
284
285 static void call_rcu_data_init(struct call_rcu_data **crdpp,
286 unsigned long flags,
287 int cpu_affinity)
288 {
289 struct call_rcu_data *crdp;
290
291 crdp = malloc(sizeof(*crdp));
292 if (crdp == NULL) {
293 fprintf(stderr, "Out of memory.\n");
294 exit(-1);
295 }
296 memset(crdp, '\0', sizeof(*crdp));
297 cds_wfq_init(&crdp->cbs);
298 crdp->qlen = 0;
299 crdp->futex = 0;
300 crdp->flags = flags;
301 cds_list_add(&crdp->list, &call_rcu_data_list);
302 crdp->cpu_affinity = cpu_affinity;
303 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
304 *crdpp = crdp;
305 if (pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp) != 0) {
306 perror("pthread_create");
307 exit(-1);
308 }
309 }
310
311 /*
312 * Return a pointer to the call_rcu_data structure for the specified
313 * CPU, returning NULL if there is none. We cannot automatically
314 * created it because the platform we are running on might not define
315 * sched_getcpu().
316 */
317
318 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
319 {
320 static int warned = 0;
321
322 if (per_cpu_call_rcu_data == NULL)
323 return NULL;
324 if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
325 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
326 warned = 1;
327 }
328 if (cpu < 0 || maxcpus <= cpu)
329 return NULL;
330 return per_cpu_call_rcu_data[cpu];
331 }
332
333 /*
334 * Return the tid corresponding to the call_rcu thread whose
335 * call_rcu_data structure is specified.
336 */
337
338 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
339 {
340 return crdp->tid;
341 }
342
343 /*
344 * Create a call_rcu_data structure (with thread) and return a pointer.
345 */
346
347 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
348 int cpu_affinity)
349 {
350 struct call_rcu_data *crdp;
351
352 call_rcu_data_init(&crdp, flags, cpu_affinity);
353 return crdp;
354 }
355
356 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
357 int cpu_affinity)
358 {
359 struct call_rcu_data *crdp;
360
361 call_rcu_lock(&call_rcu_mutex);
362 crdp = __create_call_rcu_data(flags, cpu_affinity);
363 call_rcu_unlock(&call_rcu_mutex);
364 return crdp;
365 }
366
367 /*
368 * Set the specified CPU to use the specified call_rcu_data structure.
369 *
370 * Use NULL to remove a CPU's call_rcu_data structure, but it is
371 * the caller's responsibility to dispose of the removed structure.
372 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
373 * (prior to NULLing it out, of course).
374 */
375
376 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
377 {
378 int warned = 0;
379
380 call_rcu_lock(&call_rcu_mutex);
381 if (cpu < 0 || maxcpus <= cpu) {
382 if (!warned) {
383 fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
384 warned = 1;
385 }
386 call_rcu_unlock(&call_rcu_mutex);
387 errno = EINVAL;
388 return -EINVAL;
389 }
390 alloc_cpu_call_rcu_data();
391 call_rcu_unlock(&call_rcu_mutex);
392 if (per_cpu_call_rcu_data == NULL) {
393 errno = ENOMEM;
394 return -ENOMEM;
395 }
396 per_cpu_call_rcu_data[cpu] = crdp;
397 return 0;
398 }
399
400 /*
401 * Return a pointer to the default call_rcu_data structure, creating
402 * one if need be. Because we never free call_rcu_data structures,
403 * we don't need to be in an RCU read-side critical section.
404 */
405
406 struct call_rcu_data *get_default_call_rcu_data(void)
407 {
408 if (default_call_rcu_data != NULL)
409 return rcu_dereference(default_call_rcu_data);
410 call_rcu_lock(&call_rcu_mutex);
411 if (default_call_rcu_data != NULL) {
412 call_rcu_unlock(&call_rcu_mutex);
413 return default_call_rcu_data;
414 }
415 call_rcu_data_init(&default_call_rcu_data, 0, -1);
416 call_rcu_unlock(&call_rcu_mutex);
417 return default_call_rcu_data;
418 }
419
420 /*
421 * Return the call_rcu_data structure that applies to the currently
422 * running thread. Any call_rcu_data structure assigned specifically
423 * to this thread has first priority, followed by any call_rcu_data
424 * structure assigned to the CPU on which the thread is running,
425 * followed by the default call_rcu_data structure. If there is not
426 * yet a default call_rcu_data structure, one will be created.
427 */
428 struct call_rcu_data *get_call_rcu_data(void)
429 {
430 int curcpu;
431 static int warned = 0;
432
433 if (thread_call_rcu_data != NULL)
434 return thread_call_rcu_data;
435 if (maxcpus <= 0)
436 return get_default_call_rcu_data();
437 curcpu = sched_getcpu();
438 if (!warned && (curcpu < 0 || maxcpus <= curcpu)) {
439 fprintf(stderr, "[error] liburcu: gcrd CPU # out of range\n");
440 warned = 1;
441 }
442 if (curcpu >= 0 && maxcpus > curcpu &&
443 per_cpu_call_rcu_data != NULL &&
444 per_cpu_call_rcu_data[curcpu] != NULL)
445 return per_cpu_call_rcu_data[curcpu];
446 return get_default_call_rcu_data();
447 }
448
449 /*
450 * Return a pointer to this task's call_rcu_data if there is one.
451 */
452
453 struct call_rcu_data *get_thread_call_rcu_data(void)
454 {
455 return thread_call_rcu_data;
456 }
457
458 /*
459 * Set this task's call_rcu_data structure as specified, regardless
460 * of whether or not this task already had one. (This allows switching
461 * to and from real-time call_rcu threads, for example.)
462 *
463 * Use NULL to remove a thread's call_rcu_data structure, but it is
464 * the caller's responsibility to dispose of the removed structure.
465 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
466 * (prior to NULLing it out, of course).
467 */
468
469 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
470 {
471 thread_call_rcu_data = crdp;
472 }
473
474 /*
475 * Create a separate call_rcu thread for each CPU. This does not
476 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
477 * function if you want that behavior.
478 */
479
480 int create_all_cpu_call_rcu_data(unsigned long flags)
481 {
482 int i;
483 struct call_rcu_data *crdp;
484 int ret;
485
486 call_rcu_lock(&call_rcu_mutex);
487 alloc_cpu_call_rcu_data();
488 call_rcu_unlock(&call_rcu_mutex);
489 if (maxcpus <= 0) {
490 errno = EINVAL;
491 return -EINVAL;
492 }
493 if (per_cpu_call_rcu_data == NULL) {
494 errno = ENOMEM;
495 return -ENOMEM;
496 }
497 for (i = 0; i < maxcpus; i++) {
498 call_rcu_lock(&call_rcu_mutex);
499 if (get_cpu_call_rcu_data(i)) {
500 call_rcu_unlock(&call_rcu_mutex);
501 continue;
502 }
503 crdp = __create_call_rcu_data(flags, i);
504 if (crdp == NULL) {
505 call_rcu_unlock(&call_rcu_mutex);
506 errno = ENOMEM;
507 return -ENOMEM;
508 }
509 call_rcu_unlock(&call_rcu_mutex);
510 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
511 /* FIXME: Leaks crdp for now. */
512 return ret; /* Can happen on race. */
513 }
514 }
515 return 0;
516 }
517
518 /*
519 * Wake up the call_rcu thread corresponding to the specified
520 * call_rcu_data structure.
521 */
522 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
523 {
524 if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
525 call_rcu_wake_up(crdp);
526 }
527
528 /*
529 * Schedule a function to be invoked after a following grace period.
530 * This is the only function that must be called -- the others are
531 * only present to allow applications to tune their use of RCU for
532 * maximum performance.
533 *
534 * Note that unless a call_rcu thread has not already been created,
535 * the first invocation of call_rcu() will create one. So, if you
536 * need the first invocation of call_rcu() to be fast, make sure
537 * to create a call_rcu thread first. One way to accomplish this is
538 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
539 */
540
541 void call_rcu(struct rcu_head *head,
542 void (*func)(struct rcu_head *head))
543 {
544 struct call_rcu_data *crdp;
545
546 cds_wfq_node_init(&head->next);
547 head->func = func;
548 crdp = get_call_rcu_data();
549 cds_wfq_enqueue(&crdp->cbs, &head->next);
550 uatomic_inc(&crdp->qlen);
551 wake_call_rcu_thread(crdp);
552 }
553
554 /*
555 * Free up the specified call_rcu_data structure, terminating the
556 * associated call_rcu thread. The caller must have previously
557 * removed the call_rcu_data structure from per-thread or per-CPU
558 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
559 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
560 * per-thread call_rcu_data structures.
561 *
562 * We silently refuse to free up the default call_rcu_data structure
563 * because that is where we put any leftover callbacks. Note that
564 * the possibility of self-spawning callbacks makes it impossible
565 * to execute all the callbacks in finite time without putting any
566 * newly spawned callbacks somewhere else. The "somewhere else" of
567 * last resort is the default call_rcu_data structure.
568 *
569 * We also silently refuse to free NULL pointers. This simplifies
570 * the calling code.
571 */
572 void call_rcu_data_free(struct call_rcu_data *crdp)
573 {
574 struct cds_wfq_node *cbs;
575 struct cds_wfq_node **cbs_tail;
576 struct cds_wfq_node **cbs_endprev;
577
578 if (crdp == NULL || crdp == default_call_rcu_data) {
579 return;
580 }
581 if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
582 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
583 wake_call_rcu_thread(crdp);
584 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
585 poll(NULL, 0, 1);
586 }
587 if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
588 while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
589 poll(NULL, 0, 1);
590 _CMM_STORE_SHARED(crdp->cbs.head, NULL);
591 cbs_tail = (struct cds_wfq_node **)
592 uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
593 cbs_endprev = (struct cds_wfq_node **)
594 uatomic_xchg(&default_call_rcu_data, cbs_tail);
595 *cbs_endprev = cbs;
596 uatomic_add(&default_call_rcu_data->qlen,
597 uatomic_read(&crdp->qlen));
598 cds_list_del(&crdp->list);
599 free(crdp);
600 }
601 }
602
603 /*
604 * Clean up all the per-CPU call_rcu threads.
605 */
606 void free_all_cpu_call_rcu_data(void)
607 {
608 int cpu;
609 struct call_rcu_data *crdp;
610
611 if (maxcpus <= 0)
612 return;
613 for (cpu = 0; cpu < maxcpus; cpu++) {
614 crdp = get_cpu_call_rcu_data(cpu);
615 if (crdp == NULL)
616 continue;
617 set_cpu_call_rcu_data(cpu, NULL);
618 call_rcu_data_free(crdp);
619 }
620 }
621
622 /*
623 * Acquire the call_rcu_mutex in order to ensure that the child sees
624 * all of the call_rcu() data structures in a consistent state.
625 * Suitable for pthread_atfork() and friends.
626 */
627 void call_rcu_before_fork(void)
628 {
629 call_rcu_lock(&call_rcu_mutex);
630 }
631
632 /*
633 * Clean up call_rcu data structures in the parent of a successful fork()
634 * that is not followed by exec() in the child. Suitable for
635 * pthread_atfork() and friends.
636 */
637 void call_rcu_after_fork_parent(void)
638 {
639 call_rcu_unlock(&call_rcu_mutex);
640 }
641
642 /*
643 * Clean up call_rcu data structures in the child of a successful fork()
644 * that is not followed by exec(). Suitable for pthread_atfork() and
645 * friends.
646 */
647 void call_rcu_after_fork_child(void)
648 {
649 struct call_rcu_data *crdp;
650
651 /* Release the mutex. */
652 call_rcu_unlock(&call_rcu_mutex);
653
654 /*
655 * Allocate a new default call_rcu_data structure in order
656 * to get a working call_rcu thread to go with it.
657 */
658 default_call_rcu_data = NULL;
659 (void)get_default_call_rcu_data();
660
661 /* Dispose of all of the rest of the call_rcu_data structures. */
662 while (call_rcu_data_list.next != call_rcu_data_list.prev) {
663 crdp = cds_list_entry(call_rcu_data_list.prev,
664 struct call_rcu_data, list);
665 if (crdp == default_call_rcu_data)
666 crdp = cds_list_entry(crdp->list.prev,
667 struct call_rcu_data, list);
668 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
669 call_rcu_data_free(crdp);
670 }
671 }
This page took 0.042443 seconds and 4 git commands to generate.