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>
45 #include <lttng/ust-ctl.h>
46 #include <urcu/tls-compat.h>
48 #include <usterr-signal-safe.h>
50 #include "tracepoint-internal.h"
51 #include "lttng-tracer-core.h"
53 #include "../libringbuffer/tlsfixup.h"
56 * Has lttng ust comm constructor been called ?
58 static int initialized
;
61 * The ust_lock/ust_unlock lock is used as a communication thread mutex.
62 * Held when handling a command, also held by fork() to deal with
63 * removal of threads, and by exit path.
66 /* Should the ust comm thread quit ? */
67 static int lttng_ust_comm_should_quit
;
70 * Wait for either of these before continuing to the main
72 * - the register_done message from sessiond daemon
73 * (will let the sessiond daemon enable sessions before main
75 * - sessiond daemon is not reachable.
76 * - timeout (ensuring applications are resilient to session
79 static sem_t constructor_wait
;
81 * Doing this for both the global and local sessiond.
83 static int sem_count
= { 2 };
86 * Counting nesting within lttng-ust. Used to ensure that calling fork()
87 * from liblttng-ust does not execute the pre/post fork handlers.
89 static DEFINE_URCU_TLS(int, lttng_ust_nest_count
);
92 * Info about socket and associated listener thread.
96 pthread_t ust_listener
; /* listener thread */
98 int constructor_sem_posted
;
103 char sock_path
[PATH_MAX
];
106 char wait_shm_path
[PATH_MAX
];
110 /* Socket from app (connect) to session daemon (listen) for communication */
111 struct sock_info global_apps
= {
119 .sock_path
= DEFAULT_GLOBAL_APPS_UNIX_SOCK
,
122 .wait_shm_path
= DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH
,
125 /* TODO: allow global_apps_sock_path override */
127 struct sock_info local_apps
= {
131 .allowed
= 0, /* Check setuid bit first */
137 static int wait_poll_fallback
;
139 static const char *cmd_name_mapping
[] = {
140 [ LTTNG_UST_RELEASE
] = "Release",
141 [ LTTNG_UST_SESSION
] = "Create Session",
142 [ LTTNG_UST_TRACER_VERSION
] = "Get Tracer Version",
144 [ LTTNG_UST_TRACEPOINT_LIST
] = "Create Tracepoint List",
145 [ LTTNG_UST_WAIT_QUIESCENT
] = "Wait for Quiescent State",
146 [ LTTNG_UST_REGISTER_DONE
] = "Registration Done",
147 [ LTTNG_UST_TRACEPOINT_FIELD_LIST
] = "Create Tracepoint Field List",
149 /* Session FD commands */
150 [ LTTNG_UST_CHANNEL
] = "Create Channel",
151 [ LTTNG_UST_SESSION_START
] = "Start Session",
152 [ LTTNG_UST_SESSION_STOP
] = "Stop Session",
154 /* Channel FD commands */
155 [ LTTNG_UST_STREAM
] = "Create Stream",
156 [ LTTNG_UST_EVENT
] = "Create Event",
158 /* Event and Channel FD commands */
159 [ LTTNG_UST_CONTEXT
] = "Create Context",
160 [ LTTNG_UST_FLUSH_BUFFER
] = "Flush Buffer",
162 /* Event, Channel and Session commands */
163 [ LTTNG_UST_ENABLE
] = "Enable",
164 [ LTTNG_UST_DISABLE
] = "Disable",
166 /* Tracepoint list commands */
167 [ LTTNG_UST_TRACEPOINT_LIST_GET
] = "List Next Tracepoint",
168 [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
] = "List Next Tracepoint Field",
170 /* Event FD commands */
171 [ LTTNG_UST_FILTER
] = "Create Filter",
174 extern void lttng_ring_buffer_client_overwrite_init(void);
175 extern void lttng_ring_buffer_client_discard_init(void);
176 extern void lttng_ring_buffer_metadata_client_init(void);
177 extern void lttng_ring_buffer_client_overwrite_exit(void);
178 extern void lttng_ring_buffer_client_discard_exit(void);
179 extern void lttng_ring_buffer_metadata_client_exit(void);
182 * Force a read (imply TLS fixup for dlopen) of TLS variables.
185 void lttng_fixup_nest_count_tls(void)
187 asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count
)));
191 void print_cmd(int cmd
, int handle
)
193 const char *cmd_name
= "Unknown";
195 if (cmd_name_mapping
[cmd
]) {
196 cmd_name
= cmd_name_mapping
[cmd
];
198 DBG("Message Received \"%s\", Handle \"%s\" (%d)", cmd_name
,
199 lttng_ust_obj_get_name(handle
), handle
);
203 int setup_local_apps(void)
205 const char *home_dir
;
210 * Disallow per-user tracing for setuid binaries.
212 if (uid
!= geteuid()) {
213 assert(local_apps
.allowed
== 0);
216 home_dir
= (const char *) getenv("HOME");
218 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
219 assert(local_apps
.allowed
== 0);
222 local_apps
.allowed
= 1;
223 snprintf(local_apps
.sock_path
, PATH_MAX
,
224 DEFAULT_HOME_APPS_UNIX_SOCK
, home_dir
);
225 snprintf(local_apps
.wait_shm_path
, PATH_MAX
,
226 DEFAULT_HOME_APPS_WAIT_SHM_PATH
, uid
);
231 int register_app_to_sessiond(int socket
)
241 uint32_t bits_per_long
;
242 char name
[16]; /* process name */
245 reg_msg
.major
= LTTNG_UST_COMM_VERSION_MAJOR
;
246 reg_msg
.minor
= LTTNG_UST_COMM_VERSION_MINOR
;
247 reg_msg
.pid
= getpid();
248 reg_msg
.ppid
= getppid();
249 reg_msg
.uid
= getuid();
250 reg_msg
.gid
= getgid();
251 reg_msg
.bits_per_long
= CAA_BITS_PER_LONG
;
252 lttng_ust_getprocname(reg_msg
.name
);
254 ret
= ustcomm_send_unix_sock(socket
, ®_msg
, sizeof(reg_msg
));
255 if (ret
>= 0 && ret
!= sizeof(reg_msg
))
261 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
265 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
268 DBG("message successfully sent");
271 if (len
== -ECONNRESET
) {
272 DBG("remote end closed connection");
277 DBG("incorrect message size: %zd", len
);
283 int handle_register_done(struct sock_info
*sock_info
)
287 if (sock_info
->constructor_sem_posted
)
289 sock_info
->constructor_sem_posted
= 1;
290 if (uatomic_read(&sem_count
) <= 0) {
293 ret
= uatomic_add_return(&sem_count
, -1);
295 ret
= sem_post(&constructor_wait
);
302 int handle_message(struct sock_info
*sock_info
,
303 int sock
, struct ustcomm_ust_msg
*lum
)
306 const struct lttng_ust_objd_ops
*ops
;
307 struct ustcomm_ust_reply lur
;
313 memset(&lur
, 0, sizeof(lur
));
315 if (lttng_ust_comm_should_quit
) {
316 ret
= -LTTNG_UST_ERR_EXITING
;
320 ops
= objd_ops(lum
->handle
);
327 case LTTNG_UST_REGISTER_DONE
:
328 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
329 ret
= handle_register_done(sock_info
);
333 case LTTNG_UST_RELEASE
:
334 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
337 ret
= lttng_ust_objd_unref(lum
->handle
);
339 case LTTNG_UST_FILTER
:
341 /* Receive filter data */
342 struct lttng_ust_filter_bytecode_node
*bytecode
;
344 if (lum
->u
.filter
.data_size
> FILTER_BYTECODE_MAX_LEN
) {
345 ERR("Filter data size is too large: %u bytes",
346 lum
->u
.filter
.data_size
);
351 if (lum
->u
.filter
.reloc_offset
> lum
->u
.filter
.data_size
) {
352 ERR("Filter reloc offset %u is not within data",
353 lum
->u
.filter
.reloc_offset
);
358 bytecode
= zmalloc(sizeof(*bytecode
) + lum
->u
.filter
.data_size
);
363 len
= ustcomm_recv_unix_sock(sock
, bytecode
->bc
.data
,
364 lum
->u
.filter
.data_size
);
366 case 0: /* orderly shutdown */
371 if (len
== lum
->u
.filter
.data_size
) {
372 DBG("filter data received");
374 } else if (len
< 0) {
375 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
376 if (len
== -ECONNRESET
) {
377 ERR("%s remote end closed connection", sock_info
->name
);
385 DBG("incorrect filter data message size: %zd", len
);
391 bytecode
->bc
.len
= lum
->u
.filter
.data_size
;
392 bytecode
->bc
.reloc_offset
= lum
->u
.filter
.reloc_offset
;
393 bytecode
->bc
.seqnum
= lum
->u
.filter
.seqnum
;
395 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
396 (unsigned long) bytecode
,
401 /* don't free bytecode if everything went fine. */
408 case LTTNG_UST_CHANNEL
:
412 len
= ustcomm_recv_channel_from_sessiond(sock
,
413 &chan_data
, lum
->u
.channel
.len
);
415 case 0: /* orderly shutdown */
419 if (len
== lum
->u
.channel
.len
) {
420 DBG("channel data received");
422 } else if (len
< 0) {
423 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
424 if (len
== -ECONNRESET
) {
425 ERR("%s remote end closed connection", sock_info
->name
);
432 DBG("incorrect channel data message size: %zd", len
);
437 args
.channel
.chan_data
= chan_data
;
439 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
440 (unsigned long) &lum
->u
,
446 case LTTNG_UST_STREAM
:
448 /* Receive shm_fd, wakeup_fd */
449 ret
= ustcomm_recv_stream_from_sessiond(sock
,
452 &args
.stream
.wakeup_fd
);
457 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
458 (unsigned long) &lum
->u
,
466 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
467 (unsigned long) &lum
->u
,
475 lur
.handle
= lum
->handle
;
479 lur
.ret_code
= LTTNG_UST_OK
;
482 * Use -LTTNG_UST_ERR as wildcard for UST internal
483 * error that are not caused by the transport, except if
484 * we already have a more precise error message to
487 if (ret
> -LTTNG_UST_ERR
) {
488 /* Translate code to UST error. */
491 lur
.ret_code
= -LTTNG_UST_ERR_EXIST
;
494 lur
.ret_code
= -LTTNG_UST_ERR_INVAL
;
497 lur
.ret_code
= -LTTNG_UST_ERR_NOENT
;
500 lur
.ret_code
= -LTTNG_UST_ERR_PERM
;
503 lur
.ret_code
= -LTTNG_UST_ERR_NOSYS
;
506 lur
.ret_code
= -LTTNG_UST_ERR
;
515 case LTTNG_UST_TRACER_VERSION
:
516 lur
.u
.version
= lum
->u
.version
;
518 case LTTNG_UST_TRACEPOINT_LIST_GET
:
519 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
523 DBG("Return value: %d", lur
.ret_val
);
524 ret
= send_reply(sock
, &lur
);
526 DBG("error sending reply");
531 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
534 if (lur
.ret_code
== LTTNG_UST_OK
) {
536 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
537 len
= ustcomm_send_unix_sock(sock
,
538 &args
.field_list
.entry
,
539 sizeof(args
.field_list
.entry
));
544 if (len
!= sizeof(args
.field_list
.entry
)) {
557 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
561 if (sock_info
->socket
!= -1) {
562 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
564 ERR("Error closing apps socket");
566 sock_info
->socket
= -1;
568 if (sock_info
->root_handle
!= -1) {
569 ret
= lttng_ust_objd_unref(sock_info
->root_handle
);
571 ERR("Error unref root handle");
573 sock_info
->root_handle
= -1;
575 sock_info
->constructor_sem_posted
= 0;
577 * wait_shm_mmap is used by listener threads outside of the
578 * ust lock, so we cannot tear it down ourselves, because we
579 * cannot join on these threads. Leave this task to the OS
582 if (!exiting
&& sock_info
->wait_shm_mmap
) {
583 ret
= munmap(sock_info
->wait_shm_mmap
, sysconf(_SC_PAGE_SIZE
));
585 ERR("Error unmapping wait shm");
587 sock_info
->wait_shm_mmap
= NULL
;
592 * Using fork to set umask in the child process (not multi-thread safe).
593 * We deal with the shm_open vs ftruncate race (happening when the
594 * sessiond owns the shm and does not let everybody modify it, to ensure
595 * safety against shm_unlink) by simply letting the mmap fail and
596 * retrying after a few seconds.
597 * For global shm, everybody has rw access to it until the sessiond
601 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
603 int wait_shm_fd
, ret
;
607 * Try to open read-only.
609 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
610 if (wait_shm_fd
>= 0) {
612 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
614 * Real-only open did not work, and it's not because the
615 * entry was not present. It's a failure that prohibits
618 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
622 * If the open failed because the file did not exist, try
623 * creating it ourself.
625 URCU_TLS(lttng_ust_nest_count
)++;
627 URCU_TLS(lttng_ust_nest_count
)--;
632 * Parent: wait for child to return, in which case the
633 * shared memory map will have been created.
636 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
641 * Try to open read-only again after creation.
643 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
644 if (wait_shm_fd
< 0) {
646 * Real-only open did not work. It's a failure
647 * that prohibits using shm.
649 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
653 } else if (pid
== 0) {
657 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
658 if (sock_info
->global
)
659 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
661 * We're alone in a child process, so we can modify the
662 * process-wide umask.
666 * Try creating shm (or get rw access).
667 * We don't do an exclusive open, because we allow other
668 * processes to create+ftruncate it concurrently.
670 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
671 O_RDWR
| O_CREAT
, create_mode
);
672 if (wait_shm_fd
>= 0) {
673 ret
= ftruncate(wait_shm_fd
, mmap_size
);
681 * For local shm, we need to have rw access to accept
682 * opening it: this means the local sessiond will be
683 * able to wake us up. For global shm, we open it even
684 * if rw access is not granted, because the root.root
685 * sessiond will be able to override all rights and wake
688 if (!sock_info
->global
&& errno
!= EACCES
) {
689 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
693 * The shm exists, but we cannot open it RW. Report
701 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
705 * Ensure that our user is the owner of the shm file for
706 * local shm. If we do not own the file, it means our
707 * sessiond will not have access to wake us up (there is
708 * probably a rogue process trying to fake our
709 * sessiond). Fallback to polling method in this case.
711 ret
= fstat(wait_shm_fd
, &statbuf
);
716 if (statbuf
.st_uid
!= getuid())
722 ret
= close(wait_shm_fd
);
724 PERROR("Error closing fd");
730 char *get_map_shm(struct sock_info
*sock_info
)
732 size_t mmap_size
= sysconf(_SC_PAGE_SIZE
);
733 int wait_shm_fd
, ret
;
736 wait_shm_fd
= get_wait_shm(sock_info
, mmap_size
);
737 if (wait_shm_fd
< 0) {
740 wait_shm_mmap
= mmap(NULL
, mmap_size
, PROT_READ
,
741 MAP_SHARED
, wait_shm_fd
, 0);
742 /* close shm fd immediately after taking the mmap reference */
743 ret
= close(wait_shm_fd
);
745 PERROR("Error closing fd");
747 if (wait_shm_mmap
== MAP_FAILED
) {
748 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
751 return wait_shm_mmap
;
758 void wait_for_sessiond(struct sock_info
*sock_info
)
763 if (lttng_ust_comm_should_quit
) {
766 if (wait_poll_fallback
) {
769 if (!sock_info
->wait_shm_mmap
) {
770 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
771 if (!sock_info
->wait_shm_mmap
)
776 DBG("Waiting for %s apps sessiond", sock_info
->name
);
777 /* Wait for futex wakeup */
778 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
779 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
780 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
782 if (errno
== EFAULT
) {
783 wait_poll_fallback
= 1;
785 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
786 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
787 "Please upgrade your kernel "
788 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
789 "mainline). LTTng-UST will use polling mode fallback.");
807 * This thread does not allocate any resource, except within
808 * handle_message, within mutex protection. This mutex protects against
810 * The other moment it allocates resources is at socket connection, which
811 * is also protected by the mutex.
814 void *ust_listener_thread(void *arg
)
816 struct sock_info
*sock_info
= arg
;
817 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
819 /* Restart trying to connect to the session daemon */
821 if (prev_connect_failed
) {
822 /* Wait for sessiond availability with pipe */
823 wait_for_sessiond(sock_info
);
827 * Sleep for 5 seconds before retrying after a
828 * sequence of failure / wait / failure. This
829 * deals with a killed or broken session daemon.
834 prev_connect_failed
= 0;
838 if (lttng_ust_comm_should_quit
) {
842 if (sock_info
->socket
!= -1) {
843 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
845 ERR("Error closing %s apps socket", sock_info
->name
);
847 sock_info
->socket
= -1;
851 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
853 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
854 prev_connect_failed
= 1;
856 * If we cannot find the sessiond daemon, don't delay
857 * constructor execution.
859 ret
= handle_register_done(sock_info
);
865 sock_info
->socket
= sock
= ret
;
868 * Create only one root handle per listener thread for the whole
869 * process lifetime, so we ensure we get ID which is statically
870 * assigned to the root handle.
872 if (sock_info
->root_handle
== -1) {
873 ret
= lttng_abi_create_root_handle();
875 ERR("Error creating root handle");
878 sock_info
->root_handle
= ret
;
881 ret
= register_app_to_sessiond(sock
);
883 ERR("Error registering to %s apps socket", sock_info
->name
);
884 prev_connect_failed
= 1;
886 * If we cannot register to the sessiond daemon, don't
887 * delay constructor execution.
889 ret
= handle_register_done(sock_info
);
898 struct ustcomm_ust_msg lum
;
900 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
902 case 0: /* orderly shutdown */
903 DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info
->name
);
905 if (lttng_ust_comm_should_quit
) {
909 * Either sessiond has shutdown or refused us by closing the socket.
910 * In either case, we don't want to delay construction execution,
911 * and we need to wait before retry.
913 prev_connect_failed
= 1;
915 * If we cannot register to the sessiond daemon, don't
916 * delay constructor execution.
918 ret
= handle_register_done(sock_info
);
923 print_cmd(lum
.cmd
, lum
.handle
);
924 ret
= handle_message(sock_info
, sock
, &lum
);
926 ERR("Error handling message for %s socket", sock_info
->name
);
931 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
933 DBG("incorrect message size (%s socket): %zd", sock_info
->name
, len
);
935 if (len
== -ECONNRESET
) {
936 DBG("%s remote end closed connection", sock_info
->name
);
945 if (lttng_ust_comm_should_quit
) {
948 /* Cleanup socket handles before trying to reconnect */
949 lttng_ust_objd_table_owner_cleanup(sock_info
);
951 goto restart
; /* try to reconnect */
954 sock_info
->thread_active
= 0;
960 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
963 int get_timeout(struct timespec
*constructor_timeout
)
965 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
969 str_delay
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
971 constructor_delay_ms
= strtol(str_delay
, NULL
, 10);
974 switch (constructor_delay_ms
) {
975 case -1:/* fall-through */
977 return constructor_delay_ms
;
983 * If we are unable to find the current time, don't wait.
985 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
989 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
990 constructor_timeout
->tv_nsec
+=
991 (constructor_delay_ms
% 1000UL) * 1000000UL;
992 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
993 constructor_timeout
->tv_sec
++;
994 constructor_timeout
->tv_nsec
-= 1000000000UL;
1000 * sessiond monitoring thread: monitor presence of global and per-user
1001 * sessiond by polling the application common named pipe.
1003 void __attribute__((constructor
)) lttng_ust_init(void)
1005 struct timespec constructor_timeout
;
1006 sigset_t sig_all_blocked
, orig_parent_mask
;
1007 pthread_attr_t thread_attr
;
1011 if (uatomic_xchg(&initialized
, 1) == 1)
1015 * Fixup interdependency between TLS fixup mutex (which happens
1016 * to be the dynamic linker mutex) and ust_lock, taken within
1019 lttng_fixup_ringbuffer_tls();
1020 lttng_fixup_vtid_tls();
1021 lttng_fixup_nest_count_tls();
1022 lttng_fixup_procname_tls();
1025 * We want precise control over the order in which we construct
1026 * our sub-libraries vs starting to receive commands from
1027 * sessiond (otherwise leading to errors when trying to create
1028 * sessiond before the init functions are completed).
1032 lttng_ring_buffer_metadata_client_init();
1033 lttng_ring_buffer_client_overwrite_init();
1034 lttng_ring_buffer_client_discard_init();
1036 timeout_mode
= get_timeout(&constructor_timeout
);
1038 ret
= sem_init(&constructor_wait
, 0, 0);
1041 ret
= setup_local_apps();
1043 DBG("local apps setup returned %d", ret
);
1046 /* A new thread created by pthread_create inherits the signal mask
1047 * from the parent. To avoid any signal being received by the
1048 * listener thread, we block all signals temporarily in the parent,
1049 * while we create the listener thread.
1051 sigfillset(&sig_all_blocked
);
1052 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
1054 ERR("pthread_sigmask: %s", strerror(ret
));
1057 ret
= pthread_attr_init(&thread_attr
);
1059 ERR("pthread_attr_init: %s", strerror(ret
));
1061 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
1063 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1067 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1068 ust_listener_thread
, &global_apps
);
1070 ERR("pthread_create global: %s", strerror(ret
));
1072 global_apps
.thread_active
= 1;
1075 if (local_apps
.allowed
) {
1077 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1078 ust_listener_thread
, &local_apps
);
1080 ERR("pthread_create local: %s", strerror(ret
));
1082 local_apps
.thread_active
= 1;
1085 handle_register_done(&local_apps
);
1087 ret
= pthread_attr_destroy(&thread_attr
);
1089 ERR("pthread_attr_destroy: %s", strerror(ret
));
1092 /* Restore original signal mask in parent */
1093 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1095 ERR("pthread_sigmask: %s", strerror(ret
));
1098 switch (timeout_mode
) {
1099 case 1: /* timeout wait */
1101 ret
= sem_timedwait(&constructor_wait
,
1102 &constructor_timeout
);
1103 } while (ret
< 0 && errno
== EINTR
);
1104 if (ret
< 0 && errno
== ETIMEDOUT
) {
1105 ERR("Timed out waiting for lttng-sessiond");
1110 case -1:/* wait forever */
1112 ret
= sem_wait(&constructor_wait
);
1113 } while (ret
< 0 && errno
== EINTR
);
1116 case 0: /* no timeout */
1122 void lttng_ust_cleanup(int exiting
)
1124 cleanup_sock_info(&global_apps
, exiting
);
1125 if (local_apps
.allowed
) {
1126 cleanup_sock_info(&local_apps
, exiting
);
1129 * The teardown in this function all affect data structures
1130 * accessed under the UST lock by the listener thread. This
1131 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1132 * that none of these threads are accessing this data at this
1135 lttng_ust_abi_exit();
1136 lttng_ust_events_exit();
1137 lttng_ring_buffer_client_discard_exit();
1138 lttng_ring_buffer_client_overwrite_exit();
1139 lttng_ring_buffer_metadata_client_exit();
1142 /* Reinitialize values for fork */
1144 lttng_ust_comm_should_quit
= 0;
1149 void __attribute__((destructor
)) lttng_ust_exit(void)
1154 * Using pthread_cancel here because:
1155 * A) we don't want to hang application teardown.
1156 * B) the thread is not allocating any resource.
1160 * Require the communication thread to quit. Synchronize with
1161 * mutexes to ensure it is not in a mutex critical section when
1162 * pthread_cancel is later called.
1165 lttng_ust_comm_should_quit
= 1;
1167 /* cancel threads */
1168 if (global_apps
.thread_active
) {
1169 ret
= pthread_cancel(global_apps
.ust_listener
);
1171 ERR("Error cancelling global ust listener thread: %s",
1174 global_apps
.thread_active
= 0;
1177 if (local_apps
.thread_active
) {
1178 ret
= pthread_cancel(local_apps
.ust_listener
);
1180 ERR("Error cancelling local ust listener thread: %s",
1183 local_apps
.thread_active
= 0;
1189 * Do NOT join threads: use of sys_futex makes it impossible to
1190 * join the threads without using async-cancel, but async-cancel
1191 * is delivered by a signal, which could hit the target thread
1192 * anywhere in its code path, including while the ust_lock() is
1193 * held, causing a deadlock for the other thread. Let the OS
1194 * cleanup the threads if there are stalled in a syscall.
1196 lttng_ust_cleanup(1);
1200 * We exclude the worker threads across fork and clone (except
1201 * CLONE_VM), because these system calls only keep the forking thread
1202 * running in the child. Therefore, we don't want to call fork or clone
1203 * in the middle of an tracepoint or ust tracing state modification.
1204 * Holding this mutex protects these structures across fork and clone.
1206 void ust_before_fork(sigset_t
*save_sigset
)
1209 * Disable signals. This is to avoid that the child intervenes
1210 * before it is properly setup for tracing. It is safer to
1211 * disable all signals, because then we know we are not breaking
1212 * anything by restoring the original mask.
1217 if (URCU_TLS(lttng_ust_nest_count
))
1219 /* Disable signals */
1220 sigfillset(&all_sigs
);
1221 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1223 PERROR("sigprocmask");
1226 rcu_bp_before_fork();
1229 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1233 DBG("process %d", getpid());
1235 /* Restore signals */
1236 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1238 PERROR("sigprocmask");
1242 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1244 if (URCU_TLS(lttng_ust_nest_count
))
1246 DBG("process %d", getpid());
1247 rcu_bp_after_fork_parent();
1248 /* Release mutexes and reenable signals */
1249 ust_after_fork_common(restore_sigset
);
1253 * After fork, in the child, we need to cleanup all the leftover state,
1254 * except the worker thread which already magically disappeared thanks
1255 * to the weird Linux fork semantics. After tyding up, we call
1256 * lttng_ust_init() again to start over as a new PID.
1258 * This is meant for forks() that have tracing in the child between the
1259 * fork and following exec call (if there is any).
1261 void ust_after_fork_child(sigset_t
*restore_sigset
)
1263 if (URCU_TLS(lttng_ust_nest_count
))
1265 DBG("process %d", getpid());
1266 /* Release urcu mutexes */
1267 rcu_bp_after_fork_child();
1268 lttng_ust_cleanup(0);
1269 lttng_context_vtid_reset();
1270 /* Release mutexes and reenable signals */
1271 ust_after_fork_common(restore_sigset
);