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>
38 #include <urcu/uatomic.h>
39 #include <urcu/futex.h>
40 #include <urcu/compiler.h>
42 #include <lttng/ust-events.h>
43 #include <lttng/ust-abi.h>
44 #include <lttng/ust.h>
45 #include <lttng/ust-error.h>
46 #include <lttng/ust-ctl.h>
47 #include <urcu/tls-compat.h>
49 #include <usterr-signal-safe.h>
51 #include "tracepoint-internal.h"
52 #include "lttng-tracer-core.h"
54 #include "../libringbuffer/tlsfixup.h"
55 #include "lttng-ust-baddr.h"
58 * Has lttng ust comm constructor been called ?
60 static int initialized
;
63 * The ust_lock/ust_unlock lock is used as a communication thread mutex.
64 * Held when handling a command, also held by fork() to deal with
65 * removal of threads, and by exit path.
67 * The UST lock is the centralized mutex across UST tracing control and
70 * ust_exit_mutex must never nest in ust_mutex.
72 static pthread_mutex_t ust_mutex
= PTHREAD_MUTEX_INITIALIZER
;
75 * ust_exit_mutex protects thread_active variable wrt thread exit. It
76 * cannot be done by ust_mutex because pthread_cancel(), which takes an
77 * internal libc lock, cannot nest within ust_mutex.
79 * It never nests within a ust_mutex.
81 static pthread_mutex_t ust_exit_mutex
= PTHREAD_MUTEX_INITIALIZER
;
83 /* Should the ust comm thread quit ? */
84 static int lttng_ust_comm_should_quit
;
87 * Return 0 on success, -1 if should quilt.
88 * The lock is taken in both cases.
92 pthread_mutex_lock(&ust_mutex
);
93 if (lttng_ust_comm_should_quit
) {
101 * ust_lock_nocheck() can be used in constructors/destructors, because
102 * they are already nested within the dynamic loader lock, and therefore
103 * have exclusive access against execution of liblttng-ust destructor.
105 void ust_lock_nocheck(void)
107 pthread_mutex_lock(&ust_mutex
);
110 void ust_unlock(void)
112 pthread_mutex_unlock(&ust_mutex
);
116 * Wait for either of these before continuing to the main
118 * - the register_done message from sessiond daemon
119 * (will let the sessiond daemon enable sessions before main
121 * - sessiond daemon is not reachable.
122 * - timeout (ensuring applications are resilient to session
125 static sem_t constructor_wait
;
127 * Doing this for both the global and local sessiond.
129 static int sem_count
= { 2 };
132 * Counting nesting within lttng-ust. Used to ensure that calling fork()
133 * from liblttng-ust does not execute the pre/post fork handlers.
135 static DEFINE_URCU_TLS(int, lttng_ust_nest_count
);
138 * Info about socket and associated listener thread.
142 pthread_t ust_listener
; /* listener thread */
144 int constructor_sem_posted
;
149 char sock_path
[PATH_MAX
];
153 char wait_shm_path
[PATH_MAX
];
155 /* Keep track of lazy state dump not performed yet. */
156 int statedump_pending
;
159 /* Socket from app (connect) to session daemon (listen) for communication */
160 struct sock_info global_apps
= {
168 .sock_path
= LTTNG_DEFAULT_RUNDIR
"/" LTTNG_UST_SOCK_FILENAME
,
172 .wait_shm_path
= "/" LTTNG_UST_WAIT_FILENAME
,
174 .statedump_pending
= 0,
177 /* TODO: allow global_apps_sock_path override */
179 struct sock_info local_apps
= {
183 .allowed
= 0, /* Check setuid bit first */
189 .statedump_pending
= 0,
192 static int wait_poll_fallback
;
194 static const char *cmd_name_mapping
[] = {
195 [ LTTNG_UST_RELEASE
] = "Release",
196 [ LTTNG_UST_SESSION
] = "Create Session",
197 [ LTTNG_UST_TRACER_VERSION
] = "Get Tracer Version",
199 [ LTTNG_UST_TRACEPOINT_LIST
] = "Create Tracepoint List",
200 [ LTTNG_UST_WAIT_QUIESCENT
] = "Wait for Quiescent State",
201 [ LTTNG_UST_REGISTER_DONE
] = "Registration Done",
202 [ LTTNG_UST_TRACEPOINT_FIELD_LIST
] = "Create Tracepoint Field List",
204 /* Session FD commands */
205 [ LTTNG_UST_CHANNEL
] = "Create Channel",
206 [ LTTNG_UST_SESSION_START
] = "Start Session",
207 [ LTTNG_UST_SESSION_STOP
] = "Stop Session",
209 /* Channel FD commands */
210 [ LTTNG_UST_STREAM
] = "Create Stream",
211 [ LTTNG_UST_EVENT
] = "Create Event",
213 /* Event and Channel FD commands */
214 [ LTTNG_UST_CONTEXT
] = "Create Context",
215 [ LTTNG_UST_FLUSH_BUFFER
] = "Flush Buffer",
217 /* Event, Channel and Session commands */
218 [ LTTNG_UST_ENABLE
] = "Enable",
219 [ LTTNG_UST_DISABLE
] = "Disable",
221 /* Tracepoint list commands */
222 [ LTTNG_UST_TRACEPOINT_LIST_GET
] = "List Next Tracepoint",
223 [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
] = "List Next Tracepoint Field",
225 /* Event FD commands */
226 [ LTTNG_UST_FILTER
] = "Create Filter",
227 [ LTTNG_UST_EXCLUSION
] = "Add exclusions to event",
230 static const char *str_timeout
;
231 static int got_timeout_env
;
233 extern void lttng_ring_buffer_client_overwrite_init(void);
234 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
235 extern void lttng_ring_buffer_client_discard_init(void);
236 extern void lttng_ring_buffer_client_discard_rt_init(void);
237 extern void lttng_ring_buffer_metadata_client_init(void);
238 extern void lttng_ring_buffer_client_overwrite_exit(void);
239 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
240 extern void lttng_ring_buffer_client_discard_exit(void);
241 extern void lttng_ring_buffer_client_discard_rt_exit(void);
242 extern void lttng_ring_buffer_metadata_client_exit(void);
245 * Returns the HOME directory path. Caller MUST NOT free(3) the returned
249 const char *get_lttng_home_dir(void)
253 val
= (const char *) getenv("LTTNG_HOME");
257 return (const char *) getenv("HOME");
261 * Force a read (imply TLS fixup for dlopen) of TLS variables.
264 void lttng_fixup_nest_count_tls(void)
266 asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count
)));
269 int lttng_get_notify_socket(void *owner
)
271 struct sock_info
*info
= owner
;
273 return info
->notify_socket
;
277 void print_cmd(int cmd
, int handle
)
279 const char *cmd_name
= "Unknown";
281 if (cmd
>= 0 && cmd
< LTTNG_ARRAY_SIZE(cmd_name_mapping
)
282 && cmd_name_mapping
[cmd
]) {
283 cmd_name
= cmd_name_mapping
[cmd
];
285 DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
287 lttng_ust_obj_get_name(handle
), handle
);
291 int setup_local_apps(void)
293 const char *home_dir
;
298 * Disallow per-user tracing for setuid binaries.
300 if (uid
!= geteuid()) {
301 assert(local_apps
.allowed
== 0);
304 home_dir
= get_lttng_home_dir();
306 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
307 assert(local_apps
.allowed
== 0);
310 local_apps
.allowed
= 1;
311 snprintf(local_apps
.sock_path
, PATH_MAX
, "%s/%s/%s",
313 LTTNG_DEFAULT_HOME_RUNDIR
,
314 LTTNG_UST_SOCK_FILENAME
);
315 snprintf(local_apps
.wait_shm_path
, PATH_MAX
, "/%s-%u",
316 LTTNG_UST_WAIT_FILENAME
,
322 * Get notify_sock timeout, in ms.
323 * -1: don't wait. 0: wait forever. >0: timeout, in ms.
326 long get_timeout(void)
328 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
330 if (!got_timeout_env
) {
331 str_timeout
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
335 constructor_delay_ms
= strtol(str_timeout
, NULL
, 10);
336 return constructor_delay_ms
;
340 long get_notify_sock_timeout(void)
342 return get_timeout();
346 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
349 int get_constructor_timeout(struct timespec
*constructor_timeout
)
351 long constructor_delay_ms
;
354 constructor_delay_ms
= get_timeout();
356 switch (constructor_delay_ms
) {
357 case -1:/* fall-through */
359 return constructor_delay_ms
;
365 * If we are unable to find the current time, don't wait.
367 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
371 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
372 constructor_timeout
->tv_nsec
+=
373 (constructor_delay_ms
% 1000UL) * 1000000UL;
374 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
375 constructor_timeout
->tv_sec
++;
376 constructor_timeout
->tv_nsec
-= 1000000000UL;
382 int register_to_sessiond(int socket
, enum ustctl_socket_type type
)
384 return ustcomm_send_reg_msg(socket
,
387 lttng_alignof(uint8_t) * CHAR_BIT
,
388 lttng_alignof(uint16_t) * CHAR_BIT
,
389 lttng_alignof(uint32_t) * CHAR_BIT
,
390 lttng_alignof(uint64_t) * CHAR_BIT
,
391 lttng_alignof(unsigned long) * CHAR_BIT
);
395 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
399 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
402 DBG("message successfully sent");
405 if (len
== -ECONNRESET
) {
406 DBG("remote end closed connection");
411 DBG("incorrect message size: %zd", len
);
417 int handle_register_done(struct sock_info
*sock_info
)
421 if (sock_info
->constructor_sem_posted
)
423 sock_info
->constructor_sem_posted
= 1;
424 if (uatomic_read(&sem_count
) <= 0) {
427 ret
= uatomic_add_return(&sem_count
, -1);
429 ret
= sem_post(&constructor_wait
);
436 * Only execute pending statedump after the constructor semaphore has
437 * been posted by each listener thread. This means statedump will only
438 * be performed after the "registration done" command is received from
439 * each session daemon the application is connected to.
441 * This ensures we don't run into deadlock issues with the dynamic
442 * loader mutex, which is held while the constructor is called and
443 * waiting on the constructor semaphore. All operations requiring this
444 * dynamic loader lock need to be postponed using this mechanism.
447 void handle_pending_statedump(struct sock_info
*sock_info
)
449 int ctor_passed
= sock_info
->constructor_sem_posted
;
451 if (ctor_passed
&& sock_info
->statedump_pending
) {
452 sock_info
->statedump_pending
= 0;
453 lttng_handle_pending_statedump(sock_info
);
458 int handle_message(struct sock_info
*sock_info
,
459 int sock
, struct ustcomm_ust_msg
*lum
)
462 const struct lttng_ust_objd_ops
*ops
;
463 struct ustcomm_ust_reply lur
;
467 memset(&lur
, 0, sizeof(lur
));
470 ret
= -LTTNG_UST_ERR_EXITING
;
474 ops
= objd_ops(lum
->handle
);
481 case LTTNG_UST_REGISTER_DONE
:
482 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
483 ret
= handle_register_done(sock_info
);
487 case LTTNG_UST_RELEASE
:
488 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
491 ret
= lttng_ust_objd_unref(lum
->handle
, 1);
493 case LTTNG_UST_FILTER
:
495 /* Receive filter data */
496 struct lttng_ust_filter_bytecode_node
*bytecode
;
498 if (lum
->u
.filter
.data_size
> FILTER_BYTECODE_MAX_LEN
) {
499 ERR("Filter data size is too large: %u bytes",
500 lum
->u
.filter
.data_size
);
505 if (lum
->u
.filter
.reloc_offset
> lum
->u
.filter
.data_size
) {
506 ERR("Filter reloc offset %u is not within data",
507 lum
->u
.filter
.reloc_offset
);
512 bytecode
= zmalloc(sizeof(*bytecode
) + lum
->u
.filter
.data_size
);
517 len
= ustcomm_recv_unix_sock(sock
, bytecode
->bc
.data
,
518 lum
->u
.filter
.data_size
);
520 case 0: /* orderly shutdown */
525 if (len
== lum
->u
.filter
.data_size
) {
526 DBG("filter data received");
528 } else if (len
< 0) {
529 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
530 if (len
== -ECONNRESET
) {
531 ERR("%s remote end closed connection", sock_info
->name
);
540 DBG("incorrect filter data message size: %zd", len
);
546 bytecode
->bc
.len
= lum
->u
.filter
.data_size
;
547 bytecode
->bc
.reloc_offset
= lum
->u
.filter
.reloc_offset
;
548 bytecode
->bc
.seqnum
= lum
->u
.filter
.seqnum
;
550 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
551 (unsigned long) bytecode
,
556 /* don't free bytecode if everything went fine. */
563 case LTTNG_UST_EXCLUSION
:
565 /* Receive exclusion names */
566 struct lttng_ust_excluder_node
*node
;
569 count
= lum
->u
.exclusion
.count
;
571 /* There are no names to read */
575 node
= zmalloc(sizeof(*node
) +
576 count
* LTTNG_UST_SYM_NAME_LEN
);
581 node
->excluder
.count
= count
;
582 len
= ustcomm_recv_unix_sock(sock
, node
->excluder
.names
,
583 count
* LTTNG_UST_SYM_NAME_LEN
);
585 case 0: /* orderly shutdown */
590 if (len
== count
* LTTNG_UST_SYM_NAME_LEN
) {
591 DBG("Exclusion data received");
593 } else if (len
< 0) {
594 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
595 if (len
== -ECONNRESET
) {
596 ERR("%s remote end closed connection", sock_info
->name
);
605 DBG("Incorrect exclusion data message size: %zd", len
);
612 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
613 (unsigned long) node
,
618 /* Don't free exclusion data if everything went fine. */
625 case LTTNG_UST_CHANNEL
:
630 len
= ustcomm_recv_channel_from_sessiond(sock
,
631 &chan_data
, lum
->u
.channel
.len
,
634 case 0: /* orderly shutdown */
638 if (len
== lum
->u
.channel
.len
) {
639 DBG("channel data received");
641 } else if (len
< 0) {
642 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
643 if (len
== -ECONNRESET
) {
644 ERR("%s remote end closed connection", sock_info
->name
);
651 DBG("incorrect channel data message size: %zd", len
);
656 args
.channel
.chan_data
= chan_data
;
657 args
.channel
.wakeup_fd
= wakeup_fd
;
659 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
660 (unsigned long) &lum
->u
,
666 case LTTNG_UST_STREAM
:
668 /* Receive shm_fd, wakeup_fd */
669 ret
= ustcomm_recv_stream_from_sessiond(sock
,
672 &args
.stream
.wakeup_fd
);
677 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
678 (unsigned long) &lum
->u
,
686 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
687 (unsigned long) &lum
->u
,
695 lur
.handle
= lum
->handle
;
699 lur
.ret_code
= LTTNG_UST_OK
;
702 * Use -LTTNG_UST_ERR as wildcard for UST internal
703 * error that are not caused by the transport, except if
704 * we already have a more precise error message to
707 if (ret
> -LTTNG_UST_ERR
) {
708 /* Translate code to UST error. */
711 lur
.ret_code
= -LTTNG_UST_ERR_EXIST
;
714 lur
.ret_code
= -LTTNG_UST_ERR_INVAL
;
717 lur
.ret_code
= -LTTNG_UST_ERR_NOENT
;
720 lur
.ret_code
= -LTTNG_UST_ERR_PERM
;
723 lur
.ret_code
= -LTTNG_UST_ERR_NOSYS
;
726 lur
.ret_code
= -LTTNG_UST_ERR
;
735 case LTTNG_UST_TRACER_VERSION
:
736 lur
.u
.version
= lum
->u
.version
;
738 case LTTNG_UST_TRACEPOINT_LIST_GET
:
739 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
743 DBG("Return value: %d", lur
.ret_val
);
744 ret
= send_reply(sock
, &lur
);
746 DBG("error sending reply");
751 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
754 if (lur
.ret_code
== LTTNG_UST_OK
) {
756 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
757 len
= ustcomm_send_unix_sock(sock
,
758 &args
.field_list
.entry
,
759 sizeof(args
.field_list
.entry
));
764 if (len
!= sizeof(args
.field_list
.entry
)) {
775 * Performed delayed statedump operations outside of the UST
776 * lock. We need to take the dynamic loader lock before we take
777 * the UST lock internally within handle_pending_statedump().
779 handle_pending_statedump(sock_info
);
785 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
789 if (sock_info
->root_handle
!= -1) {
790 ret
= lttng_ust_objd_unref(sock_info
->root_handle
, 1);
792 ERR("Error unref root handle");
794 sock_info
->root_handle
= -1;
796 sock_info
->constructor_sem_posted
= 0;
799 * wait_shm_mmap, socket and notify socket are used by listener
800 * threads outside of the ust lock, so we cannot tear them down
801 * ourselves, because we cannot join on these threads. Leave
802 * responsibility of cleaning up these resources to the OS
808 if (sock_info
->socket
!= -1) {
809 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
811 ERR("Error closing ust cmd socket");
813 sock_info
->socket
= -1;
815 if (sock_info
->notify_socket
!= -1) {
816 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
818 ERR("Error closing ust notify socket");
820 sock_info
->notify_socket
= -1;
822 if (sock_info
->wait_shm_mmap
) {
823 ret
= munmap(sock_info
->wait_shm_mmap
, sysconf(_SC_PAGE_SIZE
));
825 ERR("Error unmapping wait shm");
827 sock_info
->wait_shm_mmap
= NULL
;
832 * Using fork to set umask in the child process (not multi-thread safe).
833 * We deal with the shm_open vs ftruncate race (happening when the
834 * sessiond owns the shm and does not let everybody modify it, to ensure
835 * safety against shm_unlink) by simply letting the mmap fail and
836 * retrying after a few seconds.
837 * For global shm, everybody has rw access to it until the sessiond
841 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
843 int wait_shm_fd
, ret
;
847 * Try to open read-only.
849 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
850 if (wait_shm_fd
>= 0) {
853 size_t bytes_read
= 0;
856 * Try to read the fd. If unable to do so, try opening
860 len
= read(wait_shm_fd
,
861 &((char *) &tmp_read
)[bytes_read
],
862 sizeof(tmp_read
) - bytes_read
);
866 } while ((len
< 0 && errno
== EINTR
)
867 || (len
> 0 && bytes_read
< sizeof(tmp_read
)));
868 if (bytes_read
!= sizeof(tmp_read
)) {
869 ret
= close(wait_shm_fd
);
871 ERR("close wait_shm_fd");
876 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
878 * Real-only open did not work, and it's not because the
879 * entry was not present. It's a failure that prohibits
882 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
888 * If the open failed because the file did not exist, or because
889 * the file was not truncated yet, try creating it ourself.
891 URCU_TLS(lttng_ust_nest_count
)++;
893 URCU_TLS(lttng_ust_nest_count
)--;
898 * Parent: wait for child to return, in which case the
899 * shared memory map will have been created.
902 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
907 * Try to open read-only again after creation.
909 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
910 if (wait_shm_fd
< 0) {
912 * Real-only open did not work. It's a failure
913 * that prohibits using shm.
915 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
919 } else if (pid
== 0) {
923 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
924 if (sock_info
->global
)
925 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
927 * We're alone in a child process, so we can modify the
928 * process-wide umask.
932 * Try creating shm (or get rw access).
933 * We don't do an exclusive open, because we allow other
934 * processes to create+ftruncate it concurrently.
936 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
937 O_RDWR
| O_CREAT
, create_mode
);
938 if (wait_shm_fd
>= 0) {
939 ret
= ftruncate(wait_shm_fd
, mmap_size
);
947 * For local shm, we need to have rw access to accept
948 * opening it: this means the local sessiond will be
949 * able to wake us up. For global shm, we open it even
950 * if rw access is not granted, because the root.root
951 * sessiond will be able to override all rights and wake
954 if (!sock_info
->global
&& errno
!= EACCES
) {
955 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
959 * The shm exists, but we cannot open it RW. Report
967 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
971 * Ensure that our user is the owner of the shm file for
972 * local shm. If we do not own the file, it means our
973 * sessiond will not have access to wake us up (there is
974 * probably a rogue process trying to fake our
975 * sessiond). Fallback to polling method in this case.
977 ret
= fstat(wait_shm_fd
, &statbuf
);
982 if (statbuf
.st_uid
!= getuid())
988 ret
= close(wait_shm_fd
);
990 PERROR("Error closing fd");
996 char *get_map_shm(struct sock_info
*sock_info
)
998 size_t mmap_size
= sysconf(_SC_PAGE_SIZE
);
999 int wait_shm_fd
, ret
;
1000 char *wait_shm_mmap
;
1002 wait_shm_fd
= get_wait_shm(sock_info
, mmap_size
);
1003 if (wait_shm_fd
< 0) {
1006 wait_shm_mmap
= mmap(NULL
, mmap_size
, PROT_READ
,
1007 MAP_SHARED
, wait_shm_fd
, 0);
1008 /* close shm fd immediately after taking the mmap reference */
1009 ret
= close(wait_shm_fd
);
1011 PERROR("Error closing fd");
1013 if (wait_shm_mmap
== MAP_FAILED
) {
1014 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1017 return wait_shm_mmap
;
1024 void wait_for_sessiond(struct sock_info
*sock_info
)
1031 if (wait_poll_fallback
) {
1034 if (!sock_info
->wait_shm_mmap
) {
1035 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
1036 if (!sock_info
->wait_shm_mmap
)
1041 DBG("Waiting for %s apps sessiond", sock_info
->name
);
1042 /* Wait for futex wakeup */
1043 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
1044 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
1045 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
1047 if (errno
== EFAULT
) {
1048 wait_poll_fallback
= 1;
1050 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1051 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1052 "Please upgrade your kernel "
1053 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1054 "mainline). LTTng-UST will use polling mode fallback.");
1072 * This thread does not allocate any resource, except within
1073 * handle_message, within mutex protection. This mutex protects against
1075 * The other moment it allocates resources is at socket connection, which
1076 * is also protected by the mutex.
1079 void *ust_listener_thread(void *arg
)
1081 struct sock_info
*sock_info
= arg
;
1082 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
1085 /* Restart trying to connect to the session daemon */
1087 if (prev_connect_failed
) {
1088 /* Wait for sessiond availability with pipe */
1089 wait_for_sessiond(sock_info
);
1093 * Sleep for 5 seconds before retrying after a
1094 * sequence of failure / wait / failure. This
1095 * deals with a killed or broken session daemon.
1100 prev_connect_failed
= 0;
1103 if (sock_info
->socket
!= -1) {
1104 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
1106 ERR("Error closing %s ust cmd socket",
1109 sock_info
->socket
= -1;
1111 if (sock_info
->notify_socket
!= -1) {
1112 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
1114 ERR("Error closing %s ust notify socket",
1117 sock_info
->notify_socket
= -1;
1121 * Register. We need to perform both connect and sending
1122 * registration message before doing the next connect otherwise
1123 * we may reach unix socket connect queue max limits and block
1124 * on the 2nd connect while the session daemon is awaiting the
1125 * first connect registration message.
1127 /* Connect cmd socket */
1128 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1130 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1131 prev_connect_failed
= 1;
1138 * If we cannot find the sessiond daemon, don't delay
1139 * constructor execution.
1141 ret
= handle_register_done(sock_info
);
1146 sock_info
->socket
= ret
;
1153 * Create only one root handle per listener thread for the whole
1154 * process lifetime, so we ensure we get ID which is statically
1155 * assigned to the root handle.
1157 if (sock_info
->root_handle
== -1) {
1158 ret
= lttng_abi_create_root_handle();
1160 ERR("Error creating root handle");
1163 sock_info
->root_handle
= ret
;
1166 ret
= register_to_sessiond(sock_info
->socket
, USTCTL_SOCKET_CMD
);
1168 ERR("Error registering to %s ust cmd socket",
1170 prev_connect_failed
= 1;
1172 * If we cannot register to the sessiond daemon, don't
1173 * delay constructor execution.
1175 ret
= handle_register_done(sock_info
);
1183 /* Connect notify socket */
1184 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1186 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1187 prev_connect_failed
= 1;
1194 * If we cannot find the sessiond daemon, don't delay
1195 * constructor execution.
1197 ret
= handle_register_done(sock_info
);
1202 sock_info
->notify_socket
= ret
;
1204 timeout
= get_notify_sock_timeout();
1207 * Give at least 10ms to sessiond to reply to
1212 ret
= ustcomm_setsockopt_rcv_timeout(sock_info
->notify_socket
,
1215 WARN("Error setting socket receive timeout");
1217 ret
= ustcomm_setsockopt_snd_timeout(sock_info
->notify_socket
,
1220 WARN("Error setting socket send timeout");
1222 } else if (timeout
< -1) {
1223 WARN("Unsupported timeout value %ld", timeout
);
1230 ret
= register_to_sessiond(sock_info
->notify_socket
,
1231 USTCTL_SOCKET_NOTIFY
);
1233 ERR("Error registering to %s ust notify socket",
1235 prev_connect_failed
= 1;
1237 * If we cannot register to the sessiond daemon, don't
1238 * delay constructor execution.
1240 ret
= handle_register_done(sock_info
);
1245 sock
= sock_info
->socket
;
1251 struct ustcomm_ust_msg lum
;
1253 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
1255 case 0: /* orderly shutdown */
1256 DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info
->name
);
1261 * Either sessiond has shutdown or refused us by closing the socket.
1262 * In either case, we don't want to delay construction execution,
1263 * and we need to wait before retry.
1265 prev_connect_failed
= 1;
1267 * If we cannot register to the sessiond daemon, don't
1268 * delay constructor execution.
1270 ret
= handle_register_done(sock_info
);
1275 print_cmd(lum
.cmd
, lum
.handle
);
1276 ret
= handle_message(sock_info
, sock
, &lum
);
1278 ERR("Error handling message for %s socket", sock_info
->name
);
1283 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
1285 DBG("incorrect message size (%s socket): %zd", sock_info
->name
, len
);
1287 if (len
== -ECONNRESET
) {
1288 DBG("%s remote end closed connection", sock_info
->name
);
1299 /* Cleanup socket handles before trying to reconnect */
1300 lttng_ust_objd_table_owner_cleanup(sock_info
);
1302 goto restart
; /* try to reconnect */
1307 pthread_mutex_lock(&ust_exit_mutex
);
1308 sock_info
->thread_active
= 0;
1309 pthread_mutex_unlock(&ust_exit_mutex
);
1314 * Weak symbol to call when the ust malloc wrapper is not loaded.
1316 __attribute__((weak
))
1317 void lttng_ust_malloc_wrapper_init(void)
1322 * sessiond monitoring thread: monitor presence of global and per-user
1323 * sessiond by polling the application common named pipe.
1325 void __attribute__((constructor
)) lttng_ust_init(void)
1327 struct timespec constructor_timeout
;
1328 sigset_t sig_all_blocked
, orig_parent_mask
;
1329 pthread_attr_t thread_attr
;
1333 if (uatomic_xchg(&initialized
, 1) == 1)
1337 * Fixup interdependency between TLS fixup mutex (which happens
1338 * to be the dynamic linker mutex) and ust_lock, taken within
1341 lttng_fixup_ringbuffer_tls();
1342 lttng_fixup_vtid_tls();
1343 lttng_fixup_nest_count_tls();
1344 lttng_fixup_procname_tls();
1347 * We want precise control over the order in which we construct
1348 * our sub-libraries vs starting to receive commands from
1349 * sessiond (otherwise leading to errors when trying to create
1350 * sessiond before the init functions are completed).
1354 lttng_ust_baddr_statedump_init();
1355 lttng_ring_buffer_metadata_client_init();
1356 lttng_ring_buffer_client_overwrite_init();
1357 lttng_ring_buffer_client_overwrite_rt_init();
1358 lttng_ring_buffer_client_discard_init();
1359 lttng_ring_buffer_client_discard_rt_init();
1360 lttng_context_init();
1362 * Invoke ust malloc wrapper init before starting other threads.
1364 lttng_ust_malloc_wrapper_init();
1366 timeout_mode
= get_constructor_timeout(&constructor_timeout
);
1368 ret
= sem_init(&constructor_wait
, 0, 0);
1371 ret
= setup_local_apps();
1373 DBG("local apps setup returned %d", ret
);
1376 /* A new thread created by pthread_create inherits the signal mask
1377 * from the parent. To avoid any signal being received by the
1378 * listener thread, we block all signals temporarily in the parent,
1379 * while we create the listener thread.
1381 sigfillset(&sig_all_blocked
);
1382 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
1384 ERR("pthread_sigmask: %s", strerror(ret
));
1387 ret
= pthread_attr_init(&thread_attr
);
1389 ERR("pthread_attr_init: %s", strerror(ret
));
1391 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
1393 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1396 pthread_mutex_lock(&ust_exit_mutex
);
1397 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1398 ust_listener_thread
, &global_apps
);
1400 ERR("pthread_create global: %s", strerror(ret
));
1402 global_apps
.thread_active
= 1;
1403 pthread_mutex_unlock(&ust_exit_mutex
);
1405 if (local_apps
.allowed
) {
1406 pthread_mutex_lock(&ust_exit_mutex
);
1407 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1408 ust_listener_thread
, &local_apps
);
1410 ERR("pthread_create local: %s", strerror(ret
));
1412 local_apps
.thread_active
= 1;
1413 pthread_mutex_unlock(&ust_exit_mutex
);
1415 handle_register_done(&local_apps
);
1417 ret
= pthread_attr_destroy(&thread_attr
);
1419 ERR("pthread_attr_destroy: %s", strerror(ret
));
1422 /* Restore original signal mask in parent */
1423 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1425 ERR("pthread_sigmask: %s", strerror(ret
));
1428 switch (timeout_mode
) {
1429 case 1: /* timeout wait */
1431 ret
= sem_timedwait(&constructor_wait
,
1432 &constructor_timeout
);
1433 } while (ret
< 0 && errno
== EINTR
);
1434 if (ret
< 0 && errno
== ETIMEDOUT
) {
1435 ERR("Timed out waiting for lttng-sessiond");
1440 case -1:/* wait forever */
1442 ret
= sem_wait(&constructor_wait
);
1443 } while (ret
< 0 && errno
== EINTR
);
1446 case 0: /* no timeout */
1452 void lttng_ust_cleanup(int exiting
)
1454 cleanup_sock_info(&global_apps
, exiting
);
1455 if (local_apps
.allowed
) {
1456 cleanup_sock_info(&local_apps
, exiting
);
1459 * The teardown in this function all affect data structures
1460 * accessed under the UST lock by the listener thread. This
1461 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1462 * that none of these threads are accessing this data at this
1465 lttng_ust_abi_exit();
1466 lttng_ust_events_exit();
1467 lttng_context_exit();
1468 lttng_ring_buffer_client_discard_rt_exit();
1469 lttng_ring_buffer_client_discard_exit();
1470 lttng_ring_buffer_client_overwrite_rt_exit();
1471 lttng_ring_buffer_client_overwrite_exit();
1472 lttng_ring_buffer_metadata_client_exit();
1473 lttng_ust_baddr_statedump_destroy();
1476 /* Reinitialize values for fork */
1478 lttng_ust_comm_should_quit
= 0;
1483 void __attribute__((destructor
)) lttng_ust_exit(void)
1488 * Using pthread_cancel here because:
1489 * A) we don't want to hang application teardown.
1490 * B) the thread is not allocating any resource.
1494 * Require the communication thread to quit. Synchronize with
1495 * mutexes to ensure it is not in a mutex critical section when
1496 * pthread_cancel is later called.
1499 lttng_ust_comm_should_quit
= 1;
1502 pthread_mutex_lock(&ust_exit_mutex
);
1503 /* cancel threads */
1504 if (global_apps
.thread_active
) {
1505 ret
= pthread_cancel(global_apps
.ust_listener
);
1507 ERR("Error cancelling global ust listener thread: %s",
1510 global_apps
.thread_active
= 0;
1513 if (local_apps
.thread_active
) {
1514 ret
= pthread_cancel(local_apps
.ust_listener
);
1516 ERR("Error cancelling local ust listener thread: %s",
1519 local_apps
.thread_active
= 0;
1522 pthread_mutex_unlock(&ust_exit_mutex
);
1525 * Do NOT join threads: use of sys_futex makes it impossible to
1526 * join the threads without using async-cancel, but async-cancel
1527 * is delivered by a signal, which could hit the target thread
1528 * anywhere in its code path, including while the ust_lock() is
1529 * held, causing a deadlock for the other thread. Let the OS
1530 * cleanup the threads if there are stalled in a syscall.
1532 lttng_ust_cleanup(1);
1536 * We exclude the worker threads across fork and clone (except
1537 * CLONE_VM), because these system calls only keep the forking thread
1538 * running in the child. Therefore, we don't want to call fork or clone
1539 * in the middle of an tracepoint or ust tracing state modification.
1540 * Holding this mutex protects these structures across fork and clone.
1542 void ust_before_fork(sigset_t
*save_sigset
)
1545 * Disable signals. This is to avoid that the child intervenes
1546 * before it is properly setup for tracing. It is safer to
1547 * disable all signals, because then we know we are not breaking
1548 * anything by restoring the original mask.
1553 if (URCU_TLS(lttng_ust_nest_count
))
1555 /* Disable signals */
1556 sigfillset(&all_sigs
);
1557 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1559 PERROR("sigprocmask");
1562 rcu_bp_before_fork();
1565 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1569 DBG("process %d", getpid());
1571 /* Restore signals */
1572 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1574 PERROR("sigprocmask");
1578 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1580 if (URCU_TLS(lttng_ust_nest_count
))
1582 DBG("process %d", getpid());
1583 rcu_bp_after_fork_parent();
1584 /* Release mutexes and reenable signals */
1585 ust_after_fork_common(restore_sigset
);
1589 * After fork, in the child, we need to cleanup all the leftover state,
1590 * except the worker thread which already magically disappeared thanks
1591 * to the weird Linux fork semantics. After tyding up, we call
1592 * lttng_ust_init() again to start over as a new PID.
1594 * This is meant for forks() that have tracing in the child between the
1595 * fork and following exec call (if there is any).
1597 void ust_after_fork_child(sigset_t
*restore_sigset
)
1599 if (URCU_TLS(lttng_ust_nest_count
))
1601 DBG("process %d", getpid());
1602 /* Release urcu mutexes */
1603 rcu_bp_after_fork_child();
1604 lttng_ust_cleanup(0);
1605 lttng_context_vtid_reset();
1606 /* Release mutexes and reenable signals */
1607 ust_after_fork_common(restore_sigset
);
1611 void lttng_ust_sockinfo_session_enabled(void *owner
)
1613 struct sock_info
*sock_info
= owner
;
1614 sock_info
->statedump_pending
= 1;