| 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 | #include <stdio.h> |
| 24 | #include <pthread.h> |
| 25 | #include <signal.h> |
| 26 | #include <assert.h> |
| 27 | #include <stdlib.h> |
| 28 | #include <string.h> |
| 29 | #include <errno.h> |
| 30 | #include <poll.h> |
| 31 | #include <sys/time.h> |
| 32 | #include <syscall.h> |
| 33 | #include <unistd.h> |
| 34 | |
| 35 | #include "config.h" |
| 36 | #include "urcu/wfqueue.h" |
| 37 | #include "urcu-call-rcu.h" |
| 38 | #include "urcu-pointer.h" |
| 39 | #include "urcu/list.h" |
| 40 | |
| 41 | /* Data structure that identifies a call_rcu thread. */ |
| 42 | |
| 43 | struct call_rcu_data { |
| 44 | struct cds_wfq_queue cbs; |
| 45 | unsigned long flags; |
| 46 | pthread_mutex_t mtx; |
| 47 | pthread_cond_t cond; |
| 48 | unsigned long qlen; |
| 49 | pthread_t tid; |
| 50 | struct cds_list_head list; |
| 51 | } __attribute__((aligned(CAA_CACHE_LINE_SIZE))); |
| 52 | |
| 53 | /* |
| 54 | * List of all call_rcu_data structures to keep valgrind happy. |
| 55 | * Protected by call_rcu_mutex. |
| 56 | */ |
| 57 | |
| 58 | CDS_LIST_HEAD(call_rcu_data_list); |
| 59 | |
| 60 | /* Link a thread using call_rcu() to its call_rcu thread. */ |
| 61 | |
| 62 | static __thread struct call_rcu_data *thread_call_rcu_data; |
| 63 | |
| 64 | /* Guard call_rcu thread creation. */ |
| 65 | |
| 66 | static pthread_mutex_t call_rcu_mutex = PTHREAD_MUTEX_INITIALIZER; |
| 67 | |
| 68 | /* If a given thread does not have its own call_rcu thread, this is default. */ |
| 69 | |
| 70 | static struct call_rcu_data *default_call_rcu_data; |
| 71 | |
| 72 | extern void synchronize_rcu(void); |
| 73 | |
| 74 | /* |
| 75 | * If the sched_getcpu() and sysconf(_SC_NPROCESSORS_CONF) calls are |
| 76 | * available, then we can have call_rcu threads assigned to individual |
| 77 | * CPUs rather than only to specific threads. |
| 78 | */ |
| 79 | |
| 80 | #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) |
| 81 | |
| 82 | /* |
| 83 | * Pointer to array of pointers to per-CPU call_rcu_data structures |
| 84 | * and # CPUs. |
| 85 | */ |
| 86 | |
| 87 | static struct call_rcu_data **per_cpu_call_rcu_data; |
| 88 | static long maxcpus; |
| 89 | |
| 90 | /* Allocate the array if it has not already been allocated. */ |
| 91 | |
| 92 | static void alloc_cpu_call_rcu_data(void) |
| 93 | { |
| 94 | struct call_rcu_data **p; |
| 95 | static int warned = 0; |
| 96 | |
| 97 | if (maxcpus != 0) |
| 98 | return; |
| 99 | maxcpus = sysconf(_SC_NPROCESSORS_CONF); |
| 100 | if (maxcpus <= 0) { |
| 101 | return; |
| 102 | } |
| 103 | p = malloc(maxcpus * sizeof(*per_cpu_call_rcu_data)); |
| 104 | if (p != NULL) { |
| 105 | memset(p, '\0', maxcpus * sizeof(*per_cpu_call_rcu_data)); |
| 106 | per_cpu_call_rcu_data = p; |
| 107 | } else { |
| 108 | if (!warned) { |
| 109 | fprintf(stderr, "[error] liburcu: unable to allocate per-CPU pointer array\n"); |
| 110 | } |
| 111 | warned = 1; |
| 112 | } |
| 113 | } |
| 114 | |
| 115 | #else /* #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ |
| 116 | |
| 117 | static const struct call_rcu_data **per_cpu_call_rcu_data = NULL; |
| 118 | static const long maxcpus = -1; |
| 119 | |
| 120 | static void alloc_cpu_call_rcu_data(void) |
| 121 | { |
| 122 | } |
| 123 | |
| 124 | static int sched_getcpu(void) |
| 125 | { |
| 126 | return -1; |
| 127 | } |
| 128 | |
| 129 | #endif /* #else #if defined(HAVE_SCHED_GETCPU) && defined(HAVE_SYSCONF) */ |
| 130 | |
| 131 | /* Acquire the specified pthread mutex. */ |
| 132 | |
| 133 | static void call_rcu_lock(pthread_mutex_t *pmp) |
| 134 | { |
| 135 | if (pthread_mutex_lock(pmp) != 0) { |
| 136 | perror("pthread_mutex_lock"); |
| 137 | exit(-1); |
| 138 | } |
| 139 | } |
| 140 | |
| 141 | /* Release the specified pthread mutex. */ |
| 142 | |
| 143 | static void call_rcu_unlock(pthread_mutex_t *pmp) |
| 144 | { |
| 145 | if (pthread_mutex_unlock(pmp) != 0) { |
| 146 | perror("pthread_mutex_unlock"); |
| 147 | exit(-1); |
| 148 | } |
| 149 | } |
| 150 | |
| 151 | /* This is the code run by each call_rcu thread. */ |
| 152 | |
| 153 | static void *call_rcu_thread(void *arg) |
| 154 | { |
| 155 | unsigned long cbcount; |
| 156 | struct cds_wfq_node *cbs; |
| 157 | struct cds_wfq_node **cbs_tail; |
| 158 | struct call_rcu_data *crdp = (struct call_rcu_data *)arg; |
| 159 | struct rcu_head *rhp; |
| 160 | |
| 161 | thread_call_rcu_data = crdp; |
| 162 | for (;;) { |
| 163 | if (&crdp->cbs.head != _CMM_LOAD_SHARED(crdp->cbs.tail)) { |
| 164 | while ((cbs = _CMM_LOAD_SHARED(crdp->cbs.head)) == NULL) |
| 165 | poll(NULL, 0, 1); |
| 166 | _CMM_STORE_SHARED(crdp->cbs.head, NULL); |
| 167 | cbs_tail = (struct cds_wfq_node **) |
| 168 | uatomic_xchg(&crdp->cbs.tail, &crdp->cbs.head); |
| 169 | synchronize_rcu(); |
| 170 | cbcount = 0; |
| 171 | do { |
| 172 | while (cbs->next == NULL && |
| 173 | &cbs->next != cbs_tail) |
| 174 | poll(NULL, 0, 1); |
| 175 | if (cbs == &crdp->cbs.dummy) { |
| 176 | cbs = cbs->next; |
| 177 | continue; |
| 178 | } |
| 179 | rhp = (struct rcu_head *)cbs; |
| 180 | cbs = cbs->next; |
| 181 | rhp->func(rhp); |
| 182 | cbcount++; |
| 183 | } while (cbs != NULL); |
| 184 | uatomic_sub(&crdp->qlen, cbcount); |
| 185 | } |
| 186 | if (crdp->flags & URCU_CALL_RCU_RT) |
| 187 | poll(NULL, 0, 10); |
| 188 | else { |
| 189 | call_rcu_lock(&crdp->mtx); |
| 190 | _CMM_STORE_SHARED(crdp->flags, |
| 191 | crdp->flags & ~URCU_CALL_RCU_RUNNING); |
| 192 | if (&crdp->cbs.head == |
| 193 | _CMM_LOAD_SHARED(crdp->cbs.tail) && |
| 194 | pthread_cond_wait(&crdp->cond, &crdp->mtx) != 0) { |
| 195 | perror("pthread_cond_wait"); |
| 196 | exit(-1); |
| 197 | } |
| 198 | _CMM_STORE_SHARED(crdp->flags, |
| 199 | crdp->flags | URCU_CALL_RCU_RUNNING); |
| 200 | poll(NULL, 0, 10); |
| 201 | call_rcu_unlock(&crdp->mtx); |
| 202 | } |
| 203 | } |
| 204 | return NULL; /* NOTREACHED */ |
| 205 | } |
| 206 | |
| 207 | /* |
| 208 | * Create both a call_rcu thread and the corresponding call_rcu_data |
| 209 | * structure, linking the structure in as specified. Caller must hold |
| 210 | * call_rcu_mutex. |
| 211 | */ |
| 212 | |
| 213 | static void call_rcu_data_init(struct call_rcu_data **crdpp, |
| 214 | unsigned long flags) |
| 215 | { |
| 216 | struct call_rcu_data *crdp; |
| 217 | |
| 218 | crdp = malloc(sizeof(*crdp)); |
| 219 | if (crdp == NULL) { |
| 220 | fprintf(stderr, "Out of memory.\n"); |
| 221 | exit(-1); |
| 222 | } |
| 223 | memset(crdp, '\0', sizeof(*crdp)); |
| 224 | cds_wfq_init(&crdp->cbs); |
| 225 | crdp->qlen = 0; |
| 226 | if (pthread_mutex_init(&crdp->mtx, NULL) != 0) { |
| 227 | perror("pthread_mutex_init"); |
| 228 | exit(-1); |
| 229 | } |
| 230 | if (pthread_cond_init(&crdp->cond, NULL) != 0) { |
| 231 | perror("pthread_cond_init"); |
| 232 | exit(-1); |
| 233 | } |
| 234 | crdp->flags = flags | URCU_CALL_RCU_RUNNING; |
| 235 | cds_list_add(&crdp->list, &call_rcu_data_list); |
| 236 | cmm_smp_mb(); /* Structure initialized before pointer is planted. */ |
| 237 | *crdpp = crdp; |
| 238 | if (pthread_create(&crdp->tid, NULL, call_rcu_thread, crdp) != 0) { |
| 239 | perror("pthread_create"); |
| 240 | exit(-1); |
| 241 | } |
| 242 | } |
| 243 | |
| 244 | /* |
| 245 | * Return a pointer to the call_rcu_data structure for the specified |
| 246 | * CPU, returning NULL if there is none. We cannot automatically |
| 247 | * created it because the platform we are running on might not define |
| 248 | * sched_getcpu(). |
| 249 | */ |
| 250 | |
| 251 | struct call_rcu_data *get_cpu_call_rcu_data(int cpu) |
| 252 | { |
| 253 | static int warned = 0; |
| 254 | |
| 255 | if (per_cpu_call_rcu_data == NULL) |
| 256 | return NULL; |
| 257 | if (!warned && maxcpus > 0 && (cpu < 0 || maxcpus <= cpu)) { |
| 258 | fprintf(stderr, "[error] liburcu: get CPU # out of range\n"); |
| 259 | warned = 1; |
| 260 | } |
| 261 | if (cpu < 0 || maxcpus <= cpu) |
| 262 | return NULL; |
| 263 | return per_cpu_call_rcu_data[cpu]; |
| 264 | } |
| 265 | |
| 266 | /* |
| 267 | * Return the tid corresponding to the call_rcu thread whose |
| 268 | * call_rcu_data structure is specified. |
| 269 | */ |
| 270 | |
| 271 | pthread_t get_call_rcu_thread(struct call_rcu_data *crdp) |
| 272 | { |
| 273 | return crdp->tid; |
| 274 | } |
| 275 | |
| 276 | /* |
| 277 | * Create a call_rcu_data structure (with thread) and return a pointer. |
| 278 | */ |
| 279 | |
| 280 | static struct call_rcu_data *__create_call_rcu_data(unsigned long flags) |
| 281 | { |
| 282 | struct call_rcu_data *crdp; |
| 283 | |
| 284 | call_rcu_data_init(&crdp, flags); |
| 285 | return crdp; |
| 286 | } |
| 287 | |
| 288 | struct call_rcu_data *create_call_rcu_data(unsigned long flags) |
| 289 | { |
| 290 | struct call_rcu_data *crdp; |
| 291 | |
| 292 | call_rcu_lock(&call_rcu_mutex); |
| 293 | crdp = __create_call_rcu_data(flags); |
| 294 | call_rcu_unlock(&call_rcu_mutex); |
| 295 | return crdp; |
| 296 | } |
| 297 | |
| 298 | /* |
| 299 | * Set the specified CPU to use the specified call_rcu_data structure. |
| 300 | */ |
| 301 | |
| 302 | int set_cpu_call_rcu_data(int cpu, struct call_rcu_data *crdp) |
| 303 | { |
| 304 | int warned = 0; |
| 305 | |
| 306 | call_rcu_lock(&call_rcu_mutex); |
| 307 | if (cpu < 0 || maxcpus <= cpu) { |
| 308 | if (!warned) { |
| 309 | fprintf(stderr, "[error] liburcu: set CPU # out of range\n"); |
| 310 | warned = 1; |
| 311 | } |
| 312 | call_rcu_unlock(&call_rcu_mutex); |
| 313 | errno = EINVAL; |
| 314 | return -EINVAL; |
| 315 | } |
| 316 | alloc_cpu_call_rcu_data(); |
| 317 | call_rcu_unlock(&call_rcu_mutex); |
| 318 | if (per_cpu_call_rcu_data == NULL) { |
| 319 | errno = ENOMEM; |
| 320 | return -ENOMEM; |
| 321 | } |
| 322 | per_cpu_call_rcu_data[cpu] = crdp; |
| 323 | return 0; |
| 324 | } |
| 325 | |
| 326 | /* |
| 327 | * Return a pointer to the default call_rcu_data structure, creating |
| 328 | * one if need be. Because we never free call_rcu_data structures, |
| 329 | * we don't need to be in an RCU read-side critical section. |
| 330 | */ |
| 331 | |
| 332 | struct call_rcu_data *get_default_call_rcu_data(void) |
| 333 | { |
| 334 | if (default_call_rcu_data != NULL) |
| 335 | return rcu_dereference(default_call_rcu_data); |
| 336 | call_rcu_lock(&call_rcu_mutex); |
| 337 | if (default_call_rcu_data != NULL) { |
| 338 | call_rcu_unlock(&call_rcu_mutex); |
| 339 | return default_call_rcu_data; |
| 340 | } |
| 341 | call_rcu_data_init(&default_call_rcu_data, 0); |
| 342 | call_rcu_unlock(&call_rcu_mutex); |
| 343 | return default_call_rcu_data; |
| 344 | } |
| 345 | |
| 346 | /* |
| 347 | * Return the call_rcu_data structure that applies to the currently |
| 348 | * running thread. Any call_rcu_data structure assigned specifically |
| 349 | * to this thread has first priority, followed by any call_rcu_data |
| 350 | * structure assigned to the CPU on which the thread is running, |
| 351 | * followed by the default call_rcu_data structure. If there is not |
| 352 | * yet a default call_rcu_data structure, one will be created. |
| 353 | */ |
| 354 | struct call_rcu_data *get_call_rcu_data(void) |
| 355 | { |
| 356 | int curcpu; |
| 357 | static int warned = 0; |
| 358 | |
| 359 | if (thread_call_rcu_data != NULL) |
| 360 | return thread_call_rcu_data; |
| 361 | if (maxcpus <= 0) |
| 362 | return get_default_call_rcu_data(); |
| 363 | curcpu = sched_getcpu(); |
| 364 | if (!warned && (curcpu < 0 || maxcpus <= curcpu)) { |
| 365 | fprintf(stderr, "[error] liburcu: gcrd CPU # out of range\n"); |
| 366 | warned = 1; |
| 367 | } |
| 368 | if (curcpu >= 0 && maxcpus > curcpu && |
| 369 | per_cpu_call_rcu_data != NULL && |
| 370 | per_cpu_call_rcu_data[curcpu] != NULL) |
| 371 | return per_cpu_call_rcu_data[curcpu]; |
| 372 | return get_default_call_rcu_data(); |
| 373 | } |
| 374 | |
| 375 | /* |
| 376 | * Return a pointer to this task's call_rcu_data if there is one. |
| 377 | */ |
| 378 | |
| 379 | struct call_rcu_data *get_thread_call_rcu_data(void) |
| 380 | { |
| 381 | return thread_call_rcu_data; |
| 382 | } |
| 383 | |
| 384 | /* |
| 385 | * Set this task's call_rcu_data structure as specified, regardless |
| 386 | * of whether or not this task already had one. (This allows switching |
| 387 | * to and from real-time call_rcu threads, for example.) |
| 388 | */ |
| 389 | |
| 390 | void set_thread_call_rcu_data(struct call_rcu_data *crdp) |
| 391 | { |
| 392 | thread_call_rcu_data = crdp; |
| 393 | } |
| 394 | |
| 395 | /* |
| 396 | * Create a separate call_rcu thread for each CPU. This does not |
| 397 | * replace a pre-existing call_rcu thread -- use the set_cpu_call_rcu_data() |
| 398 | * function if you want that behavior. |
| 399 | */ |
| 400 | |
| 401 | int create_all_cpu_call_rcu_data(unsigned long flags) |
| 402 | { |
| 403 | int i; |
| 404 | struct call_rcu_data *crdp; |
| 405 | int ret; |
| 406 | |
| 407 | call_rcu_lock(&call_rcu_mutex); |
| 408 | alloc_cpu_call_rcu_data(); |
| 409 | call_rcu_unlock(&call_rcu_mutex); |
| 410 | if (maxcpus <= 0) { |
| 411 | errno = EINVAL; |
| 412 | return -EINVAL; |
| 413 | } |
| 414 | if (per_cpu_call_rcu_data == NULL) { |
| 415 | errno = ENOMEM; |
| 416 | return -ENOMEM; |
| 417 | } |
| 418 | for (i = 0; i < maxcpus; i++) { |
| 419 | call_rcu_lock(&call_rcu_mutex); |
| 420 | if (get_cpu_call_rcu_data(i)) { |
| 421 | call_rcu_unlock(&call_rcu_mutex); |
| 422 | continue; |
| 423 | } |
| 424 | crdp = __create_call_rcu_data(flags); |
| 425 | if (crdp == NULL) { |
| 426 | call_rcu_unlock(&call_rcu_mutex); |
| 427 | errno = ENOMEM; |
| 428 | return -ENOMEM; |
| 429 | } |
| 430 | call_rcu_unlock(&call_rcu_mutex); |
| 431 | if ((ret = set_cpu_call_rcu_data(i, crdp)) != 0) { |
| 432 | /* FIXME: Leaks crdp for now. */ |
| 433 | return ret; /* Can happen on race. */ |
| 434 | } |
| 435 | } |
| 436 | return 0; |
| 437 | } |
| 438 | |
| 439 | /* |
| 440 | * Schedule a function to be invoked after a following grace period. |
| 441 | * This is the only function that must be called -- the others are |
| 442 | * only present to allow applications to tune their use of RCU for |
| 443 | * maximum performance. |
| 444 | * |
| 445 | * Note that unless a call_rcu thread has not already been created, |
| 446 | * the first invocation of call_rcu() will create one. So, if you |
| 447 | * need the first invocation of call_rcu() to be fast, make sure |
| 448 | * to create a call_rcu thread first. One way to accomplish this is |
| 449 | * "get_call_rcu_data();", and another is create_all_cpu_call_rcu_data(). |
| 450 | */ |
| 451 | |
| 452 | void call_rcu(struct rcu_head *head, |
| 453 | void (*func)(struct rcu_head *head)) |
| 454 | { |
| 455 | struct call_rcu_data *crdp; |
| 456 | |
| 457 | cds_wfq_node_init(&head->next); |
| 458 | head->func = func; |
| 459 | crdp = get_call_rcu_data(); |
| 460 | cds_wfq_enqueue(&crdp->cbs, &head->next); |
| 461 | uatomic_inc(&crdp->qlen); |
| 462 | if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RT)) { |
| 463 | call_rcu_lock(&crdp->mtx); |
| 464 | if (!(_CMM_LOAD_SHARED(crdp->flags) & URCU_CALL_RCU_RUNNING)) { |
| 465 | if (pthread_cond_signal(&crdp->cond) != 0) { |
| 466 | perror("pthread_cond_signal"); |
| 467 | exit(-1); |
| 468 | } |
| 469 | } |
| 470 | call_rcu_unlock(&crdp->mtx); |
| 471 | } |
| 472 | } |