| 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 | |
| 23 | #define _GNU_SOURCE |
| 24 | #define _LGPL_SOURCE |
| 25 | #include <stdio.h> |
| 26 | #include <pthread.h> |
| 27 | #include <signal.h> |
| 28 | #include <assert.h> |
| 29 | #include <stdlib.h> |
| 30 | #include <stdint.h> |
| 31 | #include <string.h> |
| 32 | #include <errno.h> |
| 33 | #include <poll.h> |
| 34 | #include <sys/time.h> |
| 35 | #include <unistd.h> |
| 36 | #include <sched.h> |
| 37 | |
| 38 | #include "config.h" |
| 39 | #include "urcu/wfcqueue.h" |
| 40 | #include "urcu-call-rcu.h" |
| 41 | #include "urcu-pointer.h" |
| 42 | #include "urcu/list.h" |
| 43 | #include "urcu/futex.h" |
| 44 | #include "urcu/tls-compat.h" |
| 45 | #include "urcu-die.h" |
| 46 | |
| 47 | /* Data structure that identifies a call_rcu thread. */ |
| 48 | |
| 49 | struct call_rcu_data { |
| 50 | /* |
| 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. |
| 56 | */ |
| 57 | struct cds_wfcq_tail cbs_tail; |
| 58 | struct cds_wfcq_head cbs_head; |
| 59 | unsigned long flags; |
| 60 | int32_t futex; |
| 61 | unsigned long qlen; /* maintained for debugging. */ |
| 62 | pthread_t tid; |
| 63 | int cpu_affinity; |
| 64 | struct cds_list_head list; |
| 65 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
| 66 | |
| 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 | |
| 77 | /* |
| 78 | * List of all call_rcu_data structures to keep valgrind happy. |
| 79 | * Protected by call_rcu_mutex. |
| 80 | */ |
| 81 | |
| 82 | static CDS_LIST_HEAD(call_rcu_data_list); |
| 83 | |
| 84 | /* Link a thread using call_rcu() to its call_rcu thread. */ |
| 85 | |
| 86 | static DEFINE_URCU_TLS(struct call_rcu_data *, thread_call_rcu_data); |
| 87 | |
| 88 | /* |
| 89 | * Guard call_rcu thread creation and atfork handlers. |
| 90 | */ |
| 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 | |
| 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 | |
| 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) |
| 120 | |
| 121 | /* |
| 122 | * Pointer to array of pointers to per-CPU call_rcu_data structures |
| 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. |
| 127 | */ |
| 128 | |
| 129 | static struct call_rcu_data **per_cpu_call_rcu_data; |
| 130 | static long maxcpus; |
| 131 | |
| 132 | static void maxcpus_reset(void) |
| 133 | { |
| 134 | maxcpus = 0; |
| 135 | } |
| 136 | |
| 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)); |
| 153 | rcu_set_pointer(&per_cpu_call_rcu_data, p); |
| 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 | |
| 162 | #else /* #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
| 163 | |
| 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; |
| 170 | static const long maxcpus = -1; |
| 171 | |
| 172 | static void maxcpus_reset(void) |
| 173 | { |
| 174 | } |
| 175 | |
| 176 | static void alloc_cpu_call_rcu_data(void) |
| 177 | { |
| 178 | } |
| 179 | |
| 180 | #endif /* #else #if defined(HAVE_SYSCONF) && defined(HAVE_SCHED_GETCPU) */ |
| 181 | |
| 182 | /* Acquire the specified pthread mutex. */ |
| 183 | |
| 184 | static void call_rcu_lock(pthread_mutex_t *pmp) |
| 185 | { |
| 186 | int ret; |
| 187 | |
| 188 | ret = pthread_mutex_lock(pmp); |
| 189 | if (ret) |
| 190 | urcu_die(ret); |
| 191 | } |
| 192 | |
| 193 | /* Release the specified pthread mutex. */ |
| 194 | |
| 195 | static void call_rcu_unlock(pthread_mutex_t *pmp) |
| 196 | { |
| 197 | int ret; |
| 198 | |
| 199 | ret = pthread_mutex_unlock(pmp); |
| 200 | if (ret) |
| 201 | urcu_die(ret); |
| 202 | } |
| 203 | |
| 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 | |
| 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(); |
| 242 | if (caa_unlikely(uatomic_read(&crdp->futex) == -1)) { |
| 243 | uatomic_set(&crdp->futex, 0); |
| 244 | futex_async(&crdp->futex, FUTEX_WAKE, 1, |
| 245 | NULL, NULL, 0); |
| 246 | } |
| 247 | } |
| 248 | |
| 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 | |
| 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; |
| 274 | struct call_rcu_data *crdp = (struct call_rcu_data *) arg; |
| 275 | int rt = !!(uatomic_read(&crdp->flags) & URCU_CALL_RCU_RT); |
| 276 | int ret; |
| 277 | |
| 278 | ret = set_thread_cpu_affinity(crdp); |
| 279 | if (ret) |
| 280 | urcu_die(errno); |
| 281 | |
| 282 | /* |
| 283 | * If callbacks take a read-side lock, we need to be registered. |
| 284 | */ |
| 285 | rcu_register_thread(); |
| 286 | |
| 287 | URCU_TLS(thread_call_rcu_data) = crdp; |
| 288 | if (!rt) { |
| 289 | uatomic_dec(&crdp->futex); |
| 290 | /* Decrement futex before reading call_rcu list */ |
| 291 | cmm_smp_mb(); |
| 292 | } |
| 293 | for (;;) { |
| 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; |
| 297 | enum cds_wfcq_ret splice_ret; |
| 298 | |
| 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); |
| 311 | rcu_register_thread(); |
| 312 | } |
| 313 | |
| 314 | cds_wfcq_init(&cbs_tmp_head, &cbs_tmp_tail); |
| 315 | splice_ret = __cds_wfcq_splice_blocking(&cbs_tmp_head, |
| 316 | &cbs_tmp_tail, &crdp->cbs_head, &crdp->cbs_tail); |
| 317 | assert(splice_ret != CDS_WFCQ_RET_WOULDBLOCK); |
| 318 | assert(splice_ret != CDS_WFCQ_RET_DEST_NON_EMPTY); |
| 319 | if (splice_ret != CDS_WFCQ_RET_SRC_EMPTY) { |
| 320 | synchronize_rcu(); |
| 321 | cbcount = 0; |
| 322 | __cds_wfcq_for_each_blocking_safe(&cbs_tmp_head, |
| 323 | &cbs_tmp_tail, cbs, cbs_tmp_n) { |
| 324 | struct rcu_head *rhp; |
| 325 | |
| 326 | rhp = caa_container_of(cbs, |
| 327 | struct rcu_head, next); |
| 328 | rhp->func(rhp); |
| 329 | cbcount++; |
| 330 | } |
| 331 | uatomic_sub(&crdp->qlen, cbcount); |
| 332 | } |
| 333 | if (uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOP) |
| 334 | break; |
| 335 | rcu_thread_offline(); |
| 336 | if (!rt) { |
| 337 | if (cds_wfcq_empty(&crdp->cbs_head, |
| 338 | &crdp->cbs_tail)) { |
| 339 | call_rcu_wait(crdp); |
| 340 | poll(NULL, 0, 10); |
| 341 | uatomic_dec(&crdp->futex); |
| 342 | /* |
| 343 | * Decrement futex before reading |
| 344 | * call_rcu list. |
| 345 | */ |
| 346 | cmm_smp_mb(); |
| 347 | } else { |
| 348 | poll(NULL, 0, 10); |
| 349 | } |
| 350 | } else { |
| 351 | poll(NULL, 0, 10); |
| 352 | } |
| 353 | rcu_thread_online(); |
| 354 | } |
| 355 | if (!rt) { |
| 356 | /* |
| 357 | * Read call_rcu list before write futex. |
| 358 | */ |
| 359 | cmm_smp_mb(); |
| 360 | uatomic_set(&crdp->futex, 0); |
| 361 | } |
| 362 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOPPED); |
| 363 | rcu_unregister_thread(); |
| 364 | return NULL; |
| 365 | } |
| 366 | |
| 367 | /* |
| 368 | * Create both a call_rcu thread and the corresponding call_rcu_data |
| 369 | * structure, linking the structure in as specified. Caller must hold |
| 370 | * call_rcu_mutex. |
| 371 | */ |
| 372 | |
| 373 | static void call_rcu_data_init(struct call_rcu_data **crdpp, |
| 374 | unsigned long flags, |
| 375 | int cpu_affinity) |
| 376 | { |
| 377 | struct call_rcu_data *crdp; |
| 378 | int ret; |
| 379 | |
| 380 | crdp = malloc(sizeof(*crdp)); |
| 381 | if (crdp == NULL) |
| 382 | urcu_die(errno); |
| 383 | memset(crdp, '\0', sizeof(*crdp)); |
| 384 | cds_wfcq_init(&crdp->cbs_head, &crdp->cbs_tail); |
| 385 | crdp->qlen = 0; |
| 386 | crdp->futex = 0; |
| 387 | crdp->flags = flags; |
| 388 | cds_list_add(&crdp->list, &call_rcu_data_list); |
| 389 | crdp->cpu_affinity = cpu_affinity; |
| 390 | cmm_smp_mb(); /* Structure initialized before pointer is planted. */ |
| 391 | *crdpp = crdp; |
| 392 | ret = pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp); |
| 393 | if (ret) |
| 394 | urcu_die(ret); |
| 395 | } |
| 396 | |
| 397 | /* |
| 398 | * Return a pointer to the call_rcu_data structure for the specified |
| 399 | * CPU, returning NULL if there is none. We cannot automatically |
| 400 | * created it because the platform we are running on might not define |
| 401 | * urcu_sched_getcpu(). |
| 402 | * |
| 403 | * The call to this function and use of the returned call_rcu_data |
| 404 | * should be protected by RCU read-side lock. |
| 405 | */ |
| 406 | |
| 407 | struct call_rcu_data *get_cpu_call_rcu_data(int cpu) |
| 408 | { |
| 409 | static int warned = 0; |
| 410 | struct call_rcu_data **pcpu_crdp; |
| 411 | |
| 412 | pcpu_crdp = rcu_dereference(per_cpu_call_rcu_data); |
| 413 | if (pcpu_crdp == NULL) |
| 414 | return NULL; |
| 415 | if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) { |
| 416 | fprintf(stderr, "[error] liburcu: get CPU # out of range\n"); |
| 417 | warned = 1; |
| 418 | } |
| 419 | if (cpu < 0 || maxcpus <= cpu) |
| 420 | return NULL; |
| 421 | return rcu_dereference(pcpu_crdp[cpu]); |
| 422 | } |
| 423 | |
| 424 | /* |
| 425 | * Return the tid corresponding to the call_rcu thread whose |
| 426 | * call_rcu_data structure is specified. |
| 427 | */ |
| 428 | |
| 429 | pthread_t get_call_rcu_thread(struct call_rcu_data *crdp) |
| 430 | { |
| 431 | return crdp->tid; |
| 432 | } |
| 433 | |
| 434 | /* |
| 435 | * Create a call_rcu_data structure (with thread) and return a pointer. |
| 436 | */ |
| 437 | |
| 438 | static struct call_rcu_data *__create_call_rcu_data(unsigned long flags, |
| 439 | int cpu_affinity) |
| 440 | { |
| 441 | struct call_rcu_data *crdp; |
| 442 | |
| 443 | call_rcu_data_init(&crdp, flags, cpu_affinity); |
| 444 | return crdp; |
| 445 | } |
| 446 | |
| 447 | struct call_rcu_data *create_call_rcu_data(unsigned long flags, |
| 448 | int cpu_affinity) |
| 449 | { |
| 450 | struct call_rcu_data *crdp; |
| 451 | |
| 452 | call_rcu_lock(&call_rcu_mutex); |
| 453 | crdp = __create_call_rcu_data(flags, cpu_affinity); |
| 454 | call_rcu_unlock(&call_rcu_mutex); |
| 455 | return crdp; |
| 456 | } |
| 457 | |
| 458 | /* |
| 459 | * Set the specified CPU to use the specified call_rcu_data structure. |
| 460 | * |
| 461 | * Use NULL to remove a CPU's call_rcu_data structure, but it is |
| 462 | * the caller's responsibility to dispose of the removed structure. |
| 463 | * Use get_cpu_call_rcu_data() to obtain a pointer to the old structure |
| 464 | * (prior to NULLing it out, of course). |
| 465 | * |
| 466 | * The caller must wait for a grace-period to pass between return from |
| 467 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the |
| 468 | * previous call rcu data as argument. |
| 469 | */ |
| 470 | |
| 471 | int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp) |
| 472 | { |
| 473 | static int warned = 0; |
| 474 | |
| 475 | call_rcu_lock(&call_rcu_mutex); |
| 476 | alloc_cpu_call_rcu_data(); |
| 477 | if (cpu < 0 || maxcpus <= cpu) { |
| 478 | if (!warned) { |
| 479 | fprintf(stderr, "[error] liburcu: set CPU # out of range\n"); |
| 480 | warned = 1; |
| 481 | } |
| 482 | call_rcu_unlock(&call_rcu_mutex); |
| 483 | errno = EINVAL; |
| 484 | return -EINVAL; |
| 485 | } |
| 486 | |
| 487 | if (per_cpu_call_rcu_data == NULL) { |
| 488 | call_rcu_unlock(&call_rcu_mutex); |
| 489 | errno = ENOMEM; |
| 490 | return -ENOMEM; |
| 491 | } |
| 492 | |
| 493 | if (per_cpu_call_rcu_data[cpu] != NULL && crdp != NULL) { |
| 494 | call_rcu_unlock(&call_rcu_mutex); |
| 495 | errno = EEXIST; |
| 496 | return -EEXIST; |
| 497 | } |
| 498 | |
| 499 | rcu_set_pointer(&per_cpu_call_rcu_data[cpu], crdp); |
| 500 | call_rcu_unlock(&call_rcu_mutex); |
| 501 | return 0; |
| 502 | } |
| 503 | |
| 504 | /* |
| 505 | * Return a pointer to the default call_rcu_data structure, creating |
| 506 | * one if need be. Because we never free call_rcu_data structures, |
| 507 | * we don't need to be in an RCU read-side critical section. |
| 508 | */ |
| 509 | |
| 510 | struct call_rcu_data *get_default_call_rcu_data(void) |
| 511 | { |
| 512 | if (default_call_rcu_data != NULL) |
| 513 | return rcu_dereference(default_call_rcu_data); |
| 514 | call_rcu_lock(&call_rcu_mutex); |
| 515 | if (default_call_rcu_data != NULL) { |
| 516 | call_rcu_unlock(&call_rcu_mutex); |
| 517 | return default_call_rcu_data; |
| 518 | } |
| 519 | call_rcu_data_init(&default_call_rcu_data, 0, -1); |
| 520 | call_rcu_unlock(&call_rcu_mutex); |
| 521 | return default_call_rcu_data; |
| 522 | } |
| 523 | |
| 524 | /* |
| 525 | * Return the call_rcu_data structure that applies to the currently |
| 526 | * running thread. Any call_rcu_data structure assigned specifically |
| 527 | * to this thread has first priority, followed by any call_rcu_data |
| 528 | * structure assigned to the CPU on which the thread is running, |
| 529 | * followed by the default call_rcu_data structure. If there is not |
| 530 | * yet a default call_rcu_data structure, one will be created. |
| 531 | * |
| 532 | * Calls to this function and use of the returned call_rcu_data should |
| 533 | * be protected by RCU read-side lock. |
| 534 | */ |
| 535 | struct call_rcu_data *get_call_rcu_data(void) |
| 536 | { |
| 537 | struct call_rcu_data *crd; |
| 538 | |
| 539 | if (URCU_TLS(thread_call_rcu_data) != NULL) |
| 540 | return URCU_TLS(thread_call_rcu_data); |
| 541 | |
| 542 | if (maxcpus > 0) { |
| 543 | crd = get_cpu_call_rcu_data(urcu_sched_getcpu()); |
| 544 | if (crd) |
| 545 | return crd; |
| 546 | } |
| 547 | |
| 548 | return get_default_call_rcu_data(); |
| 549 | } |
| 550 | |
| 551 | /* |
| 552 | * Return a pointer to this task's call_rcu_data if there is one. |
| 553 | */ |
| 554 | |
| 555 | struct call_rcu_data *get_thread_call_rcu_data(void) |
| 556 | { |
| 557 | return URCU_TLS(thread_call_rcu_data); |
| 558 | } |
| 559 | |
| 560 | /* |
| 561 | * Set this task's call_rcu_data structure as specified, regardless |
| 562 | * of whether or not this task already had one. (This allows switching |
| 563 | * to and from real-time call_rcu threads, for example.) |
| 564 | * |
| 565 | * Use NULL to remove a thread's call_rcu_data structure, but it is |
| 566 | * the caller's responsibility to dispose of the removed structure. |
| 567 | * Use get_thread_call_rcu_data() to obtain a pointer to the old structure |
| 568 | * (prior to NULLing it out, of course). |
| 569 | */ |
| 570 | |
| 571 | void set_thread_call_rcu_data(struct call_rcu_data *crdp) |
| 572 | { |
| 573 | URCU_TLS(thread_call_rcu_data) = crdp; |
| 574 | } |
| 575 | |
| 576 | /* |
| 577 | * Create a separate call_rcu thread for each CPU. This does not |
| 578 | * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data() |
| 579 | * function if you want that behavior. Should be paired with |
| 580 | * free_all_cpu_call_rcu_data() to teardown these call_rcu worker |
| 581 | * threads. |
| 582 | */ |
| 583 | |
| 584 | int create_all_cpu_call_rcu_data(unsigned long flags) |
| 585 | { |
| 586 | int i; |
| 587 | struct call_rcu_data *crdp; |
| 588 | int ret; |
| 589 | |
| 590 | call_rcu_lock(&call_rcu_mutex); |
| 591 | alloc_cpu_call_rcu_data(); |
| 592 | call_rcu_unlock(&call_rcu_mutex); |
| 593 | if (maxcpus <= 0) { |
| 594 | errno = EINVAL; |
| 595 | return -EINVAL; |
| 596 | } |
| 597 | if (per_cpu_call_rcu_data == NULL) { |
| 598 | errno = ENOMEM; |
| 599 | return -ENOMEM; |
| 600 | } |
| 601 | for (i = 0; i < maxcpus; i++) { |
| 602 | call_rcu_lock(&call_rcu_mutex); |
| 603 | if (get_cpu_call_rcu_data(i)) { |
| 604 | call_rcu_unlock(&call_rcu_mutex); |
| 605 | continue; |
| 606 | } |
| 607 | crdp = __create_call_rcu_data(flags, i); |
| 608 | if (crdp == NULL) { |
| 609 | call_rcu_unlock(&call_rcu_mutex); |
| 610 | errno = ENOMEM; |
| 611 | return -ENOMEM; |
| 612 | } |
| 613 | call_rcu_unlock(&call_rcu_mutex); |
| 614 | if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) { |
| 615 | call_rcu_data_free(crdp); |
| 616 | |
| 617 | /* it has been created by other thread */ |
| 618 | if (ret == -EEXIST) |
| 619 | continue; |
| 620 | |
| 621 | return ret; |
| 622 | } |
| 623 | } |
| 624 | return 0; |
| 625 | } |
| 626 | |
| 627 | /* |
| 628 | * Wake up the call_rcu thread corresponding to the specified |
| 629 | * call_rcu_data structure. |
| 630 | */ |
| 631 | static void wake_call_rcu_thread(struct call_rcu_data *crdp) |
| 632 | { |
| 633 | if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT)) |
| 634 | call_rcu_wake_up(crdp); |
| 635 | } |
| 636 | |
| 637 | static void _call_rcu(struct rcu_head *head, |
| 638 | void (*func)(struct rcu_head *head), |
| 639 | struct call_rcu_data *crdp) |
| 640 | { |
| 641 | cds_wfcq_node_init(&head->next); |
| 642 | head->func = func; |
| 643 | cds_wfcq_enqueue(&crdp->cbs_head, &crdp->cbs_tail, &head->next); |
| 644 | uatomic_inc(&crdp->qlen); |
| 645 | wake_call_rcu_thread(crdp); |
| 646 | } |
| 647 | |
| 648 | /* |
| 649 | * Schedule a function to be invoked after a following grace period. |
| 650 | * This is the only function that must be called -- the others are |
| 651 | * only present to allow applications to tune their use of RCU for |
| 652 | * maximum performance. |
| 653 | * |
| 654 | * Note that unless a call_rcu thread has not already been created, |
| 655 | * the first invocation of call_rcu() will create one. So, if you |
| 656 | * need the first invocation of call_rcu() to be fast, make sure |
| 657 | * to create a call_rcu thread first. One way to accomplish this is |
| 658 | * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data(). |
| 659 | * |
| 660 | * call_rcu must be called by registered RCU read-side threads. |
| 661 | */ |
| 662 | void call_rcu(struct rcu_head *head, |
| 663 | void (*func)(struct rcu_head *head)) |
| 664 | { |
| 665 | struct call_rcu_data *crdp; |
| 666 | |
| 667 | /* Holding rcu read-side lock across use of per-cpu crdp */ |
| 668 | rcu_read_lock(); |
| 669 | crdp = get_call_rcu_data(); |
| 670 | _call_rcu(head, func, crdp); |
| 671 | rcu_read_unlock(); |
| 672 | } |
| 673 | |
| 674 | /* |
| 675 | * Free up the specified call_rcu_data structure, terminating the |
| 676 | * associated call_rcu thread. The caller must have previously |
| 677 | * removed the call_rcu_data structure from per-thread or per-CPU |
| 678 | * usage. For example, set_cpu_call_rcu_data(cpu, NULL) for per-CPU |
| 679 | * call_rcu_data structures or set_thread_call_rcu_data(NULL) for |
| 680 | * per-thread call_rcu_data structures. |
| 681 | * |
| 682 | * We silently refuse to free up the default call_rcu_data structure |
| 683 | * because that is where we put any leftover callbacks. Note that |
| 684 | * the possibility of self-spawning callbacks makes it impossible |
| 685 | * to execute all the callbacks in finite time without putting any |
| 686 | * newly spawned callbacks somewhere else. The "somewhere else" of |
| 687 | * last resort is the default call_rcu_data structure. |
| 688 | * |
| 689 | * We also silently refuse to free NULL pointers. This simplifies |
| 690 | * the calling code. |
| 691 | * |
| 692 | * The caller must wait for a grace-period to pass between return from |
| 693 | * set_cpu_call_rcu_data() and call to call_rcu_data_free() passing the |
| 694 | * previous call rcu data as argument. |
| 695 | * |
| 696 | * Note: introducing __cds_wfcq_splice_blocking() in this function fixed |
| 697 | * a list corruption bug in the 0.7.x series. The equivalent fix |
| 698 | * appeared in 0.6.8 for the stable-0.6 branch. |
| 699 | */ |
| 700 | void call_rcu_data_free(struct call_rcu_data *crdp) |
| 701 | { |
| 702 | if (crdp == NULL || crdp == default_call_rcu_data) { |
| 703 | return; |
| 704 | } |
| 705 | if ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) { |
| 706 | uatomic_or(&crdp->flags, URCU_CALL_RCU_STOP); |
| 707 | wake_call_rcu_thread(crdp); |
| 708 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_STOPPED) == 0) |
| 709 | poll(NULL, 0, 1); |
| 710 | } |
| 711 | if (!cds_wfcq_empty(&crdp->cbs_head, &crdp->cbs_tail)) { |
| 712 | /* Create default call rcu data if need be */ |
| 713 | (void) get_default_call_rcu_data(); |
| 714 | __cds_wfcq_splice_blocking(&default_call_rcu_data->cbs_head, |
| 715 | &default_call_rcu_data->cbs_tail, |
| 716 | &crdp->cbs_head, &crdp->cbs_tail); |
| 717 | uatomic_add(&default_call_rcu_data->qlen, |
| 718 | uatomic_read(&crdp->qlen)); |
| 719 | wake_call_rcu_thread(default_call_rcu_data); |
| 720 | } |
| 721 | |
| 722 | call_rcu_lock(&call_rcu_mutex); |
| 723 | cds_list_del(&crdp->list); |
| 724 | call_rcu_unlock(&call_rcu_mutex); |
| 725 | |
| 726 | free(crdp); |
| 727 | } |
| 728 | |
| 729 | /* |
| 730 | * Clean up all the per-CPU call_rcu threads. |
| 731 | */ |
| 732 | void free_all_cpu_call_rcu_data(void) |
| 733 | { |
| 734 | int cpu; |
| 735 | struct call_rcu_data **crdp; |
| 736 | static int warned = 0; |
| 737 | |
| 738 | if (maxcpus <= 0) |
| 739 | return; |
| 740 | |
| 741 | crdp = malloc(sizeof(*crdp) * maxcpus); |
| 742 | if (!crdp) { |
| 743 | if (!warned) { |
| 744 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); |
| 745 | } |
| 746 | warned = 1; |
| 747 | return; |
| 748 | } |
| 749 | |
| 750 | for (cpu = 0; cpu < maxcpus; cpu++) { |
| 751 | crdp[cpu] = get_cpu_call_rcu_data(cpu); |
| 752 | if (crdp[cpu] == NULL) |
| 753 | continue; |
| 754 | set_cpu_call_rcu_data(cpu, NULL); |
| 755 | } |
| 756 | /* |
| 757 | * Wait for call_rcu sites acting as RCU readers of the |
| 758 | * call_rcu_data to become quiescent. |
| 759 | */ |
| 760 | synchronize_rcu(); |
| 761 | for (cpu = 0; cpu < maxcpus; cpu++) { |
| 762 | if (crdp[cpu] == NULL) |
| 763 | continue; |
| 764 | call_rcu_data_free(crdp[cpu]); |
| 765 | } |
| 766 | free(crdp); |
| 767 | } |
| 768 | |
| 769 | static |
| 770 | void _rcu_barrier_complete(struct rcu_head *head) |
| 771 | { |
| 772 | struct call_rcu_completion_work *work; |
| 773 | struct call_rcu_completion *completion; |
| 774 | |
| 775 | work = caa_container_of(head, struct call_rcu_completion_work, head); |
| 776 | completion = work->completion; |
| 777 | uatomic_dec(&completion->barrier_count); |
| 778 | call_rcu_completion_wake_up(completion); |
| 779 | free(work); |
| 780 | } |
| 781 | |
| 782 | /* |
| 783 | * Wait for all in-flight call_rcu callbacks to complete execution. |
| 784 | */ |
| 785 | void rcu_barrier(void) |
| 786 | { |
| 787 | struct call_rcu_data *crdp; |
| 788 | struct call_rcu_completion completion; |
| 789 | int count = 0; |
| 790 | int was_online; |
| 791 | |
| 792 | /* Put in offline state in QSBR. */ |
| 793 | was_online = rcu_read_ongoing(); |
| 794 | if (was_online) |
| 795 | rcu_thread_offline(); |
| 796 | /* |
| 797 | * Calling a rcu_barrier() within a RCU read-side critical |
| 798 | * section is an error. |
| 799 | */ |
| 800 | if (rcu_read_ongoing()) { |
| 801 | static int warned = 0; |
| 802 | |
| 803 | if (!warned) { |
| 804 | fprintf(stderr, "[error] liburcu: rcu_barrier() called from within RCU read-side critical section.\n"); |
| 805 | } |
| 806 | warned = 1; |
| 807 | goto online; |
| 808 | } |
| 809 | |
| 810 | call_rcu_lock(&call_rcu_mutex); |
| 811 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) |
| 812 | count++; |
| 813 | |
| 814 | completion.barrier_count = count; |
| 815 | |
| 816 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { |
| 817 | struct call_rcu_completion_work *work; |
| 818 | |
| 819 | work = calloc(sizeof(*work), 1); |
| 820 | if (!work) |
| 821 | urcu_die(errno); |
| 822 | work->completion = &completion; |
| 823 | _call_rcu(&work->head, _rcu_barrier_complete, crdp); |
| 824 | } |
| 825 | call_rcu_unlock(&call_rcu_mutex); |
| 826 | |
| 827 | /* Wait for them */ |
| 828 | for (;;) { |
| 829 | uatomic_dec(&completion.futex); |
| 830 | /* Decrement futex before reading barrier_count */ |
| 831 | cmm_smp_mb(); |
| 832 | if (!uatomic_read(&completion.barrier_count)) |
| 833 | break; |
| 834 | call_rcu_completion_wait(&completion); |
| 835 | } |
| 836 | online: |
| 837 | if (was_online) |
| 838 | rcu_thread_online(); |
| 839 | } |
| 840 | |
| 841 | /* |
| 842 | * Acquire the call_rcu_mutex in order to ensure that the child sees |
| 843 | * all of the call_rcu() data structures in a consistent state. Ensure |
| 844 | * that all call_rcu threads are in a quiescent state across fork. |
| 845 | * Suitable for pthread_atfork() and friends. |
| 846 | */ |
| 847 | void call_rcu_before_fork(void) |
| 848 | { |
| 849 | struct call_rcu_data *crdp; |
| 850 | |
| 851 | call_rcu_lock(&call_rcu_mutex); |
| 852 | |
| 853 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { |
| 854 | uatomic_or(&crdp->flags, URCU_CALL_RCU_PAUSE); |
| 855 | cmm_smp_mb__after_uatomic_or(); |
| 856 | wake_call_rcu_thread(crdp); |
| 857 | } |
| 858 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) { |
| 859 | while ((uatomic_read(&crdp->flags) & URCU_CALL_RCU_PAUSED) == 0) |
| 860 | poll(NULL, 0, 1); |
| 861 | } |
| 862 | } |
| 863 | |
| 864 | /* |
| 865 | * Clean up call_rcu data structures in the parent of a successful fork() |
| 866 | * that is not followed by exec() in the child. Suitable for |
| 867 | * pthread_atfork() and friends. |
| 868 | */ |
| 869 | void call_rcu_after_fork_parent(void) |
| 870 | { |
| 871 | struct call_rcu_data *crdp; |
| 872 | |
| 873 | cds_list_for_each_entry(crdp, &call_rcu_data_list, list) |
| 874 | uatomic_and(&crdp->flags, ~URCU_CALL_RCU_PAUSE); |
| 875 | call_rcu_unlock(&call_rcu_mutex); |
| 876 | } |
| 877 | |
| 878 | /* |
| 879 | * Clean up call_rcu data structures in the child of a successful fork() |
| 880 | * that is not followed by exec(). Suitable for pthread_atfork() and |
| 881 | * friends. |
| 882 | */ |
| 883 | void call_rcu_after_fork_child(void) |
| 884 | { |
| 885 | struct call_rcu_data *crdp, *next; |
| 886 | |
| 887 | /* Release the mutex. */ |
| 888 | call_rcu_unlock(&call_rcu_mutex); |
| 889 | |
| 890 | /* Do nothing when call_rcu() has not been used */ |
| 891 | if (cds_list_empty(&call_rcu_data_list)) |
| 892 | return; |
| 893 | |
| 894 | /* |
| 895 | * Allocate a new default call_rcu_data structure in order |
| 896 | * to get a working call_rcu thread to go with it. |
| 897 | */ |
| 898 | default_call_rcu_data = NULL; |
| 899 | (void)get_default_call_rcu_data(); |
| 900 | |
| 901 | /* Cleanup call_rcu_data pointers before use */ |
| 902 | maxcpus_reset(); |
| 903 | free(per_cpu_call_rcu_data); |
| 904 | rcu_set_pointer(&per_cpu_call_rcu_data, NULL); |
| 905 | URCU_TLS(thread_call_rcu_data) = NULL; |
| 906 | |
| 907 | /* |
| 908 | * Dispose of all of the rest of the call_rcu_data structures. |
| 909 | * Leftover call_rcu callbacks will be merged into the new |
| 910 | * default call_rcu thread queue. |
| 911 | */ |
| 912 | cds_list_for_each_entry_safe(crdp, next, &call_rcu_data_list, list) { |
| 913 | if (crdp == default_call_rcu_data) |
| 914 | continue; |
| 915 | uatomic_set(&crdp->flags, URCU_CALL_RCU_STOPPED); |
| 916 | call_rcu_data_free(crdp); |
| 917 | } |
| 918 | } |