call_rcu: add back delay in list non-empty case
[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 int 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 static void call_rcu_wait(struct call_rcu_data *crdp)
92 {
93 /* Read call_rcu list before read futex */
94 cmm_smp_mb();
95 if (uatomic_read(&crdp->futex) == -1)
96 futex_async(&crdp->futex, FUTEX_WAIT, -1,
97 NULL, NULL, 0);
98 }
99
100 static void call_rcu_wake_up(struct call_rcu_data *crdp)
101 {
102 /* Write to call_rcu list before reading/writing futex */
103 cmm_smp_mb();
104 if (unlikely(uatomic_read(&crdp->futex) == -1)) {
105 uatomic_set(&crdp->futex, 0);
106 futex_async(&crdp->futex, FUTEX_WAKE, 1,
107 NULL, NULL, 0);
108 }
109 }
110
111 /* Allocate the array if it has not already been allocated. */
112
113 static void alloc_cpu_call_rcu_data(void)
114 {
115 struct call_rcu_data **p;
116 static int warned = 0;
117
118 if (maxcpus != 0)
119 return;
120 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
121 if (maxcpus <= 0) {
122 return;
123 }
124 p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data));
125 if (p != NULL) {
126 memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data));
127 per_cpu_call_rcu_data = p;
128 } else {
129 if (!warned) {
130 fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
131 }
132 warned = 1;
133 }
134 }
135
136 #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
137
138 static const struct call_rcu_data **per_cpu_call_rcu_data = NULL;
139 static const long maxcpus = -1;
140
141 static void alloc_cpu_call_rcu_data(void)
142 {
143 }
144
145 static int sched_getcpu(void)
146 {
147 return -1;
148 }
149
150 #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */
151
152 /* Acquire the specified pthread mutex. */
153
154 static void call_rcu_lock(pthread_mutex_t *pmp)
155 {
156 if (pthread_mutex_lock(pmp) != 0) {
157 perror("pthread_mutex_lock");
158 exit(-1);
159 }
160 }
161
162 /* Release the specified pthread mutex. */
163
164 static void call_rcu_unlock(pthread_mutex_t *pmp)
165 {
166 if (pthread_mutex_unlock(pmp) != 0) {
167 perror("pthread_mutex_unlock");
168 exit(-1);
169 }
170 }
171
172 #if HAVE_SCHED_SETAFFINITY
173 static
174 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
175 {
176 cpu_set_t mask;
177
178 if (crdp->cpu_affinity < 0)
179 return 0;
180
181 CPU_ZERO(&mask);
182 CPU_SET(crdp->cpu_affinity, &mask);
183 #if SCHED_SETAFFINITY_ARGS == 2
184 return sched_setaffinity(0, &mask);
185 #else
186 return sched_setaffinity(0, sizeof(mask), &mask);
187 #endif
188 }
189 #else
190 static
191 int set_thread_cpu_affinity(struct call_rcu_data *crdp)
192 {
193 return 0;
194 }
195 #endif
196
197 /* This is the code run by each call_rcu thread. */
198
199 static void *call_rcu_thread(void *arg)
200 {
201 unsigned long cbcount;
202 struct cds_wfq_node *cbs;
203 struct cds_wfq_node **cbs_tail;
204 struct call_rcu_data *crdp = (struct call_rcu_data *)arg;
205 struct rcu_head *rhp;
206 int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
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 if (!rt) {
215 uatomic_dec(&crdp->futex);
216 /* Decrement futex before reading call_rcu list */
217 cmm_smp_mb();
218 }
219 for (;;) {
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 (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
244 break;
245 if (!rt) {
246 if (&crdp->cbs.head
247 == _CMM_LOAD_SHARED(crdp->cbs.tail)) {
248 call_rcu_wait(crdp);
249 poll(NULL, 0, 10);
250 uatomic_dec(&crdp->futex);
251 /*
252 * Decrement futex before reading
253 * call_rcu list.
254 */
255 cmm_smp_mb();
256 } else {
257 poll(NULL, 0, 10);
258 }
259 } else {
260 poll(NULL, 0, 10);
261 }
262 }
263 if (!rt) {
264 /*
265 * Read call_rcu list before write futex.
266 */
267 cmm_smp_mb();
268 uatomic_set(&crdp->futex, 0);
269 }
270 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
271 return NULL;
272 }
273
274 /*
275 * Create both a call_rcu thread and the corresponding call_rcu_data
276 * structure, linking the structure in as specified. Caller must hold
277 * call_rcu_mutex.
278 */
279
280 static void call_rcu_data_init(struct call_rcu_data **crdpp,
281 unsigned long flags,
282 int cpu_affinity)
283 {
284 struct call_rcu_data *crdp;
285
286 crdp = malloc(sizeof(*crdp));
287 if (crdp == NULL) {
288 fprintf(stderr, "Out of memory.\n");
289 exit(-1);
290 }
291 memset(crdp, '\0', sizeof(*crdp));
292 cds_wfq_init(&crdp->cbs);
293 crdp->qlen = 0;
294 crdp->futex = 0;
295 crdp->flags = flags;
296 cds_list_add(&crdp->list, &call_rcu_data_list);
297 crdp->cpu_affinity = cpu_affinity;
298 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
299 *crdpp = crdp;
300 if (pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp) != 0) {
301 perror("pthread_create");
302 exit(-1);
303 }
304 }
305
306 /*
307 * Return a pointer to the call_rcu_data structure for the specified
308 * CPU, returning NULL if there is none. We cannot automatically
309 * created it because the platform we are running on might not define
310 * sched_getcpu().
311 */
312
313 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
314 {
315 static int warned = 0;
316
317 if (per_cpu_call_rcu_data == NULL)
318 return NULL;
319 if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
320 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
321 warned = 1;
322 }
323 if (cpu < 0 || maxcpus <= cpu)
324 return NULL;
325 return per_cpu_call_rcu_data[cpu];
326 }
327
328 /*
329 * Return the tid corresponding to the call_rcu thread whose
330 * call_rcu_data structure is specified.
331 */
332
333 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
334 {
335 return crdp->tid;
336 }
337
338 /*
339 * Create a call_rcu_data structure (with thread) and return a pointer.
340 */
341
342 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
343 int cpu_affinity)
344 {
345 struct call_rcu_data *crdp;
346
347 call_rcu_data_init(&crdp, flags, cpu_affinity);
348 return crdp;
349 }
350
351 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
352 int cpu_affinity)
353 {
354 struct call_rcu_data *crdp;
355
356 call_rcu_lock(&call_rcu_mutex);
357 crdp = __create_call_rcu_data(flags, cpu_affinity);
358 call_rcu_unlock(&call_rcu_mutex);
359 return crdp;
360 }
361
362 /*
363 * Set the specified CPU to use the specified call_rcu_data structure.
364 *
365 * Use NULL to remove a CPU's call_rcu_data structure, but it is
366 * the caller's responsibility to dispose of the removed structure.
367 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
368 * (prior to NULLing it out, of course).
369 */
370
371 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
372 {
373 int warned = 0;
374
375 call_rcu_lock(&call_rcu_mutex);
376 if (cpu < 0 || maxcpus <= cpu) {
377 if (!warned) {
378 fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
379 warned = 1;
380 }
381 call_rcu_unlock(&call_rcu_mutex);
382 errno = EINVAL;
383 return -EINVAL;
384 }
385 alloc_cpu_call_rcu_data();
386 call_rcu_unlock(&call_rcu_mutex);
387 if (per_cpu_call_rcu_data == NULL) {
388 errno = ENOMEM;
389 return -ENOMEM;
390 }
391 per_cpu_call_rcu_data[cpu] = crdp;
392 return 0;
393 }
394
395 /*
396 * Return a pointer to the default call_rcu_data structure, creating
397 * one if need be. Because we never free call_rcu_data structures,
398 * we don't need to be in an RCU read-side critical section.
399 */
400
401 struct call_rcu_data *get_default_call_rcu_data(void)
402 {
403 if (default_call_rcu_data != NULL)
404 return rcu_dereference(default_call_rcu_data);
405 call_rcu_lock(&call_rcu_mutex);
406 if (default_call_rcu_data != NULL) {
407 call_rcu_unlock(&call_rcu_mutex);
408 return default_call_rcu_data;
409 }
410 call_rcu_data_init(&default_call_rcu_data, 0, -1);
411 call_rcu_unlock(&call_rcu_mutex);
412 return default_call_rcu_data;
413 }
414
415 /*
416 * Return the call_rcu_data structure that applies to the currently
417 * running thread. Any call_rcu_data structure assigned specifically
418 * to this thread has first priority, followed by any call_rcu_data
419 * structure assigned to the CPU on which the thread is running,
420 * followed by the default call_rcu_data structure. If there is not
421 * yet a default call_rcu_data structure, one will be created.
422 */
423 struct call_rcu_data *get_call_rcu_data(void)
424 {
425 int curcpu;
426 static int warned = 0;
427
428 if (thread_call_rcu_data != NULL)
429 return thread_call_rcu_data;
430 if (maxcpus <= 0)
431 return get_default_call_rcu_data();
432 curcpu = sched_getcpu();
433 if (!warned && (curcpu < 0 || maxcpus <= curcpu)) {
434 fprintf(stderr, "[error] liburcu: gcrd CPU # out of range\n");
435 warned = 1;
436 }
437 if (curcpu >= 0 && maxcpus > curcpu &&
438 per_cpu_call_rcu_data != NULL &&
439 per_cpu_call_rcu_data[curcpu] != NULL)
440 return per_cpu_call_rcu_data[curcpu];
441 return get_default_call_rcu_data();
442 }
443
444 /*
445 * Return a pointer to this task's call_rcu_data if there is one.
446 */
447
448 struct call_rcu_data *get_thread_call_rcu_data(void)
449 {
450 return thread_call_rcu_data;
451 }
452
453 /*
454 * Set this task's call_rcu_data structure as specified, regardless
455 * of whether or not this task already had one. (This allows switching
456 * to and from real-time call_rcu threads, for example.)
457 *
458 * Use NULL to remove a thread's call_rcu_data structure, but it is
459 * the caller's responsibility to dispose of the removed structure.
460 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
461 * (prior to NULLing it out, of course).
462 */
463
464 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
465 {
466 thread_call_rcu_data = crdp;
467 }
468
469 /*
470 * Create a separate call_rcu thread for each CPU. This does not
471 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
472 * function if you want that behavior.
473 */
474
475 int create_all_cpu_call_rcu_data(unsigned long flags)
476 {
477 int i;
478 struct call_rcu_data *crdp;
479 int ret;
480
481 call_rcu_lock(&call_rcu_mutex);
482 alloc_cpu_call_rcu_data();
483 call_rcu_unlock(&call_rcu_mutex);
484 if (maxcpus <= 0) {
485 errno = EINVAL;
486 return -EINVAL;
487 }
488 if (per_cpu_call_rcu_data == NULL) {
489 errno = ENOMEM;
490 return -ENOMEM;
491 }
492 for (i = 0; i < maxcpus; i++) {
493 call_rcu_lock(&call_rcu_mutex);
494 if (get_cpu_call_rcu_data(i)) {
495 call_rcu_unlock(&call_rcu_mutex);
496 continue;
497 }
498 crdp = __create_call_rcu_data(flags, i);
499 if (crdp == NULL) {
500 call_rcu_unlock(&call_rcu_mutex);
501 errno = ENOMEM;
502 return -ENOMEM;
503 }
504 call_rcu_unlock(&call_rcu_mutex);
505 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
506 /* FIXME: Leaks crdp for now. */
507 return ret; /* Can happen on race. */
508 }
509 }
510 return 0;
511 }
512
513 /*
514 * Wake up the call_rcu thread corresponding to the specified
515 * call_rcu_data structure.
516 */
517 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
518 {
519 if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
520 call_rcu_wake_up(crdp);
521 }
522
523 /*
524 * Schedule a function to be invoked after a following grace period.
525 * This is the only function that must be called -- the others are
526 * only present to allow applications to tune their use of RCU for
527 * maximum performance.
528 *
529 * Note that unless a call_rcu thread has not already been created,
530 * the first invocation of call_rcu() will create one. So, if you
531 * need the first invocation of call_rcu() to be fast, make sure
532 * to create a call_rcu thread first. One way to accomplish this is
533 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
534 */
535
536 void call_rcu(struct rcu_head *head,
537 void (*func)(struct rcu_head *head))
538 {
539 struct call_rcu_data *crdp;
540
541 cds_wfq_node_init(&head->next);
542 head->func = func;
543 crdp = get_call_rcu_data();
544 cds_wfq_enqueue(&crdp->cbs, &head->next);
545 uatomic_inc(&crdp->qlen);
546 wake_call_rcu_thread(crdp);
547 }
548
549 /*
550 * Free up the specified call_rcu_data structure, terminating the
551 * associated call_rcu thread. The caller must have previously
552 * removed the call_rcu_data structure from per-thread or per-CPU
553 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
554 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
555 * per-thread call_rcu_data structures.
556 *
557 * We silently refuse to free up the default call_rcu_data structure
558 * because that is where we put any leftover callbacks. Note that
559 * the possibility of self-spawning callbacks makes it impossible
560 * to execute all the callbacks in finite time without putting any
561 * newly spawned callbacks somewhere else. The "somewhere else" of
562 * last resort is the default call_rcu_data structure.
563 *
564 * We also silently refuse to free NULL pointers. This simplifies
565 * the calling code.
566 */
567 void call_rcu_data_free(struct call_rcu_data *crdp)
568 {
569 struct cds_wfq_node *cbs;
570 struct cds_wfq_node **cbs_tail;
571 struct cds_wfq_node **cbs_endprev;
572
573 if (crdp == NULL || crdp == default_call_rcu_data) {
574 return;
575 }
576 if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
577 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
578 wake_call_rcu_thread(crdp);
579 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
580 poll(NULL, 0, 1);
581 }
582 if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
583 while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
584 poll(NULL, 0, 1);
585 _CMM_STORE_SHARED(crdp->cbs.head, NULL);
586 cbs_tail = (struct cds_wfq_node **)
587 uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
588 cbs_endprev = (struct cds_wfq_node **)
589 uatomic_xchg(&default_call_rcu_data, cbs_tail);
590 *cbs_endprev = cbs;
591 uatomic_add(&default_call_rcu_data->qlen,
592 uatomic_read(&crdp->qlen));
593 cds_list_del(&crdp->list);
594 free(crdp);
595 }
596 }
597
598 /*
599 * Clean up all the per-CPU call_rcu threads.
600 */
601 void free_all_cpu_call_rcu_data(void)
602 {
603 int cpu;
604 struct call_rcu_data *crdp;
605
606 if (maxcpus <= 0)
607 return;
608 for (cpu = 0; cpu < maxcpus; cpu++) {
609 crdp = get_cpu_call_rcu_data(cpu);
610 if (crdp == NULL)
611 continue;
612 set_cpu_call_rcu_data(cpu, NULL);
613 call_rcu_data_free(crdp);
614 }
615 }
616
617 /*
618 * Acquire the call_rcu_mutex in order to ensure that the child sees
619 * all of the call_rcu() data structures in a consistent state.
620 * Suitable for pthread_atfork() and friends.
621 */
622 void call_rcu_before_fork(void)
623 {
624 call_rcu_lock(&call_rcu_mutex);
625 }
626
627 /*
628 * Clean up call_rcu data structures in the parent of a successful fork()
629 * that is not followed by exec() in the child. Suitable for
630 * pthread_atfork() and friends.
631 */
632 void call_rcu_after_fork_parent(void)
633 {
634 call_rcu_unlock(&call_rcu_mutex);
635 }
636
637 /*
638 * Clean up call_rcu data structures in the child of a successful fork()
639 * that is not followed by exec(). Suitable for pthread_atfork() and
640 * friends.
641 */
642 void call_rcu_after_fork_child(void)
643 {
644 struct call_rcu_data *crdp;
645
646 /* Release the mutex. */
647 call_rcu_unlock(&call_rcu_mutex);
648
649 /*
650 * Allocate a new default call_rcu_data structure in order
651 * to get a working call_rcu thread to go with it.
652 */
653 default_call_rcu_data = NULL;
654 (void)get_default_call_rcu_data();
655
656 /* Dispose of all of the rest of the call_rcu_data structures. */
657 while (call_rcu_data_list.next != call_rcu_data_list.prev) {
658 crdp = cds_list_entry(call_rcu_data_list.prev,
659 struct call_rcu_data, list);
660 if (crdp == default_call_rcu_data)
661 crdp = cds_list_entry(crdp->list.prev,
662 struct call_rcu_data, list);
663 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
664 call_rcu_data_free(crdp);
665 }
666 }
This page took 0.041562 seconds and 4 git commands to generate.