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_node
*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
->bc
.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
->bc
.len
= lum
->u
.filter
.data_size
;
341 bytecode
->bc
.reloc_offset
= lum
->u
.filter
.reloc_offset
;
342 bytecode
->bc
.seqnum
= lum
->u
.filter
.seqnum
;
344 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
345 (unsigned long) bytecode
,
350 /* don't free bytecode if everything went fine. */
359 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
360 (unsigned long) &lum
->u
,
368 lur
.handle
= lum
->handle
;
372 lur
.ret_code
= LTTNG_UST_OK
;
375 * Use -LTTNG_UST_ERR as wildcard for UST internal
376 * error that are not caused by the transport, except if
377 * we already have a more precise error message to
380 if (ret
> -LTTNG_UST_ERR
) {
381 /* Translate code to UST error. */
384 lur
.ret_code
= -LTTNG_UST_ERR_EXIST
;
387 lur
.ret_code
= -LTTNG_UST_ERR_INVAL
;
390 lur
.ret_code
= -LTTNG_UST_ERR_NOENT
;
393 lur
.ret_code
= -LTTNG_UST_ERR_PERM
;
396 lur
.ret_code
= -LTTNG_UST_ERR_NOSYS
;
399 lur
.ret_code
= -LTTNG_UST_ERR
;
408 case LTTNG_UST_STREAM
:
410 * Special-case reply to send stream info.
413 lur
.u
.stream
.memory_map_size
= *args
.stream
.memory_map_size
;
414 shm_fd
= *args
.stream
.shm_fd
;
415 wait_fd
= *args
.stream
.wait_fd
;
417 case LTTNG_UST_METADATA
:
418 case LTTNG_UST_CHANNEL
:
419 lur
.u
.channel
.memory_map_size
= *args
.channel
.memory_map_size
;
420 shm_fd
= *args
.channel
.shm_fd
;
421 wait_fd
= *args
.channel
.wait_fd
;
423 case LTTNG_UST_TRACER_VERSION
:
424 lur
.u
.version
= lum
->u
.version
;
426 case LTTNG_UST_TRACEPOINT_LIST_GET
:
427 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
431 ret
= send_reply(sock
, &lur
);
433 DBG("error sending reply");
437 if ((lum
->cmd
== LTTNG_UST_STREAM
438 || lum
->cmd
== LTTNG_UST_CHANNEL
439 || lum
->cmd
== LTTNG_UST_METADATA
)
440 && lur
.ret_code
== LTTNG_UST_OK
) {
443 /* we also need to send the file descriptors. */
444 ret
= ustcomm_send_fds_unix_sock(sock
,
452 * The sessiond expects 2 file descriptors, even upon
455 ret
= ustcomm_send_fds_unix_sock(sock
,
459 perror("send wait_fd");
468 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
471 if (lur
.ret_code
== LTTNG_UST_OK
) {
473 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
474 len
= ustcomm_send_unix_sock(sock
,
475 &args
.field_list
.entry
,
476 sizeof(args
.field_list
.entry
));
481 if (len
!= sizeof(args
.field_list
.entry
)) {
488 * We still have the memory map reference, and the fds have been
489 * sent to the sessiond. We can therefore close those fds. Note
490 * that we keep the write side of the wait_fd open, but close
493 if (lur
.ret_code
== LTTNG_UST_OK
) {
495 case LTTNG_UST_STREAM
:
499 PERROR("Error closing stream shm_fd");
501 *args
.stream
.shm_fd
= -1;
504 ret
= close(wait_fd
);
506 PERROR("Error closing stream wait_fd");
508 *args
.stream
.wait_fd
= -1;
511 case LTTNG_UST_METADATA
:
512 case LTTNG_UST_CHANNEL
:
516 PERROR("Error closing channel shm_fd");
518 *args
.channel
.shm_fd
= -1;
521 ret
= close(wait_fd
);
523 PERROR("Error closing channel wait_fd");
525 *args
.channel
.wait_fd
= -1;
537 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
541 if (sock_info
->socket
!= -1) {
542 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
544 ERR("Error closing apps socket");
546 sock_info
->socket
= -1;
548 if (sock_info
->root_handle
!= -1) {
549 ret
= lttng_ust_objd_unref(sock_info
->root_handle
);
551 ERR("Error unref root handle");
553 sock_info
->root_handle
= -1;
555 sock_info
->constructor_sem_posted
= 0;
557 * wait_shm_mmap is used by listener threads outside of the
558 * ust lock, so we cannot tear it down ourselves, because we
559 * cannot join on these threads. Leave this task to the OS
562 if (!exiting
&& sock_info
->wait_shm_mmap
) {
563 ret
= munmap(sock_info
->wait_shm_mmap
, sysconf(_SC_PAGE_SIZE
));
565 ERR("Error unmapping wait shm");
567 sock_info
->wait_shm_mmap
= NULL
;
572 * Using fork to set umask in the child process (not multi-thread safe).
573 * We deal with the shm_open vs ftruncate race (happening when the
574 * sessiond owns the shm and does not let everybody modify it, to ensure
575 * safety against shm_unlink) by simply letting the mmap fail and
576 * retrying after a few seconds.
577 * For global shm, everybody has rw access to it until the sessiond
581 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
583 int wait_shm_fd
, ret
;
587 * Try to open read-only.
589 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
590 if (wait_shm_fd
>= 0) {
592 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
594 * Real-only open did not work, and it's not because the
595 * entry was not present. It's a failure that prohibits
598 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
602 * If the open failed because the file did not exist, try
603 * creating it ourself.
605 lttng_ust_nest_count
++;
607 lttng_ust_nest_count
--;
612 * Parent: wait for child to return, in which case the
613 * shared memory map will have been created.
616 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
621 * Try to open read-only again after creation.
623 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
624 if (wait_shm_fd
< 0) {
626 * Real-only open did not work. It's a failure
627 * that prohibits using shm.
629 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
633 } else if (pid
== 0) {
637 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
638 if (sock_info
->global
)
639 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
641 * We're alone in a child process, so we can modify the
642 * process-wide umask.
646 * Try creating shm (or get rw access).
647 * We don't do an exclusive open, because we allow other
648 * processes to create+ftruncate it concurrently.
650 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
651 O_RDWR
| O_CREAT
, create_mode
);
652 if (wait_shm_fd
>= 0) {
653 ret
= ftruncate(wait_shm_fd
, mmap_size
);
661 * For local shm, we need to have rw access to accept
662 * opening it: this means the local sessiond will be
663 * able to wake us up. For global shm, we open it even
664 * if rw access is not granted, because the root.root
665 * sessiond will be able to override all rights and wake
668 if (!sock_info
->global
&& errno
!= EACCES
) {
669 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
673 * The shm exists, but we cannot open it RW. Report
681 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
685 * Ensure that our user is the owner of the shm file for
686 * local shm. If we do not own the file, it means our
687 * sessiond will not have access to wake us up (there is
688 * probably a rogue process trying to fake our
689 * sessiond). Fallback to polling method in this case.
691 ret
= fstat(wait_shm_fd
, &statbuf
);
696 if (statbuf
.st_uid
!= getuid())
702 ret
= close(wait_shm_fd
);
704 PERROR("Error closing fd");
710 char *get_map_shm(struct sock_info
*sock_info
)
712 size_t mmap_size
= sysconf(_SC_PAGE_SIZE
);
713 int wait_shm_fd
, ret
;
716 wait_shm_fd
= get_wait_shm(sock_info
, mmap_size
);
717 if (wait_shm_fd
< 0) {
720 wait_shm_mmap
= mmap(NULL
, mmap_size
, PROT_READ
,
721 MAP_SHARED
, wait_shm_fd
, 0);
722 /* close shm fd immediately after taking the mmap reference */
723 ret
= close(wait_shm_fd
);
725 PERROR("Error closing fd");
727 if (wait_shm_mmap
== MAP_FAILED
) {
728 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
731 return wait_shm_mmap
;
738 void wait_for_sessiond(struct sock_info
*sock_info
)
743 if (lttng_ust_comm_should_quit
) {
746 if (wait_poll_fallback
) {
749 if (!sock_info
->wait_shm_mmap
) {
750 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
751 if (!sock_info
->wait_shm_mmap
)
756 DBG("Waiting for %s apps sessiond", sock_info
->name
);
757 /* Wait for futex wakeup */
758 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
759 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
760 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
762 if (errno
== EFAULT
) {
763 wait_poll_fallback
= 1;
765 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
766 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
767 "Please upgrade your kernel "
768 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
769 "mainline). LTTng-UST will use polling mode fallback.");
787 * This thread does not allocate any resource, except within
788 * handle_message, within mutex protection. This mutex protects against
790 * The other moment it allocates resources is at socket connection, which
791 * is also protected by the mutex.
794 void *ust_listener_thread(void *arg
)
796 struct sock_info
*sock_info
= arg
;
797 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
799 /* Restart trying to connect to the session daemon */
801 if (prev_connect_failed
) {
802 /* Wait for sessiond availability with pipe */
803 wait_for_sessiond(sock_info
);
807 * Sleep for 5 seconds before retrying after a
808 * sequence of failure / wait / failure. This
809 * deals with a killed or broken session daemon.
814 prev_connect_failed
= 0;
818 if (lttng_ust_comm_should_quit
) {
823 if (sock_info
->socket
!= -1) {
824 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
826 ERR("Error closing %s apps socket", sock_info
->name
);
828 sock_info
->socket
= -1;
832 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
834 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
835 prev_connect_failed
= 1;
837 * If we cannot find the sessiond daemon, don't delay
838 * constructor execution.
840 ret
= handle_register_done(sock_info
);
846 sock_info
->socket
= sock
= ret
;
849 * Create only one root handle per listener thread for the whole
850 * process lifetime, so we ensure we get ID which is statically
851 * assigned to the root handle.
853 if (sock_info
->root_handle
== -1) {
854 ret
= lttng_abi_create_root_handle();
856 ERR("Error creating root handle");
860 sock_info
->root_handle
= ret
;
863 ret
= register_app_to_sessiond(sock
);
865 ERR("Error registering to %s apps socket", sock_info
->name
);
866 prev_connect_failed
= 1;
868 * If we cannot register to the sessiond daemon, don't
869 * delay constructor execution.
871 ret
= handle_register_done(sock_info
);
880 struct ustcomm_ust_msg lum
;
882 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
884 case 0: /* orderly shutdown */
885 DBG("%s ltt-sessiond has performed an orderly shutdown", sock_info
->name
);
888 * Either sessiond has shutdown or refused us by closing the socket.
889 * In either case, we don't want to delay construction execution,
890 * and we need to wait before retry.
892 prev_connect_failed
= 1;
894 * If we cannot register to the sessiond daemon, don't
895 * delay constructor execution.
897 ret
= handle_register_done(sock_info
);
902 DBG("message received");
903 ret
= handle_message(sock_info
, sock
, &lum
);
905 ERR("Error handling message for %s socket", sock_info
->name
);
910 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
912 DBG("incorrect message size (%s socket): %zd", sock_info
->name
, len
);
914 if (len
== -ECONNRESET
) {
915 DBG("%s remote end closed connection", sock_info
->name
);
924 /* Cleanup socket handles before trying to reconnect */
925 lttng_ust_objd_table_owner_cleanup(sock_info
);
927 goto restart
; /* try to reconnect */
933 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
936 int get_timeout(struct timespec
*constructor_timeout
)
938 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
942 str_delay
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
944 constructor_delay_ms
= strtol(str_delay
, NULL
, 10);
947 switch (constructor_delay_ms
) {
948 case -1:/* fall-through */
950 return constructor_delay_ms
;
956 * If we are unable to find the current time, don't wait.
958 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
962 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
963 constructor_timeout
->tv_nsec
+=
964 (constructor_delay_ms
% 1000UL) * 1000000UL;
965 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
966 constructor_timeout
->tv_sec
++;
967 constructor_timeout
->tv_nsec
-= 1000000000UL;
973 * sessiond monitoring thread: monitor presence of global and per-user
974 * sessiond by polling the application common named pipe.
976 void __attribute__((constructor
)) lttng_ust_init(void)
978 struct timespec constructor_timeout
;
979 sigset_t sig_all_blocked
, orig_parent_mask
;
980 pthread_attr_t thread_attr
;
984 if (uatomic_xchg(&initialized
, 1) == 1)
988 * Fixup interdependency between TLS fixup mutex (which happens
989 * to be the dynamic linker mutex) and ust_lock, taken within
992 lttng_fixup_event_tls();
993 lttng_fixup_ringbuffer_tls();
994 lttng_fixup_vtid_tls();
995 lttng_fixup_nest_count_tls();
996 lttng_fixup_procname_tls();
999 * We want precise control over the order in which we construct
1000 * our sub-libraries vs starting to receive commands from
1001 * sessiond (otherwise leading to errors when trying to create
1002 * sessiond before the init functions are completed).
1006 ltt_ring_buffer_metadata_client_init();
1007 ltt_ring_buffer_client_overwrite_init();
1008 ltt_ring_buffer_client_discard_init();
1010 timeout_mode
= get_timeout(&constructor_timeout
);
1012 ret
= sem_init(&constructor_wait
, 0, 0);
1015 ret
= setup_local_apps();
1017 DBG("local apps setup returned %d", ret
);
1020 /* A new thread created by pthread_create inherits the signal mask
1021 * from the parent. To avoid any signal being received by the
1022 * listener thread, we block all signals temporarily in the parent,
1023 * while we create the listener thread.
1025 sigfillset(&sig_all_blocked
);
1026 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
1028 ERR("pthread_sigmask: %s", strerror(ret
));
1031 ret
= pthread_attr_init(&thread_attr
);
1033 ERR("pthread_attr_init: %s", strerror(ret
));
1035 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
1037 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1040 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1041 ust_listener_thread
, &global_apps
);
1043 ERR("pthread_create global: %s", strerror(ret
));
1045 if (local_apps
.allowed
) {
1046 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1047 ust_listener_thread
, &local_apps
);
1049 ERR("pthread_create local: %s", strerror(ret
));
1052 handle_register_done(&local_apps
);
1054 ret
= pthread_attr_destroy(&thread_attr
);
1056 ERR("pthread_attr_destroy: %s", strerror(ret
));
1059 /* Restore original signal mask in parent */
1060 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1062 ERR("pthread_sigmask: %s", strerror(ret
));
1065 switch (timeout_mode
) {
1066 case 1: /* timeout wait */
1068 ret
= sem_timedwait(&constructor_wait
,
1069 &constructor_timeout
);
1070 } while (ret
< 0 && errno
== EINTR
);
1071 if (ret
< 0 && errno
== ETIMEDOUT
) {
1072 ERR("Timed out waiting for ltt-sessiond");
1077 case -1:/* wait forever */
1079 ret
= sem_wait(&constructor_wait
);
1080 } while (ret
< 0 && errno
== EINTR
);
1083 case 0: /* no timeout */
1089 void lttng_ust_cleanup(int exiting
)
1091 cleanup_sock_info(&global_apps
, exiting
);
1092 if (local_apps
.allowed
) {
1093 cleanup_sock_info(&local_apps
, exiting
);
1096 * The teardown in this function all affect data structures
1097 * accessed under the UST lock by the listener thread. This
1098 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1099 * that none of these threads are accessing this data at this
1102 lttng_ust_abi_exit();
1103 lttng_ust_events_exit();
1104 ltt_ring_buffer_client_discard_exit();
1105 ltt_ring_buffer_client_overwrite_exit();
1106 ltt_ring_buffer_metadata_client_exit();
1109 /* Reinitialize values for fork */
1111 lttng_ust_comm_should_quit
= 0;
1116 void __attribute__((destructor
)) lttng_ust_exit(void)
1121 * Using pthread_cancel here because:
1122 * A) we don't want to hang application teardown.
1123 * B) the thread is not allocating any resource.
1127 * Require the communication thread to quit. Synchronize with
1128 * mutexes to ensure it is not in a mutex critical section when
1129 * pthread_cancel is later called.
1132 lttng_ust_comm_should_quit
= 1;
1135 /* cancel threads */
1136 ret
= pthread_cancel(global_apps
.ust_listener
);
1138 ERR("Error cancelling global ust listener thread: %s",
1141 if (local_apps
.allowed
) {
1142 ret
= pthread_cancel(local_apps
.ust_listener
);
1144 ERR("Error cancelling local ust listener thread: %s",
1149 * Do NOT join threads: use of sys_futex makes it impossible to
1150 * join the threads without using async-cancel, but async-cancel
1151 * is delivered by a signal, which could hit the target thread
1152 * anywhere in its code path, including while the ust_lock() is
1153 * held, causing a deadlock for the other thread. Let the OS
1154 * cleanup the threads if there are stalled in a syscall.
1156 lttng_ust_cleanup(1);
1160 * We exclude the worker threads across fork and clone (except
1161 * CLONE_VM), because these system calls only keep the forking thread
1162 * running in the child. Therefore, we don't want to call fork or clone
1163 * in the middle of an tracepoint or ust tracing state modification.
1164 * Holding this mutex protects these structures across fork and clone.
1166 void ust_before_fork(sigset_t
*save_sigset
)
1169 * Disable signals. This is to avoid that the child intervenes
1170 * before it is properly setup for tracing. It is safer to
1171 * disable all signals, because then we know we are not breaking
1172 * anything by restoring the original mask.
1177 if (lttng_ust_nest_count
)
1179 /* Disable signals */
1180 sigfillset(&all_sigs
);
1181 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1183 PERROR("sigprocmask");
1186 rcu_bp_before_fork();
1189 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1193 DBG("process %d", getpid());
1195 /* Restore signals */
1196 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1198 PERROR("sigprocmask");
1202 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1204 if (lttng_ust_nest_count
)
1206 DBG("process %d", getpid());
1207 rcu_bp_after_fork_parent();
1208 /* Release mutexes and reenable signals */
1209 ust_after_fork_common(restore_sigset
);
1213 * After fork, in the child, we need to cleanup all the leftover state,
1214 * except the worker thread which already magically disappeared thanks
1215 * to the weird Linux fork semantics. After tyding up, we call
1216 * lttng_ust_init() again to start over as a new PID.
1218 * This is meant for forks() that have tracing in the child between the
1219 * fork and following exec call (if there is any).
1221 void ust_after_fork_child(sigset_t
*restore_sigset
)
1223 if (lttng_ust_nest_count
)
1225 DBG("process %d", getpid());
1226 /* Release urcu mutexes */
1227 rcu_bp_after_fork_child();
1228 lttng_ust_cleanup(0);
1229 lttng_context_vtid_reset();
1230 /* Release mutexes and reenable signals */
1231 ust_after_fork_common(restore_sigset
);