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" | |
a47dd11c | 39 | #include "compat-getcpu.h" |
5161f31e | 40 | #include "urcu/wfcqueue.h" |
b57aee66 PM |
41 | #include "urcu-call-rcu.h" |
42 | #include "urcu-pointer.h" | |
3c24913f | 43 | #include "urcu/list.h" |
41849996 | 44 | #include "urcu/futex.h" |
bd252a04 | 45 | #include "urcu/tls-compat.h" |
81dd9134 | 46 | #include "urcu/ref.h" |
4a6d7378 | 47 | #include "urcu-die.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 | ||
3c24913f PM |
84 | /* |
85 | * List of all call_rcu_data structures to keep valgrind happy. | |
86 | * Protected by call_rcu_mutex. | |
87 | */ | |
88 | ||
bab44e28 | 89 | static CDS_LIST_HEAD(call_rcu_data_list); |
3c24913f | 90 | |
b57aee66 PM |
91 | /* Link a thread using call_rcu() to its call_rcu thread. */ |
92 | ||
bd252a04 | 93 | static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data); |
b57aee66 | 94 | |
e85451a1 MD |
95 | /* |
96 | * Guard call_rcu thread creation and atfork handlers. | |
97 | */ | |
b57aee66 PM |
98 | static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER; |
99 | ||
100 | /* If a given thread does not have its own call_rcu thread, this is default. */ | |
101 | ||
102 | static struct call_rcu_data *default_call_rcu_data; | |
103 | ||
b57aee66 PM |
104 | /* |
105 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are | |
106 | * available, then we can have call_rcu threads assigned to individual | |
107 | * CPUs rather than only to specific threads. | |
108 | */ | |
109 | ||
a47dd11c | 110 | #if defined(HAVE_SYSCONF) && (defined(HAVE_SCHED_GETCPU) || defined(HAVE_GETCPUID)) |
b57aee66 PM |
111 | |
112 | /* | |
113 | * Pointer to array of pointers to per-CPU call_rcu_data structures | |
618b2595 MD |
114 | * and # CPUs. per_cpu_call_rcu_data is a RCU-protected pointer to an |
115 | * array of RCU-protected pointers to call_rcu_data. call_rcu acts as a | |
116 | * RCU read-side and reads per_cpu_call_rcu_data and the per-cpu pointer | |
117 | * without mutex. The call_rcu_mutex protects updates. | |
b57aee66 PM |
118 | */ |
119 | ||
120 | static struct call_rcu_data **per_cpu_call_rcu_data; | |
121 | static long maxcpus; | |
122 | ||
60af049d LJ |
123 | static void maxcpus_reset(void) |
124 | { | |
125 | maxcpus = 0; | |
126 | } | |
127 | ||
b57aee66 PM |
128 | /* Allocate the array if it has not already been allocated. */ |
129 | ||
130 | static void alloc_cpu_call_rcu_data(void) | |
131 | { | |
132 | struct call_rcu_data **p; | |
133 | static int warned = 0; | |
134 | ||
135 | if (maxcpus != 0) | |
136 | return; | |
137 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); | |
138 | if (maxcpus <= 0) { | |
139 | return; | |
140 | } | |
141 | p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
142 | if (p != NULL) { | |
143 | memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data)); | |
618b2595 | 144 | rcu_set_pointer(&per_cpu_call_rcu_data, p); |
b57aee66 PM |
145 | } else { |
146 | if (!warned) { | |
147 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
148 | } | |
149 | warned = 1; | |
150 | } | |
151 | } | |
152 | ||
63b495d8 | 153 | #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
b57aee66 | 154 | |
f9437098 MD |
155 | /* |
156 | * per_cpu_call_rcu_data should be constant, but some functions below, used both | |
157 | * for cases where cpu number is available and not available, assume it it not | |
158 | * constant. | |
159 | */ | |
160 | static struct call_rcu_data **per_cpu_call_rcu_data = NULL; | |
b57aee66 PM |
161 | static const long maxcpus = -1; |
162 | ||
60af049d LJ |
163 | static void maxcpus_reset(void) |
164 | { | |
165 | } | |
166 | ||
b57aee66 PM |
167 | static void alloc_cpu_call_rcu_data(void) |
168 | { | |
169 | } | |
170 | ||
63b495d8 | 171 | #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
b57aee66 PM |
172 | |
173 | /* Acquire the specified pthread mutex. */ | |
174 | ||
175 | static void call_rcu_lock(pthread_mutex_t *pmp) | |
176 | { | |
4a6d7378 MD |
177 | int ret; |
178 | ||
179 | ret = pthread_mutex_lock(pmp); | |
180 | if (ret) | |
181 | urcu_die(ret); | |
b57aee66 PM |
182 | } |
183 | ||
184 | /* Release the specified pthread mutex. */ | |
185 | ||
186 | static void call_rcu_unlock(pthread_mutex_t *pmp) | |
187 | { | |
4a6d7378 MD |
188 | int ret; |
189 | ||
190 | ret = pthread_mutex_unlock(pmp); | |
191 | if (ret) | |
192 | urcu_die(ret); | |
b57aee66 PM |
193 | } |
194 | ||
8e3690db MD |
195 | /* |
196 | * Periodically retry setting CPU affinity if we migrate. | |
197 | * Losing affinity can be caused by CPU hotunplug/hotplug, or by | |
198 | * cpuset(7). | |
199 | */ | |
c1d2c60b MD |
200 | #if HAVE_SCHED_SETAFFINITY |
201 | static | |
202 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
203 | { | |
204 | cpu_set_t mask; | |
8e3690db | 205 | int ret; |
c1d2c60b MD |
206 | |
207 | if (crdp->cpu_affinity < 0) | |
208 | return 0; | |
8e3690db MD |
209 | if (++crdp->gp_count & SET_AFFINITY_CHECK_PERIOD_MASK) |
210 | return 0; | |
211 | if (urcu_sched_getcpu() == crdp->cpu_affinity) | |
212 | return 0; | |
c1d2c60b MD |
213 | |
214 | CPU_ZERO(&mask); | |
215 | CPU_SET(crdp->cpu_affinity, &mask); | |
216 | #if SCHED_SETAFFINITY_ARGS == 2 | |
8e3690db | 217 | ret = sched_setaffinity(0, &mask); |
c1d2c60b | 218 | #else |
8e3690db | 219 | ret = sched_setaffinity(0, sizeof(mask), &mask); |
c1d2c60b | 220 | #endif |
8e3690db MD |
221 | /* |
222 | * EINVAL is fine: can be caused by hotunplugged CPUs, or by | |
223 | * cpuset(7). This is why we should always retry if we detect | |
224 | * migration. | |
225 | */ | |
226 | if (ret && errno == EINVAL) { | |
227 | ret = 0; | |
228 | errno = 0; | |
229 | } | |
230 | return ret; | |
c1d2c60b MD |
231 | } |
232 | #else | |
233 | static | |
234 | int set_thread_cpu_affinity(struct call_rcu_data *crdp) | |
235 | { | |
236 | return 0; | |
237 | } | |
238 | #endif | |
239 | ||
03fe58b3 MD |
240 | static void call_rcu_wait(struct call_rcu_data *crdp) |
241 | { | |
242 | /* Read call_rcu list before read futex */ | |
243 | cmm_smp_mb(); | |
b0a841b4 MD |
244 | if (uatomic_read(&crdp->futex) != -1) |
245 | return; | |
246 | while (futex_async(&crdp->futex, FUTEX_WAIT, -1, | |
247 | NULL, NULL, 0)) { | |
248 | switch (errno) { | |
249 | case EWOULDBLOCK: | |
250 | /* Value already changed. */ | |
251 | return; | |
252 | case EINTR: | |
253 | /* Retry if interrupted by signal. */ | |
254 | break; /* Get out of switch. */ | |
255 | default: | |
256 | /* Unexpected error. */ | |
257 | urcu_die(errno); | |
258 | } | |
259 | } | |
03fe58b3 MD |
260 | } |
261 | ||
262 | static void call_rcu_wake_up(struct call_rcu_data *crdp) | |
263 | { | |
264 | /* Write to call_rcu list before reading/writing futex */ | |
265 | cmm_smp_mb(); | |
a0b7f7ea | 266 | if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) { |
03fe58b3 | 267 | uatomic_set(&crdp->futex, 0); |
b0a841b4 MD |
268 | if (futex_async(&crdp->futex, FUTEX_WAKE, 1, |
269 | NULL, NULL, 0) < 0) | |
270 | urcu_die(errno); | |
03fe58b3 MD |
271 | } |
272 | } | |
273 | ||
b7f721d9 MD |
274 | static void call_rcu_completion_wait(struct call_rcu_completion *completion) |
275 | { | |
276 | /* Read completion barrier count before read futex */ | |
277 | cmm_smp_mb(); | |
b0a841b4 MD |
278 | if (uatomic_read(&completion->futex) != -1) |
279 | return; | |
280 | while (futex_async(&completion->futex, FUTEX_WAIT, -1, | |
281 | NULL, NULL, 0)) { | |
282 | switch (errno) { | |
283 | case EWOULDBLOCK: | |
284 | /* Value already changed. */ | |
285 | return; | |
286 | case EINTR: | |
287 | /* Retry if interrupted by signal. */ | |
288 | break; /* Get out of switch. */ | |
289 | default: | |
290 | /* Unexpected error. */ | |
291 | urcu_die(errno); | |
292 | } | |
293 | } | |
b7f721d9 MD |
294 | } |
295 | ||
296 | static void call_rcu_completion_wake_up(struct call_rcu_completion *completion) | |
297 | { | |
298 | /* Write to completion barrier count before reading/writing futex */ | |
299 | cmm_smp_mb(); | |
300 | if (caa_unlikely(uatomic_read(&completion->futex) == -1)) { | |
301 | uatomic_set(&completion->futex, 0); | |
b0a841b4 MD |
302 | if (futex_async(&completion->futex, FUTEX_WAKE, 1, |
303 | NULL, NULL, 0) < 0) | |
304 | urcu_die(errno); | |
b7f721d9 MD |
305 | } |
306 | } | |
307 | ||
b57aee66 PM |
308 | /* This is the code run by each call_rcu thread. */ |
309 | ||
310 | static void *call_rcu_thread(void *arg) | |
311 | { | |
312 | unsigned long cbcount; | |
5161f31e | 313 | struct call_rcu_data *crdp = (struct call_rcu_data *) arg; |
2870aa1e | 314 | int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT); |
b57aee66 | 315 | |
8e3690db | 316 | if (set_thread_cpu_affinity(crdp)) |
4a6d7378 | 317 | urcu_die(errno); |
c1d2c60b | 318 | |
765f3ead MD |
319 | /* |
320 | * If callbacks take a read-side lock, we need to be registered. | |
321 | */ | |
322 | rcu_register_thread(); | |
323 | ||
bd252a04 | 324 | URCU_TLS(thread_call_rcu_data) = crdp; |
bc94ca9b MD |
325 | if (!rt) { |
326 | uatomic_dec(&crdp->futex); | |
327 | /* Decrement futex before reading call_rcu list */ | |
328 | cmm_smp_mb(); | |
329 | } | |
b57aee66 | 330 | for (;;) { |
5161f31e MD |
331 | struct cds_wfcq_head cbs_tmp_head; |
332 | struct cds_wfcq_tail cbs_tmp_tail; | |
333 | struct cds_wfcq_node *cbs, *cbs_tmp_n; | |
ae25b7e2 | 334 | enum cds_wfcq_ret splice_ret; |
5161f31e | 335 | |
8e3690db MD |
336 | if (set_thread_cpu_affinity(crdp)) |
337 | urcu_die(errno); | |
338 | ||
e85451a1 MD |
339 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) { |
340 | /* | |
341 | * Pause requested. Become quiescent: remove | |
342 | * ourself from all global lists, and don't | |
343 | * process any callback. The callback lists may | |
344 | * still be non-empty though. | |
345 | */ | |
346 | rcu_unregister_thread(); | |
347 | cmm_smp_mb__before_uatomic_or(); | |
348 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSED); | |
349 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSE) != 0) | |
109105b7 | 350 | (void) poll(NULL, 0, 1); |
fc236e5e KF |
351 | uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSED); |
352 | cmm_smp_mb__after_uatomic_and(); | |
e85451a1 MD |
353 | rcu_register_thread(); |
354 | } | |
355 | ||
5161f31e | 356 | cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail); |
ae25b7e2 MD |
357 | splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head, |
358 | &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail); | |
359 | assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK); | |
360 | assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY); | |
361 | if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) { | |
b57aee66 PM |
362 | synchronize_rcu(); |
363 | cbcount = 0; | |
5161f31e MD |
364 | __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head, |
365 | &cbs_tmp_tail, cbs, cbs_tmp_n) { | |
366 | struct rcu_head *rhp; | |
367 | ||
368 | rhp = caa_container_of(cbs, | |
369 | struct rcu_head, next); | |
b57aee66 PM |
370 | rhp->func(rhp); |
371 | cbcount++; | |
5161f31e | 372 | } |
b57aee66 PM |
373 | uatomic_sub(&crdp->qlen, cbcount); |
374 | } | |
bc94ca9b MD |
375 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP) |
376 | break; | |
765f3ead | 377 | rcu_thread_offline(); |
bc94ca9b | 378 | if (!rt) { |
5161f31e MD |
379 | if (cds_wfcq_empty(&crdp->cbs_head, |
380 | &crdp->cbs_tail)) { | |
bc94ca9b | 381 | call_rcu_wait(crdp); |
109105b7 | 382 | (void) poll(NULL, 0, 10); |
bc94ca9b | 383 | uatomic_dec(&crdp->futex); |
c768e45e | 384 | /* |
bc94ca9b MD |
385 | * Decrement futex before reading |
386 | * call_rcu list. | |
c768e45e MD |
387 | */ |
388 | cmm_smp_mb(); | |
ccbac24d | 389 | } else { |
109105b7 | 390 | (void) poll(NULL, 0, 10); |
c768e45e | 391 | } |
bc94ca9b | 392 | } else { |
109105b7 | 393 | (void) poll(NULL, 0, 10); |
b57aee66 | 394 | } |
765f3ead | 395 | rcu_thread_online(); |
bc94ca9b MD |
396 | } |
397 | if (!rt) { | |
398 | /* | |
399 | * Read call_rcu list before write futex. | |
400 | */ | |
401 | cmm_smp_mb(); | |
402 | uatomic_set(&crdp->futex, 0); | |
b57aee66 | 403 | } |
2870aa1e | 404 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED); |
765f3ead | 405 | rcu_unregister_thread(); |
7106ddf8 | 406 | return NULL; |
b57aee66 PM |
407 | } |
408 | ||
409 | /* | |
410 | * Create both a call_rcu thread and the corresponding call_rcu_data | |
3c24913f PM |
411 | * structure, linking the structure in as specified. Caller must hold |
412 | * call_rcu_mutex. | |
b57aee66 PM |
413 | */ |
414 | ||
3c24913f | 415 | static void call_rcu_data_init(struct call_rcu_data **crdpp, |
c1d2c60b MD |
416 | unsigned long flags, |
417 | int cpu_affinity) | |
b57aee66 PM |
418 | { |
419 | struct call_rcu_data *crdp; | |
4a6d7378 | 420 | int ret; |
b57aee66 PM |
421 | |
422 | crdp = malloc(sizeof(*crdp)); | |
4a6d7378 MD |
423 | if (crdp == NULL) |
424 | urcu_die(errno); | |
b57aee66 | 425 | memset(crdp, '\0', sizeof(*crdp)); |
5161f31e | 426 | cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail); |
b57aee66 | 427 | crdp->qlen = 0; |
263e3cf9 MD |
428 | crdp->futex = 0; |
429 | crdp->flags = flags; | |
3c24913f | 430 | cds_list_add(&crdp->list, &call_rcu_data_list); |
c1d2c60b | 431 | crdp->cpu_affinity = cpu_affinity; |
8e3690db | 432 | crdp->gp_count = 0; |
b57aee66 PM |
433 | cmm_smp_mb(); /* Structure initialized before pointer is planted. */ |
434 | *crdpp = crdp; | |
4a6d7378 MD |
435 | ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp); |
436 | if (ret) | |
437 | urcu_die(ret); | |
b57aee66 PM |
438 | } |
439 | ||
440 | /* | |
441 | * Return a pointer to the call_rcu_data structure for the specified | |
442 | * CPU, returning NULL if there is none. We cannot automatically | |
443 | * created it because the platform we are running on might not define | |
63b495d8 | 444 | * urcu_sched_getcpu(). |
618b2595 MD |
445 | * |
446 | * The call to this function and use of the returned call_rcu_data | |
447 | * should be protected by RCU read-side lock. | |
b57aee66 PM |
448 | */ |
449 | ||
450 | struct call_rcu_data *get_cpu_call_rcu_data(int cpu) | |
451 | { | |
452 | static int warned = 0; | |
618b2595 | 453 | struct call_rcu_data **pcpu_crdp; |
b57aee66 | 454 | |
618b2595 MD |
455 | pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data); |
456 | if (pcpu_crdp == NULL) | |
b57aee66 PM |
457 | return NULL; |
458 | if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) { | |
459 | fprintf(stderr, "[error] liburcu: get CPU # out of range\n"); | |
460 | warned = 1; | |
461 | } | |
462 | if (cpu < 0 || maxcpus <= cpu) | |
463 | return NULL; | |
618b2595 | 464 | return rcu_dereference(pcpu_crdp[cpu]); |
b57aee66 PM |
465 | } |
466 | ||
467 | /* | |
468 | * Return the tid corresponding to the call_rcu thread whose | |
469 | * call_rcu_data structure is specified. | |
470 | */ | |
471 | ||
472 | pthread_t get_call_rcu_thread(struct call_rcu_data *crdp) | |
473 | { | |
474 | return crdp->tid; | |
475 | } | |
476 | ||
477 | /* | |
478 | * Create a call_rcu_data structure (with thread) and return a pointer. | |
479 | */ | |
480 | ||
c1d2c60b MD |
481 | static struct call_rcu_data *__create_call_rcu_data(unsigned long flags, |
482 | int cpu_affinity) | |
b57aee66 PM |
483 | { |
484 | struct call_rcu_data *crdp; | |
485 | ||
c1d2c60b | 486 | call_rcu_data_init(&crdp, flags, cpu_affinity); |
b57aee66 PM |
487 | return crdp; |
488 | } | |
489 | ||
c1d2c60b MD |
490 | struct call_rcu_data *create_call_rcu_data(unsigned long flags, |
491 | int cpu_affinity) | |
3c24913f PM |
492 | { |
493 | struct call_rcu_data *crdp; | |
494 | ||
495 | call_rcu_lock(&call_rcu_mutex); | |
c1d2c60b | 496 | crdp = __create_call_rcu_data(flags, cpu_affinity); |
3c24913f PM |
497 | call_rcu_unlock(&call_rcu_mutex); |
498 | return crdp; | |
499 | } | |
500 | ||
b57aee66 PM |
501 | /* |
502 | * Set the specified CPU to use the specified call_rcu_data structure. | |
7106ddf8 PM |
503 | * |
504 | * Use NULL to remove a CPU's call_rcu_data structure, but it is | |
505 | * the caller's responsibility to dispose of the removed structure. | |
506 | * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure | |
507 | * (prior to NULLing it out, of course). | |
f9da0936 MD |
508 | * |
509 | * The caller must wait for a grace-period to pass between return from | |
510 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
511 | * previous call rcu data as argument. | |
b57aee66 PM |
512 | */ |
513 | ||
514 | int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp) | |
515 | { | |
dcfc8165 | 516 | static int warned = 0; |
b57aee66 PM |
517 | |
518 | call_rcu_lock(&call_rcu_mutex); | |
f3776786 | 519 | alloc_cpu_call_rcu_data(); |
b57aee66 PM |
520 | if (cpu < 0 || maxcpus <= cpu) { |
521 | if (!warned) { | |
522 | fprintf(stderr, "[error] liburcu: set CPU # out of range\n"); | |
523 | warned = 1; | |
524 | } | |
525 | call_rcu_unlock(&call_rcu_mutex); | |
526 | errno = EINVAL; | |
527 | return -EINVAL; | |
528 | } | |
53a55535 | 529 | |
b57aee66 | 530 | if (per_cpu_call_rcu_data == NULL) { |
3670fef2 | 531 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
532 | errno = ENOMEM; |
533 | return -ENOMEM; | |
534 | } | |
53a55535 LJ |
535 | |
536 | if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) { | |
537 | call_rcu_unlock(&call_rcu_mutex); | |
538 | errno = EEXIST; | |
539 | return -EEXIST; | |
540 | } | |
541 | ||
618b2595 | 542 | rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp); |
3670fef2 | 543 | call_rcu_unlock(&call_rcu_mutex); |
b57aee66 PM |
544 | return 0; |
545 | } | |
546 | ||
547 | /* | |
548 | * Return a pointer to the default call_rcu_data structure, creating | |
549 | * one if need be. Because we never free call_rcu_data structures, | |
550 | * we don't need to be in an RCU read-side critical section. | |
551 | */ | |
552 | ||
553 | struct call_rcu_data *get_default_call_rcu_data(void) | |
554 | { | |
555 | if (default_call_rcu_data != NULL) | |
556 | return rcu_dereference(default_call_rcu_data); | |
557 | call_rcu_lock(&call_rcu_mutex); | |
558 | if (default_call_rcu_data != NULL) { | |
559 | call_rcu_unlock(&call_rcu_mutex); | |
560 | return default_call_rcu_data; | |
561 | } | |
c1d2c60b | 562 | call_rcu_data_init(&default_call_rcu_data, 0, -1); |
b57aee66 PM |
563 | call_rcu_unlock(&call_rcu_mutex); |
564 | return default_call_rcu_data; | |
565 | } | |
566 | ||
567 | /* | |
568 | * Return the call_rcu_data structure that applies to the currently | |
569 | * running thread. Any call_rcu_data structure assigned specifically | |
570 | * to this thread has first priority, followed by any call_rcu_data | |
571 | * structure assigned to the CPU on which the thread is running, | |
572 | * followed by the default call_rcu_data structure. If there is not | |
573 | * yet a default call_rcu_data structure, one will be created. | |
618b2595 MD |
574 | * |
575 | * Calls to this function and use of the returned call_rcu_data should | |
576 | * be protected by RCU read-side lock. | |
b57aee66 PM |
577 | */ |
578 | struct call_rcu_data *get_call_rcu_data(void) | |
579 | { | |
9744f3bb | 580 | struct call_rcu_data *crd; |
b57aee66 | 581 | |
bd252a04 MD |
582 | if (URCU_TLS(thread_call_rcu_data) != NULL) |
583 | return URCU_TLS(thread_call_rcu_data); | |
9744f3bb LJ |
584 | |
585 | if (maxcpus > 0) { | |
63b495d8 | 586 | crd = get_cpu_call_rcu_data(urcu_sched_getcpu()); |
9744f3bb LJ |
587 | if (crd) |
588 | return crd; | |
b57aee66 | 589 | } |
9744f3bb | 590 | |
b57aee66 PM |
591 | return get_default_call_rcu_data(); |
592 | } | |
593 | ||
594 | /* | |
595 | * Return a pointer to this task's call_rcu_data if there is one. | |
596 | */ | |
597 | ||
598 | struct call_rcu_data *get_thread_call_rcu_data(void) | |
599 | { | |
bd252a04 | 600 | return URCU_TLS(thread_call_rcu_data); |
b57aee66 PM |
601 | } |
602 | ||
603 | /* | |
604 | * Set this task's call_rcu_data structure as specified, regardless | |
605 | * of whether or not this task already had one. (This allows switching | |
606 | * to and from real-time call_rcu threads, for example.) | |
7106ddf8 PM |
607 | * |
608 | * Use NULL to remove a thread's call_rcu_data structure, but it is | |
609 | * the caller's responsibility to dispose of the removed structure. | |
610 | * Use get_thread_call_rcu_data() to obtain a pointer to the old structure | |
611 | * (prior to NULLing it out, of course). | |
b57aee66 PM |
612 | */ |
613 | ||
614 | void set_thread_call_rcu_data(struct call_rcu_data *crdp) | |
615 | { | |
bd252a04 | 616 | URCU_TLS(thread_call_rcu_data) = crdp; |
b57aee66 PM |
617 | } |
618 | ||
619 | /* | |
620 | * Create a separate call_rcu thread for each CPU. This does not | |
621 | * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data() | |
0938c541 MD |
622 | * function if you want that behavior. Should be paired with |
623 | * free_all_cpu_call_rcu_data() to teardown these call_rcu worker | |
624 | * threads. | |
b57aee66 PM |
625 | */ |
626 | ||
627 | int create_all_cpu_call_rcu_data(unsigned long flags) | |
628 | { | |
629 | int i; | |
630 | struct call_rcu_data *crdp; | |
631 | int ret; | |
632 | ||
633 | call_rcu_lock(&call_rcu_mutex); | |
634 | alloc_cpu_call_rcu_data(); | |
635 | call_rcu_unlock(&call_rcu_mutex); | |
636 | if (maxcpus <= 0) { | |
637 | errno = EINVAL; | |
638 | return -EINVAL; | |
639 | } | |
640 | if (per_cpu_call_rcu_data == NULL) { | |
641 | errno = ENOMEM; | |
642 | return -ENOMEM; | |
643 | } | |
644 | for (i = 0; i < maxcpus; i++) { | |
645 | call_rcu_lock(&call_rcu_mutex); | |
646 | if (get_cpu_call_rcu_data(i)) { | |
647 | call_rcu_unlock(&call_rcu_mutex); | |
648 | continue; | |
649 | } | |
c1d2c60b | 650 | crdp = __create_call_rcu_data(flags, i); |
b57aee66 PM |
651 | if (crdp == NULL) { |
652 | call_rcu_unlock(&call_rcu_mutex); | |
653 | errno = ENOMEM; | |
654 | return -ENOMEM; | |
655 | } | |
656 | call_rcu_unlock(&call_rcu_mutex); | |
657 | if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) { | |
356c8794 LJ |
658 | call_rcu_data_free(crdp); |
659 | ||
660 | /* it has been created by other thread */ | |
661 | if (ret == -EEXIST) | |
662 | continue; | |
663 | ||
664 | return ret; | |
b57aee66 PM |
665 | } |
666 | } | |
667 | return 0; | |
668 | } | |
669 | ||
7106ddf8 PM |
670 | /* |
671 | * Wake up the call_rcu thread corresponding to the specified | |
672 | * call_rcu_data structure. | |
673 | */ | |
674 | static void wake_call_rcu_thread(struct call_rcu_data *crdp) | |
675 | { | |
263e3cf9 MD |
676 | if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT)) |
677 | call_rcu_wake_up(crdp); | |
7106ddf8 PM |
678 | } |
679 | ||
b7f721d9 MD |
680 | static void _call_rcu(struct rcu_head *head, |
681 | void (*func)(struct rcu_head *head), | |
682 | struct call_rcu_data *crdp) | |
683 | { | |
684 | cds_wfcq_node_init(&head->next); | |
685 | head->func = func; | |
686 | cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next); | |
687 | uatomic_inc(&crdp->qlen); | |
688 | wake_call_rcu_thread(crdp); | |
689 | } | |
690 | ||
b57aee66 PM |
691 | /* |
692 | * Schedule a function to be invoked after a following grace period. | |
693 | * This is the only function that must be called -- the others are | |
694 | * only present to allow applications to tune their use of RCU for | |
695 | * maximum performance. | |
696 | * | |
697 | * Note that unless a call_rcu thread has not already been created, | |
698 | * the first invocation of call_rcu() will create one. So, if you | |
699 | * need the first invocation of call_rcu() to be fast, make sure | |
700 | * to create a call_rcu thread first. One way to accomplish this is | |
701 | * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data(). | |
618b2595 MD |
702 | * |
703 | * call_rcu must be called by registered RCU read-side threads. | |
b57aee66 | 704 | */ |
b57aee66 PM |
705 | void call_rcu(struct rcu_head *head, |
706 | void (*func)(struct rcu_head *head)) | |
707 | { | |
708 | struct call_rcu_data *crdp; | |
709 | ||
618b2595 | 710 | /* Holding rcu read-side lock across use of per-cpu crdp */ |
ffb4aaa7 | 711 | _rcu_read_lock(); |
b57aee66 | 712 | crdp = get_call_rcu_data(); |
b7f721d9 | 713 | _call_rcu(head, func, crdp); |
ffb4aaa7 | 714 | _rcu_read_unlock(); |
7106ddf8 PM |
715 | } |
716 | ||
717 | /* | |
718 | * Free up the specified call_rcu_data structure, terminating the | |
719 | * associated call_rcu thread. The caller must have previously | |
720 | * removed the call_rcu_data structure from per-thread or per-CPU | |
721 | * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU | |
722 | * call_rcu_data structures or set_thread_call_rcu_data(NULL) for | |
723 | * per-thread call_rcu_data structures. | |
724 | * | |
725 | * We silently refuse to free up the default call_rcu_data structure | |
726 | * because that is where we put any leftover callbacks. Note that | |
727 | * the possibility of self-spawning callbacks makes it impossible | |
728 | * to execute all the callbacks in finite time without putting any | |
729 | * newly spawned callbacks somewhere else. The "somewhere else" of | |
730 | * last resort is the default call_rcu_data structure. | |
731 | * | |
732 | * We also silently refuse to free NULL pointers. This simplifies | |
733 | * the calling code. | |
f9da0936 MD |
734 | * |
735 | * The caller must wait for a grace-period to pass between return from | |
736 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the | |
737 | * previous call rcu data as argument. | |
03e5118f MD |
738 | * |
739 | * Note: introducing __cds_wfcq_splice_blocking() in this function fixed | |
740 | * a list corruption bug in the 0.7.x series. The equivalent fix | |
741 | * appeared in 0.6.8 for the stable-0.6 branch. | |
7106ddf8 PM |
742 | */ |
743 | void call_rcu_data_free(struct call_rcu_data *crdp) | |
744 | { | |
7106ddf8 PM |
745 | if (crdp == NULL || crdp == default_call_rcu_data) { |
746 | return; | |
747 | } | |
2870aa1e PB |
748 | if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) { |
749 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP); | |
7106ddf8 | 750 | wake_call_rcu_thread(crdp); |
2870aa1e | 751 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) |
109105b7 | 752 | (void) poll(NULL, 0, 1); |
7106ddf8 | 753 | } |
5161f31e | 754 | if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) { |
698d0778 MD |
755 | /* Create default call rcu data if need be */ |
756 | (void) get_default_call_rcu_data(); | |
5161f31e MD |
757 | __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head, |
758 | &default_call_rcu_data->cbs_tail, | |
759 | &crdp->cbs_head, &crdp->cbs_tail); | |
7106ddf8 PM |
760 | uatomic_add(&default_call_rcu_data->qlen, |
761 | uatomic_read(&crdp->qlen)); | |
1e92aa15 | 762 | wake_call_rcu_thread(default_call_rcu_data); |
7106ddf8 | 763 | } |
59dc9e9d | 764 | |
b75dffe6 | 765 | call_rcu_lock(&call_rcu_mutex); |
59dc9e9d | 766 | cds_list_del(&crdp->list); |
b75dffe6 LJ |
767 | call_rcu_unlock(&call_rcu_mutex); |
768 | ||
59dc9e9d | 769 | free(crdp); |
7106ddf8 PM |
770 | } |
771 | ||
772 | /* | |
773 | * Clean up all the per-CPU call_rcu threads. | |
774 | */ | |
775 | void free_all_cpu_call_rcu_data(void) | |
776 | { | |
777 | int cpu; | |
618b2595 MD |
778 | struct call_rcu_data **crdp; |
779 | static int warned = 0; | |
7106ddf8 PM |
780 | |
781 | if (maxcpus <= 0) | |
782 | return; | |
618b2595 MD |
783 | |
784 | crdp = malloc(sizeof(*crdp) * maxcpus); | |
785 | if (!crdp) { | |
786 | if (!warned) { | |
787 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); | |
788 | } | |
789 | warned = 1; | |
d31150b4 | 790 | return; |
618b2595 MD |
791 | } |
792 | ||
7106ddf8 | 793 | for (cpu = 0; cpu < maxcpus; cpu++) { |
618b2595 MD |
794 | crdp[cpu] = get_cpu_call_rcu_data(cpu); |
795 | if (crdp[cpu] == NULL) | |
7106ddf8 PM |
796 | continue; |
797 | set_cpu_call_rcu_data(cpu, NULL); | |
7106ddf8 | 798 | } |
618b2595 MD |
799 | /* |
800 | * Wait for call_rcu sites acting as RCU readers of the | |
801 | * call_rcu_data to become quiescent. | |
802 | */ | |
803 | synchronize_rcu(); | |
804 | for (cpu = 0; cpu < maxcpus; cpu++) { | |
805 | if (crdp[cpu] == NULL) | |
806 | continue; | |
807 | call_rcu_data_free(crdp[cpu]); | |
808 | } | |
809 | free(crdp); | |
7106ddf8 PM |
810 | } |
811 | ||
81dd9134 KF |
812 | static |
813 | void free_completion(struct urcu_ref *ref) | |
814 | { | |
815 | struct call_rcu_completion *completion; | |
816 | ||
817 | completion = caa_container_of(ref, struct call_rcu_completion, ref); | |
818 | free(completion); | |
819 | } | |
820 | ||
b7f721d9 MD |
821 | static |
822 | void _rcu_barrier_complete(struct rcu_head *head) | |
823 | { | |
824 | struct call_rcu_completion_work *work; | |
825 | struct call_rcu_completion *completion; | |
826 | ||
827 | work = caa_container_of(head, struct call_rcu_completion_work, head); | |
828 | completion = work->completion; | |
81dd9134 KF |
829 | if (!uatomic_sub_return(&completion->barrier_count, 1)) |
830 | call_rcu_completion_wake_up(completion); | |
831 | urcu_ref_put(&completion->ref, free_completion); | |
b7f721d9 MD |
832 | free(work); |
833 | } | |
834 | ||
835 | /* | |
836 | * Wait for all in-flight call_rcu callbacks to complete execution. | |
837 | */ | |
838 | void rcu_barrier(void) | |
839 | { | |
840 | struct call_rcu_data *crdp; | |
81dd9134 | 841 | struct call_rcu_completion *completion; |
63f0fc6d | 842 | int count = 0; |
b7f721d9 MD |
843 | int was_online; |
844 | ||
845 | /* Put in offline state in QSBR. */ | |
ffb4aaa7 | 846 | was_online = _rcu_read_ongoing(); |
b7f721d9 MD |
847 | if (was_online) |
848 | rcu_thread_offline(); | |
849 | /* | |
850 | * Calling a rcu_barrier() within a RCU read-side critical | |
851 | * section is an error. | |
852 | */ | |
ffb4aaa7 | 853 | if (_rcu_read_ongoing()) { |
b7f721d9 MD |
854 | static int warned = 0; |
855 | ||
856 | if (!warned) { | |
857 | fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n"); | |
858 | } | |
859 | warned = 1; | |
860 | goto online; | |
861 | } | |
862 | ||
81dd9134 KF |
863 | completion = calloc(sizeof(*completion), 1); |
864 | if (!completion) | |
865 | urcu_die(errno); | |
866 | ||
b7f721d9 MD |
867 | call_rcu_lock(&call_rcu_mutex); |
868 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) | |
869 | count++; | |
870 | ||
81dd9134 KF |
871 | /* Referenced by rcu_barrier() and each call_rcu thread. */ |
872 | urcu_ref_set(&completion->ref, count + 1); | |
873 | completion->barrier_count = count; | |
b7f721d9 MD |
874 | |
875 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
876 | struct call_rcu_completion_work *work; | |
877 | ||
878 | work = calloc(sizeof(*work), 1); | |
63f0fc6d MD |
879 | if (!work) |
880 | urcu_die(errno); | |
81dd9134 | 881 | work->completion = completion; |
b7f721d9 | 882 | _call_rcu(&work->head, _rcu_barrier_complete, crdp); |
b7f721d9 MD |
883 | } |
884 | call_rcu_unlock(&call_rcu_mutex); | |
885 | ||
b7f721d9 MD |
886 | /* Wait for them */ |
887 | for (;;) { | |
81dd9134 | 888 | uatomic_dec(&completion->futex); |
b7f721d9 MD |
889 | /* Decrement futex before reading barrier_count */ |
890 | cmm_smp_mb(); | |
81dd9134 | 891 | if (!uatomic_read(&completion->barrier_count)) |
b7f721d9 | 892 | break; |
81dd9134 | 893 | call_rcu_completion_wait(completion); |
b7f721d9 | 894 | } |
81dd9134 KF |
895 | |
896 | urcu_ref_put(&completion->ref, free_completion); | |
897 | ||
b7f721d9 MD |
898 | online: |
899 | if (was_online) | |
900 | rcu_thread_online(); | |
901 | } | |
902 | ||
81ad2e19 PM |
903 | /* |
904 | * Acquire the call_rcu_mutex in order to ensure that the child sees | |
e85451a1 MD |
905 | * all of the call_rcu() data structures in a consistent state. Ensure |
906 | * that all call_rcu threads are in a quiescent state across fork. | |
81ad2e19 PM |
907 | * Suitable for pthread_atfork() and friends. |
908 | */ | |
909 | void call_rcu_before_fork(void) | |
910 | { | |
e85451a1 MD |
911 | struct call_rcu_data *crdp; |
912 | ||
81ad2e19 | 913 | call_rcu_lock(&call_rcu_mutex); |
e85451a1 MD |
914 | |
915 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
916 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE); | |
917 | cmm_smp_mb__after_uatomic_or(); | |
918 | wake_call_rcu_thread(crdp); | |
919 | } | |
920 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { | |
921 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0) | |
109105b7 | 922 | (void) poll(NULL, 0, 1); |
e85451a1 | 923 | } |
81ad2e19 PM |
924 | } |
925 | ||
926 | /* | |
927 | * Clean up call_rcu data structures in the parent of a successful fork() | |
928 | * that is not followed by exec() in the child. Suitable for | |
929 | * pthread_atfork() and friends. | |
930 | */ | |
931 | void call_rcu_after_fork_parent(void) | |
932 | { | |
e85451a1 MD |
933 | struct call_rcu_data *crdp; |
934 | ||
935 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) | |
936 | uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE); | |
fc236e5e KF |
937 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { |
938 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) != 0) | |
109105b7 | 939 | (void) poll(NULL, 0, 1); |
fc236e5e | 940 | } |
81ad2e19 PM |
941 | call_rcu_unlock(&call_rcu_mutex); |
942 | } | |
943 | ||
7106ddf8 PM |
944 | /* |
945 | * Clean up call_rcu data structures in the child of a successful fork() | |
81ad2e19 PM |
946 | * that is not followed by exec(). Suitable for pthread_atfork() and |
947 | * friends. | |
7106ddf8 PM |
948 | */ |
949 | void call_rcu_after_fork_child(void) | |
950 | { | |
077ff173 | 951 | struct call_rcu_data *crdp, *next; |
7106ddf8 | 952 | |
81ad2e19 PM |
953 | /* Release the mutex. */ |
954 | call_rcu_unlock(&call_rcu_mutex); | |
955 | ||
ad1b9909 LJ |
956 | /* Do nothing when call_rcu() has not been used */ |
957 | if (cds_list_empty(&call_rcu_data_list)) | |
958 | return; | |
959 | ||
7106ddf8 PM |
960 | /* |
961 | * Allocate a new default call_rcu_data structure in order | |
962 | * to get a working call_rcu thread to go with it. | |
963 | */ | |
964 | default_call_rcu_data = NULL; | |
965 | (void)get_default_call_rcu_data(); | |
966 | ||
60af049d LJ |
967 | /* Cleanup call_rcu_data pointers before use */ |
968 | maxcpus_reset(); | |
969 | free(per_cpu_call_rcu_data); | |
618b2595 | 970 | rcu_set_pointer(&per_cpu_call_rcu_data, NULL); |
bd252a04 | 971 | URCU_TLS(thread_call_rcu_data) = NULL; |
60af049d | 972 | |
e85451a1 MD |
973 | /* |
974 | * Dispose of all of the rest of the call_rcu_data structures. | |
975 | * Leftover call_rcu callbacks will be merged into the new | |
976 | * default call_rcu thread queue. | |
977 | */ | |
077ff173 | 978 | cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) { |
7106ddf8 | 979 | if (crdp == default_call_rcu_data) |
077ff173 | 980 | continue; |
2870aa1e | 981 | uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED); |
7106ddf8 | 982 | call_rcu_data_free(crdp); |
b57aee66 PM |
983 | } |
984 | } |