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
24 #include <sys/types.h>
25 #include <sys/socket.h>
28 #include <sys/types.h>
34 #include <semaphore.h>
38 #include <urcu/uatomic.h>
39 #include <urcu/futex.h>
40 #include <urcu/compiler.h>
42 #include <lttng/ust-events.h>
43 #include <lttng/ust-abi.h>
44 #include <lttng/ust.h>
45 #include <lttng/ust-error.h>
46 #include <lttng/ust-ctl.h>
47 #include <urcu/tls-compat.h>
49 #include <usterr-signal-safe.h>
51 #include "tracepoint-internal.h"
52 #include "lttng-tracer-core.h"
54 #include "../libringbuffer/tlsfixup.h"
55 #include "lttng-ust-statedump.h"
57 #include "../libringbuffer/getcpu.h"
61 * Has lttng ust comm constructor been called ?
63 static int initialized
;
66 * The ust_lock/ust_unlock lock is used as a communication thread mutex.
67 * Held when handling a command, also held by fork() to deal with
68 * removal of threads, and by exit path.
70 * The UST lock is the centralized mutex across UST tracing control and
73 * ust_exit_mutex must never nest in ust_mutex.
75 * ust_fork_mutex must never nest in ust_mutex.
77 * ust_mutex_nest is a per-thread nesting counter, allowing the perf
78 * counter lazy initialization called by events within the statedump,
79 * which traces while the ust_mutex is held.
81 * ust_lock nests within the dynamic loader lock (within glibc) because
82 * it is taken within the library constructor.
84 static pthread_mutex_t ust_mutex
= PTHREAD_MUTEX_INITIALIZER
;
86 /* Allow nesting the ust_mutex within the same thread. */
87 static DEFINE_URCU_TLS(int, ust_mutex_nest
);
90 * ust_exit_mutex protects thread_active variable wrt thread exit. It
91 * cannot be done by ust_mutex because pthread_cancel(), which takes an
92 * internal libc lock, cannot nest within ust_mutex.
94 * It never nests within a ust_mutex.
96 static pthread_mutex_t ust_exit_mutex
= PTHREAD_MUTEX_INITIALIZER
;
99 * ust_fork_mutex protects base address statedump tracing against forks. It
100 * prevents the dynamic loader lock to be taken (by base address statedump
101 * tracing) while a fork is happening, thus preventing deadlock issues with
102 * the dynamic loader lock.
104 static pthread_mutex_t ust_fork_mutex
= PTHREAD_MUTEX_INITIALIZER
;
106 /* Should the ust comm thread quit ? */
107 static int lttng_ust_comm_should_quit
;
110 * Return 0 on success, -1 if should quit.
111 * The lock is taken in both cases.
116 sigset_t sig_all_blocked
, orig_mask
;
119 sigfillset(&sig_all_blocked
);
120 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
122 ERR("pthread_sigmask: %s", strerror(ret
));
124 if (!URCU_TLS(ust_mutex_nest
)++)
125 pthread_mutex_lock(&ust_mutex
);
126 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
128 ERR("pthread_sigmask: %s", strerror(ret
));
130 if (lttng_ust_comm_should_quit
) {
138 * ust_lock_nocheck() can be used in constructors/destructors, because
139 * they are already nested within the dynamic loader lock, and therefore
140 * have exclusive access against execution of liblttng-ust destructor.
143 void ust_lock_nocheck(void)
145 sigset_t sig_all_blocked
, orig_mask
;
148 sigfillset(&sig_all_blocked
);
149 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
151 ERR("pthread_sigmask: %s", strerror(ret
));
153 if (!URCU_TLS(ust_mutex_nest
)++)
154 pthread_mutex_lock(&ust_mutex
);
155 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
157 ERR("pthread_sigmask: %s", strerror(ret
));
164 void ust_unlock(void)
166 sigset_t sig_all_blocked
, orig_mask
;
169 sigfillset(&sig_all_blocked
);
170 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
172 ERR("pthread_sigmask: %s", strerror(ret
));
174 if (!--URCU_TLS(ust_mutex_nest
))
175 pthread_mutex_unlock(&ust_mutex
);
176 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
178 ERR("pthread_sigmask: %s", strerror(ret
));
183 * Wait for either of these before continuing to the main
185 * - the register_done message from sessiond daemon
186 * (will let the sessiond daemon enable sessions before main
188 * - sessiond daemon is not reachable.
189 * - timeout (ensuring applications are resilient to session
192 static sem_t constructor_wait
;
194 * Doing this for both the global and local sessiond.
196 static int sem_count
= { 2 };
199 * Counting nesting within lttng-ust. Used to ensure that calling fork()
200 * from liblttng-ust does not execute the pre/post fork handlers.
202 static DEFINE_URCU_TLS(int, lttng_ust_nest_count
);
205 * Info about socket and associated listener thread.
209 pthread_t ust_listener
; /* listener thread */
211 int constructor_sem_posted
;
216 char sock_path
[PATH_MAX
];
220 char wait_shm_path
[PATH_MAX
];
222 /* Keep track of lazy state dump not performed yet. */
223 int statedump_pending
;
226 /* Socket from app (connect) to session daemon (listen) for communication */
227 struct sock_info global_apps
= {
235 .sock_path
= LTTNG_DEFAULT_RUNDIR
"/" LTTNG_UST_SOCK_FILENAME
,
239 .wait_shm_path
= "/" LTTNG_UST_WAIT_FILENAME
,
241 .statedump_pending
= 0,
244 /* TODO: allow global_apps_sock_path override */
246 struct sock_info local_apps
= {
250 .allowed
= 0, /* Check setuid bit first */
256 .statedump_pending
= 0,
259 static int wait_poll_fallback
;
261 static const char *cmd_name_mapping
[] = {
262 [ LTTNG_UST_RELEASE
] = "Release",
263 [ LTTNG_UST_SESSION
] = "Create Session",
264 [ LTTNG_UST_TRACER_VERSION
] = "Get Tracer Version",
266 [ LTTNG_UST_TRACEPOINT_LIST
] = "Create Tracepoint List",
267 [ LTTNG_UST_WAIT_QUIESCENT
] = "Wait for Quiescent State",
268 [ LTTNG_UST_REGISTER_DONE
] = "Registration Done",
269 [ LTTNG_UST_TRACEPOINT_FIELD_LIST
] = "Create Tracepoint Field List",
271 /* Session FD commands */
272 [ LTTNG_UST_CHANNEL
] = "Create Channel",
273 [ LTTNG_UST_SESSION_START
] = "Start Session",
274 [ LTTNG_UST_SESSION_STOP
] = "Stop Session",
276 /* Channel FD commands */
277 [ LTTNG_UST_STREAM
] = "Create Stream",
278 [ LTTNG_UST_EVENT
] = "Create Event",
280 /* Event and Channel FD commands */
281 [ LTTNG_UST_CONTEXT
] = "Create Context",
282 [ LTTNG_UST_FLUSH_BUFFER
] = "Flush Buffer",
284 /* Event, Channel and Session commands */
285 [ LTTNG_UST_ENABLE
] = "Enable",
286 [ LTTNG_UST_DISABLE
] = "Disable",
288 /* Tracepoint list commands */
289 [ LTTNG_UST_TRACEPOINT_LIST_GET
] = "List Next Tracepoint",
290 [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
] = "List Next Tracepoint Field",
292 /* Event FD commands */
293 [ LTTNG_UST_FILTER
] = "Create Filter",
294 [ LTTNG_UST_EXCLUSION
] = "Add exclusions to event",
297 static const char *str_timeout
;
298 static int got_timeout_env
;
300 extern void lttng_ring_buffer_client_overwrite_init(void);
301 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
302 extern void lttng_ring_buffer_client_discard_init(void);
303 extern void lttng_ring_buffer_client_discard_rt_init(void);
304 extern void lttng_ring_buffer_metadata_client_init(void);
305 extern void lttng_ring_buffer_client_overwrite_exit(void);
306 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
307 extern void lttng_ring_buffer_client_discard_exit(void);
308 extern void lttng_ring_buffer_client_discard_rt_exit(void);
309 extern void lttng_ring_buffer_metadata_client_exit(void);
312 * Returns the HOME directory path. Caller MUST NOT free(3) the returned
316 const char *get_lttng_home_dir(void)
320 val
= (const char *) lttng_secure_getenv("LTTNG_HOME");
324 return (const char *) lttng_secure_getenv("HOME");
328 * Force a read (imply TLS fixup for dlopen) of TLS variables.
331 void lttng_fixup_nest_count_tls(void)
333 asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count
)));
337 void lttng_fixup_ust_mutex_nest_tls(void)
339 asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest
)));
346 void lttng_fixup_urcu_bp_tls(void)
352 int lttng_get_notify_socket(void *owner
)
354 struct sock_info
*info
= owner
;
356 return info
->notify_socket
;
360 void print_cmd(int cmd
, int handle
)
362 const char *cmd_name
= "Unknown";
364 if (cmd
>= 0 && cmd
< LTTNG_ARRAY_SIZE(cmd_name_mapping
)
365 && cmd_name_mapping
[cmd
]) {
366 cmd_name
= cmd_name_mapping
[cmd
];
368 DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
370 lttng_ust_obj_get_name(handle
), handle
);
374 int setup_local_apps(void)
376 const char *home_dir
;
381 * Disallow per-user tracing for setuid binaries.
383 if (uid
!= geteuid()) {
384 assert(local_apps
.allowed
== 0);
387 home_dir
= get_lttng_home_dir();
389 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
390 assert(local_apps
.allowed
== 0);
393 local_apps
.allowed
= 1;
394 snprintf(local_apps
.sock_path
, PATH_MAX
, "%s/%s/%s",
396 LTTNG_DEFAULT_HOME_RUNDIR
,
397 LTTNG_UST_SOCK_FILENAME
);
398 snprintf(local_apps
.wait_shm_path
, PATH_MAX
, "/%s-%u",
399 LTTNG_UST_WAIT_FILENAME
,
405 * Get notify_sock timeout, in ms.
406 * -1: wait forever. 0: don't wait. >0: timeout, in ms.
409 long get_timeout(void)
411 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
413 if (!got_timeout_env
) {
414 str_timeout
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
418 constructor_delay_ms
= strtol(str_timeout
, NULL
, 10);
419 return constructor_delay_ms
;
423 long get_notify_sock_timeout(void)
425 return get_timeout();
429 * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
432 int get_constructor_timeout(struct timespec
*constructor_timeout
)
434 long constructor_delay_ms
;
437 constructor_delay_ms
= get_timeout();
439 switch (constructor_delay_ms
) {
440 case -1:/* fall-through */
442 return constructor_delay_ms
;
448 * If we are unable to find the current time, don't wait.
450 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
455 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
456 constructor_timeout
->tv_nsec
+=
457 (constructor_delay_ms
% 1000UL) * 1000000UL;
458 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
459 constructor_timeout
->tv_sec
++;
460 constructor_timeout
->tv_nsec
-= 1000000000UL;
462 /* Timeout wait (constructor_delay_ms). */
467 int register_to_sessiond(int socket
, enum ustctl_socket_type type
)
469 return ustcomm_send_reg_msg(socket
,
472 lttng_alignof(uint8_t) * CHAR_BIT
,
473 lttng_alignof(uint16_t) * CHAR_BIT
,
474 lttng_alignof(uint32_t) * CHAR_BIT
,
475 lttng_alignof(uint64_t) * CHAR_BIT
,
476 lttng_alignof(unsigned long) * CHAR_BIT
);
480 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
484 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
487 DBG("message successfully sent");
490 if (len
== -ECONNRESET
) {
491 DBG("remote end closed connection");
496 DBG("incorrect message size: %zd", len
);
502 int handle_register_done(struct sock_info
*sock_info
)
506 if (sock_info
->constructor_sem_posted
)
508 sock_info
->constructor_sem_posted
= 1;
509 if (uatomic_read(&sem_count
) <= 0) {
512 ret
= uatomic_add_return(&sem_count
, -1);
514 ret
= sem_post(&constructor_wait
);
521 * Only execute pending statedump after the constructor semaphore has
522 * been posted by each listener thread. This means statedump will only
523 * be performed after the "registration done" command is received from
524 * each session daemon the application is connected to.
526 * This ensures we don't run into deadlock issues with the dynamic
527 * loader mutex, which is held while the constructor is called and
528 * waiting on the constructor semaphore. All operations requiring this
529 * dynamic loader lock need to be postponed using this mechanism.
532 void handle_pending_statedump(struct sock_info
*sock_info
)
534 int ctor_passed
= sock_info
->constructor_sem_posted
;
536 if (ctor_passed
&& sock_info
->statedump_pending
) {
537 sock_info
->statedump_pending
= 0;
538 pthread_mutex_lock(&ust_fork_mutex
);
539 lttng_handle_pending_statedump(sock_info
);
540 pthread_mutex_unlock(&ust_fork_mutex
);
545 int handle_message(struct sock_info
*sock_info
,
546 int sock
, struct ustcomm_ust_msg
*lum
)
549 const struct lttng_ust_objd_ops
*ops
;
550 struct ustcomm_ust_reply lur
;
554 memset(&lur
, 0, sizeof(lur
));
557 ret
= -LTTNG_UST_ERR_EXITING
;
561 ops
= objd_ops(lum
->handle
);
568 case LTTNG_UST_REGISTER_DONE
:
569 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
570 ret
= handle_register_done(sock_info
);
574 case LTTNG_UST_RELEASE
:
575 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
578 ret
= lttng_ust_objd_unref(lum
->handle
, 1);
580 case LTTNG_UST_FILTER
:
582 /* Receive filter data */
583 struct lttng_ust_filter_bytecode_node
*bytecode
;
585 if (lum
->u
.filter
.data_size
> FILTER_BYTECODE_MAX_LEN
) {
586 ERR("Filter data size is too large: %u bytes",
587 lum
->u
.filter
.data_size
);
592 if (lum
->u
.filter
.reloc_offset
> lum
->u
.filter
.data_size
) {
593 ERR("Filter reloc offset %u is not within data",
594 lum
->u
.filter
.reloc_offset
);
599 bytecode
= zmalloc(sizeof(*bytecode
) + lum
->u
.filter
.data_size
);
604 len
= ustcomm_recv_unix_sock(sock
, bytecode
->bc
.data
,
605 lum
->u
.filter
.data_size
);
607 case 0: /* orderly shutdown */
612 if (len
== lum
->u
.filter
.data_size
) {
613 DBG("filter data received");
615 } else if (len
< 0) {
616 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
617 if (len
== -ECONNRESET
) {
618 ERR("%s remote end closed connection", sock_info
->name
);
627 DBG("incorrect filter data message size: %zd", len
);
633 bytecode
->bc
.len
= lum
->u
.filter
.data_size
;
634 bytecode
->bc
.reloc_offset
= lum
->u
.filter
.reloc_offset
;
635 bytecode
->bc
.seqnum
= lum
->u
.filter
.seqnum
;
637 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
638 (unsigned long) bytecode
,
643 /* don't free bytecode if everything went fine. */
650 case LTTNG_UST_EXCLUSION
:
652 /* Receive exclusion names */
653 struct lttng_ust_excluder_node
*node
;
656 count
= lum
->u
.exclusion
.count
;
658 /* There are no names to read */
662 node
= zmalloc(sizeof(*node
) +
663 count
* LTTNG_UST_SYM_NAME_LEN
);
668 node
->excluder
.count
= count
;
669 len
= ustcomm_recv_unix_sock(sock
, node
->excluder
.names
,
670 count
* LTTNG_UST_SYM_NAME_LEN
);
672 case 0: /* orderly shutdown */
677 if (len
== count
* LTTNG_UST_SYM_NAME_LEN
) {
678 DBG("Exclusion data received");
680 } else if (len
< 0) {
681 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
682 if (len
== -ECONNRESET
) {
683 ERR("%s remote end closed connection", sock_info
->name
);
692 DBG("Incorrect exclusion data message size: %zd", len
);
699 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
700 (unsigned long) node
,
705 /* Don't free exclusion data if everything went fine. */
712 case LTTNG_UST_CHANNEL
:
717 len
= ustcomm_recv_channel_from_sessiond(sock
,
718 &chan_data
, lum
->u
.channel
.len
,
721 case 0: /* orderly shutdown */
725 if (len
== lum
->u
.channel
.len
) {
726 DBG("channel data received");
728 } else if (len
< 0) {
729 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
730 if (len
== -ECONNRESET
) {
731 ERR("%s remote end closed connection", sock_info
->name
);
738 DBG("incorrect channel data message size: %zd", len
);
743 args
.channel
.chan_data
= chan_data
;
744 args
.channel
.wakeup_fd
= wakeup_fd
;
746 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
747 (unsigned long) &lum
->u
,
753 case LTTNG_UST_STREAM
:
755 /* Receive shm_fd, wakeup_fd */
756 ret
= ustcomm_recv_stream_from_sessiond(sock
,
759 &args
.stream
.wakeup_fd
);
764 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
765 (unsigned long) &lum
->u
,
773 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
774 (unsigned long) &lum
->u
,
782 lur
.handle
= lum
->handle
;
786 lur
.ret_code
= LTTNG_UST_OK
;
789 * Use -LTTNG_UST_ERR as wildcard for UST internal
790 * error that are not caused by the transport, except if
791 * we already have a more precise error message to
794 if (ret
> -LTTNG_UST_ERR
) {
795 /* Translate code to UST error. */
798 lur
.ret_code
= -LTTNG_UST_ERR_EXIST
;
801 lur
.ret_code
= -LTTNG_UST_ERR_INVAL
;
804 lur
.ret_code
= -LTTNG_UST_ERR_NOENT
;
807 lur
.ret_code
= -LTTNG_UST_ERR_PERM
;
810 lur
.ret_code
= -LTTNG_UST_ERR_NOSYS
;
813 lur
.ret_code
= -LTTNG_UST_ERR
;
822 case LTTNG_UST_TRACER_VERSION
:
823 lur
.u
.version
= lum
->u
.version
;
825 case LTTNG_UST_TRACEPOINT_LIST_GET
:
826 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
830 DBG("Return value: %d", lur
.ret_val
);
831 ret
= send_reply(sock
, &lur
);
833 DBG("error sending reply");
838 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
841 if (lur
.ret_code
== LTTNG_UST_OK
) {
843 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
844 len
= ustcomm_send_unix_sock(sock
,
845 &args
.field_list
.entry
,
846 sizeof(args
.field_list
.entry
));
851 if (len
!= sizeof(args
.field_list
.entry
)) {
862 * Performed delayed statedump operations outside of the UST
863 * lock. We need to take the dynamic loader lock before we take
864 * the UST lock internally within handle_pending_statedump().
866 handle_pending_statedump(sock_info
);
872 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
876 if (sock_info
->root_handle
!= -1) {
877 ret
= lttng_ust_objd_unref(sock_info
->root_handle
, 1);
879 ERR("Error unref root handle");
881 sock_info
->root_handle
= -1;
883 sock_info
->constructor_sem_posted
= 0;
886 * wait_shm_mmap, socket and notify socket are used by listener
887 * threads outside of the ust lock, so we cannot tear them down
888 * ourselves, because we cannot join on these threads. Leave
889 * responsibility of cleaning up these resources to the OS
895 if (sock_info
->socket
!= -1) {
896 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
898 ERR("Error closing ust cmd socket");
900 sock_info
->socket
= -1;
902 if (sock_info
->notify_socket
!= -1) {
903 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
905 ERR("Error closing ust notify socket");
907 sock_info
->notify_socket
= -1;
909 if (sock_info
->wait_shm_mmap
) {
912 page_size
= sysconf(_SC_PAGE_SIZE
);
914 ret
= munmap(sock_info
->wait_shm_mmap
, page_size
);
916 ERR("Error unmapping wait shm");
919 sock_info
->wait_shm_mmap
= NULL
;
924 * Using fork to set umask in the child process (not multi-thread safe).
925 * We deal with the shm_open vs ftruncate race (happening when the
926 * sessiond owns the shm and does not let everybody modify it, to ensure
927 * safety against shm_unlink) by simply letting the mmap fail and
928 * retrying after a few seconds.
929 * For global shm, everybody has rw access to it until the sessiond
933 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
935 int wait_shm_fd
, ret
;
939 * Try to open read-only.
941 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
942 if (wait_shm_fd
>= 0) {
945 size_t bytes_read
= 0;
948 * Try to read the fd. If unable to do so, try opening
952 len
= read(wait_shm_fd
,
953 &((char *) &tmp_read
)[bytes_read
],
954 sizeof(tmp_read
) - bytes_read
);
958 } while ((len
< 0 && errno
== EINTR
)
959 || (len
> 0 && bytes_read
< sizeof(tmp_read
)));
960 if (bytes_read
!= sizeof(tmp_read
)) {
961 ret
= close(wait_shm_fd
);
963 ERR("close wait_shm_fd");
968 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
970 * Real-only open did not work, and it's not because the
971 * entry was not present. It's a failure that prohibits
974 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
980 * If the open failed because the file did not exist, or because
981 * the file was not truncated yet, try creating it ourself.
983 URCU_TLS(lttng_ust_nest_count
)++;
985 URCU_TLS(lttng_ust_nest_count
)--;
990 * Parent: wait for child to return, in which case the
991 * shared memory map will have been created.
994 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
999 * Try to open read-only again after creation.
1001 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
1002 if (wait_shm_fd
< 0) {
1004 * Real-only open did not work. It's a failure
1005 * that prohibits using shm.
1007 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
1011 } else if (pid
== 0) {
1015 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
1016 if (sock_info
->global
)
1017 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
1019 * We're alone in a child process, so we can modify the
1020 * process-wide umask.
1022 umask(~create_mode
);
1024 * Try creating shm (or get rw access).
1025 * We don't do an exclusive open, because we allow other
1026 * processes to create+ftruncate it concurrently.
1028 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
1029 O_RDWR
| O_CREAT
, create_mode
);
1030 if (wait_shm_fd
>= 0) {
1031 ret
= ftruncate(wait_shm_fd
, mmap_size
);
1033 PERROR("ftruncate");
1034 _exit(EXIT_FAILURE
);
1036 _exit(EXIT_SUCCESS
);
1039 * For local shm, we need to have rw access to accept
1040 * opening it: this means the local sessiond will be
1041 * able to wake us up. For global shm, we open it even
1042 * if rw access is not granted, because the root.root
1043 * sessiond will be able to override all rights and wake
1046 if (!sock_info
->global
&& errno
!= EACCES
) {
1047 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
1048 _exit(EXIT_FAILURE
);
1051 * The shm exists, but we cannot open it RW. Report
1054 _exit(EXIT_SUCCESS
);
1059 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
1060 struct stat statbuf
;
1063 * Ensure that our user is the owner of the shm file for
1064 * local shm. If we do not own the file, it means our
1065 * sessiond will not have access to wake us up (there is
1066 * probably a rogue process trying to fake our
1067 * sessiond). Fallback to polling method in this case.
1069 ret
= fstat(wait_shm_fd
, &statbuf
);
1074 if (statbuf
.st_uid
!= getuid())
1080 ret
= close(wait_shm_fd
);
1082 PERROR("Error closing fd");
1088 char *get_map_shm(struct sock_info
*sock_info
)
1091 int wait_shm_fd
, ret
;
1092 char *wait_shm_mmap
;
1094 page_size
= sysconf(_SC_PAGE_SIZE
);
1095 if (page_size
< 0) {
1099 wait_shm_fd
= get_wait_shm(sock_info
, page_size
);
1100 if (wait_shm_fd
< 0) {
1103 wait_shm_mmap
= mmap(NULL
, page_size
, PROT_READ
,
1104 MAP_SHARED
, wait_shm_fd
, 0);
1105 /* close shm fd immediately after taking the mmap reference */
1106 ret
= close(wait_shm_fd
);
1108 PERROR("Error closing fd");
1110 if (wait_shm_mmap
== MAP_FAILED
) {
1111 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1114 return wait_shm_mmap
;
1121 void wait_for_sessiond(struct sock_info
*sock_info
)
1126 if (wait_poll_fallback
) {
1129 if (!sock_info
->wait_shm_mmap
) {
1130 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
1131 if (!sock_info
->wait_shm_mmap
)
1136 DBG("Waiting for %s apps sessiond", sock_info
->name
);
1137 /* Wait for futex wakeup */
1138 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
))
1141 while (futex_async((int32_t *) sock_info
->wait_shm_mmap
,
1142 FUTEX_WAIT
, 0, NULL
, NULL
, 0)) {
1145 /* Value already changed. */
1148 /* Retry if interrupted by signal. */
1149 break; /* Get out of switch. */
1151 wait_poll_fallback
= 1;
1153 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1154 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1155 "Please upgrade your kernel "
1156 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1157 "mainline). LTTng-UST will use polling mode fallback.");
1176 * This thread does not allocate any resource, except within
1177 * handle_message, within mutex protection. This mutex protects against
1179 * The other moment it allocates resources is at socket connection, which
1180 * is also protected by the mutex.
1183 void *ust_listener_thread(void *arg
)
1185 struct sock_info
*sock_info
= arg
;
1186 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
1189 /* Restart trying to connect to the session daemon */
1191 if (prev_connect_failed
) {
1192 /* Wait for sessiond availability with pipe */
1193 wait_for_sessiond(sock_info
);
1197 * Sleep for 5 seconds before retrying after a
1198 * sequence of failure / wait / failure. This
1199 * deals with a killed or broken session daemon.
1204 prev_connect_failed
= 0;
1207 if (sock_info
->socket
!= -1) {
1208 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
1210 ERR("Error closing %s ust cmd socket",
1213 sock_info
->socket
= -1;
1215 if (sock_info
->notify_socket
!= -1) {
1216 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
1218 ERR("Error closing %s ust notify socket",
1221 sock_info
->notify_socket
= -1;
1225 * Register. We need to perform both connect and sending
1226 * registration message before doing the next connect otherwise
1227 * we may reach unix socket connect queue max limits and block
1228 * on the 2nd connect while the session daemon is awaiting the
1229 * first connect registration message.
1231 /* Connect cmd socket */
1232 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1234 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1235 prev_connect_failed
= 1;
1242 * If we cannot find the sessiond daemon, don't delay
1243 * constructor execution.
1245 ret
= handle_register_done(sock_info
);
1250 sock_info
->socket
= ret
;
1257 * Create only one root handle per listener thread for the whole
1258 * process lifetime, so we ensure we get ID which is statically
1259 * assigned to the root handle.
1261 if (sock_info
->root_handle
== -1) {
1262 ret
= lttng_abi_create_root_handle();
1264 ERR("Error creating root handle");
1267 sock_info
->root_handle
= ret
;
1270 ret
= register_to_sessiond(sock_info
->socket
, USTCTL_SOCKET_CMD
);
1272 ERR("Error registering to %s ust cmd socket",
1274 prev_connect_failed
= 1;
1276 * If we cannot register to the sessiond daemon, don't
1277 * delay constructor execution.
1279 ret
= handle_register_done(sock_info
);
1287 /* Connect notify socket */
1288 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1290 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1291 prev_connect_failed
= 1;
1298 * If we cannot find the sessiond daemon, don't delay
1299 * constructor execution.
1301 ret
= handle_register_done(sock_info
);
1306 sock_info
->notify_socket
= ret
;
1308 timeout
= get_notify_sock_timeout();
1311 * Give at least 10ms to sessiond to reply to
1316 ret
= ustcomm_setsockopt_rcv_timeout(sock_info
->notify_socket
,
1319 WARN("Error setting socket receive timeout");
1321 ret
= ustcomm_setsockopt_snd_timeout(sock_info
->notify_socket
,
1324 WARN("Error setting socket send timeout");
1326 } else if (timeout
< -1) {
1327 WARN("Unsupported timeout value %ld", timeout
);
1334 ret
= register_to_sessiond(sock_info
->notify_socket
,
1335 USTCTL_SOCKET_NOTIFY
);
1337 ERR("Error registering to %s ust notify socket",
1339 prev_connect_failed
= 1;
1341 * If we cannot register to the sessiond daemon, don't
1342 * delay constructor execution.
1344 ret
= handle_register_done(sock_info
);
1349 sock
= sock_info
->socket
;
1355 struct ustcomm_ust_msg lum
;
1357 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
1359 case 0: /* orderly shutdown */
1360 DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info
->name
);
1365 * Either sessiond has shutdown or refused us by closing the socket.
1366 * In either case, we don't want to delay construction execution,
1367 * and we need to wait before retry.
1369 prev_connect_failed
= 1;
1371 * If we cannot register to the sessiond daemon, don't
1372 * delay constructor execution.
1374 ret
= handle_register_done(sock_info
);
1379 print_cmd(lum
.cmd
, lum
.handle
);
1380 ret
= handle_message(sock_info
, sock
, &lum
);
1382 ERR("Error handling message for %s socket", sock_info
->name
);
1387 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
1389 DBG("incorrect message size (%s socket): %zd", sock_info
->name
, len
);
1391 if (len
== -ECONNRESET
) {
1392 DBG("%s remote end closed connection", sock_info
->name
);
1403 /* Cleanup socket handles before trying to reconnect */
1404 lttng_ust_objd_table_owner_cleanup(sock_info
);
1406 goto restart
; /* try to reconnect */
1411 pthread_mutex_lock(&ust_exit_mutex
);
1412 sock_info
->thread_active
= 0;
1413 pthread_mutex_unlock(&ust_exit_mutex
);
1418 * Weak symbol to call when the ust malloc wrapper is not loaded.
1420 __attribute__((weak
))
1421 void lttng_ust_malloc_wrapper_init(void)
1426 * sessiond monitoring thread: monitor presence of global and per-user
1427 * sessiond by polling the application common named pipe.
1429 void __attribute__((constructor
)) lttng_ust_init(void)
1431 struct timespec constructor_timeout
;
1432 sigset_t sig_all_blocked
, orig_parent_mask
;
1433 pthread_attr_t thread_attr
;
1437 if (uatomic_xchg(&initialized
, 1) == 1)
1441 * Fixup interdependency between TLS fixup mutex (which happens
1442 * to be the dynamic linker mutex) and ust_lock, taken within
1445 lttng_fixup_urcu_bp_tls();
1446 lttng_fixup_ringbuffer_tls();
1447 lttng_fixup_vtid_tls();
1448 lttng_fixup_nest_count_tls();
1449 lttng_fixup_procname_tls();
1450 lttng_fixup_ust_mutex_nest_tls();
1453 * We want precise control over the order in which we construct
1454 * our sub-libraries vs starting to receive commands from
1455 * sessiond (otherwise leading to errors when trying to create
1456 * sessiond before the init functions are completed).
1460 lttng_ust_clock_init();
1461 lttng_ust_getcpu_init();
1462 lttng_ust_statedump_init();
1463 lttng_ring_buffer_metadata_client_init();
1464 lttng_ring_buffer_client_overwrite_init();
1465 lttng_ring_buffer_client_overwrite_rt_init();
1466 lttng_ring_buffer_client_discard_init();
1467 lttng_ring_buffer_client_discard_rt_init();
1468 lttng_perf_counter_init();
1469 lttng_context_init();
1471 * Invoke ust malloc wrapper init before starting other threads.
1473 lttng_ust_malloc_wrapper_init();
1475 timeout_mode
= get_constructor_timeout(&constructor_timeout
);
1477 ret
= sem_init(&constructor_wait
, 0, 0);
1480 ret
= setup_local_apps();
1482 DBG("local apps setup returned %d", ret
);
1485 /* A new thread created by pthread_create inherits the signal mask
1486 * from the parent. To avoid any signal being received by the
1487 * listener thread, we block all signals temporarily in the parent,
1488 * while we create the listener thread.
1490 sigfillset(&sig_all_blocked
);
1491 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
1493 ERR("pthread_sigmask: %s", strerror(ret
));
1496 ret
= pthread_attr_init(&thread_attr
);
1498 ERR("pthread_attr_init: %s", strerror(ret
));
1500 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
1502 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1505 pthread_mutex_lock(&ust_exit_mutex
);
1506 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1507 ust_listener_thread
, &global_apps
);
1509 ERR("pthread_create global: %s", strerror(ret
));
1511 global_apps
.thread_active
= 1;
1512 pthread_mutex_unlock(&ust_exit_mutex
);
1514 if (local_apps
.allowed
) {
1515 pthread_mutex_lock(&ust_exit_mutex
);
1516 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1517 ust_listener_thread
, &local_apps
);
1519 ERR("pthread_create local: %s", strerror(ret
));
1521 local_apps
.thread_active
= 1;
1522 pthread_mutex_unlock(&ust_exit_mutex
);
1524 handle_register_done(&local_apps
);
1526 ret
= pthread_attr_destroy(&thread_attr
);
1528 ERR("pthread_attr_destroy: %s", strerror(ret
));
1531 /* Restore original signal mask in parent */
1532 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1534 ERR("pthread_sigmask: %s", strerror(ret
));
1537 switch (timeout_mode
) {
1538 case 1: /* timeout wait */
1540 ret
= sem_timedwait(&constructor_wait
,
1541 &constructor_timeout
);
1542 } while (ret
< 0 && errno
== EINTR
);
1543 if (ret
< 0 && errno
== ETIMEDOUT
) {
1544 ERR("Timed out waiting for lttng-sessiond");
1549 case -1:/* wait forever */
1551 ret
= sem_wait(&constructor_wait
);
1552 } while (ret
< 0 && errno
== EINTR
);
1555 case 0: /* no timeout */
1561 void lttng_ust_cleanup(int exiting
)
1563 cleanup_sock_info(&global_apps
, exiting
);
1564 cleanup_sock_info(&local_apps
, exiting
);
1566 * The teardown in this function all affect data structures
1567 * accessed under the UST lock by the listener thread. This
1568 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1569 * that none of these threads are accessing this data at this
1572 lttng_ust_abi_exit();
1573 lttng_ust_events_exit();
1574 lttng_context_exit();
1575 lttng_perf_counter_exit();
1576 lttng_ring_buffer_client_discard_rt_exit();
1577 lttng_ring_buffer_client_discard_exit();
1578 lttng_ring_buffer_client_overwrite_rt_exit();
1579 lttng_ring_buffer_client_overwrite_exit();
1580 lttng_ring_buffer_metadata_client_exit();
1581 lttng_ust_statedump_destroy();
1584 /* Reinitialize values for fork */
1586 lttng_ust_comm_should_quit
= 0;
1591 void __attribute__((destructor
)) lttng_ust_exit(void)
1596 * Using pthread_cancel here because:
1597 * A) we don't want to hang application teardown.
1598 * B) the thread is not allocating any resource.
1602 * Require the communication thread to quit. Synchronize with
1603 * mutexes to ensure it is not in a mutex critical section when
1604 * pthread_cancel is later called.
1607 lttng_ust_comm_should_quit
= 1;
1610 pthread_mutex_lock(&ust_exit_mutex
);
1611 /* cancel threads */
1612 if (global_apps
.thread_active
) {
1613 ret
= pthread_cancel(global_apps
.ust_listener
);
1615 ERR("Error cancelling global ust listener thread: %s",
1618 global_apps
.thread_active
= 0;
1621 if (local_apps
.thread_active
) {
1622 ret
= pthread_cancel(local_apps
.ust_listener
);
1624 ERR("Error cancelling local ust listener thread: %s",
1627 local_apps
.thread_active
= 0;
1630 pthread_mutex_unlock(&ust_exit_mutex
);
1633 * Do NOT join threads: use of sys_futex makes it impossible to
1634 * join the threads without using async-cancel, but async-cancel
1635 * is delivered by a signal, which could hit the target thread
1636 * anywhere in its code path, including while the ust_lock() is
1637 * held, causing a deadlock for the other thread. Let the OS
1638 * cleanup the threads if there are stalled in a syscall.
1640 lttng_ust_cleanup(1);
1644 * We exclude the worker threads across fork and clone (except
1645 * CLONE_VM), because these system calls only keep the forking thread
1646 * running in the child. Therefore, we don't want to call fork or clone
1647 * in the middle of an tracepoint or ust tracing state modification.
1648 * Holding this mutex protects these structures across fork and clone.
1650 void ust_before_fork(sigset_t
*save_sigset
)
1653 * Disable signals. This is to avoid that the child intervenes
1654 * before it is properly setup for tracing. It is safer to
1655 * disable all signals, because then we know we are not breaking
1656 * anything by restoring the original mask.
1661 if (URCU_TLS(lttng_ust_nest_count
))
1663 /* Disable signals */
1664 sigfillset(&all_sigs
);
1665 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1667 PERROR("sigprocmask");
1670 pthread_mutex_lock(&ust_fork_mutex
);
1673 rcu_bp_before_fork();
1676 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1680 DBG("process %d", getpid());
1683 pthread_mutex_unlock(&ust_fork_mutex
);
1685 /* Restore signals */
1686 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1688 PERROR("sigprocmask");
1692 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1694 if (URCU_TLS(lttng_ust_nest_count
))
1696 DBG("process %d", getpid());
1697 rcu_bp_after_fork_parent();
1698 /* Release mutexes and reenable signals */
1699 ust_after_fork_common(restore_sigset
);
1703 * After fork, in the child, we need to cleanup all the leftover state,
1704 * except the worker thread which already magically disappeared thanks
1705 * to the weird Linux fork semantics. After tyding up, we call
1706 * lttng_ust_init() again to start over as a new PID.
1708 * This is meant for forks() that have tracing in the child between the
1709 * fork and following exec call (if there is any).
1711 void ust_after_fork_child(sigset_t
*restore_sigset
)
1713 if (URCU_TLS(lttng_ust_nest_count
))
1715 DBG("process %d", getpid());
1716 /* Release urcu mutexes */
1717 rcu_bp_after_fork_child();
1718 lttng_ust_cleanup(0);
1719 lttng_context_vtid_reset();
1720 /* Release mutexes and reenable signals */
1721 ust_after_fork_common(restore_sigset
);
1725 void lttng_ust_sockinfo_session_enabled(void *owner
)
1727 struct sock_info
*sock_info
= owner
;
1728 sock_info
->statedump_pending
= 1;