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
)));
338 int lttng_get_notify_socket(void *owner
)
340 struct sock_info
*info
= owner
;
342 return info
->notify_socket
;
346 void print_cmd(int cmd
, int handle
)
348 const char *cmd_name
= "Unknown";
350 if (cmd
>= 0 && cmd
< LTTNG_ARRAY_SIZE(cmd_name_mapping
)
351 && cmd_name_mapping
[cmd
]) {
352 cmd_name
= cmd_name_mapping
[cmd
];
354 DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
356 lttng_ust_obj_get_name(handle
), handle
);
360 int setup_local_apps(void)
362 const char *home_dir
;
367 * Disallow per-user tracing for setuid binaries.
369 if (uid
!= geteuid()) {
370 assert(local_apps
.allowed
== 0);
373 home_dir
= get_lttng_home_dir();
375 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
376 assert(local_apps
.allowed
== 0);
379 local_apps
.allowed
= 1;
380 snprintf(local_apps
.sock_path
, PATH_MAX
, "%s/%s/%s",
382 LTTNG_DEFAULT_HOME_RUNDIR
,
383 LTTNG_UST_SOCK_FILENAME
);
384 snprintf(local_apps
.wait_shm_path
, PATH_MAX
, "/%s-%u",
385 LTTNG_UST_WAIT_FILENAME
,
391 * Get notify_sock timeout, in ms.
392 * -1: don't wait. 0: wait forever. >0: timeout, in ms.
395 long get_timeout(void)
397 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
399 if (!got_timeout_env
) {
400 str_timeout
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
404 constructor_delay_ms
= strtol(str_timeout
, NULL
, 10);
405 return constructor_delay_ms
;
409 long get_notify_sock_timeout(void)
411 return get_timeout();
415 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
418 int get_constructor_timeout(struct timespec
*constructor_timeout
)
420 long constructor_delay_ms
;
423 constructor_delay_ms
= get_timeout();
425 switch (constructor_delay_ms
) {
426 case -1:/* fall-through */
428 return constructor_delay_ms
;
434 * If we are unable to find the current time, don't wait.
436 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
440 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
441 constructor_timeout
->tv_nsec
+=
442 (constructor_delay_ms
% 1000UL) * 1000000UL;
443 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
444 constructor_timeout
->tv_sec
++;
445 constructor_timeout
->tv_nsec
-= 1000000000UL;
451 int register_to_sessiond(int socket
, enum ustctl_socket_type type
)
453 return ustcomm_send_reg_msg(socket
,
456 lttng_alignof(uint8_t) * CHAR_BIT
,
457 lttng_alignof(uint16_t) * CHAR_BIT
,
458 lttng_alignof(uint32_t) * CHAR_BIT
,
459 lttng_alignof(uint64_t) * CHAR_BIT
,
460 lttng_alignof(unsigned long) * CHAR_BIT
);
464 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
468 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
471 DBG("message successfully sent");
474 if (len
== -ECONNRESET
) {
475 DBG("remote end closed connection");
480 DBG("incorrect message size: %zd", len
);
486 int handle_register_done(struct sock_info
*sock_info
)
490 if (sock_info
->constructor_sem_posted
)
492 sock_info
->constructor_sem_posted
= 1;
493 if (uatomic_read(&sem_count
) <= 0) {
496 ret
= uatomic_add_return(&sem_count
, -1);
498 ret
= sem_post(&constructor_wait
);
505 * Only execute pending statedump after the constructor semaphore has
506 * been posted by each listener thread. This means statedump will only
507 * be performed after the "registration done" command is received from
508 * each session daemon the application is connected to.
510 * This ensures we don't run into deadlock issues with the dynamic
511 * loader mutex, which is held while the constructor is called and
512 * waiting on the constructor semaphore. All operations requiring this
513 * dynamic loader lock need to be postponed using this mechanism.
516 void handle_pending_statedump(struct sock_info
*sock_info
)
518 int ctor_passed
= sock_info
->constructor_sem_posted
;
520 if (ctor_passed
&& sock_info
->statedump_pending
) {
521 sock_info
->statedump_pending
= 0;
522 pthread_mutex_lock(&ust_fork_mutex
);
523 lttng_handle_pending_statedump(sock_info
);
524 pthread_mutex_unlock(&ust_fork_mutex
);
529 int handle_message(struct sock_info
*sock_info
,
530 int sock
, struct ustcomm_ust_msg
*lum
)
533 const struct lttng_ust_objd_ops
*ops
;
534 struct ustcomm_ust_reply lur
;
538 memset(&lur
, 0, sizeof(lur
));
541 ret
= -LTTNG_UST_ERR_EXITING
;
545 ops
= objd_ops(lum
->handle
);
552 case LTTNG_UST_REGISTER_DONE
:
553 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
554 ret
= handle_register_done(sock_info
);
558 case LTTNG_UST_RELEASE
:
559 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
562 ret
= lttng_ust_objd_unref(lum
->handle
, 1);
564 case LTTNG_UST_FILTER
:
566 /* Receive filter data */
567 struct lttng_ust_filter_bytecode_node
*bytecode
;
569 if (lum
->u
.filter
.data_size
> FILTER_BYTECODE_MAX_LEN
) {
570 ERR("Filter data size is too large: %u bytes",
571 lum
->u
.filter
.data_size
);
576 if (lum
->u
.filter
.reloc_offset
> lum
->u
.filter
.data_size
) {
577 ERR("Filter reloc offset %u is not within data",
578 lum
->u
.filter
.reloc_offset
);
583 bytecode
= zmalloc(sizeof(*bytecode
) + lum
->u
.filter
.data_size
);
588 len
= ustcomm_recv_unix_sock(sock
, bytecode
->bc
.data
,
589 lum
->u
.filter
.data_size
);
591 case 0: /* orderly shutdown */
596 if (len
== lum
->u
.filter
.data_size
) {
597 DBG("filter data received");
599 } else if (len
< 0) {
600 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
601 if (len
== -ECONNRESET
) {
602 ERR("%s remote end closed connection", sock_info
->name
);
611 DBG("incorrect filter data message size: %zd", len
);
617 bytecode
->bc
.len
= lum
->u
.filter
.data_size
;
618 bytecode
->bc
.reloc_offset
= lum
->u
.filter
.reloc_offset
;
619 bytecode
->bc
.seqnum
= lum
->u
.filter
.seqnum
;
621 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
622 (unsigned long) bytecode
,
627 /* don't free bytecode if everything went fine. */
634 case LTTNG_UST_EXCLUSION
:
636 /* Receive exclusion names */
637 struct lttng_ust_excluder_node
*node
;
640 count
= lum
->u
.exclusion
.count
;
642 /* There are no names to read */
646 node
= zmalloc(sizeof(*node
) +
647 count
* LTTNG_UST_SYM_NAME_LEN
);
652 node
->excluder
.count
= count
;
653 len
= ustcomm_recv_unix_sock(sock
, node
->excluder
.names
,
654 count
* LTTNG_UST_SYM_NAME_LEN
);
656 case 0: /* orderly shutdown */
661 if (len
== count
* LTTNG_UST_SYM_NAME_LEN
) {
662 DBG("Exclusion data received");
664 } else if (len
< 0) {
665 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
666 if (len
== -ECONNRESET
) {
667 ERR("%s remote end closed connection", sock_info
->name
);
676 DBG("Incorrect exclusion data message size: %zd", len
);
683 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
684 (unsigned long) node
,
689 /* Don't free exclusion data if everything went fine. */
696 case LTTNG_UST_CHANNEL
:
701 len
= ustcomm_recv_channel_from_sessiond(sock
,
702 &chan_data
, lum
->u
.channel
.len
,
705 case 0: /* orderly shutdown */
709 if (len
== lum
->u
.channel
.len
) {
710 DBG("channel data received");
712 } else if (len
< 0) {
713 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
714 if (len
== -ECONNRESET
) {
715 ERR("%s remote end closed connection", sock_info
->name
);
722 DBG("incorrect channel data message size: %zd", len
);
727 args
.channel
.chan_data
= chan_data
;
728 args
.channel
.wakeup_fd
= wakeup_fd
;
730 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
731 (unsigned long) &lum
->u
,
737 case LTTNG_UST_STREAM
:
739 /* Receive shm_fd, wakeup_fd */
740 ret
= ustcomm_recv_stream_from_sessiond(sock
,
743 &args
.stream
.wakeup_fd
);
748 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
749 (unsigned long) &lum
->u
,
757 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
758 (unsigned long) &lum
->u
,
766 lur
.handle
= lum
->handle
;
770 lur
.ret_code
= LTTNG_UST_OK
;
773 * Use -LTTNG_UST_ERR as wildcard for UST internal
774 * error that are not caused by the transport, except if
775 * we already have a more precise error message to
778 if (ret
> -LTTNG_UST_ERR
) {
779 /* Translate code to UST error. */
782 lur
.ret_code
= -LTTNG_UST_ERR_EXIST
;
785 lur
.ret_code
= -LTTNG_UST_ERR_INVAL
;
788 lur
.ret_code
= -LTTNG_UST_ERR_NOENT
;
791 lur
.ret_code
= -LTTNG_UST_ERR_PERM
;
794 lur
.ret_code
= -LTTNG_UST_ERR_NOSYS
;
797 lur
.ret_code
= -LTTNG_UST_ERR
;
806 case LTTNG_UST_TRACER_VERSION
:
807 lur
.u
.version
= lum
->u
.version
;
809 case LTTNG_UST_TRACEPOINT_LIST_GET
:
810 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
814 DBG("Return value: %d", lur
.ret_val
);
815 ret
= send_reply(sock
, &lur
);
817 DBG("error sending reply");
822 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
825 if (lur
.ret_code
== LTTNG_UST_OK
) {
827 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
828 len
= ustcomm_send_unix_sock(sock
,
829 &args
.field_list
.entry
,
830 sizeof(args
.field_list
.entry
));
835 if (len
!= sizeof(args
.field_list
.entry
)) {
846 * Performed delayed statedump operations outside of the UST
847 * lock. We need to take the dynamic loader lock before we take
848 * the UST lock internally within handle_pending_statedump().
850 handle_pending_statedump(sock_info
);
856 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
860 if (sock_info
->root_handle
!= -1) {
861 ret
= lttng_ust_objd_unref(sock_info
->root_handle
, 1);
863 ERR("Error unref root handle");
865 sock_info
->root_handle
= -1;
867 sock_info
->constructor_sem_posted
= 0;
870 * wait_shm_mmap, socket and notify socket are used by listener
871 * threads outside of the ust lock, so we cannot tear them down
872 * ourselves, because we cannot join on these threads. Leave
873 * responsibility of cleaning up these resources to the OS
879 if (sock_info
->socket
!= -1) {
880 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
882 ERR("Error closing ust cmd socket");
884 sock_info
->socket
= -1;
886 if (sock_info
->notify_socket
!= -1) {
887 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
889 ERR("Error closing ust notify socket");
891 sock_info
->notify_socket
= -1;
893 if (sock_info
->wait_shm_mmap
) {
896 page_size
= sysconf(_SC_PAGE_SIZE
);
898 ret
= munmap(sock_info
->wait_shm_mmap
, page_size
);
900 ERR("Error unmapping wait shm");
903 sock_info
->wait_shm_mmap
= NULL
;
908 * Using fork to set umask in the child process (not multi-thread safe).
909 * We deal with the shm_open vs ftruncate race (happening when the
910 * sessiond owns the shm and does not let everybody modify it, to ensure
911 * safety against shm_unlink) by simply letting the mmap fail and
912 * retrying after a few seconds.
913 * For global shm, everybody has rw access to it until the sessiond
917 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
919 int wait_shm_fd
, ret
;
923 * Try to open read-only.
925 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
926 if (wait_shm_fd
>= 0) {
929 size_t bytes_read
= 0;
932 * Try to read the fd. If unable to do so, try opening
936 len
= read(wait_shm_fd
,
937 &((char *) &tmp_read
)[bytes_read
],
938 sizeof(tmp_read
) - bytes_read
);
942 } while ((len
< 0 && errno
== EINTR
)
943 || (len
> 0 && bytes_read
< sizeof(tmp_read
)));
944 if (bytes_read
!= sizeof(tmp_read
)) {
945 ret
= close(wait_shm_fd
);
947 ERR("close wait_shm_fd");
952 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
954 * Real-only open did not work, and it's not because the
955 * entry was not present. It's a failure that prohibits
958 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
964 * If the open failed because the file did not exist, or because
965 * the file was not truncated yet, try creating it ourself.
967 URCU_TLS(lttng_ust_nest_count
)++;
969 URCU_TLS(lttng_ust_nest_count
)--;
974 * Parent: wait for child to return, in which case the
975 * shared memory map will have been created.
978 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
983 * Try to open read-only again after creation.
985 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
986 if (wait_shm_fd
< 0) {
988 * Real-only open did not work. It's a failure
989 * that prohibits using shm.
991 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
995 } else if (pid
== 0) {
999 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
1000 if (sock_info
->global
)
1001 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
1003 * We're alone in a child process, so we can modify the
1004 * process-wide umask.
1006 umask(~create_mode
);
1008 * Try creating shm (or get rw access).
1009 * We don't do an exclusive open, because we allow other
1010 * processes to create+ftruncate it concurrently.
1012 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
1013 O_RDWR
| O_CREAT
, create_mode
);
1014 if (wait_shm_fd
>= 0) {
1015 ret
= ftruncate(wait_shm_fd
, mmap_size
);
1017 PERROR("ftruncate");
1018 _exit(EXIT_FAILURE
);
1020 _exit(EXIT_SUCCESS
);
1023 * For local shm, we need to have rw access to accept
1024 * opening it: this means the local sessiond will be
1025 * able to wake us up. For global shm, we open it even
1026 * if rw access is not granted, because the root.root
1027 * sessiond will be able to override all rights and wake
1030 if (!sock_info
->global
&& errno
!= EACCES
) {
1031 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
1032 _exit(EXIT_FAILURE
);
1035 * The shm exists, but we cannot open it RW. Report
1038 _exit(EXIT_SUCCESS
);
1043 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
1044 struct stat statbuf
;
1047 * Ensure that our user is the owner of the shm file for
1048 * local shm. If we do not own the file, it means our
1049 * sessiond will not have access to wake us up (there is
1050 * probably a rogue process trying to fake our
1051 * sessiond). Fallback to polling method in this case.
1053 ret
= fstat(wait_shm_fd
, &statbuf
);
1058 if (statbuf
.st_uid
!= getuid())
1064 ret
= close(wait_shm_fd
);
1066 PERROR("Error closing fd");
1072 char *get_map_shm(struct sock_info
*sock_info
)
1075 int wait_shm_fd
, ret
;
1076 char *wait_shm_mmap
;
1078 page_size
= sysconf(_SC_PAGE_SIZE
);
1079 if (page_size
< 0) {
1083 wait_shm_fd
= get_wait_shm(sock_info
, page_size
);
1084 if (wait_shm_fd
< 0) {
1087 wait_shm_mmap
= mmap(NULL
, page_size
, PROT_READ
,
1088 MAP_SHARED
, wait_shm_fd
, 0);
1089 /* close shm fd immediately after taking the mmap reference */
1090 ret
= close(wait_shm_fd
);
1092 PERROR("Error closing fd");
1094 if (wait_shm_mmap
== MAP_FAILED
) {
1095 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1098 return wait_shm_mmap
;
1105 void wait_for_sessiond(struct sock_info
*sock_info
)
1112 if (wait_poll_fallback
) {
1115 if (!sock_info
->wait_shm_mmap
) {
1116 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
1117 if (!sock_info
->wait_shm_mmap
)
1122 DBG("Waiting for %s apps sessiond", sock_info
->name
);
1123 /* Wait for futex wakeup */
1124 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
1125 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
1126 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
1128 if (errno
== EFAULT
) {
1129 wait_poll_fallback
= 1;
1131 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1132 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1133 "Please upgrade your kernel "
1134 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1135 "mainline). LTTng-UST will use polling mode fallback.");
1153 * This thread does not allocate any resource, except within
1154 * handle_message, within mutex protection. This mutex protects against
1156 * The other moment it allocates resources is at socket connection, which
1157 * is also protected by the mutex.
1160 void *ust_listener_thread(void *arg
)
1162 struct sock_info
*sock_info
= arg
;
1163 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
1166 /* Restart trying to connect to the session daemon */
1168 if (prev_connect_failed
) {
1169 /* Wait for sessiond availability with pipe */
1170 wait_for_sessiond(sock_info
);
1174 * Sleep for 5 seconds before retrying after a
1175 * sequence of failure / wait / failure. This
1176 * deals with a killed or broken session daemon.
1181 prev_connect_failed
= 0;
1184 if (sock_info
->socket
!= -1) {
1185 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
1187 ERR("Error closing %s ust cmd socket",
1190 sock_info
->socket
= -1;
1192 if (sock_info
->notify_socket
!= -1) {
1193 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
1195 ERR("Error closing %s ust notify socket",
1198 sock_info
->notify_socket
= -1;
1202 * Register. We need to perform both connect and sending
1203 * registration message before doing the next connect otherwise
1204 * we may reach unix socket connect queue max limits and block
1205 * on the 2nd connect while the session daemon is awaiting the
1206 * first connect registration message.
1208 /* Connect cmd socket */
1209 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1211 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1212 prev_connect_failed
= 1;
1219 * If we cannot find the sessiond daemon, don't delay
1220 * constructor execution.
1222 ret
= handle_register_done(sock_info
);
1227 sock_info
->socket
= ret
;
1234 * Create only one root handle per listener thread for the whole
1235 * process lifetime, so we ensure we get ID which is statically
1236 * assigned to the root handle.
1238 if (sock_info
->root_handle
== -1) {
1239 ret
= lttng_abi_create_root_handle();
1241 ERR("Error creating root handle");
1244 sock_info
->root_handle
= ret
;
1247 ret
= register_to_sessiond(sock_info
->socket
, USTCTL_SOCKET_CMD
);
1249 ERR("Error registering to %s ust cmd socket",
1251 prev_connect_failed
= 1;
1253 * If we cannot register to the sessiond daemon, don't
1254 * delay constructor execution.
1256 ret
= handle_register_done(sock_info
);
1264 /* Connect notify socket */
1265 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1267 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1268 prev_connect_failed
= 1;
1275 * If we cannot find the sessiond daemon, don't delay
1276 * constructor execution.
1278 ret
= handle_register_done(sock_info
);
1283 sock_info
->notify_socket
= ret
;
1285 timeout
= get_notify_sock_timeout();
1288 * Give at least 10ms to sessiond to reply to
1293 ret
= ustcomm_setsockopt_rcv_timeout(sock_info
->notify_socket
,
1296 WARN("Error setting socket receive timeout");
1298 ret
= ustcomm_setsockopt_snd_timeout(sock_info
->notify_socket
,
1301 WARN("Error setting socket send timeout");
1303 } else if (timeout
< -1) {
1304 WARN("Unsupported timeout value %ld", timeout
);
1311 ret
= register_to_sessiond(sock_info
->notify_socket
,
1312 USTCTL_SOCKET_NOTIFY
);
1314 ERR("Error registering to %s ust notify socket",
1316 prev_connect_failed
= 1;
1318 * If we cannot register to the sessiond daemon, don't
1319 * delay constructor execution.
1321 ret
= handle_register_done(sock_info
);
1326 sock
= sock_info
->socket
;
1332 struct ustcomm_ust_msg lum
;
1334 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
1336 case 0: /* orderly shutdown */
1337 DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info
->name
);
1342 * Either sessiond has shutdown or refused us by closing the socket.
1343 * In either case, we don't want to delay construction execution,
1344 * and we need to wait before retry.
1346 prev_connect_failed
= 1;
1348 * If we cannot register to the sessiond daemon, don't
1349 * delay constructor execution.
1351 ret
= handle_register_done(sock_info
);
1356 print_cmd(lum
.cmd
, lum
.handle
);
1357 ret
= handle_message(sock_info
, sock
, &lum
);
1359 ERR("Error handling message for %s socket", sock_info
->name
);
1364 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
1366 DBG("incorrect message size (%s socket): %zd", sock_info
->name
, len
);
1368 if (len
== -ECONNRESET
) {
1369 DBG("%s remote end closed connection", sock_info
->name
);
1380 /* Cleanup socket handles before trying to reconnect */
1381 lttng_ust_objd_table_owner_cleanup(sock_info
);
1383 goto restart
; /* try to reconnect */
1388 pthread_mutex_lock(&ust_exit_mutex
);
1389 sock_info
->thread_active
= 0;
1390 pthread_mutex_unlock(&ust_exit_mutex
);
1395 * Weak symbol to call when the ust malloc wrapper is not loaded.
1397 __attribute__((weak
))
1398 void lttng_ust_malloc_wrapper_init(void)
1403 * sessiond monitoring thread: monitor presence of global and per-user
1404 * sessiond by polling the application common named pipe.
1406 void __attribute__((constructor
)) lttng_ust_init(void)
1408 struct timespec constructor_timeout
;
1409 sigset_t sig_all_blocked
, orig_parent_mask
;
1410 pthread_attr_t thread_attr
;
1414 if (uatomic_xchg(&initialized
, 1) == 1)
1418 * Fixup interdependency between TLS fixup mutex (which happens
1419 * to be the dynamic linker mutex) and ust_lock, taken within
1422 lttng_fixup_ringbuffer_tls();
1423 lttng_fixup_vtid_tls();
1424 lttng_fixup_nest_count_tls();
1425 lttng_fixup_procname_tls();
1426 lttng_fixup_ust_mutex_nest_tls();
1429 * We want precise control over the order in which we construct
1430 * our sub-libraries vs starting to receive commands from
1431 * sessiond (otherwise leading to errors when trying to create
1432 * sessiond before the init functions are completed).
1436 lttng_ust_baddr_statedump_init();
1437 lttng_ring_buffer_metadata_client_init();
1438 lttng_ring_buffer_client_overwrite_init();
1439 lttng_ring_buffer_client_overwrite_rt_init();
1440 lttng_ring_buffer_client_discard_init();
1441 lttng_ring_buffer_client_discard_rt_init();
1442 lttng_perf_counter_init();
1443 lttng_context_init();
1445 * Invoke ust malloc wrapper init before starting other threads.
1447 lttng_ust_malloc_wrapper_init();
1449 timeout_mode
= get_constructor_timeout(&constructor_timeout
);
1451 ret
= sem_init(&constructor_wait
, 0, 0);
1454 ret
= setup_local_apps();
1456 DBG("local apps setup returned %d", ret
);
1459 /* A new thread created by pthread_create inherits the signal mask
1460 * from the parent. To avoid any signal being received by the
1461 * listener thread, we block all signals temporarily in the parent,
1462 * while we create the listener thread.
1464 sigfillset(&sig_all_blocked
);
1465 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
1467 ERR("pthread_sigmask: %s", strerror(ret
));
1470 ret
= pthread_attr_init(&thread_attr
);
1472 ERR("pthread_attr_init: %s", strerror(ret
));
1474 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
1476 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1479 pthread_mutex_lock(&ust_exit_mutex
);
1480 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1481 ust_listener_thread
, &global_apps
);
1483 ERR("pthread_create global: %s", strerror(ret
));
1485 global_apps
.thread_active
= 1;
1486 pthread_mutex_unlock(&ust_exit_mutex
);
1488 if (local_apps
.allowed
) {
1489 pthread_mutex_lock(&ust_exit_mutex
);
1490 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1491 ust_listener_thread
, &local_apps
);
1493 ERR("pthread_create local: %s", strerror(ret
));
1495 local_apps
.thread_active
= 1;
1496 pthread_mutex_unlock(&ust_exit_mutex
);
1498 handle_register_done(&local_apps
);
1500 ret
= pthread_attr_destroy(&thread_attr
);
1502 ERR("pthread_attr_destroy: %s", strerror(ret
));
1505 /* Restore original signal mask in parent */
1506 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1508 ERR("pthread_sigmask: %s", strerror(ret
));
1511 switch (timeout_mode
) {
1512 case 1: /* timeout wait */
1514 ret
= sem_timedwait(&constructor_wait
,
1515 &constructor_timeout
);
1516 } while (ret
< 0 && errno
== EINTR
);
1517 if (ret
< 0 && errno
== ETIMEDOUT
) {
1518 ERR("Timed out waiting for lttng-sessiond");
1523 case -1:/* wait forever */
1525 ret
= sem_wait(&constructor_wait
);
1526 } while (ret
< 0 && errno
== EINTR
);
1529 case 0: /* no timeout */
1535 void lttng_ust_cleanup(int exiting
)
1537 cleanup_sock_info(&global_apps
, exiting
);
1538 if (local_apps
.allowed
) {
1539 cleanup_sock_info(&local_apps
, exiting
);
1542 * The teardown in this function all affect data structures
1543 * accessed under the UST lock by the listener thread. This
1544 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1545 * that none of these threads are accessing this data at this
1548 lttng_ust_abi_exit();
1549 lttng_ust_events_exit();
1550 lttng_context_exit();
1551 lttng_perf_counter_exit();
1552 lttng_ring_buffer_client_discard_rt_exit();
1553 lttng_ring_buffer_client_discard_exit();
1554 lttng_ring_buffer_client_overwrite_rt_exit();
1555 lttng_ring_buffer_client_overwrite_exit();
1556 lttng_ring_buffer_metadata_client_exit();
1557 lttng_ust_baddr_statedump_destroy();
1560 /* Reinitialize values for fork */
1562 lttng_ust_comm_should_quit
= 0;
1567 void __attribute__((destructor
)) lttng_ust_exit(void)
1572 * Using pthread_cancel here because:
1573 * A) we don't want to hang application teardown.
1574 * B) the thread is not allocating any resource.
1578 * Require the communication thread to quit. Synchronize with
1579 * mutexes to ensure it is not in a mutex critical section when
1580 * pthread_cancel is later called.
1583 lttng_ust_comm_should_quit
= 1;
1586 pthread_mutex_lock(&ust_exit_mutex
);
1587 /* cancel threads */
1588 if (global_apps
.thread_active
) {
1589 ret
= pthread_cancel(global_apps
.ust_listener
);
1591 ERR("Error cancelling global ust listener thread: %s",
1594 global_apps
.thread_active
= 0;
1597 if (local_apps
.thread_active
) {
1598 ret
= pthread_cancel(local_apps
.ust_listener
);
1600 ERR("Error cancelling local ust listener thread: %s",
1603 local_apps
.thread_active
= 0;
1606 pthread_mutex_unlock(&ust_exit_mutex
);
1609 * Do NOT join threads: use of sys_futex makes it impossible to
1610 * join the threads without using async-cancel, but async-cancel
1611 * is delivered by a signal, which could hit the target thread
1612 * anywhere in its code path, including while the ust_lock() is
1613 * held, causing a deadlock for the other thread. Let the OS
1614 * cleanup the threads if there are stalled in a syscall.
1616 lttng_ust_cleanup(1);
1620 * We exclude the worker threads across fork and clone (except
1621 * CLONE_VM), because these system calls only keep the forking thread
1622 * running in the child. Therefore, we don't want to call fork or clone
1623 * in the middle of an tracepoint or ust tracing state modification.
1624 * Holding this mutex protects these structures across fork and clone.
1626 void ust_before_fork(sigset_t
*save_sigset
)
1629 * Disable signals. This is to avoid that the child intervenes
1630 * before it is properly setup for tracing. It is safer to
1631 * disable all signals, because then we know we are not breaking
1632 * anything by restoring the original mask.
1637 if (URCU_TLS(lttng_ust_nest_count
))
1639 /* Disable signals */
1640 sigfillset(&all_sigs
);
1641 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1643 PERROR("sigprocmask");
1646 pthread_mutex_lock(&ust_fork_mutex
);
1649 rcu_bp_before_fork();
1652 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1656 DBG("process %d", getpid());
1659 pthread_mutex_unlock(&ust_fork_mutex
);
1661 /* Restore signals */
1662 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1664 PERROR("sigprocmask");
1668 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1670 if (URCU_TLS(lttng_ust_nest_count
))
1672 DBG("process %d", getpid());
1673 rcu_bp_after_fork_parent();
1674 /* Release mutexes and reenable signals */
1675 ust_after_fork_common(restore_sigset
);
1679 * After fork, in the child, we need to cleanup all the leftover state,
1680 * except the worker thread which already magically disappeared thanks
1681 * to the weird Linux fork semantics. After tyding up, we call
1682 * lttng_ust_init() again to start over as a new PID.
1684 * This is meant for forks() that have tracing in the child between the
1685 * fork and following exec call (if there is any).
1687 void ust_after_fork_child(sigset_t
*restore_sigset
)
1689 if (URCU_TLS(lttng_ust_nest_count
))
1691 DBG("process %d", getpid());
1692 /* Release urcu mutexes */
1693 rcu_bp_after_fork_child();
1694 lttng_ust_cleanup(0);
1695 lttng_context_vtid_reset();
1696 /* Release mutexes and reenable signals */
1697 ust_after_fork_common(restore_sigset
);
1701 void lttng_ust_sockinfo_session_enabled(void *owner
)
1703 struct sock_info
*sock_info
= owner
;
1704 sock_info
->statedump_pending
= 1;