projects / lttng-ust.git / blob
? search:
summary | shortlog | log | commit | commitdiff | tree
blame | history | raw | HEAD
Revert "Use initial-exec TLS model"
[lttng-ust.git] / liblttng-ust / lttng-ust-comm.c
1 /*
2 * lttng-ust-comm.c
3 *
4 * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
5 * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
6 *
7 * This library is free software; you can redistribute it and/or
8 * modify it under the terms of the GNU Lesser General Public
9 * License as published by the Free Software Foundation; only
10 * version 2.1 of the License.
11 *
12 * This library is distributed in the hope that it will be useful,
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15 * Lesser General Public License for more details.
16 *
17 * You should have received a copy of the GNU Lesser General Public
18 * License along with this library; if not, write to the Free Software
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
20 */
21
22 #define _LGPL_SOURCE
23 #define _GNU_SOURCE
24 #include <sys/types.h>
25 #include <sys/socket.h>
26 #include <sys/mman.h>
27 #include <sys/stat.h>
28 #include <sys/types.h>
29 #include <sys/wait.h>
30 #include <fcntl.h>
31 #include <unistd.h>
32 #include <errno.h>
33 #include <pthread.h>
34 #include <semaphore.h>
35 #include <time.h>
36 #include <assert.h>
37 #include <signal.h>
38 #include <limits.h>
39 #include <urcu/uatomic.h>
40 #include <urcu/futex.h>
41 #include <urcu/compiler.h>
42
43 #include <lttng/ust-events.h>
44 #include <lttng/ust-abi.h>
45 #include <lttng/ust.h>
46 #include <lttng/ust-error.h>
47 #include <lttng/ust-ctl.h>
48 #include <urcu/tls-compat.h>
49 #include <ust-comm.h>
50 #include <ust-fd.h>
51 #include <usterr-signal-safe.h>
52 #include <helper.h>
53 #include "tracepoint-internal.h"
54 #include "lttng-tracer-core.h"
55 #include "compat.h"
56 #include "../libringbuffer/rb-init.h"
57 #include "lttng-ust-statedump.h"
58 #include "clock.h"
59 #include "../libringbuffer/getcpu.h"
60 #include "getenv.h"
61
62 /*
63 * Has lttng ust comm constructor been called ?
64 */
65 static int initialized;
66
67 /*
68 * The ust_lock/ust_unlock lock is used as a communication thread mutex.
69 * Held when handling a command, also held by fork() to deal with
70 * removal of threads, and by exit path.
71 *
72 * The UST lock is the centralized mutex across UST tracing control and
73 * probe registration.
74 *
75 * ust_exit_mutex must never nest in ust_mutex.
76 *
77 * ust_fork_mutex must never nest in ust_mutex.
78 *
79 * ust_mutex_nest is a per-thread nesting counter, allowing the perf
80 * counter lazy initialization called by events within the statedump,
81 * which traces while the ust_mutex is held.
82 *
83 * ust_lock nests within the dynamic loader lock (within glibc) because
84 * it is taken within the library constructor.
85 */
86 static pthread_mutex_t ust_mutex = PTHREAD_MUTEX_INITIALIZER;
87
88 /* Allow nesting the ust_mutex within the same thread. */
89 static DEFINE_URCU_TLS(int, ust_mutex_nest);
90
91 /*
92 * ust_exit_mutex protects thread_active variable wrt thread exit. It
93 * cannot be done by ust_mutex because pthread_cancel(), which takes an
94 * internal libc lock, cannot nest within ust_mutex.
95 *
96 * It never nests within a ust_mutex.
97 */
98 static pthread_mutex_t ust_exit_mutex = PTHREAD_MUTEX_INITIALIZER;
99
100 /*
101 * ust_fork_mutex protects base address statedump tracing against forks. It
102 * prevents the dynamic loader lock to be taken (by base address statedump
103 * tracing) while a fork is happening, thus preventing deadlock issues with
104 * the dynamic loader lock.
105 */
106 static pthread_mutex_t ust_fork_mutex = PTHREAD_MUTEX_INITIALIZER;
107
108 /* Should the ust comm thread quit ? */
109 static int lttng_ust_comm_should_quit;
110
111 /*
112 * This variable can be tested by applications to check whether
113 * lttng-ust is loaded. They simply have to define their own
114 * "lttng_ust_loaded" weak symbol, and test it. It is set to 1 by the
115 * library constructor.
116 */
117 int lttng_ust_loaded __attribute__((weak));
118
119 /*
120 * Return 0 on success, -1 if should quit.
121 * The lock is taken in both cases.
122 * Signal-safe.
123 */
124 int ust_lock(void)
125 {
126 sigset_t sig_all_blocked, orig_mask;
127 int ret, oldstate;
128
129 ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
130 if (ret) {
131 ERR("pthread_setcancelstate: %s", strerror(ret));
132 }
133 if (oldstate != PTHREAD_CANCEL_ENABLE) {
134 ERR("pthread_setcancelstate: unexpected oldstate");
135 }
136 sigfillset(&sig_all_blocked);
137 ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
138 if (ret) {
139 ERR("pthread_sigmask: %s", strerror(ret));
140 }
141 if (!URCU_TLS(ust_mutex_nest)++)
142 pthread_mutex_lock(&ust_mutex);
143 ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
144 if (ret) {
145 ERR("pthread_sigmask: %s", strerror(ret));
146 }
147 if (lttng_ust_comm_should_quit) {
148 return -1;
149 } else {
150 return 0;
151 }
152 }
153
154 /*
155 * ust_lock_nocheck() can be used in constructors/destructors, because
156 * they are already nested within the dynamic loader lock, and therefore
157 * have exclusive access against execution of liblttng-ust destructor.
158 * Signal-safe.
159 */
160 void ust_lock_nocheck(void)
161 {
162 sigset_t sig_all_blocked, orig_mask;
163 int ret, oldstate;
164
165 ret = pthread_setcancelstate(PTHREAD_CANCEL_DISABLE, &oldstate);
166 if (ret) {
167 ERR("pthread_setcancelstate: %s", strerror(ret));
168 }
169 if (oldstate != PTHREAD_CANCEL_ENABLE) {
170 ERR("pthread_setcancelstate: unexpected oldstate");
171 }
172 sigfillset(&sig_all_blocked);
173 ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
174 if (ret) {
175 ERR("pthread_sigmask: %s", strerror(ret));
176 }
177 if (!URCU_TLS(ust_mutex_nest)++)
178 pthread_mutex_lock(&ust_mutex);
179 ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
180 if (ret) {
181 ERR("pthread_sigmask: %s", strerror(ret));
182 }
183 }
184
185 /*
186 * Signal-safe.
187 */
188 void ust_unlock(void)
189 {
190 sigset_t sig_all_blocked, orig_mask;
191 int ret, oldstate;
192
193 sigfillset(&sig_all_blocked);
194 ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_mask);
195 if (ret) {
196 ERR("pthread_sigmask: %s", strerror(ret));
197 }
198 if (!--URCU_TLS(ust_mutex_nest))
199 pthread_mutex_unlock(&ust_mutex);
200 ret = pthread_sigmask(SIG_SETMASK, &orig_mask, NULL);
201 if (ret) {
202 ERR("pthread_sigmask: %s", strerror(ret));
203 }
204 ret = pthread_setcancelstate(PTHREAD_CANCEL_ENABLE, &oldstate);
205 if (ret) {
206 ERR("pthread_setcancelstate: %s", strerror(ret));
207 }
208 if (oldstate != PTHREAD_CANCEL_DISABLE) {
209 ERR("pthread_setcancelstate: unexpected oldstate");
210 }
211 }
212
213 /*
214 * Wait for either of these before continuing to the main
215 * program:
216 * - the register_done message from sessiond daemon
217 * (will let the sessiond daemon enable sessions before main
218 * starts.)
219 * - sessiond daemon is not reachable.
220 * - timeout (ensuring applications are resilient to session
221 * daemon problems).
222 */
223 static sem_t constructor_wait;
224 /*
225 * Doing this for both the global and local sessiond.
226 */
227 static int sem_count = { 2 };
228
229 /*
230 * Counting nesting within lttng-ust. Used to ensure that calling fork()
231 * from liblttng-ust does not execute the pre/post fork handlers.
232 */
233 static DEFINE_URCU_TLS(int, lttng_ust_nest_count);
234
235 /*
236 * Info about socket and associated listener thread.
237 */
238 struct sock_info {
239 const char *name;
240 pthread_t ust_listener; /* listener thread */
241 int root_handle;
242 int constructor_sem_posted;
243 int allowed;
244 int global;
245 int thread_active;
246
247 char sock_path[PATH_MAX];
248 int socket;
249 int notify_socket;
250
251 char wait_shm_path[PATH_MAX];
252 char *wait_shm_mmap;
253 /* Keep track of lazy state dump not performed yet. */
254 int statedump_pending;
255 };
256
257 /* Socket from app (connect) to session daemon (listen) for communication */
258 struct sock_info global_apps = {
259 .name = "global",
260 .global = 1,
261
262 .root_handle = -1,
263 .allowed = 1,
264 .thread_active = 0,
265
266 .sock_path = LTTNG_DEFAULT_RUNDIR "/" LTTNG_UST_SOCK_FILENAME,
267 .socket = -1,
268 .notify_socket = -1,
269
270 .wait_shm_path = "/" LTTNG_UST_WAIT_FILENAME,
271
272 .statedump_pending = 0,
273 };
274
275 /* TODO: allow global_apps_sock_path override */
276
277 struct sock_info local_apps = {
278 .name = "local",
279 .global = 0,
280 .root_handle = -1,
281 .allowed = 0, /* Check setuid bit first */
282 .thread_active = 0,
283
284 .socket = -1,
285 .notify_socket = -1,
286
287 .statedump_pending = 0,
288 };
289
290 static int wait_poll_fallback;
291
292 static const char *cmd_name_mapping[] = {
293 [ LTTNG_UST_RELEASE ] = "Release",
294 [ LTTNG_UST_SESSION ] = "Create Session",
295 [ LTTNG_UST_TRACER_VERSION ] = "Get Tracer Version",
296
297 [ LTTNG_UST_TRACEPOINT_LIST ] = "Create Tracepoint List",
298 [ LTTNG_UST_WAIT_QUIESCENT ] = "Wait for Quiescent State",
299 [ LTTNG_UST_REGISTER_DONE ] = "Registration Done",
300 [ LTTNG_UST_TRACEPOINT_FIELD_LIST ] = "Create Tracepoint Field List",
301
302 /* Session FD commands */
303 [ LTTNG_UST_CHANNEL ] = "Create Channel",
304 [ LTTNG_UST_SESSION_START ] = "Start Session",
305 [ LTTNG_UST_SESSION_STOP ] = "Stop Session",
306
307 /* Channel FD commands */
308 [ LTTNG_UST_STREAM ] = "Create Stream",
309 [ LTTNG_UST_EVENT ] = "Create Event",
310
311 /* Event and Channel FD commands */
312 [ LTTNG_UST_CONTEXT ] = "Create Context",
313 [ LTTNG_UST_FLUSH_BUFFER ] = "Flush Buffer",
314
315 /* Event, Channel and Session commands */
316 [ LTTNG_UST_ENABLE ] = "Enable",
317 [ LTTNG_UST_DISABLE ] = "Disable",
318
319 /* Tracepoint list commands */
320 [ LTTNG_UST_TRACEPOINT_LIST_GET ] = "List Next Tracepoint",
321 [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET ] = "List Next Tracepoint Field",
322
323 /* Event FD commands */
324 [ LTTNG_UST_FILTER ] = "Create Filter",
325 [ LTTNG_UST_EXCLUSION ] = "Add exclusions to event",
326 };
327
328 static const char *str_timeout;
329 static int got_timeout_env;
330
331 extern void lttng_ring_buffer_client_overwrite_init(void);
332 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
333 extern void lttng_ring_buffer_client_discard_init(void);
334 extern void lttng_ring_buffer_client_discard_rt_init(void);
335 extern void lttng_ring_buffer_metadata_client_init(void);
336 extern void lttng_ring_buffer_client_overwrite_exit(void);
337 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
338 extern void lttng_ring_buffer_client_discard_exit(void);
339 extern void lttng_ring_buffer_client_discard_rt_exit(void);
340 extern void lttng_ring_buffer_metadata_client_exit(void);
341
342 ssize_t lttng_ust_read(int fd, void *buf, size_t len)
343 {
344 ssize_t ret;
345 size_t copied = 0, to_copy = len;
346
347 do {
348 ret = read(fd, buf + copied, to_copy);
349 if (ret > 0) {
350 copied += ret;
351 to_copy -= ret;
352 }
353 } while ((ret > 0 && to_copy > 0)
354 || (ret < 0 && errno == EINTR));
355 if (ret > 0) {
356 ret = copied;
357 }
358 return ret;
359 }
360 /*
361 * Returns the HOME directory path. Caller MUST NOT free(3) the returned
362 * pointer.
363 */
364 static
365 const char *get_lttng_home_dir(void)
366 {
367 const char *val;
368
369 val = (const char *) lttng_getenv("LTTNG_HOME");
370 if (val != NULL) {
371 return val;
372 }
373 return (const char *) lttng_getenv("HOME");
374 }
375
376 /*
377 * Force a read (imply TLS fixup for dlopen) of TLS variables.
378 */
379 static
380 void lttng_fixup_nest_count_tls(void)
381 {
382 asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count)));
383 }
384
385 static
386 void lttng_fixup_ust_mutex_nest_tls(void)
387 {
388 asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest)));
389 }
390
391 /*
392 * Fixup urcu bp TLS.
393 */
394 static
395 void lttng_fixup_urcu_bp_tls(void)
396 {
397 rcu_read_lock();
398 rcu_read_unlock();
399 }
400
401 void lttng_ust_fixup_tls(void)
402 {
403 lttng_fixup_urcu_bp_tls();
404 lttng_fixup_ringbuffer_tls();
405 lttng_fixup_vtid_tls();
406 lttng_fixup_nest_count_tls();
407 lttng_fixup_procname_tls();
408 lttng_fixup_ust_mutex_nest_tls();
409 lttng_ust_fixup_fd_tracker_tls();
410 }
411
412 int lttng_get_notify_socket(void *owner)
413 {
414 struct sock_info *info = owner;
415
416 return info->notify_socket;
417 }
418
419 static
420 void print_cmd(int cmd, int handle)
421 {
422 const char *cmd_name = "Unknown";
423
424 if (cmd >= 0 && cmd < LTTNG_ARRAY_SIZE(cmd_name_mapping)
425 && cmd_name_mapping[cmd]) {
426 cmd_name = cmd_name_mapping[cmd];
427 }
428 DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
429 cmd_name, cmd,
430 lttng_ust_obj_get_name(handle), handle);
431 }
432
433 static
434 int setup_local_apps(void)
435 {
436 const char *home_dir;
437 uid_t uid;
438
439 uid = getuid();
440 /*
441 * Disallow per-user tracing for setuid binaries.
442 */
443 if (uid != geteuid()) {
444 assert(local_apps.allowed == 0);
445 return 0;
446 }
447 home_dir = get_lttng_home_dir();
448 if (!home_dir) {
449 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
450 assert(local_apps.allowed == 0);
451 return -ENOENT;
452 }
453 local_apps.allowed = 1;
454 snprintf(local_apps.sock_path, PATH_MAX, "%s/%s/%s",
455 home_dir,
456 LTTNG_DEFAULT_HOME_RUNDIR,
457 LTTNG_UST_SOCK_FILENAME);
458 snprintf(local_apps.wait_shm_path, PATH_MAX, "/%s-%u",
459 LTTNG_UST_WAIT_FILENAME,
460 uid);
461 return 0;
462 }
463
464 /*
465 * Get socket timeout, in ms.
466 * -1: wait forever. 0: don't wait. >0: timeout, in ms.
467 */
468 static
469 long get_timeout(void)
470 {
471 long constructor_delay_ms = LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS;
472
473 if (!got_timeout_env) {
474 str_timeout = lttng_getenv("LTTNG_UST_REGISTER_TIMEOUT");
475 got_timeout_env = 1;
476 }
477 if (str_timeout)
478 constructor_delay_ms = strtol(str_timeout, NULL, 10);
479 /* All negative values are considered as "-1". */
480 if (constructor_delay_ms < -1)
481 constructor_delay_ms = -1;
482 return constructor_delay_ms;
483 }
484
485 /* Timeout for notify socket send and recv. */
486 static
487 long get_notify_sock_timeout(void)
488 {
489 return get_timeout();
490 }
491
492 /* Timeout for connecting to cmd and notify sockets. */
493 static
494 long get_connect_sock_timeout(void)
495 {
496 return get_timeout();
497 }
498
499 /*
500 * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
501 */
502 static
503 int get_constructor_timeout(struct timespec *constructor_timeout)
504 {
505 long constructor_delay_ms;
506 int ret;
507
508 constructor_delay_ms = get_timeout();
509
510 switch (constructor_delay_ms) {
511 case -1:/* fall-through */
512 case 0:
513 return constructor_delay_ms;
514 default:
515 break;
516 }
517
518 /*
519 * If we are unable to find the current time, don't wait.
520 */
521 ret = clock_gettime(CLOCK_REALTIME, constructor_timeout);
522 if (ret) {
523 /* Don't wait. */
524 return 0;
525 }
526 constructor_timeout->tv_sec += constructor_delay_ms / 1000UL;
527 constructor_timeout->tv_nsec +=
528 (constructor_delay_ms % 1000UL) * 1000000UL;
529 if (constructor_timeout->tv_nsec >= 1000000000UL) {
530 constructor_timeout->tv_sec++;
531 constructor_timeout->tv_nsec -= 1000000000UL;
532 }
533 /* Timeout wait (constructor_delay_ms). */
534 return 1;
535 }
536
537 static
538 void get_allow_blocking(void)
539 {
540 const char *str_allow_blocking =
541 lttng_getenv("LTTNG_UST_ALLOW_BLOCKING");
542
543 if (str_allow_blocking) {
544 DBG("%s environment variable is set",
545 "LTTNG_UST_ALLOW_BLOCKING");
546 lttng_ust_ringbuffer_set_allow_blocking();
547 }
548 }
549
550 static
551 int register_to_sessiond(int socket, enum ustctl_socket_type type)
552 {
553 return ustcomm_send_reg_msg(socket,
554 type,
555 CAA_BITS_PER_LONG,
556 lttng_alignof(uint8_t) * CHAR_BIT,
557 lttng_alignof(uint16_t) * CHAR_BIT,
558 lttng_alignof(uint32_t) * CHAR_BIT,
559 lttng_alignof(uint64_t) * CHAR_BIT,
560 lttng_alignof(unsigned long) * CHAR_BIT);
561 }
562
563 static
564 int send_reply(int sock, struct ustcomm_ust_reply *lur)
565 {
566 ssize_t len;
567
568 len = ustcomm_send_unix_sock(sock, lur, sizeof(*lur));
569 switch (len) {
570 case sizeof(*lur):
571 DBG("message successfully sent");
572 return 0;
573 default:
574 if (len == -ECONNRESET) {
575 DBG("remote end closed connection");
576 return 0;
577 }
578 if (len < 0)
579 return len;
580 DBG("incorrect message size: %zd", len);
581 return -EINVAL;
582 }
583 }
584
585 static
586 int handle_register_done(struct sock_info *sock_info)
587 {
588 int ret;
589
590 if (sock_info->constructor_sem_posted)
591 return 0;
592 sock_info->constructor_sem_posted = 1;
593 if (uatomic_read(&sem_count) <= 0) {
594 return 0;
595 }
596 ret = uatomic_add_return(&sem_count, -1);
597 if (ret == 0) {
598 ret = sem_post(&constructor_wait);
599 assert(!ret);
600 }
601 return 0;
602 }
603
604 /*
605 * Only execute pending statedump after the constructor semaphore has
606 * been posted by each listener thread. This means statedump will only
607 * be performed after the "registration done" command is received from
608 * each session daemon the application is connected to.
609 *
610 * This ensures we don't run into deadlock issues with the dynamic
611 * loader mutex, which is held while the constructor is called and
612 * waiting on the constructor semaphore. All operations requiring this
613 * dynamic loader lock need to be postponed using this mechanism.
614 */
615 static
616 void handle_pending_statedump(struct sock_info *sock_info)
617 {
618 int ctor_passed = sock_info->constructor_sem_posted;
619
620 if (ctor_passed && sock_info->statedump_pending) {
621 sock_info->statedump_pending = 0;
622 pthread_mutex_lock(&ust_fork_mutex);
623 lttng_handle_pending_statedump(sock_info);
624 pthread_mutex_unlock(&ust_fork_mutex);
625 }
626 }
627
628 static
629 int handle_message(struct sock_info *sock_info,
630 int sock, struct ustcomm_ust_msg *lum)
631 {
632 int ret = 0;
633 const struct lttng_ust_objd_ops *ops;
634 struct ustcomm_ust_reply lur;
635 union ust_args args;
636 char ctxstr[LTTNG_UST_SYM_NAME_LEN]; /* App context string. */
637 ssize_t len;
638
639 memset(&lur, 0, sizeof(lur));
640
641 if (ust_lock()) {
642 ret = -LTTNG_UST_ERR_EXITING;
643 goto error;
644 }
645
646 ops = objd_ops(lum->handle);
647 if (!ops) {
648 ret = -ENOENT;
649 goto error;
650 }
651
652 switch (lum->cmd) {
653 case LTTNG_UST_REGISTER_DONE:
654 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
655 ret = handle_register_done(sock_info);
656 else
657 ret = -EINVAL;
658 break;
659 case LTTNG_UST_RELEASE:
660 if (lum->handle == LTTNG_UST_ROOT_HANDLE)
661 ret = -EPERM;
662 else
663 ret = lttng_ust_objd_unref(lum->handle, 1);
664 break;
665 case LTTNG_UST_FILTER:
666 {
667 /* Receive filter data */
668 struct lttng_ust_filter_bytecode_node *bytecode;
669
670 if (lum->u.filter.data_size > FILTER_BYTECODE_MAX_LEN) {
671 ERR("Filter data size is too large: %u bytes",
672 lum->u.filter.data_size);
673 ret = -EINVAL;
674 goto error;
675 }
676
677 if (lum->u.filter.reloc_offset > lum->u.filter.data_size) {
678 ERR("Filter reloc offset %u is not within data",
679 lum->u.filter.reloc_offset);
680 ret = -EINVAL;
681 goto error;
682 }
683
684 bytecode = zmalloc(sizeof(*bytecode) + lum->u.filter.data_size);
685 if (!bytecode) {
686 ret = -ENOMEM;
687 goto error;
688 }
689 len = ustcomm_recv_unix_sock(sock, bytecode->bc.data,
690 lum->u.filter.data_size);
691 switch (len) {
692 case 0: /* orderly shutdown */
693 ret = 0;
694 free(bytecode);
695 goto error;
696 default:
697 if (len == lum->u.filter.data_size) {
698 DBG("filter data received");
699 break;
700 } else if (len < 0) {
701 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
702 if (len == -ECONNRESET) {
703 ERR("%s remote end closed connection", sock_info->name);
704 ret = len;
705 free(bytecode);
706 goto error;
707 }
708 ret = len;
709 free(bytecode);
710 goto error;
711 } else {
712 DBG("incorrect filter data message size: %zd", len);
713 ret = -EINVAL;
714 free(bytecode);
715 goto error;
716 }
717 }
718 bytecode->bc.len = lum->u.filter.data_size;
719 bytecode->bc.reloc_offset = lum->u.filter.reloc_offset;
720 bytecode->bc.seqnum = lum->u.filter.seqnum;
721 if (ops->cmd) {
722 ret = ops->cmd(lum->handle, lum->cmd,
723 (unsigned long) bytecode,
724 &args, sock_info);
725 if (ret) {
726 free(bytecode);
727 }
728 /* don't free bytecode if everything went fine. */
729 } else {
730 ret = -ENOSYS;
731 free(bytecode);
732 }
733 break;
734 }
735 case LTTNG_UST_EXCLUSION:
736 {
737 /* Receive exclusion names */
738 struct lttng_ust_excluder_node *node;
739 unsigned int count;
740
741 count = lum->u.exclusion.count;
742 if (count == 0) {
743 /* There are no names to read */
744 ret = 0;
745 goto error;
746 }
747 node = zmalloc(sizeof(*node) +
748 count * LTTNG_UST_SYM_NAME_LEN);
749 if (!node) {
750 ret = -ENOMEM;
751 goto error;
752 }
753 node->excluder.count = count;
754 len = ustcomm_recv_unix_sock(sock, node->excluder.names,
755 count * LTTNG_UST_SYM_NAME_LEN);
756 switch (len) {
757 case 0: /* orderly shutdown */
758 ret = 0;
759 free(node);
760 goto error;
761 default:
762 if (len == count * LTTNG_UST_SYM_NAME_LEN) {
763 DBG("Exclusion data received");
764 break;
765 } else if (len < 0) {
766 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
767 if (len == -ECONNRESET) {
768 ERR("%s remote end closed connection", sock_info->name);
769 ret = len;
770 free(node);
771 goto error;
772 }
773 ret = len;
774 free(node);
775 goto error;
776 } else {
777 DBG("Incorrect exclusion data message size: %zd", len);
778 ret = -EINVAL;
779 free(node);
780 goto error;
781 }
782 }
783 if (ops->cmd) {
784 ret = ops->cmd(lum->handle, lum->cmd,
785 (unsigned long) node,
786 &args, sock_info);
787 if (ret) {
788 free(node);
789 }
790 /* Don't free exclusion data if everything went fine. */
791 } else {
792 ret = -ENOSYS;
793 free(node);
794 }
795 break;
796 }
797 case LTTNG_UST_CHANNEL:
798 {
799 void *chan_data;
800 int wakeup_fd;
801
802 len = ustcomm_recv_channel_from_sessiond(sock,
803 &chan_data, lum->u.channel.len,
804 &wakeup_fd);
805 switch (len) {
806 case 0: /* orderly shutdown */
807 ret = 0;
808 goto error;
809 default:
810 if (len == lum->u.channel.len) {
811 DBG("channel data received");
812 break;
813 } else if (len < 0) {
814 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
815 if (len == -ECONNRESET) {
816 ERR("%s remote end closed connection", sock_info->name);
817 ret = len;
818 goto error;
819 }
820 ret = len;
821 goto error;
822 } else {
823 DBG("incorrect channel data message size: %zd", len);
824 ret = -EINVAL;
825 goto error;
826 }
827 }
828 args.channel.chan_data = chan_data;
829 args.channel.wakeup_fd = wakeup_fd;
830 if (ops->cmd)
831 ret = ops->cmd(lum->handle, lum->cmd,
832 (unsigned long) &lum->u,
833 &args, sock_info);
834 else
835 ret = -ENOSYS;
836 break;
837 }
838 case LTTNG_UST_STREAM:
839 {
840 /* Receive shm_fd, wakeup_fd */
841 ret = ustcomm_recv_stream_from_sessiond(sock,
842 &lum->u.stream.len,
843 &args.stream.shm_fd,
844 &args.stream.wakeup_fd);
845 if (ret) {
846 goto error;
847 }
848 if (ops->cmd)
849 ret = ops->cmd(lum->handle, lum->cmd,
850 (unsigned long) &lum->u,
851 &args, sock_info);
852 else
853 ret = -ENOSYS;
854 break;
855 }
856 case LTTNG_UST_CONTEXT:
857 switch (lum->u.context.ctx) {
858 case LTTNG_UST_CONTEXT_APP_CONTEXT:
859 {
860 char *p;
861 size_t ctxlen, recvlen;
862
863 ctxlen = strlen("$app.") + lum->u.context.u.app_ctx.provider_name_len - 1
864 + strlen(":") + lum->u.context.u.app_ctx.ctx_name_len;
865 if (ctxlen >= LTTNG_UST_SYM_NAME_LEN) {
866 ERR("Application context string length size is too large: %zu bytes",
867 ctxlen);
868 ret = -EINVAL;
869 goto error;
870 }
871 strcpy(ctxstr, "$app.");
872 p = &ctxstr[strlen("$app.")];
873 recvlen = ctxlen - strlen("$app.");
874 len = ustcomm_recv_unix_sock(sock, p, recvlen);
875 switch (len) {
876 case 0: /* orderly shutdown */
877 ret = 0;
878 goto error;
879 default:
880 if (len == recvlen) {
881 DBG("app context data received");
882 break;
883 } else if (len < 0) {
884 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
885 if (len == -ECONNRESET) {
886 ERR("%s remote end closed connection", sock_info->name);
887 ret = len;
888 goto error;
889 }
890 ret = len;
891 goto error;
892 } else {
893 DBG("incorrect app context data message size: %zd", len);
894 ret = -EINVAL;
895 goto error;
896 }
897 }
898 /* Put : between provider and ctxname. */
899 p[lum->u.context.u.app_ctx.provider_name_len - 1] = ':';
900 args.app_context.ctxname = ctxstr;
901 break;
902 }
903 default:
904 break;
905 }
906 if (ops->cmd) {
907 ret = ops->cmd(lum->handle, lum->cmd,
908 (unsigned long) &lum->u,
909 &args, sock_info);
910 } else {
911 ret = -ENOSYS;
912 }
913 break;
914 default:
915 if (ops->cmd)
916 ret = ops->cmd(lum->handle, lum->cmd,
917 (unsigned long) &lum->u,
918 &args, sock_info);
919 else
920 ret = -ENOSYS;
921 break;
922 }
923
924 lur.handle = lum->handle;
925 lur.cmd = lum->cmd;
926 lur.ret_val = ret;
927 if (ret >= 0) {
928 lur.ret_code = LTTNG_UST_OK;
929 } else {
930 /*
931 * Use -LTTNG_UST_ERR as wildcard for UST internal
932 * error that are not caused by the transport, except if
933 * we already have a more precise error message to
934 * report.
935 */
936 if (ret > -LTTNG_UST_ERR) {
937 /* Translate code to UST error. */
938 switch (ret) {
939 case -EEXIST:
940 lur.ret_code = -LTTNG_UST_ERR_EXIST;
941 break;
942 case -EINVAL:
943 lur.ret_code = -LTTNG_UST_ERR_INVAL;
944 break;
945 case -ENOENT:
946 lur.ret_code = -LTTNG_UST_ERR_NOENT;
947 break;
948 case -EPERM:
949 lur.ret_code = -LTTNG_UST_ERR_PERM;
950 break;
951 case -ENOSYS:
952 lur.ret_code = -LTTNG_UST_ERR_NOSYS;
953 break;
954 default:
955 lur.ret_code = -LTTNG_UST_ERR;
956 break;
957 }
958 } else {
959 lur.ret_code = ret;
960 }
961 }
962 if (ret >= 0) {
963 switch (lum->cmd) {
964 case LTTNG_UST_TRACER_VERSION:
965 lur.u.version = lum->u.version;
966 break;
967 case LTTNG_UST_TRACEPOINT_LIST_GET:
968 memcpy(&lur.u.tracepoint, &lum->u.tracepoint, sizeof(lur.u.tracepoint));
969 break;
970 }
971 }
972 DBG("Return value: %d", lur.ret_val);
973
974 ust_unlock();
975
976 /*
977 * Performed delayed statedump operations outside of the UST
978 * lock. We need to take the dynamic loader lock before we take
979 * the UST lock internally within handle_pending_statedump().
980 */
981 handle_pending_statedump(sock_info);
982
983 if (ust_lock()) {
984 ret = -LTTNG_UST_ERR_EXITING;
985 goto error;
986 }
987
988 ret = send_reply(sock, &lur);
989 if (ret < 0) {
990 DBG("error sending reply");
991 goto error;
992 }
993
994 /*
995 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
996 * after the reply.
997 */
998 if (lur.ret_code == LTTNG_UST_OK) {
999 switch (lum->cmd) {
1000 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET:
1001 len = ustcomm_send_unix_sock(sock,
1002 &args.field_list.entry,
1003 sizeof(args.field_list.entry));
1004 if (len < 0) {
1005 ret = len;
1006 goto error;
1007 }
1008 if (len != sizeof(args.field_list.entry)) {
1009 ret = -EINVAL;
1010 goto error;
1011 }
1012 }
1013 }
1014
1015 error:
1016 ust_unlock();
1017
1018 return ret;
1019 }
1020
1021 static
1022 void cleanup_sock_info(struct sock_info *sock_info, int exiting)
1023 {
1024 int ret;
1025
1026 if (sock_info->root_handle != -1) {
1027 ret = lttng_ust_objd_unref(sock_info->root_handle, 1);
1028 if (ret) {
1029 ERR("Error unref root handle");
1030 }
1031 sock_info->root_handle = -1;
1032 }
1033 sock_info->constructor_sem_posted = 0;
1034
1035 /*
1036 * wait_shm_mmap, socket and notify socket are used by listener
1037 * threads outside of the ust lock, so we cannot tear them down
1038 * ourselves, because we cannot join on these threads. Leave
1039 * responsibility of cleaning up these resources to the OS
1040 * process exit.
1041 */
1042 if (exiting)
1043 return;
1044
1045 if (sock_info->socket != -1) {
1046 ret = ustcomm_close_unix_sock(sock_info->socket);
1047 if (ret) {
1048 ERR("Error closing ust cmd socket");
1049 }
1050 sock_info->socket = -1;
1051 }
1052 if (sock_info->notify_socket != -1) {
1053 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1054 if (ret) {
1055 ERR("Error closing ust notify socket");
1056 }
1057 sock_info->notify_socket = -1;
1058 }
1059 if (sock_info->wait_shm_mmap) {
1060 long page_size;
1061
1062 page_size = sysconf(_SC_PAGE_SIZE);
1063 if (page_size <= 0) {
1064 if (!page_size) {
1065 errno = EINVAL;
1066 }
1067 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1068 } else {
1069 ret = munmap(sock_info->wait_shm_mmap, page_size);
1070 if (ret) {
1071 ERR("Error unmapping wait shm");
1072 }
1073 }
1074 sock_info->wait_shm_mmap = NULL;
1075 }
1076 }
1077
1078 /*
1079 * Using fork to set umask in the child process (not multi-thread safe).
1080 * We deal with the shm_open vs ftruncate race (happening when the
1081 * sessiond owns the shm and does not let everybody modify it, to ensure
1082 * safety against shm_unlink) by simply letting the mmap fail and
1083 * retrying after a few seconds.
1084 * For global shm, everybody has rw access to it until the sessiond
1085 * starts.
1086 */
1087 static
1088 int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
1089 {
1090 int wait_shm_fd, ret;
1091 pid_t pid;
1092
1093 /*
1094 * Try to open read-only.
1095 */
1096 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1097 if (wait_shm_fd >= 0) {
1098 int32_t tmp_read;
1099 ssize_t len;
1100 size_t bytes_read = 0;
1101
1102 /*
1103 * Try to read the fd. If unable to do so, try opening
1104 * it in write mode.
1105 */
1106 do {
1107 len = read(wait_shm_fd,
1108 &((char *) &tmp_read)[bytes_read],
1109 sizeof(tmp_read) - bytes_read);
1110 if (len > 0) {
1111 bytes_read += len;
1112 }
1113 } while ((len < 0 && errno == EINTR)
1114 || (len > 0 && bytes_read < sizeof(tmp_read)));
1115 if (bytes_read != sizeof(tmp_read)) {
1116 ret = close(wait_shm_fd);
1117 if (ret) {
1118 ERR("close wait_shm_fd");
1119 }
1120 goto open_write;
1121 }
1122 goto end;
1123 } else if (wait_shm_fd < 0 && errno != ENOENT) {
1124 /*
1125 * Real-only open did not work, and it's not because the
1126 * entry was not present. It's a failure that prohibits
1127 * using shm.
1128 */
1129 ERR("Error opening shm %s", sock_info->wait_shm_path);
1130 goto end;
1131 }
1132
1133 open_write:
1134 /*
1135 * If the open failed because the file did not exist, or because
1136 * the file was not truncated yet, try creating it ourself.
1137 */
1138 URCU_TLS(lttng_ust_nest_count)++;
1139 pid = fork();
1140 URCU_TLS(lttng_ust_nest_count)--;
1141 if (pid > 0) {
1142 int status;
1143
1144 /*
1145 * Parent: wait for child to return, in which case the
1146 * shared memory map will have been created.
1147 */
1148 pid = wait(&status);
1149 if (pid < 0 || !WIFEXITED(status) || WEXITSTATUS(status) != 0) {
1150 wait_shm_fd = -1;
1151 goto end;
1152 }
1153 /*
1154 * Try to open read-only again after creation.
1155 */
1156 wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
1157 if (wait_shm_fd < 0) {
1158 /*
1159 * Real-only open did not work. It's a failure
1160 * that prohibits using shm.
1161 */
1162 ERR("Error opening shm %s", sock_info->wait_shm_path);
1163 goto end;
1164 }
1165 goto end;
1166 } else if (pid == 0) {
1167 int create_mode;
1168
1169 /* Child */
1170 create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
1171 if (sock_info->global)
1172 create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
1173 /*
1174 * We're alone in a child process, so we can modify the
1175 * process-wide umask.
1176 */
1177 umask(~create_mode);
1178 /*
1179 * Try creating shm (or get rw access).
1180 * We don't do an exclusive open, because we allow other
1181 * processes to create+ftruncate it concurrently.
1182 */
1183 wait_shm_fd = shm_open(sock_info->wait_shm_path,
1184 O_RDWR | O_CREAT, create_mode);
1185 if (wait_shm_fd >= 0) {
1186 ret = ftruncate(wait_shm_fd, mmap_size);
1187 if (ret) {
1188 PERROR("ftruncate");
1189 _exit(EXIT_FAILURE);
1190 }
1191 _exit(EXIT_SUCCESS);
1192 }
1193 /*
1194 * For local shm, we need to have rw access to accept
1195 * opening it: this means the local sessiond will be
1196 * able to wake us up. For global shm, we open it even
1197 * if rw access is not granted, because the root.root
1198 * sessiond will be able to override all rights and wake
1199 * us up.
1200 */
1201 if (!sock_info->global && errno != EACCES) {
1202 ERR("Error opening shm %s", sock_info->wait_shm_path);
1203 _exit(EXIT_FAILURE);
1204 }
1205 /*
1206 * The shm exists, but we cannot open it RW. Report
1207 * success.
1208 */
1209 _exit(EXIT_SUCCESS);
1210 } else {
1211 return -1;
1212 }
1213 end:
1214 if (wait_shm_fd >= 0 && !sock_info->global) {
1215 struct stat statbuf;
1216
1217 /*
1218 * Ensure that our user is the owner of the shm file for
1219 * local shm. If we do not own the file, it means our
1220 * sessiond will not have access to wake us up (there is
1221 * probably a rogue process trying to fake our
1222 * sessiond). Fallback to polling method in this case.
1223 */
1224 ret = fstat(wait_shm_fd, &statbuf);
1225 if (ret) {
1226 PERROR("fstat");
1227 goto error_close;
1228 }
1229 if (statbuf.st_uid != getuid())
1230 goto error_close;
1231 }
1232 return wait_shm_fd;
1233
1234 error_close:
1235 ret = close(wait_shm_fd);
1236 if (ret) {
1237 PERROR("Error closing fd");
1238 }
1239 return -1;
1240 }
1241
1242 static
1243 char *get_map_shm(struct sock_info *sock_info)
1244 {
1245 long page_size;
1246 int wait_shm_fd, ret;
1247 char *wait_shm_mmap;
1248
1249 page_size = sysconf(_SC_PAGE_SIZE);
1250 if (page_size <= 0) {
1251 if (!page_size) {
1252 errno = EINVAL;
1253 }
1254 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1255 goto error;
1256 }
1257
1258 lttng_ust_lock_fd_tracker();
1259 wait_shm_fd = get_wait_shm(sock_info, page_size);
1260 if (wait_shm_fd < 0) {
1261 lttng_ust_unlock_fd_tracker();
1262 goto error;
1263 }
1264 lttng_ust_add_fd_to_tracker(wait_shm_fd);
1265 lttng_ust_unlock_fd_tracker();
1266
1267 wait_shm_mmap = mmap(NULL, page_size, PROT_READ,
1268 MAP_SHARED, wait_shm_fd, 0);
1269
1270 /* close shm fd immediately after taking the mmap reference */
1271 lttng_ust_lock_fd_tracker();
1272 ret = close(wait_shm_fd);
1273 if (!ret) {
1274 lttng_ust_delete_fd_from_tracker(wait_shm_fd);
1275 } else {
1276 PERROR("Error closing fd");
1277 }
1278 lttng_ust_unlock_fd_tracker();
1279
1280 if (wait_shm_mmap == MAP_FAILED) {
1281 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1282 goto error;
1283 }
1284 return wait_shm_mmap;
1285
1286 error:
1287 return NULL;
1288 }
1289
1290 static
1291 void wait_for_sessiond(struct sock_info *sock_info)
1292 {
1293 if (ust_lock()) {
1294 goto quit;
1295 }
1296 if (wait_poll_fallback) {
1297 goto error;
1298 }
1299 if (!sock_info->wait_shm_mmap) {
1300 sock_info->wait_shm_mmap = get_map_shm(sock_info);
1301 if (!sock_info->wait_shm_mmap)
1302 goto error;
1303 }
1304 ust_unlock();
1305
1306 DBG("Waiting for %s apps sessiond", sock_info->name);
1307 /* Wait for futex wakeup */
1308 if (uatomic_read((int32_t *) sock_info->wait_shm_mmap))
1309 goto end_wait;
1310
1311 while (futex_async((int32_t *) sock_info->wait_shm_mmap,
1312 FUTEX_WAIT, 0, NULL, NULL, 0)) {
1313 switch (errno) {
1314 case EWOULDBLOCK:
1315 /* Value already changed. */
1316 goto end_wait;
1317 case EINTR:
1318 /* Retry if interrupted by signal. */
1319 break; /* Get out of switch. */
1320 case EFAULT:
1321 wait_poll_fallback = 1;
1322 DBG(
1323 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1324 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1325 "Please upgrade your kernel "
1326 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1327 "mainline). LTTng-UST will use polling mode fallback.");
1328 if (ust_debug())
1329 PERROR("futex");
1330 goto end_wait;
1331 }
1332 }
1333 end_wait:
1334 return;
1335
1336 quit:
1337 ust_unlock();
1338 return;
1339
1340 error:
1341 ust_unlock();
1342 return;
1343 }
1344
1345 /*
1346 * This thread does not allocate any resource, except within
1347 * handle_message, within mutex protection. This mutex protects against
1348 * fork and exit.
1349 * The other moment it allocates resources is at socket connection, which
1350 * is also protected by the mutex.
1351 */
1352 static
1353 void *ust_listener_thread(void *arg)
1354 {
1355 struct sock_info *sock_info = arg;
1356 int sock, ret, prev_connect_failed = 0, has_waited = 0;
1357 long timeout;
1358
1359 lttng_ust_fixup_tls();
1360 /*
1361 * If available, add '-ust' to the end of this thread's
1362 * process name
1363 */
1364 ret = lttng_ust_setustprocname();
1365 if (ret) {
1366 ERR("Unable to set UST process name");
1367 }
1368
1369 /* Restart trying to connect to the session daemon */
1370 restart:
1371 if (prev_connect_failed) {
1372 /* Wait for sessiond availability with pipe */
1373 wait_for_sessiond(sock_info);
1374 if (has_waited) {
1375 has_waited = 0;
1376 /*
1377 * Sleep for 5 seconds before retrying after a
1378 * sequence of failure / wait / failure. This
1379 * deals with a killed or broken session daemon.
1380 */
1381 sleep(5);
1382 } else {
1383 has_waited = 1;
1384 }
1385 prev_connect_failed = 0;
1386 }
1387
1388 if (sock_info->socket != -1) {
1389 /* FD tracker is updated by ustcomm_close_unix_sock() */
1390 ret = ustcomm_close_unix_sock(sock_info->socket);
1391 if (ret) {
1392 ERR("Error closing %s ust cmd socket",
1393 sock_info->name);
1394 }
1395 sock_info->socket = -1;
1396 }
1397 if (sock_info->notify_socket != -1) {
1398 /* FD tracker is updated by ustcomm_close_unix_sock() */
1399 ret = ustcomm_close_unix_sock(sock_info->notify_socket);
1400 if (ret) {
1401 ERR("Error closing %s ust notify socket",
1402 sock_info->name);
1403 }
1404 sock_info->notify_socket = -1;
1405 }
1406
1407 if (ust_lock()) {
1408 goto quit;
1409 }
1410
1411 /*
1412 * Register. We need to perform both connect and sending
1413 * registration message before doing the next connect otherwise
1414 * we may reach unix socket connect queue max limits and block
1415 * on the 2nd connect while the session daemon is awaiting the
1416 * first connect registration message.
1417 */
1418 /* Connect cmd socket */
1419 lttng_ust_lock_fd_tracker();
1420 ret = ustcomm_connect_unix_sock(sock_info->sock_path,
1421 get_connect_sock_timeout());
1422 if (ret < 0) {
1423 lttng_ust_unlock_fd_tracker();
1424 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1425 prev_connect_failed = 1;
1426
1427 /*
1428 * If we cannot find the sessiond daemon, don't delay
1429 * constructor execution.
1430 */
1431 ret = handle_register_done(sock_info);
1432 assert(!ret);
1433 ust_unlock();
1434 goto restart;
1435 }
1436 lttng_ust_add_fd_to_tracker(ret);
1437 lttng_ust_unlock_fd_tracker();
1438 sock_info->socket = ret;
1439
1440 ust_unlock();
1441 /*
1442 * Unlock/relock ust lock because connect is blocking (with
1443 * timeout). Don't delay constructors on the ust lock for too
1444 * long.
1445 */
1446 if (ust_lock()) {
1447 goto quit;
1448 }
1449
1450 /*
1451 * Create only one root handle per listener thread for the whole
1452 * process lifetime, so we ensure we get ID which is statically
1453 * assigned to the root handle.
1454 */
1455 if (sock_info->root_handle == -1) {
1456 ret = lttng_abi_create_root_handle();
1457 if (ret < 0) {
1458 ERR("Error creating root handle");
1459 goto quit;
1460 }
1461 sock_info->root_handle = ret;
1462 }
1463
1464 ret = register_to_sessiond(sock_info->socket, USTCTL_SOCKET_CMD);
1465 if (ret < 0) {
1466 ERR("Error registering to %s ust cmd socket",
1467 sock_info->name);
1468 prev_connect_failed = 1;
1469 /*
1470 * If we cannot register to the sessiond daemon, don't
1471 * delay constructor execution.
1472 */
1473 ret = handle_register_done(sock_info);
1474 assert(!ret);
1475 ust_unlock();
1476 goto restart;
1477 }
1478
1479 ust_unlock();
1480 /*
1481 * Unlock/relock ust lock because connect is blocking (with
1482 * timeout). Don't delay constructors on the ust lock for too
1483 * long.
1484 */
1485 if (ust_lock()) {
1486 goto quit;
1487 }
1488
1489 /* Connect notify socket */
1490 lttng_ust_lock_fd_tracker();
1491 ret = ustcomm_connect_unix_sock(sock_info->sock_path,
1492 get_connect_sock_timeout());
1493 if (ret < 0) {
1494 lttng_ust_unlock_fd_tracker();
1495 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info->name);
1496 prev_connect_failed = 1;
1497
1498 /*
1499 * If we cannot find the sessiond daemon, don't delay
1500 * constructor execution.
1501 */
1502 ret = handle_register_done(sock_info);
1503 assert(!ret);
1504 ust_unlock();
1505 goto restart;
1506 }
1507 lttng_ust_add_fd_to_tracker(ret);
1508 lttng_ust_unlock_fd_tracker();
1509 sock_info->notify_socket = ret;
1510
1511 ust_unlock();
1512 /*
1513 * Unlock/relock ust lock because connect is blocking (with
1514 * timeout). Don't delay constructors on the ust lock for too
1515 * long.
1516 */
1517 if (ust_lock()) {
1518 goto quit;
1519 }
1520
1521 timeout = get_notify_sock_timeout();
1522 if (timeout >= 0) {
1523 /*
1524 * Give at least 10ms to sessiond to reply to
1525 * notifications.
1526 */
1527 if (timeout < 10)
1528 timeout = 10;
1529 ret = ustcomm_setsockopt_rcv_timeout(sock_info->notify_socket,
1530 timeout);
1531 if (ret < 0) {
1532 WARN("Error setting socket receive timeout");
1533 }
1534 ret = ustcomm_setsockopt_snd_timeout(sock_info->notify_socket,
1535 timeout);
1536 if (ret < 0) {
1537 WARN("Error setting socket send timeout");
1538 }
1539 } else if (timeout < -1) {
1540 WARN("Unsupported timeout value %ld", timeout);
1541 }
1542
1543 ret = register_to_sessiond(sock_info->notify_socket,
1544 USTCTL_SOCKET_NOTIFY);
1545 if (ret < 0) {
1546 ERR("Error registering to %s ust notify socket",
1547 sock_info->name);
1548 prev_connect_failed = 1;
1549 /*
1550 * If we cannot register to the sessiond daemon, don't
1551 * delay constructor execution.
1552 */
1553 ret = handle_register_done(sock_info);
1554 assert(!ret);
1555 ust_unlock();
1556 goto restart;
1557 }
1558 sock = sock_info->socket;
1559
1560 ust_unlock();
1561
1562 for (;;) {
1563 ssize_t len;
1564 struct ustcomm_ust_msg lum;
1565
1566 len = ustcomm_recv_unix_sock(sock, &lum, sizeof(lum));
1567 switch (len) {
1568 case 0: /* orderly shutdown */
1569 DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info->name);
1570 if (ust_lock()) {
1571 goto quit;
1572 }
1573 /*
1574 * Either sessiond has shutdown or refused us by closing the socket.
1575 * In either case, we don't want to delay construction execution,
1576 * and we need to wait before retry.
1577 */
1578 prev_connect_failed = 1;
1579 /*
1580 * If we cannot register to the sessiond daemon, don't
1581 * delay constructor execution.
1582 */
1583 ret = handle_register_done(sock_info);
1584 assert(!ret);
1585 ust_unlock();
1586 goto end;
1587 case sizeof(lum):
1588 print_cmd(lum.cmd, lum.handle);
1589 ret = handle_message(sock_info, sock, &lum);
1590 if (ret) {
1591 ERR("Error handling message for %s socket",
1592 sock_info->name);
1593 /*
1594 * Close socket if protocol error is
1595 * detected.
1596 */
1597 goto end;
1598 }
1599 continue;
1600 default:
1601 if (len < 0) {
1602 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len);
1603 } else {
1604 DBG("incorrect message size (%s socket): %zd", sock_info->name, len);
1605 }
1606 if (len == -ECONNRESET) {
1607 DBG("%s remote end closed connection", sock_info->name);
1608 goto end;
1609 }
1610 goto end;
1611 }
1612
1613 }
1614 end:
1615 if (ust_lock()) {
1616 goto quit;
1617 }
1618 /* Cleanup socket handles before trying to reconnect */
1619 lttng_ust_objd_table_owner_cleanup(sock_info);
1620 ust_unlock();
1621 goto restart; /* try to reconnect */
1622
1623 quit:
1624 ust_unlock();
1625
1626 pthread_mutex_lock(&ust_exit_mutex);
1627 sock_info->thread_active = 0;
1628 pthread_mutex_unlock(&ust_exit_mutex);
1629 return NULL;
1630 }
1631
1632 /*
1633 * Weak symbol to call when the ust malloc wrapper is not loaded.
1634 */
1635 __attribute__((weak))
1636 void lttng_ust_malloc_wrapper_init(void)
1637 {
1638 }
1639
1640 /*
1641 * sessiond monitoring thread: monitor presence of global and per-user
1642 * sessiond by polling the application common named pipe.
1643 */
1644 void __attribute__((constructor)) lttng_ust_init(void)
1645 {
1646 struct timespec constructor_timeout;
1647 sigset_t sig_all_blocked, orig_parent_mask;
1648 pthread_attr_t thread_attr;
1649 int timeout_mode;
1650 int ret;
1651
1652 if (uatomic_xchg(&initialized, 1) == 1)
1653 return;
1654
1655 /*
1656 * Fixup interdependency between TLS fixup mutex (which happens
1657 * to be the dynamic linker mutex) and ust_lock, taken within
1658 * the ust lock.
1659 */
1660 lttng_ust_fixup_tls();
1661
1662 lttng_ust_loaded = 1;
1663
1664 /*
1665 * We want precise control over the order in which we construct
1666 * our sub-libraries vs starting to receive commands from
1667 * sessiond (otherwise leading to errors when trying to create
1668 * sessiond before the init functions are completed).
1669 */
1670 init_usterr();
1671 lttng_ust_getenv_init(); /* Needs init_usterr() to be completed. */
1672 init_tracepoint();
1673 lttng_ust_init_fd_tracker();
1674 lttng_ust_clock_init();
1675 lttng_ust_getcpu_init();
1676 lttng_ust_statedump_init();
1677 lttng_ring_buffer_metadata_client_init();
1678 lttng_ring_buffer_client_overwrite_init();
1679 lttng_ring_buffer_client_overwrite_rt_init();
1680 lttng_ring_buffer_client_discard_init();
1681 lttng_ring_buffer_client_discard_rt_init();
1682 lttng_perf_counter_init();
1683 /*
1684 * Invoke ust malloc wrapper init before starting other threads.
1685 */
1686 lttng_ust_malloc_wrapper_init();
1687
1688 timeout_mode = get_constructor_timeout(&constructor_timeout);
1689
1690 get_allow_blocking();
1691
1692 ret = sem_init(&constructor_wait, 0, 0);
1693 if (ret) {
1694 PERROR("sem_init");
1695 }
1696
1697 ret = setup_local_apps();
1698 if (ret) {
1699 DBG("local apps setup returned %d", ret);
1700 }
1701
1702 /* A new thread created by pthread_create inherits the signal mask
1703 * from the parent. To avoid any signal being received by the
1704 * listener thread, we block all signals temporarily in the parent,
1705 * while we create the listener thread.
1706 */
1707 sigfillset(&sig_all_blocked);
1708 ret = pthread_sigmask(SIG_SETMASK, &sig_all_blocked, &orig_parent_mask);
1709 if (ret) {
1710 ERR("pthread_sigmask: %s", strerror(ret));
1711 }
1712
1713 ret = pthread_attr_init(&thread_attr);
1714 if (ret) {
1715 ERR("pthread_attr_init: %s", strerror(ret));
1716 }
1717 ret = pthread_attr_setdetachstate(&thread_attr, PTHREAD_CREATE_DETACHED);
1718 if (ret) {
1719 ERR("pthread_attr_setdetachstate: %s", strerror(ret));
1720 }
1721
1722 pthread_mutex_lock(&ust_exit_mutex);
1723 ret = pthread_create(&global_apps.ust_listener, &thread_attr,
1724 ust_listener_thread, &global_apps);
1725 if (ret) {
1726 ERR("pthread_create global: %s", strerror(ret));
1727 }
1728 global_apps.thread_active = 1;
1729 pthread_mutex_unlock(&ust_exit_mutex);
1730
1731 if (local_apps.allowed) {
1732 pthread_mutex_lock(&ust_exit_mutex);
1733 ret = pthread_create(&local_apps.ust_listener, &thread_attr,
1734 ust_listener_thread, &local_apps);
1735 if (ret) {
1736 ERR("pthread_create local: %s", strerror(ret));
1737 }
1738 local_apps.thread_active = 1;
1739 pthread_mutex_unlock(&ust_exit_mutex);
1740 } else {
1741 handle_register_done(&local_apps);
1742 }
1743 ret = pthread_attr_destroy(&thread_attr);
1744 if (ret) {
1745 ERR("pthread_attr_destroy: %s", strerror(ret));
1746 }
1747
1748 /* Restore original signal mask in parent */
1749 ret = pthread_sigmask(SIG_SETMASK, &orig_parent_mask, NULL);
1750 if (ret) {
1751 ERR("pthread_sigmask: %s", strerror(ret));
1752 }
1753
1754 switch (timeout_mode) {
1755 case 1: /* timeout wait */
1756 do {
1757 ret = sem_timedwait(&constructor_wait,
1758 &constructor_timeout);
1759 } while (ret < 0 && errno == EINTR);
1760 if (ret < 0) {
1761 switch (errno) {
1762 case ETIMEDOUT:
1763 ERR("Timed out waiting for lttng-sessiond");
1764 break;
1765 case EINVAL:
1766 PERROR("sem_timedwait");
1767 break;
1768 default:
1769 ERR("Unexpected error \"%s\" returned by sem_timedwait",
1770 strerror(errno));
1771 }
1772 }
1773 break;
1774 case -1:/* wait forever */
1775 do {
1776 ret = sem_wait(&constructor_wait);
1777 } while (ret < 0 && errno == EINTR);
1778 if (ret < 0) {
1779 switch (errno) {
1780 case EINVAL:
1781 PERROR("sem_wait");
1782 break;
1783 default:
1784 ERR("Unexpected error \"%s\" returned by sem_wait",
1785 strerror(errno));
1786 }
1787 }
1788 break;
1789 case 0: /* no timeout */
1790 break;
1791 }
1792 }
1793
1794 static
1795 void lttng_ust_cleanup(int exiting)
1796 {
1797 cleanup_sock_info(&global_apps, exiting);
1798 cleanup_sock_info(&local_apps, exiting);
1799 local_apps.allowed = 0;
1800 /*
1801 * The teardown in this function all affect data structures
1802 * accessed under the UST lock by the listener thread. This
1803 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1804 * that none of these threads are accessing this data at this
1805 * point.
1806 */
1807 lttng_ust_abi_exit();
1808 lttng_ust_events_exit();
1809 lttng_perf_counter_exit();
1810 lttng_ring_buffer_client_discard_rt_exit();
1811 lttng_ring_buffer_client_discard_exit();
1812 lttng_ring_buffer_client_overwrite_rt_exit();
1813 lttng_ring_buffer_client_overwrite_exit();
1814 lttng_ring_buffer_metadata_client_exit();
1815 lttng_ust_statedump_destroy();
1816 exit_tracepoint();
1817 if (!exiting) {
1818 /* Reinitialize values for fork */
1819 sem_count = 2;
1820 lttng_ust_comm_should_quit = 0;
1821 initialized = 0;
1822 }
1823 }
1824
1825 void __attribute__((destructor)) lttng_ust_exit(void)
1826 {
1827 int ret;
1828
1829 /*
1830 * Using pthread_cancel here because:
1831 * A) we don't want to hang application teardown.
1832 * B) the thread is not allocating any resource.
1833 */
1834
1835 /*
1836 * Require the communication thread to quit. Synchronize with
1837 * mutexes to ensure it is not in a mutex critical section when
1838 * pthread_cancel is later called.
1839 */
1840 ust_lock_nocheck();
1841 lttng_ust_comm_should_quit = 1;
1842 ust_unlock();
1843
1844 pthread_mutex_lock(&ust_exit_mutex);
1845 /* cancel threads */
1846 if (global_apps.thread_active) {
1847 ret = pthread_cancel(global_apps.ust_listener);
1848 if (ret) {
1849 ERR("Error cancelling global ust listener thread: %s",
1850 strerror(ret));
1851 } else {
1852 global_apps.thread_active = 0;
1853 }
1854 }
1855 if (local_apps.thread_active) {
1856 ret = pthread_cancel(local_apps.ust_listener);
1857 if (ret) {
1858 ERR("Error cancelling local ust listener thread: %s",
1859 strerror(ret));
1860 } else {
1861 local_apps.thread_active = 0;
1862 }
1863 }
1864 pthread_mutex_unlock(&ust_exit_mutex);
1865
1866 /*
1867 * Do NOT join threads: use of sys_futex makes it impossible to
1868 * join the threads without using async-cancel, but async-cancel
1869 * is delivered by a signal, which could hit the target thread
1870 * anywhere in its code path, including while the ust_lock() is
1871 * held, causing a deadlock for the other thread. Let the OS
1872 * cleanup the threads if there are stalled in a syscall.
1873 */
1874 lttng_ust_cleanup(1);
1875 }
1876
1877 /*
1878 * We exclude the worker threads across fork and clone (except
1879 * CLONE_VM), because these system calls only keep the forking thread
1880 * running in the child. Therefore, we don't want to call fork or clone
1881 * in the middle of an tracepoint or ust tracing state modification.
1882 * Holding this mutex protects these structures across fork and clone.
1883 */
1884 void ust_before_fork(sigset_t *save_sigset)
1885 {
1886 /*
1887 * Disable signals. This is to avoid that the child intervenes
1888 * before it is properly setup for tracing. It is safer to
1889 * disable all signals, because then we know we are not breaking
1890 * anything by restoring the original mask.
1891 */
1892 sigset_t all_sigs;
1893 int ret;
1894
1895 /* Fixup lttng-ust TLS. */
1896 lttng_ust_fixup_tls();
1897
1898 if (URCU_TLS(lttng_ust_nest_count))
1899 return;
1900 /* Disable signals */
1901 sigfillset(&all_sigs);
1902 ret = sigprocmask(SIG_BLOCK, &all_sigs, save_sigset);
1903 if (ret == -1) {
1904 PERROR("sigprocmask");
1905 }
1906
1907 pthread_mutex_lock(&ust_fork_mutex);
1908
1909 ust_lock_nocheck();
1910 rcu_bp_before_fork();
1911 }
1912
1913 static void ust_after_fork_common(sigset_t *restore_sigset)
1914 {
1915 int ret;
1916
1917 DBG("process %d", getpid());
1918 ust_unlock();
1919
1920 pthread_mutex_unlock(&ust_fork_mutex);
1921
1922 /* Restore signals */
1923 ret = sigprocmask(SIG_SETMASK, restore_sigset, NULL);
1924 if (ret == -1) {
1925 PERROR("sigprocmask");
1926 }
1927 }
1928
1929 void ust_after_fork_parent(sigset_t *restore_sigset)
1930 {
1931 if (URCU_TLS(lttng_ust_nest_count))
1932 return;
1933 DBG("process %d", getpid());
1934 rcu_bp_after_fork_parent();
1935 /* Release mutexes and reenable signals */
1936 ust_after_fork_common(restore_sigset);
1937 }
1938
1939 /*
1940 * After fork, in the child, we need to cleanup all the leftover state,
1941 * except the worker thread which already magically disappeared thanks
1942 * to the weird Linux fork semantics. After tyding up, we call
1943 * lttng_ust_init() again to start over as a new PID.
1944 *
1945 * This is meant for forks() that have tracing in the child between the
1946 * fork and following exec call (if there is any).
1947 */
1948 void ust_after_fork_child(sigset_t *restore_sigset)
1949 {
1950 if (URCU_TLS(lttng_ust_nest_count))
1951 return;
1952 lttng_context_vtid_reset();
1953 DBG("process %d", getpid());
1954 /* Release urcu mutexes */
1955 rcu_bp_after_fork_child();
1956 lttng_ust_cleanup(0);
1957 /* Release mutexes and reenable signals */
1958 ust_after_fork_common(restore_sigset);
1959 lttng_ust_init();
1960 }
1961
1962 void lttng_ust_sockinfo_session_enabled(void *owner)
1963 {
1964 struct sock_info *sock_info = owner;
1965 sock_info->statedump_pending = 1;
1966 }
LTTng 2.0 Userspace Tracer
This page took 0.122052 seconds and 4 git commands to generate.