Update README following API changes
[urcu.git] / src / urcu-call-rcu-impl.h
CommitLineData
b57aee66
PM
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
5161f31e 23#define _LGPL_SOURCE
b57aee66
PM
24#include <stdio.h>
25#include <pthread.h>
26#include <signal.h>
27#include <assert.h>
28#include <stdlib.h>
6d841bc2 29#include <stdint.h>
b57aee66
PM
30#include <string.h>
31#include <errno.h>
32#include <poll.h>
33#include <sys/time.h>
b57aee66 34#include <unistd.h>
c1d2c60b 35#include <sched.h>
b57aee66 36
a47dd11c 37#include "compat-getcpu.h"
4477a870
MD
38#include <urcu/wfcqueue.h>
39#include <urcu/call-rcu.h>
40#include <urcu/pointer.h>
41#include <urcu/list.h>
42#include <urcu/futex.h>
43#include <urcu/tls-compat.h>
44#include <urcu/ref.h>
4a6d7378 45#include "urcu-die.h"
4477a870 46#include "urcu-utils.h"
b57aee66 47
8e3690db
MD
48#define SET_AFFINITY_CHECK_PERIOD (1U << 8) /* 256 */
49#define SET_AFFINITY_CHECK_PERIOD_MASK (SET_AFFINITY_CHECK_PERIOD - 1)
50
b57aee66
PM
51/* Data structure that identifies a call_rcu thread. */
52
53struct call_rcu_data {
5161f31e 54 /*
0b8ab7df
MD
55 * We do not align head on a different cache-line than tail
56 * mainly because call_rcu callback-invocation threads use
57 * batching ("splice") to get an entire list of callbacks, which
58 * effectively empties the queue, and requires to touch the tail
59 * anyway.
5161f31e 60 */
b9f893b6 61 struct cds_wfcq_tail cbs_tail;
0b8ab7df 62 struct cds_wfcq_head cbs_head;
b57aee66 63 unsigned long flags;
6d841bc2 64 int32_t futex;
73987721 65 unsigned long qlen; /* maintained for debugging. */
b57aee66 66 pthread_t tid;
c1d2c60b 67 int cpu_affinity;
8e3690db 68 unsigned long gp_count;
3c24913f 69 struct cds_list_head list;
b57aee66
PM
70} __attribute__((aligned(CAA_CACHE_LINE_SIZE)));
71
b7f721d9
MD
72struct call_rcu_completion {
73 int barrier_count;
74 int32_t futex;
81dd9134 75 struct urcu_ref ref;
b7f721d9
MD
76};
77
78struct call_rcu_completion_work {
79 struct rcu_head head;
80 struct call_rcu_completion *completion;
81};
82
3c24913f
PM
83/*
84 * List of all call_rcu_data structures to keep valgrind happy.
85 * Protected by call_rcu_mutex.
86 */
87
bab44e28 88static CDS_LIST_HEAD(call_rcu_data_list);
3c24913f 89
b57aee66
PM
90/* Link a thread using call_rcu() to its call_rcu thread. */
91
2f661865 92static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data);
b57aee66 93
e85451a1
MD
94/*
95 * Guard call_rcu thread creation and atfork handlers.
96 */
b57aee66
PM
97static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER;
98
99/* If a given thread does not have its own call_rcu thread, this is default. */
100
101static struct call_rcu_data *default_call_rcu_data;
102
d0ec0ed2
MD
103static struct urcu_atfork *registered_rculfhash_atfork;
104static unsigned long registered_rculfhash_atfork_refcount;
105
b57aee66
PM
106/*
107 * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are
108 * available, then we can have call_rcu threads assigned to individual
109 * CPUs rather than only to specific threads.
110 */
111
a47dd11c 112#if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID))
b57aee66
PM
113
114/*
115 * Pointer to array of pointers to per-CPU call_rcu_data structures
618b2595
MD
116 * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an
117 * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a
118 * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer
119 * without mutex. The call_rcu_mutex protects updates.
b57aee66
PM
120 */
121
122static struct call_rcu_data **per_cpu_call_rcu_data;
123static long maxcpus;
124
60af049d
LJ
125static void maxcpus_reset(void)
126{
127 maxcpus = 0;
128}
129
b57aee66
PM
130/* Allocate the array if it has not already been allocated. */
131
132static void alloc_cpu_call_rcu_data(void)
133{
134 struct call_rcu_data **p;
135 static int warned = 0;
136
137 if (maxcpus != 0)
138 return;
139 maxcpus = sysconf(_SC_NPROCESSORS_CONF);
140 if (maxcpus <= 0) {
141 return;
142 }
143 p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data));
144 if (p != NULL) {
145 memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data));
618b2595 146 rcu_set_pointer(&per_cpu_call_rcu_data, p);
b57aee66
PM
147 } else {
148 if (!warned) {
149 fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
150 }
151 warned = 1;
152 }
153}
154
63b495d8 155#else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
b57aee66 156
f9437098
MD
157/*
158 * per_cpu_call_rcu_data should be constant, but some functions below, used both
159 * for cases where cpu number is available and not available, assume it it not
160 * constant.
161 */
162static struct call_rcu_data **per_cpu_call_rcu_data = NULL;
b57aee66
PM
163static const long maxcpus = -1;
164
60af049d
LJ
165static void maxcpus_reset(void)
166{
167}
168
b57aee66
PM
169static void alloc_cpu_call_rcu_data(void)
170{
171}
172
63b495d8 173#endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */
b57aee66
PM
174
175/* Acquire the specified pthread mutex. */
176
177static void call_rcu_lock(pthread_mutex_t *pmp)
178{
4a6d7378
MD
179 int ret;
180
181 ret = pthread_mutex_lock(pmp);
182 if (ret)
183 urcu_die(ret);
b57aee66
PM
184}
185
186/* Release the specified pthread mutex. */
187
188static void call_rcu_unlock(pthread_mutex_t *pmp)
189{
4a6d7378
MD
190 int ret;
191
192 ret = pthread_mutex_unlock(pmp);
193 if (ret)
194 urcu_die(ret);
b57aee66
PM
195}
196
8e3690db
MD
197/*
198 * Periodically retry setting CPU affinity if we migrate.
199 * Losing affinity can be caused by CPU hotunplug/hotplug, or by
200 * cpuset(7).
201 */
c1d2c60b
MD
202#if HAVE_SCHED_SETAFFINITY
203static
204int set_thread_cpu_affinity(struct call_rcu_data *crdp)
205{
206 cpu_set_t mask;
8e3690db 207 int ret;
c1d2c60b
MD
208
209 if (crdp->cpu_affinity < 0)
210 return 0;
8e3690db
MD
211 if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK)
212 return 0;
213 if (urcu_sched_getcpu() == crdp->cpu_affinity)
214 return 0;
c1d2c60b
MD
215
216 CPU_ZERO(&mask);
217 CPU_SET(crdp->cpu_affinity, &mask);
218#if SCHED_SETAFFINITY_ARGS == 2
8e3690db 219 ret = sched_setaffinity(0, &mask);
c1d2c60b 220#else
8e3690db 221 ret = sched_setaffinity(0, sizeof(mask), &mask);
c1d2c60b 222#endif
8e3690db
MD
223 /*
224 * EINVAL is fine: can be caused by hotunplugged CPUs, or by
225 * cpuset(7). This is why we should always retry if we detect
226 * migration.
227 */
228 if (ret && errno == EINVAL) {
229 ret = 0;
230 errno = 0;
231 }
232 return ret;
c1d2c60b
MD
233}
234#else
235static
236int set_thread_cpu_affinity(struct call_rcu_data *crdp)
237{
238 return 0;
239}
240#endif
241
03fe58b3
MD
242static void call_rcu_wait(struct call_rcu_data *crdp)
243{
244 /* Read call_rcu list before read futex */
245 cmm_smp_mb();
b0a841b4
MD
246 if (uatomic_read(&crdp->futex) != -1)
247 return;
248 while (futex_async(&crdp->futex, FUTEX_WAIT, -1,
249 NULL, NULL, 0)) {
250 switch (errno) {
251 case EWOULDBLOCK:
252 /* Value already changed. */
253 return;
254 case EINTR:
255 /* Retry if interrupted by signal. */
256 break; /* Get out of switch. */
257 default:
258 /* Unexpected error. */
259 urcu_die(errno);
260 }
261 }
03fe58b3
MD
262}
263
264static void call_rcu_wake_up(struct call_rcu_data *crdp)
265{
266 /* Write to call_rcu list before reading/writing futex */
267 cmm_smp_mb();
a0b7f7ea 268 if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) {
03fe58b3 269 uatomic_set(&crdp->futex, 0);
b0a841b4
MD
270 if (futex_async(&crdp->futex, FUTEX_WAKE, 1,
271 NULL, NULL, 0) < 0)
272 urcu_die(errno);
03fe58b3
MD
273 }
274}
275
b7f721d9
MD
276static void call_rcu_completion_wait(struct call_rcu_completion *completion)
277{
278 /* Read completion barrier count before read futex */
279 cmm_smp_mb();
b0a841b4
MD
280 if (uatomic_read(&completion->futex) != -1)
281 return;
282 while (futex_async(&completion->futex, FUTEX_WAIT, -1,
283 NULL, NULL, 0)) {
284 switch (errno) {
285 case EWOULDBLOCK:
286 /* Value already changed. */
287 return;
288 case EINTR:
289 /* Retry if interrupted by signal. */
290 break; /* Get out of switch. */
291 default:
292 /* Unexpected error. */
293 urcu_die(errno);
294 }
295 }
b7f721d9
MD
296}
297
298static void call_rcu_completion_wake_up(struct call_rcu_completion *completion)
299{
300 /* Write to completion barrier count before reading/writing futex */
301 cmm_smp_mb();
302 if (caa_unlikely(uatomic_read(&completion->futex) == -1)) {
303 uatomic_set(&completion->futex, 0);
b0a841b4
MD
304 if (futex_async(&completion->futex, FUTEX_WAKE, 1,
305 NULL, NULL, 0) < 0)
306 urcu_die(errno);
b7f721d9
MD
307 }
308}
309
b57aee66
PM
310/* This is the code run by each call_rcu thread. */
311
312static void *call_rcu_thread(void *arg)
313{
314 unsigned long cbcount;
5161f31e 315 struct call_rcu_data *crdp = (struct call_rcu_data *) arg;
2870aa1e 316 int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT);
b57aee66 317
8e3690db 318 if (set_thread_cpu_affinity(crdp))
4a6d7378 319 urcu_die(errno);
c1d2c60b 320
765f3ead
MD
321 /*
322 * If callbacks take a read-side lock, we need to be registered.
323 */
324 rcu_register_thread();
325
bd252a04 326 URCU_TLS(thread_call_rcu_data) = crdp;
bc94ca9b
MD
327 if (!rt) {
328 uatomic_dec(&crdp->futex);
329 /* Decrement futex before reading call_rcu list */
330 cmm_smp_mb();
331 }
b57aee66 332 for (;;) {
5161f31e
MD
333 struct cds_wfcq_head cbs_tmp_head;
334 struct cds_wfcq_tail cbs_tmp_tail;
335 struct cds_wfcq_node *cbs, *cbs_tmp_n;
ae25b7e2 336 enum cds_wfcq_ret splice_ret;
5161f31e 337
8e3690db
MD
338 if (set_thread_cpu_affinity(crdp))
339 urcu_die(errno);
340
e85451a1
MD
341 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) {
342 /*
343 * Pause requested. Become quiescent: remove
344 * ourself from all global lists, and don't
345 * process any callback. The callback lists may
346 * still be non-empty though.
347 */
348 rcu_unregister_thread();
349 cmm_smp_mb__before_uatomic_or();
350 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED);
351 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0)
109105b7 352 (void) poll(NULL, 0, 1);
fc236e5e
KF
353 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED);
354 cmm_smp_mb__after_uatomic_and();
e85451a1
MD
355 rcu_register_thread();
356 }
357
5161f31e 358 cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail);
ae25b7e2
MD
359 splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head,
360 &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail);
361 assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK);
362 assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY);
363 if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) {
b57aee66
PM
364 synchronize_rcu();
365 cbcount = 0;
5161f31e
MD
366 __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head,
367 &cbs_tmp_tail, cbs, cbs_tmp_n) {
368 struct rcu_head *rhp;
369
370 rhp = caa_container_of(cbs,
371 struct rcu_head, next);
b57aee66
PM
372 rhp->func(rhp);
373 cbcount++;
5161f31e 374 }
b57aee66
PM
375 uatomic_sub(&crdp->qlen, cbcount);
376 }
bc94ca9b
MD
377 if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP)
378 break;
765f3ead 379 rcu_thread_offline();
bc94ca9b 380 if (!rt) {
5161f31e
MD
381 if (cds_wfcq_empty(&crdp->cbs_head,
382 &crdp->cbs_tail)) {
bc94ca9b 383 call_rcu_wait(crdp);
109105b7 384 (void) poll(NULL, 0, 10);
bc94ca9b 385 uatomic_dec(&crdp->futex);
c768e45e 386 /*
bc94ca9b
MD
387 * Decrement futex before reading
388 * call_rcu list.
c768e45e
MD
389 */
390 cmm_smp_mb();
ccbac24d 391 } else {
109105b7 392 (void) poll(NULL, 0, 10);
c768e45e 393 }
bc94ca9b 394 } else {
109105b7 395 (void) poll(NULL, 0, 10);
b57aee66 396 }
765f3ead 397 rcu_thread_online();
bc94ca9b
MD
398 }
399 if (!rt) {
400 /*
401 * Read call_rcu list before write futex.
402 */
403 cmm_smp_mb();
404 uatomic_set(&crdp->futex, 0);
b57aee66 405 }
2870aa1e 406 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED);
765f3ead 407 rcu_unregister_thread();
7106ddf8 408 return NULL;
b57aee66
PM
409}
410
411/*
412 * Create both a call_rcu thread and the corresponding call_rcu_data
3c24913f
PM
413 * structure, linking the structure in as specified. Caller must hold
414 * call_rcu_mutex.
b57aee66
PM
415 */
416
3c24913f 417static void call_rcu_data_init(struct call_rcu_data **crdpp,
c1d2c60b
MD
418 unsigned long flags,
419 int cpu_affinity)
b57aee66
PM
420{
421 struct call_rcu_data *crdp;
4a6d7378 422 int ret;
b57aee66
PM
423
424 crdp = malloc(sizeof(*crdp));
4a6d7378
MD
425 if (crdp == NULL)
426 urcu_die(errno);
b57aee66 427 memset(crdp, '\0', sizeof(*crdp));
5161f31e 428 cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail);
b57aee66 429 crdp->qlen = 0;
263e3cf9
MD
430 crdp->futex = 0;
431 crdp->flags = flags;
3c24913f 432 cds_list_add(&crdp->list, &call_rcu_data_list);
c1d2c60b 433 crdp->cpu_affinity = cpu_affinity;
8e3690db 434 crdp->gp_count = 0;
b57aee66
PM
435 cmm_smp_mb(); /* Structure initialized before pointer is planted. */
436 *crdpp = crdp;
4a6d7378
MD
437 ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp);
438 if (ret)
439 urcu_die(ret);
b57aee66
PM
440}
441
442/*
443 * Return a pointer to the call_rcu_data structure for the specified
444 * CPU, returning NULL if there is none. We cannot automatically
445 * created it because the platform we are running on might not define
63b495d8 446 * urcu_sched_getcpu().
618b2595
MD
447 *
448 * The call to this function and use of the returned call_rcu_data
449 * should be protected by RCU read-side lock.
b57aee66
PM
450 */
451
452struct call_rcu_data *get_cpu_call_rcu_data(int cpu)
453{
454 static int warned = 0;
618b2595 455 struct call_rcu_data **pcpu_crdp;
b57aee66 456
618b2595
MD
457 pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data);
458 if (pcpu_crdp == NULL)
b57aee66
PM
459 return NULL;
460 if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) {
461 fprintf(stderr, "[error] liburcu: get CPU # out of range\n");
462 warned = 1;
463 }
464 if (cpu < 0 || maxcpus <= cpu)
465 return NULL;
618b2595 466 return rcu_dereference(pcpu_crdp[cpu]);
b57aee66 467}
4477a870
MD
468__attribute__((alias(urcu_stringify(get_cpu_call_rcu_data))))
469struct call_rcu_data *alias_get_cpu_call_rcu_data();
b57aee66
PM
470
471/*
472 * Return the tid corresponding to the call_rcu thread whose
473 * call_rcu_data structure is specified.
474 */
475
476pthread_t get_call_rcu_thread(struct call_rcu_data *crdp)
477{
478 return crdp->tid;
479}
4477a870
MD
480__attribute__((alias(urcu_stringify(get_call_rcu_thread))))
481pthread_t alias_get_call_rcu_thread();
b57aee66
PM
482
483/*
484 * Create a call_rcu_data structure (with thread) and return a pointer.
485 */
486
c1d2c60b
MD
487static struct call_rcu_data *__create_call_rcu_data(unsigned long flags,
488 int cpu_affinity)
b57aee66
PM
489{
490 struct call_rcu_data *crdp;
491
c1d2c60b 492 call_rcu_data_init(&crdp, flags, cpu_affinity);
b57aee66
PM
493 return crdp;
494}
495
4477a870
MD
496__attribute__((alias(urcu_stringify(create_call_rcu_data))))
497struct call_rcu_data *alias_create_call_rcu_data();
c1d2c60b
MD
498struct call_rcu_data *create_call_rcu_data(unsigned long flags,
499 int cpu_affinity)
3c24913f
PM
500{
501 struct call_rcu_data *crdp;
502
503 call_rcu_lock(&call_rcu_mutex);
c1d2c60b 504 crdp = __create_call_rcu_data(flags, cpu_affinity);
3c24913f
PM
505 call_rcu_unlock(&call_rcu_mutex);
506 return crdp;
507}
508
b57aee66
PM
509/*
510 * Set the specified CPU to use the specified call_rcu_data structure.
7106ddf8
PM
511 *
512 * Use NULL to remove a CPU's call_rcu_data structure, but it is
513 * the caller's responsibility to dispose of the removed structure.
514 * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure
515 * (prior to NULLing it out, of course).
f9da0936
MD
516 *
517 * The caller must wait for a grace-period to pass between return from
518 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
519 * previous call rcu data as argument.
b57aee66
PM
520 */
521
522int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp)
523{
dcfc8165 524 static int warned = 0;
b57aee66
PM
525
526 call_rcu_lock(&call_rcu_mutex);
f3776786 527 alloc_cpu_call_rcu_data();
b57aee66
PM
528 if (cpu < 0 || maxcpus <= cpu) {
529 if (!warned) {
530 fprintf(stderr, "[error] liburcu: set CPU # out of range\n");
531 warned = 1;
532 }
533 call_rcu_unlock(&call_rcu_mutex);
534 errno = EINVAL;
535 return -EINVAL;
536 }
53a55535 537
b57aee66 538 if (per_cpu_call_rcu_data == NULL) {
3670fef2 539 call_rcu_unlock(&call_rcu_mutex);
b57aee66
PM
540 errno = ENOMEM;
541 return -ENOMEM;
542 }
53a55535
LJ
543
544 if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) {
545 call_rcu_unlock(&call_rcu_mutex);
546 errno = EEXIST;
547 return -EEXIST;
548 }
549
618b2595 550 rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp);
3670fef2 551 call_rcu_unlock(&call_rcu_mutex);
b57aee66
PM
552 return 0;
553}
4477a870
MD
554__attribute__((alias(urcu_stringify(set_cpu_call_rcu_data))))
555int alias_set_cpu_call_rcu_data();
b57aee66
PM
556
557/*
558 * Return a pointer to the default call_rcu_data structure, creating
559 * one if need be. Because we never free call_rcu_data structures,
560 * we don't need to be in an RCU read-side critical section.
561 */
562
563struct call_rcu_data *get_default_call_rcu_data(void)
564{
565 if (default_call_rcu_data != NULL)
566 return rcu_dereference(default_call_rcu_data);
567 call_rcu_lock(&call_rcu_mutex);
568 if (default_call_rcu_data != NULL) {
569 call_rcu_unlock(&call_rcu_mutex);
570 return default_call_rcu_data;
571 }
c1d2c60b 572 call_rcu_data_init(&default_call_rcu_data, 0, -1);
b57aee66
PM
573 call_rcu_unlock(&call_rcu_mutex);
574 return default_call_rcu_data;
575}
4477a870
MD
576__attribute__((alias(urcu_stringify(get_default_call_rcu_data))))
577struct call_rcu_data *alias_get_default_call_rcu_data();
b57aee66
PM
578
579/*
580 * Return the call_rcu_data structure that applies to the currently
581 * running thread. Any call_rcu_data structure assigned specifically
582 * to this thread has first priority, followed by any call_rcu_data
583 * structure assigned to the CPU on which the thread is running,
584 * followed by the default call_rcu_data structure. If there is not
585 * yet a default call_rcu_data structure, one will be created.
618b2595
MD
586 *
587 * Calls to this function and use of the returned call_rcu_data should
588 * be protected by RCU read-side lock.
b57aee66
PM
589 */
590struct call_rcu_data *get_call_rcu_data(void)
591{
9744f3bb 592 struct call_rcu_data *crd;
b57aee66 593
bd252a04
MD
594 if (URCU_TLS(thread_call_rcu_data) != NULL)
595 return URCU_TLS(thread_call_rcu_data);
9744f3bb
LJ
596
597 if (maxcpus > 0) {
63b495d8 598 crd = get_cpu_call_rcu_data(urcu_sched_getcpu());
9744f3bb
LJ
599 if (crd)
600 return crd;
b57aee66 601 }
9744f3bb 602
b57aee66
PM
603 return get_default_call_rcu_data();
604}
4477a870
MD
605__attribute__((alias(urcu_stringify(get_call_rcu_data))))
606struct call_rcu_data *alias_get_call_rcu_data();
b57aee66
PM
607
608/*
609 * Return a pointer to this task's call_rcu_data if there is one.
610 */
611
612struct call_rcu_data *get_thread_call_rcu_data(void)
613{
bd252a04 614 return URCU_TLS(thread_call_rcu_data);
b57aee66 615}
4477a870
MD
616__attribute__((alias(urcu_stringify(get_thread_call_rcu_data))))
617struct call_rcu_data *alias_get_thread_call_rcu_data();
b57aee66
PM
618
619/*
620 * Set this task's call_rcu_data structure as specified, regardless
621 * of whether or not this task already had one. (This allows switching
622 * to and from real-time call_rcu threads, for example.)
7106ddf8
PM
623 *
624 * Use NULL to remove a thread's call_rcu_data structure, but it is
625 * the caller's responsibility to dispose of the removed structure.
626 * Use get_thread_call_rcu_data() to obtain a pointer to the old structure
627 * (prior to NULLing it out, of course).
b57aee66
PM
628 */
629
630void set_thread_call_rcu_data(struct call_rcu_data *crdp)
631{
bd252a04 632 URCU_TLS(thread_call_rcu_data) = crdp;
b57aee66 633}
4477a870
MD
634__attribute__((alias(urcu_stringify(set_thread_call_rcu_data))))
635void alias_set_thread_call_rcu_data();
b57aee66
PM
636
637/*
638 * Create a separate call_rcu thread for each CPU. This does not
639 * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data()
0938c541
MD
640 * function if you want that behavior. Should be paired with
641 * free_all_cpu_call_rcu_data() to teardown these call_rcu worker
642 * threads.
b57aee66
PM
643 */
644
645int create_all_cpu_call_rcu_data(unsigned long flags)
646{
647 int i;
648 struct call_rcu_data *crdp;
649 int ret;
650
651 call_rcu_lock(&call_rcu_mutex);
652 alloc_cpu_call_rcu_data();
653 call_rcu_unlock(&call_rcu_mutex);
654 if (maxcpus <= 0) {
655 errno = EINVAL;
656 return -EINVAL;
657 }
658 if (per_cpu_call_rcu_data == NULL) {
659 errno = ENOMEM;
660 return -ENOMEM;
661 }
662 for (i = 0; i < maxcpus; i++) {
663 call_rcu_lock(&call_rcu_mutex);
664 if (get_cpu_call_rcu_data(i)) {
665 call_rcu_unlock(&call_rcu_mutex);
666 continue;
667 }
c1d2c60b 668 crdp = __create_call_rcu_data(flags, i);
b57aee66
PM
669 if (crdp == NULL) {
670 call_rcu_unlock(&call_rcu_mutex);
671 errno = ENOMEM;
672 return -ENOMEM;
673 }
674 call_rcu_unlock(&call_rcu_mutex);
675 if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) {
356c8794
LJ
676 call_rcu_data_free(crdp);
677
678 /* it has been created by other thread */
679 if (ret == -EEXIST)
680 continue;
681
682 return ret;
b57aee66
PM
683 }
684 }
685 return 0;
686}
4477a870
MD
687__attribute__((alias(urcu_stringify(create_all_cpu_call_rcu_data))))
688int alias_create_all_cpu_call_rcu_data();
b57aee66 689
7106ddf8
PM
690/*
691 * Wake up the call_rcu thread corresponding to the specified
692 * call_rcu_data structure.
693 */
694static void wake_call_rcu_thread(struct call_rcu_data *crdp)
695{
263e3cf9
MD
696 if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT))
697 call_rcu_wake_up(crdp);
7106ddf8
PM
698}
699
b7f721d9
MD
700static void _call_rcu(struct rcu_head *head,
701 void (*func)(struct rcu_head *head),
702 struct call_rcu_data *crdp)
703{
704 cds_wfcq_node_init(&head->next);
705 head->func = func;
706 cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next);
707 uatomic_inc(&crdp->qlen);
708 wake_call_rcu_thread(crdp);
709}
710
b57aee66
PM
711/*
712 * Schedule a function to be invoked after a following grace period.
713 * This is the only function that must be called -- the others are
714 * only present to allow applications to tune their use of RCU for
715 * maximum performance.
716 *
717 * Note that unless a call_rcu thread has not already been created,
718 * the first invocation of call_rcu() will create one. So, if you
719 * need the first invocation of call_rcu() to be fast, make sure
720 * to create a call_rcu thread first. One way to accomplish this is
721 * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data().
618b2595
MD
722 *
723 * call_rcu must be called by registered RCU read-side threads.
b57aee66 724 */
b57aee66
PM
725void call_rcu(struct rcu_head *head,
726 void (*func)(struct rcu_head *head))
727{
728 struct call_rcu_data *crdp;
729
618b2595 730 /* Holding rcu read-side lock across use of per-cpu crdp */
ffb4aaa7 731 _rcu_read_lock();
b57aee66 732 crdp = get_call_rcu_data();
b7f721d9 733 _call_rcu(head, func, crdp);
ffb4aaa7 734 _rcu_read_unlock();
7106ddf8 735}
4477a870 736__attribute__((alias(urcu_stringify(call_rcu)))) void alias_call_rcu();
7106ddf8
PM
737
738/*
739 * Free up the specified call_rcu_data structure, terminating the
740 * associated call_rcu thread. The caller must have previously
741 * removed the call_rcu_data structure from per-thread or per-CPU
742 * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU
743 * call_rcu_data structures or set_thread_call_rcu_data(NULL) for
744 * per-thread call_rcu_data structures.
745 *
746 * We silently refuse to free up the default call_rcu_data structure
747 * because that is where we put any leftover callbacks. Note that
748 * the possibility of self-spawning callbacks makes it impossible
749 * to execute all the callbacks in finite time without putting any
750 * newly spawned callbacks somewhere else. The "somewhere else" of
751 * last resort is the default call_rcu_data structure.
752 *
753 * We also silently refuse to free NULL pointers. This simplifies
754 * the calling code.
f9da0936
MD
755 *
756 * The caller must wait for a grace-period to pass between return from
757 * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the
758 * previous call rcu data as argument.
03e5118f
MD
759 *
760 * Note: introducing __cds_wfcq_splice_blocking() in this function fixed
761 * a list corruption bug in the 0.7.x series. The equivalent fix
762 * appeared in 0.6.8 for the stable-0.6 branch.
7106ddf8
PM
763 */
764void call_rcu_data_free(struct call_rcu_data *crdp)
765{
7106ddf8
PM
766 if (crdp == NULL || crdp == default_call_rcu_data) {
767 return;
768 }
2870aa1e
PB
769 if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) {
770 uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP);
7106ddf8 771 wake_call_rcu_thread(crdp);
2870aa1e 772 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0)
109105b7 773 (void) poll(NULL, 0, 1);
7106ddf8 774 }
5161f31e 775 if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) {
698d0778
MD
776 /* Create default call rcu data if need be */
777 (void) get_default_call_rcu_data();
5161f31e
MD
778 __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head,
779 &default_call_rcu_data->cbs_tail,
780 &crdp->cbs_head, &crdp->cbs_tail);
7106ddf8
PM
781 uatomic_add(&default_call_rcu_data->qlen,
782 uatomic_read(&crdp->qlen));
1e92aa15 783 wake_call_rcu_thread(default_call_rcu_data);
7106ddf8 784 }
59dc9e9d 785
b75dffe6 786 call_rcu_lock(&call_rcu_mutex);
59dc9e9d 787 cds_list_del(&crdp->list);
b75dffe6
LJ
788 call_rcu_unlock(&call_rcu_mutex);
789
59dc9e9d 790 free(crdp);
7106ddf8 791}
4477a870
MD
792__attribute__((alias(urcu_stringify(call_rcu_data_free))))
793void alias_call_rcu_data_free();
7106ddf8
PM
794
795/*
796 * Clean up all the per-CPU call_rcu threads.
797 */
798void free_all_cpu_call_rcu_data(void)
799{
800 int cpu;
618b2595
MD
801 struct call_rcu_data **crdp;
802 static int warned = 0;
7106ddf8
PM
803
804 if (maxcpus <= 0)
805 return;
618b2595
MD
806
807 crdp = malloc(sizeof(*crdp) * maxcpus);
808 if (!crdp) {
809 if (!warned) {
810 fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n");
811 }
812 warned = 1;
d31150b4 813 return;
618b2595
MD
814 }
815
7106ddf8 816 for (cpu = 0; cpu < maxcpus; cpu++) {
618b2595
MD
817 crdp[cpu] = get_cpu_call_rcu_data(cpu);
818 if (crdp[cpu] == NULL)
7106ddf8
PM
819 continue;
820 set_cpu_call_rcu_data(cpu, NULL);
7106ddf8 821 }
618b2595
MD
822 /*
823 * Wait for call_rcu sites acting as RCU readers of the
824 * call_rcu_data to become quiescent.
825 */
826 synchronize_rcu();
827 for (cpu = 0; cpu < maxcpus; cpu++) {
828 if (crdp[cpu] == NULL)
829 continue;
830 call_rcu_data_free(crdp[cpu]);
831 }
832 free(crdp);
7106ddf8 833}
4477a870
MD
834#ifdef RCU_QSBR
835/* ABI6 has a non-namespaced free_all_cpu_call_rcu_data for qsbr */
836#undef free_all_cpu_call_rcu_data
837__attribute__((alias("urcu_qsbr_free_all_cpu_call_rcu_data")))
838void free_all_cpu_call_rcu_data();
839#define free_all_cpu_call_rcu_data urcu_qsbr_free_all_cpu_call_rcu_data
840#else
841__attribute__((alias(urcu_stringify(free_all_cpu_call_rcu_data))))
842void alias_free_all_cpu_call_rcu_data();
843#endif
7106ddf8 844
81dd9134
KF
845static
846void free_completion(struct urcu_ref *ref)
847{
848 struct call_rcu_completion *completion;
849
850 completion = caa_container_of(ref, struct call_rcu_completion, ref);
851 free(completion);
852}
853
b7f721d9
MD
854static
855void _rcu_barrier_complete(struct rcu_head *head)
856{
857 struct call_rcu_completion_work *work;
858 struct call_rcu_completion *completion;
859
860 work = caa_container_of(head, struct call_rcu_completion_work, head);
861 completion = work->completion;
81dd9134
KF
862 if (!uatomic_sub_return(&completion->barrier_count, 1))
863 call_rcu_completion_wake_up(completion);
864 urcu_ref_put(&completion->ref, free_completion);
b7f721d9
MD
865 free(work);
866}
867
868/*
869 * Wait for all in-flight call_rcu callbacks to complete execution.
870 */
871void rcu_barrier(void)
872{
873 struct call_rcu_data *crdp;
81dd9134 874 struct call_rcu_completion *completion;
63f0fc6d 875 int count = 0;
b7f721d9
MD
876 int was_online;
877
878 /* Put in offline state in QSBR. */
ffb4aaa7 879 was_online = _rcu_read_ongoing();
b7f721d9
MD
880 if (was_online)
881 rcu_thread_offline();
882 /*
883 * Calling a rcu_barrier() within a RCU read-side critical
884 * section is an error.
885 */
ffb4aaa7 886 if (_rcu_read_ongoing()) {
b7f721d9
MD
887 static int warned = 0;
888
889 if (!warned) {
890 fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n");
891 }
892 warned = 1;
893 goto online;
894 }
895
81dd9134
KF
896 completion = calloc(sizeof(*completion), 1);
897 if (!completion)
898 urcu_die(errno);
899
b7f721d9
MD
900 call_rcu_lock(&call_rcu_mutex);
901 cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
902 count++;
903
81dd9134
KF
904 /* Referenced by rcu_barrier() and each call_rcu thread. */
905 urcu_ref_set(&completion->ref, count + 1);
906 completion->barrier_count = count;
b7f721d9
MD
907
908 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
909 struct call_rcu_completion_work *work;
910
911 work = calloc(sizeof(*work), 1);
63f0fc6d
MD
912 if (!work)
913 urcu_die(errno);
81dd9134 914 work->completion = completion;
b7f721d9 915 _call_rcu(&work->head, _rcu_barrier_complete, crdp);
b7f721d9
MD
916 }
917 call_rcu_unlock(&call_rcu_mutex);
918
b7f721d9
MD
919 /* Wait for them */
920 for (;;) {
81dd9134 921 uatomic_dec(&completion->futex);
b7f721d9
MD
922 /* Decrement futex before reading barrier_count */
923 cmm_smp_mb();
81dd9134 924 if (!uatomic_read(&completion->barrier_count))
b7f721d9 925 break;
81dd9134 926 call_rcu_completion_wait(completion);
b7f721d9 927 }
81dd9134
KF
928
929 urcu_ref_put(&completion->ref, free_completion);
930
b7f721d9
MD
931online:
932 if (was_online)
933 rcu_thread_online();
934}
4477a870
MD
935__attribute__((alias(urcu_stringify(rcu_barrier))))
936void alias_rcu_barrier();
b7f721d9 937
81ad2e19
PM
938/*
939 * Acquire the call_rcu_mutex in order to ensure that the child sees
e85451a1
MD
940 * all of the call_rcu() data structures in a consistent state. Ensure
941 * that all call_rcu threads are in a quiescent state across fork.
81ad2e19
PM
942 * Suitable for pthread_atfork() and friends.
943 */
944void call_rcu_before_fork(void)
945{
e85451a1 946 struct call_rcu_data *crdp;
d0ec0ed2 947 struct urcu_atfork *atfork;
e85451a1 948
81ad2e19 949 call_rcu_lock(&call_rcu_mutex);
e85451a1 950
d0ec0ed2
MD
951 atfork = registered_rculfhash_atfork;
952 if (atfork)
953 atfork->before_fork(atfork->priv);
954
e85451a1
MD
955 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
956 uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE);
957 cmm_smp_mb__after_uatomic_or();
958 wake_call_rcu_thread(crdp);
959 }
960 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
961 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0)
109105b7 962 (void) poll(NULL, 0, 1);
e85451a1 963 }
81ad2e19 964}
4477a870
MD
965__attribute__((alias(urcu_stringify(call_rcu_before_fork))))
966void alias_call_rcu_before_fork();
81ad2e19
PM
967
968/*
969 * Clean up call_rcu data structures in the parent of a successful fork()
970 * that is not followed by exec() in the child. Suitable for
971 * pthread_atfork() and friends.
972 */
973void call_rcu_after_fork_parent(void)
974{
e85451a1 975 struct call_rcu_data *crdp;
d0ec0ed2 976 struct urcu_atfork *atfork;
e85451a1
MD
977
978 cds_list_for_each_entry(crdp, &call_rcu_data_list, list)
979 uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE);
fc236e5e
KF
980 cds_list_for_each_entry(crdp, &call_rcu_data_list, list) {
981 while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0)
109105b7 982 (void) poll(NULL, 0, 1);
fc236e5e 983 }
d0ec0ed2
MD
984 atfork = registered_rculfhash_atfork;
985 if (atfork)
986 atfork->after_fork_parent(atfork->priv);
81ad2e19
PM
987 call_rcu_unlock(&call_rcu_mutex);
988}
4477a870
MD
989__attribute__((alias(urcu_stringify(call_rcu_after_fork_parent))))
990void alias_call_rcu_after_fork_parent();
81ad2e19 991
7106ddf8
PM
992/*
993 * Clean up call_rcu data structures in the child of a successful fork()
81ad2e19
PM
994 * that is not followed by exec(). Suitable for pthread_atfork() and
995 * friends.
7106ddf8
PM
996 */
997void call_rcu_after_fork_child(void)
998{
077ff173 999 struct call_rcu_data *crdp, *next;
d0ec0ed2 1000 struct urcu_atfork *atfork;
7106ddf8 1001
81ad2e19
PM
1002 /* Release the mutex. */
1003 call_rcu_unlock(&call_rcu_mutex);
1004
d0ec0ed2
MD
1005 atfork = registered_rculfhash_atfork;
1006 if (atfork)
1007 atfork->after_fork_child(atfork->priv);
1008
ad1b9909
LJ
1009 /* Do nothing when call_rcu() has not been used */
1010 if (cds_list_empty(&call_rcu_data_list))
1011 return;
1012
7106ddf8
PM
1013 /*
1014 * Allocate a new default call_rcu_data structure in order
1015 * to get a working call_rcu thread to go with it.
1016 */
1017 default_call_rcu_data = NULL;
1018 (void)get_default_call_rcu_data();
1019
60af049d
LJ
1020 /* Cleanup call_rcu_data pointers before use */
1021 maxcpus_reset();
1022 free(per_cpu_call_rcu_data);
618b2595 1023 rcu_set_pointer(&per_cpu_call_rcu_data, NULL);
bd252a04 1024 URCU_TLS(thread_call_rcu_data) = NULL;
60af049d 1025
e85451a1
MD
1026 /*
1027 * Dispose of all of the rest of the call_rcu_data structures.
1028 * Leftover call_rcu callbacks will be merged into the new
1029 * default call_rcu thread queue.
1030 */
077ff173 1031 cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) {
7106ddf8 1032 if (crdp == default_call_rcu_data)
077ff173 1033 continue;
2870aa1e 1034 uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED);
7106ddf8 1035 call_rcu_data_free(crdp);
b57aee66
PM
1036 }
1037}
4477a870
MD
1038__attribute__((alias(urcu_stringify(call_rcu_after_fork_child))))
1039void alias_call_rcu_after_fork_child();
d0ec0ed2
MD
1040
1041void urcu_register_rculfhash_atfork(struct urcu_atfork *atfork)
1042{
1043 call_rcu_lock(&call_rcu_mutex);
1044 if (registered_rculfhash_atfork_refcount++)
1045 goto end;
1046 registered_rculfhash_atfork = atfork;
1047end:
1048 call_rcu_unlock(&call_rcu_mutex);
1049}
4477a870
MD
1050__attribute__((alias(urcu_stringify(urcu_register_rculfhash_atfork))))
1051void alias_urcu_register_rculfhash_atfork();
d0ec0ed2
MD
1052
1053void urcu_unregister_rculfhash_atfork(struct urcu_atfork *atfork)
1054{
1055 call_rcu_lock(&call_rcu_mutex);
1056 if (--registered_rculfhash_atfork_refcount)
1057 goto end;
1058 registered_rculfhash_atfork = NULL;
1059end:
1060 call_rcu_unlock(&call_rcu_mutex);
1061}
4477a870
MD
1062__attribute__((alias(urcu_stringify(urcu_unregister_rculfhash_atfork))))
1063void alias_urcu_unregister_rculfhash_atfork();
This page took 0.096468 seconds and 4 git commands to generate.