SoW-2020-0002: Trace Hit Counters: trigger error reporting integration
[lttng-tools.git] / src / bin / lttng-sessiond / dispatch.c
1 /*
2 * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
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 <stddef.h>
11 #include <stdlib.h>
12 #include <urcu.h>
13 #include <common/futex.h>
14 #include <common/macros.h>
15
16 #include "dispatch.h"
17 #include "ust-app.h"
18 #include "testpoint.h"
19 #include "fd-limit.h"
20 #include "health-sessiond.h"
21 #include "lttng-sessiond.h"
22 #include "thread.h"
23
24 struct thread_notifiers {
25 struct ust_cmd_queue *ust_cmd_queue;
26 int apps_cmd_pipe_write_fd;
27 int apps_cmd_notify_pipe_write_fd;
28 int dispatch_thread_exit;
29 };
30
31 /*
32 * For each tracing session, update newly registered apps. The session list
33 * lock MUST be acquired before calling this.
34 */
35 static void update_ust_app(int app_sock)
36 {
37 struct ltt_session *sess, *stmp;
38 const struct ltt_session_list *session_list = session_get_list();
39 struct ust_app *app;
40
41 /* Consumer is in an ERROR state. Stop any application update. */
42 if (uatomic_read(&ust_consumerd_state) == CONSUMER_ERROR) {
43 /* Stop the update process since the consumer is dead. */
44 return;
45 }
46
47 rcu_read_lock();
48 assert(app_sock >= 0);
49 app = ust_app_find_by_sock(app_sock);
50 if (app == NULL) {
51 /*
52 * Application can be unregistered before so
53 * this is possible hence simply stopping the
54 * update.
55 */
56 DBG3("UST app update failed to find app sock %d",
57 app_sock);
58 goto unlock_rcu;
59 }
60
61 /* Update all tokens for the app */
62 ust_app_global_update_tokens(app);
63
64 /* For all tracing session(s) */
65 cds_list_for_each_entry_safe(sess, stmp, &session_list->head, list) {
66 if (!session_get(sess)) {
67 continue;
68 }
69 session_lock(sess);
70 if (!sess->active || !sess->ust_session) {
71 goto unlock_session;
72 }
73
74 ust_app_global_update(sess->ust_session, app);
75 unlock_session:
76 session_unlock(sess);
77 session_put(sess);
78 }
79
80 unlock_rcu:
81 rcu_read_unlock();
82
83 }
84
85 /*
86 * Sanitize the wait queue of the dispatch registration thread meaning removing
87 * invalid nodes from it. This is to avoid memory leaks for the case the UST
88 * notify socket is never received.
89 */
90 static void sanitize_wait_queue(struct ust_reg_wait_queue *wait_queue)
91 {
92 int ret, nb_fd = 0, i;
93 unsigned int fd_added = 0;
94 struct lttng_poll_event events;
95 struct ust_reg_wait_node *wait_node = NULL, *tmp_wait_node;
96
97 assert(wait_queue);
98
99 lttng_poll_init(&events);
100
101 /* Just skip everything for an empty queue. */
102 if (!wait_queue->count) {
103 goto end;
104 }
105
106 ret = lttng_poll_create(&events, wait_queue->count, LTTNG_CLOEXEC);
107 if (ret < 0) {
108 goto error_create;
109 }
110
111 cds_list_for_each_entry_safe(wait_node, tmp_wait_node,
112 &wait_queue->head, head) {
113 assert(wait_node->app);
114 ret = lttng_poll_add(&events, wait_node->app->sock,
115 LPOLLHUP | LPOLLERR);
116 if (ret < 0) {
117 goto error;
118 }
119
120 fd_added = 1;
121 }
122
123 if (!fd_added) {
124 goto end;
125 }
126
127 /*
128 * Poll but don't block so we can quickly identify the faulty events and
129 * clean them afterwards from the wait queue.
130 */
131 ret = lttng_poll_wait(&events, 0);
132 if (ret < 0) {
133 goto error;
134 }
135 nb_fd = ret;
136
137 for (i = 0; i < nb_fd; i++) {
138 /* Get faulty FD. */
139 uint32_t revents = LTTNG_POLL_GETEV(&events, i);
140 int pollfd = LTTNG_POLL_GETFD(&events, i);
141
142 cds_list_for_each_entry_safe(wait_node, tmp_wait_node,
143 &wait_queue->head, head) {
144 if (pollfd == wait_node->app->sock &&
145 (revents & (LPOLLHUP | LPOLLERR))) {
146 cds_list_del(&wait_node->head);
147 wait_queue->count--;
148 ust_app_destroy(wait_node->app);
149 free(wait_node);
150 /*
151 * Silence warning of use-after-free in
152 * cds_list_for_each_entry_safe which uses
153 * __typeof__(*wait_node).
154 */
155 wait_node = NULL;
156 break;
157 } else {
158 ERR("Unexpected poll events %u for sock %d", revents, pollfd);
159 goto error;
160 }
161 }
162 }
163
164 if (nb_fd > 0) {
165 DBG("Wait queue sanitized, %d node were cleaned up", nb_fd);
166 }
167
168 end:
169 lttng_poll_clean(&events);
170 return;
171
172 error:
173 lttng_poll_clean(&events);
174 error_create:
175 ERR("Unable to sanitize wait queue");
176 return;
177 }
178
179 /*
180 * Send a socket to a thread This is called from the dispatch UST registration
181 * thread once all sockets are set for the application.
182 *
183 * The sock value can be invalid, we don't really care, the thread will handle
184 * it and make the necessary cleanup if so.
185 *
186 * On success, return 0 else a negative value being the errno message of the
187 * write().
188 */
189 static int send_socket_to_thread(int fd, int sock)
190 {
191 ssize_t ret;
192
193 /*
194 * It's possible that the FD is set as invalid with -1 concurrently just
195 * before calling this function being a shutdown state of the thread.
196 */
197 if (fd < 0) {
198 ret = -EBADF;
199 goto error;
200 }
201
202 ret = lttng_write(fd, &sock, sizeof(sock));
203 if (ret < sizeof(sock)) {
204 PERROR("write apps pipe %d", fd);
205 if (ret < 0) {
206 ret = -errno;
207 }
208 goto error;
209 }
210
211 /* All good. Don't send back the write positive ret value. */
212 ret = 0;
213 error:
214 return (int) ret;
215 }
216
217 static void cleanup_ust_dispatch_thread(void *data)
218 {
219 free(data);
220 }
221
222 /*
223 * Dispatch request from the registration threads to the application
224 * communication thread.
225 */
226 static void *thread_dispatch_ust_registration(void *data)
227 {
228 int ret, err = -1;
229 struct cds_wfcq_node *node;
230 struct ust_command *ust_cmd = NULL;
231 struct ust_reg_wait_node *wait_node = NULL, *tmp_wait_node;
232 struct ust_reg_wait_queue wait_queue = {
233 .count = 0,
234 };
235 struct thread_notifiers *notifiers = data;
236
237 rcu_register_thread();
238
239 health_register(health_sessiond, HEALTH_SESSIOND_TYPE_APP_REG_DISPATCH);
240
241 if (testpoint(sessiond_thread_app_reg_dispatch)) {
242 goto error_testpoint;
243 }
244
245 health_code_update();
246
247 CDS_INIT_LIST_HEAD(&wait_queue.head);
248
249 DBG("[thread] Dispatch UST command started");
250
251 for (;;) {
252 health_code_update();
253
254 /* Atomically prepare the queue futex */
255 futex_nto1_prepare(&notifiers->ust_cmd_queue->futex);
256
257 if (CMM_LOAD_SHARED(notifiers->dispatch_thread_exit)) {
258 break;
259 }
260
261 do {
262 struct ust_app *app = NULL;
263 ust_cmd = NULL;
264
265 /*
266 * Make sure we don't have node(s) that have hung up before receiving
267 * the notify socket. This is to clean the list in order to avoid
268 * memory leaks from notify socket that are never seen.
269 */
270 sanitize_wait_queue(&wait_queue);
271
272 health_code_update();
273 /* Dequeue command for registration */
274 node = cds_wfcq_dequeue_blocking(
275 &notifiers->ust_cmd_queue->head,
276 &notifiers->ust_cmd_queue->tail);
277 if (node == NULL) {
278 DBG("Woken up but nothing in the UST command queue");
279 /* Continue thread execution */
280 break;
281 }
282
283 ust_cmd = caa_container_of(node, struct ust_command, node);
284
285 DBG("Dispatching UST registration pid:%d ppid:%d uid:%d"
286 " gid:%d sock:%d name:%s (version %d.%d)",
287 ust_cmd->reg_msg.pid, ust_cmd->reg_msg.ppid,
288 ust_cmd->reg_msg.uid, ust_cmd->reg_msg.gid,
289 ust_cmd->sock, ust_cmd->reg_msg.name,
290 ust_cmd->reg_msg.major, ust_cmd->reg_msg.minor);
291
292 if (ust_cmd->reg_msg.type == USTCTL_SOCKET_CMD) {
293 wait_node = zmalloc(sizeof(*wait_node));
294 if (!wait_node) {
295 PERROR("zmalloc wait_node dispatch");
296 ret = close(ust_cmd->sock);
297 if (ret < 0) {
298 PERROR("close ust sock dispatch %d", ust_cmd->sock);
299 }
300 lttng_fd_put(LTTNG_FD_APPS, 1);
301 free(ust_cmd);
302 ust_cmd = NULL;
303 goto error;
304 }
305 CDS_INIT_LIST_HEAD(&wait_node->head);
306
307 /* Create application object if socket is CMD. */
308 wait_node->app = ust_app_create(&ust_cmd->reg_msg,
309 ust_cmd->sock);
310 if (!wait_node->app) {
311 ret = close(ust_cmd->sock);
312 if (ret < 0) {
313 PERROR("close ust sock dispatch %d", ust_cmd->sock);
314 }
315 lttng_fd_put(LTTNG_FD_APPS, 1);
316 free(wait_node);
317 wait_node = NULL;
318 free(ust_cmd);
319 ust_cmd = NULL;
320 continue;
321 }
322 /*
323 * Add application to the wait queue so we can set the notify
324 * socket before putting this object in the global ht.
325 */
326 cds_list_add(&wait_node->head, &wait_queue.head);
327 wait_queue.count++;
328
329 free(ust_cmd);
330 ust_cmd = NULL;
331 /*
332 * We have to continue here since we don't have the notify
333 * socket and the application MUST be added to the hash table
334 * only at that moment.
335 */
336 continue;
337 } else {
338 /*
339 * Look for the application in the local wait queue and set the
340 * notify socket if found.
341 */
342 cds_list_for_each_entry_safe(wait_node, tmp_wait_node,
343 &wait_queue.head, head) {
344 health_code_update();
345 if (wait_node->app->pid == ust_cmd->reg_msg.pid) {
346 wait_node->app->notify_sock = ust_cmd->sock;
347 cds_list_del(&wait_node->head);
348 wait_queue.count--;
349 app = wait_node->app;
350 free(wait_node);
351 wait_node = NULL;
352 DBG3("UST app notify socket %d is set", ust_cmd->sock);
353 break;
354 }
355 }
356
357 /*
358 * With no application at this stage the received socket is
359 * basically useless so close it before we free the cmd data
360 * structure for good.
361 */
362 if (!app) {
363 ret = close(ust_cmd->sock);
364 if (ret < 0) {
365 PERROR("close ust sock dispatch %d", ust_cmd->sock);
366 }
367 lttng_fd_put(LTTNG_FD_APPS, 1);
368 }
369 free(ust_cmd);
370 ust_cmd = NULL;
371 }
372
373 if (app) {
374 /*
375 * @session_lock_list
376 *
377 * Lock the global session list so from the register up to the
378 * registration done message, no thread can see the application
379 * and change its state.
380 */
381 session_lock_list();
382 rcu_read_lock();
383
384 /*
385 * Add application to the global hash table. This needs to be
386 * done before the update to the UST registry can locate the
387 * application.
388 */
389 ust_app_add(app);
390
391 /* Set app version. This call will print an error if needed. */
392 (void) ust_app_version(app);
393
394 (void) ust_app_setup_trigger_group(app);
395
396 /* Send notify socket through the notify pipe. */
397 ret = send_socket_to_thread(
398 notifiers->apps_cmd_notify_pipe_write_fd,
399 app->notify_sock);
400 if (ret < 0) {
401 rcu_read_unlock();
402 session_unlock_list();
403 /*
404 * No notify thread, stop the UST tracing. However, this is
405 * not an internal error of the this thread thus setting
406 * the health error code to a normal exit.
407 */
408 err = 0;
409 goto error;
410 }
411
412 /*
413 * Update newly registered application with the tracing
414 * registry info already enabled information.
415 */
416 update_ust_app(app->sock);
417
418 /*
419 * Don't care about return value. Let the manage apps threads
420 * handle app unregistration upon socket close.
421 */
422 (void) ust_app_register_done(app);
423
424 /*
425 * Even if the application socket has been closed, send the app
426 * to the thread and unregistration will take place at that
427 * place.
428 */
429 ret = send_socket_to_thread(
430 notifiers->apps_cmd_pipe_write_fd,
431 app->sock);
432 if (ret < 0) {
433 rcu_read_unlock();
434 session_unlock_list();
435 /*
436 * No apps. thread, stop the UST tracing. However, this is
437 * not an internal error of the this thread thus setting
438 * the health error code to a normal exit.
439 */
440 err = 0;
441 goto error;
442 }
443
444 rcu_read_unlock();
445 session_unlock_list();
446 }
447 } while (node != NULL);
448
449 health_poll_entry();
450 /* Futex wait on queue. Blocking call on futex() */
451 futex_nto1_wait(&notifiers->ust_cmd_queue->futex);
452 health_poll_exit();
453 }
454 /* Normal exit, no error */
455 err = 0;
456
457 error:
458 /* Clean up wait queue. */
459 cds_list_for_each_entry_safe(wait_node, tmp_wait_node,
460 &wait_queue.head, head) {
461 cds_list_del(&wait_node->head);
462 wait_queue.count--;
463 free(wait_node);
464 }
465
466 /* Empty command queue. */
467 for (;;) {
468 /* Dequeue command for registration */
469 node = cds_wfcq_dequeue_blocking(
470 &notifiers->ust_cmd_queue->head,
471 &notifiers->ust_cmd_queue->tail);
472 if (node == NULL) {
473 break;
474 }
475 ust_cmd = caa_container_of(node, struct ust_command, node);
476 ret = close(ust_cmd->sock);
477 if (ret < 0) {
478 PERROR("close ust sock exit dispatch %d", ust_cmd->sock);
479 }
480 lttng_fd_put(LTTNG_FD_APPS, 1);
481 free(ust_cmd);
482 }
483
484 error_testpoint:
485 DBG("Dispatch thread dying");
486 if (err) {
487 health_error();
488 ERR("Health error occurred in %s", __func__);
489 }
490 health_unregister(health_sessiond);
491 rcu_unregister_thread();
492 return NULL;
493 }
494
495 static bool shutdown_ust_dispatch_thread(void *data)
496 {
497 struct thread_notifiers *notifiers = data;
498
499 CMM_STORE_SHARED(notifiers->dispatch_thread_exit, 1);
500 futex_nto1_wake(&notifiers->ust_cmd_queue->futex);
501 return true;
502 }
503
504 bool launch_ust_dispatch_thread(struct ust_cmd_queue *cmd_queue,
505 int apps_cmd_pipe_write_fd,
506 int apps_cmd_notify_pipe_write_fd)
507 {
508 struct lttng_thread *thread;
509 struct thread_notifiers *notifiers;
510
511 notifiers = zmalloc(sizeof(*notifiers));
512 if (!notifiers) {
513 goto error;
514 }
515 notifiers->ust_cmd_queue = cmd_queue;
516 notifiers->apps_cmd_pipe_write_fd = apps_cmd_pipe_write_fd;
517 notifiers->apps_cmd_notify_pipe_write_fd = apps_cmd_notify_pipe_write_fd;
518
519 thread = lttng_thread_create("UST registration dispatch",
520 thread_dispatch_ust_registration,
521 shutdown_ust_dispatch_thread,
522 cleanup_ust_dispatch_thread,
523 notifiers);
524 if (!thread) {
525 goto error;
526 }
527 lttng_thread_put(thread);
528 return true;
529 error:
530 free(notifiers);
531 return false;
532 }
This page took 0.063747 seconds and 5 git commands to generate.