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 |
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 PM |
36 | |
37 | #include "config.h" | |
38 | #include "urcu/wfqueue.h" | |
39 | #include "urcu-call-rcu.h" | |
40 | #include "urcu-pointer.h" | |
3c24913f | 41 | #include "urcu/list.h" |
41849996 | 42 | #include "urcu/futex.h" |
bd252a04 | 43 | #include "urcu/tls-compat.h" |
4a6d7378 | 44 | #include "urcu-die.h" |
b57aee66 PM |
45 | |
46 | /* Data structure that identifies a call_rcu thread. */ | |
47 | ||
48 | struct call_rcu_data { | |
49 | struct cds_wfq_queue cbs; | |
50 | unsigned long flags; | |
6d841bc2 | 51 | int32_t futex; |
73987721 | 52 | unsigned long qlen; /* maintained for debugging. */ |
b57aee66 | 53 | pthread_t tid; |
c1d2c60b | 54 | int cpu_affinity; |
3c24913f | 55 | struct cds_list_head list; |
b57aee66 PM |
56 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
57 | ||
3c24913f PM |
58 | /* |
59 | * List of all call_rcu_data structures to keep valgrind happy. | |
60 | * Protected by call_rcu_mutex. | |
61 | */ | |
62 | ||
63 | CDS_LIST_HEAD(call_rcu_data_list); | |
64 | ||
b57aee66 PM |
65 | /* Link a thread using call_rcu() to its call_rcu thread. */ |
66 | ||
bd252a04 | 67 | static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data); |
b57aee66 | 68 | |
750cfec5 MD |
69 | /* |
70 | * Guard call_rcu thread creation and atfork handlers. | |
71 | */ | |
b57aee66 PM |
72 | static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER; |
73 | ||
74 | /* If a given thread does not have its own call_rcu thread, this is default. */ | |
75 | ||
76 | static struct call_rcu_data *default_call_rcu_data; | |
77 | ||
b57aee66 PM |
78 | /* |
79 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are | |
80 | * available, then we can have call_rcu threads assigned to individual | |
81 | * CPUs rather than only to specific threads. | |
82 | */ | |
83 | ||
84 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) | |
85 | ||
86 | /* | |
87 | * Pointer to array of pointers to per-CPU call_rcu_data structures | |
618b2595 MD |
88 | * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an |
89 | * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a | |
90 | * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer | |
91 | * without mutex. The call_rcu_mutex protects updates. | |
b57aee66 PM |
92 | */ |
93 | ||
94 | static struct call_rcu_data **per_cpu_call_rcu_data; | |
95 | static long maxcpus; | |
96 | ||
60af049d LJ |
97 | static void maxcpus_reset(void) |
98 | { | |
99 | maxcpus = 0; | |
100 | } | |
101 | ||
b57aee66 PM |
102 | /* Allocate the array if it has not already been allocated. */ |
103 | ||
104 | static void alloc_cpu_call_rcu_data(void) | |
105 | { | |
106 | struct call_rcu_data **p; | |
107 | static int warned = 0; | |
108 | ||
109 | if (maxcpus != 0) | |
110 | return; | |
111 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
112 | if (maxcpus <= 0) { | |
113 | return; | |
114 | } | |
115 | p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
116 | if (p != NULL) { | |
117 | memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
618b2595 | 118 | rcu_set_pointer(&per_cpu_call_rcu_data, p); |
b57aee66 PM |
119 | } else { |
120 | if (!warned) { | |
121 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
122 | } | |
123 | warned = 1; | |
124 | } | |
125 | } | |
126 | ||
127 | #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
128 | ||
f9437098 MD |
129 | /* |
130 | * per_cpu_call_rcu_data should be constant, but some functions below, used both | |
131 | * for cases where cpu number is available and not available, assume it it not | |
132 | * constant. | |
133 | */ | |
134 | static struct call_rcu_data **per_cpu_call_rcu_data = NULL; | |
b57aee66 PM |
135 | static const long maxcpus = -1; |
136 | ||
60af049d LJ |
137 | static void maxcpus_reset(void) |
138 | { | |
139 | } | |
140 | ||
b57aee66 PM |
141 | static void alloc_cpu_call_rcu_data(void) |
142 | { | |
143 | } | |
144 | ||
145 | static int sched_getcpu(void) | |
146 | { | |
147 | return -1; | |
148 | } | |
149 | ||
150 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ | |
151 | ||
152 | /* Acquire the specified pthread mutex. */ | |
153 | ||
154 | static void call_rcu_lock(pthread_mutex_t *pmp) | |
155 | { | |
4a6d7378 MD |
156 | int ret; |
157 | ||
158 | ret = pthread_mutex_lock(pmp); | |
159 | if (ret) | |
160 | urcu_die(ret); | |
b57aee66 PM |
161 | } |
162 | ||
163 | /* Release the specified pthread mutex. */ | |
164 | ||
165 | static void call_rcu_unlock(pthread_mutex_t *pmp) | |
166 | { | |
4a6d7378 MD |
167 | int ret; |
168 | ||
169 | ret = pthread_mutex_unlock(pmp); | |
170 | if (ret) | |
171 | urcu_die(ret); | |
b57aee66 PM |
172 | } |
173 | ||
c1d2c60b MD |
174 | #if HAVE_SCHED_SETAFFINITY |
175 | static | |
176 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
177 | { | |
178 | cpu_set_t mask; | |
179 | ||
180 | if (crdp->cpu_affinity < 0) | |
181 | return 0; | |
182 | ||
183 | CPU_ZERO(&mask); | |
184 | CPU_SET(crdp->cpu_affinity, &mask); | |
185 | #if SCHED_SETAFFINITY_ARGS == 2 | |
186 | return sched_setaffinity(0, &mask); | |
187 | #else | |
188 | return sched_setaffinity(0, sizeof(mask), &mask); | |
189 | #endif | |
190 | } | |
191 | #else | |
192 | static | |
193 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
194 | { | |
195 | return 0; | |
196 | } | |
197 | #endif | |
198 | ||
03fe58b3 MD |
199 | static void call_rcu_wait(struct call_rcu_data *crdp) |
200 | { | |
201 | /* Read call_rcu list before read futex */ | |
202 | cmm_smp_mb(); | |
203 | if (uatomic_read(&crdp->futex) == -1) | |
204 | futex_async(&crdp->futex, FUTEX_WAIT, -1, | |
205 | NULL, NULL, 0); | |
206 | } | |
207 | ||
208 | static void call_rcu_wake_up(struct call_rcu_data *crdp) | |
209 | { | |
210 | /* Write to call_rcu list before reading/writing futex */ | |
211 | cmm_smp_mb(); | |
a0b7f7ea | 212 | if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) { |
03fe58b3 MD |
213 | uatomic_set(&crdp->futex, 0); |
214 | futex_async(&crdp->futex, FUTEX_WAKE, 1, | |
215 | NULL, NULL, 0); | |
216 | } | |
217 | } | |
218 | ||
b57aee66 PM |
219 | /* This is the code run by each call_rcu thread. */ |
220 | ||
221 | static void *call_rcu_thread(void *arg) | |
222 | { | |
223 | unsigned long cbcount; | |
224 | struct cds_wfq_node *cbs; | |
225 | struct cds_wfq_node **cbs_tail; | |
226 | struct call_rcu_data *crdp = (struct call_rcu_data *)arg; | |
227 | struct rcu_head *rhp; | |
2870aa1e | 228 | int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT); |
4a6d7378 | 229 | int ret; |
b57aee66 | 230 | |
4a6d7378 MD |
231 | ret = set_thread_cpu_affinity(crdp); |
232 | if (ret) | |
233 | urcu_die(errno); | |
c1d2c60b | 234 | |
765f3ead MD |
235 | /* |
236 | * If callbacks take a read-side lock, we need to be registered. | |
237 | */ | |
238 | rcu_register_thread(); | |
239 | ||
bd252a04 | 240 | URCU_TLS(thread_call_rcu_data) = crdp; |
bc94ca9b MD |
241 | if (!rt) { |
242 | uatomic_dec(&crdp->futex); | |
243 | /* Decrement futex before reading call_rcu list */ | |
244 | cmm_smp_mb(); | |
245 | } | |
b57aee66 | 246 | for (;;) { |
750cfec5 MD |
247 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) { |
248 | /* | |
249 | * Pause requested. Become quiescent: remove | |
250 | * ourself from all global lists, and don't | |
251 | * process any callback. The callback lists may | |
252 | * still be non-empty though. | |
253 | */ | |
254 | rcu_unregister_thread(); | |
255 | cmm_smp_mb__before_uatomic_or(); | |
256 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED); | |
257 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0) | |
258 | poll(NULL, 0, 1); | |
259 | rcu_register_thread(); | |
260 | } | |
261 | ||
b57aee66 PM |
262 | if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) { |
263 | while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL) | |
264 | poll(NULL, 0, 1); | |
265 | _CMM_STORE_SHARED(crdp->cbs.head, NULL); | |
266 | cbs_tail = (struct cds_wfq_node **) | |
267 | uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head); | |
268 | synchronize_rcu(); | |
269 | cbcount = 0; | |
270 | do { | |
271 | while (cbs->next == NULL && | |
272 | &cbs->next != cbs_tail) | |
273 | poll(NULL, 0, 1); | |
274 | if (cbs == &crdp->cbs.dummy) { | |
275 | cbs = cbs->next; | |
276 | continue; | |
277 | } | |
278 | rhp = (struct rcu_head *)cbs; | |
279 | cbs = cbs->next; | |
280 | rhp->func(rhp); | |
281 | cbcount++; | |
282 | } while (cbs != NULL); | |
283 | uatomic_sub(&crdp->qlen, cbcount); | |
284 | } | |
bc94ca9b MD |
285 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP) |
286 | break; | |
765f3ead | 287 | rcu_thread_offline(); |
bc94ca9b MD |
288 | if (!rt) { |
289 | if (&crdp->cbs.head | |
290 | == _CMM_LOAD_SHARED(crdp->cbs.tail)) { | |
291 | call_rcu_wait(crdp); | |
292 | poll(NULL, 0, 10); | |
293 | uatomic_dec(&crdp->futex); | |
c768e45e | 294 | /* |
bc94ca9b MD |
295 | * Decrement futex before reading |
296 | * call_rcu list. | |
c768e45e MD |
297 | */ |
298 | cmm_smp_mb(); | |
ccbac24d MD |
299 | } else { |
300 | poll(NULL, 0, 10); | |
c768e45e | 301 | } |
bc94ca9b MD |
302 | } else { |
303 | poll(NULL, 0, 10); | |
b57aee66 | 304 | } |
765f3ead | 305 | rcu_thread_online(); |
bc94ca9b MD |
306 | } |
307 | if (!rt) { | |
308 | /* | |
309 | * Read call_rcu list before write futex. | |
310 | */ | |
311 | cmm_smp_mb(); | |
312 | uatomic_set(&crdp->futex, 0); | |
b57aee66 | 313 | } |
2870aa1e | 314 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED); |
765f3ead | 315 | rcu_unregister_thread(); |
7106ddf8 | 316 | return NULL; |
b57aee66 PM |
317 | } |
318 | ||
319 | /* | |
320 | * Create both a call_rcu thread and the corresponding call_rcu_data | |
3c24913f PM |
321 | * structure, linking the structure in as specified. Caller must hold |
322 | * call_rcu_mutex. | |
b57aee66 PM |
323 | */ |
324 | ||
3c24913f | 325 | static void call_rcu_data_init(struct call_rcu_data **crdpp, |
c1d2c60b MD |
326 | unsigned long flags, |
327 | int cpu_affinity) | |
b57aee66 PM |
328 | { |
329 | struct call_rcu_data *crdp; | |
4a6d7378 | 330 | int ret; |
b57aee66 PM |
331 | |
332 | crdp = malloc(sizeof(*crdp)); | |
4a6d7378 MD |
333 | if (crdp == NULL) |
334 | urcu_die(errno); | |
b57aee66 PM |
335 | memset(crdp, '\0', sizeof(*crdp)); |
336 | cds_wfq_init(&crdp->cbs); | |
337 | crdp->qlen = 0; | |
263e3cf9 MD |
338 | crdp->futex = 0; |
339 | crdp->flags = flags; | |
3c24913f | 340 | cds_list_add(&crdp->list, &call_rcu_data_list); |
c1d2c60b | 341 | crdp->cpu_affinity = cpu_affinity; |
b57aee66 PM |
342 | cmm_smp_mb(); /* Structure initialized before pointer is planted. */ |
343 | *crdpp = crdp; | |
4a6d7378 MD |
344 | ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp); |
345 | if (ret) | |
346 | urcu_die(ret); | |
b57aee66 PM |
347 | } |
348 | ||
349 | /* | |
350 | * Return a pointer to the call_rcu_data structure for the specified | |
351 | * CPU, returning NULL if there is none. We cannot automatically | |
352 | * created it because the platform we are running on might not define | |
353 | * sched_getcpu(). | |
618b2595 MD |
354 | * |
355 | * The call to this function and use of the returned call_rcu_data | |
356 | * should be protected by RCU read-side lock. | |
b57aee66 PM |
357 | */ |
358 | ||
359 | struct call_rcu_data *get_cpu_call_rcu_data(int cpu) | |
360 | { | |
361 | static int warned = 0; | |
618b2595 | 362 | struct call_rcu_data **pcpu_crdp; |
b57aee66 | 363 | |
618b2595 MD |
364 | pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data); |
365 | if (pcpu_crdp == NULL) | |
b57aee66 PM |
366 | return NULL; |
367 | if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) { | |
368 | fprintf(stderr, "[error] liburcu: get CPU # out of range\n"); | |
369 | warned = 1; | |
370 | } | |
371 | if (cpu < 0 || maxcpus <= cpu) | |
372 | return NULL; | |
618b2595 | 373 | return rcu_dereference(pcpu_crdp[cpu]); |
b57aee66 PM |
374 | } |
375 | ||
376 | /* | |
377 | * Return the tid corresponding to the call_rcu thread whose | |
378 | * call_rcu_data structure is specified. | |
379 | */ | |
380 | ||
381 | pthread_t get_call_rcu_thread(struct call_rcu_data *crdp) | |
382 | { | |
383 | return crdp->tid; | |
384 | } | |
385 | ||
386 | /* | |
387 | * Create a call_rcu_data structure (with thread) and return a pointer. | |
388 | */ | |
389 | ||
c1d2c60b MD |
390 | static struct call_rcu_data *__create_call_rcu_data(unsigned long flags, |
391 | int cpu_affinity) | |
b57aee66 PM |
392 | { |
393 | struct call_rcu_data *crdp; | |
394 | ||
c1d2c60b | 395 | call_rcu_data_init(&crdp, flags, cpu_affinity); |
b57aee66 PM |
396 | return crdp; |
397 | } | |
398 | ||
c1d2c60b MD |
399 | struct call_rcu_data *create_call_rcu_data(unsigned long flags, |
400 | int cpu_affinity) | |
3c24913f PM |
401 | { |
402 | struct call_rcu_data *crdp; | |
403 | ||
404 | call_rcu_lock(&call_rcu_mutex); | |
c1d2c60b | 405 | crdp = __create_call_rcu_data(flags, cpu_affinity); |
3c24913f PM |
406 | call_rcu_unlock(&call_rcu_mutex); |
407 | return crdp; | |
408 | } | |
409 | ||
b57aee66 PM |
410 | /* |
411 | * Set the specified CPU to use the specified call_rcu_data structure. | |
7106ddf8 PM |
412 | * |
413 | * Use NULL to remove a CPU's call_rcu_data structure, but it is | |
414 | * the caller's responsibility to dispose of the removed structure. | |
415 | * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure | |
416 | * (prior to NULLing it out, of course). | |
f9da0936 MD |
417 | * |
418 | * The caller must wait for a grace-period to pass between return from | |
419 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
420 | * previous call rcu data as argument. | |
b57aee66 PM |
421 | */ |
422 | ||
423 | int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp) | |
424 | { | |
dcfc8165 | 425 | static int warned = 0; |
b57aee66 PM |
426 | |
427 | call_rcu_lock(&call_rcu_mutex); | |
f3776786 | 428 | alloc_cpu_call_rcu_data(); |
b57aee66 PM |
429 | if (cpu < 0 || maxcpus <= cpu) { |
430 | if (!warned) { | |
431 | fprintf(stderr, "[error] liburcu: set CPU # out of range\n"); | |
432 | warned = 1; | |
433 | } | |
434 | call_rcu_unlock(&call_rcu_mutex); | |
435 | errno = EINVAL; | |
436 | return -EINVAL; | |
437 | } | |
53a55535 | 438 | |
b57aee66 | 439 | if (per_cpu_call_rcu_data == NULL) { |
3670fef2 | 440 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
441 | errno = ENOMEM; |
442 | return -ENOMEM; | |
443 | } | |
53a55535 LJ |
444 | |
445 | if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) { | |
446 | call_rcu_unlock(&call_rcu_mutex); | |
447 | errno = EEXIST; | |
448 | return -EEXIST; | |
449 | } | |
450 | ||
618b2595 | 451 | rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp); |
3670fef2 | 452 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
453 | return 0; |
454 | } | |
455 | ||
456 | /* | |
457 | * Return a pointer to the default call_rcu_data structure, creating | |
458 | * one if need be. Because we never free call_rcu_data structures, | |
459 | * we don't need to be in an RCU read-side critical section. | |
460 | */ | |
461 | ||
462 | struct call_rcu_data *get_default_call_rcu_data(void) | |
463 | { | |
464 | if (default_call_rcu_data != NULL) | |
465 | return rcu_dereference(default_call_rcu_data); | |
466 | call_rcu_lock(&call_rcu_mutex); | |
467 | if (default_call_rcu_data != NULL) { | |
468 | call_rcu_unlock(&call_rcu_mutex); | |
469 | return default_call_rcu_data; | |
470 | } | |
c1d2c60b | 471 | call_rcu_data_init(&default_call_rcu_data, 0, -1); |
b57aee66 PM |
472 | call_rcu_unlock(&call_rcu_mutex); |
473 | return default_call_rcu_data; | |
474 | } | |
475 | ||
476 | /* | |
477 | * Return the call_rcu_data structure that applies to the currently | |
478 | * running thread. Any call_rcu_data structure assigned specifically | |
479 | * to this thread has first priority, followed by any call_rcu_data | |
480 | * structure assigned to the CPU on which the thread is running, | |
481 | * followed by the default call_rcu_data structure. If there is not | |
482 | * yet a default call_rcu_data structure, one will be created. | |
618b2595 MD |
483 | * |
484 | * Calls to this function and use of the returned call_rcu_data should | |
485 | * be protected by RCU read-side lock. | |
b57aee66 PM |
486 | */ |
487 | struct call_rcu_data *get_call_rcu_data(void) | |
488 | { | |
9744f3bb | 489 | struct call_rcu_data *crd; |
b57aee66 | 490 | |
bd252a04 MD |
491 | if (URCU_TLS(thread_call_rcu_data) != NULL) |
492 | return URCU_TLS(thread_call_rcu_data); | |
9744f3bb LJ |
493 | |
494 | if (maxcpus > 0) { | |
495 | crd = get_cpu_call_rcu_data(sched_getcpu()); | |
496 | if (crd) | |
497 | return crd; | |
b57aee66 | 498 | } |
9744f3bb | 499 | |
b57aee66 PM |
500 | return get_default_call_rcu_data(); |
501 | } | |
502 | ||
503 | /* | |
504 | * Return a pointer to this task's call_rcu_data if there is one. | |
505 | */ | |
506 | ||
507 | struct call_rcu_data *get_thread_call_rcu_data(void) | |
508 | { | |
bd252a04 | 509 | return URCU_TLS(thread_call_rcu_data); |
b57aee66 PM |
510 | } |
511 | ||
512 | /* | |
513 | * Set this task's call_rcu_data structure as specified, regardless | |
514 | * of whether or not this task already had one. (This allows switching | |
515 | * to and from real-time call_rcu threads, for example.) | |
7106ddf8 PM |
516 | * |
517 | * Use NULL to remove a thread's call_rcu_data structure, but it is | |
518 | * the caller's responsibility to dispose of the removed structure. | |
519 | * Use get_thread_call_rcu_data() to obtain a pointer to the old structure | |
520 | * (prior to NULLing it out, of course). | |
b57aee66 PM |
521 | */ |
522 | ||
523 | void set_thread_call_rcu_data(struct call_rcu_data *crdp) | |
524 | { | |
bd252a04 | 525 | URCU_TLS(thread_call_rcu_data) = crdp; |
b57aee66 PM |
526 | } |
527 | ||
528 | /* | |
529 | * Create a separate call_rcu thread for each CPU. This does not | |
530 | * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data() | |
0938c541 MD |
531 | * function if you want that behavior. Should be paired with |
532 | * free_all_cpu_call_rcu_data() to teardown these call_rcu worker | |
533 | * threads. | |
b57aee66 PM |
534 | */ |
535 | ||
536 | int create_all_cpu_call_rcu_data(unsigned long flags) | |
537 | { | |
538 | int i; | |
539 | struct call_rcu_data *crdp; | |
540 | int ret; | |
541 | ||
542 | call_rcu_lock(&call_rcu_mutex); | |
543 | alloc_cpu_call_rcu_data(); | |
544 | call_rcu_unlock(&call_rcu_mutex); | |
545 | if (maxcpus <= 0) { | |
546 | errno = EINVAL; | |
547 | return -EINVAL; | |
548 | } | |
549 | if (per_cpu_call_rcu_data == NULL) { | |
550 | errno = ENOMEM; | |
551 | return -ENOMEM; | |
552 | } | |
553 | for (i = 0; i < maxcpus; i++) { | |
554 | call_rcu_lock(&call_rcu_mutex); | |
555 | if (get_cpu_call_rcu_data(i)) { | |
556 | call_rcu_unlock(&call_rcu_mutex); | |
557 | continue; | |
558 | } | |
c1d2c60b | 559 | crdp = __create_call_rcu_data(flags, i); |
b57aee66 PM |
560 | if (crdp == NULL) { |
561 | call_rcu_unlock(&call_rcu_mutex); | |
562 | errno = ENOMEM; | |
563 | return -ENOMEM; | |
564 | } | |
565 | call_rcu_unlock(&call_rcu_mutex); | |
566 | if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) { | |
356c8794 LJ |
567 | call_rcu_data_free(crdp); |
568 | ||
569 | /* it has been created by other thread */ | |
570 | if (ret == -EEXIST) | |
571 | continue; | |
572 | ||
573 | return ret; | |
b57aee66 PM |
574 | } |
575 | } | |
576 | return 0; | |
577 | } | |
578 | ||
7106ddf8 PM |
579 | /* |
580 | * Wake up the call_rcu thread corresponding to the specified | |
581 | * call_rcu_data structure. | |
582 | */ | |
583 | static void wake_call_rcu_thread(struct call_rcu_data *crdp) | |
584 | { | |
263e3cf9 MD |
585 | if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT)) |
586 | call_rcu_wake_up(crdp); | |
7106ddf8 PM |
587 | } |
588 | ||
b57aee66 PM |
589 | /* |
590 | * Schedule a function to be invoked after a following grace period. | |
591 | * This is the only function that must be called -- the others are | |
592 | * only present to allow applications to tune their use of RCU for | |
593 | * maximum performance. | |
594 | * | |
595 | * Note that unless a call_rcu thread has not already been created, | |
596 | * the first invocation of call_rcu() will create one. So, if you | |
597 | * need the first invocation of call_rcu() to be fast, make sure | |
598 | * to create a call_rcu thread first. One way to accomplish this is | |
599 | * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data(). | |
618b2595 MD |
600 | * |
601 | * call_rcu must be called by registered RCU read-side threads. | |
b57aee66 PM |
602 | */ |
603 | ||
604 | void call_rcu(struct rcu_head *head, | |
605 | void (*func)(struct rcu_head *head)) | |
606 | { | |
607 | struct call_rcu_data *crdp; | |
608 | ||
609 | cds_wfq_node_init(&head->next); | |
610 | head->func = func; | |
618b2595 MD |
611 | /* Holding rcu read-side lock across use of per-cpu crdp */ |
612 | rcu_read_lock(); | |
b57aee66 PM |
613 | crdp = get_call_rcu_data(); |
614 | cds_wfq_enqueue(&crdp->cbs, &head->next); | |
615 | uatomic_inc(&crdp->qlen); | |
7106ddf8 | 616 | wake_call_rcu_thread(crdp); |
618b2595 | 617 | rcu_read_unlock(); |
7106ddf8 PM |
618 | } |
619 | ||
620 | /* | |
621 | * Free up the specified call_rcu_data structure, terminating the | |
622 | * associated call_rcu thread. The caller must have previously | |
623 | * removed the call_rcu_data structure from per-thread or per-CPU | |
624 | * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU | |
625 | * call_rcu_data structures or set_thread_call_rcu_data(NULL) for | |
626 | * per-thread call_rcu_data structures. | |
627 | * | |
628 | * We silently refuse to free up the default call_rcu_data structure | |
629 | * because that is where we put any leftover callbacks. Note that | |
630 | * the possibility of self-spawning callbacks makes it impossible | |
631 | * to execute all the callbacks in finite time without putting any | |
632 | * newly spawned callbacks somewhere else. The "somewhere else" of | |
633 | * last resort is the default call_rcu_data structure. | |
634 | * | |
635 | * We also silently refuse to free NULL pointers. This simplifies | |
636 | * the calling code. | |
f9da0936 MD |
637 | * |
638 | * The caller must wait for a grace-period to pass between return from | |
639 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
640 | * previous call rcu data as argument. | |
7106ddf8 PM |
641 | */ |
642 | void call_rcu_data_free(struct call_rcu_data *crdp) | |
643 | { | |
644 | struct cds_wfq_node *cbs; | |
645 | struct cds_wfq_node **cbs_tail; | |
646 | struct cds_wfq_node **cbs_endprev; | |
647 | ||
648 | if (crdp == NULL || crdp == default_call_rcu_data) { | |
649 | return; | |
650 | } | |
2870aa1e PB |
651 | if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) { |
652 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP); | |
7106ddf8 | 653 | wake_call_rcu_thread(crdp); |
2870aa1e | 654 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) |
7106ddf8 PM |
655 | poll(NULL, 0, 1); |
656 | } | |
657 | if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) { | |
658 | while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL) | |
659 | poll(NULL, 0, 1); | |
660 | _CMM_STORE_SHARED(crdp->cbs.head, NULL); | |
661 | cbs_tail = (struct cds_wfq_node **) | |
662 | uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head); | |
698d0778 MD |
663 | /* Create default call rcu data if need be */ |
664 | (void) get_default_call_rcu_data(); | |
7106ddf8 | 665 | cbs_endprev = (struct cds_wfq_node **) |
19cf3ae1 MD |
666 | uatomic_xchg(&default_call_rcu_data->cbs.tail, |
667 | cbs_tail); | |
668 | _CMM_STORE_SHARED(*cbs_endprev, cbs); | |
7106ddf8 PM |
669 | uatomic_add(&default_call_rcu_data->qlen, |
670 | uatomic_read(&crdp->qlen)); | |
1e92aa15 | 671 | wake_call_rcu_thread(default_call_rcu_data); |
7106ddf8 | 672 | } |
59dc9e9d | 673 | |
b75dffe6 | 674 | call_rcu_lock(&call_rcu_mutex); |
59dc9e9d | 675 | cds_list_del(&crdp->list); |
b75dffe6 LJ |
676 | call_rcu_unlock(&call_rcu_mutex); |
677 | ||
59dc9e9d | 678 | free(crdp); |
7106ddf8 PM |
679 | } |
680 | ||
681 | /* | |
682 | * Clean up all the per-CPU call_rcu threads. | |
683 | */ | |
684 | void free_all_cpu_call_rcu_data(void) | |
685 | { | |
686 | int cpu; | |
618b2595 MD |
687 | struct call_rcu_data **crdp; |
688 | static int warned = 0; | |
7106ddf8 PM |
689 | |
690 | if (maxcpus <= 0) | |
691 | return; | |
618b2595 MD |
692 | |
693 | crdp = malloc(sizeof(*crdp) * maxcpus); | |
694 | if (!crdp) { | |
695 | if (!warned) { | |
696 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
697 | } | |
698 | warned = 1; | |
d31150b4 | 699 | return; |
618b2595 MD |
700 | } |
701 | ||
7106ddf8 | 702 | for (cpu = 0; cpu < maxcpus; cpu++) { |
618b2595 MD |
703 | crdp[cpu] = get_cpu_call_rcu_data(cpu); |
704 | if (crdp[cpu] == NULL) | |
7106ddf8 PM |
705 | continue; |
706 | set_cpu_call_rcu_data(cpu, NULL); | |
7106ddf8 | 707 | } |
618b2595 MD |
708 | /* |
709 | * Wait for call_rcu sites acting as RCU readers of the | |
710 | * call_rcu_data to become quiescent. | |
711 | */ | |
712 | synchronize_rcu(); | |
713 | for (cpu = 0; cpu < maxcpus; cpu++) { | |
714 | if (crdp[cpu] == NULL) | |
715 | continue; | |
716 | call_rcu_data_free(crdp[cpu]); | |
717 | } | |
718 | free(crdp); | |
7106ddf8 PM |
719 | } |
720 | ||
81ad2e19 PM |
721 | /* |
722 | * Acquire the call_rcu_mutex in order to ensure that the child sees | |
750cfec5 MD |
723 | * all of the call_rcu() data structures in a consistent state. Ensure |
724 | * that all call_rcu threads are in a quiescent state across fork. | |
81ad2e19 PM |
725 | * Suitable for pthread_atfork() and friends. |
726 | */ | |
727 | void call_rcu_before_fork(void) | |
728 | { | |
750cfec5 MD |
729 | struct call_rcu_data *crdp; |
730 | ||
81ad2e19 | 731 | call_rcu_lock(&call_rcu_mutex); |
750cfec5 MD |
732 | |
733 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
734 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE); | |
735 | cmm_smp_mb__after_uatomic_or(); | |
736 | wake_call_rcu_thread(crdp); | |
737 | } | |
738 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
739 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0) | |
740 | poll(NULL, 0, 1); | |
741 | } | |
81ad2e19 PM |
742 | } |
743 | ||
744 | /* | |
745 | * Clean up call_rcu data structures in the parent of a successful fork() | |
746 | * that is not followed by exec() in the child. Suitable for | |
747 | * pthread_atfork() and friends. | |
748 | */ | |
749 | void call_rcu_after_fork_parent(void) | |
750 | { | |
750cfec5 MD |
751 | struct call_rcu_data *crdp; |
752 | ||
753 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) | |
754 | uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE); | |
81ad2e19 PM |
755 | call_rcu_unlock(&call_rcu_mutex); |
756 | } | |
757 | ||
7106ddf8 PM |
758 | /* |
759 | * Clean up call_rcu data structures in the child of a successful fork() | |
81ad2e19 PM |
760 | * that is not followed by exec(). Suitable for pthread_atfork() and |
761 | * friends. | |
7106ddf8 PM |
762 | */ |
763 | void call_rcu_after_fork_child(void) | |
764 | { | |
077ff173 | 765 | struct call_rcu_data *crdp, *next; |
7106ddf8 | 766 | |
81ad2e19 PM |
767 | /* Release the mutex. */ |
768 | call_rcu_unlock(&call_rcu_mutex); | |
769 | ||
ad1b9909 LJ |
770 | /* Do nothing when call_rcu() has not been used */ |
771 | if (cds_list_empty(&call_rcu_data_list)) | |
772 | return; | |
773 | ||
7106ddf8 PM |
774 | /* |
775 | * Allocate a new default call_rcu_data structure in order | |
776 | * to get a working call_rcu thread to go with it. | |
777 | */ | |
778 | default_call_rcu_data = NULL; | |
779 | (void)get_default_call_rcu_data(); | |
780 | ||
60af049d LJ |
781 | /* Cleanup call_rcu_data pointers before use */ |
782 | maxcpus_reset(); | |
783 | free(per_cpu_call_rcu_data); | |
618b2595 | 784 | rcu_set_pointer(&per_cpu_call_rcu_data, NULL); |
bd252a04 | 785 | URCU_TLS(thread_call_rcu_data) = NULL; |
60af049d | 786 | |
750cfec5 MD |
787 | /* |
788 | * Dispose of all of the rest of the call_rcu_data structures. | |
789 | * Leftover call_rcu callbacks will be merged into the new | |
790 | * default call_rcu thread queue. | |
791 | */ | |
077ff173 | 792 | cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) { |
7106ddf8 | 793 | if (crdp == default_call_rcu_data) |
077ff173 | 794 | continue; |
2870aa1e | 795 | uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED); |
7106ddf8 | 796 | call_rcu_data_free(crdp); |
b57aee66 PM |
797 | } |
798 | } |