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>
45 #include <usterr-signal-safe.h>
47 #include "tracepoint-internal.h"
48 #include "ltt-tracer-core.h"
50 #include "../libringbuffer/tlsfixup.h"
53 * Has lttng ust comm constructor been called ?
55 static int initialized
;
58 * The ust_lock/ust_unlock lock is used as a communication thread mutex.
59 * Held when handling a command, also held by fork() to deal with
60 * removal of threads, and by exit path.
63 /* Should the ust comm thread quit ? */
64 static int lttng_ust_comm_should_quit
;
67 * Wait for either of these before continuing to the main
69 * - the register_done message from sessiond daemon
70 * (will let the sessiond daemon enable sessions before main
72 * - sessiond daemon is not reachable.
73 * - timeout (ensuring applications are resilient to session
76 static sem_t constructor_wait
;
78 * Doing this for both the global and local sessiond.
80 static int sem_count
= { 2 };
83 * Counting nesting within lttng-ust. Used to ensure that calling fork()
84 * from liblttng-ust does not execute the pre/post fork handlers.
86 static int __thread lttng_ust_nest_count
;
89 * Info about socket and associated listener thread.
93 pthread_t ust_listener
; /* listener thread */
95 int constructor_sem_posted
;
99 char sock_path
[PATH_MAX
];
102 char wait_shm_path
[PATH_MAX
];
106 /* Socket from app (connect) to session daemon (listen) for communication */
107 struct sock_info global_apps
= {
114 .sock_path
= DEFAULT_GLOBAL_APPS_UNIX_SOCK
,
117 .wait_shm_path
= DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH
,
120 /* TODO: allow global_apps_sock_path override */
122 struct sock_info local_apps
= {
126 .allowed
= 0, /* Check setuid bit first */
131 static int wait_poll_fallback
;
133 extern void ltt_ring_buffer_client_overwrite_init(void);
134 extern void ltt_ring_buffer_client_discard_init(void);
135 extern void ltt_ring_buffer_metadata_client_init(void);
136 extern void ltt_ring_buffer_client_overwrite_exit(void);
137 extern void ltt_ring_buffer_client_discard_exit(void);
138 extern void ltt_ring_buffer_metadata_client_exit(void);
141 * Force a read (imply TLS fixup for dlopen) of TLS variables.
144 void lttng_fixup_nest_count_tls(void)
146 asm volatile ("" : : "m" (lttng_ust_nest_count
));
150 int setup_local_apps(void)
152 const char *home_dir
;
157 * Disallow per-user tracing for setuid binaries.
159 if (uid
!= geteuid()) {
160 assert(local_apps
.allowed
== 0);
163 home_dir
= (const char *) getenv("HOME");
165 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
166 assert(local_apps
.allowed
== 0);
169 local_apps
.allowed
= 1;
170 snprintf(local_apps
.sock_path
, PATH_MAX
,
171 DEFAULT_HOME_APPS_UNIX_SOCK
, home_dir
);
172 snprintf(local_apps
.wait_shm_path
, PATH_MAX
,
173 DEFAULT_HOME_APPS_WAIT_SHM_PATH
, uid
);
178 int register_app_to_sessiond(int socket
)
188 uint32_t bits_per_long
;
189 char name
[16]; /* process name */
192 reg_msg
.major
= LTTNG_UST_COMM_VERSION_MAJOR
;
193 reg_msg
.minor
= LTTNG_UST_COMM_VERSION_MINOR
;
194 reg_msg
.pid
= getpid();
195 reg_msg
.ppid
= getppid();
196 reg_msg
.uid
= getuid();
197 reg_msg
.gid
= getgid();
198 reg_msg
.bits_per_long
= CAA_BITS_PER_LONG
;
199 lttng_ust_getprocname(reg_msg
.name
);
201 ret
= ustcomm_send_unix_sock(socket
, ®_msg
, sizeof(reg_msg
));
202 if (ret
>= 0 && ret
!= sizeof(reg_msg
))
208 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
212 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
215 DBG("message successfully sent");
218 if (errno
== ECONNRESET
) {
219 printf("remote end closed connection\n");
224 printf("incorrect message size: %zd\n", len
);
230 int handle_register_done(struct sock_info
*sock_info
)
234 if (sock_info
->constructor_sem_posted
)
236 sock_info
->constructor_sem_posted
= 1;
237 if (uatomic_read(&sem_count
) <= 0) {
240 ret
= uatomic_add_return(&sem_count
, -1);
242 ret
= sem_post(&constructor_wait
);
249 int handle_message(struct sock_info
*sock_info
,
250 int sock
, struct ustcomm_ust_msg
*lum
)
253 const struct lttng_ust_objd_ops
*ops
;
254 struct ustcomm_ust_reply lur
;
261 memset(&lur
, 0, sizeof(lur
));
263 if (lttng_ust_comm_should_quit
) {
268 ops
= objd_ops(lum
->handle
);
275 case LTTNG_UST_REGISTER_DONE
:
276 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
277 ret
= handle_register_done(sock_info
);
281 case LTTNG_UST_RELEASE
:
282 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
285 ret
= lttng_ust_objd_unref(lum
->handle
);
287 case LTTNG_UST_FILTER
:
289 /* Receive filter data */
290 struct lttng_ust_filter_bytecode
*bytecode
;
292 if (lum
->u
.filter
.data_size
> FILTER_BYTECODE_MAX_LEN
) {
293 ERR("Filter data size is too large: %u bytes\n",
294 lum
->u
.filter
.data_size
);
299 if (lum
->u
.filter
.reloc_offset
> lum
->u
.filter
.data_size
) {
300 ERR("Filter reloc offset %u is not within data\n",
301 lum
->u
.filter
.reloc_offset
);
306 bytecode
= zmalloc(sizeof(*bytecode
) + lum
->u
.filter
.data_size
);
311 len
= ustcomm_recv_unix_sock(sock
, bytecode
->data
,
312 lum
->u
.filter
.data_size
);
314 case 0: /* orderly shutdown */
319 DBG("Receive failed from lttng-sessiond with errno %d", errno
);
320 if (errno
== ECONNRESET
) {
321 ERR("%s remote end closed connection\n", sock_info
->name
);
329 if (len
== lum
->u
.filter
.data_size
) {
330 DBG("filter data received\n");
333 ERR("incorrect filter data message size: %zd\n", len
);
339 bytecode
->len
= lum
->u
.filter
.data_size
;
340 bytecode
->reloc_offset
= lum
->u
.filter
.reloc_offset
;
342 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
343 (unsigned long) bytecode
,
348 /* don't free bytecode if everything went fine. */
357 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
358 (unsigned long) &lum
->u
,
366 lur
.handle
= lum
->handle
;
370 lur
.ret_code
= USTCOMM_OK
;
372 //lur.ret_code = USTCOMM_SESSION_FAIL;
377 case LTTNG_UST_STREAM
:
379 * Special-case reply to send stream info.
382 lur
.u
.stream
.memory_map_size
= *args
.stream
.memory_map_size
;
383 shm_fd
= *args
.stream
.shm_fd
;
384 wait_fd
= *args
.stream
.wait_fd
;
386 case LTTNG_UST_METADATA
:
387 case LTTNG_UST_CHANNEL
:
388 lur
.u
.channel
.memory_map_size
= *args
.channel
.memory_map_size
;
389 shm_fd
= *args
.channel
.shm_fd
;
390 wait_fd
= *args
.channel
.wait_fd
;
392 case LTTNG_UST_TRACER_VERSION
:
393 lur
.u
.version
= lum
->u
.version
;
395 case LTTNG_UST_TRACEPOINT_LIST_GET
:
396 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
400 ret
= send_reply(sock
, &lur
);
402 perror("error sending reply");
406 if ((lum
->cmd
== LTTNG_UST_STREAM
407 || lum
->cmd
== LTTNG_UST_CHANNEL
408 || lum
->cmd
== LTTNG_UST_METADATA
)
409 && lur
.ret_code
== USTCOMM_OK
) {
412 /* we also need to send the file descriptors. */
413 ret
= ustcomm_send_fds_unix_sock(sock
,
417 perror("send shm_fd");
421 * The sessiond expects 2 file descriptors, even upon
424 ret
= ustcomm_send_fds_unix_sock(sock
,
428 perror("send wait_fd");
437 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
440 if (lur
.ret_code
== USTCOMM_OK
) {
442 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
443 len
= ustcomm_send_unix_sock(sock
,
444 &args
.field_list
.entry
,
445 sizeof(args
.field_list
.entry
));
446 if (len
!= sizeof(args
.field_list
.entry
)) {
453 * We still have the memory map reference, and the fds have been
454 * sent to the sessiond. We can therefore close those fds. Note
455 * that we keep the write side of the wait_fd open, but close
458 if (lur
.ret_code
== USTCOMM_OK
) {
460 case LTTNG_UST_STREAM
:
464 PERROR("Error closing stream shm_fd");
466 *args
.stream
.shm_fd
= -1;
469 ret
= close(wait_fd
);
471 PERROR("Error closing stream wait_fd");
473 *args
.stream
.wait_fd
= -1;
476 case LTTNG_UST_METADATA
:
477 case LTTNG_UST_CHANNEL
:
481 PERROR("Error closing channel shm_fd");
483 *args
.channel
.shm_fd
= -1;
486 ret
= close(wait_fd
);
488 PERROR("Error closing channel wait_fd");
490 *args
.channel
.wait_fd
= -1;
502 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
506 if (sock_info
->socket
!= -1) {
507 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
509 ERR("Error closing apps socket");
511 sock_info
->socket
= -1;
513 if (sock_info
->root_handle
!= -1) {
514 ret
= lttng_ust_objd_unref(sock_info
->root_handle
);
516 ERR("Error unref root handle");
518 sock_info
->root_handle
= -1;
520 sock_info
->constructor_sem_posted
= 0;
522 * wait_shm_mmap is used by listener threads outside of the
523 * ust lock, so we cannot tear it down ourselves, because we
524 * cannot join on these threads. Leave this task to the OS
527 if (!exiting
&& sock_info
->wait_shm_mmap
) {
528 ret
= munmap(sock_info
->wait_shm_mmap
, sysconf(_SC_PAGE_SIZE
));
530 ERR("Error unmapping wait shm");
532 sock_info
->wait_shm_mmap
= NULL
;
537 * Using fork to set umask in the child process (not multi-thread safe).
538 * We deal with the shm_open vs ftruncate race (happening when the
539 * sessiond owns the shm and does not let everybody modify it, to ensure
540 * safety against shm_unlink) by simply letting the mmap fail and
541 * retrying after a few seconds.
542 * For global shm, everybody has rw access to it until the sessiond
546 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
548 int wait_shm_fd
, ret
;
552 * Try to open read-only.
554 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
555 if (wait_shm_fd
>= 0) {
557 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
559 * Real-only open did not work, and it's not because the
560 * entry was not present. It's a failure that prohibits
563 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
567 * If the open failed because the file did not exist, try
568 * creating it ourself.
570 lttng_ust_nest_count
++;
572 lttng_ust_nest_count
--;
577 * Parent: wait for child to return, in which case the
578 * shared memory map will have been created.
581 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
586 * Try to open read-only again after creation.
588 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
589 if (wait_shm_fd
< 0) {
591 * Real-only open did not work. It's a failure
592 * that prohibits using shm.
594 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
598 } else if (pid
== 0) {
602 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
603 if (sock_info
->global
)
604 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
606 * We're alone in a child process, so we can modify the
607 * process-wide umask.
611 * Try creating shm (or get rw access).
612 * We don't do an exclusive open, because we allow other
613 * processes to create+ftruncate it concurrently.
615 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
616 O_RDWR
| O_CREAT
, create_mode
);
617 if (wait_shm_fd
>= 0) {
618 ret
= ftruncate(wait_shm_fd
, mmap_size
);
626 * For local shm, we need to have rw access to accept
627 * opening it: this means the local sessiond will be
628 * able to wake us up. For global shm, we open it even
629 * if rw access is not granted, because the root.root
630 * sessiond will be able to override all rights and wake
633 if (!sock_info
->global
&& errno
!= EACCES
) {
634 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
638 * The shm exists, but we cannot open it RW. Report
646 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
650 * Ensure that our user is the owner of the shm file for
651 * local shm. If we do not own the file, it means our
652 * sessiond will not have access to wake us up (there is
653 * probably a rogue process trying to fake our
654 * sessiond). Fallback to polling method in this case.
656 ret
= fstat(wait_shm_fd
, &statbuf
);
661 if (statbuf
.st_uid
!= getuid())
667 ret
= close(wait_shm_fd
);
669 PERROR("Error closing fd");
675 char *get_map_shm(struct sock_info
*sock_info
)
677 size_t mmap_size
= sysconf(_SC_PAGE_SIZE
);
678 int wait_shm_fd
, ret
;
681 wait_shm_fd
= get_wait_shm(sock_info
, mmap_size
);
682 if (wait_shm_fd
< 0) {
685 wait_shm_mmap
= mmap(NULL
, mmap_size
, PROT_READ
,
686 MAP_SHARED
, wait_shm_fd
, 0);
687 /* close shm fd immediately after taking the mmap reference */
688 ret
= close(wait_shm_fd
);
690 PERROR("Error closing fd");
692 if (wait_shm_mmap
== MAP_FAILED
) {
693 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
696 return wait_shm_mmap
;
703 void wait_for_sessiond(struct sock_info
*sock_info
)
708 if (lttng_ust_comm_should_quit
) {
711 if (wait_poll_fallback
) {
714 if (!sock_info
->wait_shm_mmap
) {
715 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
716 if (!sock_info
->wait_shm_mmap
)
721 DBG("Waiting for %s apps sessiond", sock_info
->name
);
722 /* Wait for futex wakeup */
723 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
724 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
725 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
727 if (errno
== EFAULT
) {
728 wait_poll_fallback
= 1;
730 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
731 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
732 "Please upgrade your kernel "
733 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
734 "mainline). LTTng-UST will use polling mode fallback.");
752 * This thread does not allocate any resource, except within
753 * handle_message, within mutex protection. This mutex protects against
755 * The other moment it allocates resources is at socket connection, which
756 * is also protected by the mutex.
759 void *ust_listener_thread(void *arg
)
761 struct sock_info
*sock_info
= arg
;
762 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
764 /* Restart trying to connect to the session daemon */
766 if (prev_connect_failed
) {
767 /* Wait for sessiond availability with pipe */
768 wait_for_sessiond(sock_info
);
772 * Sleep for 5 seconds before retrying after a
773 * sequence of failure / wait / failure. This
774 * deals with a killed or broken session daemon.
779 prev_connect_failed
= 0;
783 if (lttng_ust_comm_should_quit
) {
788 if (sock_info
->socket
!= -1) {
789 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
791 ERR("Error closing %s apps socket", sock_info
->name
);
793 sock_info
->socket
= -1;
797 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
799 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
800 prev_connect_failed
= 1;
802 * If we cannot find the sessiond daemon, don't delay
803 * constructor execution.
805 ret
= handle_register_done(sock_info
);
811 sock_info
->socket
= sock
= ret
;
814 * Create only one root handle per listener thread for the whole
817 if (sock_info
->root_handle
== -1) {
818 ret
= lttng_abi_create_root_handle();
820 ERR("Error creating root handle");
824 sock_info
->root_handle
= ret
;
827 ret
= register_app_to_sessiond(sock
);
829 ERR("Error registering to %s apps socket", sock_info
->name
);
830 prev_connect_failed
= 1;
832 * If we cannot register to the sessiond daemon, don't
833 * delay constructor execution.
835 ret
= handle_register_done(sock_info
);
844 struct ustcomm_ust_msg lum
;
846 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
848 case 0: /* orderly shutdown */
849 DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info
->name
);
852 * Either sessiond has shutdown or refused us by closing the socket.
853 * In either case, we don't want to delay construction execution,
854 * and we need to wait before retry.
856 prev_connect_failed
= 1;
858 * If we cannot register to the sessiond daemon, don't
859 * delay constructor execution.
861 ret
= handle_register_done(sock_info
);
866 DBG("message received\n");
867 ret
= handle_message(sock_info
, sock
, &lum
);
869 ERR("Error handling message for %s socket", sock_info
->name
);
873 DBG("Receive failed from lttng-sessiond with errno %d", errno
);
874 if (errno
== ECONNRESET
) {
875 ERR("%s remote end closed connection\n", sock_info
->name
);
880 ERR("incorrect message size (%s socket): %zd\n", sock_info
->name
, len
);
886 goto restart
; /* try to reconnect */
892 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
895 int get_timeout(struct timespec
*constructor_timeout
)
897 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
901 str_delay
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
903 constructor_delay_ms
= strtol(str_delay
, NULL
, 10);
906 switch (constructor_delay_ms
) {
907 case -1:/* fall-through */
909 return constructor_delay_ms
;
915 * If we are unable to find the current time, don't wait.
917 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
921 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
922 constructor_timeout
->tv_nsec
+=
923 (constructor_delay_ms
% 1000UL) * 1000000UL;
924 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
925 constructor_timeout
->tv_sec
++;
926 constructor_timeout
->tv_nsec
-= 1000000000UL;
932 * sessiond monitoring thread: monitor presence of global and per-user
933 * sessiond by polling the application common named pipe.
937 void __attribute__((constructor
)) lttng_ust_init(void)
939 struct timespec constructor_timeout
;
940 sigset_t sig_all_blocked
, orig_parent_mask
;
941 pthread_attr_t thread_attr
;
945 if (uatomic_xchg(&initialized
, 1) == 1)
949 * Fixup interdependency between TLS fixup mutex (which happens
950 * to be the dynamic linker mutex) and ust_lock, taken within
953 lttng_fixup_event_tls();
954 lttng_fixup_ringbuffer_tls();
955 lttng_fixup_vtid_tls();
956 lttng_fixup_nest_count_tls();
957 lttng_fixup_procname_tls();
960 * We want precise control over the order in which we construct
961 * our sub-libraries vs starting to receive commands from
962 * sessiond (otherwise leading to errors when trying to create
963 * sessiond before the init functions are completed).
967 ltt_ring_buffer_metadata_client_init();
968 ltt_ring_buffer_client_overwrite_init();
969 ltt_ring_buffer_client_discard_init();
971 timeout_mode
= get_timeout(&constructor_timeout
);
973 ret
= sem_init(&constructor_wait
, 0, 0);
976 ret
= setup_local_apps();
978 DBG("local apps setup returned %d", ret
);
981 /* A new thread created by pthread_create inherits the signal mask
982 * from the parent. To avoid any signal being received by the
983 * listener thread, we block all signals temporarily in the parent,
984 * while we create the listener thread.
986 sigfillset(&sig_all_blocked
);
987 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
989 ERR("pthread_sigmask: %s", strerror(ret
));
992 ret
= pthread_attr_init(&thread_attr
);
994 ERR("pthread_attr_init: %s", strerror(ret
));
996 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
998 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1001 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1002 ust_listener_thread
, &global_apps
);
1004 ERR("pthread_create global: %s", strerror(ret
));
1006 if (local_apps
.allowed
) {
1007 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1008 ust_listener_thread
, &local_apps
);
1010 ERR("pthread_create local: %s", strerror(ret
));
1013 handle_register_done(&local_apps
);
1015 ret
= pthread_attr_destroy(&thread_attr
);
1017 ERR("pthread_attr_destroy: %s", strerror(ret
));
1020 /* Restore original signal mask in parent */
1021 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1023 ERR("pthread_sigmask: %s", strerror(ret
));
1026 switch (timeout_mode
) {
1027 case 1: /* timeout wait */
1029 ret
= sem_timedwait(&constructor_wait
,
1030 &constructor_timeout
);
1031 } while (ret
< 0 && errno
== EINTR
);
1032 if (ret
< 0 && errno
== ETIMEDOUT
) {
1033 ERR("Timed out waiting for ltt-sessiond");
1038 case -1:/* wait forever */
1040 ret
= sem_wait(&constructor_wait
);
1041 } while (ret
< 0 && errno
== EINTR
);
1044 case 0: /* no timeout */
1050 void lttng_ust_cleanup(int exiting
)
1052 cleanup_sock_info(&global_apps
, exiting
);
1053 if (local_apps
.allowed
) {
1054 cleanup_sock_info(&local_apps
, exiting
);
1057 * The teardown in this function all affect data structures
1058 * accessed under the UST lock by the listener thread. This
1059 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1060 * that none of these threads are accessing this data at this
1063 lttng_ust_abi_exit();
1064 lttng_ust_events_exit();
1065 ltt_ring_buffer_client_discard_exit();
1066 ltt_ring_buffer_client_overwrite_exit();
1067 ltt_ring_buffer_metadata_client_exit();
1070 /* Reinitialize values for fork */
1072 lttng_ust_comm_should_quit
= 0;
1077 void __attribute__((destructor
)) lttng_ust_exit(void)
1082 * Using pthread_cancel here because:
1083 * A) we don't want to hang application teardown.
1084 * B) the thread is not allocating any resource.
1088 * Require the communication thread to quit. Synchronize with
1089 * mutexes to ensure it is not in a mutex critical section when
1090 * pthread_cancel is later called.
1093 lttng_ust_comm_should_quit
= 1;
1096 /* cancel threads */
1097 ret
= pthread_cancel(global_apps
.ust_listener
);
1099 ERR("Error cancelling global ust listener thread: %s",
1102 if (local_apps
.allowed
) {
1103 ret
= pthread_cancel(local_apps
.ust_listener
);
1105 ERR("Error cancelling local ust listener thread: %s",
1110 * Do NOT join threads: use of sys_futex makes it impossible to
1111 * join the threads without using async-cancel, but async-cancel
1112 * is delivered by a signal, which could hit the target thread
1113 * anywhere in its code path, including while the ust_lock() is
1114 * held, causing a deadlock for the other thread. Let the OS
1115 * cleanup the threads if there are stalled in a syscall.
1117 lttng_ust_cleanup(1);
1121 * We exclude the worker threads across fork and clone (except
1122 * CLONE_VM), because these system calls only keep the forking thread
1123 * running in the child. Therefore, we don't want to call fork or clone
1124 * in the middle of an tracepoint or ust tracing state modification.
1125 * Holding this mutex protects these structures across fork and clone.
1127 void ust_before_fork(sigset_t
*save_sigset
)
1130 * Disable signals. This is to avoid that the child intervenes
1131 * before it is properly setup for tracing. It is safer to
1132 * disable all signals, because then we know we are not breaking
1133 * anything by restoring the original mask.
1138 if (lttng_ust_nest_count
)
1140 /* Disable signals */
1141 sigfillset(&all_sigs
);
1142 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1144 PERROR("sigprocmask");
1147 rcu_bp_before_fork();
1150 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1154 DBG("process %d", getpid());
1156 /* Restore signals */
1157 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1159 PERROR("sigprocmask");
1163 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1165 if (lttng_ust_nest_count
)
1167 DBG("process %d", getpid());
1168 rcu_bp_after_fork_parent();
1169 /* Release mutexes and reenable signals */
1170 ust_after_fork_common(restore_sigset
);
1174 * After fork, in the child, we need to cleanup all the leftover state,
1175 * except the worker thread which already magically disappeared thanks
1176 * to the weird Linux fork semantics. After tyding up, we call
1177 * lttng_ust_init() again to start over as a new PID.
1179 * This is meant for forks() that have tracing in the child between the
1180 * fork and following exec call (if there is any).
1182 void ust_after_fork_child(sigset_t
*restore_sigset
)
1184 if (lttng_ust_nest_count
)
1186 DBG("process %d", getpid());
1187 /* Release urcu mutexes */
1188 rcu_bp_after_fork_child();
1189 lttng_ust_cleanup(0);
1190 lttng_context_vtid_reset();
1191 /* Release mutexes and reenable signals */
1192 ust_after_fork_common(restore_sigset
);