4 * Copyright (C) 2011 David Goulet <david.goulet@polymtl.ca>
5 * Copyright (C) 2011 Mathieu Desnoyers <mathieu.desnoyers@efficios.com>
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.
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.
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
23 #include <sys/types.h>
24 #include <sys/socket.h>
27 #include <sys/types.h>
33 #include <semaphore.h>
37 #include <urcu/uatomic.h>
38 #include <urcu/futex.h>
39 #include <urcu/compiler.h>
41 #include <lttng/ust-events.h>
42 #include <lttng/ust-abi.h>
43 #include <lttng/ust.h>
44 #include <lttng/ust-error.h>
46 #include <usterr-signal-safe.h>
48 #include "tracepoint-internal.h"
49 #include "lttng-tracer-core.h"
51 #include "../libringbuffer/tlsfixup.h"
54 * Has lttng ust comm constructor been called ?
56 static int initialized
;
59 * The ust_lock/ust_unlock lock is used as a communication thread mutex.
60 * Held when handling a command, also held by fork() to deal with
61 * removal of threads, and by exit path.
64 /* Should the ust comm thread quit ? */
65 static int lttng_ust_comm_should_quit
;
68 * Wait for either of these before continuing to the main
70 * - the register_done message from sessiond daemon
71 * (will let the sessiond daemon enable sessions before main
73 * - sessiond daemon is not reachable.
74 * - timeout (ensuring applications are resilient to session
77 static sem_t constructor_wait
;
79 * Doing this for both the global and local sessiond.
81 static int sem_count
= { 2 };
84 * Counting nesting within lttng-ust. Used to ensure that calling fork()
85 * from liblttng-ust does not execute the pre/post fork handlers.
87 static int __thread lttng_ust_nest_count
;
90 * Info about socket and associated listener thread.
94 pthread_t ust_listener
; /* listener thread */
96 int constructor_sem_posted
;
101 char sock_path
[PATH_MAX
];
104 char wait_shm_path
[PATH_MAX
];
108 /* Socket from app (connect) to session daemon (listen) for communication */
109 struct sock_info global_apps
= {
117 .sock_path
= DEFAULT_GLOBAL_APPS_UNIX_SOCK
,
120 .wait_shm_path
= DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH
,
123 /* TODO: allow global_apps_sock_path override */
125 struct sock_info local_apps
= {
129 .allowed
= 0, /* Check setuid bit first */
135 static int wait_poll_fallback
;
137 extern void lttng_ring_buffer_client_overwrite_init(void);
138 extern void lttng_ring_buffer_client_discard_init(void);
139 extern void lttng_ring_buffer_metadata_client_init(void);
140 extern void lttng_ring_buffer_client_overwrite_exit(void);
141 extern void lttng_ring_buffer_client_discard_exit(void);
142 extern void lttng_ring_buffer_metadata_client_exit(void);
145 * Force a read (imply TLS fixup for dlopen) of TLS variables.
148 void lttng_fixup_nest_count_tls(void)
150 asm volatile ("" : : "m" (lttng_ust_nest_count
));
154 int setup_local_apps(void)
156 const char *home_dir
;
161 * Disallow per-user tracing for setuid binaries.
163 if (uid
!= geteuid()) {
164 assert(local_apps
.allowed
== 0);
167 home_dir
= (const char *) getenv("HOME");
169 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
170 assert(local_apps
.allowed
== 0);
173 local_apps
.allowed
= 1;
174 snprintf(local_apps
.sock_path
, PATH_MAX
,
175 DEFAULT_HOME_APPS_UNIX_SOCK
, home_dir
);
176 snprintf(local_apps
.wait_shm_path
, PATH_MAX
,
177 DEFAULT_HOME_APPS_WAIT_SHM_PATH
, uid
);
182 int register_app_to_sessiond(int socket
)
192 uint32_t bits_per_long
;
193 char name
[16]; /* process name */
196 reg_msg
.major
= LTTNG_UST_COMM_VERSION_MAJOR
;
197 reg_msg
.minor
= LTTNG_UST_COMM_VERSION_MINOR
;
198 reg_msg
.pid
= getpid();
199 reg_msg
.ppid
= getppid();
200 reg_msg
.uid
= getuid();
201 reg_msg
.gid
= getgid();
202 reg_msg
.bits_per_long
= CAA_BITS_PER_LONG
;
203 lttng_ust_getprocname(reg_msg
.name
);
205 ret
= ustcomm_send_unix_sock(socket
, ®_msg
, sizeof(reg_msg
));
206 if (ret
>= 0 && ret
!= sizeof(reg_msg
))
212 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
216 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
219 DBG("message successfully sent");
222 if (len
== -ECONNRESET
) {
223 DBG("remote end closed connection");
228 DBG("incorrect message size: %zd", len
);
234 int handle_register_done(struct sock_info
*sock_info
)
238 if (sock_info
->constructor_sem_posted
)
240 sock_info
->constructor_sem_posted
= 1;
241 if (uatomic_read(&sem_count
) <= 0) {
244 ret
= uatomic_add_return(&sem_count
, -1);
246 ret
= sem_post(&constructor_wait
);
253 int handle_message(struct sock_info
*sock_info
,
254 int sock
, struct ustcomm_ust_msg
*lum
)
257 const struct lttng_ust_objd_ops
*ops
;
258 struct ustcomm_ust_reply lur
;
265 memset(&lur
, 0, sizeof(lur
));
267 if (lttng_ust_comm_should_quit
) {
272 ops
= objd_ops(lum
->handle
);
279 case LTTNG_UST_REGISTER_DONE
:
280 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
281 ret
= handle_register_done(sock_info
);
285 case LTTNG_UST_RELEASE
:
286 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
289 ret
= lttng_ust_objd_unref(lum
->handle
);
291 case LTTNG_UST_FILTER
:
293 /* Receive filter data */
294 struct lttng_ust_filter_bytecode_node
*bytecode
;
296 if (lum
->u
.filter
.data_size
> FILTER_BYTECODE_MAX_LEN
) {
297 ERR("Filter data size is too large: %u bytes",
298 lum
->u
.filter
.data_size
);
303 if (lum
->u
.filter
.reloc_offset
> lum
->u
.filter
.data_size
) {
304 ERR("Filter reloc offset %u is not within data",
305 lum
->u
.filter
.reloc_offset
);
310 bytecode
= zmalloc(sizeof(*bytecode
) + lum
->u
.filter
.data_size
);
315 len
= ustcomm_recv_unix_sock(sock
, bytecode
->bc
.data
,
316 lum
->u
.filter
.data_size
);
318 case 0: /* orderly shutdown */
323 if (len
== lum
->u
.filter
.data_size
) {
324 DBG("filter data received");
326 } else if (len
< 0) {
327 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
328 if (len
== -ECONNRESET
) {
329 ERR("%s remote end closed connection", sock_info
->name
);
337 DBG("incorrect filter data message size: %zd", len
);
343 bytecode
->bc
.len
= lum
->u
.filter
.data_size
;
344 bytecode
->bc
.reloc_offset
= lum
->u
.filter
.reloc_offset
;
345 bytecode
->bc
.seqnum
= lum
->u
.filter
.seqnum
;
347 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
348 (unsigned long) bytecode
,
353 /* don't free bytecode if everything went fine. */
362 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
363 (unsigned long) &lum
->u
,
371 lur
.handle
= lum
->handle
;
375 lur
.ret_code
= LTTNG_UST_OK
;
378 * Use -LTTNG_UST_ERR as wildcard for UST internal
379 * error that are not caused by the transport, except if
380 * we already have a more precise error message to
383 if (ret
> -LTTNG_UST_ERR
) {
384 /* Translate code to UST error. */
387 lur
.ret_code
= -LTTNG_UST_ERR_EXIST
;
390 lur
.ret_code
= -LTTNG_UST_ERR_INVAL
;
393 lur
.ret_code
= -LTTNG_UST_ERR_NOENT
;
396 lur
.ret_code
= -LTTNG_UST_ERR_PERM
;
399 lur
.ret_code
= -LTTNG_UST_ERR_NOSYS
;
402 lur
.ret_code
= -LTTNG_UST_ERR
;
411 case LTTNG_UST_STREAM
:
413 * Special-case reply to send stream info.
416 lur
.u
.stream
.memory_map_size
= *args
.stream
.memory_map_size
;
417 shm_fd
= *args
.stream
.shm_fd
;
418 wait_fd
= *args
.stream
.wait_fd
;
420 case LTTNG_UST_METADATA
:
421 case LTTNG_UST_CHANNEL
:
422 lur
.u
.channel
.memory_map_size
= *args
.channel
.memory_map_size
;
423 shm_fd
= *args
.channel
.shm_fd
;
424 wait_fd
= *args
.channel
.wait_fd
;
426 case LTTNG_UST_TRACER_VERSION
:
427 lur
.u
.version
= lum
->u
.version
;
429 case LTTNG_UST_TRACEPOINT_LIST_GET
:
430 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
434 ret
= send_reply(sock
, &lur
);
436 DBG("error sending reply");
440 if ((lum
->cmd
== LTTNG_UST_STREAM
441 || lum
->cmd
== LTTNG_UST_CHANNEL
442 || lum
->cmd
== LTTNG_UST_METADATA
)
443 && lur
.ret_code
== LTTNG_UST_OK
) {
446 /* we also need to send the file descriptors. */
447 ret
= ustcomm_send_fds_unix_sock(sock
,
455 * The sessiond expects 2 file descriptors, even upon
458 ret
= ustcomm_send_fds_unix_sock(sock
,
462 perror("send wait_fd");
471 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
474 if (lur
.ret_code
== LTTNG_UST_OK
) {
476 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
477 len
= ustcomm_send_unix_sock(sock
,
478 &args
.field_list
.entry
,
479 sizeof(args
.field_list
.entry
));
484 if (len
!= sizeof(args
.field_list
.entry
)) {
491 * We still have the memory map reference, and the fds have been
492 * sent to the sessiond. We can therefore close those fds. Note
493 * that we keep the write side of the wait_fd open, but close
496 if (lur
.ret_code
== LTTNG_UST_OK
) {
498 case LTTNG_UST_STREAM
:
502 PERROR("Error closing stream shm_fd");
504 *args
.stream
.shm_fd
= -1;
507 ret
= close(wait_fd
);
509 PERROR("Error closing stream wait_fd");
511 *args
.stream
.wait_fd
= -1;
514 case LTTNG_UST_METADATA
:
515 case LTTNG_UST_CHANNEL
:
519 PERROR("Error closing channel shm_fd");
521 *args
.channel
.shm_fd
= -1;
524 ret
= close(wait_fd
);
526 PERROR("Error closing channel wait_fd");
528 *args
.channel
.wait_fd
= -1;
540 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
544 if (sock_info
->socket
!= -1) {
545 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
547 ERR("Error closing apps socket");
549 sock_info
->socket
= -1;
551 if (sock_info
->root_handle
!= -1) {
552 ret
= lttng_ust_objd_unref(sock_info
->root_handle
);
554 ERR("Error unref root handle");
556 sock_info
->root_handle
= -1;
558 sock_info
->constructor_sem_posted
= 0;
560 * wait_shm_mmap is used by listener threads outside of the
561 * ust lock, so we cannot tear it down ourselves, because we
562 * cannot join on these threads. Leave this task to the OS
565 if (!exiting
&& sock_info
->wait_shm_mmap
) {
566 ret
= munmap(sock_info
->wait_shm_mmap
, sysconf(_SC_PAGE_SIZE
));
568 ERR("Error unmapping wait shm");
570 sock_info
->wait_shm_mmap
= NULL
;
575 * Using fork to set umask in the child process (not multi-thread safe).
576 * We deal with the shm_open vs ftruncate race (happening when the
577 * sessiond owns the shm and does not let everybody modify it, to ensure
578 * safety against shm_unlink) by simply letting the mmap fail and
579 * retrying after a few seconds.
580 * For global shm, everybody has rw access to it until the sessiond
584 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
586 int wait_shm_fd
, ret
;
590 * Try to open read-only.
592 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
593 if (wait_shm_fd
>= 0) {
595 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
597 * Real-only open did not work, and it's not because the
598 * entry was not present. It's a failure that prohibits
601 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
605 * If the open failed because the file did not exist, try
606 * creating it ourself.
608 lttng_ust_nest_count
++;
610 lttng_ust_nest_count
--;
615 * Parent: wait for child to return, in which case the
616 * shared memory map will have been created.
619 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
624 * Try to open read-only again after creation.
626 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
627 if (wait_shm_fd
< 0) {
629 * Real-only open did not work. It's a failure
630 * that prohibits using shm.
632 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
636 } else if (pid
== 0) {
640 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
641 if (sock_info
->global
)
642 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
644 * We're alone in a child process, so we can modify the
645 * process-wide umask.
649 * Try creating shm (or get rw access).
650 * We don't do an exclusive open, because we allow other
651 * processes to create+ftruncate it concurrently.
653 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
654 O_RDWR
| O_CREAT
, create_mode
);
655 if (wait_shm_fd
>= 0) {
656 ret
= ftruncate(wait_shm_fd
, mmap_size
);
664 * For local shm, we need to have rw access to accept
665 * opening it: this means the local sessiond will be
666 * able to wake us up. For global shm, we open it even
667 * if rw access is not granted, because the root.root
668 * sessiond will be able to override all rights and wake
671 if (!sock_info
->global
&& errno
!= EACCES
) {
672 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
676 * The shm exists, but we cannot open it RW. Report
684 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
688 * Ensure that our user is the owner of the shm file for
689 * local shm. If we do not own the file, it means our
690 * sessiond will not have access to wake us up (there is
691 * probably a rogue process trying to fake our
692 * sessiond). Fallback to polling method in this case.
694 ret
= fstat(wait_shm_fd
, &statbuf
);
699 if (statbuf
.st_uid
!= getuid())
705 ret
= close(wait_shm_fd
);
707 PERROR("Error closing fd");
713 char *get_map_shm(struct sock_info
*sock_info
)
715 size_t mmap_size
= sysconf(_SC_PAGE_SIZE
);
716 int wait_shm_fd
, ret
;
719 wait_shm_fd
= get_wait_shm(sock_info
, mmap_size
);
720 if (wait_shm_fd
< 0) {
723 wait_shm_mmap
= mmap(NULL
, mmap_size
, PROT_READ
,
724 MAP_SHARED
, wait_shm_fd
, 0);
725 /* close shm fd immediately after taking the mmap reference */
726 ret
= close(wait_shm_fd
);
728 PERROR("Error closing fd");
730 if (wait_shm_mmap
== MAP_FAILED
) {
731 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
734 return wait_shm_mmap
;
741 void wait_for_sessiond(struct sock_info
*sock_info
)
746 if (lttng_ust_comm_should_quit
) {
749 if (wait_poll_fallback
) {
752 if (!sock_info
->wait_shm_mmap
) {
753 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
754 if (!sock_info
->wait_shm_mmap
)
759 DBG("Waiting for %s apps sessiond", sock_info
->name
);
760 /* Wait for futex wakeup */
761 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
762 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
763 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
765 if (errno
== EFAULT
) {
766 wait_poll_fallback
= 1;
768 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
769 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
770 "Please upgrade your kernel "
771 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
772 "mainline). LTTng-UST will use polling mode fallback.");
790 * This thread does not allocate any resource, except within
791 * handle_message, within mutex protection. This mutex protects against
793 * The other moment it allocates resources is at socket connection, which
794 * is also protected by the mutex.
797 void *ust_listener_thread(void *arg
)
799 struct sock_info
*sock_info
= arg
;
800 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
802 /* Restart trying to connect to the session daemon */
804 if (prev_connect_failed
) {
805 /* Wait for sessiond availability with pipe */
806 wait_for_sessiond(sock_info
);
810 * Sleep for 5 seconds before retrying after a
811 * sequence of failure / wait / failure. This
812 * deals with a killed or broken session daemon.
817 prev_connect_failed
= 0;
821 if (lttng_ust_comm_should_quit
) {
825 if (sock_info
->socket
!= -1) {
826 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
828 ERR("Error closing %s apps socket", sock_info
->name
);
830 sock_info
->socket
= -1;
834 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
836 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
837 prev_connect_failed
= 1;
839 * If we cannot find the sessiond daemon, don't delay
840 * constructor execution.
842 ret
= handle_register_done(sock_info
);
848 sock_info
->socket
= sock
= ret
;
851 * Create only one root handle per listener thread for the whole
852 * process lifetime, so we ensure we get ID which is statically
853 * assigned to the root handle.
855 if (sock_info
->root_handle
== -1) {
856 ret
= lttng_abi_create_root_handle();
858 ERR("Error creating root handle");
861 sock_info
->root_handle
= ret
;
864 ret
= register_app_to_sessiond(sock
);
866 ERR("Error registering to %s apps socket", sock_info
->name
);
867 prev_connect_failed
= 1;
869 * If we cannot register to the sessiond daemon, don't
870 * delay constructor execution.
872 ret
= handle_register_done(sock_info
);
881 struct ustcomm_ust_msg lum
;
883 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
885 case 0: /* orderly shutdown */
886 DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info
->name
);
888 if (lttng_ust_comm_should_quit
) {
892 * Either sessiond has shutdown or refused us by closing the socket.
893 * In either case, we don't want to delay construction execution,
894 * and we need to wait before retry.
896 prev_connect_failed
= 1;
898 * If we cannot register to the sessiond daemon, don't
899 * delay constructor execution.
901 ret
= handle_register_done(sock_info
);
906 DBG("message received");
907 ret
= handle_message(sock_info
, sock
, &lum
);
909 ERR("Error handling message for %s socket", sock_info
->name
);
914 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
916 DBG("incorrect message size (%s socket): %zd", sock_info
->name
, len
);
918 if (len
== -ECONNRESET
) {
919 DBG("%s remote end closed connection", sock_info
->name
);
928 if (lttng_ust_comm_should_quit
) {
931 /* Cleanup socket handles before trying to reconnect */
932 lttng_ust_objd_table_owner_cleanup(sock_info
);
934 goto restart
; /* try to reconnect */
937 sock_info
->thread_active
= 0;
943 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
946 int get_timeout(struct timespec
*constructor_timeout
)
948 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
952 str_delay
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
954 constructor_delay_ms
= strtol(str_delay
, NULL
, 10);
957 switch (constructor_delay_ms
) {
958 case -1:/* fall-through */
960 return constructor_delay_ms
;
966 * If we are unable to find the current time, don't wait.
968 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
972 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
973 constructor_timeout
->tv_nsec
+=
974 (constructor_delay_ms
% 1000UL) * 1000000UL;
975 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
976 constructor_timeout
->tv_sec
++;
977 constructor_timeout
->tv_nsec
-= 1000000000UL;
983 * sessiond monitoring thread: monitor presence of global and per-user
984 * sessiond by polling the application common named pipe.
986 void __attribute__((constructor
)) lttng_ust_init(void)
988 struct timespec constructor_timeout
;
989 sigset_t sig_all_blocked
, orig_parent_mask
;
990 pthread_attr_t thread_attr
;
994 if (uatomic_xchg(&initialized
, 1) == 1)
998 * Fixup interdependency between TLS fixup mutex (which happens
999 * to be the dynamic linker mutex) and ust_lock, taken within
1002 lttng_fixup_event_tls();
1003 lttng_fixup_ringbuffer_tls();
1004 lttng_fixup_vtid_tls();
1005 lttng_fixup_nest_count_tls();
1006 lttng_fixup_procname_tls();
1009 * We want precise control over the order in which we construct
1010 * our sub-libraries vs starting to receive commands from
1011 * sessiond (otherwise leading to errors when trying to create
1012 * sessiond before the init functions are completed).
1016 lttng_ring_buffer_metadata_client_init();
1017 lttng_ring_buffer_client_overwrite_init();
1018 lttng_ring_buffer_client_discard_init();
1020 timeout_mode
= get_timeout(&constructor_timeout
);
1022 ret
= sem_init(&constructor_wait
, 0, 0);
1025 ret
= setup_local_apps();
1027 DBG("local apps setup returned %d", ret
);
1030 /* A new thread created by pthread_create inherits the signal mask
1031 * from the parent. To avoid any signal being received by the
1032 * listener thread, we block all signals temporarily in the parent,
1033 * while we create the listener thread.
1035 sigfillset(&sig_all_blocked
);
1036 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
1038 ERR("pthread_sigmask: %s", strerror(ret
));
1041 ret
= pthread_attr_init(&thread_attr
);
1043 ERR("pthread_attr_init: %s", strerror(ret
));
1045 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
1047 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1051 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1052 ust_listener_thread
, &global_apps
);
1054 ERR("pthread_create global: %s", strerror(ret
));
1056 global_apps
.thread_active
= 1;
1059 if (local_apps
.allowed
) {
1061 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1062 ust_listener_thread
, &local_apps
);
1064 ERR("pthread_create local: %s", strerror(ret
));
1066 local_apps
.thread_active
= 1;
1069 handle_register_done(&local_apps
);
1071 ret
= pthread_attr_destroy(&thread_attr
);
1073 ERR("pthread_attr_destroy: %s", strerror(ret
));
1076 /* Restore original signal mask in parent */
1077 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1079 ERR("pthread_sigmask: %s", strerror(ret
));
1082 switch (timeout_mode
) {
1083 case 1: /* timeout wait */
1085 ret
= sem_timedwait(&constructor_wait
,
1086 &constructor_timeout
);
1087 } while (ret
< 0 && errno
== EINTR
);
1088 if (ret
< 0 && errno
== ETIMEDOUT
) {
1089 ERR("Timed out waiting for lttng-sessiond");
1094 case -1:/* wait forever */
1096 ret
= sem_wait(&constructor_wait
);
1097 } while (ret
< 0 && errno
== EINTR
);
1100 case 0: /* no timeout */
1106 void lttng_ust_cleanup(int exiting
)
1108 cleanup_sock_info(&global_apps
, exiting
);
1109 if (local_apps
.allowed
) {
1110 cleanup_sock_info(&local_apps
, exiting
);
1113 * The teardown in this function all affect data structures
1114 * accessed under the UST lock by the listener thread. This
1115 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1116 * that none of these threads are accessing this data at this
1119 lttng_ust_abi_exit();
1120 lttng_ust_events_exit();
1121 lttng_ring_buffer_client_discard_exit();
1122 lttng_ring_buffer_client_overwrite_exit();
1123 lttng_ring_buffer_metadata_client_exit();
1126 /* Reinitialize values for fork */
1128 lttng_ust_comm_should_quit
= 0;
1133 void __attribute__((destructor
)) lttng_ust_exit(void)
1138 * Using pthread_cancel here because:
1139 * A) we don't want to hang application teardown.
1140 * B) the thread is not allocating any resource.
1144 * Require the communication thread to quit. Synchronize with
1145 * mutexes to ensure it is not in a mutex critical section when
1146 * pthread_cancel is later called.
1149 lttng_ust_comm_should_quit
= 1;
1151 /* cancel threads */
1152 if (global_apps
.thread_active
) {
1153 ret
= pthread_cancel(global_apps
.ust_listener
);
1155 ERR("Error cancelling global ust listener thread: %s",
1158 global_apps
.thread_active
= 0;
1161 if (local_apps
.thread_active
) {
1162 ret
= pthread_cancel(local_apps
.ust_listener
);
1164 ERR("Error cancelling local ust listener thread: %s",
1167 local_apps
.thread_active
= 0;
1173 * Do NOT join threads: use of sys_futex makes it impossible to
1174 * join the threads without using async-cancel, but async-cancel
1175 * is delivered by a signal, which could hit the target thread
1176 * anywhere in its code path, including while the ust_lock() is
1177 * held, causing a deadlock for the other thread. Let the OS
1178 * cleanup the threads if there are stalled in a syscall.
1180 lttng_ust_cleanup(1);
1184 * We exclude the worker threads across fork and clone (except
1185 * CLONE_VM), because these system calls only keep the forking thread
1186 * running in the child. Therefore, we don't want to call fork or clone
1187 * in the middle of an tracepoint or ust tracing state modification.
1188 * Holding this mutex protects these structures across fork and clone.
1190 void ust_before_fork(sigset_t
*save_sigset
)
1193 * Disable signals. This is to avoid that the child intervenes
1194 * before it is properly setup for tracing. It is safer to
1195 * disable all signals, because then we know we are not breaking
1196 * anything by restoring the original mask.
1201 if (lttng_ust_nest_count
)
1203 /* Disable signals */
1204 sigfillset(&all_sigs
);
1205 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1207 PERROR("sigprocmask");
1210 rcu_bp_before_fork();
1213 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1217 DBG("process %d", getpid());
1219 /* Restore signals */
1220 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1222 PERROR("sigprocmask");
1226 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1228 if (lttng_ust_nest_count
)
1230 DBG("process %d", getpid());
1231 rcu_bp_after_fork_parent();
1232 /* Release mutexes and reenable signals */
1233 ust_after_fork_common(restore_sigset
);
1237 * After fork, in the child, we need to cleanup all the leftover state,
1238 * except the worker thread which already magically disappeared thanks
1239 * to the weird Linux fork semantics. After tyding up, we call
1240 * lttng_ust_init() again to start over as a new PID.
1242 * This is meant for forks() that have tracing in the child between the
1243 * fork and following exec call (if there is any).
1245 void ust_after_fork_child(sigset_t
*restore_sigset
)
1247 if (lttng_ust_nest_count
)
1249 DBG("process %d", getpid());
1250 /* Release urcu mutexes */
1251 rcu_bp_after_fork_child();
1252 lttng_ust_cleanup(0);
1253 lttng_context_vtid_reset();
1254 /* Release mutexes and reenable signals */
1255 ust_after_fork_common(restore_sigset
);