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 | ||
5161f31e | 23 | #define _LGPL_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 | 36 | |
a47dd11c | 37 | #include "compat-getcpu.h" |
4477a870 MD |
38 | #include <urcu/wfcqueue.h> |
39 | #include <urcu/call-rcu.h> | |
40 | #include <urcu/pointer.h> | |
41 | #include <urcu/list.h> | |
42 | #include <urcu/futex.h> | |
43 | #include <urcu/tls-compat.h> | |
44 | #include <urcu/ref.h> | |
4a6d7378 | 45 | #include "urcu-die.h" |
4477a870 | 46 | #include "urcu-utils.h" |
6b6204e4 | 47 | #include "compat-smp.h" |
b57aee66 | 48 | |
8e3690db MD |
49 | #define SET_AFFINITY_CHECK_PERIOD (1U << 8) /* 256 */ |
50 | #define SET_AFFINITY_CHECK_PERIOD_MASK (SET_AFFINITY_CHECK_PERIOD - 1) | |
51 | ||
b57aee66 PM |
52 | /* Data structure that identifies a call_rcu thread. */ |
53 | ||
54 | struct call_rcu_data { | |
5161f31e | 55 | /* |
0b8ab7df MD |
56 | * We do not align head on a different cache-line than tail |
57 | * mainly because call_rcu callback-invocation threads use | |
58 | * batching ("splice") to get an entire list of callbacks, which | |
59 | * effectively empties the queue, and requires to touch the tail | |
60 | * anyway. | |
5161f31e | 61 | */ |
b9f893b6 | 62 | struct cds_wfcq_tail cbs_tail; |
0b8ab7df | 63 | struct cds_wfcq_head cbs_head; |
b57aee66 | 64 | unsigned long flags; |
6d841bc2 | 65 | int32_t futex; |
73987721 | 66 | unsigned long qlen; /* maintained for debugging. */ |
b57aee66 | 67 | pthread_t tid; |
c1d2c60b | 68 | int cpu_affinity; |
8e3690db | 69 | unsigned long gp_count; |
3c24913f | 70 | struct cds_list_head list; |
b57aee66 PM |
71 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
72 | ||
b7f721d9 MD |
73 | struct call_rcu_completion { |
74 | int barrier_count; | |
75 | int32_t futex; | |
81dd9134 | 76 | struct urcu_ref ref; |
b7f721d9 MD |
77 | }; |
78 | ||
79 | struct call_rcu_completion_work { | |
80 | struct rcu_head head; | |
81 | struct call_rcu_completion *completion; | |
82 | }; | |
83 | ||
caefd1ee MD |
84 | enum crdf_flags { |
85 | CRDF_FLAG_JOIN_THREAD = (1 << 0), | |
86 | }; | |
87 | ||
3c24913f PM |
88 | /* |
89 | * List of all call_rcu_data structures to keep valgrind happy. | |
90 | * Protected by call_rcu_mutex. | |
91 | */ | |
92 | ||
bab44e28 | 93 | static CDS_LIST_HEAD(call_rcu_data_list); |
3c24913f | 94 | |
b57aee66 PM |
95 | /* Link a thread using call_rcu() to its call_rcu thread. */ |
96 | ||
2f661865 | 97 | static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data); |
b57aee66 | 98 | |
e85451a1 MD |
99 | /* |
100 | * Guard call_rcu thread creation and atfork handlers. | |
101 | */ | |
b57aee66 PM |
102 | static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER; |
103 | ||
104 | /* If a given thread does not have its own call_rcu thread, this is default. */ | |
105 | ||
106 | static struct call_rcu_data *default_call_rcu_data; | |
107 | ||
d0ec0ed2 | 108 | static struct urcu_atfork *registered_rculfhash_atfork; |
d0ec0ed2 | 109 | |
b57aee66 PM |
110 | /* |
111 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are | |
112 | * available, then we can have call_rcu threads assigned to individual | |
113 | * CPUs rather than only to specific threads. | |
114 | */ | |
115 | ||
a47dd11c | 116 | #if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID)) |
b57aee66 PM |
117 | |
118 | /* | |
119 | * Pointer to array of pointers to per-CPU call_rcu_data structures | |
618b2595 MD |
120 | * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an |
121 | * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a | |
122 | * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer | |
123 | * without mutex. The call_rcu_mutex protects updates. | |
b57aee66 PM |
124 | */ |
125 | ||
126 | static struct call_rcu_data **per_cpu_call_rcu_data; | |
6b6204e4 | 127 | static long cpus_array_len; |
b57aee66 | 128 | |
6b6204e4 | 129 | static void cpus_array_len_reset(void) |
60af049d | 130 | { |
6b6204e4 | 131 | cpus_array_len = 0; |
60af049d LJ |
132 | } |
133 | ||
b57aee66 PM |
134 | /* Allocate the array if it has not already been allocated. */ |
135 | ||
136 | static void alloc_cpu_call_rcu_data(void) | |
137 | { | |
138 | struct call_rcu_data **p; | |
139 | static int warned = 0; | |
140 | ||
6b6204e4 | 141 | if (cpus_array_len != 0) |
b57aee66 | 142 | return; |
6b6204e4 MJ |
143 | cpus_array_len = get_possible_cpus_array_len(); |
144 | if (cpus_array_len <= 0) { | |
b57aee66 PM |
145 | return; |
146 | } | |
6b6204e4 | 147 | p = malloc(cpus_array_len * sizeof(*per_cpu_call_rcu_data)); |
b57aee66 | 148 | if (p != NULL) { |
6b6204e4 | 149 | memset(p, '\0', cpus_array_len * sizeof(*per_cpu_call_rcu_data)); |
618b2595 | 150 | rcu_set_pointer(&per_cpu_call_rcu_data, p); |
b57aee66 PM |
151 | } else { |
152 | if (!warned) { | |
153 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
154 | } | |
155 | warned = 1; | |
156 | } | |
157 | } | |
158 | ||
63b495d8 | 159 | #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
b57aee66 | 160 | |
f9437098 MD |
161 | /* |
162 | * per_cpu_call_rcu_data should be constant, but some functions below, used both | |
163 | * for cases where cpu number is available and not available, assume it it not | |
164 | * constant. | |
165 | */ | |
166 | static struct call_rcu_data **per_cpu_call_rcu_data = NULL; | |
6b6204e4 | 167 | static const long cpus_array_len = -1; |
b57aee66 | 168 | |
6b6204e4 | 169 | static void cpus_array_len_reset(void) |
60af049d LJ |
170 | { |
171 | } | |
172 | ||
b57aee66 PM |
173 | static void alloc_cpu_call_rcu_data(void) |
174 | { | |
175 | } | |
176 | ||
63b495d8 | 177 | #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
b57aee66 PM |
178 | |
179 | /* Acquire the specified pthread mutex. */ | |
180 | ||
181 | static void call_rcu_lock(pthread_mutex_t *pmp) | |
182 | { | |
4a6d7378 MD |
183 | int ret; |
184 | ||
185 | ret = pthread_mutex_lock(pmp); | |
186 | if (ret) | |
187 | urcu_die(ret); | |
b57aee66 PM |
188 | } |
189 | ||
190 | /* Release the specified pthread mutex. */ | |
191 | ||
192 | static void call_rcu_unlock(pthread_mutex_t *pmp) | |
193 | { | |
4a6d7378 MD |
194 | int ret; |
195 | ||
196 | ret = pthread_mutex_unlock(pmp); | |
197 | if (ret) | |
198 | urcu_die(ret); | |
b57aee66 PM |
199 | } |
200 | ||
8e3690db MD |
201 | /* |
202 | * Periodically retry setting CPU affinity if we migrate. | |
203 | * Losing affinity can be caused by CPU hotunplug/hotplug, or by | |
204 | * cpuset(7). | |
205 | */ | |
2388c075 | 206 | #ifdef HAVE_SCHED_SETAFFINITY |
c1d2c60b MD |
207 | static |
208 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
209 | { | |
210 | cpu_set_t mask; | |
8e3690db | 211 | int ret; |
c1d2c60b MD |
212 | |
213 | if (crdp->cpu_affinity < 0) | |
214 | return 0; | |
8e3690db MD |
215 | if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK) |
216 | return 0; | |
217 | if (urcu_sched_getcpu() == crdp->cpu_affinity) | |
218 | return 0; | |
c1d2c60b MD |
219 | |
220 | CPU_ZERO(&mask); | |
221 | CPU_SET(crdp->cpu_affinity, &mask); | |
8e3690db | 222 | ret = sched_setaffinity(0, sizeof(mask), &mask); |
0614a2e6 | 223 | |
8e3690db MD |
224 | /* |
225 | * EINVAL is fine: can be caused by hotunplugged CPUs, or by | |
226 | * cpuset(7). This is why we should always retry if we detect | |
227 | * migration. | |
228 | */ | |
229 | if (ret && errno == EINVAL) { | |
230 | ret = 0; | |
231 | errno = 0; | |
232 | } | |
233 | return ret; | |
c1d2c60b MD |
234 | } |
235 | #else | |
236 | static | |
a142df4e | 237 | int set_thread_cpu_affinity(struct call_rcu_data *crdp __attribute__((unused))) |
c1d2c60b MD |
238 | { |
239 | return 0; | |
240 | } | |
241 | #endif | |
242 | ||
03fe58b3 MD |
243 | static void call_rcu_wait(struct call_rcu_data *crdp) |
244 | { | |
245 | /* Read call_rcu list before read futex */ | |
246 | cmm_smp_mb(); | |
4b783ba2 MD |
247 | while (uatomic_read(&crdp->futex) == -1) { |
248 | if (!futex_async(&crdp->futex, FUTEX_WAIT, -1, NULL, NULL, 0)) { | |
249 | /* | |
250 | * Prior queued wakeups queued by unrelated code | |
251 | * using the same address can cause futex wait to | |
252 | * return 0 even through the futex value is still | |
253 | * -1 (spurious wakeups). Check the value again | |
254 | * in user-space to validate whether it really | |
255 | * differs from -1. | |
256 | */ | |
257 | continue; | |
258 | } | |
b0a841b4 | 259 | switch (errno) { |
4b783ba2 | 260 | case EAGAIN: |
b0a841b4 MD |
261 | /* Value already changed. */ |
262 | return; | |
263 | case EINTR: | |
264 | /* Retry if interrupted by signal. */ | |
4b783ba2 | 265 | break; /* Get out of switch. Check again. */ |
b0a841b4 MD |
266 | default: |
267 | /* Unexpected error. */ | |
268 | urcu_die(errno); | |
269 | } | |
270 | } | |
03fe58b3 MD |
271 | } |
272 | ||
273 | static void call_rcu_wake_up(struct call_rcu_data *crdp) | |
274 | { | |
275 | /* Write to call_rcu list before reading/writing futex */ | |
276 | cmm_smp_mb(); | |
a0b7f7ea | 277 | if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) { |
03fe58b3 | 278 | uatomic_set(&crdp->futex, 0); |
b0a841b4 MD |
279 | if (futex_async(&crdp->futex, FUTEX_WAKE, 1, |
280 | NULL, NULL, 0) < 0) | |
281 | urcu_die(errno); | |
03fe58b3 MD |
282 | } |
283 | } | |
284 | ||
b7f721d9 MD |
285 | static void call_rcu_completion_wait(struct call_rcu_completion *completion) |
286 | { | |
287 | /* Read completion barrier count before read futex */ | |
288 | cmm_smp_mb(); | |
4b783ba2 MD |
289 | while (uatomic_read(&completion->futex) == -1) { |
290 | if (!futex_async(&completion->futex, FUTEX_WAIT, -1, NULL, NULL, 0)) { | |
291 | /* | |
292 | * Prior queued wakeups queued by unrelated code | |
293 | * using the same address can cause futex wait to | |
294 | * return 0 even through the futex value is still | |
295 | * -1 (spurious wakeups). Check the value again | |
296 | * in user-space to validate whether it really | |
297 | * differs from -1. | |
298 | */ | |
299 | continue; | |
300 | } | |
b0a841b4 | 301 | switch (errno) { |
4b783ba2 | 302 | case EAGAIN: |
b0a841b4 MD |
303 | /* Value already changed. */ |
304 | return; | |
305 | case EINTR: | |
306 | /* Retry if interrupted by signal. */ | |
4b783ba2 | 307 | break; /* Get out of switch. Check again. */ |
b0a841b4 MD |
308 | default: |
309 | /* Unexpected error. */ | |
310 | urcu_die(errno); | |
311 | } | |
312 | } | |
b7f721d9 MD |
313 | } |
314 | ||
315 | static void call_rcu_completion_wake_up(struct call_rcu_completion *completion) | |
316 | { | |
317 | /* Write to completion barrier count before reading/writing futex */ | |
318 | cmm_smp_mb(); | |
319 | if (caa_unlikely(uatomic_read(&completion->futex) == -1)) { | |
320 | uatomic_set(&completion->futex, 0); | |
b0a841b4 MD |
321 | if (futex_async(&completion->futex, FUTEX_WAKE, 1, |
322 | NULL, NULL, 0) < 0) | |
323 | urcu_die(errno); | |
b7f721d9 MD |
324 | } |
325 | } | |
326 | ||
b57aee66 PM |
327 | /* This is the code run by each call_rcu thread. */ |
328 | ||
329 | static void *call_rcu_thread(void *arg) | |
330 | { | |
331 | unsigned long cbcount; | |
5161f31e | 332 | struct call_rcu_data *crdp = (struct call_rcu_data *) arg; |
2870aa1e | 333 | int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT); |
b57aee66 | 334 | |
8e3690db | 335 | if (set_thread_cpu_affinity(crdp)) |
4a6d7378 | 336 | urcu_die(errno); |
c1d2c60b | 337 | |
765f3ead MD |
338 | /* |
339 | * If callbacks take a read-side lock, we need to be registered. | |
340 | */ | |
341 | rcu_register_thread(); | |
342 | ||
bd252a04 | 343 | URCU_TLS(thread_call_rcu_data) = crdp; |
bc94ca9b MD |
344 | if (!rt) { |
345 | uatomic_dec(&crdp->futex); | |
346 | /* Decrement futex before reading call_rcu list */ | |
347 | cmm_smp_mb(); | |
348 | } | |
b57aee66 | 349 | for (;;) { |
5161f31e MD |
350 | struct cds_wfcq_head cbs_tmp_head; |
351 | struct cds_wfcq_tail cbs_tmp_tail; | |
352 | struct cds_wfcq_node *cbs, *cbs_tmp_n; | |
ae25b7e2 | 353 | enum cds_wfcq_ret splice_ret; |
5161f31e | 354 | |
8e3690db MD |
355 | if (set_thread_cpu_affinity(crdp)) |
356 | urcu_die(errno); | |
357 | ||
e85451a1 MD |
358 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) { |
359 | /* | |
360 | * Pause requested. Become quiescent: remove | |
361 | * ourself from all global lists, and don't | |
362 | * process any callback. The callback lists may | |
363 | * still be non-empty though. | |
364 | */ | |
365 | rcu_unregister_thread(); | |
366 | cmm_smp_mb__before_uatomic_or(); | |
367 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED); | |
368 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0) | |
109105b7 | 369 | (void) poll(NULL, 0, 1); |
fc236e5e KF |
370 | uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED); |
371 | cmm_smp_mb__after_uatomic_and(); | |
e85451a1 MD |
372 | rcu_register_thread(); |
373 | } | |
374 | ||
5161f31e | 375 | cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail); |
ae25b7e2 MD |
376 | splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head, |
377 | &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail); | |
378 | assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK); | |
379 | assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY); | |
380 | if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) { | |
b57aee66 PM |
381 | synchronize_rcu(); |
382 | cbcount = 0; | |
5161f31e MD |
383 | __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head, |
384 | &cbs_tmp_tail, cbs, cbs_tmp_n) { | |
385 | struct rcu_head *rhp; | |
386 | ||
387 | rhp = caa_container_of(cbs, | |
388 | struct rcu_head, next); | |
b57aee66 PM |
389 | rhp->func(rhp); |
390 | cbcount++; | |
5161f31e | 391 | } |
b57aee66 PM |
392 | uatomic_sub(&crdp->qlen, cbcount); |
393 | } | |
bc94ca9b MD |
394 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP) |
395 | break; | |
765f3ead | 396 | rcu_thread_offline(); |
bc94ca9b | 397 | if (!rt) { |
5161f31e MD |
398 | if (cds_wfcq_empty(&crdp->cbs_head, |
399 | &crdp->cbs_tail)) { | |
bc94ca9b | 400 | call_rcu_wait(crdp); |
109105b7 | 401 | (void) poll(NULL, 0, 10); |
bc94ca9b | 402 | uatomic_dec(&crdp->futex); |
c768e45e | 403 | /* |
bc94ca9b MD |
404 | * Decrement futex before reading |
405 | * call_rcu list. | |
c768e45e MD |
406 | */ |
407 | cmm_smp_mb(); | |
ccbac24d | 408 | } else { |
109105b7 | 409 | (void) poll(NULL, 0, 10); |
c768e45e | 410 | } |
bc94ca9b | 411 | } else { |
109105b7 | 412 | (void) poll(NULL, 0, 10); |
b57aee66 | 413 | } |
765f3ead | 414 | rcu_thread_online(); |
bc94ca9b MD |
415 | } |
416 | if (!rt) { | |
417 | /* | |
418 | * Read call_rcu list before write futex. | |
419 | */ | |
420 | cmm_smp_mb(); | |
421 | uatomic_set(&crdp->futex, 0); | |
b57aee66 | 422 | } |
2870aa1e | 423 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED); |
765f3ead | 424 | rcu_unregister_thread(); |
7106ddf8 | 425 | return NULL; |
b57aee66 PM |
426 | } |
427 | ||
428 | /* | |
429 | * Create both a call_rcu thread and the corresponding call_rcu_data | |
3c24913f PM |
430 | * structure, linking the structure in as specified. Caller must hold |
431 | * call_rcu_mutex. | |
b57aee66 PM |
432 | */ |
433 | ||
3c24913f | 434 | static void call_rcu_data_init(struct call_rcu_data **crdpp, |
c1d2c60b MD |
435 | unsigned long flags, |
436 | int cpu_affinity) | |
b57aee66 PM |
437 | { |
438 | struct call_rcu_data *crdp; | |
4a6d7378 | 439 | int ret; |
b57aee66 PM |
440 | |
441 | crdp = malloc(sizeof(*crdp)); | |
4a6d7378 MD |
442 | if (crdp == NULL) |
443 | urcu_die(errno); | |
b57aee66 | 444 | memset(crdp, '\0', sizeof(*crdp)); |
5161f31e | 445 | cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail); |
b57aee66 | 446 | crdp->qlen = 0; |
263e3cf9 MD |
447 | crdp->futex = 0; |
448 | crdp->flags = flags; | |
3c24913f | 449 | cds_list_add(&crdp->list, &call_rcu_data_list); |
c1d2c60b | 450 | crdp->cpu_affinity = cpu_affinity; |
8e3690db | 451 | crdp->gp_count = 0; |
c2f6c243 MD |
452 | rcu_set_pointer(crdpp, crdp); |
453 | ||
4a6d7378 MD |
454 | ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp); |
455 | if (ret) | |
456 | urcu_die(ret); | |
b57aee66 PM |
457 | } |
458 | ||
459 | /* | |
460 | * Return a pointer to the call_rcu_data structure for the specified | |
461 | * CPU, returning NULL if there is none. We cannot automatically | |
462 | * created it because the platform we are running on might not define | |
63b495d8 | 463 | * urcu_sched_getcpu(). |
618b2595 MD |
464 | * |
465 | * The call to this function and use of the returned call_rcu_data | |
466 | * should be protected by RCU read-side lock. | |
b57aee66 PM |
467 | */ |
468 | ||
469 | struct call_rcu_data *get_cpu_call_rcu_data(int cpu) | |
470 | { | |
471 | static int warned = 0; | |
618b2595 | 472 | struct call_rcu_data **pcpu_crdp; |
b57aee66 | 473 | |
618b2595 MD |
474 | pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data); |
475 | if (pcpu_crdp == NULL) | |
b57aee66 | 476 | return NULL; |
6b6204e4 | 477 | if (!warned && cpus_array_len > 0 && (cpu < 0 || cpus_array_len <= cpu)) { |
b57aee66 PM |
478 | fprintf(stderr, "[error] liburcu: get CPU # out of range\n"); |
479 | warned = 1; | |
480 | } | |
6b6204e4 | 481 | if (cpu < 0 || cpus_array_len <= cpu) |
b57aee66 | 482 | return NULL; |
618b2595 | 483 | return rcu_dereference(pcpu_crdp[cpu]); |
b57aee66 PM |
484 | } |
485 | ||
486 | /* | |
487 | * Return the tid corresponding to the call_rcu thread whose | |
488 | * call_rcu_data structure is specified. | |
489 | */ | |
490 | ||
491 | pthread_t get_call_rcu_thread(struct call_rcu_data *crdp) | |
492 | { | |
493 | return crdp->tid; | |
494 | } | |
495 | ||
496 | /* | |
497 | * Create a call_rcu_data structure (with thread) and return a pointer. | |
498 | */ | |
499 | ||
c1d2c60b MD |
500 | static struct call_rcu_data *__create_call_rcu_data(unsigned long flags, |
501 | int cpu_affinity) | |
b57aee66 PM |
502 | { |
503 | struct call_rcu_data *crdp; | |
504 | ||
c1d2c60b | 505 | call_rcu_data_init(&crdp, flags, cpu_affinity); |
b57aee66 PM |
506 | return crdp; |
507 | } | |
508 | ||
c1d2c60b MD |
509 | struct call_rcu_data *create_call_rcu_data(unsigned long flags, |
510 | int cpu_affinity) | |
3c24913f PM |
511 | { |
512 | struct call_rcu_data *crdp; | |
513 | ||
514 | call_rcu_lock(&call_rcu_mutex); | |
c1d2c60b | 515 | crdp = __create_call_rcu_data(flags, cpu_affinity); |
3c24913f PM |
516 | call_rcu_unlock(&call_rcu_mutex); |
517 | return crdp; | |
518 | } | |
519 | ||
b57aee66 PM |
520 | /* |
521 | * Set the specified CPU to use the specified call_rcu_data structure. | |
7106ddf8 PM |
522 | * |
523 | * Use NULL to remove a CPU's call_rcu_data structure, but it is | |
524 | * the caller's responsibility to dispose of the removed structure. | |
525 | * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure | |
526 | * (prior to NULLing it out, of course). | |
f9da0936 MD |
527 | * |
528 | * The caller must wait for a grace-period to pass between return from | |
529 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
530 | * previous call rcu data as argument. | |
b57aee66 PM |
531 | */ |
532 | ||
533 | int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp) | |
534 | { | |
dcfc8165 | 535 | static int warned = 0; |
b57aee66 PM |
536 | |
537 | call_rcu_lock(&call_rcu_mutex); | |
f3776786 | 538 | alloc_cpu_call_rcu_data(); |
6b6204e4 | 539 | if (cpu < 0 || cpus_array_len <= cpu) { |
b57aee66 PM |
540 | if (!warned) { |
541 | fprintf(stderr, "[error] liburcu: set CPU # out of range\n"); | |
542 | warned = 1; | |
543 | } | |
544 | call_rcu_unlock(&call_rcu_mutex); | |
545 | errno = EINVAL; | |
546 | return -EINVAL; | |
547 | } | |
53a55535 | 548 | |
b57aee66 | 549 | if (per_cpu_call_rcu_data == NULL) { |
3670fef2 | 550 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
551 | errno = ENOMEM; |
552 | return -ENOMEM; | |
553 | } | |
53a55535 LJ |
554 | |
555 | if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) { | |
556 | call_rcu_unlock(&call_rcu_mutex); | |
557 | errno = EEXIST; | |
558 | return -EEXIST; | |
559 | } | |
560 | ||
618b2595 | 561 | rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp); |
3670fef2 | 562 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
563 | return 0; |
564 | } | |
565 | ||
566 | /* | |
567 | * Return a pointer to the default call_rcu_data structure, creating | |
c2f6c243 MD |
568 | * one if need be. |
569 | * | |
570 | * The call to this function with intent to use the returned | |
571 | * call_rcu_data should be protected by RCU read-side lock. | |
b57aee66 PM |
572 | */ |
573 | ||
574 | struct call_rcu_data *get_default_call_rcu_data(void) | |
575 | { | |
c2f6c243 MD |
576 | struct call_rcu_data *crdp; |
577 | ||
578 | crdp = rcu_dereference(default_call_rcu_data); | |
579 | if (crdp != NULL) | |
580 | return crdp; | |
581 | ||
b57aee66 | 582 | call_rcu_lock(&call_rcu_mutex); |
c2f6c243 MD |
583 | if (default_call_rcu_data == NULL) |
584 | call_rcu_data_init(&default_call_rcu_data, 0, -1); | |
585 | crdp = default_call_rcu_data; | |
b57aee66 | 586 | call_rcu_unlock(&call_rcu_mutex); |
c2f6c243 MD |
587 | |
588 | return crdp; | |
b57aee66 PM |
589 | } |
590 | ||
591 | /* | |
592 | * Return the call_rcu_data structure that applies to the currently | |
593 | * running thread. Any call_rcu_data structure assigned specifically | |
594 | * to this thread has first priority, followed by any call_rcu_data | |
595 | * structure assigned to the CPU on which the thread is running, | |
596 | * followed by the default call_rcu_data structure. If there is not | |
597 | * yet a default call_rcu_data structure, one will be created. | |
618b2595 MD |
598 | * |
599 | * Calls to this function and use of the returned call_rcu_data should | |
600 | * be protected by RCU read-side lock. | |
b57aee66 PM |
601 | */ |
602 | struct call_rcu_data *get_call_rcu_data(void) | |
603 | { | |
9744f3bb | 604 | struct call_rcu_data *crd; |
b57aee66 | 605 | |
bd252a04 MD |
606 | if (URCU_TLS(thread_call_rcu_data) != NULL) |
607 | return URCU_TLS(thread_call_rcu_data); | |
9744f3bb | 608 | |
6b6204e4 | 609 | if (cpus_array_len > 0) { |
63b495d8 | 610 | crd = get_cpu_call_rcu_data(urcu_sched_getcpu()); |
9744f3bb LJ |
611 | if (crd) |
612 | return crd; | |
b57aee66 | 613 | } |
9744f3bb | 614 | |
b57aee66 PM |
615 | return get_default_call_rcu_data(); |
616 | } | |
617 | ||
618 | /* | |
619 | * Return a pointer to this task's call_rcu_data if there is one. | |
620 | */ | |
621 | ||
622 | struct call_rcu_data *get_thread_call_rcu_data(void) | |
623 | { | |
bd252a04 | 624 | return URCU_TLS(thread_call_rcu_data); |
b57aee66 PM |
625 | } |
626 | ||
627 | /* | |
628 | * Set this task's call_rcu_data structure as specified, regardless | |
629 | * of whether or not this task already had one. (This allows switching | |
630 | * to and from real-time call_rcu threads, for example.) | |
7106ddf8 PM |
631 | * |
632 | * Use NULL to remove a thread's call_rcu_data structure, but it is | |
633 | * the caller's responsibility to dispose of the removed structure. | |
634 | * Use get_thread_call_rcu_data() to obtain a pointer to the old structure | |
635 | * (prior to NULLing it out, of course). | |
b57aee66 PM |
636 | */ |
637 | ||
638 | void set_thread_call_rcu_data(struct call_rcu_data *crdp) | |
639 | { | |
bd252a04 | 640 | URCU_TLS(thread_call_rcu_data) = crdp; |
b57aee66 PM |
641 | } |
642 | ||
643 | /* | |
644 | * Create a separate call_rcu thread for each CPU. This does not | |
645 | * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data() | |
0938c541 MD |
646 | * function if you want that behavior. Should be paired with |
647 | * free_all_cpu_call_rcu_data() to teardown these call_rcu worker | |
648 | * threads. | |
b57aee66 PM |
649 | */ |
650 | ||
651 | int create_all_cpu_call_rcu_data(unsigned long flags) | |
652 | { | |
653 | int i; | |
654 | struct call_rcu_data *crdp; | |
655 | int ret; | |
656 | ||
657 | call_rcu_lock(&call_rcu_mutex); | |
658 | alloc_cpu_call_rcu_data(); | |
659 | call_rcu_unlock(&call_rcu_mutex); | |
6b6204e4 | 660 | if (cpus_array_len <= 0) { |
b57aee66 PM |
661 | errno = EINVAL; |
662 | return -EINVAL; | |
663 | } | |
664 | if (per_cpu_call_rcu_data == NULL) { | |
665 | errno = ENOMEM; | |
666 | return -ENOMEM; | |
667 | } | |
6b6204e4 | 668 | for (i = 0; i < cpus_array_len; i++) { |
b57aee66 PM |
669 | call_rcu_lock(&call_rcu_mutex); |
670 | if (get_cpu_call_rcu_data(i)) { | |
671 | call_rcu_unlock(&call_rcu_mutex); | |
672 | continue; | |
673 | } | |
c1d2c60b | 674 | crdp = __create_call_rcu_data(flags, i); |
b57aee66 PM |
675 | if (crdp == NULL) { |
676 | call_rcu_unlock(&call_rcu_mutex); | |
677 | errno = ENOMEM; | |
678 | return -ENOMEM; | |
679 | } | |
680 | call_rcu_unlock(&call_rcu_mutex); | |
681 | if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) { | |
356c8794 LJ |
682 | call_rcu_data_free(crdp); |
683 | ||
684 | /* it has been created by other thread */ | |
685 | if (ret == -EEXIST) | |
686 | continue; | |
687 | ||
688 | return ret; | |
b57aee66 PM |
689 | } |
690 | } | |
691 | return 0; | |
692 | } | |
693 | ||
7106ddf8 PM |
694 | /* |
695 | * Wake up the call_rcu thread corresponding to the specified | |
696 | * call_rcu_data structure. | |
697 | */ | |
698 | static void wake_call_rcu_thread(struct call_rcu_data *crdp) | |
699 | { | |
263e3cf9 MD |
700 | if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT)) |
701 | call_rcu_wake_up(crdp); | |
7106ddf8 PM |
702 | } |
703 | ||
b7f721d9 MD |
704 | static void _call_rcu(struct rcu_head *head, |
705 | void (*func)(struct rcu_head *head), | |
706 | struct call_rcu_data *crdp) | |
707 | { | |
708 | cds_wfcq_node_init(&head->next); | |
709 | head->func = func; | |
710 | cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next); | |
711 | uatomic_inc(&crdp->qlen); | |
712 | wake_call_rcu_thread(crdp); | |
713 | } | |
714 | ||
b57aee66 PM |
715 | /* |
716 | * Schedule a function to be invoked after a following grace period. | |
717 | * This is the only function that must be called -- the others are | |
718 | * only present to allow applications to tune their use of RCU for | |
719 | * maximum performance. | |
720 | * | |
721 | * Note that unless a call_rcu thread has not already been created, | |
722 | * the first invocation of call_rcu() will create one. So, if you | |
723 | * need the first invocation of call_rcu() to be fast, make sure | |
724 | * to create a call_rcu thread first. One way to accomplish this is | |
725 | * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data(). | |
618b2595 MD |
726 | * |
727 | * call_rcu must be called by registered RCU read-side threads. | |
b57aee66 | 728 | */ |
b57aee66 PM |
729 | void call_rcu(struct rcu_head *head, |
730 | void (*func)(struct rcu_head *head)) | |
731 | { | |
732 | struct call_rcu_data *crdp; | |
733 | ||
618b2595 | 734 | /* Holding rcu read-side lock across use of per-cpu crdp */ |
ffb4aaa7 | 735 | _rcu_read_lock(); |
b57aee66 | 736 | crdp = get_call_rcu_data(); |
b7f721d9 | 737 | _call_rcu(head, func, crdp); |
ffb4aaa7 | 738 | _rcu_read_unlock(); |
7106ddf8 PM |
739 | } |
740 | ||
741 | /* | |
742 | * Free up the specified call_rcu_data structure, terminating the | |
743 | * associated call_rcu thread. The caller must have previously | |
744 | * removed the call_rcu_data structure from per-thread or per-CPU | |
745 | * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU | |
746 | * call_rcu_data structures or set_thread_call_rcu_data(NULL) for | |
747 | * per-thread call_rcu_data structures. | |
748 | * | |
749 | * We silently refuse to free up the default call_rcu_data structure | |
750 | * because that is where we put any leftover callbacks. Note that | |
751 | * the possibility of self-spawning callbacks makes it impossible | |
752 | * to execute all the callbacks in finite time without putting any | |
753 | * newly spawned callbacks somewhere else. The "somewhere else" of | |
754 | * last resort is the default call_rcu_data structure. | |
755 | * | |
756 | * We also silently refuse to free NULL pointers. This simplifies | |
757 | * the calling code. | |
f9da0936 MD |
758 | * |
759 | * The caller must wait for a grace-period to pass between return from | |
760 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
761 | * previous call rcu data as argument. | |
03e5118f MD |
762 | * |
763 | * Note: introducing __cds_wfcq_splice_blocking() in this function fixed | |
764 | * a list corruption bug in the 0.7.x series. The equivalent fix | |
765 | * appeared in 0.6.8 for the stable-0.6 branch. | |
7106ddf8 | 766 | */ |
caefd1ee MD |
767 | static |
768 | void _call_rcu_data_free(struct call_rcu_data *crdp, unsigned int flags) | |
7106ddf8 | 769 | { |
7106ddf8 PM |
770 | if (crdp == NULL || crdp == default_call_rcu_data) { |
771 | return; | |
772 | } | |
2870aa1e PB |
773 | if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) { |
774 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP); | |
7106ddf8 | 775 | wake_call_rcu_thread(crdp); |
2870aa1e | 776 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) |
109105b7 | 777 | (void) poll(NULL, 0, 1); |
7106ddf8 | 778 | } |
51b98c3c | 779 | call_rcu_lock(&call_rcu_mutex); |
5161f31e | 780 | if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) { |
51b98c3c PM |
781 | call_rcu_unlock(&call_rcu_mutex); |
782 | /* Create default call rcu data if need be. */ | |
783 | /* CBs queued here will be handed to the default list. */ | |
698d0778 | 784 | (void) get_default_call_rcu_data(); |
51b98c3c | 785 | call_rcu_lock(&call_rcu_mutex); |
5161f31e MD |
786 | __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head, |
787 | &default_call_rcu_data->cbs_tail, | |
788 | &crdp->cbs_head, &crdp->cbs_tail); | |
7106ddf8 PM |
789 | uatomic_add(&default_call_rcu_data->qlen, |
790 | uatomic_read(&crdp->qlen)); | |
1e92aa15 | 791 | wake_call_rcu_thread(default_call_rcu_data); |
7106ddf8 | 792 | } |
59dc9e9d LJ |
793 | |
794 | cds_list_del(&crdp->list); | |
b75dffe6 LJ |
795 | call_rcu_unlock(&call_rcu_mutex); |
796 | ||
caefd1ee MD |
797 | if (flags & CRDF_FLAG_JOIN_THREAD) { |
798 | int ret; | |
799 | ||
800 | ret = pthread_join(get_call_rcu_thread(crdp), NULL); | |
801 | if (ret) | |
802 | urcu_die(ret); | |
803 | } | |
59dc9e9d | 804 | free(crdp); |
7106ddf8 PM |
805 | } |
806 | ||
caefd1ee MD |
807 | void call_rcu_data_free(struct call_rcu_data *crdp) |
808 | { | |
809 | _call_rcu_data_free(crdp, CRDF_FLAG_JOIN_THREAD); | |
810 | } | |
811 | ||
7106ddf8 PM |
812 | /* |
813 | * Clean up all the per-CPU call_rcu threads. | |
814 | */ | |
815 | void free_all_cpu_call_rcu_data(void) | |
816 | { | |
817 | int cpu; | |
618b2595 MD |
818 | struct call_rcu_data **crdp; |
819 | static int warned = 0; | |
7106ddf8 | 820 | |
6b6204e4 | 821 | if (cpus_array_len <= 0) |
7106ddf8 | 822 | return; |
618b2595 | 823 | |
6b6204e4 | 824 | crdp = malloc(sizeof(*crdp) * cpus_array_len); |
618b2595 MD |
825 | if (!crdp) { |
826 | if (!warned) { | |
827 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
828 | } | |
829 | warned = 1; | |
d31150b4 | 830 | return; |
618b2595 MD |
831 | } |
832 | ||
6b6204e4 | 833 | for (cpu = 0; cpu < cpus_array_len; cpu++) { |
618b2595 MD |
834 | crdp[cpu] = get_cpu_call_rcu_data(cpu); |
835 | if (crdp[cpu] == NULL) | |
7106ddf8 PM |
836 | continue; |
837 | set_cpu_call_rcu_data(cpu, NULL); | |
7106ddf8 | 838 | } |
618b2595 MD |
839 | /* |
840 | * Wait for call_rcu sites acting as RCU readers of the | |
841 | * call_rcu_data to become quiescent. | |
842 | */ | |
843 | synchronize_rcu(); | |
6b6204e4 | 844 | for (cpu = 0; cpu < cpus_array_len; cpu++) { |
618b2595 MD |
845 | if (crdp[cpu] == NULL) |
846 | continue; | |
847 | call_rcu_data_free(crdp[cpu]); | |
848 | } | |
849 | free(crdp); | |
7106ddf8 PM |
850 | } |
851 | ||
81dd9134 KF |
852 | static |
853 | void free_completion(struct urcu_ref *ref) | |
854 | { | |
855 | struct call_rcu_completion *completion; | |
856 | ||
857 | completion = caa_container_of(ref, struct call_rcu_completion, ref); | |
858 | free(completion); | |
859 | } | |
860 | ||
b7f721d9 MD |
861 | static |
862 | void _rcu_barrier_complete(struct rcu_head *head) | |
863 | { | |
864 | struct call_rcu_completion_work *work; | |
865 | struct call_rcu_completion *completion; | |
866 | ||
867 | work = caa_container_of(head, struct call_rcu_completion_work, head); | |
868 | completion = work->completion; | |
81dd9134 KF |
869 | if (!uatomic_sub_return(&completion->barrier_count, 1)) |
870 | call_rcu_completion_wake_up(completion); | |
871 | urcu_ref_put(&completion->ref, free_completion); | |
b7f721d9 MD |
872 | free(work); |
873 | } | |
874 | ||
875 | /* | |
876 | * Wait for all in-flight call_rcu callbacks to complete execution. | |
877 | */ | |
878 | void rcu_barrier(void) | |
879 | { | |
880 | struct call_rcu_data *crdp; | |
81dd9134 | 881 | struct call_rcu_completion *completion; |
63f0fc6d | 882 | int count = 0; |
b7f721d9 MD |
883 | int was_online; |
884 | ||
885 | /* Put in offline state in QSBR. */ | |
ffb4aaa7 | 886 | was_online = _rcu_read_ongoing(); |
b7f721d9 MD |
887 | if (was_online) |
888 | rcu_thread_offline(); | |
889 | /* | |
890 | * Calling a rcu_barrier() within a RCU read-side critical | |
891 | * section is an error. | |
892 | */ | |
ffb4aaa7 | 893 | if (_rcu_read_ongoing()) { |
b7f721d9 MD |
894 | static int warned = 0; |
895 | ||
896 | if (!warned) { | |
897 | fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n"); | |
898 | } | |
899 | warned = 1; | |
900 | goto online; | |
901 | } | |
902 | ||
81dd9134 KF |
903 | completion = calloc(sizeof(*completion), 1); |
904 | if (!completion) | |
905 | urcu_die(errno); | |
906 | ||
b7f721d9 MD |
907 | call_rcu_lock(&call_rcu_mutex); |
908 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) | |
909 | count++; | |
910 | ||
81dd9134 KF |
911 | /* Referenced by rcu_barrier() and each call_rcu thread. */ |
912 | urcu_ref_set(&completion->ref, count + 1); | |
913 | completion->barrier_count = count; | |
b7f721d9 MD |
914 | |
915 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
916 | struct call_rcu_completion_work *work; | |
917 | ||
918 | work = calloc(sizeof(*work), 1); | |
63f0fc6d MD |
919 | if (!work) |
920 | urcu_die(errno); | |
81dd9134 | 921 | work->completion = completion; |
b7f721d9 | 922 | _call_rcu(&work->head, _rcu_barrier_complete, crdp); |
b7f721d9 MD |
923 | } |
924 | call_rcu_unlock(&call_rcu_mutex); | |
925 | ||
b7f721d9 MD |
926 | /* Wait for them */ |
927 | for (;;) { | |
81dd9134 | 928 | uatomic_dec(&completion->futex); |
b7f721d9 MD |
929 | /* Decrement futex before reading barrier_count */ |
930 | cmm_smp_mb(); | |
81dd9134 | 931 | if (!uatomic_read(&completion->barrier_count)) |
b7f721d9 | 932 | break; |
81dd9134 | 933 | call_rcu_completion_wait(completion); |
b7f721d9 | 934 | } |
81dd9134 KF |
935 | |
936 | urcu_ref_put(&completion->ref, free_completion); | |
937 | ||
b7f721d9 MD |
938 | online: |
939 | if (was_online) | |
940 | rcu_thread_online(); | |
941 | } | |
942 | ||
81ad2e19 PM |
943 | /* |
944 | * Acquire the call_rcu_mutex in order to ensure that the child sees | |
e85451a1 MD |
945 | * all of the call_rcu() data structures in a consistent state. Ensure |
946 | * that all call_rcu threads are in a quiescent state across fork. | |
81ad2e19 PM |
947 | * Suitable for pthread_atfork() and friends. |
948 | */ | |
949 | void call_rcu_before_fork(void) | |
950 | { | |
e85451a1 | 951 | struct call_rcu_data *crdp; |
d0ec0ed2 | 952 | struct urcu_atfork *atfork; |
e85451a1 | 953 | |
81ad2e19 | 954 | call_rcu_lock(&call_rcu_mutex); |
e85451a1 | 955 | |
d0ec0ed2 MD |
956 | atfork = registered_rculfhash_atfork; |
957 | if (atfork) | |
958 | atfork->before_fork(atfork->priv); | |
959 | ||
e85451a1 MD |
960 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { |
961 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE); | |
962 | cmm_smp_mb__after_uatomic_or(); | |
963 | wake_call_rcu_thread(crdp); | |
964 | } | |
965 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
966 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0) | |
109105b7 | 967 | (void) poll(NULL, 0, 1); |
e85451a1 | 968 | } |
81ad2e19 PM |
969 | } |
970 | ||
971 | /* | |
972 | * Clean up call_rcu data structures in the parent of a successful fork() | |
973 | * that is not followed by exec() in the child. Suitable for | |
974 | * pthread_atfork() and friends. | |
975 | */ | |
976 | void call_rcu_after_fork_parent(void) | |
977 | { | |
e85451a1 | 978 | struct call_rcu_data *crdp; |
d0ec0ed2 | 979 | struct urcu_atfork *atfork; |
e85451a1 MD |
980 | |
981 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) | |
982 | uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE); | |
fc236e5e KF |
983 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { |
984 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0) | |
109105b7 | 985 | (void) poll(NULL, 0, 1); |
fc236e5e | 986 | } |
d0ec0ed2 MD |
987 | atfork = registered_rculfhash_atfork; |
988 | if (atfork) | |
989 | atfork->after_fork_parent(atfork->priv); | |
81ad2e19 PM |
990 | call_rcu_unlock(&call_rcu_mutex); |
991 | } | |
992 | ||
7106ddf8 PM |
993 | /* |
994 | * Clean up call_rcu data structures in the child of a successful fork() | |
81ad2e19 PM |
995 | * that is not followed by exec(). Suitable for pthread_atfork() and |
996 | * friends. | |
7106ddf8 PM |
997 | */ |
998 | void call_rcu_after_fork_child(void) | |
999 | { | |
077ff173 | 1000 | struct call_rcu_data *crdp, *next; |
d0ec0ed2 | 1001 | struct urcu_atfork *atfork; |
7106ddf8 | 1002 | |
81ad2e19 PM |
1003 | /* Release the mutex. */ |
1004 | call_rcu_unlock(&call_rcu_mutex); | |
1005 | ||
d0ec0ed2 MD |
1006 | atfork = registered_rculfhash_atfork; |
1007 | if (atfork) | |
1008 | atfork->after_fork_child(atfork->priv); | |
1009 | ||
ad1b9909 LJ |
1010 | /* Do nothing when call_rcu() has not been used */ |
1011 | if (cds_list_empty(&call_rcu_data_list)) | |
1012 | return; | |
1013 | ||
7106ddf8 PM |
1014 | /* |
1015 | * Allocate a new default call_rcu_data structure in order | |
1016 | * to get a working call_rcu thread to go with it. | |
1017 | */ | |
1018 | default_call_rcu_data = NULL; | |
1019 | (void)get_default_call_rcu_data(); | |
1020 | ||
60af049d | 1021 | /* Cleanup call_rcu_data pointers before use */ |
6b6204e4 | 1022 | cpus_array_len_reset(); |
60af049d | 1023 | free(per_cpu_call_rcu_data); |
618b2595 | 1024 | rcu_set_pointer(&per_cpu_call_rcu_data, NULL); |
bd252a04 | 1025 | URCU_TLS(thread_call_rcu_data) = NULL; |
60af049d | 1026 | |
e85451a1 MD |
1027 | /* |
1028 | * Dispose of all of the rest of the call_rcu_data structures. | |
1029 | * Leftover call_rcu callbacks will be merged into the new | |
1030 | * default call_rcu thread queue. | |
1031 | */ | |
077ff173 | 1032 | cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) { |
7106ddf8 | 1033 | if (crdp == default_call_rcu_data) |
077ff173 | 1034 | continue; |
2870aa1e | 1035 | uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED); |
caefd1ee MD |
1036 | /* |
1037 | * Do not join the thread because it does not exist in | |
1038 | * the child. | |
1039 | */ | |
1040 | _call_rcu_data_free(crdp, 0); | |
b57aee66 PM |
1041 | } |
1042 | } | |
d0ec0ed2 MD |
1043 | |
1044 | void urcu_register_rculfhash_atfork(struct urcu_atfork *atfork) | |
1045 | { | |
9f75d2d9 MD |
1046 | if (CMM_LOAD_SHARED(registered_rculfhash_atfork)) |
1047 | return; | |
d0ec0ed2 | 1048 | call_rcu_lock(&call_rcu_mutex); |
9f75d2d9 MD |
1049 | if (!registered_rculfhash_atfork) |
1050 | registered_rculfhash_atfork = atfork; | |
d0ec0ed2 MD |
1051 | call_rcu_unlock(&call_rcu_mutex); |
1052 | } | |
1053 | ||
9f75d2d9 MD |
1054 | /* |
1055 | * This unregistration function is deprecated, meant only for internal | |
1056 | * use by rculfhash. | |
1057 | */ | |
1058 | __attribute__((noreturn)) | |
70469b43 | 1059 | void urcu_unregister_rculfhash_atfork(struct urcu_atfork *atfork __attribute__((unused))) |
d0ec0ed2 | 1060 | { |
9f75d2d9 | 1061 | urcu_die(EPERM); |
d0ec0ed2 | 1062 | } |
c2f6c243 MD |
1063 | |
1064 | /* | |
1065 | * Teardown the default call_rcu worker thread if there are no queued | |
1066 | * callbacks on process exit. This prevents leaking memory. | |
1067 | * | |
1068 | * Here is how an application can ensure graceful teardown of this | |
1069 | * worker thread: | |
1070 | * | |
1071 | * - An application queuing call_rcu callbacks should invoke | |
1072 | * rcu_barrier() before it exits. | |
1073 | * - When chaining call_rcu callbacks, the number of calls to | |
1074 | * rcu_barrier() on application exit must match at least the maximum | |
1075 | * number of chained callbacks. | |
1076 | * - If an application chains callbacks endlessly, it would have to be | |
1077 | * modified to stop chaining callbacks when it detects an application | |
1078 | * exit (e.g. with a flag), and wait for quiescence with rcu_barrier() | |
1079 | * after setting that flag. | |
1080 | * - The statements above apply to a library which queues call_rcu | |
1081 | * callbacks, only it needs to invoke rcu_barrier in its library | |
1082 | * destructor. | |
1083 | * | |
1084 | * Note that this function does not presume it is being called when the | |
1085 | * application is single-threaded even though this is invoked from a | |
1086 | * destructor: this function synchronizes against concurrent calls to | |
1087 | * get_default_call_rcu_data(). | |
1088 | */ | |
1089 | static void urcu_call_rcu_exit(void) | |
1090 | { | |
1091 | struct call_rcu_data *crdp; | |
1092 | bool teardown = true; | |
1093 | ||
1094 | if (default_call_rcu_data == NULL) | |
1095 | return; | |
1096 | call_rcu_lock(&call_rcu_mutex); | |
1097 | /* | |
1098 | * If the application leaves callbacks in the default call_rcu | |
1099 | * worker queue, keep the default worker in place. | |
1100 | */ | |
1101 | crdp = default_call_rcu_data; | |
1102 | if (!crdp) { | |
1103 | teardown = false; | |
1104 | goto unlock; | |
1105 | } | |
1106 | if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) { | |
1107 | teardown = false; | |
1108 | goto unlock; | |
1109 | } | |
1110 | rcu_set_pointer(&default_call_rcu_data, NULL); | |
1111 | unlock: | |
1112 | call_rcu_unlock(&call_rcu_mutex); | |
1113 | if (teardown) { | |
1114 | synchronize_rcu(); | |
1115 | call_rcu_data_free(crdp); | |
1116 | } | |
1117 | } |