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 | ||
b7f721d9 MD |
67 | struct call_rcu_completion { |
68 | int barrier_count; | |
69 | int32_t futex; | |
70 | }; | |
71 | ||
72 | struct call_rcu_completion_work { | |
73 | struct rcu_head head; | |
74 | struct call_rcu_completion *completion; | |
75 | }; | |
76 | ||
3c24913f PM |
77 | /* |
78 | * List of all call_rcu_data structures to keep valgrind happy. | |
79 | * Protected by call_rcu_mutex. | |
80 | */ | |
81 | ||
bab44e28 | 82 | static CDS_LIST_HEAD(call_rcu_data_list); |
3c24913f | 83 | |
b57aee66 PM |
84 | /* Link a thread using call_rcu() to its call_rcu thread. */ |
85 | ||
bd252a04 | 86 | static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data); |
b57aee66 | 87 | |
e85451a1 MD |
88 | /* |
89 | * Guard call_rcu thread creation and atfork handlers. | |
90 | */ | |
b57aee66 PM |
91 | static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER; |
92 | ||
93 | /* If a given thread does not have its own call_rcu thread, this is default. */ | |
94 | ||
95 | static struct call_rcu_data *default_call_rcu_data; | |
96 | ||
b57aee66 PM |
97 | /* |
98 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are | |
99 | * available, then we can have call_rcu threads assigned to individual | |
100 | * CPUs rather than only to specific threads. | |
101 | */ | |
102 | ||
63b495d8 MD |
103 | #ifdef HAVE_SCHED_GETCPU |
104 | ||
105 | static int urcu_sched_getcpu(void) | |
106 | { | |
107 | return sched_getcpu(); | |
108 | } | |
109 | ||
110 | #else /* #ifdef HAVE_SCHED_GETCPU */ | |
111 | ||
112 | static int urcu_sched_getcpu(void) | |
113 | { | |
114 | return -1; | |
115 | } | |
116 | ||
117 | #endif /* #else #ifdef HAVE_SCHED_GETCPU */ | |
118 | ||
119 | #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) | |
b57aee66 PM |
120 | |
121 | /* | |
122 | * Pointer to array of pointers to per-CPU call_rcu_data structures | |
618b2595 MD |
123 | * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an |
124 | * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a | |
125 | * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer | |
126 | * without mutex. The call_rcu_mutex protects updates. | |
b57aee66 PM |
127 | */ |
128 | ||
129 | static struct call_rcu_data **per_cpu_call_rcu_data; | |
130 | static long maxcpus; | |
131 | ||
60af049d LJ |
132 | static void maxcpus_reset(void) |
133 | { | |
134 | maxcpus = 0; | |
135 | } | |
136 | ||
b57aee66 PM |
137 | /* Allocate the array if it has not already been allocated. */ |
138 | ||
139 | static void alloc_cpu_call_rcu_data(void) | |
140 | { | |
141 | struct call_rcu_data **p; | |
142 | static int warned = 0; | |
143 | ||
144 | if (maxcpus != 0) | |
145 | return; | |
146 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
147 | if (maxcpus <= 0) { | |
148 | return; | |
149 | } | |
150 | p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
151 | if (p != NULL) { | |
152 | memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
618b2595 | 153 | rcu_set_pointer(&per_cpu_call_rcu_data, p); |
b57aee66 PM |
154 | } else { |
155 | if (!warned) { | |
156 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
157 | } | |
158 | warned = 1; | |
159 | } | |
160 | } | |
161 | ||
63b495d8 | 162 | #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
b57aee66 | 163 | |
f9437098 MD |
164 | /* |
165 | * per_cpu_call_rcu_data should be constant, but some functions below, used both | |
166 | * for cases where cpu number is available and not available, assume it it not | |
167 | * constant. | |
168 | */ | |
169 | static struct call_rcu_data **per_cpu_call_rcu_data = NULL; | |
b57aee66 PM |
170 | static const long maxcpus = -1; |
171 | ||
60af049d LJ |
172 | static void maxcpus_reset(void) |
173 | { | |
174 | } | |
175 | ||
b57aee66 PM |
176 | static void alloc_cpu_call_rcu_data(void) |
177 | { | |
178 | } | |
179 | ||
63b495d8 | 180 | #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
b57aee66 PM |
181 | |
182 | /* Acquire the specified pthread mutex. */ | |
183 | ||
184 | static void call_rcu_lock(pthread_mutex_t *pmp) | |
185 | { | |
4a6d7378 MD |
186 | int ret; |
187 | ||
188 | ret = pthread_mutex_lock(pmp); | |
189 | if (ret) | |
190 | urcu_die(ret); | |
b57aee66 PM |
191 | } |
192 | ||
193 | /* Release the specified pthread mutex. */ | |
194 | ||
195 | static void call_rcu_unlock(pthread_mutex_t *pmp) | |
196 | { | |
4a6d7378 MD |
197 | int ret; |
198 | ||
199 | ret = pthread_mutex_unlock(pmp); | |
200 | if (ret) | |
201 | urcu_die(ret); | |
b57aee66 PM |
202 | } |
203 | ||
c1d2c60b MD |
204 | #if HAVE_SCHED_SETAFFINITY |
205 | static | |
206 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
207 | { | |
208 | cpu_set_t mask; | |
209 | ||
210 | if (crdp->cpu_affinity < 0) | |
211 | return 0; | |
212 | ||
213 | CPU_ZERO(&mask); | |
214 | CPU_SET(crdp->cpu_affinity, &mask); | |
215 | #if SCHED_SETAFFINITY_ARGS == 2 | |
216 | return sched_setaffinity(0, &mask); | |
217 | #else | |
218 | return sched_setaffinity(0, sizeof(mask), &mask); | |
219 | #endif | |
220 | } | |
221 | #else | |
222 | static | |
223 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
224 | { | |
225 | return 0; | |
226 | } | |
227 | #endif | |
228 | ||
03fe58b3 MD |
229 | static void call_rcu_wait(struct call_rcu_data *crdp) |
230 | { | |
231 | /* Read call_rcu list before read futex */ | |
232 | cmm_smp_mb(); | |
233 | if (uatomic_read(&crdp->futex) == -1) | |
234 | futex_async(&crdp->futex, FUTEX_WAIT, -1, | |
235 | NULL, NULL, 0); | |
236 | } | |
237 | ||
238 | static void call_rcu_wake_up(struct call_rcu_data *crdp) | |
239 | { | |
240 | /* Write to call_rcu list before reading/writing futex */ | |
241 | cmm_smp_mb(); | |
a0b7f7ea | 242 | if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) { |
03fe58b3 MD |
243 | uatomic_set(&crdp->futex, 0); |
244 | futex_async(&crdp->futex, FUTEX_WAKE, 1, | |
245 | NULL, NULL, 0); | |
246 | } | |
247 | } | |
248 | ||
b7f721d9 MD |
249 | static void call_rcu_completion_wait(struct call_rcu_completion *completion) |
250 | { | |
251 | /* Read completion barrier count before read futex */ | |
252 | cmm_smp_mb(); | |
253 | if (uatomic_read(&completion->futex) == -1) | |
254 | futex_async(&completion->futex, FUTEX_WAIT, -1, | |
255 | NULL, NULL, 0); | |
256 | } | |
257 | ||
258 | static void call_rcu_completion_wake_up(struct call_rcu_completion *completion) | |
259 | { | |
260 | /* Write to completion barrier count before reading/writing futex */ | |
261 | cmm_smp_mb(); | |
262 | if (caa_unlikely(uatomic_read(&completion->futex) == -1)) { | |
263 | uatomic_set(&completion->futex, 0); | |
264 | futex_async(&completion->futex, FUTEX_WAKE, 1, | |
265 | NULL, NULL, 0); | |
266 | } | |
267 | } | |
268 | ||
b57aee66 PM |
269 | /* This is the code run by each call_rcu thread. */ |
270 | ||
271 | static void *call_rcu_thread(void *arg) | |
272 | { | |
273 | unsigned long cbcount; | |
5161f31e | 274 | struct call_rcu_data *crdp = (struct call_rcu_data *) arg; |
2870aa1e | 275 | int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT); |
4a6d7378 | 276 | int ret; |
b57aee66 | 277 | |
4a6d7378 MD |
278 | ret = set_thread_cpu_affinity(crdp); |
279 | if (ret) | |
280 | urcu_die(errno); | |
c1d2c60b | 281 | |
765f3ead MD |
282 | /* |
283 | * If callbacks take a read-side lock, we need to be registered. | |
284 | */ | |
285 | rcu_register_thread(); | |
286 | ||
bd252a04 | 287 | URCU_TLS(thread_call_rcu_data) = crdp; |
bc94ca9b MD |
288 | if (!rt) { |
289 | uatomic_dec(&crdp->futex); | |
290 | /* Decrement futex before reading call_rcu list */ | |
291 | cmm_smp_mb(); | |
292 | } | |
b57aee66 | 293 | for (;;) { |
5161f31e MD |
294 | struct cds_wfcq_head cbs_tmp_head; |
295 | struct cds_wfcq_tail cbs_tmp_tail; | |
296 | struct cds_wfcq_node *cbs, *cbs_tmp_n; | |
ae25b7e2 | 297 | enum cds_wfcq_ret splice_ret; |
5161f31e | 298 | |
e85451a1 MD |
299 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) { |
300 | /* | |
301 | * Pause requested. Become quiescent: remove | |
302 | * ourself from all global lists, and don't | |
303 | * process any callback. The callback lists may | |
304 | * still be non-empty though. | |
305 | */ | |
306 | rcu_unregister_thread(); | |
307 | cmm_smp_mb__before_uatomic_or(); | |
308 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED); | |
309 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0) | |
310 | poll(NULL, 0, 1); | |
db2b0858 KF |
311 | uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED); |
312 | cmm_smp_mb__after_uatomic_and(); | |
e85451a1 MD |
313 | rcu_register_thread(); |
314 | } | |
315 | ||
5161f31e | 316 | cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail); |
ae25b7e2 MD |
317 | splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head, |
318 | &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail); | |
319 | assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK); | |
320 | assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY); | |
321 | if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) { | |
b57aee66 PM |
322 | synchronize_rcu(); |
323 | cbcount = 0; | |
5161f31e MD |
324 | __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head, |
325 | &cbs_tmp_tail, cbs, cbs_tmp_n) { | |
326 | struct rcu_head *rhp; | |
327 | ||
328 | rhp = caa_container_of(cbs, | |
329 | struct rcu_head, next); | |
b57aee66 PM |
330 | rhp->func(rhp); |
331 | cbcount++; | |
5161f31e | 332 | } |
b57aee66 PM |
333 | uatomic_sub(&crdp->qlen, cbcount); |
334 | } | |
bc94ca9b MD |
335 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP) |
336 | break; | |
765f3ead | 337 | rcu_thread_offline(); |
bc94ca9b | 338 | if (!rt) { |
5161f31e MD |
339 | if (cds_wfcq_empty(&crdp->cbs_head, |
340 | &crdp->cbs_tail)) { | |
bc94ca9b MD |
341 | call_rcu_wait(crdp); |
342 | poll(NULL, 0, 10); | |
343 | uatomic_dec(&crdp->futex); | |
c768e45e | 344 | /* |
bc94ca9b MD |
345 | * Decrement futex before reading |
346 | * call_rcu list. | |
c768e45e MD |
347 | */ |
348 | cmm_smp_mb(); | |
ccbac24d MD |
349 | } else { |
350 | poll(NULL, 0, 10); | |
c768e45e | 351 | } |
bc94ca9b MD |
352 | } else { |
353 | poll(NULL, 0, 10); | |
b57aee66 | 354 | } |
765f3ead | 355 | rcu_thread_online(); |
bc94ca9b MD |
356 | } |
357 | if (!rt) { | |
358 | /* | |
359 | * Read call_rcu list before write futex. | |
360 | */ | |
361 | cmm_smp_mb(); | |
362 | uatomic_set(&crdp->futex, 0); | |
b57aee66 | 363 | } |
2870aa1e | 364 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED); |
765f3ead | 365 | rcu_unregister_thread(); |
7106ddf8 | 366 | return NULL; |
b57aee66 PM |
367 | } |
368 | ||
369 | /* | |
370 | * Create both a call_rcu thread and the corresponding call_rcu_data | |
3c24913f PM |
371 | * structure, linking the structure in as specified. Caller must hold |
372 | * call_rcu_mutex. | |
b57aee66 PM |
373 | */ |
374 | ||
3c24913f | 375 | static void call_rcu_data_init(struct call_rcu_data **crdpp, |
c1d2c60b MD |
376 | unsigned long flags, |
377 | int cpu_affinity) | |
b57aee66 PM |
378 | { |
379 | struct call_rcu_data *crdp; | |
4a6d7378 | 380 | int ret; |
b57aee66 PM |
381 | |
382 | crdp = malloc(sizeof(*crdp)); | |
4a6d7378 MD |
383 | if (crdp == NULL) |
384 | urcu_die(errno); | |
b57aee66 | 385 | memset(crdp, '\0', sizeof(*crdp)); |
5161f31e | 386 | cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail); |
b57aee66 | 387 | crdp->qlen = 0; |
263e3cf9 MD |
388 | crdp->futex = 0; |
389 | crdp->flags = flags; | |
3c24913f | 390 | cds_list_add(&crdp->list, &call_rcu_data_list); |
c1d2c60b | 391 | crdp->cpu_affinity = cpu_affinity; |
b57aee66 PM |
392 | cmm_smp_mb(); /* Structure initialized before pointer is planted. */ |
393 | *crdpp = crdp; | |
4a6d7378 MD |
394 | ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp); |
395 | if (ret) | |
396 | urcu_die(ret); | |
b57aee66 PM |
397 | } |
398 | ||
399 | /* | |
400 | * Return a pointer to the call_rcu_data structure for the specified | |
401 | * CPU, returning NULL if there is none. We cannot automatically | |
402 | * created it because the platform we are running on might not define | |
63b495d8 | 403 | * urcu_sched_getcpu(). |
618b2595 MD |
404 | * |
405 | * The call to this function and use of the returned call_rcu_data | |
406 | * should be protected by RCU read-side lock. | |
b57aee66 PM |
407 | */ |
408 | ||
409 | struct call_rcu_data *get_cpu_call_rcu_data(int cpu) | |
410 | { | |
411 | static int warned = 0; | |
618b2595 | 412 | struct call_rcu_data **pcpu_crdp; |
b57aee66 | 413 | |
618b2595 MD |
414 | pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data); |
415 | if (pcpu_crdp == NULL) | |
b57aee66 PM |
416 | return NULL; |
417 | if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) { | |
418 | fprintf(stderr, "[error] liburcu: get CPU # out of range\n"); | |
419 | warned = 1; | |
420 | } | |
421 | if (cpu < 0 || maxcpus <= cpu) | |
422 | return NULL; | |
618b2595 | 423 | return rcu_dereference(pcpu_crdp[cpu]); |
b57aee66 PM |
424 | } |
425 | ||
426 | /* | |
427 | * Return the tid corresponding to the call_rcu thread whose | |
428 | * call_rcu_data structure is specified. | |
429 | */ | |
430 | ||
431 | pthread_t get_call_rcu_thread(struct call_rcu_data *crdp) | |
432 | { | |
433 | return crdp->tid; | |
434 | } | |
435 | ||
436 | /* | |
437 | * Create a call_rcu_data structure (with thread) and return a pointer. | |
438 | */ | |
439 | ||
c1d2c60b MD |
440 | static struct call_rcu_data *__create_call_rcu_data(unsigned long flags, |
441 | int cpu_affinity) | |
b57aee66 PM |
442 | { |
443 | struct call_rcu_data *crdp; | |
444 | ||
c1d2c60b | 445 | call_rcu_data_init(&crdp, flags, cpu_affinity); |
b57aee66 PM |
446 | return crdp; |
447 | } | |
448 | ||
c1d2c60b MD |
449 | struct call_rcu_data *create_call_rcu_data(unsigned long flags, |
450 | int cpu_affinity) | |
3c24913f PM |
451 | { |
452 | struct call_rcu_data *crdp; | |
453 | ||
454 | call_rcu_lock(&call_rcu_mutex); | |
c1d2c60b | 455 | crdp = __create_call_rcu_data(flags, cpu_affinity); |
3c24913f PM |
456 | call_rcu_unlock(&call_rcu_mutex); |
457 | return crdp; | |
458 | } | |
459 | ||
b57aee66 PM |
460 | /* |
461 | * Set the specified CPU to use the specified call_rcu_data structure. | |
7106ddf8 PM |
462 | * |
463 | * Use NULL to remove a CPU's call_rcu_data structure, but it is | |
464 | * the caller's responsibility to dispose of the removed structure. | |
465 | * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure | |
466 | * (prior to NULLing it out, of course). | |
f9da0936 MD |
467 | * |
468 | * The caller must wait for a grace-period to pass between return from | |
469 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
470 | * previous call rcu data as argument. | |
b57aee66 PM |
471 | */ |
472 | ||
473 | int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp) | |
474 | { | |
dcfc8165 | 475 | static int warned = 0; |
b57aee66 PM |
476 | |
477 | call_rcu_lock(&call_rcu_mutex); | |
f3776786 | 478 | alloc_cpu_call_rcu_data(); |
b57aee66 PM |
479 | if (cpu < 0 || maxcpus <= cpu) { |
480 | if (!warned) { | |
481 | fprintf(stderr, "[error] liburcu: set CPU # out of range\n"); | |
482 | warned = 1; | |
483 | } | |
484 | call_rcu_unlock(&call_rcu_mutex); | |
485 | errno = EINVAL; | |
486 | return -EINVAL; | |
487 | } | |
53a55535 | 488 | |
b57aee66 | 489 | if (per_cpu_call_rcu_data == NULL) { |
3670fef2 | 490 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
491 | errno = ENOMEM; |
492 | return -ENOMEM; | |
493 | } | |
53a55535 LJ |
494 | |
495 | if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) { | |
496 | call_rcu_unlock(&call_rcu_mutex); | |
497 | errno = EEXIST; | |
498 | return -EEXIST; | |
499 | } | |
500 | ||
618b2595 | 501 | rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp); |
3670fef2 | 502 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
503 | return 0; |
504 | } | |
505 | ||
506 | /* | |
507 | * Return a pointer to the default call_rcu_data structure, creating | |
508 | * one if need be. Because we never free call_rcu_data structures, | |
509 | * we don't need to be in an RCU read-side critical section. | |
510 | */ | |
511 | ||
512 | struct call_rcu_data *get_default_call_rcu_data(void) | |
513 | { | |
514 | if (default_call_rcu_data != NULL) | |
515 | return rcu_dereference(default_call_rcu_data); | |
516 | call_rcu_lock(&call_rcu_mutex); | |
517 | if (default_call_rcu_data != NULL) { | |
518 | call_rcu_unlock(&call_rcu_mutex); | |
519 | return default_call_rcu_data; | |
520 | } | |
c1d2c60b | 521 | call_rcu_data_init(&default_call_rcu_data, 0, -1); |
b57aee66 PM |
522 | call_rcu_unlock(&call_rcu_mutex); |
523 | return default_call_rcu_data; | |
524 | } | |
525 | ||
526 | /* | |
527 | * Return the call_rcu_data structure that applies to the currently | |
528 | * running thread. Any call_rcu_data structure assigned specifically | |
529 | * to this thread has first priority, followed by any call_rcu_data | |
530 | * structure assigned to the CPU on which the thread is running, | |
531 | * followed by the default call_rcu_data structure. If there is not | |
532 | * yet a default call_rcu_data structure, one will be created. | |
618b2595 MD |
533 | * |
534 | * Calls to this function and use of the returned call_rcu_data should | |
535 | * be protected by RCU read-side lock. | |
b57aee66 PM |
536 | */ |
537 | struct call_rcu_data *get_call_rcu_data(void) | |
538 | { | |
9744f3bb | 539 | struct call_rcu_data *crd; |
b57aee66 | 540 | |
bd252a04 MD |
541 | if (URCU_TLS(thread_call_rcu_data) != NULL) |
542 | return URCU_TLS(thread_call_rcu_data); | |
9744f3bb LJ |
543 | |
544 | if (maxcpus > 0) { | |
63b495d8 | 545 | crd = get_cpu_call_rcu_data(urcu_sched_getcpu()); |
9744f3bb LJ |
546 | if (crd) |
547 | return crd; | |
b57aee66 | 548 | } |
9744f3bb | 549 | |
b57aee66 PM |
550 | return get_default_call_rcu_data(); |
551 | } | |
552 | ||
553 | /* | |
554 | * Return a pointer to this task's call_rcu_data if there is one. | |
555 | */ | |
556 | ||
557 | struct call_rcu_data *get_thread_call_rcu_data(void) | |
558 | { | |
bd252a04 | 559 | return URCU_TLS(thread_call_rcu_data); |
b57aee66 PM |
560 | } |
561 | ||
562 | /* | |
563 | * Set this task's call_rcu_data structure as specified, regardless | |
564 | * of whether or not this task already had one. (This allows switching | |
565 | * to and from real-time call_rcu threads, for example.) | |
7106ddf8 PM |
566 | * |
567 | * Use NULL to remove a thread's call_rcu_data structure, but it is | |
568 | * the caller's responsibility to dispose of the removed structure. | |
569 | * Use get_thread_call_rcu_data() to obtain a pointer to the old structure | |
570 | * (prior to NULLing it out, of course). | |
b57aee66 PM |
571 | */ |
572 | ||
573 | void set_thread_call_rcu_data(struct call_rcu_data *crdp) | |
574 | { | |
bd252a04 | 575 | URCU_TLS(thread_call_rcu_data) = crdp; |
b57aee66 PM |
576 | } |
577 | ||
578 | /* | |
579 | * Create a separate call_rcu thread for each CPU. This does not | |
580 | * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data() | |
0938c541 MD |
581 | * function if you want that behavior. Should be paired with |
582 | * free_all_cpu_call_rcu_data() to teardown these call_rcu worker | |
583 | * threads. | |
b57aee66 PM |
584 | */ |
585 | ||
586 | int create_all_cpu_call_rcu_data(unsigned long flags) | |
587 | { | |
588 | int i; | |
589 | struct call_rcu_data *crdp; | |
590 | int ret; | |
591 | ||
592 | call_rcu_lock(&call_rcu_mutex); | |
593 | alloc_cpu_call_rcu_data(); | |
594 | call_rcu_unlock(&call_rcu_mutex); | |
595 | if (maxcpus <= 0) { | |
596 | errno = EINVAL; | |
597 | return -EINVAL; | |
598 | } | |
599 | if (per_cpu_call_rcu_data == NULL) { | |
600 | errno = ENOMEM; | |
601 | return -ENOMEM; | |
602 | } | |
603 | for (i = 0; i < maxcpus; i++) { | |
604 | call_rcu_lock(&call_rcu_mutex); | |
605 | if (get_cpu_call_rcu_data(i)) { | |
606 | call_rcu_unlock(&call_rcu_mutex); | |
607 | continue; | |
608 | } | |
c1d2c60b | 609 | crdp = __create_call_rcu_data(flags, i); |
b57aee66 PM |
610 | if (crdp == NULL) { |
611 | call_rcu_unlock(&call_rcu_mutex); | |
612 | errno = ENOMEM; | |
613 | return -ENOMEM; | |
614 | } | |
615 | call_rcu_unlock(&call_rcu_mutex); | |
616 | if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) { | |
356c8794 LJ |
617 | call_rcu_data_free(crdp); |
618 | ||
619 | /* it has been created by other thread */ | |
620 | if (ret == -EEXIST) | |
621 | continue; | |
622 | ||
623 | return ret; | |
b57aee66 PM |
624 | } |
625 | } | |
626 | return 0; | |
627 | } | |
628 | ||
7106ddf8 PM |
629 | /* |
630 | * Wake up the call_rcu thread corresponding to the specified | |
631 | * call_rcu_data structure. | |
632 | */ | |
633 | static void wake_call_rcu_thread(struct call_rcu_data *crdp) | |
634 | { | |
263e3cf9 MD |
635 | if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT)) |
636 | call_rcu_wake_up(crdp); | |
7106ddf8 PM |
637 | } |
638 | ||
b7f721d9 MD |
639 | static void _call_rcu(struct rcu_head *head, |
640 | void (*func)(struct rcu_head *head), | |
641 | struct call_rcu_data *crdp) | |
642 | { | |
643 | cds_wfcq_node_init(&head->next); | |
644 | head->func = func; | |
645 | cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next); | |
646 | uatomic_inc(&crdp->qlen); | |
647 | wake_call_rcu_thread(crdp); | |
648 | } | |
649 | ||
b57aee66 PM |
650 | /* |
651 | * Schedule a function to be invoked after a following grace period. | |
652 | * This is the only function that must be called -- the others are | |
653 | * only present to allow applications to tune their use of RCU for | |
654 | * maximum performance. | |
655 | * | |
656 | * Note that unless a call_rcu thread has not already been created, | |
657 | * the first invocation of call_rcu() will create one. So, if you | |
658 | * need the first invocation of call_rcu() to be fast, make sure | |
659 | * to create a call_rcu thread first. One way to accomplish this is | |
660 | * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data(). | |
618b2595 MD |
661 | * |
662 | * call_rcu must be called by registered RCU read-side threads. | |
b57aee66 | 663 | */ |
b57aee66 PM |
664 | void call_rcu(struct rcu_head *head, |
665 | void (*func)(struct rcu_head *head)) | |
666 | { | |
667 | struct call_rcu_data *crdp; | |
668 | ||
618b2595 MD |
669 | /* Holding rcu read-side lock across use of per-cpu crdp */ |
670 | rcu_read_lock(); | |
b57aee66 | 671 | crdp = get_call_rcu_data(); |
b7f721d9 | 672 | _call_rcu(head, func, crdp); |
618b2595 | 673 | rcu_read_unlock(); |
7106ddf8 PM |
674 | } |
675 | ||
676 | /* | |
677 | * Free up the specified call_rcu_data structure, terminating the | |
678 | * associated call_rcu thread. The caller must have previously | |
679 | * removed the call_rcu_data structure from per-thread or per-CPU | |
680 | * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU | |
681 | * call_rcu_data structures or set_thread_call_rcu_data(NULL) for | |
682 | * per-thread call_rcu_data structures. | |
683 | * | |
684 | * We silently refuse to free up the default call_rcu_data structure | |
685 | * because that is where we put any leftover callbacks. Note that | |
686 | * the possibility of self-spawning callbacks makes it impossible | |
687 | * to execute all the callbacks in finite time without putting any | |
688 | * newly spawned callbacks somewhere else. The "somewhere else" of | |
689 | * last resort is the default call_rcu_data structure. | |
690 | * | |
691 | * We also silently refuse to free NULL pointers. This simplifies | |
692 | * the calling code. | |
f9da0936 MD |
693 | * |
694 | * The caller must wait for a grace-period to pass between return from | |
695 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
696 | * previous call rcu data as argument. | |
03e5118f MD |
697 | * |
698 | * Note: introducing __cds_wfcq_splice_blocking() in this function fixed | |
699 | * a list corruption bug in the 0.7.x series. The equivalent fix | |
700 | * appeared in 0.6.8 for the stable-0.6 branch. | |
7106ddf8 PM |
701 | */ |
702 | void call_rcu_data_free(struct call_rcu_data *crdp) | |
703 | { | |
7106ddf8 PM |
704 | if (crdp == NULL || crdp == default_call_rcu_data) { |
705 | return; | |
706 | } | |
2870aa1e PB |
707 | if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) { |
708 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP); | |
7106ddf8 | 709 | wake_call_rcu_thread(crdp); |
2870aa1e | 710 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) |
7106ddf8 PM |
711 | poll(NULL, 0, 1); |
712 | } | |
5161f31e | 713 | if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) { |
698d0778 MD |
714 | /* Create default call rcu data if need be */ |
715 | (void) get_default_call_rcu_data(); | |
5161f31e MD |
716 | __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head, |
717 | &default_call_rcu_data->cbs_tail, | |
718 | &crdp->cbs_head, &crdp->cbs_tail); | |
7106ddf8 PM |
719 | uatomic_add(&default_call_rcu_data->qlen, |
720 | uatomic_read(&crdp->qlen)); | |
1e92aa15 | 721 | wake_call_rcu_thread(default_call_rcu_data); |
7106ddf8 | 722 | } |
59dc9e9d | 723 | |
b75dffe6 | 724 | call_rcu_lock(&call_rcu_mutex); |
59dc9e9d | 725 | cds_list_del(&crdp->list); |
b75dffe6 LJ |
726 | call_rcu_unlock(&call_rcu_mutex); |
727 | ||
59dc9e9d | 728 | free(crdp); |
7106ddf8 PM |
729 | } |
730 | ||
731 | /* | |
732 | * Clean up all the per-CPU call_rcu threads. | |
733 | */ | |
734 | void free_all_cpu_call_rcu_data(void) | |
735 | { | |
736 | int cpu; | |
618b2595 MD |
737 | struct call_rcu_data **crdp; |
738 | static int warned = 0; | |
7106ddf8 PM |
739 | |
740 | if (maxcpus <= 0) | |
741 | return; | |
618b2595 MD |
742 | |
743 | crdp = malloc(sizeof(*crdp) * maxcpus); | |
744 | if (!crdp) { | |
745 | if (!warned) { | |
746 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
747 | } | |
748 | warned = 1; | |
d31150b4 | 749 | return; |
618b2595 MD |
750 | } |
751 | ||
7106ddf8 | 752 | for (cpu = 0; cpu < maxcpus; cpu++) { |
618b2595 MD |
753 | crdp[cpu] = get_cpu_call_rcu_data(cpu); |
754 | if (crdp[cpu] == NULL) | |
7106ddf8 PM |
755 | continue; |
756 | set_cpu_call_rcu_data(cpu, NULL); | |
7106ddf8 | 757 | } |
618b2595 MD |
758 | /* |
759 | * Wait for call_rcu sites acting as RCU readers of the | |
760 | * call_rcu_data to become quiescent. | |
761 | */ | |
762 | synchronize_rcu(); | |
763 | for (cpu = 0; cpu < maxcpus; cpu++) { | |
764 | if (crdp[cpu] == NULL) | |
765 | continue; | |
766 | call_rcu_data_free(crdp[cpu]); | |
767 | } | |
768 | free(crdp); | |
7106ddf8 PM |
769 | } |
770 | ||
b7f721d9 MD |
771 | static |
772 | void _rcu_barrier_complete(struct rcu_head *head) | |
773 | { | |
774 | struct call_rcu_completion_work *work; | |
775 | struct call_rcu_completion *completion; | |
776 | ||
777 | work = caa_container_of(head, struct call_rcu_completion_work, head); | |
778 | completion = work->completion; | |
779 | uatomic_dec(&completion->barrier_count); | |
780 | call_rcu_completion_wake_up(completion); | |
781 | free(work); | |
782 | } | |
783 | ||
784 | /* | |
785 | * Wait for all in-flight call_rcu callbacks to complete execution. | |
786 | */ | |
787 | void rcu_barrier(void) | |
788 | { | |
789 | struct call_rcu_data *crdp; | |
790 | struct call_rcu_completion completion; | |
63f0fc6d | 791 | int count = 0; |
b7f721d9 MD |
792 | int was_online; |
793 | ||
794 | /* Put in offline state in QSBR. */ | |
795 | was_online = rcu_read_ongoing(); | |
796 | if (was_online) | |
797 | rcu_thread_offline(); | |
798 | /* | |
799 | * Calling a rcu_barrier() within a RCU read-side critical | |
800 | * section is an error. | |
801 | */ | |
802 | if (rcu_read_ongoing()) { | |
803 | static int warned = 0; | |
804 | ||
805 | if (!warned) { | |
806 | fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n"); | |
807 | } | |
808 | warned = 1; | |
809 | goto online; | |
810 | } | |
811 | ||
812 | call_rcu_lock(&call_rcu_mutex); | |
813 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) | |
814 | count++; | |
815 | ||
816 | completion.barrier_count = count; | |
817 | ||
818 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
819 | struct call_rcu_completion_work *work; | |
820 | ||
821 | work = calloc(sizeof(*work), 1); | |
63f0fc6d MD |
822 | if (!work) |
823 | urcu_die(errno); | |
b7f721d9 MD |
824 | work->completion = &completion; |
825 | _call_rcu(&work->head, _rcu_barrier_complete, crdp); | |
b7f721d9 MD |
826 | } |
827 | call_rcu_unlock(&call_rcu_mutex); | |
828 | ||
b7f721d9 MD |
829 | /* Wait for them */ |
830 | for (;;) { | |
831 | uatomic_dec(&completion.futex); | |
832 | /* Decrement futex before reading barrier_count */ | |
833 | cmm_smp_mb(); | |
834 | if (!uatomic_read(&completion.barrier_count)) | |
835 | break; | |
836 | call_rcu_completion_wait(&completion); | |
837 | } | |
838 | online: | |
839 | if (was_online) | |
840 | rcu_thread_online(); | |
841 | } | |
842 | ||
81ad2e19 PM |
843 | /* |
844 | * Acquire the call_rcu_mutex in order to ensure that the child sees | |
e85451a1 MD |
845 | * all of the call_rcu() data structures in a consistent state. Ensure |
846 | * that all call_rcu threads are in a quiescent state across fork. | |
81ad2e19 PM |
847 | * Suitable for pthread_atfork() and friends. |
848 | */ | |
849 | void call_rcu_before_fork(void) | |
850 | { | |
e85451a1 MD |
851 | struct call_rcu_data *crdp; |
852 | ||
81ad2e19 | 853 | call_rcu_lock(&call_rcu_mutex); |
e85451a1 MD |
854 | |
855 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
856 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE); | |
857 | cmm_smp_mb__after_uatomic_or(); | |
858 | wake_call_rcu_thread(crdp); | |
859 | } | |
860 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
861 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0) | |
862 | poll(NULL, 0, 1); | |
863 | } | |
81ad2e19 PM |
864 | } |
865 | ||
866 | /* | |
867 | * Clean up call_rcu data structures in the parent of a successful fork() | |
868 | * that is not followed by exec() in the child. Suitable for | |
869 | * pthread_atfork() and friends. | |
870 | */ | |
871 | void call_rcu_after_fork_parent(void) | |
872 | { | |
e85451a1 MD |
873 | struct call_rcu_data *crdp; |
874 | ||
875 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) | |
876 | uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE); | |
db2b0858 KF |
877 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { |
878 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0) | |
879 | poll(NULL, 0, 1); | |
880 | } | |
81ad2e19 PM |
881 | call_rcu_unlock(&call_rcu_mutex); |
882 | } | |
883 | ||
7106ddf8 PM |
884 | /* |
885 | * Clean up call_rcu data structures in the child of a successful fork() | |
81ad2e19 PM |
886 | * that is not followed by exec(). Suitable for pthread_atfork() and |
887 | * friends. | |
7106ddf8 PM |
888 | */ |
889 | void call_rcu_after_fork_child(void) | |
890 | { | |
077ff173 | 891 | struct call_rcu_data *crdp, *next; |
7106ddf8 | 892 | |
81ad2e19 PM |
893 | /* Release the mutex. */ |
894 | call_rcu_unlock(&call_rcu_mutex); | |
895 | ||
ad1b9909 LJ |
896 | /* Do nothing when call_rcu() has not been used */ |
897 | if (cds_list_empty(&call_rcu_data_list)) | |
898 | return; | |
899 | ||
7106ddf8 PM |
900 | /* |
901 | * Allocate a new default call_rcu_data structure in order | |
902 | * to get a working call_rcu thread to go with it. | |
903 | */ | |
904 | default_call_rcu_data = NULL; | |
905 | (void)get_default_call_rcu_data(); | |
906 | ||
60af049d LJ |
907 | /* Cleanup call_rcu_data pointers before use */ |
908 | maxcpus_reset(); | |
909 | free(per_cpu_call_rcu_data); | |
618b2595 | 910 | rcu_set_pointer(&per_cpu_call_rcu_data, NULL); |
bd252a04 | 911 | URCU_TLS(thread_call_rcu_data) = NULL; |
60af049d | 912 | |
e85451a1 MD |
913 | /* |
914 | * Dispose of all of the rest of the call_rcu_data structures. | |
915 | * Leftover call_rcu callbacks will be merged into the new | |
916 | * default call_rcu thread queue. | |
917 | */ | |
077ff173 | 918 | cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) { |
7106ddf8 | 919 | if (crdp == default_call_rcu_data) |
077ff173 | 920 | continue; |
2870aa1e | 921 | uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED); |
7106ddf8 | 922 | call_rcu_data_free(crdp); |
b57aee66 PM |
923 | } |
924 | } |