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