| 1 | /* |
| 2 | * SPDX-License-Identifier: LGPL-2.1-or-later |
| 3 | * |
| 4 | * Copyright (c) 2009 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 5 | * Copyright (c) 2009 Paul E. McKenney, IBM Corporation. |
| 6 | * |
| 7 | * Userspace RCU library for LTTng-UST, derived from liburcu "bulletproof" version. |
| 8 | */ |
| 9 | |
| 10 | #define _LGPL_SOURCE |
| 11 | #include <stdio.h> |
| 12 | #include <pthread.h> |
| 13 | #include <signal.h> |
| 14 | #include <assert.h> |
| 15 | #include <stdlib.h> |
| 16 | #include <string.h> |
| 17 | #include <errno.h> |
| 18 | #include <poll.h> |
| 19 | #include <unistd.h> |
| 20 | #include <stdbool.h> |
| 21 | #include <sys/mman.h> |
| 22 | |
| 23 | #include <urcu/arch.h> |
| 24 | #include <urcu/wfcqueue.h> |
| 25 | #include <lttng/urcu/static/urcu-ust.h> |
| 26 | #include <lttng/urcu/pointer.h> |
| 27 | #include <urcu/tls-compat.h> |
| 28 | |
| 29 | /* Do not #define _LGPL_SOURCE to ensure we can emit the wrapper symbols */ |
| 30 | #undef _LGPL_SOURCE |
| 31 | #include <lttng/urcu/urcu-ust.h> |
| 32 | #define _LGPL_SOURCE |
| 33 | |
| 34 | #ifndef MAP_ANONYMOUS |
| 35 | #define MAP_ANONYMOUS MAP_ANON |
| 36 | #endif |
| 37 | |
| 38 | #ifdef __linux__ |
| 39 | static |
| 40 | void *mremap_wrapper(void *old_address, size_t old_size, |
| 41 | size_t new_size, int flags) |
| 42 | { |
| 43 | return mremap(old_address, old_size, new_size, flags); |
| 44 | } |
| 45 | #else |
| 46 | |
| 47 | #define MREMAP_MAYMOVE 1 |
| 48 | #define MREMAP_FIXED 2 |
| 49 | |
| 50 | /* |
| 51 | * mremap wrapper for non-Linux systems not allowing MAYMOVE. |
| 52 | * This is not generic. |
| 53 | */ |
| 54 | static |
| 55 | void *mremap_wrapper(void *old_address, size_t old_size, |
| 56 | size_t new_size, int flags) |
| 57 | { |
| 58 | assert(!(flags & MREMAP_MAYMOVE)); |
| 59 | |
| 60 | return MAP_FAILED; |
| 61 | } |
| 62 | #endif |
| 63 | |
| 64 | /* Sleep delay in ms */ |
| 65 | #define RCU_SLEEP_DELAY_MS 10 |
| 66 | #define INIT_NR_THREADS 8 |
| 67 | #define ARENA_INIT_ALLOC \ |
| 68 | sizeof(struct registry_chunk) \ |
| 69 | + INIT_NR_THREADS * sizeof(struct lttng_ust_urcu_reader) |
| 70 | |
| 71 | /* |
| 72 | * Active attempts to check for reader Q.S. before calling sleep(). |
| 73 | */ |
| 74 | #define RCU_QS_ACTIVE_ATTEMPTS 100 |
| 75 | |
| 76 | static |
| 77 | int lttng_ust_urcu_refcount; |
| 78 | |
| 79 | /* If the headers do not support membarrier system call, fall back smp_mb. */ |
| 80 | #ifdef __NR_membarrier |
| 81 | # define membarrier(...) syscall(__NR_membarrier, __VA_ARGS__) |
| 82 | #else |
| 83 | # define membarrier(...) -ENOSYS |
| 84 | #endif |
| 85 | |
| 86 | enum membarrier_cmd { |
| 87 | MEMBARRIER_CMD_QUERY = 0, |
| 88 | MEMBARRIER_CMD_SHARED = (1 << 0), |
| 89 | /* reserved for MEMBARRIER_CMD_SHARED_EXPEDITED (1 << 1) */ |
| 90 | /* reserved for MEMBARRIER_CMD_PRIVATE (1 << 2) */ |
| 91 | MEMBARRIER_CMD_PRIVATE_EXPEDITED = (1 << 3), |
| 92 | MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED = (1 << 4), |
| 93 | }; |
| 94 | |
| 95 | static |
| 96 | void _lttng_ust_urcu_init(void) |
| 97 | __attribute__((constructor)); |
| 98 | static |
| 99 | void __attribute__((destructor)) lttng_ust_urcu_exit(void); |
| 100 | |
| 101 | #ifndef CONFIG_RCU_FORCE_SYS_MEMBARRIER |
| 102 | int lttng_ust_urcu_has_sys_membarrier; |
| 103 | #endif |
| 104 | |
| 105 | /* |
| 106 | * rcu_gp_lock ensures mutual exclusion between threads calling |
| 107 | * synchronize_rcu(). |
| 108 | */ |
| 109 | static pthread_mutex_t rcu_gp_lock = PTHREAD_MUTEX_INITIALIZER; |
| 110 | /* |
| 111 | * rcu_registry_lock ensures mutual exclusion between threads |
| 112 | * registering and unregistering themselves to/from the registry, and |
| 113 | * with threads reading that registry from synchronize_rcu(). However, |
| 114 | * this lock is not held all the way through the completion of awaiting |
| 115 | * for the grace period. It is sporadically released between iterations |
| 116 | * on the registry. |
| 117 | * rcu_registry_lock may nest inside rcu_gp_lock. |
| 118 | */ |
| 119 | static pthread_mutex_t rcu_registry_lock = PTHREAD_MUTEX_INITIALIZER; |
| 120 | |
| 121 | static pthread_mutex_t init_lock = PTHREAD_MUTEX_INITIALIZER; |
| 122 | static int initialized; |
| 123 | |
| 124 | static pthread_key_t lttng_ust_urcu_key; |
| 125 | |
| 126 | struct lttng_ust_urcu_gp lttng_ust_urcu_gp = { .ctr = LTTNG_UST_URCU_GP_COUNT }; |
| 127 | |
| 128 | /* |
| 129 | * Pointer to registry elements. Written to only by each individual reader. Read |
| 130 | * by both the reader and the writers. |
| 131 | */ |
| 132 | DEFINE_URCU_TLS(struct lttng_ust_urcu_reader *, lttng_ust_urcu_reader); |
| 133 | |
| 134 | static CDS_LIST_HEAD(registry); |
| 135 | |
| 136 | struct registry_chunk { |
| 137 | size_t data_len; /* data length */ |
| 138 | size_t used; /* amount of data used */ |
| 139 | struct cds_list_head node; /* chunk_list node */ |
| 140 | char data[]; |
| 141 | }; |
| 142 | |
| 143 | struct registry_arena { |
| 144 | struct cds_list_head chunk_list; |
| 145 | }; |
| 146 | |
| 147 | static struct registry_arena registry_arena = { |
| 148 | .chunk_list = CDS_LIST_HEAD_INIT(registry_arena.chunk_list), |
| 149 | }; |
| 150 | |
| 151 | /* Saved fork signal mask, protected by rcu_gp_lock */ |
| 152 | static sigset_t saved_fork_signal_mask; |
| 153 | |
| 154 | static void mutex_lock(pthread_mutex_t *mutex) |
| 155 | { |
| 156 | int ret; |
| 157 | |
| 158 | #ifndef DISTRUST_SIGNALS_EXTREME |
| 159 | ret = pthread_mutex_lock(mutex); |
| 160 | if (ret) |
| 161 | abort(); |
| 162 | #else /* #ifndef DISTRUST_SIGNALS_EXTREME */ |
| 163 | while ((ret = pthread_mutex_trylock(mutex)) != 0) { |
| 164 | if (ret != EBUSY && ret != EINTR) |
| 165 | abort(); |
| 166 | poll(NULL,0,10); |
| 167 | } |
| 168 | #endif /* #else #ifndef DISTRUST_SIGNALS_EXTREME */ |
| 169 | } |
| 170 | |
| 171 | static void mutex_unlock(pthread_mutex_t *mutex) |
| 172 | { |
| 173 | int ret; |
| 174 | |
| 175 | ret = pthread_mutex_unlock(mutex); |
| 176 | if (ret) |
| 177 | abort(); |
| 178 | } |
| 179 | |
| 180 | static void smp_mb_master(void) |
| 181 | { |
| 182 | if (caa_likely(lttng_ust_urcu_has_sys_membarrier)) { |
| 183 | if (membarrier(MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0)) |
| 184 | abort(); |
| 185 | } else { |
| 186 | cmm_smp_mb(); |
| 187 | } |
| 188 | } |
| 189 | |
| 190 | /* |
| 191 | * Always called with rcu_registry lock held. Releases this lock between |
| 192 | * iterations and grabs it again. Holds the lock when it returns. |
| 193 | */ |
| 194 | static void wait_for_readers(struct cds_list_head *input_readers, |
| 195 | struct cds_list_head *cur_snap_readers, |
| 196 | struct cds_list_head *qsreaders) |
| 197 | { |
| 198 | unsigned int wait_loops = 0; |
| 199 | struct lttng_ust_urcu_reader *index, *tmp; |
| 200 | |
| 201 | /* |
| 202 | * Wait for each thread URCU_TLS(lttng_ust_urcu_reader).ctr to either |
| 203 | * indicate quiescence (not nested), or observe the current |
| 204 | * rcu_gp.ctr value. |
| 205 | */ |
| 206 | for (;;) { |
| 207 | if (wait_loops < RCU_QS_ACTIVE_ATTEMPTS) |
| 208 | wait_loops++; |
| 209 | |
| 210 | cds_list_for_each_entry_safe(index, tmp, input_readers, node) { |
| 211 | switch (lttng_ust_urcu_reader_state(&index->ctr)) { |
| 212 | case LTTNG_UST_URCU_READER_ACTIVE_CURRENT: |
| 213 | if (cur_snap_readers) { |
| 214 | cds_list_move(&index->node, |
| 215 | cur_snap_readers); |
| 216 | break; |
| 217 | } |
| 218 | /* Fall-through */ |
| 219 | case LTTNG_UST_URCU_READER_INACTIVE: |
| 220 | cds_list_move(&index->node, qsreaders); |
| 221 | break; |
| 222 | case LTTNG_UST_URCU_READER_ACTIVE_OLD: |
| 223 | /* |
| 224 | * Old snapshot. Leaving node in |
| 225 | * input_readers will make us busy-loop |
| 226 | * until the snapshot becomes current or |
| 227 | * the reader becomes inactive. |
| 228 | */ |
| 229 | break; |
| 230 | } |
| 231 | } |
| 232 | |
| 233 | if (cds_list_empty(input_readers)) { |
| 234 | break; |
| 235 | } else { |
| 236 | /* Temporarily unlock the registry lock. */ |
| 237 | mutex_unlock(&rcu_registry_lock); |
| 238 | if (wait_loops >= RCU_QS_ACTIVE_ATTEMPTS) |
| 239 | (void) poll(NULL, 0, RCU_SLEEP_DELAY_MS); |
| 240 | else |
| 241 | caa_cpu_relax(); |
| 242 | /* Re-lock the registry lock before the next loop. */ |
| 243 | mutex_lock(&rcu_registry_lock); |
| 244 | } |
| 245 | } |
| 246 | } |
| 247 | |
| 248 | void lttng_ust_urcu_synchronize_rcu(void) |
| 249 | { |
| 250 | CDS_LIST_HEAD(cur_snap_readers); |
| 251 | CDS_LIST_HEAD(qsreaders); |
| 252 | sigset_t newmask, oldmask; |
| 253 | int ret; |
| 254 | |
| 255 | ret = sigfillset(&newmask); |
| 256 | assert(!ret); |
| 257 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
| 258 | assert(!ret); |
| 259 | |
| 260 | mutex_lock(&rcu_gp_lock); |
| 261 | |
| 262 | mutex_lock(&rcu_registry_lock); |
| 263 | |
| 264 | if (cds_list_empty(®istry)) |
| 265 | goto out; |
| 266 | |
| 267 | /* All threads should read qparity before accessing data structure |
| 268 | * where new ptr points to. */ |
| 269 | /* Write new ptr before changing the qparity */ |
| 270 | smp_mb_master(); |
| 271 | |
| 272 | /* |
| 273 | * Wait for readers to observe original parity or be quiescent. |
| 274 | * wait_for_readers() can release and grab again rcu_registry_lock |
| 275 | * interally. |
| 276 | */ |
| 277 | wait_for_readers(®istry, &cur_snap_readers, &qsreaders); |
| 278 | |
| 279 | /* |
| 280 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
| 281 | * model easier to understand. It does not have a big performance impact |
| 282 | * anyway, given this is the write-side. |
| 283 | */ |
| 284 | cmm_smp_mb(); |
| 285 | |
| 286 | /* Switch parity: 0 -> 1, 1 -> 0 */ |
| 287 | CMM_STORE_SHARED(lttng_ust_urcu_gp.ctr, lttng_ust_urcu_gp.ctr ^ LTTNG_UST_URCU_GP_CTR_PHASE); |
| 288 | |
| 289 | /* |
| 290 | * Must commit qparity update to memory before waiting for other parity |
| 291 | * quiescent state. Failure to do so could result in the writer waiting |
| 292 | * forever while new readers are always accessing data (no progress). |
| 293 | * Ensured by CMM_STORE_SHARED and CMM_LOAD_SHARED. |
| 294 | */ |
| 295 | |
| 296 | /* |
| 297 | * Adding a cmm_smp_mb() which is _not_ formally required, but makes the |
| 298 | * model easier to understand. It does not have a big performance impact |
| 299 | * anyway, given this is the write-side. |
| 300 | */ |
| 301 | cmm_smp_mb(); |
| 302 | |
| 303 | /* |
| 304 | * Wait for readers to observe new parity or be quiescent. |
| 305 | * wait_for_readers() can release and grab again rcu_registry_lock |
| 306 | * interally. |
| 307 | */ |
| 308 | wait_for_readers(&cur_snap_readers, NULL, &qsreaders); |
| 309 | |
| 310 | /* |
| 311 | * Put quiescent reader list back into registry. |
| 312 | */ |
| 313 | cds_list_splice(&qsreaders, ®istry); |
| 314 | |
| 315 | /* |
| 316 | * Finish waiting for reader threads before letting the old ptr being |
| 317 | * freed. |
| 318 | */ |
| 319 | smp_mb_master(); |
| 320 | out: |
| 321 | mutex_unlock(&rcu_registry_lock); |
| 322 | mutex_unlock(&rcu_gp_lock); |
| 323 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
| 324 | assert(!ret); |
| 325 | } |
| 326 | |
| 327 | /* |
| 328 | * library wrappers to be used by non-LGPL compatible source code. |
| 329 | */ |
| 330 | |
| 331 | void lttng_ust_urcu_read_lock(void) |
| 332 | { |
| 333 | _lttng_ust_urcu_read_lock(); |
| 334 | } |
| 335 | |
| 336 | void lttng_ust_urcu_read_unlock(void) |
| 337 | { |
| 338 | _lttng_ust_urcu_read_unlock(); |
| 339 | } |
| 340 | |
| 341 | int lttng_ust_urcu_read_ongoing(void) |
| 342 | { |
| 343 | return _lttng_ust_urcu_read_ongoing(); |
| 344 | } |
| 345 | |
| 346 | /* |
| 347 | * Only grow for now. If empty, allocate a ARENA_INIT_ALLOC sized chunk. |
| 348 | * Else, try expanding the last chunk. If this fails, allocate a new |
| 349 | * chunk twice as big as the last chunk. |
| 350 | * Memory used by chunks _never_ moves. A chunk could theoretically be |
| 351 | * freed when all "used" slots are released, but we don't do it at this |
| 352 | * point. |
| 353 | */ |
| 354 | static |
| 355 | void expand_arena(struct registry_arena *arena) |
| 356 | { |
| 357 | struct registry_chunk *new_chunk, *last_chunk; |
| 358 | size_t old_chunk_len, new_chunk_len; |
| 359 | |
| 360 | /* No chunk. */ |
| 361 | if (cds_list_empty(&arena->chunk_list)) { |
| 362 | assert(ARENA_INIT_ALLOC >= |
| 363 | sizeof(struct registry_chunk) |
| 364 | + sizeof(struct lttng_ust_urcu_reader)); |
| 365 | new_chunk_len = ARENA_INIT_ALLOC; |
| 366 | new_chunk = (struct registry_chunk *) mmap(NULL, |
| 367 | new_chunk_len, |
| 368 | PROT_READ | PROT_WRITE, |
| 369 | MAP_ANONYMOUS | MAP_PRIVATE, |
| 370 | -1, 0); |
| 371 | if (new_chunk == MAP_FAILED) |
| 372 | abort(); |
| 373 | memset(new_chunk, 0, new_chunk_len); |
| 374 | new_chunk->data_len = |
| 375 | new_chunk_len - sizeof(struct registry_chunk); |
| 376 | cds_list_add_tail(&new_chunk->node, &arena->chunk_list); |
| 377 | return; /* We're done. */ |
| 378 | } |
| 379 | |
| 380 | /* Try expanding last chunk. */ |
| 381 | last_chunk = cds_list_entry(arena->chunk_list.prev, |
| 382 | struct registry_chunk, node); |
| 383 | old_chunk_len = |
| 384 | last_chunk->data_len + sizeof(struct registry_chunk); |
| 385 | new_chunk_len = old_chunk_len << 1; |
| 386 | |
| 387 | /* Don't allow memory mapping to move, just expand. */ |
| 388 | new_chunk = mremap_wrapper(last_chunk, old_chunk_len, |
| 389 | new_chunk_len, 0); |
| 390 | if (new_chunk != MAP_FAILED) { |
| 391 | /* Should not have moved. */ |
| 392 | assert(new_chunk == last_chunk); |
| 393 | memset((char *) last_chunk + old_chunk_len, 0, |
| 394 | new_chunk_len - old_chunk_len); |
| 395 | last_chunk->data_len = |
| 396 | new_chunk_len - sizeof(struct registry_chunk); |
| 397 | return; /* We're done. */ |
| 398 | } |
| 399 | |
| 400 | /* Remap did not succeed, we need to add a new chunk. */ |
| 401 | new_chunk = (struct registry_chunk *) mmap(NULL, |
| 402 | new_chunk_len, |
| 403 | PROT_READ | PROT_WRITE, |
| 404 | MAP_ANONYMOUS | MAP_PRIVATE, |
| 405 | -1, 0); |
| 406 | if (new_chunk == MAP_FAILED) |
| 407 | abort(); |
| 408 | memset(new_chunk, 0, new_chunk_len); |
| 409 | new_chunk->data_len = |
| 410 | new_chunk_len - sizeof(struct registry_chunk); |
| 411 | cds_list_add_tail(&new_chunk->node, &arena->chunk_list); |
| 412 | } |
| 413 | |
| 414 | static |
| 415 | struct lttng_ust_urcu_reader *arena_alloc(struct registry_arena *arena) |
| 416 | { |
| 417 | struct registry_chunk *chunk; |
| 418 | struct lttng_ust_urcu_reader *rcu_reader_reg; |
| 419 | int expand_done = 0; /* Only allow to expand once per alloc */ |
| 420 | size_t len = sizeof(struct lttng_ust_urcu_reader); |
| 421 | |
| 422 | retry: |
| 423 | cds_list_for_each_entry(chunk, &arena->chunk_list, node) { |
| 424 | if (chunk->data_len - chunk->used < len) |
| 425 | continue; |
| 426 | /* Find spot */ |
| 427 | for (rcu_reader_reg = (struct lttng_ust_urcu_reader *) &chunk->data[0]; |
| 428 | rcu_reader_reg < (struct lttng_ust_urcu_reader *) &chunk->data[chunk->data_len]; |
| 429 | rcu_reader_reg++) { |
| 430 | if (!rcu_reader_reg->alloc) { |
| 431 | rcu_reader_reg->alloc = 1; |
| 432 | chunk->used += len; |
| 433 | return rcu_reader_reg; |
| 434 | } |
| 435 | } |
| 436 | } |
| 437 | |
| 438 | if (!expand_done) { |
| 439 | expand_arena(arena); |
| 440 | expand_done = 1; |
| 441 | goto retry; |
| 442 | } |
| 443 | |
| 444 | return NULL; |
| 445 | } |
| 446 | |
| 447 | /* Called with signals off and mutex locked */ |
| 448 | static |
| 449 | void add_thread(void) |
| 450 | { |
| 451 | struct lttng_ust_urcu_reader *rcu_reader_reg; |
| 452 | int ret; |
| 453 | |
| 454 | rcu_reader_reg = arena_alloc(®istry_arena); |
| 455 | if (!rcu_reader_reg) |
| 456 | abort(); |
| 457 | ret = pthread_setspecific(lttng_ust_urcu_key, rcu_reader_reg); |
| 458 | if (ret) |
| 459 | abort(); |
| 460 | |
| 461 | /* Add to registry */ |
| 462 | rcu_reader_reg->tid = pthread_self(); |
| 463 | assert(rcu_reader_reg->ctr == 0); |
| 464 | cds_list_add(&rcu_reader_reg->node, ®istry); |
| 465 | /* |
| 466 | * Reader threads are pointing to the reader registry. This is |
| 467 | * why its memory should never be relocated. |
| 468 | */ |
| 469 | URCU_TLS(lttng_ust_urcu_reader) = rcu_reader_reg; |
| 470 | } |
| 471 | |
| 472 | /* Called with mutex locked */ |
| 473 | static |
| 474 | void cleanup_thread(struct registry_chunk *chunk, |
| 475 | struct lttng_ust_urcu_reader *rcu_reader_reg) |
| 476 | { |
| 477 | rcu_reader_reg->ctr = 0; |
| 478 | cds_list_del(&rcu_reader_reg->node); |
| 479 | rcu_reader_reg->tid = 0; |
| 480 | rcu_reader_reg->alloc = 0; |
| 481 | chunk->used -= sizeof(struct lttng_ust_urcu_reader); |
| 482 | } |
| 483 | |
| 484 | static |
| 485 | struct registry_chunk *find_chunk(struct lttng_ust_urcu_reader *rcu_reader_reg) |
| 486 | { |
| 487 | struct registry_chunk *chunk; |
| 488 | |
| 489 | cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) { |
| 490 | if (rcu_reader_reg < (struct lttng_ust_urcu_reader *) &chunk->data[0]) |
| 491 | continue; |
| 492 | if (rcu_reader_reg >= (struct lttng_ust_urcu_reader *) &chunk->data[chunk->data_len]) |
| 493 | continue; |
| 494 | return chunk; |
| 495 | } |
| 496 | return NULL; |
| 497 | } |
| 498 | |
| 499 | /* Called with signals off and mutex locked */ |
| 500 | static |
| 501 | void remove_thread(struct lttng_ust_urcu_reader *rcu_reader_reg) |
| 502 | { |
| 503 | cleanup_thread(find_chunk(rcu_reader_reg), rcu_reader_reg); |
| 504 | URCU_TLS(lttng_ust_urcu_reader) = NULL; |
| 505 | } |
| 506 | |
| 507 | /* Disable signals, take mutex, add to registry */ |
| 508 | void lttng_ust_urcu_register(void) |
| 509 | { |
| 510 | sigset_t newmask, oldmask; |
| 511 | int ret; |
| 512 | |
| 513 | ret = sigfillset(&newmask); |
| 514 | if (ret) |
| 515 | abort(); |
| 516 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
| 517 | if (ret) |
| 518 | abort(); |
| 519 | |
| 520 | /* |
| 521 | * Check if a signal concurrently registered our thread since |
| 522 | * the check in rcu_read_lock(). |
| 523 | */ |
| 524 | if (URCU_TLS(lttng_ust_urcu_reader)) |
| 525 | goto end; |
| 526 | |
| 527 | /* |
| 528 | * Take care of early registration before lttng_ust_urcu constructor. |
| 529 | */ |
| 530 | _lttng_ust_urcu_init(); |
| 531 | |
| 532 | mutex_lock(&rcu_registry_lock); |
| 533 | add_thread(); |
| 534 | mutex_unlock(&rcu_registry_lock); |
| 535 | end: |
| 536 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
| 537 | if (ret) |
| 538 | abort(); |
| 539 | } |
| 540 | |
| 541 | void lttng_ust_urcu_register_thread(void) |
| 542 | { |
| 543 | if (caa_unlikely(!URCU_TLS(lttng_ust_urcu_reader))) |
| 544 | lttng_ust_urcu_register(); /* If not yet registered. */ |
| 545 | } |
| 546 | |
| 547 | /* Disable signals, take mutex, remove from registry */ |
| 548 | static |
| 549 | void lttng_ust_urcu_unregister(struct lttng_ust_urcu_reader *rcu_reader_reg) |
| 550 | { |
| 551 | sigset_t newmask, oldmask; |
| 552 | int ret; |
| 553 | |
| 554 | ret = sigfillset(&newmask); |
| 555 | if (ret) |
| 556 | abort(); |
| 557 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
| 558 | if (ret) |
| 559 | abort(); |
| 560 | |
| 561 | mutex_lock(&rcu_registry_lock); |
| 562 | remove_thread(rcu_reader_reg); |
| 563 | mutex_unlock(&rcu_registry_lock); |
| 564 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
| 565 | if (ret) |
| 566 | abort(); |
| 567 | lttng_ust_urcu_exit(); |
| 568 | } |
| 569 | |
| 570 | /* |
| 571 | * Remove thread from the registry when it exits, and flag it as |
| 572 | * destroyed so garbage collection can take care of it. |
| 573 | */ |
| 574 | static |
| 575 | void lttng_ust_urcu_thread_exit_notifier(void *rcu_key) |
| 576 | { |
| 577 | lttng_ust_urcu_unregister(rcu_key); |
| 578 | } |
| 579 | |
| 580 | #ifdef CONFIG_RCU_FORCE_SYS_MEMBARRIER |
| 581 | static |
| 582 | void lttng_ust_urcu_sys_membarrier_status(bool available) |
| 583 | { |
| 584 | if (!available) |
| 585 | abort(); |
| 586 | } |
| 587 | #else |
| 588 | static |
| 589 | void lttng_ust_urcu_sys_membarrier_status(bool available) |
| 590 | { |
| 591 | if (!available) |
| 592 | return; |
| 593 | lttng_ust_urcu_has_sys_membarrier = 1; |
| 594 | } |
| 595 | #endif |
| 596 | |
| 597 | static |
| 598 | void lttng_ust_urcu_sys_membarrier_init(void) |
| 599 | { |
| 600 | bool available = false; |
| 601 | int mask; |
| 602 | |
| 603 | mask = membarrier(MEMBARRIER_CMD_QUERY, 0); |
| 604 | if (mask >= 0) { |
| 605 | if (mask & MEMBARRIER_CMD_PRIVATE_EXPEDITED) { |
| 606 | if (membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0)) |
| 607 | abort(); |
| 608 | available = true; |
| 609 | } |
| 610 | } |
| 611 | lttng_ust_urcu_sys_membarrier_status(available); |
| 612 | } |
| 613 | |
| 614 | static |
| 615 | void _lttng_ust_urcu_init(void) |
| 616 | { |
| 617 | mutex_lock(&init_lock); |
| 618 | if (!lttng_ust_urcu_refcount++) { |
| 619 | int ret; |
| 620 | |
| 621 | ret = pthread_key_create(<tng_ust_urcu_key, |
| 622 | lttng_ust_urcu_thread_exit_notifier); |
| 623 | if (ret) |
| 624 | abort(); |
| 625 | lttng_ust_urcu_sys_membarrier_init(); |
| 626 | initialized = 1; |
| 627 | } |
| 628 | mutex_unlock(&init_lock); |
| 629 | } |
| 630 | |
| 631 | static |
| 632 | void lttng_ust_urcu_exit(void) |
| 633 | { |
| 634 | mutex_lock(&init_lock); |
| 635 | if (!--lttng_ust_urcu_refcount) { |
| 636 | struct registry_chunk *chunk, *tmp; |
| 637 | int ret; |
| 638 | |
| 639 | cds_list_for_each_entry_safe(chunk, tmp, |
| 640 | ®istry_arena.chunk_list, node) { |
| 641 | munmap((void *) chunk, chunk->data_len |
| 642 | + sizeof(struct registry_chunk)); |
| 643 | } |
| 644 | CDS_INIT_LIST_HEAD(®istry_arena.chunk_list); |
| 645 | ret = pthread_key_delete(lttng_ust_urcu_key); |
| 646 | if (ret) |
| 647 | abort(); |
| 648 | } |
| 649 | mutex_unlock(&init_lock); |
| 650 | } |
| 651 | |
| 652 | /* |
| 653 | * Holding the rcu_gp_lock and rcu_registry_lock across fork will make |
| 654 | * sure we fork() don't race with a concurrent thread executing with |
| 655 | * any of those locks held. This ensures that the registry and data |
| 656 | * protected by rcu_gp_lock are in a coherent state in the child. |
| 657 | */ |
| 658 | void lttng_ust_urcu_before_fork(void) |
| 659 | { |
| 660 | sigset_t newmask, oldmask; |
| 661 | int ret; |
| 662 | |
| 663 | ret = sigfillset(&newmask); |
| 664 | assert(!ret); |
| 665 | ret = pthread_sigmask(SIG_BLOCK, &newmask, &oldmask); |
| 666 | assert(!ret); |
| 667 | mutex_lock(&rcu_gp_lock); |
| 668 | mutex_lock(&rcu_registry_lock); |
| 669 | saved_fork_signal_mask = oldmask; |
| 670 | } |
| 671 | |
| 672 | void lttng_ust_urcu_after_fork_parent(void) |
| 673 | { |
| 674 | sigset_t oldmask; |
| 675 | int ret; |
| 676 | |
| 677 | oldmask = saved_fork_signal_mask; |
| 678 | mutex_unlock(&rcu_registry_lock); |
| 679 | mutex_unlock(&rcu_gp_lock); |
| 680 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
| 681 | assert(!ret); |
| 682 | } |
| 683 | |
| 684 | /* |
| 685 | * Prune all entries from registry except our own thread. Fits the Linux |
| 686 | * fork behavior. Called with rcu_gp_lock and rcu_registry_lock held. |
| 687 | */ |
| 688 | static |
| 689 | void lttng_ust_urcu_prune_registry(void) |
| 690 | { |
| 691 | struct registry_chunk *chunk; |
| 692 | struct lttng_ust_urcu_reader *rcu_reader_reg; |
| 693 | |
| 694 | cds_list_for_each_entry(chunk, ®istry_arena.chunk_list, node) { |
| 695 | for (rcu_reader_reg = (struct lttng_ust_urcu_reader *) &chunk->data[0]; |
| 696 | rcu_reader_reg < (struct lttng_ust_urcu_reader *) &chunk->data[chunk->data_len]; |
| 697 | rcu_reader_reg++) { |
| 698 | if (!rcu_reader_reg->alloc) |
| 699 | continue; |
| 700 | if (rcu_reader_reg->tid == pthread_self()) |
| 701 | continue; |
| 702 | cleanup_thread(chunk, rcu_reader_reg); |
| 703 | } |
| 704 | } |
| 705 | } |
| 706 | |
| 707 | void lttng_ust_urcu_after_fork_child(void) |
| 708 | { |
| 709 | sigset_t oldmask; |
| 710 | int ret; |
| 711 | |
| 712 | lttng_ust_urcu_prune_registry(); |
| 713 | oldmask = saved_fork_signal_mask; |
| 714 | mutex_unlock(&rcu_registry_lock); |
| 715 | mutex_unlock(&rcu_gp_lock); |
| 716 | ret = pthread_sigmask(SIG_SETMASK, &oldmask, NULL); |
| 717 | assert(!ret); |
| 718 | } |