call_rcu: handle retry without wait correctly
[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 }
257 } else {
258 poll(NULL, 0, 10);
259 }
260 }
261 if (!rt) {
262 /*
263 * Read call_rcu list before write futex.
264 */
265 cmm_smp_mb();
266 uatomic_set(&crdp->futex, 0);
267 }
268 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
269 return NULL;
270 }
271
272 /*
273 * Create both a call_rcu thread and the corresponding call_rcu_data
274 * structure, linking the structure in as specified. Caller must hold
275 * call_rcu_mutex.
276 */
277
278 static void call_rcu_data_init(struct call_rcu_data **crdpp,
279 unsigned long flags,
280 int cpu_affinity)
281 {
282 struct call_rcu_data *crdp;
283
284 crdp = malloc(sizeof(*crdp));
285 if (crdp == NULL) {
286 fprintf(stderr, "Out of memory.\n");
287 exit(-1);
288 }
289 memset(crdp, '\0', sizeof(*crdp));
290 cds_wfq_init(&crdp->cbs);
291 crdp->qlen = 0;
292 crdp->futex = 0;
293 crdp->flags = flags;
294 cds_list_add(&crdp->list, &call_rcu_data_list);
295 crdp->cpu_affinity = cpu_affinity;
296 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
297 *crdpp = crdp;
298 if (pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp) != 0) {
299 perror("pthread_create");
300 exit(-1);
301 }
302 }
303
304 /*
305 * Return a pointer to the call_rcu_data structure for the specified
306 * CPU, returning NULL if there is none. We cannot automatically
307 * created it because the platform we are running on might not define
308 * sched_getcpu().
309 */
310
311 struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
312 {
313 static int warned = 0;
314
315 if (per_cpu_call_rcu_data == NULL)
316 return NULL;
317 if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
318 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
319 warned = 1;
320 }
321 if (cpu < 0 || maxcpus <= cpu)
322 return NULL;
323 return per_cpu_call_rcu_data[cpu];
324 }
325
326 /*
327 * Return the tid corresponding to the call_rcu thread whose
328 * call_rcu_data structure is specified.
329 */
330
331 pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
332 {
333 return crdp->tid;
334 }
335
336 /*
337 * Create a call_rcu_data structure (with thread) and return a pointer.
338 */
339
340 static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
341 int cpu_affinity)
342 {
343 struct call_rcu_data *crdp;
344
345 call_rcu_data_init(&crdp, flags, cpu_affinity);
346 return crdp;
347 }
348
349 struct call_rcu_data *create_call_rcu_data(unsigned long flags,
350 int cpu_affinity)
351 {
352 struct call_rcu_data *crdp;
353
354 call_rcu_lock(&call_rcu_mutex);
355 crdp = __create_call_rcu_data(flags, cpu_affinity);
356 call_rcu_unlock(&call_rcu_mutex);
357 return crdp;
358 }
359
360 /*
361 * Set the specified CPU to use the specified call_rcu_data structure.
362 *
363 * Use NULL to remove a CPU's call_rcu_data structure, but it is
364 * the caller's responsibility to dispose of the removed structure.
365 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
366 * (prior to NULLing it out, of course).
367 */
368
369 int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
370 {
371 int warned = 0;
372
373 call_rcu_lock(&call_rcu_mutex);
374 if (cpu < 0 || maxcpus <= cpu) {
375 if (!warned) {
376 fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
377 warned = 1;
378 }
379 call_rcu_unlock(&call_rcu_mutex);
380 errno = EINVAL;
381 return -EINVAL;
382 }
383 alloc_cpu_call_rcu_data();
384 call_rcu_unlock(&call_rcu_mutex);
385 if (per_cpu_call_rcu_data == NULL) {
386 errno = ENOMEM;
387 return -ENOMEM;
388 }
389 per_cpu_call_rcu_data[cpu] = crdp;
390 return 0;
391 }
392
393 /*
394 * Return a pointer to the default call_rcu_data structure, creating
395 * one if need be. Because we never free call_rcu_data structures,
396 * we don't need to be in an RCU read-side critical section.
397 */
398
399 struct call_rcu_data *get_default_call_rcu_data(void)
400 {
401 if (default_call_rcu_data != NULL)
402 return rcu_dereference(default_call_rcu_data);
403 call_rcu_lock(&call_rcu_mutex);
404 if (default_call_rcu_data != NULL) {
405 call_rcu_unlock(&call_rcu_mutex);
406 return default_call_rcu_data;
407 }
408 call_rcu_data_init(&default_call_rcu_data, 0, -1);
409 call_rcu_unlock(&call_rcu_mutex);
410 return default_call_rcu_data;
411 }
412
413 /*
414 * Return the call_rcu_data structure that applies to the currently
415 * running thread. Any call_rcu_data structure assigned specifically
416 * to this thread has first priority, followed by any call_rcu_data
417 * structure assigned to the CPU on which the thread is running,
418 * followed by the default call_rcu_data structure. If there is not
419 * yet a default call_rcu_data structure, one will be created.
420 */
421 struct call_rcu_data *get_call_rcu_data(void)
422 {
423 int curcpu;
424 static int warned = 0;
425
426 if (thread_call_rcu_data != NULL)
427 return thread_call_rcu_data;
428 if (maxcpus <= 0)
429 return get_default_call_rcu_data();
430 curcpu = sched_getcpu();
431 if (!warned && (curcpu < 0 || maxcpus <= curcpu)) {
432 fprintf(stderr, "[error] liburcu: gcrd CPU # out of range\n");
433 warned = 1;
434 }
435 if (curcpu >= 0 && maxcpus > curcpu &&
436 per_cpu_call_rcu_data != NULL &&
437 per_cpu_call_rcu_data[curcpu] != NULL)
438 return per_cpu_call_rcu_data[curcpu];
439 return get_default_call_rcu_data();
440 }
441
442 /*
443 * Return a pointer to this task's call_rcu_data if there is one.
444 */
445
446 struct call_rcu_data *get_thread_call_rcu_data(void)
447 {
448 return thread_call_rcu_data;
449 }
450
451 /*
452 * Set this task's call_rcu_data structure as specified, regardless
453 * of whether or not this task already had one. (This allows switching
454 * to and from real-time call_rcu threads, for example.)
455 *
456 * Use NULL to remove a thread's call_rcu_data structure, but it is
457 * the caller's responsibility to dispose of the removed structure.
458 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
459 * (prior to NULLing it out, of course).
460 */
461
462 void set_thread_call_rcu_data(struct call_rcu_data *crdp)
463 {
464 thread_call_rcu_data = crdp;
465 }
466
467 /*
468 * Create a separate call_rcu thread for each CPU. This does not
469 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
470 * function if you want that behavior.
471 */
472
473 int create_all_cpu_call_rcu_data(unsigned long flags)
474 {
475 int i;
476 struct call_rcu_data *crdp;
477 int ret;
478
479 call_rcu_lock(&call_rcu_mutex);
480 alloc_cpu_call_rcu_data();
481 call_rcu_unlock(&call_rcu_mutex);
482 if (maxcpus <= 0) {
483 errno = EINVAL;
484 return -EINVAL;
485 }
486 if (per_cpu_call_rcu_data == NULL) {
487 errno = ENOMEM;
488 return -ENOMEM;
489 }
490 for (i = 0; i < maxcpus; i++) {
491 call_rcu_lock(&call_rcu_mutex);
492 if (get_cpu_call_rcu_data(i)) {
493 call_rcu_unlock(&call_rcu_mutex);
494 continue;
495 }
496 crdp = __create_call_rcu_data(flags, i);
497 if (crdp == NULL) {
498 call_rcu_unlock(&call_rcu_mutex);
499 errno = ENOMEM;
500 return -ENOMEM;
501 }
502 call_rcu_unlock(&call_rcu_mutex);
503 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
504 /* FIXME: Leaks crdp for now. */
505 return ret; /* Can happen on race. */
506 }
507 }
508 return 0;
509 }
510
511 /*
512 * Wake up the call_rcu thread corresponding to the specified
513 * call_rcu_data structure.
514 */
515 static void wake_call_rcu_thread(struct call_rcu_data *crdp)
516 {
517 if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
518 call_rcu_wake_up(crdp);
519 }
520
521 /*
522 * Schedule a function to be invoked after a following grace period.
523 * This is the only function that must be called -- the others are
524 * only present to allow applications to tune their use of RCU for
525 * maximum performance.
526 *
527 * Note that unless a call_rcu thread has not already been created,
528 * the first invocation of call_rcu() will create one. So, if you
529 * need the first invocation of call_rcu() to be fast, make sure
530 * to create a call_rcu thread first. One way to accomplish this is
531 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
532 */
533
534 void call_rcu(struct rcu_head *head,
535 void (*func)(struct rcu_head *head))
536 {
537 struct call_rcu_data *crdp;
538
539 cds_wfq_node_init(&head->next);
540 head->func = func;
541 crdp = get_call_rcu_data();
542 cds_wfq_enqueue(&crdp->cbs, &head->next);
543 uatomic_inc(&crdp->qlen);
544 wake_call_rcu_thread(crdp);
545 }
546
547 /*
548 * Free up the specified call_rcu_data structure, terminating the
549 * associated call_rcu thread. The caller must have previously
550 * removed the call_rcu_data structure from per-thread or per-CPU
551 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
552 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
553 * per-thread call_rcu_data structures.
554 *
555 * We silently refuse to free up the default call_rcu_data structure
556 * because that is where we put any leftover callbacks. Note that
557 * the possibility of self-spawning callbacks makes it impossible
558 * to execute all the callbacks in finite time without putting any
559 * newly spawned callbacks somewhere else. The "somewhere else" of
560 * last resort is the default call_rcu_data structure.
561 *
562 * We also silently refuse to free NULL pointers. This simplifies
563 * the calling code.
564 */
565 void call_rcu_data_free(struct call_rcu_data *crdp)
566 {
567 struct cds_wfq_node *cbs;
568 struct cds_wfq_node **cbs_tail;
569 struct cds_wfq_node **cbs_endprev;
570
571 if (crdp == NULL || crdp == default_call_rcu_data) {
572 return;
573 }
574 if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
575 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
576 wake_call_rcu_thread(crdp);
577 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
578 poll(NULL, 0, 1);
579 }
580 if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) {
581 while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL)
582 poll(NULL, 0, 1);
583 _CMM_STORE_SHARED(crdp->cbs.head, NULL);
584 cbs_tail = (struct cds_wfq_node **)
585 uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head);
586 cbs_endprev = (struct cds_wfq_node **)
587 uatomic_xchg(&default_call_rcu_data, cbs_tail);
588 *cbs_endprev = cbs;
589 uatomic_add(&default_call_rcu_data->qlen,
590 uatomic_read(&crdp->qlen));
591 cds_list_del(&crdp->list);
592 free(crdp);
593 }
594 }
595
596 /*
597 * Clean up all the per-CPU call_rcu threads.
598 */
599 void free_all_cpu_call_rcu_data(void)
600 {
601 int cpu;
602 struct call_rcu_data *crdp;
603
604 if (maxcpus <= 0)
605 return;
606 for (cpu = 0; cpu < maxcpus; cpu++) {
607 crdp = get_cpu_call_rcu_data(cpu);
608 if (crdp == NULL)
609 continue;
610 set_cpu_call_rcu_data(cpu, NULL);
611 call_rcu_data_free(crdp);
612 }
613 }
614
615 /*
616 * Acquire the call_rcu_mutex in order to ensure that the child sees
617 * all of the call_rcu() data structures in a consistent state.
618 * Suitable for pthread_atfork() and friends.
619 */
620 void call_rcu_before_fork(void)
621 {
622 call_rcu_lock(&call_rcu_mutex);
623 }
624
625 /*
626 * Clean up call_rcu data structures in the parent of a successful fork()
627 * that is not followed by exec() in the child. Suitable for
628 * pthread_atfork() and friends.
629 */
630 void call_rcu_after_fork_parent(void)
631 {
632 call_rcu_unlock(&call_rcu_mutex);
633 }
634
635 /*
636 * Clean up call_rcu data structures in the child of a successful fork()
637 * that is not followed by exec(). Suitable for pthread_atfork() and
638 * friends.
639 */
640 void call_rcu_after_fork_child(void)
641 {
642 struct call_rcu_data *crdp;
643
644 /* Release the mutex. */
645 call_rcu_unlock(&call_rcu_mutex);
646
647 /*
648 * Allocate a new default call_rcu_data structure in order
649 * to get a working call_rcu thread to go with it.
650 */
651 default_call_rcu_data = NULL;
652 (void)get_default_call_rcu_data();
653
654 /* Dispose of all of the rest of the call_rcu_data structures. */
655 while (call_rcu_data_list.next != call_rcu_data_list.prev) {
656 crdp = cds_list_entry(call_rcu_data_list.prev,
657 struct call_rcu_data, list);
658 if (crdp == default_call_rcu_data)
659 crdp = cds_list_entry(crdp->list.prev,
660 struct call_rcu_data, list);
661 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
662 call_rcu_data_free(crdp);
663 }
664 }
This page took 0.043139 seconds and 5 git commands to generate.