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 "ltt-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
;
100 char sock_path
[PATH_MAX
];
103 char wait_shm_path
[PATH_MAX
];
107 /* Socket from app (connect) to session daemon (listen) for communication */
108 struct sock_info global_apps
= {
115 .sock_path
= DEFAULT_GLOBAL_APPS_UNIX_SOCK
,
118 .wait_shm_path
= DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH
,
121 /* TODO: allow global_apps_sock_path override */
123 struct sock_info local_apps
= {
127 .allowed
= 0, /* Check setuid bit first */
132 static int wait_poll_fallback
;
134 extern void ltt_ring_buffer_client_overwrite_init(void);
135 extern void ltt_ring_buffer_client_discard_init(void);
136 extern void ltt_ring_buffer_metadata_client_init(void);
137 extern void ltt_ring_buffer_client_overwrite_exit(void);
138 extern void ltt_ring_buffer_client_discard_exit(void);
139 extern void ltt_ring_buffer_metadata_client_exit(void);
142 * Force a read (imply TLS fixup for dlopen) of TLS variables.
145 void lttng_fixup_nest_count_tls(void)
147 asm volatile ("" : : "m" (lttng_ust_nest_count
));
151 int setup_local_apps(void)
153 const char *home_dir
;
158 * Disallow per-user tracing for setuid binaries.
160 if (uid
!= geteuid()) {
161 assert(local_apps
.allowed
== 0);
164 home_dir
= (const char *) getenv("HOME");
166 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
167 assert(local_apps
.allowed
== 0);
170 local_apps
.allowed
= 1;
171 snprintf(local_apps
.sock_path
, PATH_MAX
,
172 DEFAULT_HOME_APPS_UNIX_SOCK
, home_dir
);
173 snprintf(local_apps
.wait_shm_path
, PATH_MAX
,
174 DEFAULT_HOME_APPS_WAIT_SHM_PATH
, uid
);
179 int register_app_to_sessiond(int socket
)
189 uint32_t bits_per_long
;
190 char name
[16]; /* process name */
193 reg_msg
.major
= LTTNG_UST_COMM_VERSION_MAJOR
;
194 reg_msg
.minor
= LTTNG_UST_COMM_VERSION_MINOR
;
195 reg_msg
.pid
= getpid();
196 reg_msg
.ppid
= getppid();
197 reg_msg
.uid
= getuid();
198 reg_msg
.gid
= getgid();
199 reg_msg
.bits_per_long
= CAA_BITS_PER_LONG
;
200 lttng_ust_getprocname(reg_msg
.name
);
202 ret
= ustcomm_send_unix_sock(socket
, ®_msg
, sizeof(reg_msg
));
203 if (ret
>= 0 && ret
!= sizeof(reg_msg
))
209 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
213 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
216 DBG("message successfully sent");
219 if (len
== -ECONNRESET
) {
220 DBG("remote end closed connection");
225 DBG("incorrect message size: %zd", len
);
231 int handle_register_done(struct sock_info
*sock_info
)
235 if (sock_info
->constructor_sem_posted
)
237 sock_info
->constructor_sem_posted
= 1;
238 if (uatomic_read(&sem_count
) <= 0) {
241 ret
= uatomic_add_return(&sem_count
, -1);
243 ret
= sem_post(&constructor_wait
);
250 int handle_message(struct sock_info
*sock_info
,
251 int sock
, struct ustcomm_ust_msg
*lum
)
254 const struct lttng_ust_objd_ops
*ops
;
255 struct ustcomm_ust_reply lur
;
262 memset(&lur
, 0, sizeof(lur
));
264 if (lttng_ust_comm_should_quit
) {
269 ops
= objd_ops(lum
->handle
);
276 case LTTNG_UST_REGISTER_DONE
:
277 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
278 ret
= handle_register_done(sock_info
);
282 case LTTNG_UST_RELEASE
:
283 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
286 ret
= lttng_ust_objd_unref(lum
->handle
);
288 case LTTNG_UST_FILTER
:
290 /* Receive filter data */
291 struct lttng_ust_filter_bytecode
*bytecode
;
293 if (lum
->u
.filter
.data_size
> FILTER_BYTECODE_MAX_LEN
) {
294 ERR("Filter data size is too large: %u bytes",
295 lum
->u
.filter
.data_size
);
300 if (lum
->u
.filter
.reloc_offset
> lum
->u
.filter
.data_size
) {
301 ERR("Filter reloc offset %u is not within data",
302 lum
->u
.filter
.reloc_offset
);
307 bytecode
= zmalloc(sizeof(*bytecode
) + lum
->u
.filter
.data_size
);
312 len
= ustcomm_recv_unix_sock(sock
, bytecode
->data
,
313 lum
->u
.filter
.data_size
);
315 case 0: /* orderly shutdown */
320 if (len
== lum
->u
.filter
.data_size
) {
321 DBG("filter data received");
323 } else if (len
< 0) {
324 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
325 if (len
== -ECONNRESET
) {
326 ERR("%s remote end closed connection", sock_info
->name
);
334 DBG("incorrect filter data message size: %zd", len
);
340 bytecode
->len
= lum
->u
.filter
.data_size
;
341 bytecode
->reloc_offset
= lum
->u
.filter
.reloc_offset
;
343 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
344 (unsigned long) bytecode
,
349 /* don't free bytecode if everything went fine. */
358 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
359 (unsigned long) &lum
->u
,
367 lur
.handle
= lum
->handle
;
371 lur
.ret_code
= LTTNG_UST_OK
;
374 * Use -LTTNG_UST_ERR as wildcard for UST internal
375 * error that are not caused by the transport, except if
376 * we already have a more precise error message to
379 if (ret
> -LTTNG_UST_ERR
)
380 lur
.ret_code
= -LTTNG_UST_ERR
;
386 case LTTNG_UST_STREAM
:
388 * Special-case reply to send stream info.
391 lur
.u
.stream
.memory_map_size
= *args
.stream
.memory_map_size
;
392 shm_fd
= *args
.stream
.shm_fd
;
393 wait_fd
= *args
.stream
.wait_fd
;
395 case LTTNG_UST_METADATA
:
396 case LTTNG_UST_CHANNEL
:
397 lur
.u
.channel
.memory_map_size
= *args
.channel
.memory_map_size
;
398 shm_fd
= *args
.channel
.shm_fd
;
399 wait_fd
= *args
.channel
.wait_fd
;
401 case LTTNG_UST_TRACER_VERSION
:
402 lur
.u
.version
= lum
->u
.version
;
404 case LTTNG_UST_TRACEPOINT_LIST_GET
:
405 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
409 ret
= send_reply(sock
, &lur
);
411 DBG("error sending reply");
415 if ((lum
->cmd
== LTTNG_UST_STREAM
416 || lum
->cmd
== LTTNG_UST_CHANNEL
417 || lum
->cmd
== LTTNG_UST_METADATA
)
418 && lur
.ret_code
== LTTNG_UST_OK
) {
421 /* we also need to send the file descriptors. */
422 ret
= ustcomm_send_fds_unix_sock(sock
,
430 * The sessiond expects 2 file descriptors, even upon
433 ret
= ustcomm_send_fds_unix_sock(sock
,
437 perror("send wait_fd");
446 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
449 if (lur
.ret_code
== LTTNG_UST_OK
) {
451 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
452 len
= ustcomm_send_unix_sock(sock
,
453 &args
.field_list
.entry
,
454 sizeof(args
.field_list
.entry
));
459 if (len
!= sizeof(args
.field_list
.entry
)) {
466 * We still have the memory map reference, and the fds have been
467 * sent to the sessiond. We can therefore close those fds. Note
468 * that we keep the write side of the wait_fd open, but close
471 if (lur
.ret_code
== LTTNG_UST_OK
) {
473 case LTTNG_UST_STREAM
:
477 PERROR("Error closing stream shm_fd");
479 *args
.stream
.shm_fd
= -1;
482 ret
= close(wait_fd
);
484 PERROR("Error closing stream wait_fd");
486 *args
.stream
.wait_fd
= -1;
489 case LTTNG_UST_METADATA
:
490 case LTTNG_UST_CHANNEL
:
494 PERROR("Error closing channel shm_fd");
496 *args
.channel
.shm_fd
= -1;
499 ret
= close(wait_fd
);
501 PERROR("Error closing channel wait_fd");
503 *args
.channel
.wait_fd
= -1;
515 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
519 if (sock_info
->socket
!= -1) {
520 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
522 ERR("Error closing apps socket");
524 sock_info
->socket
= -1;
526 if (sock_info
->root_handle
!= -1) {
527 ret
= lttng_ust_objd_unref(sock_info
->root_handle
);
529 ERR("Error unref root handle");
531 sock_info
->root_handle
= -1;
533 sock_info
->constructor_sem_posted
= 0;
535 * wait_shm_mmap is used by listener threads outside of the
536 * ust lock, so we cannot tear it down ourselves, because we
537 * cannot join on these threads. Leave this task to the OS
540 if (!exiting
&& sock_info
->wait_shm_mmap
) {
541 ret
= munmap(sock_info
->wait_shm_mmap
, sysconf(_SC_PAGE_SIZE
));
543 ERR("Error unmapping wait shm");
545 sock_info
->wait_shm_mmap
= NULL
;
550 * Using fork to set umask in the child process (not multi-thread safe).
551 * We deal with the shm_open vs ftruncate race (happening when the
552 * sessiond owns the shm and does not let everybody modify it, to ensure
553 * safety against shm_unlink) by simply letting the mmap fail and
554 * retrying after a few seconds.
555 * For global shm, everybody has rw access to it until the sessiond
559 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
561 int wait_shm_fd
, ret
;
565 * Try to open read-only.
567 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
568 if (wait_shm_fd
>= 0) {
570 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
572 * Real-only open did not work, and it's not because the
573 * entry was not present. It's a failure that prohibits
576 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
580 * If the open failed because the file did not exist, try
581 * creating it ourself.
583 lttng_ust_nest_count
++;
585 lttng_ust_nest_count
--;
590 * Parent: wait for child to return, in which case the
591 * shared memory map will have been created.
594 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
599 * Try to open read-only again after creation.
601 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
602 if (wait_shm_fd
< 0) {
604 * Real-only open did not work. It's a failure
605 * that prohibits using shm.
607 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
611 } else if (pid
== 0) {
615 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
616 if (sock_info
->global
)
617 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
619 * We're alone in a child process, so we can modify the
620 * process-wide umask.
624 * Try creating shm (or get rw access).
625 * We don't do an exclusive open, because we allow other
626 * processes to create+ftruncate it concurrently.
628 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
629 O_RDWR
| O_CREAT
, create_mode
);
630 if (wait_shm_fd
>= 0) {
631 ret
= ftruncate(wait_shm_fd
, mmap_size
);
639 * For local shm, we need to have rw access to accept
640 * opening it: this means the local sessiond will be
641 * able to wake us up. For global shm, we open it even
642 * if rw access is not granted, because the root.root
643 * sessiond will be able to override all rights and wake
646 if (!sock_info
->global
&& errno
!= EACCES
) {
647 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
651 * The shm exists, but we cannot open it RW. Report
659 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
663 * Ensure that our user is the owner of the shm file for
664 * local shm. If we do not own the file, it means our
665 * sessiond will not have access to wake us up (there is
666 * probably a rogue process trying to fake our
667 * sessiond). Fallback to polling method in this case.
669 ret
= fstat(wait_shm_fd
, &statbuf
);
674 if (statbuf
.st_uid
!= getuid())
680 ret
= close(wait_shm_fd
);
682 PERROR("Error closing fd");
688 char *get_map_shm(struct sock_info
*sock_info
)
690 size_t mmap_size
= sysconf(_SC_PAGE_SIZE
);
691 int wait_shm_fd
, ret
;
694 wait_shm_fd
= get_wait_shm(sock_info
, mmap_size
);
695 if (wait_shm_fd
< 0) {
698 wait_shm_mmap
= mmap(NULL
, mmap_size
, PROT_READ
,
699 MAP_SHARED
, wait_shm_fd
, 0);
700 /* close shm fd immediately after taking the mmap reference */
701 ret
= close(wait_shm_fd
);
703 PERROR("Error closing fd");
705 if (wait_shm_mmap
== MAP_FAILED
) {
706 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
709 return wait_shm_mmap
;
716 void wait_for_sessiond(struct sock_info
*sock_info
)
721 if (lttng_ust_comm_should_quit
) {
724 if (wait_poll_fallback
) {
727 if (!sock_info
->wait_shm_mmap
) {
728 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
729 if (!sock_info
->wait_shm_mmap
)
734 DBG("Waiting for %s apps sessiond", sock_info
->name
);
735 /* Wait for futex wakeup */
736 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
737 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
738 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
740 if (errno
== EFAULT
) {
741 wait_poll_fallback
= 1;
743 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
744 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
745 "Please upgrade your kernel "
746 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
747 "mainline). LTTng-UST will use polling mode fallback.");
765 * This thread does not allocate any resource, except within
766 * handle_message, within mutex protection. This mutex protects against
768 * The other moment it allocates resources is at socket connection, which
769 * is also protected by the mutex.
772 void *ust_listener_thread(void *arg
)
774 struct sock_info
*sock_info
= arg
;
775 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
777 /* Restart trying to connect to the session daemon */
779 if (prev_connect_failed
) {
780 /* Wait for sessiond availability with pipe */
781 wait_for_sessiond(sock_info
);
785 * Sleep for 5 seconds before retrying after a
786 * sequence of failure / wait / failure. This
787 * deals with a killed or broken session daemon.
792 prev_connect_failed
= 0;
796 if (lttng_ust_comm_should_quit
) {
801 if (sock_info
->socket
!= -1) {
802 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
804 ERR("Error closing %s apps socket", sock_info
->name
);
806 sock_info
->socket
= -1;
810 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
812 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
813 prev_connect_failed
= 1;
815 * If we cannot find the sessiond daemon, don't delay
816 * constructor execution.
818 ret
= handle_register_done(sock_info
);
824 sock_info
->socket
= sock
= ret
;
827 * Create only one root handle per listener thread for the whole
828 * process lifetime, so we ensure we get ID which is statically
829 * assigned to the root handle.
831 if (sock_info
->root_handle
== -1) {
832 ret
= lttng_abi_create_root_handle();
834 ERR("Error creating root handle");
838 sock_info
->root_handle
= ret
;
841 ret
= register_app_to_sessiond(sock
);
843 ERR("Error registering to %s apps socket", sock_info
->name
);
844 prev_connect_failed
= 1;
846 * If we cannot register to the sessiond daemon, don't
847 * delay constructor execution.
849 ret
= handle_register_done(sock_info
);
858 struct ustcomm_ust_msg lum
;
860 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
862 case 0: /* orderly shutdown */
863 DBG("%s ltt-sessiond has performed an orderly shutdown", sock_info
->name
);
866 * Either sessiond has shutdown or refused us by closing the socket.
867 * In either case, we don't want to delay construction execution,
868 * and we need to wait before retry.
870 prev_connect_failed
= 1;
872 * If we cannot register to the sessiond daemon, don't
873 * delay constructor execution.
875 ret
= handle_register_done(sock_info
);
880 DBG("message received");
881 ret
= handle_message(sock_info
, sock
, &lum
);
883 ERR("Error handling message for %s socket", sock_info
->name
);
888 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
890 DBG("incorrect message size (%s socket): %zd", sock_info
->name
, len
);
892 if (len
== -ECONNRESET
) {
893 DBG("%s remote end closed connection", sock_info
->name
);
902 /* Cleanup socket handles before trying to reconnect */
903 lttng_ust_objd_table_owner_cleanup(sock_info
);
905 goto restart
; /* try to reconnect */
911 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
914 int get_timeout(struct timespec
*constructor_timeout
)
916 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
920 str_delay
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
922 constructor_delay_ms
= strtol(str_delay
, NULL
, 10);
925 switch (constructor_delay_ms
) {
926 case -1:/* fall-through */
928 return constructor_delay_ms
;
934 * If we are unable to find the current time, don't wait.
936 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
940 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
941 constructor_timeout
->tv_nsec
+=
942 (constructor_delay_ms
% 1000UL) * 1000000UL;
943 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
944 constructor_timeout
->tv_sec
++;
945 constructor_timeout
->tv_nsec
-= 1000000000UL;
951 * sessiond monitoring thread: monitor presence of global and per-user
952 * sessiond by polling the application common named pipe.
954 void __attribute__((constructor
)) lttng_ust_init(void)
956 struct timespec constructor_timeout
;
957 sigset_t sig_all_blocked
, orig_parent_mask
;
958 pthread_attr_t thread_attr
;
962 if (uatomic_xchg(&initialized
, 1) == 1)
966 * Fixup interdependency between TLS fixup mutex (which happens
967 * to be the dynamic linker mutex) and ust_lock, taken within
970 lttng_fixup_event_tls();
971 lttng_fixup_ringbuffer_tls();
972 lttng_fixup_vtid_tls();
973 lttng_fixup_nest_count_tls();
974 lttng_fixup_procname_tls();
977 * We want precise control over the order in which we construct
978 * our sub-libraries vs starting to receive commands from
979 * sessiond (otherwise leading to errors when trying to create
980 * sessiond before the init functions are completed).
984 ltt_ring_buffer_metadata_client_init();
985 ltt_ring_buffer_client_overwrite_init();
986 ltt_ring_buffer_client_discard_init();
988 timeout_mode
= get_timeout(&constructor_timeout
);
990 ret
= sem_init(&constructor_wait
, 0, 0);
993 ret
= setup_local_apps();
995 DBG("local apps setup returned %d", ret
);
998 /* A new thread created by pthread_create inherits the signal mask
999 * from the parent. To avoid any signal being received by the
1000 * listener thread, we block all signals temporarily in the parent,
1001 * while we create the listener thread.
1003 sigfillset(&sig_all_blocked
);
1004 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
1006 ERR("pthread_sigmask: %s", strerror(ret
));
1009 ret
= pthread_attr_init(&thread_attr
);
1011 ERR("pthread_attr_init: %s", strerror(ret
));
1013 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
1015 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1018 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1019 ust_listener_thread
, &global_apps
);
1021 ERR("pthread_create global: %s", strerror(ret
));
1023 if (local_apps
.allowed
) {
1024 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1025 ust_listener_thread
, &local_apps
);
1027 ERR("pthread_create local: %s", strerror(ret
));
1030 handle_register_done(&local_apps
);
1032 ret
= pthread_attr_destroy(&thread_attr
);
1034 ERR("pthread_attr_destroy: %s", strerror(ret
));
1037 /* Restore original signal mask in parent */
1038 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1040 ERR("pthread_sigmask: %s", strerror(ret
));
1043 switch (timeout_mode
) {
1044 case 1: /* timeout wait */
1046 ret
= sem_timedwait(&constructor_wait
,
1047 &constructor_timeout
);
1048 } while (ret
< 0 && errno
== EINTR
);
1049 if (ret
< 0 && errno
== ETIMEDOUT
) {
1050 ERR("Timed out waiting for ltt-sessiond");
1055 case -1:/* wait forever */
1057 ret
= sem_wait(&constructor_wait
);
1058 } while (ret
< 0 && errno
== EINTR
);
1061 case 0: /* no timeout */
1067 void lttng_ust_cleanup(int exiting
)
1069 cleanup_sock_info(&global_apps
, exiting
);
1070 if (local_apps
.allowed
) {
1071 cleanup_sock_info(&local_apps
, exiting
);
1074 * The teardown in this function all affect data structures
1075 * accessed under the UST lock by the listener thread. This
1076 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1077 * that none of these threads are accessing this data at this
1080 lttng_ust_abi_exit();
1081 lttng_ust_events_exit();
1082 ltt_ring_buffer_client_discard_exit();
1083 ltt_ring_buffer_client_overwrite_exit();
1084 ltt_ring_buffer_metadata_client_exit();
1087 /* Reinitialize values for fork */
1089 lttng_ust_comm_should_quit
= 0;
1094 void __attribute__((destructor
)) lttng_ust_exit(void)
1099 * Using pthread_cancel here because:
1100 * A) we don't want to hang application teardown.
1101 * B) the thread is not allocating any resource.
1105 * Require the communication thread to quit. Synchronize with
1106 * mutexes to ensure it is not in a mutex critical section when
1107 * pthread_cancel is later called.
1110 lttng_ust_comm_should_quit
= 1;
1113 /* cancel threads */
1114 ret
= pthread_cancel(global_apps
.ust_listener
);
1116 ERR("Error cancelling global ust listener thread: %s",
1119 if (local_apps
.allowed
) {
1120 ret
= pthread_cancel(local_apps
.ust_listener
);
1122 ERR("Error cancelling local ust listener thread: %s",
1127 * Do NOT join threads: use of sys_futex makes it impossible to
1128 * join the threads without using async-cancel, but async-cancel
1129 * is delivered by a signal, which could hit the target thread
1130 * anywhere in its code path, including while the ust_lock() is
1131 * held, causing a deadlock for the other thread. Let the OS
1132 * cleanup the threads if there are stalled in a syscall.
1134 lttng_ust_cleanup(1);
1138 * We exclude the worker threads across fork and clone (except
1139 * CLONE_VM), because these system calls only keep the forking thread
1140 * running in the child. Therefore, we don't want to call fork or clone
1141 * in the middle of an tracepoint or ust tracing state modification.
1142 * Holding this mutex protects these structures across fork and clone.
1144 void ust_before_fork(sigset_t
*save_sigset
)
1147 * Disable signals. This is to avoid that the child intervenes
1148 * before it is properly setup for tracing. It is safer to
1149 * disable all signals, because then we know we are not breaking
1150 * anything by restoring the original mask.
1155 if (lttng_ust_nest_count
)
1157 /* Disable signals */
1158 sigfillset(&all_sigs
);
1159 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1161 PERROR("sigprocmask");
1164 rcu_bp_before_fork();
1167 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1171 DBG("process %d", getpid());
1173 /* Restore signals */
1174 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1176 PERROR("sigprocmask");
1180 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1182 if (lttng_ust_nest_count
)
1184 DBG("process %d", getpid());
1185 rcu_bp_after_fork_parent();
1186 /* Release mutexes and reenable signals */
1187 ust_after_fork_common(restore_sigset
);
1191 * After fork, in the child, we need to cleanup all the leftover state,
1192 * except the worker thread which already magically disappeared thanks
1193 * to the weird Linux fork semantics. After tyding up, we call
1194 * lttng_ust_init() again to start over as a new PID.
1196 * This is meant for forks() that have tracing in the child between the
1197 * fork and following exec call (if there is any).
1199 void ust_after_fork_child(sigset_t
*restore_sigset
)
1201 if (lttng_ust_nest_count
)
1203 DBG("process %d", getpid());
1204 /* Release urcu mutexes */
1205 rcu_bp_after_fork_child();
1206 lttng_ust_cleanup(0);
1207 lttng_context_vtid_reset();
1208 /* Release mutexes and reenable signals */
1209 ust_after_fork_common(restore_sigset
);