Commit | Line | Data |
---|---|---|
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 | ||
c1d2c60b | 23 | #define _GNU_SOURCE |
5161f31e | 24 | #define _LGPL_SOURCE |
b57aee66 PM |
25 | #include <stdio.h> |
26 | #include <pthread.h> | |
27 | #include <signal.h> | |
28 | #include <assert.h> | |
29 | #include <stdlib.h> | |
6d841bc2 | 30 | #include <stdint.h> |
b57aee66 PM |
31 | #include <string.h> |
32 | #include <errno.h> | |
33 | #include <poll.h> | |
34 | #include <sys/time.h> | |
b57aee66 | 35 | #include <unistd.h> |
c1d2c60b | 36 | #include <sched.h> |
b57aee66 PM |
37 | |
38 | #include "config.h" | |
5161f31e | 39 | #include "urcu/wfcqueue.h" |
b57aee66 PM |
40 | #include "urcu-call-rcu.h" |
41 | #include "urcu-pointer.h" | |
3c24913f | 42 | #include "urcu/list.h" |
41849996 | 43 | #include "urcu/futex.h" |
bd252a04 | 44 | #include "urcu/tls-compat.h" |
4a6d7378 | 45 | #include "urcu-die.h" |
b57aee66 PM |
46 | |
47 | /* Data structure that identifies a call_rcu thread. */ | |
48 | ||
49 | struct call_rcu_data { | |
5161f31e | 50 | /* |
0b8ab7df MD |
51 | * We do not align head on a different cache-line than tail |
52 | * mainly because call_rcu callback-invocation threads use | |
53 | * batching ("splice") to get an entire list of callbacks, which | |
54 | * effectively empties the queue, and requires to touch the tail | |
55 | * anyway. | |
5161f31e | 56 | */ |
b9f893b6 | 57 | struct cds_wfcq_tail cbs_tail; |
0b8ab7df | 58 | struct cds_wfcq_head cbs_head; |
b57aee66 | 59 | unsigned long flags; |
6d841bc2 | 60 | int32_t futex; |
73987721 | 61 | unsigned long qlen; /* maintained for debugging. */ |
b57aee66 | 62 | pthread_t tid; |
c1d2c60b | 63 | int cpu_affinity; |
3c24913f | 64 | struct cds_list_head list; |
b57aee66 PM |
65 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
66 | ||
3c24913f PM |
67 | /* |
68 | * List of all call_rcu_data structures to keep valgrind happy. | |
69 | * Protected by call_rcu_mutex. | |
70 | */ | |
71 | ||
bab44e28 | 72 | static CDS_LIST_HEAD(call_rcu_data_list); |
3c24913f | 73 | |
b57aee66 PM |
74 | /* Link a thread using call_rcu() to its call_rcu thread. */ |
75 | ||
bd252a04 | 76 | static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data); |
b57aee66 PM |
77 | |
78 | /* Guard call_rcu thread creation. */ | |
79 | ||
80 | static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER; | |
81 | ||
82 | /* If a given thread does not have its own call_rcu thread, this is default. */ | |
83 | ||
84 | static struct call_rcu_data *default_call_rcu_data; | |
85 | ||
b57aee66 PM |
86 | /* |
87 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are | |
88 | * available, then we can have call_rcu threads assigned to individual | |
89 | * CPUs rather than only to specific threads. | |
90 | */ | |
91 | ||
92 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) | |
93 | ||
94 | /* | |
95 | * Pointer to array of pointers to per-CPU call_rcu_data structures | |
618b2595 MD |
96 | * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an |
97 | * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a | |
98 | * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer | |
99 | * without mutex. The call_rcu_mutex protects updates. | |
b57aee66 PM |
100 | */ |
101 | ||
102 | static struct call_rcu_data **per_cpu_call_rcu_data; | |
103 | static long maxcpus; | |
104 | ||
60af049d LJ |
105 | static void maxcpus_reset(void) |
106 | { | |
107 | maxcpus = 0; | |
108 | } | |
109 | ||
b57aee66 PM |
110 | /* Allocate the array if it has not already been allocated. */ |
111 | ||
112 | static void alloc_cpu_call_rcu_data(void) | |
113 | { | |
114 | struct call_rcu_data **p; | |
115 | static int warned = 0; | |
116 | ||
117 | if (maxcpus != 0) | |
118 | return; | |
119 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
120 | if (maxcpus <= 0) { | |
121 | return; | |
122 | } | |
123 | p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
124 | if (p != NULL) { | |
125 | memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
618b2595 | 126 | rcu_set_pointer(&per_cpu_call_rcu_data, p); |
b57aee66 PM |
127 | } else { |
128 | if (!warned) { | |
129 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
130 | } | |
131 | warned = 1; | |
132 | } | |
133 | } | |
134 | ||
135 | #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
136 | ||
f9437098 MD |
137 | /* |
138 | * per_cpu_call_rcu_data should be constant, but some functions below, used both | |
139 | * for cases where cpu number is available and not available, assume it it not | |
140 | * constant. | |
141 | */ | |
142 | static struct call_rcu_data **per_cpu_call_rcu_data = NULL; | |
b57aee66 PM |
143 | static const long maxcpus = -1; |
144 | ||
60af049d LJ |
145 | static void maxcpus_reset(void) |
146 | { | |
147 | } | |
148 | ||
b57aee66 PM |
149 | static void alloc_cpu_call_rcu_data(void) |
150 | { | |
151 | } | |
152 | ||
153 | static int sched_getcpu(void) | |
154 | { | |
155 | return -1; | |
156 | } | |
157 | ||
158 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
159 | ||
160 | /* Acquire the specified pthread mutex. */ | |
161 | ||
162 | static void call_rcu_lock(pthread_mutex_t *pmp) | |
163 | { | |
4a6d7378 MD |
164 | int ret; |
165 | ||
166 | ret = pthread_mutex_lock(pmp); | |
167 | if (ret) | |
168 | urcu_die(ret); | |
b57aee66 PM |
169 | } |
170 | ||
171 | /* Release the specified pthread mutex. */ | |
172 | ||
173 | static void call_rcu_unlock(pthread_mutex_t *pmp) | |
174 | { | |
4a6d7378 MD |
175 | int ret; |
176 | ||
177 | ret = pthread_mutex_unlock(pmp); | |
178 | if (ret) | |
179 | urcu_die(ret); | |
b57aee66 PM |
180 | } |
181 | ||
c1d2c60b MD |
182 | #if HAVE_SCHED_SETAFFINITY |
183 | static | |
184 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
185 | { | |
186 | cpu_set_t mask; | |
187 | ||
188 | if (crdp->cpu_affinity < 0) | |
189 | return 0; | |
190 | ||
191 | CPU_ZERO(&mask); | |
192 | CPU_SET(crdp->cpu_affinity, &mask); | |
193 | #if SCHED_SETAFFINITY_ARGS == 2 | |
194 | return sched_setaffinity(0, &mask); | |
195 | #else | |
196 | return sched_setaffinity(0, sizeof(mask), &mask); | |
197 | #endif | |
198 | } | |
199 | #else | |
200 | static | |
201 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
202 | { | |
203 | return 0; | |
204 | } | |
205 | #endif | |
206 | ||
03fe58b3 MD |
207 | static void call_rcu_wait(struct call_rcu_data *crdp) |
208 | { | |
209 | /* Read call_rcu list before read futex */ | |
210 | cmm_smp_mb(); | |
211 | if (uatomic_read(&crdp->futex) == -1) | |
212 | futex_async(&crdp->futex, FUTEX_WAIT, -1, | |
213 | NULL, NULL, 0); | |
214 | } | |
215 | ||
216 | static void call_rcu_wake_up(struct call_rcu_data *crdp) | |
217 | { | |
218 | /* Write to call_rcu list before reading/writing futex */ | |
219 | cmm_smp_mb(); | |
a0b7f7ea | 220 | if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) { |
03fe58b3 MD |
221 | uatomic_set(&crdp->futex, 0); |
222 | futex_async(&crdp->futex, FUTEX_WAKE, 1, | |
223 | NULL, NULL, 0); | |
224 | } | |
225 | } | |
226 | ||
b57aee66 PM |
227 | /* This is the code run by each call_rcu thread. */ |
228 | ||
229 | static void *call_rcu_thread(void *arg) | |
230 | { | |
231 | unsigned long cbcount; | |
5161f31e | 232 | struct call_rcu_data *crdp = (struct call_rcu_data *) arg; |
2870aa1e | 233 | int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT); |
4a6d7378 | 234 | int ret; |
b57aee66 | 235 | |
4a6d7378 MD |
236 | ret = set_thread_cpu_affinity(crdp); |
237 | if (ret) | |
238 | urcu_die(errno); | |
c1d2c60b | 239 | |
765f3ead MD |
240 | /* |
241 | * If callbacks take a read-side lock, we need to be registered. | |
242 | */ | |
243 | rcu_register_thread(); | |
244 | ||
bd252a04 | 245 | URCU_TLS(thread_call_rcu_data) = crdp; |
bc94ca9b MD |
246 | if (!rt) { |
247 | uatomic_dec(&crdp->futex); | |
248 | /* Decrement futex before reading call_rcu list */ | |
249 | cmm_smp_mb(); | |
250 | } | |
b57aee66 | 251 | for (;;) { |
5161f31e MD |
252 | struct cds_wfcq_head cbs_tmp_head; |
253 | struct cds_wfcq_tail cbs_tmp_tail; | |
254 | struct cds_wfcq_node *cbs, *cbs_tmp_n; | |
255 | ||
256 | cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail); | |
257 | __cds_wfcq_splice_blocking(&cbs_tmp_head, &cbs_tmp_tail, | |
258 | &crdp->cbs_head, &crdp->cbs_tail); | |
259 | if (!cds_wfcq_empty(&cbs_tmp_head, &cbs_tmp_tail)) { | |
b57aee66 PM |
260 | synchronize_rcu(); |
261 | cbcount = 0; | |
5161f31e MD |
262 | __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head, |
263 | &cbs_tmp_tail, cbs, cbs_tmp_n) { | |
264 | struct rcu_head *rhp; | |
265 | ||
266 | rhp = caa_container_of(cbs, | |
267 | struct rcu_head, next); | |
b57aee66 PM |
268 | rhp->func(rhp); |
269 | cbcount++; | |
5161f31e | 270 | } |
b57aee66 PM |
271 | uatomic_sub(&crdp->qlen, cbcount); |
272 | } | |
bc94ca9b MD |
273 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP) |
274 | break; | |
765f3ead | 275 | rcu_thread_offline(); |
bc94ca9b | 276 | if (!rt) { |
5161f31e MD |
277 | if (cds_wfcq_empty(&crdp->cbs_head, |
278 | &crdp->cbs_tail)) { | |
bc94ca9b MD |
279 | call_rcu_wait(crdp); |
280 | poll(NULL, 0, 10); | |
281 | uatomic_dec(&crdp->futex); | |
c768e45e | 282 | /* |
bc94ca9b MD |
283 | * Decrement futex before reading |
284 | * call_rcu list. | |
c768e45e MD |
285 | */ |
286 | cmm_smp_mb(); | |
ccbac24d MD |
287 | } else { |
288 | poll(NULL, 0, 10); | |
c768e45e | 289 | } |
bc94ca9b MD |
290 | } else { |
291 | poll(NULL, 0, 10); | |
b57aee66 | 292 | } |
765f3ead | 293 | rcu_thread_online(); |
bc94ca9b MD |
294 | } |
295 | if (!rt) { | |
296 | /* | |
297 | * Read call_rcu list before write futex. | |
298 | */ | |
299 | cmm_smp_mb(); | |
300 | uatomic_set(&crdp->futex, 0); | |
b57aee66 | 301 | } |
2870aa1e | 302 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED); |
765f3ead | 303 | rcu_unregister_thread(); |
7106ddf8 | 304 | return NULL; |
b57aee66 PM |
305 | } |
306 | ||
307 | /* | |
308 | * Create both a call_rcu thread and the corresponding call_rcu_data | |
3c24913f PM |
309 | * structure, linking the structure in as specified. Caller must hold |
310 | * call_rcu_mutex. | |
b57aee66 PM |
311 | */ |
312 | ||
3c24913f | 313 | static void call_rcu_data_init(struct call_rcu_data **crdpp, |
c1d2c60b MD |
314 | unsigned long flags, |
315 | int cpu_affinity) | |
b57aee66 PM |
316 | { |
317 | struct call_rcu_data *crdp; | |
4a6d7378 | 318 | int ret; |
b57aee66 PM |
319 | |
320 | crdp = malloc(sizeof(*crdp)); | |
4a6d7378 MD |
321 | if (crdp == NULL) |
322 | urcu_die(errno); | |
b57aee66 | 323 | memset(crdp, '\0', sizeof(*crdp)); |
5161f31e | 324 | cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail); |
b57aee66 | 325 | crdp->qlen = 0; |
263e3cf9 MD |
326 | crdp->futex = 0; |
327 | crdp->flags = flags; | |
3c24913f | 328 | cds_list_add(&crdp->list, &call_rcu_data_list); |
c1d2c60b | 329 | crdp->cpu_affinity = cpu_affinity; |
b57aee66 PM |
330 | cmm_smp_mb(); /* Structure initialized before pointer is planted. */ |
331 | *crdpp = crdp; | |
4a6d7378 MD |
332 | ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp); |
333 | if (ret) | |
334 | urcu_die(ret); | |
b57aee66 PM |
335 | } |
336 | ||
337 | /* | |
338 | * Return a pointer to the call_rcu_data structure for the specified | |
339 | * CPU, returning NULL if there is none. We cannot automatically | |
340 | * created it because the platform we are running on might not define | |
341 | * sched_getcpu(). | |
618b2595 MD |
342 | * |
343 | * The call to this function and use of the returned call_rcu_data | |
344 | * should be protected by RCU read-side lock. | |
b57aee66 PM |
345 | */ |
346 | ||
347 | struct call_rcu_data *get_cpu_call_rcu_data(int cpu) | |
348 | { | |
349 | static int warned = 0; | |
618b2595 | 350 | struct call_rcu_data **pcpu_crdp; |
b57aee66 | 351 | |
618b2595 MD |
352 | pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data); |
353 | if (pcpu_crdp == NULL) | |
b57aee66 PM |
354 | return NULL; |
355 | if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) { | |
356 | fprintf(stderr, "[error] liburcu: get CPU # out of range\n"); | |
357 | warned = 1; | |
358 | } | |
359 | if (cpu < 0 || maxcpus <= cpu) | |
360 | return NULL; | |
618b2595 | 361 | return rcu_dereference(pcpu_crdp[cpu]); |
b57aee66 PM |
362 | } |
363 | ||
364 | /* | |
365 | * Return the tid corresponding to the call_rcu thread whose | |
366 | * call_rcu_data structure is specified. | |
367 | */ | |
368 | ||
369 | pthread_t get_call_rcu_thread(struct call_rcu_data *crdp) | |
370 | { | |
371 | return crdp->tid; | |
372 | } | |
373 | ||
374 | /* | |
375 | * Create a call_rcu_data structure (with thread) and return a pointer. | |
376 | */ | |
377 | ||
c1d2c60b MD |
378 | static struct call_rcu_data *__create_call_rcu_data(unsigned long flags, |
379 | int cpu_affinity) | |
b57aee66 PM |
380 | { |
381 | struct call_rcu_data *crdp; | |
382 | ||
c1d2c60b | 383 | call_rcu_data_init(&crdp, flags, cpu_affinity); |
b57aee66 PM |
384 | return crdp; |
385 | } | |
386 | ||
c1d2c60b MD |
387 | struct call_rcu_data *create_call_rcu_data(unsigned long flags, |
388 | int cpu_affinity) | |
3c24913f PM |
389 | { |
390 | struct call_rcu_data *crdp; | |
391 | ||
392 | call_rcu_lock(&call_rcu_mutex); | |
c1d2c60b | 393 | crdp = __create_call_rcu_data(flags, cpu_affinity); |
3c24913f PM |
394 | call_rcu_unlock(&call_rcu_mutex); |
395 | return crdp; | |
396 | } | |
397 | ||
b57aee66 PM |
398 | /* |
399 | * Set the specified CPU to use the specified call_rcu_data structure. | |
7106ddf8 PM |
400 | * |
401 | * Use NULL to remove a CPU's call_rcu_data structure, but it is | |
402 | * the caller's responsibility to dispose of the removed structure. | |
403 | * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure | |
404 | * (prior to NULLing it out, of course). | |
f9da0936 MD |
405 | * |
406 | * The caller must wait for a grace-period to pass between return from | |
407 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
408 | * previous call rcu data as argument. | |
b57aee66 PM |
409 | */ |
410 | ||
411 | int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp) | |
412 | { | |
dcfc8165 | 413 | static int warned = 0; |
b57aee66 PM |
414 | |
415 | call_rcu_lock(&call_rcu_mutex); | |
f3776786 | 416 | alloc_cpu_call_rcu_data(); |
b57aee66 PM |
417 | if (cpu < 0 || maxcpus <= cpu) { |
418 | if (!warned) { | |
419 | fprintf(stderr, "[error] liburcu: set CPU # out of range\n"); | |
420 | warned = 1; | |
421 | } | |
422 | call_rcu_unlock(&call_rcu_mutex); | |
423 | errno = EINVAL; | |
424 | return -EINVAL; | |
425 | } | |
53a55535 | 426 | |
b57aee66 | 427 | if (per_cpu_call_rcu_data == NULL) { |
3670fef2 | 428 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
429 | errno = ENOMEM; |
430 | return -ENOMEM; | |
431 | } | |
53a55535 LJ |
432 | |
433 | if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) { | |
434 | call_rcu_unlock(&call_rcu_mutex); | |
435 | errno = EEXIST; | |
436 | return -EEXIST; | |
437 | } | |
438 | ||
618b2595 | 439 | rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp); |
3670fef2 | 440 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
441 | return 0; |
442 | } | |
443 | ||
444 | /* | |
445 | * Return a pointer to the default call_rcu_data structure, creating | |
446 | * one if need be. Because we never free call_rcu_data structures, | |
447 | * we don't need to be in an RCU read-side critical section. | |
448 | */ | |
449 | ||
450 | struct call_rcu_data *get_default_call_rcu_data(void) | |
451 | { | |
452 | if (default_call_rcu_data != NULL) | |
453 | return rcu_dereference(default_call_rcu_data); | |
454 | call_rcu_lock(&call_rcu_mutex); | |
455 | if (default_call_rcu_data != NULL) { | |
456 | call_rcu_unlock(&call_rcu_mutex); | |
457 | return default_call_rcu_data; | |
458 | } | |
c1d2c60b | 459 | call_rcu_data_init(&default_call_rcu_data, 0, -1); |
b57aee66 PM |
460 | call_rcu_unlock(&call_rcu_mutex); |
461 | return default_call_rcu_data; | |
462 | } | |
463 | ||
464 | /* | |
465 | * Return the call_rcu_data structure that applies to the currently | |
466 | * running thread. Any call_rcu_data structure assigned specifically | |
467 | * to this thread has first priority, followed by any call_rcu_data | |
468 | * structure assigned to the CPU on which the thread is running, | |
469 | * followed by the default call_rcu_data structure. If there is not | |
470 | * yet a default call_rcu_data structure, one will be created. | |
618b2595 MD |
471 | * |
472 | * Calls to this function and use of the returned call_rcu_data should | |
473 | * be protected by RCU read-side lock. | |
b57aee66 PM |
474 | */ |
475 | struct call_rcu_data *get_call_rcu_data(void) | |
476 | { | |
9744f3bb | 477 | struct call_rcu_data *crd; |
b57aee66 | 478 | |
bd252a04 MD |
479 | if (URCU_TLS(thread_call_rcu_data) != NULL) |
480 | return URCU_TLS(thread_call_rcu_data); | |
9744f3bb LJ |
481 | |
482 | if (maxcpus > 0) { | |
483 | crd = get_cpu_call_rcu_data(sched_getcpu()); | |
484 | if (crd) | |
485 | return crd; | |
b57aee66 | 486 | } |
9744f3bb | 487 | |
b57aee66 PM |
488 | return get_default_call_rcu_data(); |
489 | } | |
490 | ||
491 | /* | |
492 | * Return a pointer to this task's call_rcu_data if there is one. | |
493 | */ | |
494 | ||
495 | struct call_rcu_data *get_thread_call_rcu_data(void) | |
496 | { | |
bd252a04 | 497 | return URCU_TLS(thread_call_rcu_data); |
b57aee66 PM |
498 | } |
499 | ||
500 | /* | |
501 | * Set this task's call_rcu_data structure as specified, regardless | |
502 | * of whether or not this task already had one. (This allows switching | |
503 | * to and from real-time call_rcu threads, for example.) | |
7106ddf8 PM |
504 | * |
505 | * Use NULL to remove a thread's call_rcu_data structure, but it is | |
506 | * the caller's responsibility to dispose of the removed structure. | |
507 | * Use get_thread_call_rcu_data() to obtain a pointer to the old structure | |
508 | * (prior to NULLing it out, of course). | |
b57aee66 PM |
509 | */ |
510 | ||
511 | void set_thread_call_rcu_data(struct call_rcu_data *crdp) | |
512 | { | |
bd252a04 | 513 | URCU_TLS(thread_call_rcu_data) = crdp; |
b57aee66 PM |
514 | } |
515 | ||
516 | /* | |
517 | * Create a separate call_rcu thread for each CPU. This does not | |
518 | * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data() | |
0938c541 MD |
519 | * function if you want that behavior. Should be paired with |
520 | * free_all_cpu_call_rcu_data() to teardown these call_rcu worker | |
521 | * threads. | |
b57aee66 PM |
522 | */ |
523 | ||
524 | int create_all_cpu_call_rcu_data(unsigned long flags) | |
525 | { | |
526 | int i; | |
527 | struct call_rcu_data *crdp; | |
528 | int ret; | |
529 | ||
530 | call_rcu_lock(&call_rcu_mutex); | |
531 | alloc_cpu_call_rcu_data(); | |
532 | call_rcu_unlock(&call_rcu_mutex); | |
533 | if (maxcpus <= 0) { | |
534 | errno = EINVAL; | |
535 | return -EINVAL; | |
536 | } | |
537 | if (per_cpu_call_rcu_data == NULL) { | |
538 | errno = ENOMEM; | |
539 | return -ENOMEM; | |
540 | } | |
541 | for (i = 0; i < maxcpus; i++) { | |
542 | call_rcu_lock(&call_rcu_mutex); | |
543 | if (get_cpu_call_rcu_data(i)) { | |
544 | call_rcu_unlock(&call_rcu_mutex); | |
545 | continue; | |
546 | } | |
c1d2c60b | 547 | crdp = __create_call_rcu_data(flags, i); |
b57aee66 PM |
548 | if (crdp == NULL) { |
549 | call_rcu_unlock(&call_rcu_mutex); | |
550 | errno = ENOMEM; | |
551 | return -ENOMEM; | |
552 | } | |
553 | call_rcu_unlock(&call_rcu_mutex); | |
554 | if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) { | |
356c8794 LJ |
555 | call_rcu_data_free(crdp); |
556 | ||
557 | /* it has been created by other thread */ | |
558 | if (ret == -EEXIST) | |
559 | continue; | |
560 | ||
561 | return ret; | |
b57aee66 PM |
562 | } |
563 | } | |
564 | return 0; | |
565 | } | |
566 | ||
7106ddf8 PM |
567 | /* |
568 | * Wake up the call_rcu thread corresponding to the specified | |
569 | * call_rcu_data structure. | |
570 | */ | |
571 | static void wake_call_rcu_thread(struct call_rcu_data *crdp) | |
572 | { | |
263e3cf9 MD |
573 | if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT)) |
574 | call_rcu_wake_up(crdp); | |
7106ddf8 PM |
575 | } |
576 | ||
b57aee66 PM |
577 | /* |
578 | * Schedule a function to be invoked after a following grace period. | |
579 | * This is the only function that must be called -- the others are | |
580 | * only present to allow applications to tune their use of RCU for | |
581 | * maximum performance. | |
582 | * | |
583 | * Note that unless a call_rcu thread has not already been created, | |
584 | * the first invocation of call_rcu() will create one. So, if you | |
585 | * need the first invocation of call_rcu() to be fast, make sure | |
586 | * to create a call_rcu thread first. One way to accomplish this is | |
587 | * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data(). | |
618b2595 MD |
588 | * |
589 | * call_rcu must be called by registered RCU read-side threads. | |
b57aee66 PM |
590 | */ |
591 | ||
592 | void call_rcu(struct rcu_head *head, | |
593 | void (*func)(struct rcu_head *head)) | |
594 | { | |
595 | struct call_rcu_data *crdp; | |
596 | ||
5161f31e | 597 | cds_wfcq_node_init(&head->next); |
b57aee66 | 598 | head->func = func; |
618b2595 MD |
599 | /* Holding rcu read-side lock across use of per-cpu crdp */ |
600 | rcu_read_lock(); | |
b57aee66 | 601 | crdp = get_call_rcu_data(); |
5161f31e | 602 | cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next); |
b57aee66 | 603 | uatomic_inc(&crdp->qlen); |
7106ddf8 | 604 | wake_call_rcu_thread(crdp); |
618b2595 | 605 | rcu_read_unlock(); |
7106ddf8 PM |
606 | } |
607 | ||
608 | /* | |
609 | * Free up the specified call_rcu_data structure, terminating the | |
610 | * associated call_rcu thread. The caller must have previously | |
611 | * removed the call_rcu_data structure from per-thread or per-CPU | |
612 | * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU | |
613 | * call_rcu_data structures or set_thread_call_rcu_data(NULL) for | |
614 | * per-thread call_rcu_data structures. | |
615 | * | |
616 | * We silently refuse to free up the default call_rcu_data structure | |
617 | * because that is where we put any leftover callbacks. Note that | |
618 | * the possibility of self-spawning callbacks makes it impossible | |
619 | * to execute all the callbacks in finite time without putting any | |
620 | * newly spawned callbacks somewhere else. The "somewhere else" of | |
621 | * last resort is the default call_rcu_data structure. | |
622 | * | |
623 | * We also silently refuse to free NULL pointers. This simplifies | |
624 | * the calling code. | |
f9da0936 MD |
625 | * |
626 | * The caller must wait for a grace-period to pass between return from | |
627 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
628 | * previous call rcu data as argument. | |
03e5118f MD |
629 | * |
630 | * Note: introducing __cds_wfcq_splice_blocking() in this function fixed | |
631 | * a list corruption bug in the 0.7.x series. The equivalent fix | |
632 | * appeared in 0.6.8 for the stable-0.6 branch. | |
7106ddf8 PM |
633 | */ |
634 | void call_rcu_data_free(struct call_rcu_data *crdp) | |
635 | { | |
7106ddf8 PM |
636 | if (crdp == NULL || crdp == default_call_rcu_data) { |
637 | return; | |
638 | } | |
2870aa1e PB |
639 | if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) { |
640 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP); | |
7106ddf8 | 641 | wake_call_rcu_thread(crdp); |
2870aa1e | 642 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) |
7106ddf8 PM |
643 | poll(NULL, 0, 1); |
644 | } | |
5161f31e | 645 | if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) { |
698d0778 MD |
646 | /* Create default call rcu data if need be */ |
647 | (void) get_default_call_rcu_data(); | |
5161f31e MD |
648 | __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head, |
649 | &default_call_rcu_data->cbs_tail, | |
650 | &crdp->cbs_head, &crdp->cbs_tail); | |
7106ddf8 PM |
651 | uatomic_add(&default_call_rcu_data->qlen, |
652 | uatomic_read(&crdp->qlen)); | |
1e92aa15 | 653 | wake_call_rcu_thread(default_call_rcu_data); |
7106ddf8 | 654 | } |
59dc9e9d | 655 | |
b75dffe6 | 656 | call_rcu_lock(&call_rcu_mutex); |
59dc9e9d | 657 | cds_list_del(&crdp->list); |
b75dffe6 LJ |
658 | call_rcu_unlock(&call_rcu_mutex); |
659 | ||
59dc9e9d | 660 | free(crdp); |
7106ddf8 PM |
661 | } |
662 | ||
663 | /* | |
664 | * Clean up all the per-CPU call_rcu threads. | |
665 | */ | |
666 | void free_all_cpu_call_rcu_data(void) | |
667 | { | |
668 | int cpu; | |
618b2595 MD |
669 | struct call_rcu_data **crdp; |
670 | static int warned = 0; | |
7106ddf8 PM |
671 | |
672 | if (maxcpus <= 0) | |
673 | return; | |
618b2595 MD |
674 | |
675 | crdp = malloc(sizeof(*crdp) * maxcpus); | |
676 | if (!crdp) { | |
677 | if (!warned) { | |
678 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
679 | } | |
680 | warned = 1; | |
d31150b4 | 681 | return; |
618b2595 MD |
682 | } |
683 | ||
7106ddf8 | 684 | for (cpu = 0; cpu < maxcpus; cpu++) { |
618b2595 MD |
685 | crdp[cpu] = get_cpu_call_rcu_data(cpu); |
686 | if (crdp[cpu] == NULL) | |
7106ddf8 PM |
687 | continue; |
688 | set_cpu_call_rcu_data(cpu, NULL); | |
7106ddf8 | 689 | } |
618b2595 MD |
690 | /* |
691 | * Wait for call_rcu sites acting as RCU readers of the | |
692 | * call_rcu_data to become quiescent. | |
693 | */ | |
694 | synchronize_rcu(); | |
695 | for (cpu = 0; cpu < maxcpus; cpu++) { | |
696 | if (crdp[cpu] == NULL) | |
697 | continue; | |
698 | call_rcu_data_free(crdp[cpu]); | |
699 | } | |
700 | free(crdp); | |
7106ddf8 PM |
701 | } |
702 | ||
81ad2e19 PM |
703 | /* |
704 | * Acquire the call_rcu_mutex in order to ensure that the child sees | |
705 | * all of the call_rcu() data structures in a consistent state. | |
706 | * Suitable for pthread_atfork() and friends. | |
707 | */ | |
708 | void call_rcu_before_fork(void) | |
709 | { | |
710 | call_rcu_lock(&call_rcu_mutex); | |
711 | } | |
712 | ||
713 | /* | |
714 | * Clean up call_rcu data structures in the parent of a successful fork() | |
715 | * that is not followed by exec() in the child. Suitable for | |
716 | * pthread_atfork() and friends. | |
717 | */ | |
718 | void call_rcu_after_fork_parent(void) | |
719 | { | |
720 | call_rcu_unlock(&call_rcu_mutex); | |
721 | } | |
722 | ||
7106ddf8 PM |
723 | /* |
724 | * Clean up call_rcu data structures in the child of a successful fork() | |
81ad2e19 PM |
725 | * that is not followed by exec(). Suitable for pthread_atfork() and |
726 | * friends. | |
7106ddf8 PM |
727 | */ |
728 | void call_rcu_after_fork_child(void) | |
729 | { | |
077ff173 | 730 | struct call_rcu_data *crdp, *next; |
7106ddf8 | 731 | |
81ad2e19 PM |
732 | /* Release the mutex. */ |
733 | call_rcu_unlock(&call_rcu_mutex); | |
734 | ||
ad1b9909 LJ |
735 | /* Do nothing when call_rcu() has not been used */ |
736 | if (cds_list_empty(&call_rcu_data_list)) | |
737 | return; | |
738 | ||
7106ddf8 PM |
739 | /* |
740 | * Allocate a new default call_rcu_data structure in order | |
741 | * to get a working call_rcu thread to go with it. | |
742 | */ | |
743 | default_call_rcu_data = NULL; | |
744 | (void)get_default_call_rcu_data(); | |
745 | ||
60af049d LJ |
746 | /* Cleanup call_rcu_data pointers before use */ |
747 | maxcpus_reset(); | |
748 | free(per_cpu_call_rcu_data); | |
618b2595 | 749 | rcu_set_pointer(&per_cpu_call_rcu_data, NULL); |
bd252a04 | 750 | URCU_TLS(thread_call_rcu_data) = NULL; |
60af049d | 751 | |
7106ddf8 | 752 | /* Dispose of all of the rest of the call_rcu_data structures. */ |
077ff173 | 753 | cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) { |
7106ddf8 | 754 | if (crdp == default_call_rcu_data) |
077ff173 | 755 | continue; |
2870aa1e | 756 | uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED); |
7106ddf8 | 757 | call_rcu_data_free(crdp); |
b57aee66 PM |
758 | } |
759 | } |