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"
54 #include "lttng-ust-baddr.h"
57 * Has lttng ust comm constructor been called ?
59 static int initialized
;
62 * The ust_lock/ust_unlock lock is used as a communication thread mutex.
63 * Held when handling a command, also held by fork() to deal with
64 * removal of threads, and by exit path.
66 * The UST lock is the centralized mutex across UST tracing control and
69 * ust_exit_mutex must never nest in ust_mutex.
71 * ust_fork_mutex must never nest in ust_mutex.
73 * ust_mutex_nest is a per-thread nesting counter, allowing the perf
74 * counter lazy initialization called by events within the statedump,
75 * which traces while the ust_mutex is held.
77 * ust_lock nests within the dynamic loader lock (within glibc) because
78 * it is taken within the library constructor.
80 static pthread_mutex_t ust_mutex
= PTHREAD_MUTEX_INITIALIZER
;
82 /* Allow nesting the ust_mutex within the same thread. */
83 static DEFINE_URCU_TLS(int, ust_mutex_nest
);
86 * ust_exit_mutex protects thread_active variable wrt thread exit. It
87 * cannot be done by ust_mutex because pthread_cancel(), which takes an
88 * internal libc lock, cannot nest within ust_mutex.
90 * It never nests within a ust_mutex.
92 static pthread_mutex_t ust_exit_mutex
= PTHREAD_MUTEX_INITIALIZER
;
95 * ust_fork_mutex protects base address statedump tracing against forks. It
96 * prevents the dynamic loader lock to be taken (by base address statedump
97 * tracing) while a fork is happening, thus preventing deadlock issues with
98 * the dynamic loader lock.
100 static pthread_mutex_t ust_fork_mutex
= PTHREAD_MUTEX_INITIALIZER
;
102 /* Should the ust comm thread quit ? */
103 static int lttng_ust_comm_should_quit
;
106 * Return 0 on success, -1 if should quit.
107 * The lock is taken in both cases.
112 sigset_t sig_all_blocked
, orig_mask
;
115 sigfillset(&sig_all_blocked
);
116 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
118 ERR("pthread_sigmask: %s", strerror(ret
));
120 if (!URCU_TLS(ust_mutex_nest
)++)
121 pthread_mutex_lock(&ust_mutex
);
122 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
124 ERR("pthread_sigmask: %s", strerror(ret
));
126 if (lttng_ust_comm_should_quit
) {
134 * ust_lock_nocheck() can be used in constructors/destructors, because
135 * they are already nested within the dynamic loader lock, and therefore
136 * have exclusive access against execution of liblttng-ust destructor.
139 void ust_lock_nocheck(void)
141 sigset_t sig_all_blocked
, orig_mask
;
144 sigfillset(&sig_all_blocked
);
145 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
147 ERR("pthread_sigmask: %s", strerror(ret
));
149 if (!URCU_TLS(ust_mutex_nest
)++)
150 pthread_mutex_lock(&ust_mutex
);
151 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
153 ERR("pthread_sigmask: %s", strerror(ret
));
160 void ust_unlock(void)
162 sigset_t sig_all_blocked
, orig_mask
;
165 sigfillset(&sig_all_blocked
);
166 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
168 ERR("pthread_sigmask: %s", strerror(ret
));
170 if (!--URCU_TLS(ust_mutex_nest
))
171 pthread_mutex_unlock(&ust_mutex
);
172 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
174 ERR("pthread_sigmask: %s", strerror(ret
));
179 * Wait for either of these before continuing to the main
181 * - the register_done message from sessiond daemon
182 * (will let the sessiond daemon enable sessions before main
184 * - sessiond daemon is not reachable.
185 * - timeout (ensuring applications are resilient to session
188 static sem_t constructor_wait
;
190 * Doing this for both the global and local sessiond.
192 static int sem_count
= { 2 };
195 * Counting nesting within lttng-ust. Used to ensure that calling fork()
196 * from liblttng-ust does not execute the pre/post fork handlers.
198 static DEFINE_URCU_TLS(int, lttng_ust_nest_count
);
201 * Info about socket and associated listener thread.
205 pthread_t ust_listener
; /* listener thread */
207 int constructor_sem_posted
;
212 char sock_path
[PATH_MAX
];
216 char wait_shm_path
[PATH_MAX
];
218 /* Keep track of lazy state dump not performed yet. */
219 int statedump_pending
;
222 /* Socket from app (connect) to session daemon (listen) for communication */
223 struct sock_info global_apps
= {
231 .sock_path
= LTTNG_DEFAULT_RUNDIR
"/" LTTNG_UST_SOCK_FILENAME
,
235 .wait_shm_path
= "/" LTTNG_UST_WAIT_FILENAME
,
237 .statedump_pending
= 0,
240 /* TODO: allow global_apps_sock_path override */
242 struct sock_info local_apps
= {
246 .allowed
= 0, /* Check setuid bit first */
252 .statedump_pending
= 0,
255 static int wait_poll_fallback
;
257 static const char *cmd_name_mapping
[] = {
258 [ LTTNG_UST_RELEASE
] = "Release",
259 [ LTTNG_UST_SESSION
] = "Create Session",
260 [ LTTNG_UST_TRACER_VERSION
] = "Get Tracer Version",
262 [ LTTNG_UST_TRACEPOINT_LIST
] = "Create Tracepoint List",
263 [ LTTNG_UST_WAIT_QUIESCENT
] = "Wait for Quiescent State",
264 [ LTTNG_UST_REGISTER_DONE
] = "Registration Done",
265 [ LTTNG_UST_TRACEPOINT_FIELD_LIST
] = "Create Tracepoint Field List",
267 /* Session FD commands */
268 [ LTTNG_UST_CHANNEL
] = "Create Channel",
269 [ LTTNG_UST_SESSION_START
] = "Start Session",
270 [ LTTNG_UST_SESSION_STOP
] = "Stop Session",
272 /* Channel FD commands */
273 [ LTTNG_UST_STREAM
] = "Create Stream",
274 [ LTTNG_UST_EVENT
] = "Create Event",
276 /* Event and Channel FD commands */
277 [ LTTNG_UST_CONTEXT
] = "Create Context",
278 [ LTTNG_UST_FLUSH_BUFFER
] = "Flush Buffer",
280 /* Event, Channel and Session commands */
281 [ LTTNG_UST_ENABLE
] = "Enable",
282 [ LTTNG_UST_DISABLE
] = "Disable",
284 /* Tracepoint list commands */
285 [ LTTNG_UST_TRACEPOINT_LIST_GET
] = "List Next Tracepoint",
286 [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
] = "List Next Tracepoint Field",
288 /* Event FD commands */
289 [ LTTNG_UST_FILTER
] = "Create Filter",
290 [ LTTNG_UST_EXCLUSION
] = "Add exclusions to event",
293 static const char *str_timeout
;
294 static int got_timeout_env
;
296 extern void lttng_ring_buffer_client_overwrite_init(void);
297 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
298 extern void lttng_ring_buffer_client_discard_init(void);
299 extern void lttng_ring_buffer_client_discard_rt_init(void);
300 extern void lttng_ring_buffer_metadata_client_init(void);
301 extern void lttng_ring_buffer_client_overwrite_exit(void);
302 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
303 extern void lttng_ring_buffer_client_discard_exit(void);
304 extern void lttng_ring_buffer_client_discard_rt_exit(void);
305 extern void lttng_ring_buffer_metadata_client_exit(void);
308 * Returns the HOME directory path. Caller MUST NOT free(3) the returned
312 const char *get_lttng_home_dir(void)
316 val
= (const char *) getenv("LTTNG_HOME");
320 return (const char *) getenv("HOME");
324 * Force a read (imply TLS fixup for dlopen) of TLS variables.
327 void lttng_fixup_nest_count_tls(void)
329 asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count
)));
333 void lttng_fixup_ust_mutex_nest_tls(void)
335 asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest
)));
342 void lttng_fixup_urcu_bp_tls(void)
348 int lttng_get_notify_socket(void *owner
)
350 struct sock_info
*info
= owner
;
352 return info
->notify_socket
;
356 void print_cmd(int cmd
, int handle
)
358 const char *cmd_name
= "Unknown";
360 if (cmd
>= 0 && cmd
< LTTNG_ARRAY_SIZE(cmd_name_mapping
)
361 && cmd_name_mapping
[cmd
]) {
362 cmd_name
= cmd_name_mapping
[cmd
];
364 DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
366 lttng_ust_obj_get_name(handle
), handle
);
370 int setup_local_apps(void)
372 const char *home_dir
;
377 * Disallow per-user tracing for setuid binaries.
379 if (uid
!= geteuid()) {
380 assert(local_apps
.allowed
== 0);
383 home_dir
= get_lttng_home_dir();
385 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
386 assert(local_apps
.allowed
== 0);
389 local_apps
.allowed
= 1;
390 snprintf(local_apps
.sock_path
, PATH_MAX
, "%s/%s/%s",
392 LTTNG_DEFAULT_HOME_RUNDIR
,
393 LTTNG_UST_SOCK_FILENAME
);
394 snprintf(local_apps
.wait_shm_path
, PATH_MAX
, "/%s-%u",
395 LTTNG_UST_WAIT_FILENAME
,
401 * Get notify_sock timeout, in ms.
402 * -1: don't wait. 0: wait forever. >0: timeout, in ms.
405 long get_timeout(void)
407 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
409 if (!got_timeout_env
) {
410 str_timeout
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
414 constructor_delay_ms
= strtol(str_timeout
, NULL
, 10);
415 return constructor_delay_ms
;
419 long get_notify_sock_timeout(void)
421 return get_timeout();
425 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
428 int get_constructor_timeout(struct timespec
*constructor_timeout
)
430 long constructor_delay_ms
;
433 constructor_delay_ms
= get_timeout();
435 switch (constructor_delay_ms
) {
436 case -1:/* fall-through */
438 return constructor_delay_ms
;
444 * If we are unable to find the current time, don't wait.
446 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
450 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
451 constructor_timeout
->tv_nsec
+=
452 (constructor_delay_ms
% 1000UL) * 1000000UL;
453 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
454 constructor_timeout
->tv_sec
++;
455 constructor_timeout
->tv_nsec
-= 1000000000UL;
461 int register_to_sessiond(int socket
, enum ustctl_socket_type type
)
463 return ustcomm_send_reg_msg(socket
,
466 lttng_alignof(uint8_t) * CHAR_BIT
,
467 lttng_alignof(uint16_t) * CHAR_BIT
,
468 lttng_alignof(uint32_t) * CHAR_BIT
,
469 lttng_alignof(uint64_t) * CHAR_BIT
,
470 lttng_alignof(unsigned long) * CHAR_BIT
);
474 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
478 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
481 DBG("message successfully sent");
484 if (len
== -ECONNRESET
) {
485 DBG("remote end closed connection");
490 DBG("incorrect message size: %zd", len
);
496 int handle_register_done(struct sock_info
*sock_info
)
500 if (sock_info
->constructor_sem_posted
)
502 sock_info
->constructor_sem_posted
= 1;
503 if (uatomic_read(&sem_count
) <= 0) {
506 ret
= uatomic_add_return(&sem_count
, -1);
508 ret
= sem_post(&constructor_wait
);
515 * Only execute pending statedump after the constructor semaphore has
516 * been posted by each listener thread. This means statedump will only
517 * be performed after the "registration done" command is received from
518 * each session daemon the application is connected to.
520 * This ensures we don't run into deadlock issues with the dynamic
521 * loader mutex, which is held while the constructor is called and
522 * waiting on the constructor semaphore. All operations requiring this
523 * dynamic loader lock need to be postponed using this mechanism.
526 void handle_pending_statedump(struct sock_info
*sock_info
)
528 int ctor_passed
= sock_info
->constructor_sem_posted
;
530 if (ctor_passed
&& sock_info
->statedump_pending
) {
531 sock_info
->statedump_pending
= 0;
532 pthread_mutex_lock(&ust_fork_mutex
);
533 lttng_handle_pending_statedump(sock_info
);
534 pthread_mutex_unlock(&ust_fork_mutex
);
539 int handle_message(struct sock_info
*sock_info
,
540 int sock
, struct ustcomm_ust_msg
*lum
)
543 const struct lttng_ust_objd_ops
*ops
;
544 struct ustcomm_ust_reply lur
;
548 memset(&lur
, 0, sizeof(lur
));
551 ret
= -LTTNG_UST_ERR_EXITING
;
555 ops
= objd_ops(lum
->handle
);
562 case LTTNG_UST_REGISTER_DONE
:
563 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
564 ret
= handle_register_done(sock_info
);
568 case LTTNG_UST_RELEASE
:
569 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
572 ret
= lttng_ust_objd_unref(lum
->handle
, 1);
574 case LTTNG_UST_FILTER
:
576 /* Receive filter data */
577 struct lttng_ust_filter_bytecode_node
*bytecode
;
579 if (lum
->u
.filter
.data_size
> FILTER_BYTECODE_MAX_LEN
) {
580 ERR("Filter data size is too large: %u bytes",
581 lum
->u
.filter
.data_size
);
586 if (lum
->u
.filter
.reloc_offset
> lum
->u
.filter
.data_size
) {
587 ERR("Filter reloc offset %u is not within data",
588 lum
->u
.filter
.reloc_offset
);
593 bytecode
= zmalloc(sizeof(*bytecode
) + lum
->u
.filter
.data_size
);
598 len
= ustcomm_recv_unix_sock(sock
, bytecode
->bc
.data
,
599 lum
->u
.filter
.data_size
);
601 case 0: /* orderly shutdown */
606 if (len
== lum
->u
.filter
.data_size
) {
607 DBG("filter data received");
609 } else if (len
< 0) {
610 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
611 if (len
== -ECONNRESET
) {
612 ERR("%s remote end closed connection", sock_info
->name
);
621 DBG("incorrect filter data message size: %zd", len
);
627 bytecode
->bc
.len
= lum
->u
.filter
.data_size
;
628 bytecode
->bc
.reloc_offset
= lum
->u
.filter
.reloc_offset
;
629 bytecode
->bc
.seqnum
= lum
->u
.filter
.seqnum
;
631 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
632 (unsigned long) bytecode
,
637 /* don't free bytecode if everything went fine. */
644 case LTTNG_UST_EXCLUSION
:
646 /* Receive exclusion names */
647 struct lttng_ust_excluder_node
*node
;
650 count
= lum
->u
.exclusion
.count
;
652 /* There are no names to read */
656 node
= zmalloc(sizeof(*node
) +
657 count
* LTTNG_UST_SYM_NAME_LEN
);
662 node
->excluder
.count
= count
;
663 len
= ustcomm_recv_unix_sock(sock
, node
->excluder
.names
,
664 count
* LTTNG_UST_SYM_NAME_LEN
);
666 case 0: /* orderly shutdown */
671 if (len
== count
* LTTNG_UST_SYM_NAME_LEN
) {
672 DBG("Exclusion data received");
674 } else if (len
< 0) {
675 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
676 if (len
== -ECONNRESET
) {
677 ERR("%s remote end closed connection", sock_info
->name
);
686 DBG("Incorrect exclusion data message size: %zd", len
);
693 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
694 (unsigned long) node
,
699 /* Don't free exclusion data if everything went fine. */
706 case LTTNG_UST_CHANNEL
:
711 len
= ustcomm_recv_channel_from_sessiond(sock
,
712 &chan_data
, lum
->u
.channel
.len
,
715 case 0: /* orderly shutdown */
719 if (len
== lum
->u
.channel
.len
) {
720 DBG("channel data received");
722 } else if (len
< 0) {
723 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
724 if (len
== -ECONNRESET
) {
725 ERR("%s remote end closed connection", sock_info
->name
);
732 DBG("incorrect channel data message size: %zd", len
);
737 args
.channel
.chan_data
= chan_data
;
738 args
.channel
.wakeup_fd
= wakeup_fd
;
740 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
741 (unsigned long) &lum
->u
,
747 case LTTNG_UST_STREAM
:
749 /* Receive shm_fd, wakeup_fd */
750 ret
= ustcomm_recv_stream_from_sessiond(sock
,
753 &args
.stream
.wakeup_fd
);
758 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
759 (unsigned long) &lum
->u
,
767 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
768 (unsigned long) &lum
->u
,
776 lur
.handle
= lum
->handle
;
780 lur
.ret_code
= LTTNG_UST_OK
;
783 * Use -LTTNG_UST_ERR as wildcard for UST internal
784 * error that are not caused by the transport, except if
785 * we already have a more precise error message to
788 if (ret
> -LTTNG_UST_ERR
) {
789 /* Translate code to UST error. */
792 lur
.ret_code
= -LTTNG_UST_ERR_EXIST
;
795 lur
.ret_code
= -LTTNG_UST_ERR_INVAL
;
798 lur
.ret_code
= -LTTNG_UST_ERR_NOENT
;
801 lur
.ret_code
= -LTTNG_UST_ERR_PERM
;
804 lur
.ret_code
= -LTTNG_UST_ERR_NOSYS
;
807 lur
.ret_code
= -LTTNG_UST_ERR
;
816 case LTTNG_UST_TRACER_VERSION
:
817 lur
.u
.version
= lum
->u
.version
;
819 case LTTNG_UST_TRACEPOINT_LIST_GET
:
820 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
824 DBG("Return value: %d", lur
.ret_val
);
825 ret
= send_reply(sock
, &lur
);
827 DBG("error sending reply");
832 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
835 if (lur
.ret_code
== LTTNG_UST_OK
) {
837 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
838 len
= ustcomm_send_unix_sock(sock
,
839 &args
.field_list
.entry
,
840 sizeof(args
.field_list
.entry
));
845 if (len
!= sizeof(args
.field_list
.entry
)) {
856 * Performed delayed statedump operations outside of the UST
857 * lock. We need to take the dynamic loader lock before we take
858 * the UST lock internally within handle_pending_statedump().
860 handle_pending_statedump(sock_info
);
866 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
870 if (sock_info
->root_handle
!= -1) {
871 ret
= lttng_ust_objd_unref(sock_info
->root_handle
, 1);
873 ERR("Error unref root handle");
875 sock_info
->root_handle
= -1;
877 sock_info
->constructor_sem_posted
= 0;
880 * wait_shm_mmap, socket and notify socket are used by listener
881 * threads outside of the ust lock, so we cannot tear them down
882 * ourselves, because we cannot join on these threads. Leave
883 * responsibility of cleaning up these resources to the OS
889 if (sock_info
->socket
!= -1) {
890 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
892 ERR("Error closing ust cmd socket");
894 sock_info
->socket
= -1;
896 if (sock_info
->notify_socket
!= -1) {
897 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
899 ERR("Error closing ust notify socket");
901 sock_info
->notify_socket
= -1;
903 if (sock_info
->wait_shm_mmap
) {
906 page_size
= sysconf(_SC_PAGE_SIZE
);
908 ret
= munmap(sock_info
->wait_shm_mmap
, page_size
);
910 ERR("Error unmapping wait shm");
913 sock_info
->wait_shm_mmap
= NULL
;
918 * Using fork to set umask in the child process (not multi-thread safe).
919 * We deal with the shm_open vs ftruncate race (happening when the
920 * sessiond owns the shm and does not let everybody modify it, to ensure
921 * safety against shm_unlink) by simply letting the mmap fail and
922 * retrying after a few seconds.
923 * For global shm, everybody has rw access to it until the sessiond
927 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
929 int wait_shm_fd
, ret
;
933 * Try to open read-only.
935 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
936 if (wait_shm_fd
>= 0) {
939 size_t bytes_read
= 0;
942 * Try to read the fd. If unable to do so, try opening
946 len
= read(wait_shm_fd
,
947 &((char *) &tmp_read
)[bytes_read
],
948 sizeof(tmp_read
) - bytes_read
);
952 } while ((len
< 0 && errno
== EINTR
)
953 || (len
> 0 && bytes_read
< sizeof(tmp_read
)));
954 if (bytes_read
!= sizeof(tmp_read
)) {
955 ret
= close(wait_shm_fd
);
957 ERR("close wait_shm_fd");
962 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
964 * Real-only open did not work, and it's not because the
965 * entry was not present. It's a failure that prohibits
968 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
974 * If the open failed because the file did not exist, or because
975 * the file was not truncated yet, try creating it ourself.
977 URCU_TLS(lttng_ust_nest_count
)++;
979 URCU_TLS(lttng_ust_nest_count
)--;
984 * Parent: wait for child to return, in which case the
985 * shared memory map will have been created.
988 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
993 * Try to open read-only again after creation.
995 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
996 if (wait_shm_fd
< 0) {
998 * Real-only open did not work. It's a failure
999 * that prohibits using shm.
1001 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
1005 } else if (pid
== 0) {
1009 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
1010 if (sock_info
->global
)
1011 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
1013 * We're alone in a child process, so we can modify the
1014 * process-wide umask.
1016 umask(~create_mode
);
1018 * Try creating shm (or get rw access).
1019 * We don't do an exclusive open, because we allow other
1020 * processes to create+ftruncate it concurrently.
1022 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
1023 O_RDWR
| O_CREAT
, create_mode
);
1024 if (wait_shm_fd
>= 0) {
1025 ret
= ftruncate(wait_shm_fd
, mmap_size
);
1027 PERROR("ftruncate");
1028 _exit(EXIT_FAILURE
);
1030 _exit(EXIT_SUCCESS
);
1033 * For local shm, we need to have rw access to accept
1034 * opening it: this means the local sessiond will be
1035 * able to wake us up. For global shm, we open it even
1036 * if rw access is not granted, because the root.root
1037 * sessiond will be able to override all rights and wake
1040 if (!sock_info
->global
&& errno
!= EACCES
) {
1041 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
1042 _exit(EXIT_FAILURE
);
1045 * The shm exists, but we cannot open it RW. Report
1048 _exit(EXIT_SUCCESS
);
1053 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
1054 struct stat statbuf
;
1057 * Ensure that our user is the owner of the shm file for
1058 * local shm. If we do not own the file, it means our
1059 * sessiond will not have access to wake us up (there is
1060 * probably a rogue process trying to fake our
1061 * sessiond). Fallback to polling method in this case.
1063 ret
= fstat(wait_shm_fd
, &statbuf
);
1068 if (statbuf
.st_uid
!= getuid())
1074 ret
= close(wait_shm_fd
);
1076 PERROR("Error closing fd");
1082 char *get_map_shm(struct sock_info
*sock_info
)
1085 int wait_shm_fd
, ret
;
1086 char *wait_shm_mmap
;
1088 page_size
= sysconf(_SC_PAGE_SIZE
);
1089 if (page_size
< 0) {
1093 wait_shm_fd
= get_wait_shm(sock_info
, page_size
);
1094 if (wait_shm_fd
< 0) {
1097 wait_shm_mmap
= mmap(NULL
, page_size
, PROT_READ
,
1098 MAP_SHARED
, wait_shm_fd
, 0);
1099 /* close shm fd immediately after taking the mmap reference */
1100 ret
= close(wait_shm_fd
);
1102 PERROR("Error closing fd");
1104 if (wait_shm_mmap
== MAP_FAILED
) {
1105 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1108 return wait_shm_mmap
;
1115 void wait_for_sessiond(struct sock_info
*sock_info
)
1122 if (wait_poll_fallback
) {
1125 if (!sock_info
->wait_shm_mmap
) {
1126 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
1127 if (!sock_info
->wait_shm_mmap
)
1132 DBG("Waiting for %s apps sessiond", sock_info
->name
);
1133 /* Wait for futex wakeup */
1134 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
1135 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
1136 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
1138 if (errno
== EFAULT
) {
1139 wait_poll_fallback
= 1;
1141 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1142 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1143 "Please upgrade your kernel "
1144 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1145 "mainline). LTTng-UST will use polling mode fallback.");
1163 * This thread does not allocate any resource, except within
1164 * handle_message, within mutex protection. This mutex protects against
1166 * The other moment it allocates resources is at socket connection, which
1167 * is also protected by the mutex.
1170 void *ust_listener_thread(void *arg
)
1172 struct sock_info
*sock_info
= arg
;
1173 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
1176 /* Restart trying to connect to the session daemon */
1178 if (prev_connect_failed
) {
1179 /* Wait for sessiond availability with pipe */
1180 wait_for_sessiond(sock_info
);
1184 * Sleep for 5 seconds before retrying after a
1185 * sequence of failure / wait / failure. This
1186 * deals with a killed or broken session daemon.
1191 prev_connect_failed
= 0;
1194 if (sock_info
->socket
!= -1) {
1195 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
1197 ERR("Error closing %s ust cmd socket",
1200 sock_info
->socket
= -1;
1202 if (sock_info
->notify_socket
!= -1) {
1203 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
1205 ERR("Error closing %s ust notify socket",
1208 sock_info
->notify_socket
= -1;
1212 * Register. We need to perform both connect and sending
1213 * registration message before doing the next connect otherwise
1214 * we may reach unix socket connect queue max limits and block
1215 * on the 2nd connect while the session daemon is awaiting the
1216 * first connect registration message.
1218 /* Connect cmd socket */
1219 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1221 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1222 prev_connect_failed
= 1;
1229 * If we cannot find the sessiond daemon, don't delay
1230 * constructor execution.
1232 ret
= handle_register_done(sock_info
);
1237 sock_info
->socket
= ret
;
1244 * Create only one root handle per listener thread for the whole
1245 * process lifetime, so we ensure we get ID which is statically
1246 * assigned to the root handle.
1248 if (sock_info
->root_handle
== -1) {
1249 ret
= lttng_abi_create_root_handle();
1251 ERR("Error creating root handle");
1254 sock_info
->root_handle
= ret
;
1257 ret
= register_to_sessiond(sock_info
->socket
, USTCTL_SOCKET_CMD
);
1259 ERR("Error registering to %s ust cmd socket",
1261 prev_connect_failed
= 1;
1263 * If we cannot register to the sessiond daemon, don't
1264 * delay constructor execution.
1266 ret
= handle_register_done(sock_info
);
1274 /* Connect notify socket */
1275 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1277 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1278 prev_connect_failed
= 1;
1285 * If we cannot find the sessiond daemon, don't delay
1286 * constructor execution.
1288 ret
= handle_register_done(sock_info
);
1293 sock_info
->notify_socket
= ret
;
1295 timeout
= get_notify_sock_timeout();
1298 * Give at least 10ms to sessiond to reply to
1303 ret
= ustcomm_setsockopt_rcv_timeout(sock_info
->notify_socket
,
1306 WARN("Error setting socket receive timeout");
1308 ret
= ustcomm_setsockopt_snd_timeout(sock_info
->notify_socket
,
1311 WARN("Error setting socket send timeout");
1313 } else if (timeout
< -1) {
1314 WARN("Unsupported timeout value %ld", timeout
);
1321 ret
= register_to_sessiond(sock_info
->notify_socket
,
1322 USTCTL_SOCKET_NOTIFY
);
1324 ERR("Error registering to %s ust notify socket",
1326 prev_connect_failed
= 1;
1328 * If we cannot register to the sessiond daemon, don't
1329 * delay constructor execution.
1331 ret
= handle_register_done(sock_info
);
1336 sock
= sock_info
->socket
;
1342 struct ustcomm_ust_msg lum
;
1344 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
1346 case 0: /* orderly shutdown */
1347 DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info
->name
);
1352 * Either sessiond has shutdown or refused us by closing the socket.
1353 * In either case, we don't want to delay construction execution,
1354 * and we need to wait before retry.
1356 prev_connect_failed
= 1;
1358 * If we cannot register to the sessiond daemon, don't
1359 * delay constructor execution.
1361 ret
= handle_register_done(sock_info
);
1366 print_cmd(lum
.cmd
, lum
.handle
);
1367 ret
= handle_message(sock_info
, sock
, &lum
);
1369 ERR("Error handling message for %s socket", sock_info
->name
);
1374 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
1376 DBG("incorrect message size (%s socket): %zd", sock_info
->name
, len
);
1378 if (len
== -ECONNRESET
) {
1379 DBG("%s remote end closed connection", sock_info
->name
);
1390 /* Cleanup socket handles before trying to reconnect */
1391 lttng_ust_objd_table_owner_cleanup(sock_info
);
1393 goto restart
; /* try to reconnect */
1398 pthread_mutex_lock(&ust_exit_mutex
);
1399 sock_info
->thread_active
= 0;
1400 pthread_mutex_unlock(&ust_exit_mutex
);
1405 * Weak symbol to call when the ust malloc wrapper is not loaded.
1407 __attribute__((weak
))
1408 void lttng_ust_malloc_wrapper_init(void)
1413 * sessiond monitoring thread: monitor presence of global and per-user
1414 * sessiond by polling the application common named pipe.
1416 void __attribute__((constructor
)) lttng_ust_init(void)
1418 struct timespec constructor_timeout
;
1419 sigset_t sig_all_blocked
, orig_parent_mask
;
1420 pthread_attr_t thread_attr
;
1424 if (uatomic_xchg(&initialized
, 1) == 1)
1428 * Fixup interdependency between TLS fixup mutex (which happens
1429 * to be the dynamic linker mutex) and ust_lock, taken within
1432 lttng_fixup_urcu_bp_tls();
1433 lttng_fixup_ringbuffer_tls();
1434 lttng_fixup_vtid_tls();
1435 lttng_fixup_nest_count_tls();
1436 lttng_fixup_procname_tls();
1437 lttng_fixup_ust_mutex_nest_tls();
1440 * We want precise control over the order in which we construct
1441 * our sub-libraries vs starting to receive commands from
1442 * sessiond (otherwise leading to errors when trying to create
1443 * sessiond before the init functions are completed).
1447 lttng_ust_baddr_statedump_init();
1448 lttng_ring_buffer_metadata_client_init();
1449 lttng_ring_buffer_client_overwrite_init();
1450 lttng_ring_buffer_client_overwrite_rt_init();
1451 lttng_ring_buffer_client_discard_init();
1452 lttng_ring_buffer_client_discard_rt_init();
1453 lttng_perf_counter_init();
1454 lttng_context_init();
1456 * Invoke ust malloc wrapper init before starting other threads.
1458 lttng_ust_malloc_wrapper_init();
1460 timeout_mode
= get_constructor_timeout(&constructor_timeout
);
1462 ret
= sem_init(&constructor_wait
, 0, 0);
1465 ret
= setup_local_apps();
1467 DBG("local apps setup returned %d", ret
);
1470 /* A new thread created by pthread_create inherits the signal mask
1471 * from the parent. To avoid any signal being received by the
1472 * listener thread, we block all signals temporarily in the parent,
1473 * while we create the listener thread.
1475 sigfillset(&sig_all_blocked
);
1476 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
1478 ERR("pthread_sigmask: %s", strerror(ret
));
1481 ret
= pthread_attr_init(&thread_attr
);
1483 ERR("pthread_attr_init: %s", strerror(ret
));
1485 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
1487 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1490 pthread_mutex_lock(&ust_exit_mutex
);
1491 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1492 ust_listener_thread
, &global_apps
);
1494 ERR("pthread_create global: %s", strerror(ret
));
1496 global_apps
.thread_active
= 1;
1497 pthread_mutex_unlock(&ust_exit_mutex
);
1499 if (local_apps
.allowed
) {
1500 pthread_mutex_lock(&ust_exit_mutex
);
1501 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1502 ust_listener_thread
, &local_apps
);
1504 ERR("pthread_create local: %s", strerror(ret
));
1506 local_apps
.thread_active
= 1;
1507 pthread_mutex_unlock(&ust_exit_mutex
);
1509 handle_register_done(&local_apps
);
1511 ret
= pthread_attr_destroy(&thread_attr
);
1513 ERR("pthread_attr_destroy: %s", strerror(ret
));
1516 /* Restore original signal mask in parent */
1517 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1519 ERR("pthread_sigmask: %s", strerror(ret
));
1522 switch (timeout_mode
) {
1523 case 1: /* timeout wait */
1525 ret
= sem_timedwait(&constructor_wait
,
1526 &constructor_timeout
);
1527 } while (ret
< 0 && errno
== EINTR
);
1528 if (ret
< 0 && errno
== ETIMEDOUT
) {
1529 ERR("Timed out waiting for lttng-sessiond");
1534 case -1:/* wait forever */
1536 ret
= sem_wait(&constructor_wait
);
1537 } while (ret
< 0 && errno
== EINTR
);
1540 case 0: /* no timeout */
1546 void lttng_ust_cleanup(int exiting
)
1548 cleanup_sock_info(&global_apps
, exiting
);
1549 if (local_apps
.allowed
) {
1550 cleanup_sock_info(&local_apps
, exiting
);
1553 * The teardown in this function all affect data structures
1554 * accessed under the UST lock by the listener thread. This
1555 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1556 * that none of these threads are accessing this data at this
1559 lttng_ust_abi_exit();
1560 lttng_ust_events_exit();
1561 lttng_context_exit();
1562 lttng_perf_counter_exit();
1563 lttng_ring_buffer_client_discard_rt_exit();
1564 lttng_ring_buffer_client_discard_exit();
1565 lttng_ring_buffer_client_overwrite_rt_exit();
1566 lttng_ring_buffer_client_overwrite_exit();
1567 lttng_ring_buffer_metadata_client_exit();
1568 lttng_ust_baddr_statedump_destroy();
1571 /* Reinitialize values for fork */
1573 lttng_ust_comm_should_quit
= 0;
1578 void __attribute__((destructor
)) lttng_ust_exit(void)
1583 * Using pthread_cancel here because:
1584 * A) we don't want to hang application teardown.
1585 * B) the thread is not allocating any resource.
1589 * Require the communication thread to quit. Synchronize with
1590 * mutexes to ensure it is not in a mutex critical section when
1591 * pthread_cancel is later called.
1594 lttng_ust_comm_should_quit
= 1;
1597 pthread_mutex_lock(&ust_exit_mutex
);
1598 /* cancel threads */
1599 if (global_apps
.thread_active
) {
1600 ret
= pthread_cancel(global_apps
.ust_listener
);
1602 ERR("Error cancelling global ust listener thread: %s",
1605 global_apps
.thread_active
= 0;
1608 if (local_apps
.thread_active
) {
1609 ret
= pthread_cancel(local_apps
.ust_listener
);
1611 ERR("Error cancelling local ust listener thread: %s",
1614 local_apps
.thread_active
= 0;
1617 pthread_mutex_unlock(&ust_exit_mutex
);
1620 * Do NOT join threads: use of sys_futex makes it impossible to
1621 * join the threads without using async-cancel, but async-cancel
1622 * is delivered by a signal, which could hit the target thread
1623 * anywhere in its code path, including while the ust_lock() is
1624 * held, causing a deadlock for the other thread. Let the OS
1625 * cleanup the threads if there are stalled in a syscall.
1627 lttng_ust_cleanup(1);
1631 * We exclude the worker threads across fork and clone (except
1632 * CLONE_VM), because these system calls only keep the forking thread
1633 * running in the child. Therefore, we don't want to call fork or clone
1634 * in the middle of an tracepoint or ust tracing state modification.
1635 * Holding this mutex protects these structures across fork and clone.
1637 void ust_before_fork(sigset_t
*save_sigset
)
1640 * Disable signals. This is to avoid that the child intervenes
1641 * before it is properly setup for tracing. It is safer to
1642 * disable all signals, because then we know we are not breaking
1643 * anything by restoring the original mask.
1648 if (URCU_TLS(lttng_ust_nest_count
))
1650 /* Disable signals */
1651 sigfillset(&all_sigs
);
1652 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1654 PERROR("sigprocmask");
1657 pthread_mutex_lock(&ust_fork_mutex
);
1660 rcu_bp_before_fork();
1663 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1667 DBG("process %d", getpid());
1670 pthread_mutex_unlock(&ust_fork_mutex
);
1672 /* Restore signals */
1673 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1675 PERROR("sigprocmask");
1679 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1681 if (URCU_TLS(lttng_ust_nest_count
))
1683 DBG("process %d", getpid());
1684 rcu_bp_after_fork_parent();
1685 /* Release mutexes and reenable signals */
1686 ust_after_fork_common(restore_sigset
);
1690 * After fork, in the child, we need to cleanup all the leftover state,
1691 * except the worker thread which already magically disappeared thanks
1692 * to the weird Linux fork semantics. After tyding up, we call
1693 * lttng_ust_init() again to start over as a new PID.
1695 * This is meant for forks() that have tracing in the child between the
1696 * fork and following exec call (if there is any).
1698 void ust_after_fork_child(sigset_t
*restore_sigset
)
1700 if (URCU_TLS(lttng_ust_nest_count
))
1702 DBG("process %d", getpid());
1703 /* Release urcu mutexes */
1704 rcu_bp_after_fork_child();
1705 lttng_ust_cleanup(0);
1706 lttng_context_vtid_reset();
1707 /* Release mutexes and reenable signals */
1708 ust_after_fork_common(restore_sigset
);
1712 void lttng_ust_sockinfo_session_enabled(void *owner
)
1714 struct sock_info
*sock_info
= owner
;
1715 sock_info
->statedump_pending
= 1;