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 | ||
254ebaf3 MD |
84 | #ifdef HAVE_SCHED_GETCPU |
85 | ||
86 | static int urcu_sched_getcpu(void) | |
87 | { | |
88 | return sched_getcpu(); | |
89 | } | |
90 | ||
91 | #else /* #ifdef HAVE_SCHED_GETCPU */ | |
92 | ||
93 | static int urcu_sched_getcpu(void) | |
94 | { | |
95 | return -1; | |
96 | } | |
97 | ||
98 | #endif /* #else #ifdef HAVE_SCHED_GETCPU */ | |
99 | ||
100 | #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) | |
b57aee66 PM |
101 | |
102 | /* | |
103 | * Pointer to array of pointers to per-CPU call_rcu_data structures | |
618b2595 MD |
104 | * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an |
105 | * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a | |
106 | * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer | |
107 | * without mutex. The call_rcu_mutex protects updates. | |
b57aee66 PM |
108 | */ |
109 | ||
110 | static struct call_rcu_data **per_cpu_call_rcu_data; | |
111 | static long maxcpus; | |
112 | ||
60af049d LJ |
113 | static void maxcpus_reset(void) |
114 | { | |
115 | maxcpus = 0; | |
116 | } | |
117 | ||
b57aee66 PM |
118 | /* Allocate the array if it has not already been allocated. */ |
119 | ||
120 | static void alloc_cpu_call_rcu_data(void) | |
121 | { | |
122 | struct call_rcu_data **p; | |
123 | static int warned = 0; | |
124 | ||
125 | if (maxcpus != 0) | |
126 | return; | |
127 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
128 | if (maxcpus <= 0) { | |
129 | return; | |
130 | } | |
131 | p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
132 | if (p != NULL) { | |
133 | memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
618b2595 | 134 | rcu_set_pointer(&per_cpu_call_rcu_data, p); |
b57aee66 PM |
135 | } else { |
136 | if (!warned) { | |
137 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
138 | } | |
139 | warned = 1; | |
140 | } | |
141 | } | |
142 | ||
254ebaf3 | 143 | #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
b57aee66 | 144 | |
f9437098 MD |
145 | /* |
146 | * per_cpu_call_rcu_data should be constant, but some functions below, used both | |
147 | * for cases where cpu number is available and not available, assume it it not | |
148 | * constant. | |
149 | */ | |
150 | static struct call_rcu_data **per_cpu_call_rcu_data = NULL; | |
b57aee66 PM |
151 | static const long maxcpus = -1; |
152 | ||
60af049d LJ |
153 | static void maxcpus_reset(void) |
154 | { | |
155 | } | |
156 | ||
b57aee66 PM |
157 | static void alloc_cpu_call_rcu_data(void) |
158 | { | |
159 | } | |
160 | ||
254ebaf3 | 161 | #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
b57aee66 PM |
162 | |
163 | /* Acquire the specified pthread mutex. */ | |
164 | ||
165 | static void call_rcu_lock(pthread_mutex_t *pmp) | |
166 | { | |
4a6d7378 MD |
167 | int ret; |
168 | ||
169 | ret = pthread_mutex_lock(pmp); | |
170 | if (ret) | |
171 | urcu_die(ret); | |
b57aee66 PM |
172 | } |
173 | ||
174 | /* Release the specified pthread mutex. */ | |
175 | ||
176 | static void call_rcu_unlock(pthread_mutex_t *pmp) | |
177 | { | |
4a6d7378 MD |
178 | int ret; |
179 | ||
180 | ret = pthread_mutex_unlock(pmp); | |
181 | if (ret) | |
182 | urcu_die(ret); | |
b57aee66 PM |
183 | } |
184 | ||
c1d2c60b MD |
185 | #if HAVE_SCHED_SETAFFINITY |
186 | static | |
187 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
188 | { | |
189 | cpu_set_t mask; | |
190 | ||
191 | if (crdp->cpu_affinity < 0) | |
192 | return 0; | |
193 | ||
194 | CPU_ZERO(&mask); | |
195 | CPU_SET(crdp->cpu_affinity, &mask); | |
196 | #if SCHED_SETAFFINITY_ARGS == 2 | |
197 | return sched_setaffinity(0, &mask); | |
198 | #else | |
199 | return sched_setaffinity(0, sizeof(mask), &mask); | |
200 | #endif | |
201 | } | |
202 | #else | |
203 | static | |
204 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
205 | { | |
206 | return 0; | |
207 | } | |
208 | #endif | |
209 | ||
03fe58b3 MD |
210 | static void call_rcu_wait(struct call_rcu_data *crdp) |
211 | { | |
212 | /* Read call_rcu list before read futex */ | |
213 | cmm_smp_mb(); | |
214 | if (uatomic_read(&crdp->futex) == -1) | |
215 | futex_async(&crdp->futex, FUTEX_WAIT, -1, | |
216 | NULL, NULL, 0); | |
217 | } | |
218 | ||
219 | static void call_rcu_wake_up(struct call_rcu_data *crdp) | |
220 | { | |
221 | /* Write to call_rcu list before reading/writing futex */ | |
222 | cmm_smp_mb(); | |
a0b7f7ea | 223 | if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) { |
03fe58b3 MD |
224 | uatomic_set(&crdp->futex, 0); |
225 | futex_async(&crdp->futex, FUTEX_WAKE, 1, | |
226 | NULL, NULL, 0); | |
227 | } | |
228 | } | |
229 | ||
b57aee66 PM |
230 | /* This is the code run by each call_rcu thread. */ |
231 | ||
232 | static void *call_rcu_thread(void *arg) | |
233 | { | |
234 | unsigned long cbcount; | |
235 | struct cds_wfq_node *cbs; | |
236 | struct cds_wfq_node **cbs_tail; | |
237 | struct call_rcu_data *crdp = (struct call_rcu_data *)arg; | |
238 | struct rcu_head *rhp; | |
2870aa1e | 239 | int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT); |
4a6d7378 | 240 | int ret; |
b57aee66 | 241 | |
4a6d7378 MD |
242 | ret = set_thread_cpu_affinity(crdp); |
243 | if (ret) | |
244 | urcu_die(errno); | |
c1d2c60b | 245 | |
765f3ead MD |
246 | /* |
247 | * If callbacks take a read-side lock, we need to be registered. | |
248 | */ | |
249 | rcu_register_thread(); | |
250 | ||
bd252a04 | 251 | URCU_TLS(thread_call_rcu_data) = crdp; |
bc94ca9b MD |
252 | if (!rt) { |
253 | uatomic_dec(&crdp->futex); | |
254 | /* Decrement futex before reading call_rcu list */ | |
255 | cmm_smp_mb(); | |
256 | } | |
b57aee66 | 257 | for (;;) { |
750cfec5 MD |
258 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) { |
259 | /* | |
260 | * Pause requested. Become quiescent: remove | |
261 | * ourself from all global lists, and don't | |
262 | * process any callback. The callback lists may | |
263 | * still be non-empty though. | |
264 | */ | |
265 | rcu_unregister_thread(); | |
266 | cmm_smp_mb__before_uatomic_or(); | |
267 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED); | |
268 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0) | |
269 | poll(NULL, 0, 1); | |
e6564e91 KF |
270 | uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED); |
271 | cmm_smp_mb__after_uatomic_and(); | |
750cfec5 MD |
272 | rcu_register_thread(); |
273 | } | |
274 | ||
b57aee66 PM |
275 | if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) { |
276 | while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL) | |
277 | poll(NULL, 0, 1); | |
278 | _CMM_STORE_SHARED(crdp->cbs.head, NULL); | |
279 | cbs_tail = (struct cds_wfq_node **) | |
280 | uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head); | |
281 | synchronize_rcu(); | |
282 | cbcount = 0; | |
283 | do { | |
284 | while (cbs->next == NULL && | |
285 | &cbs->next != cbs_tail) | |
286 | poll(NULL, 0, 1); | |
287 | if (cbs == &crdp->cbs.dummy) { | |
288 | cbs = cbs->next; | |
289 | continue; | |
290 | } | |
291 | rhp = (struct rcu_head *)cbs; | |
292 | cbs = cbs->next; | |
293 | rhp->func(rhp); | |
294 | cbcount++; | |
295 | } while (cbs != NULL); | |
296 | uatomic_sub(&crdp->qlen, cbcount); | |
297 | } | |
bc94ca9b MD |
298 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP) |
299 | break; | |
765f3ead | 300 | rcu_thread_offline(); |
bc94ca9b MD |
301 | if (!rt) { |
302 | if (&crdp->cbs.head | |
303 | == _CMM_LOAD_SHARED(crdp->cbs.tail)) { | |
304 | call_rcu_wait(crdp); | |
305 | poll(NULL, 0, 10); | |
306 | uatomic_dec(&crdp->futex); | |
c768e45e | 307 | /* |
bc94ca9b MD |
308 | * Decrement futex before reading |
309 | * call_rcu list. | |
c768e45e MD |
310 | */ |
311 | cmm_smp_mb(); | |
ccbac24d MD |
312 | } else { |
313 | poll(NULL, 0, 10); | |
c768e45e | 314 | } |
bc94ca9b MD |
315 | } else { |
316 | poll(NULL, 0, 10); | |
b57aee66 | 317 | } |
765f3ead | 318 | rcu_thread_online(); |
bc94ca9b MD |
319 | } |
320 | if (!rt) { | |
321 | /* | |
322 | * Read call_rcu list before write futex. | |
323 | */ | |
324 | cmm_smp_mb(); | |
325 | uatomic_set(&crdp->futex, 0); | |
b57aee66 | 326 | } |
2870aa1e | 327 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED); |
765f3ead | 328 | rcu_unregister_thread(); |
7106ddf8 | 329 | return NULL; |
b57aee66 PM |
330 | } |
331 | ||
332 | /* | |
333 | * Create both a call_rcu thread and the corresponding call_rcu_data | |
3c24913f PM |
334 | * structure, linking the structure in as specified. Caller must hold |
335 | * call_rcu_mutex. | |
b57aee66 PM |
336 | */ |
337 | ||
3c24913f | 338 | static void call_rcu_data_init(struct call_rcu_data **crdpp, |
c1d2c60b MD |
339 | unsigned long flags, |
340 | int cpu_affinity) | |
b57aee66 PM |
341 | { |
342 | struct call_rcu_data *crdp; | |
4a6d7378 | 343 | int ret; |
b57aee66 PM |
344 | |
345 | crdp = malloc(sizeof(*crdp)); | |
4a6d7378 MD |
346 | if (crdp == NULL) |
347 | urcu_die(errno); | |
b57aee66 PM |
348 | memset(crdp, '\0', sizeof(*crdp)); |
349 | cds_wfq_init(&crdp->cbs); | |
350 | crdp->qlen = 0; | |
263e3cf9 MD |
351 | crdp->futex = 0; |
352 | crdp->flags = flags; | |
3c24913f | 353 | cds_list_add(&crdp->list, &call_rcu_data_list); |
c1d2c60b | 354 | crdp->cpu_affinity = cpu_affinity; |
b57aee66 PM |
355 | cmm_smp_mb(); /* Structure initialized before pointer is planted. */ |
356 | *crdpp = crdp; | |
4a6d7378 MD |
357 | ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp); |
358 | if (ret) | |
359 | urcu_die(ret); | |
b57aee66 PM |
360 | } |
361 | ||
362 | /* | |
363 | * Return a pointer to the call_rcu_data structure for the specified | |
364 | * CPU, returning NULL if there is none. We cannot automatically | |
365 | * created it because the platform we are running on might not define | |
254ebaf3 | 366 | * urcu_sched_getcpu(). |
618b2595 MD |
367 | * |
368 | * The call to this function and use of the returned call_rcu_data | |
369 | * should be protected by RCU read-side lock. | |
b57aee66 PM |
370 | */ |
371 | ||
372 | struct call_rcu_data *get_cpu_call_rcu_data(int cpu) | |
373 | { | |
374 | static int warned = 0; | |
618b2595 | 375 | struct call_rcu_data **pcpu_crdp; |
b57aee66 | 376 | |
618b2595 MD |
377 | pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data); |
378 | if (pcpu_crdp == NULL) | |
b57aee66 PM |
379 | return NULL; |
380 | if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) { | |
381 | fprintf(stderr, "[error] liburcu: get CPU # out of range\n"); | |
382 | warned = 1; | |
383 | } | |
384 | if (cpu < 0 || maxcpus <= cpu) | |
385 | return NULL; | |
618b2595 | 386 | return rcu_dereference(pcpu_crdp[cpu]); |
b57aee66 PM |
387 | } |
388 | ||
389 | /* | |
390 | * Return the tid corresponding to the call_rcu thread whose | |
391 | * call_rcu_data structure is specified. | |
392 | */ | |
393 | ||
394 | pthread_t get_call_rcu_thread(struct call_rcu_data *crdp) | |
395 | { | |
396 | return crdp->tid; | |
397 | } | |
398 | ||
399 | /* | |
400 | * Create a call_rcu_data structure (with thread) and return a pointer. | |
401 | */ | |
402 | ||
c1d2c60b MD |
403 | static struct call_rcu_data *__create_call_rcu_data(unsigned long flags, |
404 | int cpu_affinity) | |
b57aee66 PM |
405 | { |
406 | struct call_rcu_data *crdp; | |
407 | ||
c1d2c60b | 408 | call_rcu_data_init(&crdp, flags, cpu_affinity); |
b57aee66 PM |
409 | return crdp; |
410 | } | |
411 | ||
c1d2c60b MD |
412 | struct call_rcu_data *create_call_rcu_data(unsigned long flags, |
413 | int cpu_affinity) | |
3c24913f PM |
414 | { |
415 | struct call_rcu_data *crdp; | |
416 | ||
417 | call_rcu_lock(&call_rcu_mutex); | |
c1d2c60b | 418 | crdp = __create_call_rcu_data(flags, cpu_affinity); |
3c24913f PM |
419 | call_rcu_unlock(&call_rcu_mutex); |
420 | return crdp; | |
421 | } | |
422 | ||
b57aee66 PM |
423 | /* |
424 | * Set the specified CPU to use the specified call_rcu_data structure. | |
7106ddf8 PM |
425 | * |
426 | * Use NULL to remove a CPU's call_rcu_data structure, but it is | |
427 | * the caller's responsibility to dispose of the removed structure. | |
428 | * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure | |
429 | * (prior to NULLing it out, of course). | |
f9da0936 MD |
430 | * |
431 | * The caller must wait for a grace-period to pass between return from | |
432 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
433 | * previous call rcu data as argument. | |
b57aee66 PM |
434 | */ |
435 | ||
436 | int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp) | |
437 | { | |
dcfc8165 | 438 | static int warned = 0; |
b57aee66 PM |
439 | |
440 | call_rcu_lock(&call_rcu_mutex); | |
f3776786 | 441 | alloc_cpu_call_rcu_data(); |
b57aee66 PM |
442 | if (cpu < 0 || maxcpus <= cpu) { |
443 | if (!warned) { | |
444 | fprintf(stderr, "[error] liburcu: set CPU # out of range\n"); | |
445 | warned = 1; | |
446 | } | |
447 | call_rcu_unlock(&call_rcu_mutex); | |
448 | errno = EINVAL; | |
449 | return -EINVAL; | |
450 | } | |
53a55535 | 451 | |
b57aee66 | 452 | if (per_cpu_call_rcu_data == NULL) { |
3670fef2 | 453 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
454 | errno = ENOMEM; |
455 | return -ENOMEM; | |
456 | } | |
53a55535 LJ |
457 | |
458 | if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) { | |
459 | call_rcu_unlock(&call_rcu_mutex); | |
460 | errno = EEXIST; | |
461 | return -EEXIST; | |
462 | } | |
463 | ||
618b2595 | 464 | rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp); |
3670fef2 | 465 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
466 | return 0; |
467 | } | |
468 | ||
469 | /* | |
470 | * Return a pointer to the default call_rcu_data structure, creating | |
471 | * one if need be. Because we never free call_rcu_data structures, | |
472 | * we don't need to be in an RCU read-side critical section. | |
473 | */ | |
474 | ||
475 | struct call_rcu_data *get_default_call_rcu_data(void) | |
476 | { | |
477 | if (default_call_rcu_data != NULL) | |
478 | return rcu_dereference(default_call_rcu_data); | |
479 | call_rcu_lock(&call_rcu_mutex); | |
480 | if (default_call_rcu_data != NULL) { | |
481 | call_rcu_unlock(&call_rcu_mutex); | |
482 | return default_call_rcu_data; | |
483 | } | |
c1d2c60b | 484 | call_rcu_data_init(&default_call_rcu_data, 0, -1); |
b57aee66 PM |
485 | call_rcu_unlock(&call_rcu_mutex); |
486 | return default_call_rcu_data; | |
487 | } | |
488 | ||
489 | /* | |
490 | * Return the call_rcu_data structure that applies to the currently | |
491 | * running thread. Any call_rcu_data structure assigned specifically | |
492 | * to this thread has first priority, followed by any call_rcu_data | |
493 | * structure assigned to the CPU on which the thread is running, | |
494 | * followed by the default call_rcu_data structure. If there is not | |
495 | * yet a default call_rcu_data structure, one will be created. | |
618b2595 MD |
496 | * |
497 | * Calls to this function and use of the returned call_rcu_data should | |
498 | * be protected by RCU read-side lock. | |
b57aee66 PM |
499 | */ |
500 | struct call_rcu_data *get_call_rcu_data(void) | |
501 | { | |
9744f3bb | 502 | struct call_rcu_data *crd; |
b57aee66 | 503 | |
bd252a04 MD |
504 | if (URCU_TLS(thread_call_rcu_data) != NULL) |
505 | return URCU_TLS(thread_call_rcu_data); | |
9744f3bb LJ |
506 | |
507 | if (maxcpus > 0) { | |
254ebaf3 | 508 | crd = get_cpu_call_rcu_data(urcu_sched_getcpu()); |
9744f3bb LJ |
509 | if (crd) |
510 | return crd; | |
b57aee66 | 511 | } |
9744f3bb | 512 | |
b57aee66 PM |
513 | return get_default_call_rcu_data(); |
514 | } | |
515 | ||
516 | /* | |
517 | * Return a pointer to this task's call_rcu_data if there is one. | |
518 | */ | |
519 | ||
520 | struct call_rcu_data *get_thread_call_rcu_data(void) | |
521 | { | |
bd252a04 | 522 | return URCU_TLS(thread_call_rcu_data); |
b57aee66 PM |
523 | } |
524 | ||
525 | /* | |
526 | * Set this task's call_rcu_data structure as specified, regardless | |
527 | * of whether or not this task already had one. (This allows switching | |
528 | * to and from real-time call_rcu threads, for example.) | |
7106ddf8 PM |
529 | * |
530 | * Use NULL to remove a thread's call_rcu_data structure, but it is | |
531 | * the caller's responsibility to dispose of the removed structure. | |
532 | * Use get_thread_call_rcu_data() to obtain a pointer to the old structure | |
533 | * (prior to NULLing it out, of course). | |
b57aee66 PM |
534 | */ |
535 | ||
536 | void set_thread_call_rcu_data(struct call_rcu_data *crdp) | |
537 | { | |
bd252a04 | 538 | URCU_TLS(thread_call_rcu_data) = crdp; |
b57aee66 PM |
539 | } |
540 | ||
541 | /* | |
542 | * Create a separate call_rcu thread for each CPU. This does not | |
543 | * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data() | |
0938c541 MD |
544 | * function if you want that behavior. Should be paired with |
545 | * free_all_cpu_call_rcu_data() to teardown these call_rcu worker | |
546 | * threads. | |
b57aee66 PM |
547 | */ |
548 | ||
549 | int create_all_cpu_call_rcu_data(unsigned long flags) | |
550 | { | |
551 | int i; | |
552 | struct call_rcu_data *crdp; | |
553 | int ret; | |
554 | ||
555 | call_rcu_lock(&call_rcu_mutex); | |
556 | alloc_cpu_call_rcu_data(); | |
557 | call_rcu_unlock(&call_rcu_mutex); | |
558 | if (maxcpus <= 0) { | |
559 | errno = EINVAL; | |
560 | return -EINVAL; | |
561 | } | |
562 | if (per_cpu_call_rcu_data == NULL) { | |
563 | errno = ENOMEM; | |
564 | return -ENOMEM; | |
565 | } | |
566 | for (i = 0; i < maxcpus; i++) { | |
567 | call_rcu_lock(&call_rcu_mutex); | |
568 | if (get_cpu_call_rcu_data(i)) { | |
569 | call_rcu_unlock(&call_rcu_mutex); | |
570 | continue; | |
571 | } | |
c1d2c60b | 572 | crdp = __create_call_rcu_data(flags, i); |
b57aee66 PM |
573 | if (crdp == NULL) { |
574 | call_rcu_unlock(&call_rcu_mutex); | |
575 | errno = ENOMEM; | |
576 | return -ENOMEM; | |
577 | } | |
578 | call_rcu_unlock(&call_rcu_mutex); | |
579 | if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) { | |
356c8794 LJ |
580 | call_rcu_data_free(crdp); |
581 | ||
582 | /* it has been created by other thread */ | |
583 | if (ret == -EEXIST) | |
584 | continue; | |
585 | ||
586 | return ret; | |
b57aee66 PM |
587 | } |
588 | } | |
589 | return 0; | |
590 | } | |
591 | ||
7106ddf8 PM |
592 | /* |
593 | * Wake up the call_rcu thread corresponding to the specified | |
594 | * call_rcu_data structure. | |
595 | */ | |
596 | static void wake_call_rcu_thread(struct call_rcu_data *crdp) | |
597 | { | |
263e3cf9 MD |
598 | if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT)) |
599 | call_rcu_wake_up(crdp); | |
7106ddf8 PM |
600 | } |
601 | ||
b57aee66 PM |
602 | /* |
603 | * Schedule a function to be invoked after a following grace period. | |
604 | * This is the only function that must be called -- the others are | |
605 | * only present to allow applications to tune their use of RCU for | |
606 | * maximum performance. | |
607 | * | |
608 | * Note that unless a call_rcu thread has not already been created, | |
609 | * the first invocation of call_rcu() will create one. So, if you | |
610 | * need the first invocation of call_rcu() to be fast, make sure | |
611 | * to create a call_rcu thread first. One way to accomplish this is | |
612 | * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data(). | |
618b2595 MD |
613 | * |
614 | * call_rcu must be called by registered RCU read-side threads. | |
b57aee66 PM |
615 | */ |
616 | ||
617 | void call_rcu(struct rcu_head *head, | |
618 | void (*func)(struct rcu_head *head)) | |
619 | { | |
620 | struct call_rcu_data *crdp; | |
621 | ||
622 | cds_wfq_node_init(&head->next); | |
623 | head->func = func; | |
618b2595 | 624 | /* Holding rcu read-side lock across use of per-cpu crdp */ |
7d214e56 | 625 | _rcu_read_lock(); |
b57aee66 PM |
626 | crdp = get_call_rcu_data(); |
627 | cds_wfq_enqueue(&crdp->cbs, &head->next); | |
628 | uatomic_inc(&crdp->qlen); | |
7106ddf8 | 629 | wake_call_rcu_thread(crdp); |
7d214e56 | 630 | _rcu_read_unlock(); |
7106ddf8 PM |
631 | } |
632 | ||
633 | /* | |
634 | * Free up the specified call_rcu_data structure, terminating the | |
635 | * associated call_rcu thread. The caller must have previously | |
636 | * removed the call_rcu_data structure from per-thread or per-CPU | |
637 | * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU | |
638 | * call_rcu_data structures or set_thread_call_rcu_data(NULL) for | |
639 | * per-thread call_rcu_data structures. | |
640 | * | |
641 | * We silently refuse to free up the default call_rcu_data structure | |
642 | * because that is where we put any leftover callbacks. Note that | |
643 | * the possibility of self-spawning callbacks makes it impossible | |
644 | * to execute all the callbacks in finite time without putting any | |
645 | * newly spawned callbacks somewhere else. The "somewhere else" of | |
646 | * last resort is the default call_rcu_data structure. | |
647 | * | |
648 | * We also silently refuse to free NULL pointers. This simplifies | |
649 | * the calling code. | |
f9da0936 MD |
650 | * |
651 | * The caller must wait for a grace-period to pass between return from | |
652 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
653 | * previous call rcu data as argument. | |
7106ddf8 PM |
654 | */ |
655 | void call_rcu_data_free(struct call_rcu_data *crdp) | |
656 | { | |
657 | struct cds_wfq_node *cbs; | |
658 | struct cds_wfq_node **cbs_tail; | |
659 | struct cds_wfq_node **cbs_endprev; | |
660 | ||
661 | if (crdp == NULL || crdp == default_call_rcu_data) { | |
662 | return; | |
663 | } | |
2870aa1e PB |
664 | if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) { |
665 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP); | |
7106ddf8 | 666 | wake_call_rcu_thread(crdp); |
2870aa1e | 667 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) |
7106ddf8 PM |
668 | poll(NULL, 0, 1); |
669 | } | |
670 | if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) { | |
671 | while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL) | |
672 | poll(NULL, 0, 1); | |
673 | _CMM_STORE_SHARED(crdp->cbs.head, NULL); | |
674 | cbs_tail = (struct cds_wfq_node **) | |
675 | uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head); | |
698d0778 MD |
676 | /* Create default call rcu data if need be */ |
677 | (void) get_default_call_rcu_data(); | |
7106ddf8 | 678 | cbs_endprev = (struct cds_wfq_node **) |
19cf3ae1 MD |
679 | uatomic_xchg(&default_call_rcu_data->cbs.tail, |
680 | cbs_tail); | |
681 | _CMM_STORE_SHARED(*cbs_endprev, cbs); | |
7106ddf8 PM |
682 | uatomic_add(&default_call_rcu_data->qlen, |
683 | uatomic_read(&crdp->qlen)); | |
1e92aa15 | 684 | wake_call_rcu_thread(default_call_rcu_data); |
7106ddf8 | 685 | } |
59dc9e9d | 686 | |
b75dffe6 | 687 | call_rcu_lock(&call_rcu_mutex); |
59dc9e9d | 688 | cds_list_del(&crdp->list); |
b75dffe6 LJ |
689 | call_rcu_unlock(&call_rcu_mutex); |
690 | ||
59dc9e9d | 691 | free(crdp); |
7106ddf8 PM |
692 | } |
693 | ||
694 | /* | |
695 | * Clean up all the per-CPU call_rcu threads. | |
696 | */ | |
697 | void free_all_cpu_call_rcu_data(void) | |
698 | { | |
699 | int cpu; | |
618b2595 MD |
700 | struct call_rcu_data **crdp; |
701 | static int warned = 0; | |
7106ddf8 PM |
702 | |
703 | if (maxcpus <= 0) | |
704 | return; | |
618b2595 MD |
705 | |
706 | crdp = malloc(sizeof(*crdp) * maxcpus); | |
707 | if (!crdp) { | |
708 | if (!warned) { | |
709 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
710 | } | |
711 | warned = 1; | |
d31150b4 | 712 | return; |
618b2595 MD |
713 | } |
714 | ||
7106ddf8 | 715 | for (cpu = 0; cpu < maxcpus; cpu++) { |
618b2595 MD |
716 | crdp[cpu] = get_cpu_call_rcu_data(cpu); |
717 | if (crdp[cpu] == NULL) | |
7106ddf8 PM |
718 | continue; |
719 | set_cpu_call_rcu_data(cpu, NULL); | |
7106ddf8 | 720 | } |
618b2595 MD |
721 | /* |
722 | * Wait for call_rcu sites acting as RCU readers of the | |
723 | * call_rcu_data to become quiescent. | |
724 | */ | |
725 | synchronize_rcu(); | |
726 | for (cpu = 0; cpu < maxcpus; cpu++) { | |
727 | if (crdp[cpu] == NULL) | |
728 | continue; | |
729 | call_rcu_data_free(crdp[cpu]); | |
730 | } | |
731 | free(crdp); | |
7106ddf8 PM |
732 | } |
733 | ||
81ad2e19 PM |
734 | /* |
735 | * Acquire the call_rcu_mutex in order to ensure that the child sees | |
750cfec5 MD |
736 | * all of the call_rcu() data structures in a consistent state. Ensure |
737 | * that all call_rcu threads are in a quiescent state across fork. | |
81ad2e19 PM |
738 | * Suitable for pthread_atfork() and friends. |
739 | */ | |
740 | void call_rcu_before_fork(void) | |
741 | { | |
750cfec5 MD |
742 | struct call_rcu_data *crdp; |
743 | ||
81ad2e19 | 744 | call_rcu_lock(&call_rcu_mutex); |
750cfec5 MD |
745 | |
746 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
747 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE); | |
748 | cmm_smp_mb__after_uatomic_or(); | |
749 | wake_call_rcu_thread(crdp); | |
750 | } | |
751 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
752 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0) | |
753 | poll(NULL, 0, 1); | |
754 | } | |
81ad2e19 PM |
755 | } |
756 | ||
757 | /* | |
758 | * Clean up call_rcu data structures in the parent of a successful fork() | |
759 | * that is not followed by exec() in the child. Suitable for | |
760 | * pthread_atfork() and friends. | |
761 | */ | |
762 | void call_rcu_after_fork_parent(void) | |
763 | { | |
750cfec5 MD |
764 | struct call_rcu_data *crdp; |
765 | ||
766 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) | |
767 | uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE); | |
e6564e91 KF |
768 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { |
769 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0) | |
770 | poll(NULL, 0, 1); | |
771 | } | |
81ad2e19 PM |
772 | call_rcu_unlock(&call_rcu_mutex); |
773 | } | |
774 | ||
7106ddf8 PM |
775 | /* |
776 | * Clean up call_rcu data structures in the child of a successful fork() | |
81ad2e19 PM |
777 | * that is not followed by exec(). Suitable for pthread_atfork() and |
778 | * friends. | |
7106ddf8 PM |
779 | */ |
780 | void call_rcu_after_fork_child(void) | |
781 | { | |
077ff173 | 782 | struct call_rcu_data *crdp, *next; |
7106ddf8 | 783 | |
81ad2e19 PM |
784 | /* Release the mutex. */ |
785 | call_rcu_unlock(&call_rcu_mutex); | |
786 | ||
ad1b9909 LJ |
787 | /* Do nothing when call_rcu() has not been used */ |
788 | if (cds_list_empty(&call_rcu_data_list)) | |
789 | return; | |
790 | ||
7106ddf8 PM |
791 | /* |
792 | * Allocate a new default call_rcu_data structure in order | |
793 | * to get a working call_rcu thread to go with it. | |
794 | */ | |
795 | default_call_rcu_data = NULL; | |
796 | (void)get_default_call_rcu_data(); | |
797 | ||
60af049d LJ |
798 | /* Cleanup call_rcu_data pointers before use */ |
799 | maxcpus_reset(); | |
800 | free(per_cpu_call_rcu_data); | |
618b2595 | 801 | rcu_set_pointer(&per_cpu_call_rcu_data, NULL); |
bd252a04 | 802 | URCU_TLS(thread_call_rcu_data) = NULL; |
60af049d | 803 | |
750cfec5 MD |
804 | /* |
805 | * Dispose of all of the rest of the call_rcu_data structures. | |
806 | * Leftover call_rcu callbacks will be merged into the new | |
807 | * default call_rcu thread queue. | |
808 | */ | |
077ff173 | 809 | cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) { |
7106ddf8 | 810 | if (crdp == default_call_rcu_data) |
077ff173 | 811 | continue; |
2870aa1e | 812 | uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED); |
7106ddf8 | 813 | call_rcu_data_free(crdp); |
b57aee66 PM |
814 | } |
815 | } |