| 1 | /* |
| 2 | * Copyright (C) 2011 EfficiOS Inc. |
| 3 | * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com> |
| 4 | * Copyright (C) 2013 Jérémie Galarneau <jeremie.galarneau@efficios.com> |
| 5 | * |
| 6 | * SPDX-License-Identifier: GPL-2.0-only |
| 7 | * |
| 8 | */ |
| 9 | |
| 10 | #include "dispatch.hpp" |
| 11 | #include "fd-limit.hpp" |
| 12 | #include "health-sessiond.hpp" |
| 13 | #include "lttng-sessiond.hpp" |
| 14 | #include "testpoint.hpp" |
| 15 | #include "thread.hpp" |
| 16 | #include "ust-app.hpp" |
| 17 | |
| 18 | #include <common/futex.hpp> |
| 19 | #include <common/macros.hpp> |
| 20 | #include <common/urcu.hpp> |
| 21 | |
| 22 | #include <stddef.h> |
| 23 | #include <stdlib.h> |
| 24 | #include <urcu.h> |
| 25 | |
| 26 | namespace { |
| 27 | struct thread_notifiers { |
| 28 | struct ust_cmd_queue *ust_cmd_queue; |
| 29 | int apps_cmd_pipe_write_fd; |
| 30 | int apps_cmd_notify_pipe_write_fd; |
| 31 | int dispatch_thread_exit; |
| 32 | }; |
| 33 | } /* namespace */ |
| 34 | |
| 35 | /* |
| 36 | * For each tracing session, update newly registered apps. The session list |
| 37 | * lock MUST be acquired before calling this. |
| 38 | */ |
| 39 | static void update_ust_app(int app_sock) |
| 40 | { |
| 41 | struct ltt_session *sess, *stmp; |
| 42 | const struct ltt_session_list *session_list = session_get_list(); |
| 43 | struct ust_app *app; |
| 44 | |
| 45 | /* Consumer is in an ERROR state. Stop any application update. */ |
| 46 | if (uatomic_read(&the_ust_consumerd_state) == CONSUMER_ERROR) { |
| 47 | /* Stop the update process since the consumer is dead. */ |
| 48 | return; |
| 49 | } |
| 50 | |
| 51 | lttng::urcu::read_lock_guard read_lock; |
| 52 | LTTNG_ASSERT(app_sock >= 0); |
| 53 | app = ust_app_find_by_sock(app_sock); |
| 54 | if (app == nullptr) { |
| 55 | /* |
| 56 | * Application can be unregistered before so |
| 57 | * this is possible hence simply stopping the |
| 58 | * update. |
| 59 | */ |
| 60 | DBG3("UST app update failed to find app sock %d", app_sock); |
| 61 | return; |
| 62 | } |
| 63 | |
| 64 | /* Update all event notifiers for the app. */ |
| 65 | ust_app_global_update_event_notifier_rules(app); |
| 66 | |
| 67 | /* For all tracing session(s) */ |
| 68 | cds_list_for_each_entry_safe (sess, stmp, &session_list->head, list) { |
| 69 | if (!session_get(sess)) { |
| 70 | continue; |
| 71 | } |
| 72 | session_lock(sess); |
| 73 | if (!sess->active || !sess->ust_session || !sess->ust_session->active) { |
| 74 | goto unlock_session; |
| 75 | } |
| 76 | |
| 77 | ust_app_global_update(sess->ust_session, app); |
| 78 | unlock_session: |
| 79 | session_unlock(sess); |
| 80 | session_put(sess); |
| 81 | } |
| 82 | } |
| 83 | |
| 84 | /* |
| 85 | * Sanitize the wait queue of the dispatch registration thread meaning removing |
| 86 | * invalid nodes from it. This is to avoid memory leaks for the case the UST |
| 87 | * notify socket is never received. |
| 88 | */ |
| 89 | static void sanitize_wait_queue(struct ust_reg_wait_queue *wait_queue) |
| 90 | { |
| 91 | int ret, nb_fd = 0, i; |
| 92 | unsigned int fd_added = 0; |
| 93 | struct lttng_poll_event events; |
| 94 | struct ust_reg_wait_node *wait_node = nullptr, *tmp_wait_node; |
| 95 | |
| 96 | LTTNG_ASSERT(wait_queue); |
| 97 | |
| 98 | lttng_poll_init(&events); |
| 99 | |
| 100 | /* Just skip everything for an empty queue. */ |
| 101 | if (!wait_queue->count) { |
| 102 | goto end; |
| 103 | } |
| 104 | |
| 105 | ret = lttng_poll_create(&events, wait_queue->count, LTTNG_CLOEXEC); |
| 106 | if (ret < 0) { |
| 107 | goto error_create; |
| 108 | } |
| 109 | |
| 110 | cds_list_for_each_entry_safe (wait_node, tmp_wait_node, &wait_queue->head, head) { |
| 111 | LTTNG_ASSERT(wait_node->app); |
| 112 | ret = lttng_poll_add(&events, wait_node->app->sock, LPOLLIN); |
| 113 | if (ret < 0) { |
| 114 | goto error; |
| 115 | } |
| 116 | |
| 117 | fd_added = 1; |
| 118 | } |
| 119 | |
| 120 | if (!fd_added) { |
| 121 | goto end; |
| 122 | } |
| 123 | |
| 124 | /* |
| 125 | * Poll but don't block so we can quickly identify the faulty events and |
| 126 | * clean them afterwards from the wait queue. |
| 127 | */ |
| 128 | ret = lttng_poll_wait(&events, 0); |
| 129 | if (ret < 0) { |
| 130 | goto error; |
| 131 | } |
| 132 | nb_fd = ret; |
| 133 | |
| 134 | for (i = 0; i < nb_fd; i++) { |
| 135 | /* Get faulty FD. */ |
| 136 | uint32_t revents = LTTNG_POLL_GETEV(&events, i); |
| 137 | int pollfd = LTTNG_POLL_GETFD(&events, i); |
| 138 | |
| 139 | cds_list_for_each_entry_safe (wait_node, tmp_wait_node, &wait_queue->head, head) { |
| 140 | if (pollfd == wait_node->app->sock && (revents & (LPOLLHUP | LPOLLERR))) { |
| 141 | cds_list_del(&wait_node->head); |
| 142 | wait_queue->count--; |
| 143 | ust_app_destroy(wait_node->app); |
| 144 | free(wait_node); |
| 145 | /* |
| 146 | * Silence warning of use-after-free in |
| 147 | * cds_list_for_each_entry_safe which uses |
| 148 | * __typeof__(*wait_node). |
| 149 | */ |
| 150 | wait_node = nullptr; |
| 151 | break; |
| 152 | } else { |
| 153 | ERR("Unexpected poll events %u for sock %d", revents, pollfd); |
| 154 | goto error; |
| 155 | } |
| 156 | } |
| 157 | } |
| 158 | |
| 159 | if (nb_fd > 0) { |
| 160 | DBG("Wait queue sanitized, %d node were cleaned up", nb_fd); |
| 161 | } |
| 162 | |
| 163 | end: |
| 164 | lttng_poll_clean(&events); |
| 165 | return; |
| 166 | |
| 167 | error: |
| 168 | lttng_poll_clean(&events); |
| 169 | error_create: |
| 170 | ERR("Unable to sanitize wait queue"); |
| 171 | return; |
| 172 | } |
| 173 | |
| 174 | /* |
| 175 | * Send a socket to a thread This is called from the dispatch UST registration |
| 176 | * thread once all sockets are set for the application. |
| 177 | * |
| 178 | * The sock value can be invalid, we don't really care, the thread will handle |
| 179 | * it and make the necessary cleanup if so. |
| 180 | * |
| 181 | * On success, return 0 else a negative value being the errno message of the |
| 182 | * write(). |
| 183 | */ |
| 184 | static int send_socket_to_thread(int fd, int sock) |
| 185 | { |
| 186 | ssize_t ret; |
| 187 | |
| 188 | /* |
| 189 | * It's possible that the FD is set as invalid with -1 concurrently just |
| 190 | * before calling this function being a shutdown state of the thread. |
| 191 | */ |
| 192 | if (fd < 0) { |
| 193 | ret = -EBADF; |
| 194 | goto error; |
| 195 | } |
| 196 | |
| 197 | ret = lttng_write(fd, &sock, sizeof(sock)); |
| 198 | if (ret < sizeof(sock)) { |
| 199 | PERROR("write apps pipe %d", fd); |
| 200 | if (ret < 0) { |
| 201 | ret = -errno; |
| 202 | } |
| 203 | goto error; |
| 204 | } |
| 205 | |
| 206 | /* All good. Don't send back the write positive ret value. */ |
| 207 | ret = 0; |
| 208 | error: |
| 209 | return (int) ret; |
| 210 | } |
| 211 | |
| 212 | static void cleanup_ust_dispatch_thread(void *data) |
| 213 | { |
| 214 | free(data); |
| 215 | } |
| 216 | |
| 217 | /* |
| 218 | * Dispatch request from the registration threads to the application |
| 219 | * communication thread. |
| 220 | */ |
| 221 | static void *thread_dispatch_ust_registration(void *data) |
| 222 | { |
| 223 | int ret, err = -1; |
| 224 | struct cds_wfcq_node *node; |
| 225 | struct ust_command *ust_cmd = nullptr; |
| 226 | struct ust_reg_wait_node *wait_node = nullptr, *tmp_wait_node; |
| 227 | struct ust_reg_wait_queue wait_queue = { |
| 228 | .count = 0, |
| 229 | .head = {}, |
| 230 | }; |
| 231 | struct thread_notifiers *notifiers = (thread_notifiers *) data; |
| 232 | |
| 233 | rcu_register_thread(); |
| 234 | |
| 235 | health_register(the_health_sessiond, HEALTH_SESSIOND_TYPE_APP_REG_DISPATCH); |
| 236 | |
| 237 | if (testpoint(sessiond_thread_app_reg_dispatch)) { |
| 238 | goto error_testpoint; |
| 239 | } |
| 240 | |
| 241 | health_code_update(); |
| 242 | |
| 243 | CDS_INIT_LIST_HEAD(&wait_queue.head); |
| 244 | |
| 245 | DBG("[thread] Dispatch UST command started"); |
| 246 | |
| 247 | for (;;) { |
| 248 | health_code_update(); |
| 249 | |
| 250 | /* Atomically prepare the queue futex */ |
| 251 | futex_nto1_prepare(¬ifiers->ust_cmd_queue->futex); |
| 252 | |
| 253 | if (CMM_LOAD_SHARED(notifiers->dispatch_thread_exit)) { |
| 254 | break; |
| 255 | } |
| 256 | |
| 257 | do { |
| 258 | struct ust_app *app = nullptr; |
| 259 | ust_cmd = nullptr; |
| 260 | |
| 261 | /* |
| 262 | * Make sure we don't have node(s) that have hung up before receiving |
| 263 | * the notify socket. This is to clean the list in order to avoid |
| 264 | * memory leaks from notify socket that are never seen. |
| 265 | */ |
| 266 | sanitize_wait_queue(&wait_queue); |
| 267 | |
| 268 | health_code_update(); |
| 269 | /* Dequeue command for registration */ |
| 270 | node = cds_wfcq_dequeue_blocking(¬ifiers->ust_cmd_queue->head, |
| 271 | ¬ifiers->ust_cmd_queue->tail); |
| 272 | if (node == nullptr) { |
| 273 | DBG("Woken up but nothing in the UST command queue"); |
| 274 | /* Continue thread execution */ |
| 275 | break; |
| 276 | } |
| 277 | |
| 278 | ust_cmd = lttng::utils::container_of(node, &ust_command::node); |
| 279 | |
| 280 | DBG("Dispatching UST registration pid:%d ppid:%d uid:%d" |
| 281 | " gid:%d sock:%d name:%s (version %d.%d)", |
| 282 | ust_cmd->reg_msg.pid, |
| 283 | ust_cmd->reg_msg.ppid, |
| 284 | ust_cmd->reg_msg.uid, |
| 285 | ust_cmd->reg_msg.gid, |
| 286 | ust_cmd->sock, |
| 287 | ust_cmd->reg_msg.name, |
| 288 | ust_cmd->reg_msg.major, |
| 289 | ust_cmd->reg_msg.minor); |
| 290 | |
| 291 | if (ust_cmd->reg_msg.type == LTTNG_UST_CTL_SOCKET_CMD) { |
| 292 | wait_node = zmalloc<ust_reg_wait_node>(); |
| 293 | if (!wait_node) { |
| 294 | PERROR("zmalloc wait_node dispatch"); |
| 295 | ret = close(ust_cmd->sock); |
| 296 | if (ret < 0) { |
| 297 | PERROR("close ust sock dispatch %d", ust_cmd->sock); |
| 298 | } |
| 299 | lttng_fd_put(LTTNG_FD_APPS, 1); |
| 300 | free(ust_cmd); |
| 301 | ust_cmd = nullptr; |
| 302 | goto error; |
| 303 | } |
| 304 | CDS_INIT_LIST_HEAD(&wait_node->head); |
| 305 | |
| 306 | /* Create application object if socket is CMD. */ |
| 307 | wait_node->app = ust_app_create(&ust_cmd->reg_msg, ust_cmd->sock); |
| 308 | if (!wait_node->app) { |
| 309 | ret = close(ust_cmd->sock); |
| 310 | if (ret < 0) { |
| 311 | PERROR("close ust sock dispatch %d", ust_cmd->sock); |
| 312 | } |
| 313 | lttng_fd_put(LTTNG_FD_APPS, 1); |
| 314 | free(wait_node); |
| 315 | wait_node = nullptr; |
| 316 | free(ust_cmd); |
| 317 | ust_cmd = nullptr; |
| 318 | continue; |
| 319 | } |
| 320 | /* |
| 321 | * Add application to the wait queue so we can set the notify |
| 322 | * socket before putting this object in the global ht. |
| 323 | */ |
| 324 | cds_list_add(&wait_node->head, &wait_queue.head); |
| 325 | wait_queue.count++; |
| 326 | |
| 327 | free(ust_cmd); |
| 328 | ust_cmd = nullptr; |
| 329 | /* |
| 330 | * We have to continue here since we don't have the notify |
| 331 | * socket and the application MUST be added to the hash table |
| 332 | * only at that moment. |
| 333 | */ |
| 334 | continue; |
| 335 | } else { |
| 336 | /* |
| 337 | * Look for the application in the local wait queue and set the |
| 338 | * notify socket if found. |
| 339 | */ |
| 340 | cds_list_for_each_entry_safe ( |
| 341 | wait_node, tmp_wait_node, &wait_queue.head, head) { |
| 342 | health_code_update(); |
| 343 | if (wait_node->app->pid == ust_cmd->reg_msg.pid) { |
| 344 | wait_node->app->notify_sock = ust_cmd->sock; |
| 345 | cds_list_del(&wait_node->head); |
| 346 | wait_queue.count--; |
| 347 | app = wait_node->app; |
| 348 | free(wait_node); |
| 349 | wait_node = nullptr; |
| 350 | DBG3("UST app notify socket %d is set", |
| 351 | ust_cmd->sock); |
| 352 | break; |
| 353 | } |
| 354 | } |
| 355 | |
| 356 | /* |
| 357 | * With no application at this stage the received socket is |
| 358 | * basically useless so close it before we free the cmd data |
| 359 | * structure for good. |
| 360 | */ |
| 361 | if (!app) { |
| 362 | ret = close(ust_cmd->sock); |
| 363 | if (ret < 0) { |
| 364 | PERROR("close ust sock dispatch %d", ust_cmd->sock); |
| 365 | } |
| 366 | lttng_fd_put(LTTNG_FD_APPS, 1); |
| 367 | } |
| 368 | free(ust_cmd); |
| 369 | ust_cmd = nullptr; |
| 370 | } |
| 371 | |
| 372 | if (app) { |
| 373 | /* |
| 374 | * @session_lock_list |
| 375 | * |
| 376 | * Lock the global session list so from the register up to the |
| 377 | * registration done message, no thread can see the application |
| 378 | * and change its state. |
| 379 | */ |
| 380 | session_lock_list(); |
| 381 | lttng::urcu::read_lock_guard read_lock; |
| 382 | |
| 383 | /* |
| 384 | * Add application to the global hash table. This needs to be |
| 385 | * done before the update to the UST registry can locate the |
| 386 | * application. |
| 387 | */ |
| 388 | ust_app_add(app); |
| 389 | |
| 390 | /* Set app version. This call will print an error if needed. */ |
| 391 | (void) ust_app_version(app); |
| 392 | |
| 393 | (void) ust_app_setup_event_notifier_group(app); |
| 394 | |
| 395 | /* Send notify socket through the notify pipe. */ |
| 396 | ret = send_socket_to_thread( |
| 397 | notifiers->apps_cmd_notify_pipe_write_fd, app->notify_sock); |
| 398 | if (ret < 0) { |
| 399 | session_unlock_list(); |
| 400 | /* |
| 401 | * No notify thread, stop the UST tracing. However, this is |
| 402 | * not an internal error of the this thread thus setting |
| 403 | * the health error code to a normal exit. |
| 404 | */ |
| 405 | err = 0; |
| 406 | goto error; |
| 407 | } |
| 408 | |
| 409 | /* |
| 410 | * Update newly registered application with the tracing |
| 411 | * registry info already enabled information. |
| 412 | */ |
| 413 | update_ust_app(app->sock); |
| 414 | |
| 415 | /* |
| 416 | * Don't care about return value. Let the manage apps threads |
| 417 | * handle app unregistration upon socket close. |
| 418 | */ |
| 419 | (void) ust_app_register_done(app); |
| 420 | |
| 421 | /* |
| 422 | * Even if the application socket has been closed, send the app |
| 423 | * to the thread and unregistration will take place at that |
| 424 | * place. |
| 425 | */ |
| 426 | ret = send_socket_to_thread(notifiers->apps_cmd_pipe_write_fd, |
| 427 | app->sock); |
| 428 | if (ret < 0) { |
| 429 | session_unlock_list(); |
| 430 | /* |
| 431 | * No apps. thread, stop the UST tracing. However, this is |
| 432 | * not an internal error of the this thread thus setting |
| 433 | * the health error code to a normal exit. |
| 434 | */ |
| 435 | err = 0; |
| 436 | goto error; |
| 437 | } |
| 438 | |
| 439 | session_unlock_list(); |
| 440 | } |
| 441 | } while (node != nullptr); |
| 442 | |
| 443 | health_poll_entry(); |
| 444 | /* Futex wait on queue. Blocking call on futex() */ |
| 445 | futex_nto1_wait(¬ifiers->ust_cmd_queue->futex); |
| 446 | health_poll_exit(); |
| 447 | } |
| 448 | /* Normal exit, no error */ |
| 449 | err = 0; |
| 450 | |
| 451 | error: |
| 452 | /* Clean up wait queue. */ |
| 453 | cds_list_for_each_entry_safe (wait_node, tmp_wait_node, &wait_queue.head, head) { |
| 454 | cds_list_del(&wait_node->head); |
| 455 | wait_queue.count--; |
| 456 | free(wait_node); |
| 457 | } |
| 458 | |
| 459 | /* Empty command queue. */ |
| 460 | for (;;) { |
| 461 | /* Dequeue command for registration */ |
| 462 | node = cds_wfcq_dequeue_blocking(¬ifiers->ust_cmd_queue->head, |
| 463 | ¬ifiers->ust_cmd_queue->tail); |
| 464 | if (node == nullptr) { |
| 465 | break; |
| 466 | } |
| 467 | ust_cmd = lttng::utils::container_of(node, &ust_command::node); |
| 468 | ret = close(ust_cmd->sock); |
| 469 | if (ret < 0) { |
| 470 | PERROR("close ust sock exit dispatch %d", ust_cmd->sock); |
| 471 | } |
| 472 | lttng_fd_put(LTTNG_FD_APPS, 1); |
| 473 | free(ust_cmd); |
| 474 | } |
| 475 | |
| 476 | error_testpoint: |
| 477 | DBG("Dispatch thread dying"); |
| 478 | if (err) { |
| 479 | health_error(); |
| 480 | ERR("Health error occurred in %s", __func__); |
| 481 | } |
| 482 | health_unregister(the_health_sessiond); |
| 483 | rcu_unregister_thread(); |
| 484 | return nullptr; |
| 485 | } |
| 486 | |
| 487 | static bool shutdown_ust_dispatch_thread(void *data) |
| 488 | { |
| 489 | struct thread_notifiers *notifiers = (thread_notifiers *) data; |
| 490 | |
| 491 | CMM_STORE_SHARED(notifiers->dispatch_thread_exit, 1); |
| 492 | futex_nto1_wake(¬ifiers->ust_cmd_queue->futex); |
| 493 | return true; |
| 494 | } |
| 495 | |
| 496 | bool launch_ust_dispatch_thread(struct ust_cmd_queue *cmd_queue, |
| 497 | int apps_cmd_pipe_write_fd, |
| 498 | int apps_cmd_notify_pipe_write_fd) |
| 499 | { |
| 500 | struct lttng_thread *thread; |
| 501 | struct thread_notifiers *notifiers; |
| 502 | |
| 503 | notifiers = zmalloc<thread_notifiers>(); |
| 504 | if (!notifiers) { |
| 505 | goto error; |
| 506 | } |
| 507 | notifiers->ust_cmd_queue = cmd_queue; |
| 508 | notifiers->apps_cmd_pipe_write_fd = apps_cmd_pipe_write_fd; |
| 509 | notifiers->apps_cmd_notify_pipe_write_fd = apps_cmd_notify_pipe_write_fd; |
| 510 | |
| 511 | thread = lttng_thread_create("UST registration dispatch", |
| 512 | thread_dispatch_ust_registration, |
| 513 | shutdown_ust_dispatch_thread, |
| 514 | cleanup_ust_dispatch_thread, |
| 515 | notifiers); |
| 516 | if (!thread) { |
| 517 | goto error; |
| 518 | } |
| 519 | lttng_thread_put(thread); |
| 520 | return true; |
| 521 | error: |
| 522 | free(notifiers); |
| 523 | return false; |
| 524 | } |