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_mutex_nest is a per-thread nesting counter, allowing the perf
72 * counter lazy initialization called by events within the statedump,
73 * which traces while the ust_mutex is held.
75 static pthread_mutex_t ust_mutex
= PTHREAD_MUTEX_INITIALIZER
;
77 /* Allow nesting the ust_mutex within the same thread. */
78 static DEFINE_URCU_TLS(int, ust_mutex_nest
);
81 * ust_exit_mutex protects thread_active variable wrt thread exit. It
82 * cannot be done by ust_mutex because pthread_cancel(), which takes an
83 * internal libc lock, cannot nest within ust_mutex.
85 * It never nests within a ust_mutex.
87 static pthread_mutex_t ust_exit_mutex
= PTHREAD_MUTEX_INITIALIZER
;
90 * ust_fork_mutex protects base address statedump tracing against forks. It
91 * prevents the dynamic loader lock to be taken (by base address statedump
92 * tracing) while a fork is happening, thus preventing deadlock issues with
93 * the dynamic loader lock.
95 static pthread_mutex_t ust_fork_mutex
= PTHREAD_MUTEX_INITIALIZER
;
97 /* Should the ust comm thread quit ? */
98 static int lttng_ust_comm_should_quit
;
101 * Return 0 on success, -1 if should quit.
102 * The lock is taken in both cases.
107 sigset_t sig_all_blocked
, orig_mask
;
110 sigfillset(&sig_all_blocked
);
111 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
113 ERR("pthread_sigmask: %s", strerror(ret
));
115 if (!URCU_TLS(ust_mutex_nest
)++)
116 pthread_mutex_lock(&ust_mutex
);
117 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
119 ERR("pthread_sigmask: %s", strerror(ret
));
121 if (lttng_ust_comm_should_quit
) {
129 * ust_lock_nocheck() can be used in constructors/destructors, because
130 * they are already nested within the dynamic loader lock, and therefore
131 * have exclusive access against execution of liblttng-ust destructor.
134 void ust_lock_nocheck(void)
136 sigset_t sig_all_blocked
, orig_mask
;
139 sigfillset(&sig_all_blocked
);
140 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
142 ERR("pthread_sigmask: %s", strerror(ret
));
144 if (!URCU_TLS(ust_mutex_nest
)++)
145 pthread_mutex_lock(&ust_mutex
);
146 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
148 ERR("pthread_sigmask: %s", strerror(ret
));
155 void ust_unlock(void)
157 sigset_t sig_all_blocked
, orig_mask
;
160 sigfillset(&sig_all_blocked
);
161 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
163 ERR("pthread_sigmask: %s", strerror(ret
));
165 if (!--URCU_TLS(ust_mutex_nest
))
166 pthread_mutex_unlock(&ust_mutex
);
167 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
169 ERR("pthread_sigmask: %s", strerror(ret
));
174 * Wait for either of these before continuing to the main
176 * - the register_done message from sessiond daemon
177 * (will let the sessiond daemon enable sessions before main
179 * - sessiond daemon is not reachable.
180 * - timeout (ensuring applications are resilient to session
183 static sem_t constructor_wait
;
185 * Doing this for both the global and local sessiond.
187 static int sem_count
= { 2 };
190 * Counting nesting within lttng-ust. Used to ensure that calling fork()
191 * from liblttng-ust does not execute the pre/post fork handlers.
193 static DEFINE_URCU_TLS(int, lttng_ust_nest_count
);
196 * Info about socket and associated listener thread.
200 pthread_t ust_listener
; /* listener thread */
202 int constructor_sem_posted
;
207 char sock_path
[PATH_MAX
];
211 char wait_shm_path
[PATH_MAX
];
213 /* Keep track of lazy state dump not performed yet. */
214 int statedump_pending
;
217 /* Socket from app (connect) to session daemon (listen) for communication */
218 struct sock_info global_apps
= {
226 .sock_path
= LTTNG_DEFAULT_RUNDIR
"/" LTTNG_UST_SOCK_FILENAME
,
230 .wait_shm_path
= "/" LTTNG_UST_WAIT_FILENAME
,
232 .statedump_pending
= 0,
235 /* TODO: allow global_apps_sock_path override */
237 struct sock_info local_apps
= {
241 .allowed
= 0, /* Check setuid bit first */
247 .statedump_pending
= 0,
250 static int wait_poll_fallback
;
252 static const char *cmd_name_mapping
[] = {
253 [ LTTNG_UST_RELEASE
] = "Release",
254 [ LTTNG_UST_SESSION
] = "Create Session",
255 [ LTTNG_UST_TRACER_VERSION
] = "Get Tracer Version",
257 [ LTTNG_UST_TRACEPOINT_LIST
] = "Create Tracepoint List",
258 [ LTTNG_UST_WAIT_QUIESCENT
] = "Wait for Quiescent State",
259 [ LTTNG_UST_REGISTER_DONE
] = "Registration Done",
260 [ LTTNG_UST_TRACEPOINT_FIELD_LIST
] = "Create Tracepoint Field List",
262 /* Session FD commands */
263 [ LTTNG_UST_CHANNEL
] = "Create Channel",
264 [ LTTNG_UST_SESSION_START
] = "Start Session",
265 [ LTTNG_UST_SESSION_STOP
] = "Stop Session",
267 /* Channel FD commands */
268 [ LTTNG_UST_STREAM
] = "Create Stream",
269 [ LTTNG_UST_EVENT
] = "Create Event",
271 /* Event and Channel FD commands */
272 [ LTTNG_UST_CONTEXT
] = "Create Context",
273 [ LTTNG_UST_FLUSH_BUFFER
] = "Flush Buffer",
275 /* Event, Channel and Session commands */
276 [ LTTNG_UST_ENABLE
] = "Enable",
277 [ LTTNG_UST_DISABLE
] = "Disable",
279 /* Tracepoint list commands */
280 [ LTTNG_UST_TRACEPOINT_LIST_GET
] = "List Next Tracepoint",
281 [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
] = "List Next Tracepoint Field",
283 /* Event FD commands */
284 [ LTTNG_UST_FILTER
] = "Create Filter",
285 [ LTTNG_UST_EXCLUSION
] = "Add exclusions to event",
288 static const char *str_timeout
;
289 static int got_timeout_env
;
291 extern void lttng_ring_buffer_client_overwrite_init(void);
292 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
293 extern void lttng_ring_buffer_client_discard_init(void);
294 extern void lttng_ring_buffer_client_discard_rt_init(void);
295 extern void lttng_ring_buffer_metadata_client_init(void);
296 extern void lttng_ring_buffer_client_overwrite_exit(void);
297 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
298 extern void lttng_ring_buffer_client_discard_exit(void);
299 extern void lttng_ring_buffer_client_discard_rt_exit(void);
300 extern void lttng_ring_buffer_metadata_client_exit(void);
303 * Returns the HOME directory path. Caller MUST NOT free(3) the returned
307 const char *get_lttng_home_dir(void)
311 val
= (const char *) getenv("LTTNG_HOME");
315 return (const char *) getenv("HOME");
319 * Force a read (imply TLS fixup for dlopen) of TLS variables.
322 void lttng_fixup_nest_count_tls(void)
324 asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count
)));
328 void lttng_fixup_ust_mutex_nest_tls(void)
330 asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest
)));
333 int lttng_get_notify_socket(void *owner
)
335 struct sock_info
*info
= owner
;
337 return info
->notify_socket
;
341 void print_cmd(int cmd
, int handle
)
343 const char *cmd_name
= "Unknown";
345 if (cmd
>= 0 && cmd
< LTTNG_ARRAY_SIZE(cmd_name_mapping
)
346 && cmd_name_mapping
[cmd
]) {
347 cmd_name
= cmd_name_mapping
[cmd
];
349 DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
351 lttng_ust_obj_get_name(handle
), handle
);
355 int setup_local_apps(void)
357 const char *home_dir
;
362 * Disallow per-user tracing for setuid binaries.
364 if (uid
!= geteuid()) {
365 assert(local_apps
.allowed
== 0);
368 home_dir
= get_lttng_home_dir();
370 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
371 assert(local_apps
.allowed
== 0);
374 local_apps
.allowed
= 1;
375 snprintf(local_apps
.sock_path
, PATH_MAX
, "%s/%s/%s",
377 LTTNG_DEFAULT_HOME_RUNDIR
,
378 LTTNG_UST_SOCK_FILENAME
);
379 snprintf(local_apps
.wait_shm_path
, PATH_MAX
, "/%s-%u",
380 LTTNG_UST_WAIT_FILENAME
,
386 * Get notify_sock timeout, in ms.
387 * -1: don't wait. 0: wait forever. >0: timeout, in ms.
390 long get_timeout(void)
392 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
394 if (!got_timeout_env
) {
395 str_timeout
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
399 constructor_delay_ms
= strtol(str_timeout
, NULL
, 10);
400 return constructor_delay_ms
;
404 long get_notify_sock_timeout(void)
406 return get_timeout();
410 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
413 int get_constructor_timeout(struct timespec
*constructor_timeout
)
415 long constructor_delay_ms
;
418 constructor_delay_ms
= get_timeout();
420 switch (constructor_delay_ms
) {
421 case -1:/* fall-through */
423 return constructor_delay_ms
;
429 * If we are unable to find the current time, don't wait.
431 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
435 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
436 constructor_timeout
->tv_nsec
+=
437 (constructor_delay_ms
% 1000UL) * 1000000UL;
438 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
439 constructor_timeout
->tv_sec
++;
440 constructor_timeout
->tv_nsec
-= 1000000000UL;
446 int register_to_sessiond(int socket
, enum ustctl_socket_type type
)
448 return ustcomm_send_reg_msg(socket
,
451 lttng_alignof(uint8_t) * CHAR_BIT
,
452 lttng_alignof(uint16_t) * CHAR_BIT
,
453 lttng_alignof(uint32_t) * CHAR_BIT
,
454 lttng_alignof(uint64_t) * CHAR_BIT
,
455 lttng_alignof(unsigned long) * CHAR_BIT
);
459 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
463 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
466 DBG("message successfully sent");
469 if (len
== -ECONNRESET
) {
470 DBG("remote end closed connection");
475 DBG("incorrect message size: %zd", len
);
481 int handle_register_done(struct sock_info
*sock_info
)
485 if (sock_info
->constructor_sem_posted
)
487 sock_info
->constructor_sem_posted
= 1;
488 if (uatomic_read(&sem_count
) <= 0) {
491 ret
= uatomic_add_return(&sem_count
, -1);
493 ret
= sem_post(&constructor_wait
);
500 * Only execute pending statedump after the constructor semaphore has
501 * been posted by each listener thread. This means statedump will only
502 * be performed after the "registration done" command is received from
503 * each session daemon the application is connected to.
505 * This ensures we don't run into deadlock issues with the dynamic
506 * loader mutex, which is held while the constructor is called and
507 * waiting on the constructor semaphore. All operations requiring this
508 * dynamic loader lock need to be postponed using this mechanism.
511 void handle_pending_statedump(struct sock_info
*sock_info
)
513 int ctor_passed
= sock_info
->constructor_sem_posted
;
515 if (ctor_passed
&& sock_info
->statedump_pending
) {
516 pthread_mutex_lock(&ust_fork_mutex
);
517 sock_info
->statedump_pending
= 0;
518 lttng_handle_pending_statedump(sock_info
);
519 pthread_mutex_unlock(&ust_fork_mutex
);
524 int handle_message(struct sock_info
*sock_info
,
525 int sock
, struct ustcomm_ust_msg
*lum
)
528 const struct lttng_ust_objd_ops
*ops
;
529 struct ustcomm_ust_reply lur
;
533 memset(&lur
, 0, sizeof(lur
));
536 ret
= -LTTNG_UST_ERR_EXITING
;
540 ops
= objd_ops(lum
->handle
);
547 case LTTNG_UST_REGISTER_DONE
:
548 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
549 ret
= handle_register_done(sock_info
);
553 case LTTNG_UST_RELEASE
:
554 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
557 ret
= lttng_ust_objd_unref(lum
->handle
, 1);
559 case LTTNG_UST_FILTER
:
561 /* Receive filter data */
562 struct lttng_ust_filter_bytecode_node
*bytecode
;
564 if (lum
->u
.filter
.data_size
> FILTER_BYTECODE_MAX_LEN
) {
565 ERR("Filter data size is too large: %u bytes",
566 lum
->u
.filter
.data_size
);
571 if (lum
->u
.filter
.reloc_offset
> lum
->u
.filter
.data_size
) {
572 ERR("Filter reloc offset %u is not within data",
573 lum
->u
.filter
.reloc_offset
);
578 bytecode
= zmalloc(sizeof(*bytecode
) + lum
->u
.filter
.data_size
);
583 len
= ustcomm_recv_unix_sock(sock
, bytecode
->bc
.data
,
584 lum
->u
.filter
.data_size
);
586 case 0: /* orderly shutdown */
591 if (len
== lum
->u
.filter
.data_size
) {
592 DBG("filter data received");
594 } else if (len
< 0) {
595 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
596 if (len
== -ECONNRESET
) {
597 ERR("%s remote end closed connection", sock_info
->name
);
606 DBG("incorrect filter data message size: %zd", len
);
612 bytecode
->bc
.len
= lum
->u
.filter
.data_size
;
613 bytecode
->bc
.reloc_offset
= lum
->u
.filter
.reloc_offset
;
614 bytecode
->bc
.seqnum
= lum
->u
.filter
.seqnum
;
616 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
617 (unsigned long) bytecode
,
622 /* don't free bytecode if everything went fine. */
629 case LTTNG_UST_EXCLUSION
:
631 /* Receive exclusion names */
632 struct lttng_ust_excluder_node
*node
;
635 count
= lum
->u
.exclusion
.count
;
637 /* There are no names to read */
641 node
= zmalloc(sizeof(*node
) +
642 count
* LTTNG_UST_SYM_NAME_LEN
);
647 node
->excluder
.count
= count
;
648 len
= ustcomm_recv_unix_sock(sock
, node
->excluder
.names
,
649 count
* LTTNG_UST_SYM_NAME_LEN
);
651 case 0: /* orderly shutdown */
656 if (len
== count
* LTTNG_UST_SYM_NAME_LEN
) {
657 DBG("Exclusion data received");
659 } else if (len
< 0) {
660 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
661 if (len
== -ECONNRESET
) {
662 ERR("%s remote end closed connection", sock_info
->name
);
671 DBG("Incorrect exclusion data message size: %zd", len
);
678 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
679 (unsigned long) node
,
684 /* Don't free exclusion data if everything went fine. */
691 case LTTNG_UST_CHANNEL
:
696 len
= ustcomm_recv_channel_from_sessiond(sock
,
697 &chan_data
, lum
->u
.channel
.len
,
700 case 0: /* orderly shutdown */
704 if (len
== lum
->u
.channel
.len
) {
705 DBG("channel data received");
707 } else if (len
< 0) {
708 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
709 if (len
== -ECONNRESET
) {
710 ERR("%s remote end closed connection", sock_info
->name
);
717 DBG("incorrect channel data message size: %zd", len
);
722 args
.channel
.chan_data
= chan_data
;
723 args
.channel
.wakeup_fd
= wakeup_fd
;
725 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
726 (unsigned long) &lum
->u
,
732 case LTTNG_UST_STREAM
:
734 /* Receive shm_fd, wakeup_fd */
735 ret
= ustcomm_recv_stream_from_sessiond(sock
,
738 &args
.stream
.wakeup_fd
);
743 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
744 (unsigned long) &lum
->u
,
752 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
753 (unsigned long) &lum
->u
,
761 lur
.handle
= lum
->handle
;
765 lur
.ret_code
= LTTNG_UST_OK
;
768 * Use -LTTNG_UST_ERR as wildcard for UST internal
769 * error that are not caused by the transport, except if
770 * we already have a more precise error message to
773 if (ret
> -LTTNG_UST_ERR
) {
774 /* Translate code to UST error. */
777 lur
.ret_code
= -LTTNG_UST_ERR_EXIST
;
780 lur
.ret_code
= -LTTNG_UST_ERR_INVAL
;
783 lur
.ret_code
= -LTTNG_UST_ERR_NOENT
;
786 lur
.ret_code
= -LTTNG_UST_ERR_PERM
;
789 lur
.ret_code
= -LTTNG_UST_ERR_NOSYS
;
792 lur
.ret_code
= -LTTNG_UST_ERR
;
801 case LTTNG_UST_TRACER_VERSION
:
802 lur
.u
.version
= lum
->u
.version
;
804 case LTTNG_UST_TRACEPOINT_LIST_GET
:
805 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
809 DBG("Return value: %d", lur
.ret_val
);
810 ret
= send_reply(sock
, &lur
);
812 DBG("error sending reply");
817 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
820 if (lur
.ret_code
== LTTNG_UST_OK
) {
822 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
823 len
= ustcomm_send_unix_sock(sock
,
824 &args
.field_list
.entry
,
825 sizeof(args
.field_list
.entry
));
830 if (len
!= sizeof(args
.field_list
.entry
)) {
841 * Performed delayed statedump operations outside of the UST
842 * lock. We need to take the dynamic loader lock before we take
843 * the UST lock internally within handle_pending_statedump().
845 handle_pending_statedump(sock_info
);
851 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
855 if (sock_info
->root_handle
!= -1) {
856 ret
= lttng_ust_objd_unref(sock_info
->root_handle
, 1);
858 ERR("Error unref root handle");
860 sock_info
->root_handle
= -1;
862 sock_info
->constructor_sem_posted
= 0;
865 * wait_shm_mmap, socket and notify socket are used by listener
866 * threads outside of the ust lock, so we cannot tear them down
867 * ourselves, because we cannot join on these threads. Leave
868 * responsibility of cleaning up these resources to the OS
874 if (sock_info
->socket
!= -1) {
875 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
877 ERR("Error closing ust cmd socket");
879 sock_info
->socket
= -1;
881 if (sock_info
->notify_socket
!= -1) {
882 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
884 ERR("Error closing ust notify socket");
886 sock_info
->notify_socket
= -1;
888 if (sock_info
->wait_shm_mmap
) {
891 page_size
= sysconf(_SC_PAGE_SIZE
);
893 ret
= munmap(sock_info
->wait_shm_mmap
, page_size
);
895 ERR("Error unmapping wait shm");
898 sock_info
->wait_shm_mmap
= NULL
;
903 * Using fork to set umask in the child process (not multi-thread safe).
904 * We deal with the shm_open vs ftruncate race (happening when the
905 * sessiond owns the shm and does not let everybody modify it, to ensure
906 * safety against shm_unlink) by simply letting the mmap fail and
907 * retrying after a few seconds.
908 * For global shm, everybody has rw access to it until the sessiond
912 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
914 int wait_shm_fd
, ret
;
918 * Try to open read-only.
920 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
921 if (wait_shm_fd
>= 0) {
924 size_t bytes_read
= 0;
927 * Try to read the fd. If unable to do so, try opening
931 len
= read(wait_shm_fd
,
932 &((char *) &tmp_read
)[bytes_read
],
933 sizeof(tmp_read
) - bytes_read
);
937 } while ((len
< 0 && errno
== EINTR
)
938 || (len
> 0 && bytes_read
< sizeof(tmp_read
)));
939 if (bytes_read
!= sizeof(tmp_read
)) {
940 ret
= close(wait_shm_fd
);
942 ERR("close wait_shm_fd");
947 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
949 * Real-only open did not work, and it's not because the
950 * entry was not present. It's a failure that prohibits
953 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
959 * If the open failed because the file did not exist, or because
960 * the file was not truncated yet, try creating it ourself.
962 URCU_TLS(lttng_ust_nest_count
)++;
964 URCU_TLS(lttng_ust_nest_count
)--;
969 * Parent: wait for child to return, in which case the
970 * shared memory map will have been created.
973 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
978 * Try to open read-only again after creation.
980 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
981 if (wait_shm_fd
< 0) {
983 * Real-only open did not work. It's a failure
984 * that prohibits using shm.
986 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
990 } else if (pid
== 0) {
994 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
995 if (sock_info
->global
)
996 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
998 * We're alone in a child process, so we can modify the
999 * process-wide umask.
1001 umask(~create_mode
);
1003 * Try creating shm (or get rw access).
1004 * We don't do an exclusive open, because we allow other
1005 * processes to create+ftruncate it concurrently.
1007 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
1008 O_RDWR
| O_CREAT
, create_mode
);
1009 if (wait_shm_fd
>= 0) {
1010 ret
= ftruncate(wait_shm_fd
, mmap_size
);
1012 PERROR("ftruncate");
1013 _exit(EXIT_FAILURE
);
1015 _exit(EXIT_SUCCESS
);
1018 * For local shm, we need to have rw access to accept
1019 * opening it: this means the local sessiond will be
1020 * able to wake us up. For global shm, we open it even
1021 * if rw access is not granted, because the root.root
1022 * sessiond will be able to override all rights and wake
1025 if (!sock_info
->global
&& errno
!= EACCES
) {
1026 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
1027 _exit(EXIT_FAILURE
);
1030 * The shm exists, but we cannot open it RW. Report
1033 _exit(EXIT_SUCCESS
);
1038 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
1039 struct stat statbuf
;
1042 * Ensure that our user is the owner of the shm file for
1043 * local shm. If we do not own the file, it means our
1044 * sessiond will not have access to wake us up (there is
1045 * probably a rogue process trying to fake our
1046 * sessiond). Fallback to polling method in this case.
1048 ret
= fstat(wait_shm_fd
, &statbuf
);
1053 if (statbuf
.st_uid
!= getuid())
1059 ret
= close(wait_shm_fd
);
1061 PERROR("Error closing fd");
1067 char *get_map_shm(struct sock_info
*sock_info
)
1070 int wait_shm_fd
, ret
;
1071 char *wait_shm_mmap
;
1073 page_size
= sysconf(_SC_PAGE_SIZE
);
1074 if (page_size
< 0) {
1078 wait_shm_fd
= get_wait_shm(sock_info
, page_size
);
1079 if (wait_shm_fd
< 0) {
1082 wait_shm_mmap
= mmap(NULL
, page_size
, PROT_READ
,
1083 MAP_SHARED
, wait_shm_fd
, 0);
1084 /* close shm fd immediately after taking the mmap reference */
1085 ret
= close(wait_shm_fd
);
1087 PERROR("Error closing fd");
1089 if (wait_shm_mmap
== MAP_FAILED
) {
1090 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1093 return wait_shm_mmap
;
1100 void wait_for_sessiond(struct sock_info
*sock_info
)
1107 if (wait_poll_fallback
) {
1110 if (!sock_info
->wait_shm_mmap
) {
1111 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
1112 if (!sock_info
->wait_shm_mmap
)
1117 DBG("Waiting for %s apps sessiond", sock_info
->name
);
1118 /* Wait for futex wakeup */
1119 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
1120 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
1121 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
1123 if (errno
== EFAULT
) {
1124 wait_poll_fallback
= 1;
1126 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1127 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1128 "Please upgrade your kernel "
1129 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1130 "mainline). LTTng-UST will use polling mode fallback.");
1148 * This thread does not allocate any resource, except within
1149 * handle_message, within mutex protection. This mutex protects against
1151 * The other moment it allocates resources is at socket connection, which
1152 * is also protected by the mutex.
1155 void *ust_listener_thread(void *arg
)
1157 struct sock_info
*sock_info
= arg
;
1158 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
1161 /* Restart trying to connect to the session daemon */
1163 if (prev_connect_failed
) {
1164 /* Wait for sessiond availability with pipe */
1165 wait_for_sessiond(sock_info
);
1169 * Sleep for 5 seconds before retrying after a
1170 * sequence of failure / wait / failure. This
1171 * deals with a killed or broken session daemon.
1176 prev_connect_failed
= 0;
1179 if (sock_info
->socket
!= -1) {
1180 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
1182 ERR("Error closing %s ust cmd socket",
1185 sock_info
->socket
= -1;
1187 if (sock_info
->notify_socket
!= -1) {
1188 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
1190 ERR("Error closing %s ust notify socket",
1193 sock_info
->notify_socket
= -1;
1197 * Register. We need to perform both connect and sending
1198 * registration message before doing the next connect otherwise
1199 * we may reach unix socket connect queue max limits and block
1200 * on the 2nd connect while the session daemon is awaiting the
1201 * first connect registration message.
1203 /* Connect cmd socket */
1204 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1206 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1207 prev_connect_failed
= 1;
1214 * If we cannot find the sessiond daemon, don't delay
1215 * constructor execution.
1217 ret
= handle_register_done(sock_info
);
1222 sock_info
->socket
= ret
;
1229 * Create only one root handle per listener thread for the whole
1230 * process lifetime, so we ensure we get ID which is statically
1231 * assigned to the root handle.
1233 if (sock_info
->root_handle
== -1) {
1234 ret
= lttng_abi_create_root_handle();
1236 ERR("Error creating root handle");
1239 sock_info
->root_handle
= ret
;
1242 ret
= register_to_sessiond(sock_info
->socket
, USTCTL_SOCKET_CMD
);
1244 ERR("Error registering to %s ust cmd socket",
1246 prev_connect_failed
= 1;
1248 * If we cannot register to the sessiond daemon, don't
1249 * delay constructor execution.
1251 ret
= handle_register_done(sock_info
);
1259 /* Connect notify socket */
1260 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1262 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1263 prev_connect_failed
= 1;
1270 * If we cannot find the sessiond daemon, don't delay
1271 * constructor execution.
1273 ret
= handle_register_done(sock_info
);
1278 sock_info
->notify_socket
= ret
;
1280 timeout
= get_notify_sock_timeout();
1283 * Give at least 10ms to sessiond to reply to
1288 ret
= ustcomm_setsockopt_rcv_timeout(sock_info
->notify_socket
,
1291 WARN("Error setting socket receive timeout");
1293 ret
= ustcomm_setsockopt_snd_timeout(sock_info
->notify_socket
,
1296 WARN("Error setting socket send timeout");
1298 } else if (timeout
< -1) {
1299 WARN("Unsupported timeout value %ld", timeout
);
1306 ret
= register_to_sessiond(sock_info
->notify_socket
,
1307 USTCTL_SOCKET_NOTIFY
);
1309 ERR("Error registering to %s ust notify socket",
1311 prev_connect_failed
= 1;
1313 * If we cannot register to the sessiond daemon, don't
1314 * delay constructor execution.
1316 ret
= handle_register_done(sock_info
);
1321 sock
= sock_info
->socket
;
1327 struct ustcomm_ust_msg lum
;
1329 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
1331 case 0: /* orderly shutdown */
1332 DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info
->name
);
1337 * Either sessiond has shutdown or refused us by closing the socket.
1338 * In either case, we don't want to delay construction execution,
1339 * and we need to wait before retry.
1341 prev_connect_failed
= 1;
1343 * If we cannot register to the sessiond daemon, don't
1344 * delay constructor execution.
1346 ret
= handle_register_done(sock_info
);
1351 print_cmd(lum
.cmd
, lum
.handle
);
1352 ret
= handle_message(sock_info
, sock
, &lum
);
1354 ERR("Error handling message for %s socket", sock_info
->name
);
1359 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
1361 DBG("incorrect message size (%s socket): %zd", sock_info
->name
, len
);
1363 if (len
== -ECONNRESET
) {
1364 DBG("%s remote end closed connection", sock_info
->name
);
1375 /* Cleanup socket handles before trying to reconnect */
1376 lttng_ust_objd_table_owner_cleanup(sock_info
);
1378 goto restart
; /* try to reconnect */
1383 pthread_mutex_lock(&ust_exit_mutex
);
1384 sock_info
->thread_active
= 0;
1385 pthread_mutex_unlock(&ust_exit_mutex
);
1390 * Weak symbol to call when the ust malloc wrapper is not loaded.
1392 __attribute__((weak
))
1393 void lttng_ust_malloc_wrapper_init(void)
1398 * sessiond monitoring thread: monitor presence of global and per-user
1399 * sessiond by polling the application common named pipe.
1401 void __attribute__((constructor
)) lttng_ust_init(void)
1403 struct timespec constructor_timeout
;
1404 sigset_t sig_all_blocked
, orig_parent_mask
;
1405 pthread_attr_t thread_attr
;
1409 if (uatomic_xchg(&initialized
, 1) == 1)
1413 * Fixup interdependency between TLS fixup mutex (which happens
1414 * to be the dynamic linker mutex) and ust_lock, taken within
1417 lttng_fixup_ringbuffer_tls();
1418 lttng_fixup_vtid_tls();
1419 lttng_fixup_nest_count_tls();
1420 lttng_fixup_procname_tls();
1421 lttng_fixup_ust_mutex_nest_tls();
1424 * We want precise control over the order in which we construct
1425 * our sub-libraries vs starting to receive commands from
1426 * sessiond (otherwise leading to errors when trying to create
1427 * sessiond before the init functions are completed).
1431 lttng_ust_baddr_statedump_init();
1432 lttng_ring_buffer_metadata_client_init();
1433 lttng_ring_buffer_client_overwrite_init();
1434 lttng_ring_buffer_client_overwrite_rt_init();
1435 lttng_ring_buffer_client_discard_init();
1436 lttng_ring_buffer_client_discard_rt_init();
1437 lttng_perf_counter_init();
1438 lttng_context_init();
1440 * Invoke ust malloc wrapper init before starting other threads.
1442 lttng_ust_malloc_wrapper_init();
1444 timeout_mode
= get_constructor_timeout(&constructor_timeout
);
1446 ret
= sem_init(&constructor_wait
, 0, 0);
1449 ret
= setup_local_apps();
1451 DBG("local apps setup returned %d", ret
);
1454 /* A new thread created by pthread_create inherits the signal mask
1455 * from the parent. To avoid any signal being received by the
1456 * listener thread, we block all signals temporarily in the parent,
1457 * while we create the listener thread.
1459 sigfillset(&sig_all_blocked
);
1460 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
1462 ERR("pthread_sigmask: %s", strerror(ret
));
1465 ret
= pthread_attr_init(&thread_attr
);
1467 ERR("pthread_attr_init: %s", strerror(ret
));
1469 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
1471 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1474 pthread_mutex_lock(&ust_exit_mutex
);
1475 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1476 ust_listener_thread
, &global_apps
);
1478 ERR("pthread_create global: %s", strerror(ret
));
1480 global_apps
.thread_active
= 1;
1481 pthread_mutex_unlock(&ust_exit_mutex
);
1483 if (local_apps
.allowed
) {
1484 pthread_mutex_lock(&ust_exit_mutex
);
1485 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1486 ust_listener_thread
, &local_apps
);
1488 ERR("pthread_create local: %s", strerror(ret
));
1490 local_apps
.thread_active
= 1;
1491 pthread_mutex_unlock(&ust_exit_mutex
);
1493 handle_register_done(&local_apps
);
1495 ret
= pthread_attr_destroy(&thread_attr
);
1497 ERR("pthread_attr_destroy: %s", strerror(ret
));
1500 /* Restore original signal mask in parent */
1501 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1503 ERR("pthread_sigmask: %s", strerror(ret
));
1506 switch (timeout_mode
) {
1507 case 1: /* timeout wait */
1509 ret
= sem_timedwait(&constructor_wait
,
1510 &constructor_timeout
);
1511 } while (ret
< 0 && errno
== EINTR
);
1512 if (ret
< 0 && errno
== ETIMEDOUT
) {
1513 ERR("Timed out waiting for lttng-sessiond");
1518 case -1:/* wait forever */
1520 ret
= sem_wait(&constructor_wait
);
1521 } while (ret
< 0 && errno
== EINTR
);
1524 case 0: /* no timeout */
1530 void lttng_ust_cleanup(int exiting
)
1532 cleanup_sock_info(&global_apps
, exiting
);
1533 if (local_apps
.allowed
) {
1534 cleanup_sock_info(&local_apps
, exiting
);
1537 * The teardown in this function all affect data structures
1538 * accessed under the UST lock by the listener thread. This
1539 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1540 * that none of these threads are accessing this data at this
1543 lttng_ust_abi_exit();
1544 lttng_ust_events_exit();
1545 lttng_context_exit();
1546 lttng_perf_counter_exit();
1547 lttng_ring_buffer_client_discard_rt_exit();
1548 lttng_ring_buffer_client_discard_exit();
1549 lttng_ring_buffer_client_overwrite_rt_exit();
1550 lttng_ring_buffer_client_overwrite_exit();
1551 lttng_ring_buffer_metadata_client_exit();
1552 lttng_ust_baddr_statedump_destroy();
1555 /* Reinitialize values for fork */
1557 lttng_ust_comm_should_quit
= 0;
1562 void __attribute__((destructor
)) lttng_ust_exit(void)
1567 * Using pthread_cancel here because:
1568 * A) we don't want to hang application teardown.
1569 * B) the thread is not allocating any resource.
1573 * Require the communication thread to quit. Synchronize with
1574 * mutexes to ensure it is not in a mutex critical section when
1575 * pthread_cancel is later called.
1578 lttng_ust_comm_should_quit
= 1;
1581 pthread_mutex_lock(&ust_exit_mutex
);
1582 /* cancel threads */
1583 if (global_apps
.thread_active
) {
1584 ret
= pthread_cancel(global_apps
.ust_listener
);
1586 ERR("Error cancelling global ust listener thread: %s",
1589 global_apps
.thread_active
= 0;
1592 if (local_apps
.thread_active
) {
1593 ret
= pthread_cancel(local_apps
.ust_listener
);
1595 ERR("Error cancelling local ust listener thread: %s",
1598 local_apps
.thread_active
= 0;
1601 pthread_mutex_unlock(&ust_exit_mutex
);
1604 * Do NOT join threads: use of sys_futex makes it impossible to
1605 * join the threads without using async-cancel, but async-cancel
1606 * is delivered by a signal, which could hit the target thread
1607 * anywhere in its code path, including while the ust_lock() is
1608 * held, causing a deadlock for the other thread. Let the OS
1609 * cleanup the threads if there are stalled in a syscall.
1611 lttng_ust_cleanup(1);
1615 * We exclude the worker threads across fork and clone (except
1616 * CLONE_VM), because these system calls only keep the forking thread
1617 * running in the child. Therefore, we don't want to call fork or clone
1618 * in the middle of an tracepoint or ust tracing state modification.
1619 * Holding this mutex protects these structures across fork and clone.
1621 void ust_before_fork(sigset_t
*save_sigset
)
1624 * Disable signals. This is to avoid that the child intervenes
1625 * before it is properly setup for tracing. It is safer to
1626 * disable all signals, because then we know we are not breaking
1627 * anything by restoring the original mask.
1632 if (URCU_TLS(lttng_ust_nest_count
))
1634 /* Disable signals */
1635 sigfillset(&all_sigs
);
1636 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1638 PERROR("sigprocmask");
1641 pthread_mutex_lock(&ust_fork_mutex
);
1644 rcu_bp_before_fork();
1647 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1651 DBG("process %d", getpid());
1654 pthread_mutex_unlock(&ust_fork_mutex
);
1656 /* Restore signals */
1657 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1659 PERROR("sigprocmask");
1663 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1665 if (URCU_TLS(lttng_ust_nest_count
))
1667 DBG("process %d", getpid());
1668 rcu_bp_after_fork_parent();
1669 /* Release mutexes and reenable signals */
1670 ust_after_fork_common(restore_sigset
);
1674 * After fork, in the child, we need to cleanup all the leftover state,
1675 * except the worker thread which already magically disappeared thanks
1676 * to the weird Linux fork semantics. After tyding up, we call
1677 * lttng_ust_init() again to start over as a new PID.
1679 * This is meant for forks() that have tracing in the child between the
1680 * fork and following exec call (if there is any).
1682 void ust_after_fork_child(sigset_t
*restore_sigset
)
1684 if (URCU_TLS(lttng_ust_nest_count
))
1686 DBG("process %d", getpid());
1687 /* Release urcu mutexes */
1688 rcu_bp_after_fork_child();
1689 lttng_ust_cleanup(0);
1690 lttng_context_vtid_reset();
1691 /* Release mutexes and reenable signals */
1692 ust_after_fork_common(restore_sigset
);
1696 void lttng_ust_sockinfo_session_enabled(void *owner
)
1698 struct sock_info
*sock_info
= owner
;
1699 sock_info
->statedump_pending
= 1;