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 | 77 | |
e85451a1 MD |
78 | /* |
79 | * Guard call_rcu thread creation and atfork handlers. | |
80 | */ | |
b57aee66 PM |
81 | static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER; |
82 | ||
83 | /* If a given thread does not have its own call_rcu thread, this is default. */ | |
84 | ||
85 | static struct call_rcu_data *default_call_rcu_data; | |
86 | ||
b57aee66 PM |
87 | /* |
88 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are | |
89 | * available, then we can have call_rcu threads assigned to individual | |
90 | * CPUs rather than only to specific threads. | |
91 | */ | |
92 | ||
63b495d8 MD |
93 | #ifdef HAVE_SCHED_GETCPU |
94 | ||
95 | static int urcu_sched_getcpu(void) | |
96 | { | |
97 | return sched_getcpu(); | |
98 | } | |
99 | ||
100 | #else /* #ifdef HAVE_SCHED_GETCPU */ | |
101 | ||
102 | static int urcu_sched_getcpu(void) | |
103 | { | |
104 | return -1; | |
105 | } | |
106 | ||
107 | #endif /* #else #ifdef HAVE_SCHED_GETCPU */ | |
108 | ||
109 | #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) | |
b57aee66 PM |
110 | |
111 | /* | |
112 | * Pointer to array of pointers to per-CPU call_rcu_data structures | |
618b2595 MD |
113 | * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an |
114 | * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a | |
115 | * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer | |
116 | * without mutex. The call_rcu_mutex protects updates. | |
b57aee66 PM |
117 | */ |
118 | ||
119 | static struct call_rcu_data **per_cpu_call_rcu_data; | |
120 | static long maxcpus; | |
121 | ||
60af049d LJ |
122 | static void maxcpus_reset(void) |
123 | { | |
124 | maxcpus = 0; | |
125 | } | |
126 | ||
b57aee66 PM |
127 | /* Allocate the array if it has not already been allocated. */ |
128 | ||
129 | static void alloc_cpu_call_rcu_data(void) | |
130 | { | |
131 | struct call_rcu_data **p; | |
132 | static int warned = 0; | |
133 | ||
134 | if (maxcpus != 0) | |
135 | return; | |
136 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
137 | if (maxcpus <= 0) { | |
138 | return; | |
139 | } | |
140 | p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
141 | if (p != NULL) { | |
142 | memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
618b2595 | 143 | rcu_set_pointer(&per_cpu_call_rcu_data, p); |
b57aee66 PM |
144 | } else { |
145 | if (!warned) { | |
146 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
147 | } | |
148 | warned = 1; | |
149 | } | |
150 | } | |
151 | ||
63b495d8 | 152 | #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
b57aee66 | 153 | |
f9437098 MD |
154 | /* |
155 | * per_cpu_call_rcu_data should be constant, but some functions below, used both | |
156 | * for cases where cpu number is available and not available, assume it it not | |
157 | * constant. | |
158 | */ | |
159 | static struct call_rcu_data **per_cpu_call_rcu_data = NULL; | |
b57aee66 PM |
160 | static const long maxcpus = -1; |
161 | ||
60af049d LJ |
162 | static void maxcpus_reset(void) |
163 | { | |
164 | } | |
165 | ||
b57aee66 PM |
166 | static void alloc_cpu_call_rcu_data(void) |
167 | { | |
168 | } | |
169 | ||
63b495d8 | 170 | #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
b57aee66 PM |
171 | |
172 | /* Acquire the specified pthread mutex. */ | |
173 | ||
174 | static void call_rcu_lock(pthread_mutex_t *pmp) | |
175 | { | |
4a6d7378 MD |
176 | int ret; |
177 | ||
178 | ret = pthread_mutex_lock(pmp); | |
179 | if (ret) | |
180 | urcu_die(ret); | |
b57aee66 PM |
181 | } |
182 | ||
183 | /* Release the specified pthread mutex. */ | |
184 | ||
185 | static void call_rcu_unlock(pthread_mutex_t *pmp) | |
186 | { | |
4a6d7378 MD |
187 | int ret; |
188 | ||
189 | ret = pthread_mutex_unlock(pmp); | |
190 | if (ret) | |
191 | urcu_die(ret); | |
b57aee66 PM |
192 | } |
193 | ||
c1d2c60b MD |
194 | #if HAVE_SCHED_SETAFFINITY |
195 | static | |
196 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
197 | { | |
198 | cpu_set_t mask; | |
199 | ||
200 | if (crdp->cpu_affinity < 0) | |
201 | return 0; | |
202 | ||
203 | CPU_ZERO(&mask); | |
204 | CPU_SET(crdp->cpu_affinity, &mask); | |
205 | #if SCHED_SETAFFINITY_ARGS == 2 | |
206 | return sched_setaffinity(0, &mask); | |
207 | #else | |
208 | return sched_setaffinity(0, sizeof(mask), &mask); | |
209 | #endif | |
210 | } | |
211 | #else | |
212 | static | |
213 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
214 | { | |
215 | return 0; | |
216 | } | |
217 | #endif | |
218 | ||
03fe58b3 MD |
219 | static void call_rcu_wait(struct call_rcu_data *crdp) |
220 | { | |
221 | /* Read call_rcu list before read futex */ | |
222 | cmm_smp_mb(); | |
223 | if (uatomic_read(&crdp->futex) == -1) | |
224 | futex_async(&crdp->futex, FUTEX_WAIT, -1, | |
225 | NULL, NULL, 0); | |
226 | } | |
227 | ||
228 | static void call_rcu_wake_up(struct call_rcu_data *crdp) | |
229 | { | |
230 | /* Write to call_rcu list before reading/writing futex */ | |
231 | cmm_smp_mb(); | |
a0b7f7ea | 232 | if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) { |
03fe58b3 MD |
233 | uatomic_set(&crdp->futex, 0); |
234 | futex_async(&crdp->futex, FUTEX_WAKE, 1, | |
235 | NULL, NULL, 0); | |
236 | } | |
237 | } | |
238 | ||
b57aee66 PM |
239 | /* This is the code run by each call_rcu thread. */ |
240 | ||
241 | static void *call_rcu_thread(void *arg) | |
242 | { | |
243 | unsigned long cbcount; | |
5161f31e | 244 | struct call_rcu_data *crdp = (struct call_rcu_data *) arg; |
2870aa1e | 245 | int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT); |
4a6d7378 | 246 | int ret; |
b57aee66 | 247 | |
4a6d7378 MD |
248 | ret = set_thread_cpu_affinity(crdp); |
249 | if (ret) | |
250 | urcu_die(errno); | |
c1d2c60b | 251 | |
765f3ead MD |
252 | /* |
253 | * If callbacks take a read-side lock, we need to be registered. | |
254 | */ | |
255 | rcu_register_thread(); | |
256 | ||
bd252a04 | 257 | URCU_TLS(thread_call_rcu_data) = crdp; |
bc94ca9b MD |
258 | if (!rt) { |
259 | uatomic_dec(&crdp->futex); | |
260 | /* Decrement futex before reading call_rcu list */ | |
261 | cmm_smp_mb(); | |
262 | } | |
b57aee66 | 263 | for (;;) { |
5161f31e MD |
264 | struct cds_wfcq_head cbs_tmp_head; |
265 | struct cds_wfcq_tail cbs_tmp_tail; | |
266 | struct cds_wfcq_node *cbs, *cbs_tmp_n; | |
ae25b7e2 | 267 | enum cds_wfcq_ret splice_ret; |
5161f31e | 268 | |
e85451a1 MD |
269 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) { |
270 | /* | |
271 | * Pause requested. Become quiescent: remove | |
272 | * ourself from all global lists, and don't | |
273 | * process any callback. The callback lists may | |
274 | * still be non-empty though. | |
275 | */ | |
276 | rcu_unregister_thread(); | |
277 | cmm_smp_mb__before_uatomic_or(); | |
278 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED); | |
279 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0) | |
280 | poll(NULL, 0, 1); | |
281 | rcu_register_thread(); | |
282 | } | |
283 | ||
5161f31e | 284 | cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail); |
ae25b7e2 MD |
285 | splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head, |
286 | &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail); | |
287 | assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK); | |
288 | assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY); | |
289 | if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) { | |
b57aee66 PM |
290 | synchronize_rcu(); |
291 | cbcount = 0; | |
5161f31e MD |
292 | __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head, |
293 | &cbs_tmp_tail, cbs, cbs_tmp_n) { | |
294 | struct rcu_head *rhp; | |
295 | ||
296 | rhp = caa_container_of(cbs, | |
297 | struct rcu_head, next); | |
b57aee66 PM |
298 | rhp->func(rhp); |
299 | cbcount++; | |
5161f31e | 300 | } |
b57aee66 PM |
301 | uatomic_sub(&crdp->qlen, cbcount); |
302 | } | |
bc94ca9b MD |
303 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP) |
304 | break; | |
765f3ead | 305 | rcu_thread_offline(); |
bc94ca9b | 306 | if (!rt) { |
5161f31e MD |
307 | if (cds_wfcq_empty(&crdp->cbs_head, |
308 | &crdp->cbs_tail)) { | |
bc94ca9b MD |
309 | call_rcu_wait(crdp); |
310 | poll(NULL, 0, 10); | |
311 | uatomic_dec(&crdp->futex); | |
c768e45e | 312 | /* |
bc94ca9b MD |
313 | * Decrement futex before reading |
314 | * call_rcu list. | |
c768e45e MD |
315 | */ |
316 | cmm_smp_mb(); | |
ccbac24d MD |
317 | } else { |
318 | poll(NULL, 0, 10); | |
c768e45e | 319 | } |
bc94ca9b MD |
320 | } else { |
321 | poll(NULL, 0, 10); | |
b57aee66 | 322 | } |
765f3ead | 323 | rcu_thread_online(); |
bc94ca9b MD |
324 | } |
325 | if (!rt) { | |
326 | /* | |
327 | * Read call_rcu list before write futex. | |
328 | */ | |
329 | cmm_smp_mb(); | |
330 | uatomic_set(&crdp->futex, 0); | |
b57aee66 | 331 | } |
2870aa1e | 332 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED); |
765f3ead | 333 | rcu_unregister_thread(); |
7106ddf8 | 334 | return NULL; |
b57aee66 PM |
335 | } |
336 | ||
337 | /* | |
338 | * Create both a call_rcu thread and the corresponding call_rcu_data | |
3c24913f PM |
339 | * structure, linking the structure in as specified. Caller must hold |
340 | * call_rcu_mutex. | |
b57aee66 PM |
341 | */ |
342 | ||
3c24913f | 343 | static void call_rcu_data_init(struct call_rcu_data **crdpp, |
c1d2c60b MD |
344 | unsigned long flags, |
345 | int cpu_affinity) | |
b57aee66 PM |
346 | { |
347 | struct call_rcu_data *crdp; | |
4a6d7378 | 348 | int ret; |
b57aee66 PM |
349 | |
350 | crdp = malloc(sizeof(*crdp)); | |
4a6d7378 MD |
351 | if (crdp == NULL) |
352 | urcu_die(errno); | |
b57aee66 | 353 | memset(crdp, '\0', sizeof(*crdp)); |
5161f31e | 354 | cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail); |
b57aee66 | 355 | crdp->qlen = 0; |
263e3cf9 MD |
356 | crdp->futex = 0; |
357 | crdp->flags = flags; | |
3c24913f | 358 | cds_list_add(&crdp->list, &call_rcu_data_list); |
c1d2c60b | 359 | crdp->cpu_affinity = cpu_affinity; |
b57aee66 PM |
360 | cmm_smp_mb(); /* Structure initialized before pointer is planted. */ |
361 | *crdpp = crdp; | |
4a6d7378 MD |
362 | ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp); |
363 | if (ret) | |
364 | urcu_die(ret); | |
b57aee66 PM |
365 | } |
366 | ||
367 | /* | |
368 | * Return a pointer to the call_rcu_data structure for the specified | |
369 | * CPU, returning NULL if there is none. We cannot automatically | |
370 | * created it because the platform we are running on might not define | |
63b495d8 | 371 | * urcu_sched_getcpu(). |
618b2595 MD |
372 | * |
373 | * The call to this function and use of the returned call_rcu_data | |
374 | * should be protected by RCU read-side lock. | |
b57aee66 PM |
375 | */ |
376 | ||
377 | struct call_rcu_data *get_cpu_call_rcu_data(int cpu) | |
378 | { | |
379 | static int warned = 0; | |
618b2595 | 380 | struct call_rcu_data **pcpu_crdp; |
b57aee66 | 381 | |
618b2595 MD |
382 | pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data); |
383 | if (pcpu_crdp == NULL) | |
b57aee66 PM |
384 | return NULL; |
385 | if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) { | |
386 | fprintf(stderr, "[error] liburcu: get CPU # out of range\n"); | |
387 | warned = 1; | |
388 | } | |
389 | if (cpu < 0 || maxcpus <= cpu) | |
390 | return NULL; | |
618b2595 | 391 | return rcu_dereference(pcpu_crdp[cpu]); |
b57aee66 PM |
392 | } |
393 | ||
394 | /* | |
395 | * Return the tid corresponding to the call_rcu thread whose | |
396 | * call_rcu_data structure is specified. | |
397 | */ | |
398 | ||
399 | pthread_t get_call_rcu_thread(struct call_rcu_data *crdp) | |
400 | { | |
401 | return crdp->tid; | |
402 | } | |
403 | ||
404 | /* | |
405 | * Create a call_rcu_data structure (with thread) and return a pointer. | |
406 | */ | |
407 | ||
c1d2c60b MD |
408 | static struct call_rcu_data *__create_call_rcu_data(unsigned long flags, |
409 | int cpu_affinity) | |
b57aee66 PM |
410 | { |
411 | struct call_rcu_data *crdp; | |
412 | ||
c1d2c60b | 413 | call_rcu_data_init(&crdp, flags, cpu_affinity); |
b57aee66 PM |
414 | return crdp; |
415 | } | |
416 | ||
c1d2c60b MD |
417 | struct call_rcu_data *create_call_rcu_data(unsigned long flags, |
418 | int cpu_affinity) | |
3c24913f PM |
419 | { |
420 | struct call_rcu_data *crdp; | |
421 | ||
422 | call_rcu_lock(&call_rcu_mutex); | |
c1d2c60b | 423 | crdp = __create_call_rcu_data(flags, cpu_affinity); |
3c24913f PM |
424 | call_rcu_unlock(&call_rcu_mutex); |
425 | return crdp; | |
426 | } | |
427 | ||
b57aee66 PM |
428 | /* |
429 | * Set the specified CPU to use the specified call_rcu_data structure. | |
7106ddf8 PM |
430 | * |
431 | * Use NULL to remove a CPU's call_rcu_data structure, but it is | |
432 | * the caller's responsibility to dispose of the removed structure. | |
433 | * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure | |
434 | * (prior to NULLing it out, of course). | |
f9da0936 MD |
435 | * |
436 | * The caller must wait for a grace-period to pass between return from | |
437 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
438 | * previous call rcu data as argument. | |
b57aee66 PM |
439 | */ |
440 | ||
441 | int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp) | |
442 | { | |
dcfc8165 | 443 | static int warned = 0; |
b57aee66 PM |
444 | |
445 | call_rcu_lock(&call_rcu_mutex); | |
f3776786 | 446 | alloc_cpu_call_rcu_data(); |
b57aee66 PM |
447 | if (cpu < 0 || maxcpus <= cpu) { |
448 | if (!warned) { | |
449 | fprintf(stderr, "[error] liburcu: set CPU # out of range\n"); | |
450 | warned = 1; | |
451 | } | |
452 | call_rcu_unlock(&call_rcu_mutex); | |
453 | errno = EINVAL; | |
454 | return -EINVAL; | |
455 | } | |
53a55535 | 456 | |
b57aee66 | 457 | if (per_cpu_call_rcu_data == NULL) { |
3670fef2 | 458 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
459 | errno = ENOMEM; |
460 | return -ENOMEM; | |
461 | } | |
53a55535 LJ |
462 | |
463 | if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) { | |
464 | call_rcu_unlock(&call_rcu_mutex); | |
465 | errno = EEXIST; | |
466 | return -EEXIST; | |
467 | } | |
468 | ||
618b2595 | 469 | rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp); |
3670fef2 | 470 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
471 | return 0; |
472 | } | |
473 | ||
474 | /* | |
475 | * Return a pointer to the default call_rcu_data structure, creating | |
476 | * one if need be. Because we never free call_rcu_data structures, | |
477 | * we don't need to be in an RCU read-side critical section. | |
478 | */ | |
479 | ||
480 | struct call_rcu_data *get_default_call_rcu_data(void) | |
481 | { | |
482 | if (default_call_rcu_data != NULL) | |
483 | return rcu_dereference(default_call_rcu_data); | |
484 | call_rcu_lock(&call_rcu_mutex); | |
485 | if (default_call_rcu_data != NULL) { | |
486 | call_rcu_unlock(&call_rcu_mutex); | |
487 | return default_call_rcu_data; | |
488 | } | |
c1d2c60b | 489 | call_rcu_data_init(&default_call_rcu_data, 0, -1); |
b57aee66 PM |
490 | call_rcu_unlock(&call_rcu_mutex); |
491 | return default_call_rcu_data; | |
492 | } | |
493 | ||
494 | /* | |
495 | * Return the call_rcu_data structure that applies to the currently | |
496 | * running thread. Any call_rcu_data structure assigned specifically | |
497 | * to this thread has first priority, followed by any call_rcu_data | |
498 | * structure assigned to the CPU on which the thread is running, | |
499 | * followed by the default call_rcu_data structure. If there is not | |
500 | * yet a default call_rcu_data structure, one will be created. | |
618b2595 MD |
501 | * |
502 | * Calls to this function and use of the returned call_rcu_data should | |
503 | * be protected by RCU read-side lock. | |
b57aee66 PM |
504 | */ |
505 | struct call_rcu_data *get_call_rcu_data(void) | |
506 | { | |
9744f3bb | 507 | struct call_rcu_data *crd; |
b57aee66 | 508 | |
bd252a04 MD |
509 | if (URCU_TLS(thread_call_rcu_data) != NULL) |
510 | return URCU_TLS(thread_call_rcu_data); | |
9744f3bb LJ |
511 | |
512 | if (maxcpus > 0) { | |
63b495d8 | 513 | crd = get_cpu_call_rcu_data(urcu_sched_getcpu()); |
9744f3bb LJ |
514 | if (crd) |
515 | return crd; | |
b57aee66 | 516 | } |
9744f3bb | 517 | |
b57aee66 PM |
518 | return get_default_call_rcu_data(); |
519 | } | |
520 | ||
521 | /* | |
522 | * Return a pointer to this task's call_rcu_data if there is one. | |
523 | */ | |
524 | ||
525 | struct call_rcu_data *get_thread_call_rcu_data(void) | |
526 | { | |
bd252a04 | 527 | return URCU_TLS(thread_call_rcu_data); |
b57aee66 PM |
528 | } |
529 | ||
530 | /* | |
531 | * Set this task's call_rcu_data structure as specified, regardless | |
532 | * of whether or not this task already had one. (This allows switching | |
533 | * to and from real-time call_rcu threads, for example.) | |
7106ddf8 PM |
534 | * |
535 | * Use NULL to remove a thread's call_rcu_data structure, but it is | |
536 | * the caller's responsibility to dispose of the removed structure. | |
537 | * Use get_thread_call_rcu_data() to obtain a pointer to the old structure | |
538 | * (prior to NULLing it out, of course). | |
b57aee66 PM |
539 | */ |
540 | ||
541 | void set_thread_call_rcu_data(struct call_rcu_data *crdp) | |
542 | { | |
bd252a04 | 543 | URCU_TLS(thread_call_rcu_data) = crdp; |
b57aee66 PM |
544 | } |
545 | ||
546 | /* | |
547 | * Create a separate call_rcu thread for each CPU. This does not | |
548 | * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data() | |
0938c541 MD |
549 | * function if you want that behavior. Should be paired with |
550 | * free_all_cpu_call_rcu_data() to teardown these call_rcu worker | |
551 | * threads. | |
b57aee66 PM |
552 | */ |
553 | ||
554 | int create_all_cpu_call_rcu_data(unsigned long flags) | |
555 | { | |
556 | int i; | |
557 | struct call_rcu_data *crdp; | |
558 | int ret; | |
559 | ||
560 | call_rcu_lock(&call_rcu_mutex); | |
561 | alloc_cpu_call_rcu_data(); | |
562 | call_rcu_unlock(&call_rcu_mutex); | |
563 | if (maxcpus <= 0) { | |
564 | errno = EINVAL; | |
565 | return -EINVAL; | |
566 | } | |
567 | if (per_cpu_call_rcu_data == NULL) { | |
568 | errno = ENOMEM; | |
569 | return -ENOMEM; | |
570 | } | |
571 | for (i = 0; i < maxcpus; i++) { | |
572 | call_rcu_lock(&call_rcu_mutex); | |
573 | if (get_cpu_call_rcu_data(i)) { | |
574 | call_rcu_unlock(&call_rcu_mutex); | |
575 | continue; | |
576 | } | |
c1d2c60b | 577 | crdp = __create_call_rcu_data(flags, i); |
b57aee66 PM |
578 | if (crdp == NULL) { |
579 | call_rcu_unlock(&call_rcu_mutex); | |
580 | errno = ENOMEM; | |
581 | return -ENOMEM; | |
582 | } | |
583 | call_rcu_unlock(&call_rcu_mutex); | |
584 | if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) { | |
356c8794 LJ |
585 | call_rcu_data_free(crdp); |
586 | ||
587 | /* it has been created by other thread */ | |
588 | if (ret == -EEXIST) | |
589 | continue; | |
590 | ||
591 | return ret; | |
b57aee66 PM |
592 | } |
593 | } | |
594 | return 0; | |
595 | } | |
596 | ||
7106ddf8 PM |
597 | /* |
598 | * Wake up the call_rcu thread corresponding to the specified | |
599 | * call_rcu_data structure. | |
600 | */ | |
601 | static void wake_call_rcu_thread(struct call_rcu_data *crdp) | |
602 | { | |
263e3cf9 MD |
603 | if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT)) |
604 | call_rcu_wake_up(crdp); | |
7106ddf8 PM |
605 | } |
606 | ||
b57aee66 PM |
607 | /* |
608 | * Schedule a function to be invoked after a following grace period. | |
609 | * This is the only function that must be called -- the others are | |
610 | * only present to allow applications to tune their use of RCU for | |
611 | * maximum performance. | |
612 | * | |
613 | * Note that unless a call_rcu thread has not already been created, | |
614 | * the first invocation of call_rcu() will create one. So, if you | |
615 | * need the first invocation of call_rcu() to be fast, make sure | |
616 | * to create a call_rcu thread first. One way to accomplish this is | |
617 | * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data(). | |
618b2595 MD |
618 | * |
619 | * call_rcu must be called by registered RCU read-side threads. | |
b57aee66 PM |
620 | */ |
621 | ||
622 | void call_rcu(struct rcu_head *head, | |
623 | void (*func)(struct rcu_head *head)) | |
624 | { | |
625 | struct call_rcu_data *crdp; | |
626 | ||
5161f31e | 627 | cds_wfcq_node_init(&head->next); |
b57aee66 | 628 | head->func = func; |
618b2595 MD |
629 | /* Holding rcu read-side lock across use of per-cpu crdp */ |
630 | rcu_read_lock(); | |
b57aee66 | 631 | crdp = get_call_rcu_data(); |
5161f31e | 632 | cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next); |
b57aee66 | 633 | uatomic_inc(&crdp->qlen); |
7106ddf8 | 634 | wake_call_rcu_thread(crdp); |
618b2595 | 635 | rcu_read_unlock(); |
7106ddf8 PM |
636 | } |
637 | ||
638 | /* | |
639 | * Free up the specified call_rcu_data structure, terminating the | |
640 | * associated call_rcu thread. The caller must have previously | |
641 | * removed the call_rcu_data structure from per-thread or per-CPU | |
642 | * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU | |
643 | * call_rcu_data structures or set_thread_call_rcu_data(NULL) for | |
644 | * per-thread call_rcu_data structures. | |
645 | * | |
646 | * We silently refuse to free up the default call_rcu_data structure | |
647 | * because that is where we put any leftover callbacks. Note that | |
648 | * the possibility of self-spawning callbacks makes it impossible | |
649 | * to execute all the callbacks in finite time without putting any | |
650 | * newly spawned callbacks somewhere else. The "somewhere else" of | |
651 | * last resort is the default call_rcu_data structure. | |
652 | * | |
653 | * We also silently refuse to free NULL pointers. This simplifies | |
654 | * the calling code. | |
f9da0936 MD |
655 | * |
656 | * The caller must wait for a grace-period to pass between return from | |
657 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
658 | * previous call rcu data as argument. | |
03e5118f MD |
659 | * |
660 | * Note: introducing __cds_wfcq_splice_blocking() in this function fixed | |
661 | * a list corruption bug in the 0.7.x series. The equivalent fix | |
662 | * appeared in 0.6.8 for the stable-0.6 branch. | |
7106ddf8 PM |
663 | */ |
664 | void call_rcu_data_free(struct call_rcu_data *crdp) | |
665 | { | |
7106ddf8 PM |
666 | if (crdp == NULL || crdp == default_call_rcu_data) { |
667 | return; | |
668 | } | |
2870aa1e PB |
669 | if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) { |
670 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP); | |
7106ddf8 | 671 | wake_call_rcu_thread(crdp); |
2870aa1e | 672 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) |
7106ddf8 PM |
673 | poll(NULL, 0, 1); |
674 | } | |
5161f31e | 675 | if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) { |
698d0778 MD |
676 | /* Create default call rcu data if need be */ |
677 | (void) get_default_call_rcu_data(); | |
5161f31e MD |
678 | __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head, |
679 | &default_call_rcu_data->cbs_tail, | |
680 | &crdp->cbs_head, &crdp->cbs_tail); | |
7106ddf8 PM |
681 | uatomic_add(&default_call_rcu_data->qlen, |
682 | uatomic_read(&crdp->qlen)); | |
1e92aa15 | 683 | wake_call_rcu_thread(default_call_rcu_data); |
7106ddf8 | 684 | } |
59dc9e9d | 685 | |
b75dffe6 | 686 | call_rcu_lock(&call_rcu_mutex); |
59dc9e9d | 687 | cds_list_del(&crdp->list); |
b75dffe6 LJ |
688 | call_rcu_unlock(&call_rcu_mutex); |
689 | ||
59dc9e9d | 690 | free(crdp); |
7106ddf8 PM |
691 | } |
692 | ||
693 | /* | |
694 | * Clean up all the per-CPU call_rcu threads. | |
695 | */ | |
696 | void free_all_cpu_call_rcu_data(void) | |
697 | { | |
698 | int cpu; | |
618b2595 MD |
699 | struct call_rcu_data **crdp; |
700 | static int warned = 0; | |
7106ddf8 PM |
701 | |
702 | if (maxcpus <= 0) | |
703 | return; | |
618b2595 MD |
704 | |
705 | crdp = malloc(sizeof(*crdp) * maxcpus); | |
706 | if (!crdp) { | |
707 | if (!warned) { | |
708 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
709 | } | |
710 | warned = 1; | |
d31150b4 | 711 | return; |
618b2595 MD |
712 | } |
713 | ||
7106ddf8 | 714 | for (cpu = 0; cpu < maxcpus; cpu++) { |
618b2595 MD |
715 | crdp[cpu] = get_cpu_call_rcu_data(cpu); |
716 | if (crdp[cpu] == NULL) | |
7106ddf8 PM |
717 | continue; |
718 | set_cpu_call_rcu_data(cpu, NULL); | |
7106ddf8 | 719 | } |
618b2595 MD |
720 | /* |
721 | * Wait for call_rcu sites acting as RCU readers of the | |
722 | * call_rcu_data to become quiescent. | |
723 | */ | |
724 | synchronize_rcu(); | |
725 | for (cpu = 0; cpu < maxcpus; cpu++) { | |
726 | if (crdp[cpu] == NULL) | |
727 | continue; | |
728 | call_rcu_data_free(crdp[cpu]); | |
729 | } | |
730 | free(crdp); | |
7106ddf8 PM |
731 | } |
732 | ||
81ad2e19 PM |
733 | /* |
734 | * Acquire the call_rcu_mutex in order to ensure that the child sees | |
e85451a1 MD |
735 | * all of the call_rcu() data structures in a consistent state. Ensure |
736 | * that all call_rcu threads are in a quiescent state across fork. | |
81ad2e19 PM |
737 | * Suitable for pthread_atfork() and friends. |
738 | */ | |
739 | void call_rcu_before_fork(void) | |
740 | { | |
e85451a1 MD |
741 | struct call_rcu_data *crdp; |
742 | ||
81ad2e19 | 743 | call_rcu_lock(&call_rcu_mutex); |
e85451a1 MD |
744 | |
745 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
746 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE); | |
747 | cmm_smp_mb__after_uatomic_or(); | |
748 | wake_call_rcu_thread(crdp); | |
749 | } | |
750 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
751 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0) | |
752 | poll(NULL, 0, 1); | |
753 | } | |
81ad2e19 PM |
754 | } |
755 | ||
756 | /* | |
757 | * Clean up call_rcu data structures in the parent of a successful fork() | |
758 | * that is not followed by exec() in the child. Suitable for | |
759 | * pthread_atfork() and friends. | |
760 | */ | |
761 | void call_rcu_after_fork_parent(void) | |
762 | { | |
e85451a1 MD |
763 | struct call_rcu_data *crdp; |
764 | ||
765 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) | |
766 | uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE); | |
81ad2e19 PM |
767 | call_rcu_unlock(&call_rcu_mutex); |
768 | } | |
769 | ||
7106ddf8 PM |
770 | /* |
771 | * Clean up call_rcu data structures in the child of a successful fork() | |
81ad2e19 PM |
772 | * that is not followed by exec(). Suitable for pthread_atfork() and |
773 | * friends. | |
7106ddf8 PM |
774 | */ |
775 | void call_rcu_after_fork_child(void) | |
776 | { | |
077ff173 | 777 | struct call_rcu_data *crdp, *next; |
7106ddf8 | 778 | |
81ad2e19 PM |
779 | /* Release the mutex. */ |
780 | call_rcu_unlock(&call_rcu_mutex); | |
781 | ||
ad1b9909 LJ |
782 | /* Do nothing when call_rcu() has not been used */ |
783 | if (cds_list_empty(&call_rcu_data_list)) | |
784 | return; | |
785 | ||
7106ddf8 PM |
786 | /* |
787 | * Allocate a new default call_rcu_data structure in order | |
788 | * to get a working call_rcu thread to go with it. | |
789 | */ | |
790 | default_call_rcu_data = NULL; | |
791 | (void)get_default_call_rcu_data(); | |
792 | ||
60af049d LJ |
793 | /* Cleanup call_rcu_data pointers before use */ |
794 | maxcpus_reset(); | |
795 | free(per_cpu_call_rcu_data); | |
618b2595 | 796 | rcu_set_pointer(&per_cpu_call_rcu_data, NULL); |
bd252a04 | 797 | URCU_TLS(thread_call_rcu_data) = NULL; |
60af049d | 798 | |
e85451a1 MD |
799 | /* |
800 | * Dispose of all of the rest of the call_rcu_data structures. | |
801 | * Leftover call_rcu callbacks will be merged into the new | |
802 | * default call_rcu thread queue. | |
803 | */ | |
077ff173 | 804 | cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) { |
7106ddf8 | 805 | if (crdp == default_call_rcu_data) |
077ff173 | 806 | continue; |
2870aa1e | 807 | uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED); |
7106ddf8 | 808 | call_rcu_data_free(crdp); |
b57aee66 PM |
809 | } |
810 | } |