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 ret
= pthread_setcancelstate(PTHREAD_CANCEL_DISABLE
, &oldstate
);
121 ERR("pthread_setcancelstate: %s", strerror(ret
));
123 if (oldstate
!= PTHREAD_CANCEL_ENABLE
) {
124 ERR("pthread_setcancelstate: unexpected oldstate");
126 sigfillset(&sig_all_blocked
);
127 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
129 ERR("pthread_sigmask: %s", strerror(ret
));
131 if (!URCU_TLS(ust_mutex_nest
)++)
132 pthread_mutex_lock(&ust_mutex
);
133 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
135 ERR("pthread_sigmask: %s", strerror(ret
));
137 if (lttng_ust_comm_should_quit
) {
145 * ust_lock_nocheck() can be used in constructors/destructors, because
146 * they are already nested within the dynamic loader lock, and therefore
147 * have exclusive access against execution of liblttng-ust destructor.
150 void ust_lock_nocheck(void)
152 sigset_t sig_all_blocked
, orig_mask
;
155 ret
= pthread_setcancelstate(PTHREAD_CANCEL_DISABLE
, &oldstate
);
157 ERR("pthread_setcancelstate: %s", strerror(ret
));
159 if (oldstate
!= PTHREAD_CANCEL_ENABLE
) {
160 ERR("pthread_setcancelstate: unexpected oldstate");
162 sigfillset(&sig_all_blocked
);
163 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
165 ERR("pthread_sigmask: %s", strerror(ret
));
167 if (!URCU_TLS(ust_mutex_nest
)++)
168 pthread_mutex_lock(&ust_mutex
);
169 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
171 ERR("pthread_sigmask: %s", strerror(ret
));
178 void ust_unlock(void)
180 sigset_t sig_all_blocked
, orig_mask
;
183 sigfillset(&sig_all_blocked
);
184 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
186 ERR("pthread_sigmask: %s", strerror(ret
));
188 if (!--URCU_TLS(ust_mutex_nest
))
189 pthread_mutex_unlock(&ust_mutex
);
190 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
192 ERR("pthread_sigmask: %s", strerror(ret
));
194 ret
= pthread_setcancelstate(PTHREAD_CANCEL_ENABLE
, &oldstate
);
196 ERR("pthread_setcancelstate: %s", strerror(ret
));
198 if (oldstate
!= PTHREAD_CANCEL_DISABLE
) {
199 ERR("pthread_setcancelstate: unexpected oldstate");
204 * Wait for either of these before continuing to the main
206 * - the register_done message from sessiond daemon
207 * (will let the sessiond daemon enable sessions before main
209 * - sessiond daemon is not reachable.
210 * - timeout (ensuring applications are resilient to session
213 static sem_t constructor_wait
;
215 * Doing this for both the global and local sessiond.
217 static int sem_count
= { 2 };
220 * Counting nesting within lttng-ust. Used to ensure that calling fork()
221 * from liblttng-ust does not execute the pre/post fork handlers.
223 static DEFINE_URCU_TLS(int, lttng_ust_nest_count
);
226 * Info about socket and associated listener thread.
230 pthread_t ust_listener
; /* listener thread */
232 int constructor_sem_posted
;
237 char sock_path
[PATH_MAX
];
241 char wait_shm_path
[PATH_MAX
];
243 /* Keep track of lazy state dump not performed yet. */
244 int statedump_pending
;
247 /* Socket from app (connect) to session daemon (listen) for communication */
248 struct sock_info global_apps
= {
256 .sock_path
= LTTNG_DEFAULT_RUNDIR
"/" LTTNG_UST_SOCK_FILENAME
,
260 .wait_shm_path
= "/" LTTNG_UST_WAIT_FILENAME
,
262 .statedump_pending
= 0,
265 /* TODO: allow global_apps_sock_path override */
267 struct sock_info local_apps
= {
271 .allowed
= 0, /* Check setuid bit first */
277 .statedump_pending
= 0,
280 static int wait_poll_fallback
;
282 static const char *cmd_name_mapping
[] = {
283 [ LTTNG_UST_RELEASE
] = "Release",
284 [ LTTNG_UST_SESSION
] = "Create Session",
285 [ LTTNG_UST_TRACER_VERSION
] = "Get Tracer Version",
287 [ LTTNG_UST_TRACEPOINT_LIST
] = "Create Tracepoint List",
288 [ LTTNG_UST_WAIT_QUIESCENT
] = "Wait for Quiescent State",
289 [ LTTNG_UST_REGISTER_DONE
] = "Registration Done",
290 [ LTTNG_UST_TRACEPOINT_FIELD_LIST
] = "Create Tracepoint Field List",
292 /* Session FD commands */
293 [ LTTNG_UST_CHANNEL
] = "Create Channel",
294 [ LTTNG_UST_SESSION_START
] = "Start Session",
295 [ LTTNG_UST_SESSION_STOP
] = "Stop Session",
297 /* Channel FD commands */
298 [ LTTNG_UST_STREAM
] = "Create Stream",
299 [ LTTNG_UST_EVENT
] = "Create Event",
301 /* Event and Channel FD commands */
302 [ LTTNG_UST_CONTEXT
] = "Create Context",
303 [ LTTNG_UST_FLUSH_BUFFER
] = "Flush Buffer",
305 /* Event, Channel and Session commands */
306 [ LTTNG_UST_ENABLE
] = "Enable",
307 [ LTTNG_UST_DISABLE
] = "Disable",
309 /* Tracepoint list commands */
310 [ LTTNG_UST_TRACEPOINT_LIST_GET
] = "List Next Tracepoint",
311 [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
] = "List Next Tracepoint Field",
313 /* Event FD commands */
314 [ LTTNG_UST_FILTER
] = "Create Filter",
315 [ LTTNG_UST_EXCLUSION
] = "Add exclusions to event",
318 static const char *str_timeout
;
319 static int got_timeout_env
;
321 extern void lttng_ring_buffer_client_overwrite_init(void);
322 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
323 extern void lttng_ring_buffer_client_discard_init(void);
324 extern void lttng_ring_buffer_client_discard_rt_init(void);
325 extern void lttng_ring_buffer_metadata_client_init(void);
326 extern void lttng_ring_buffer_client_overwrite_exit(void);
327 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
328 extern void lttng_ring_buffer_client_discard_exit(void);
329 extern void lttng_ring_buffer_client_discard_rt_exit(void);
330 extern void lttng_ring_buffer_metadata_client_exit(void);
333 * Returns the HOME directory path. Caller MUST NOT free(3) the returned
337 const char *get_lttng_home_dir(void)
341 val
= (const char *) lttng_secure_getenv("LTTNG_HOME");
345 return (const char *) lttng_secure_getenv("HOME");
349 * Force a read (imply TLS fixup for dlopen) of TLS variables.
352 void lttng_fixup_nest_count_tls(void)
354 asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count
)));
358 void lttng_fixup_ust_mutex_nest_tls(void)
360 asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest
)));
367 void lttng_fixup_urcu_bp_tls(void)
373 int lttng_get_notify_socket(void *owner
)
375 struct sock_info
*info
= owner
;
377 return info
->notify_socket
;
381 void print_cmd(int cmd
, int handle
)
383 const char *cmd_name
= "Unknown";
385 if (cmd
>= 0 && cmd
< LTTNG_ARRAY_SIZE(cmd_name_mapping
)
386 && cmd_name_mapping
[cmd
]) {
387 cmd_name
= cmd_name_mapping
[cmd
];
389 DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
391 lttng_ust_obj_get_name(handle
), handle
);
395 int setup_local_apps(void)
397 const char *home_dir
;
402 * Disallow per-user tracing for setuid binaries.
404 if (uid
!= geteuid()) {
405 assert(local_apps
.allowed
== 0);
408 home_dir
= get_lttng_home_dir();
410 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
411 assert(local_apps
.allowed
== 0);
414 local_apps
.allowed
= 1;
415 snprintf(local_apps
.sock_path
, PATH_MAX
, "%s/%s/%s",
417 LTTNG_DEFAULT_HOME_RUNDIR
,
418 LTTNG_UST_SOCK_FILENAME
);
419 snprintf(local_apps
.wait_shm_path
, PATH_MAX
, "/%s-%u",
420 LTTNG_UST_WAIT_FILENAME
,
426 * Get notify_sock timeout, in ms.
427 * -1: wait forever. 0: don't wait. >0: timeout, in ms.
430 long get_timeout(void)
432 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
434 if (!got_timeout_env
) {
435 str_timeout
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
439 constructor_delay_ms
= strtol(str_timeout
, NULL
, 10);
440 /* All negative values are considered as "-1". */
441 if (constructor_delay_ms
< -1)
442 constructor_delay_ms
= -1;
443 return constructor_delay_ms
;
447 long get_notify_sock_timeout(void)
449 return get_timeout();
453 * Return values: -1: wait forever. 0: don't wait. 1: timeout wait.
456 int get_constructor_timeout(struct timespec
*constructor_timeout
)
458 long constructor_delay_ms
;
461 constructor_delay_ms
= get_timeout();
463 switch (constructor_delay_ms
) {
464 case -1:/* fall-through */
466 return constructor_delay_ms
;
472 * If we are unable to find the current time, don't wait.
474 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
479 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
480 constructor_timeout
->tv_nsec
+=
481 (constructor_delay_ms
% 1000UL) * 1000000UL;
482 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
483 constructor_timeout
->tv_sec
++;
484 constructor_timeout
->tv_nsec
-= 1000000000UL;
486 /* Timeout wait (constructor_delay_ms). */
491 int register_to_sessiond(int socket
, enum ustctl_socket_type type
)
493 return ustcomm_send_reg_msg(socket
,
496 lttng_alignof(uint8_t) * CHAR_BIT
,
497 lttng_alignof(uint16_t) * CHAR_BIT
,
498 lttng_alignof(uint32_t) * CHAR_BIT
,
499 lttng_alignof(uint64_t) * CHAR_BIT
,
500 lttng_alignof(unsigned long) * CHAR_BIT
);
504 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
508 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
511 DBG("message successfully sent");
514 if (len
== -ECONNRESET
) {
515 DBG("remote end closed connection");
520 DBG("incorrect message size: %zd", len
);
526 int handle_register_done(struct sock_info
*sock_info
)
530 if (sock_info
->constructor_sem_posted
)
532 sock_info
->constructor_sem_posted
= 1;
533 if (uatomic_read(&sem_count
) <= 0) {
536 ret
= uatomic_add_return(&sem_count
, -1);
538 ret
= sem_post(&constructor_wait
);
545 * Only execute pending statedump after the constructor semaphore has
546 * been posted by each listener thread. This means statedump will only
547 * be performed after the "registration done" command is received from
548 * each session daemon the application is connected to.
550 * This ensures we don't run into deadlock issues with the dynamic
551 * loader mutex, which is held while the constructor is called and
552 * waiting on the constructor semaphore. All operations requiring this
553 * dynamic loader lock need to be postponed using this mechanism.
556 void handle_pending_statedump(struct sock_info
*sock_info
)
558 int ctor_passed
= sock_info
->constructor_sem_posted
;
560 if (ctor_passed
&& sock_info
->statedump_pending
) {
561 sock_info
->statedump_pending
= 0;
562 pthread_mutex_lock(&ust_fork_mutex
);
563 lttng_handle_pending_statedump(sock_info
);
564 pthread_mutex_unlock(&ust_fork_mutex
);
569 int handle_message(struct sock_info
*sock_info
,
570 int sock
, struct ustcomm_ust_msg
*lum
)
573 const struct lttng_ust_objd_ops
*ops
;
574 struct ustcomm_ust_reply lur
;
578 memset(&lur
, 0, sizeof(lur
));
581 ret
= -LTTNG_UST_ERR_EXITING
;
585 ops
= objd_ops(lum
->handle
);
592 case LTTNG_UST_REGISTER_DONE
:
593 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
594 ret
= handle_register_done(sock_info
);
598 case LTTNG_UST_RELEASE
:
599 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
602 ret
= lttng_ust_objd_unref(lum
->handle
, 1);
604 case LTTNG_UST_FILTER
:
606 /* Receive filter data */
607 struct lttng_ust_filter_bytecode_node
*bytecode
;
609 if (lum
->u
.filter
.data_size
> FILTER_BYTECODE_MAX_LEN
) {
610 ERR("Filter data size is too large: %u bytes",
611 lum
->u
.filter
.data_size
);
616 if (lum
->u
.filter
.reloc_offset
> lum
->u
.filter
.data_size
) {
617 ERR("Filter reloc offset %u is not within data",
618 lum
->u
.filter
.reloc_offset
);
623 bytecode
= zmalloc(sizeof(*bytecode
) + lum
->u
.filter
.data_size
);
628 len
= ustcomm_recv_unix_sock(sock
, bytecode
->bc
.data
,
629 lum
->u
.filter
.data_size
);
631 case 0: /* orderly shutdown */
636 if (len
== lum
->u
.filter
.data_size
) {
637 DBG("filter data received");
639 } else if (len
< 0) {
640 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
641 if (len
== -ECONNRESET
) {
642 ERR("%s remote end closed connection", sock_info
->name
);
651 DBG("incorrect filter data message size: %zd", len
);
657 bytecode
->bc
.len
= lum
->u
.filter
.data_size
;
658 bytecode
->bc
.reloc_offset
= lum
->u
.filter
.reloc_offset
;
659 bytecode
->bc
.seqnum
= lum
->u
.filter
.seqnum
;
661 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
662 (unsigned long) bytecode
,
667 /* don't free bytecode if everything went fine. */
674 case LTTNG_UST_EXCLUSION
:
676 /* Receive exclusion names */
677 struct lttng_ust_excluder_node
*node
;
680 count
= lum
->u
.exclusion
.count
;
682 /* There are no names to read */
686 node
= zmalloc(sizeof(*node
) +
687 count
* LTTNG_UST_SYM_NAME_LEN
);
692 node
->excluder
.count
= count
;
693 len
= ustcomm_recv_unix_sock(sock
, node
->excluder
.names
,
694 count
* LTTNG_UST_SYM_NAME_LEN
);
696 case 0: /* orderly shutdown */
701 if (len
== count
* LTTNG_UST_SYM_NAME_LEN
) {
702 DBG("Exclusion data received");
704 } else if (len
< 0) {
705 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
706 if (len
== -ECONNRESET
) {
707 ERR("%s remote end closed connection", sock_info
->name
);
716 DBG("Incorrect exclusion data message size: %zd", len
);
723 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
724 (unsigned long) node
,
729 /* Don't free exclusion data if everything went fine. */
736 case LTTNG_UST_CHANNEL
:
741 len
= ustcomm_recv_channel_from_sessiond(sock
,
742 &chan_data
, lum
->u
.channel
.len
,
745 case 0: /* orderly shutdown */
749 if (len
== lum
->u
.channel
.len
) {
750 DBG("channel data received");
752 } else if (len
< 0) {
753 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
754 if (len
== -ECONNRESET
) {
755 ERR("%s remote end closed connection", sock_info
->name
);
762 DBG("incorrect channel data message size: %zd", len
);
767 args
.channel
.chan_data
= chan_data
;
768 args
.channel
.wakeup_fd
= wakeup_fd
;
770 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
771 (unsigned long) &lum
->u
,
777 case LTTNG_UST_STREAM
:
779 /* Receive shm_fd, wakeup_fd */
780 ret
= ustcomm_recv_stream_from_sessiond(sock
,
783 &args
.stream
.wakeup_fd
);
788 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
789 (unsigned long) &lum
->u
,
797 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
798 (unsigned long) &lum
->u
,
805 lur
.handle
= lum
->handle
;
809 lur
.ret_code
= LTTNG_UST_OK
;
812 * Use -LTTNG_UST_ERR as wildcard for UST internal
813 * error that are not caused by the transport, except if
814 * we already have a more precise error message to
817 if (ret
> -LTTNG_UST_ERR
) {
818 /* Translate code to UST error. */
821 lur
.ret_code
= -LTTNG_UST_ERR_EXIST
;
824 lur
.ret_code
= -LTTNG_UST_ERR_INVAL
;
827 lur
.ret_code
= -LTTNG_UST_ERR_NOENT
;
830 lur
.ret_code
= -LTTNG_UST_ERR_PERM
;
833 lur
.ret_code
= -LTTNG_UST_ERR_NOSYS
;
836 lur
.ret_code
= -LTTNG_UST_ERR
;
845 case LTTNG_UST_TRACER_VERSION
:
846 lur
.u
.version
= lum
->u
.version
;
848 case LTTNG_UST_TRACEPOINT_LIST_GET
:
849 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
853 DBG("Return value: %d", lur
.ret_val
);
854 ret
= send_reply(sock
, &lur
);
856 DBG("error sending reply");
861 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
864 if (lur
.ret_code
== LTTNG_UST_OK
) {
866 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
867 len
= ustcomm_send_unix_sock(sock
,
868 &args
.field_list
.entry
,
869 sizeof(args
.field_list
.entry
));
874 if (len
!= sizeof(args
.field_list
.entry
)) {
885 * Performed delayed statedump operations outside of the UST
886 * lock. We need to take the dynamic loader lock before we take
887 * the UST lock internally within handle_pending_statedump().
889 handle_pending_statedump(sock_info
);
895 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
899 if (sock_info
->root_handle
!= -1) {
900 ret
= lttng_ust_objd_unref(sock_info
->root_handle
, 1);
902 ERR("Error unref root handle");
904 sock_info
->root_handle
= -1;
906 sock_info
->constructor_sem_posted
= 0;
909 * wait_shm_mmap, socket and notify socket are used by listener
910 * threads outside of the ust lock, so we cannot tear them down
911 * ourselves, because we cannot join on these threads. Leave
912 * responsibility of cleaning up these resources to the OS
918 if (sock_info
->socket
!= -1) {
919 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
921 ERR("Error closing ust cmd socket");
923 sock_info
->socket
= -1;
925 if (sock_info
->notify_socket
!= -1) {
926 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
928 ERR("Error closing ust notify socket");
930 sock_info
->notify_socket
= -1;
932 if (sock_info
->wait_shm_mmap
) {
935 page_size
= sysconf(_SC_PAGE_SIZE
);
936 if (page_size
<= 0) {
940 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
942 ret
= munmap(sock_info
->wait_shm_mmap
, page_size
);
944 ERR("Error unmapping wait shm");
947 sock_info
->wait_shm_mmap
= NULL
;
952 * Using fork to set umask in the child process (not multi-thread safe).
953 * We deal with the shm_open vs ftruncate race (happening when the
954 * sessiond owns the shm and does not let everybody modify it, to ensure
955 * safety against shm_unlink) by simply letting the mmap fail and
956 * retrying after a few seconds.
957 * For global shm, everybody has rw access to it until the sessiond
961 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
963 int wait_shm_fd
, ret
;
967 * Try to open read-only.
969 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
970 if (wait_shm_fd
>= 0) {
973 size_t bytes_read
= 0;
976 * Try to read the fd. If unable to do so, try opening
980 len
= read(wait_shm_fd
,
981 &((char *) &tmp_read
)[bytes_read
],
982 sizeof(tmp_read
) - bytes_read
);
986 } while ((len
< 0 && errno
== EINTR
)
987 || (len
> 0 && bytes_read
< sizeof(tmp_read
)));
988 if (bytes_read
!= sizeof(tmp_read
)) {
989 ret
= close(wait_shm_fd
);
991 ERR("close wait_shm_fd");
996 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
998 * Real-only open did not work, and it's not because the
999 * entry was not present. It's a failure that prohibits
1002 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
1008 * If the open failed because the file did not exist, or because
1009 * the file was not truncated yet, try creating it ourself.
1011 URCU_TLS(lttng_ust_nest_count
)++;
1013 URCU_TLS(lttng_ust_nest_count
)--;
1018 * Parent: wait for child to return, in which case the
1019 * shared memory map will have been created.
1021 pid
= wait(&status
);
1022 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
1027 * Try to open read-only again after creation.
1029 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
1030 if (wait_shm_fd
< 0) {
1032 * Real-only open did not work. It's a failure
1033 * that prohibits using shm.
1035 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
1039 } else if (pid
== 0) {
1043 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
1044 if (sock_info
->global
)
1045 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
1047 * We're alone in a child process, so we can modify the
1048 * process-wide umask.
1050 umask(~create_mode
);
1052 * Try creating shm (or get rw access).
1053 * We don't do an exclusive open, because we allow other
1054 * processes to create+ftruncate it concurrently.
1056 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
1057 O_RDWR
| O_CREAT
, create_mode
);
1058 if (wait_shm_fd
>= 0) {
1059 ret
= ftruncate(wait_shm_fd
, mmap_size
);
1061 PERROR("ftruncate");
1062 _exit(EXIT_FAILURE
);
1064 _exit(EXIT_SUCCESS
);
1067 * For local shm, we need to have rw access to accept
1068 * opening it: this means the local sessiond will be
1069 * able to wake us up. For global shm, we open it even
1070 * if rw access is not granted, because the root.root
1071 * sessiond will be able to override all rights and wake
1074 if (!sock_info
->global
&& errno
!= EACCES
) {
1075 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
1076 _exit(EXIT_FAILURE
);
1079 * The shm exists, but we cannot open it RW. Report
1082 _exit(EXIT_SUCCESS
);
1087 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
1088 struct stat statbuf
;
1091 * Ensure that our user is the owner of the shm file for
1092 * local shm. If we do not own the file, it means our
1093 * sessiond will not have access to wake us up (there is
1094 * probably a rogue process trying to fake our
1095 * sessiond). Fallback to polling method in this case.
1097 ret
= fstat(wait_shm_fd
, &statbuf
);
1102 if (statbuf
.st_uid
!= getuid())
1108 ret
= close(wait_shm_fd
);
1110 PERROR("Error closing fd");
1116 char *get_map_shm(struct sock_info
*sock_info
)
1119 int wait_shm_fd
, ret
;
1120 char *wait_shm_mmap
;
1122 page_size
= sysconf(_SC_PAGE_SIZE
);
1123 if (page_size
<= 0) {
1127 PERROR("Error in sysconf(_SC_PAGE_SIZE)");
1131 wait_shm_fd
= get_wait_shm(sock_info
, page_size
);
1132 if (wait_shm_fd
< 0) {
1135 wait_shm_mmap
= mmap(NULL
, page_size
, PROT_READ
,
1136 MAP_SHARED
, wait_shm_fd
, 0);
1137 /* close shm fd immediately after taking the mmap reference */
1138 ret
= close(wait_shm_fd
);
1140 PERROR("Error closing fd");
1142 if (wait_shm_mmap
== MAP_FAILED
) {
1143 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1146 return wait_shm_mmap
;
1153 void wait_for_sessiond(struct sock_info
*sock_info
)
1158 if (wait_poll_fallback
) {
1161 if (!sock_info
->wait_shm_mmap
) {
1162 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
1163 if (!sock_info
->wait_shm_mmap
)
1168 DBG("Waiting for %s apps sessiond", sock_info
->name
);
1169 /* Wait for futex wakeup */
1170 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
))
1173 while (futex_async((int32_t *) sock_info
->wait_shm_mmap
,
1174 FUTEX_WAIT
, 0, NULL
, NULL
, 0)) {
1177 /* Value already changed. */
1180 /* Retry if interrupted by signal. */
1181 break; /* Get out of switch. */
1183 wait_poll_fallback
= 1;
1185 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1186 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1187 "Please upgrade your kernel "
1188 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1189 "mainline). LTTng-UST will use polling mode fallback.");
1208 * This thread does not allocate any resource, except within
1209 * handle_message, within mutex protection. This mutex protects against
1211 * The other moment it allocates resources is at socket connection, which
1212 * is also protected by the mutex.
1215 void *ust_listener_thread(void *arg
)
1217 struct sock_info
*sock_info
= arg
;
1218 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
1221 /* Restart trying to connect to the session daemon */
1223 if (prev_connect_failed
) {
1224 /* Wait for sessiond availability with pipe */
1225 wait_for_sessiond(sock_info
);
1229 * Sleep for 5 seconds before retrying after a
1230 * sequence of failure / wait / failure. This
1231 * deals with a killed or broken session daemon.
1237 prev_connect_failed
= 0;
1240 if (sock_info
->socket
!= -1) {
1241 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
1243 ERR("Error closing %s ust cmd socket",
1246 sock_info
->socket
= -1;
1248 if (sock_info
->notify_socket
!= -1) {
1249 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
1251 ERR("Error closing %s ust notify socket",
1254 sock_info
->notify_socket
= -1;
1258 * Register. We need to perform both connect and sending
1259 * registration message before doing the next connect otherwise
1260 * we may reach unix socket connect queue max limits and block
1261 * on the 2nd connect while the session daemon is awaiting the
1262 * first connect registration message.
1264 /* Connect cmd 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
->socket
= ret
;
1290 * Create only one root handle per listener thread for the whole
1291 * process lifetime, so we ensure we get ID which is statically
1292 * assigned to the root handle.
1294 if (sock_info
->root_handle
== -1) {
1295 ret
= lttng_abi_create_root_handle();
1297 ERR("Error creating root handle");
1300 sock_info
->root_handle
= ret
;
1303 ret
= register_to_sessiond(sock_info
->socket
, USTCTL_SOCKET_CMD
);
1305 ERR("Error registering to %s ust cmd socket",
1307 prev_connect_failed
= 1;
1309 * If we cannot register to the sessiond daemon, don't
1310 * delay constructor execution.
1312 ret
= handle_register_done(sock_info
);
1320 /* Connect notify socket */
1321 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1323 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1324 prev_connect_failed
= 1;
1331 * If we cannot find the sessiond daemon, don't delay
1332 * constructor execution.
1334 ret
= handle_register_done(sock_info
);
1339 sock_info
->notify_socket
= ret
;
1341 timeout
= get_notify_sock_timeout();
1344 * Give at least 10ms to sessiond to reply to
1349 ret
= ustcomm_setsockopt_rcv_timeout(sock_info
->notify_socket
,
1352 WARN("Error setting socket receive timeout");
1354 ret
= ustcomm_setsockopt_snd_timeout(sock_info
->notify_socket
,
1357 WARN("Error setting socket send timeout");
1359 } else if (timeout
< -1) {
1360 WARN("Unsupported timeout value %ld", timeout
);
1367 ret
= register_to_sessiond(sock_info
->notify_socket
,
1368 USTCTL_SOCKET_NOTIFY
);
1370 ERR("Error registering to %s ust notify socket",
1372 prev_connect_failed
= 1;
1374 * If we cannot register to the sessiond daemon, don't
1375 * delay constructor execution.
1377 ret
= handle_register_done(sock_info
);
1382 sock
= sock_info
->socket
;
1388 struct ustcomm_ust_msg lum
;
1390 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
1392 case 0: /* orderly shutdown */
1393 DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info
->name
);
1398 * Either sessiond has shutdown or refused us by closing the socket.
1399 * In either case, we don't want to delay construction execution,
1400 * and we need to wait before retry.
1402 prev_connect_failed
= 1;
1404 * If we cannot register to the sessiond daemon, don't
1405 * delay constructor execution.
1407 ret
= handle_register_done(sock_info
);
1412 print_cmd(lum
.cmd
, lum
.handle
);
1413 ret
= handle_message(sock_info
, sock
, &lum
);
1415 ERR("Error handling message for %s socket",
1418 * Close socket if protocol error is
1426 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
1428 DBG("incorrect message size (%s socket): %zd", sock_info
->name
, len
);
1430 if (len
== -ECONNRESET
) {
1431 DBG("%s remote end closed connection", sock_info
->name
);
1442 /* Cleanup socket handles before trying to reconnect */
1443 lttng_ust_objd_table_owner_cleanup(sock_info
);
1445 goto restart
; /* try to reconnect */
1450 pthread_mutex_lock(&ust_exit_mutex
);
1451 sock_info
->thread_active
= 0;
1452 pthread_mutex_unlock(&ust_exit_mutex
);
1457 * Weak symbol to call when the ust malloc wrapper is not loaded.
1459 __attribute__((weak
))
1460 void lttng_ust_malloc_wrapper_init(void)
1465 * sessiond monitoring thread: monitor presence of global and per-user
1466 * sessiond by polling the application common named pipe.
1468 void __attribute__((constructor
)) lttng_ust_init(void)
1470 struct timespec constructor_timeout
;
1471 sigset_t sig_all_blocked
, orig_parent_mask
;
1472 pthread_attr_t thread_attr
;
1476 if (uatomic_xchg(&initialized
, 1) == 1)
1480 * Fixup interdependency between TLS fixup mutex (which happens
1481 * to be the dynamic linker mutex) and ust_lock, taken within
1484 lttng_fixup_urcu_bp_tls();
1485 lttng_fixup_ringbuffer_tls();
1486 lttng_fixup_vtid_tls();
1487 lttng_fixup_nest_count_tls();
1488 lttng_fixup_procname_tls();
1489 lttng_fixup_ust_mutex_nest_tls();
1492 * We want precise control over the order in which we construct
1493 * our sub-libraries vs starting to receive commands from
1494 * sessiond (otherwise leading to errors when trying to create
1495 * sessiond before the init functions are completed).
1499 lttng_ust_clock_init();
1500 lttng_ust_getcpu_init();
1501 lttng_ust_statedump_init();
1502 lttng_ring_buffer_metadata_client_init();
1503 lttng_ring_buffer_client_overwrite_init();
1504 lttng_ring_buffer_client_overwrite_rt_init();
1505 lttng_ring_buffer_client_discard_init();
1506 lttng_ring_buffer_client_discard_rt_init();
1507 lttng_perf_counter_init();
1508 lttng_context_init();
1510 * Invoke ust malloc wrapper init before starting other threads.
1512 lttng_ust_malloc_wrapper_init();
1514 timeout_mode
= get_constructor_timeout(&constructor_timeout
);
1516 ret
= sem_init(&constructor_wait
, 0, 0);
1519 ret
= setup_local_apps();
1521 DBG("local apps setup returned %d", ret
);
1524 /* A new thread created by pthread_create inherits the signal mask
1525 * from the parent. To avoid any signal being received by the
1526 * listener thread, we block all signals temporarily in the parent,
1527 * while we create the listener thread.
1529 sigfillset(&sig_all_blocked
);
1530 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
1532 ERR("pthread_sigmask: %s", strerror(ret
));
1535 ret
= pthread_attr_init(&thread_attr
);
1537 ERR("pthread_attr_init: %s", strerror(ret
));
1539 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
1541 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1544 pthread_mutex_lock(&ust_exit_mutex
);
1545 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1546 ust_listener_thread
, &global_apps
);
1548 ERR("pthread_create global: %s", strerror(ret
));
1550 global_apps
.thread_active
= 1;
1551 pthread_mutex_unlock(&ust_exit_mutex
);
1553 if (local_apps
.allowed
) {
1554 pthread_mutex_lock(&ust_exit_mutex
);
1555 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1556 ust_listener_thread
, &local_apps
);
1558 ERR("pthread_create local: %s", strerror(ret
));
1560 local_apps
.thread_active
= 1;
1561 pthread_mutex_unlock(&ust_exit_mutex
);
1563 handle_register_done(&local_apps
);
1565 ret
= pthread_attr_destroy(&thread_attr
);
1567 ERR("pthread_attr_destroy: %s", strerror(ret
));
1570 /* Restore original signal mask in parent */
1571 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1573 ERR("pthread_sigmask: %s", strerror(ret
));
1576 switch (timeout_mode
) {
1577 case 1: /* timeout wait */
1579 ret
= sem_timedwait(&constructor_wait
,
1580 &constructor_timeout
);
1581 } while (ret
< 0 && errno
== EINTR
);
1582 if (ret
< 0 && errno
== ETIMEDOUT
) {
1583 ERR("Timed out waiting for lttng-sessiond");
1588 case -1:/* wait forever */
1590 ret
= sem_wait(&constructor_wait
);
1591 } while (ret
< 0 && errno
== EINTR
);
1594 case 0: /* no timeout */
1600 void lttng_ust_cleanup(int exiting
)
1602 cleanup_sock_info(&global_apps
, exiting
);
1603 cleanup_sock_info(&local_apps
, exiting
);
1605 * The teardown in this function all affect data structures
1606 * accessed under the UST lock by the listener thread. This
1607 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1608 * that none of these threads are accessing this data at this
1611 lttng_ust_abi_exit();
1612 lttng_ust_events_exit();
1613 lttng_context_exit();
1614 lttng_perf_counter_exit();
1615 lttng_ring_buffer_client_discard_rt_exit();
1616 lttng_ring_buffer_client_discard_exit();
1617 lttng_ring_buffer_client_overwrite_rt_exit();
1618 lttng_ring_buffer_client_overwrite_exit();
1619 lttng_ring_buffer_metadata_client_exit();
1620 lttng_ust_statedump_destroy();
1623 /* Reinitialize values for fork */
1625 lttng_ust_comm_should_quit
= 0;
1630 void __attribute__((destructor
)) lttng_ust_exit(void)
1635 * Using pthread_cancel here because:
1636 * A) we don't want to hang application teardown.
1637 * B) the thread is not allocating any resource.
1641 * Require the communication thread to quit. Synchronize with
1642 * mutexes to ensure it is not in a mutex critical section when
1643 * pthread_cancel is later called.
1646 lttng_ust_comm_should_quit
= 1;
1649 pthread_mutex_lock(&ust_exit_mutex
);
1650 /* cancel threads */
1651 if (global_apps
.thread_active
) {
1652 ret
= pthread_cancel(global_apps
.ust_listener
);
1654 ERR("Error cancelling global ust listener thread: %s",
1657 global_apps
.thread_active
= 0;
1660 if (local_apps
.thread_active
) {
1661 ret
= pthread_cancel(local_apps
.ust_listener
);
1663 ERR("Error cancelling local ust listener thread: %s",
1666 local_apps
.thread_active
= 0;
1669 pthread_mutex_unlock(&ust_exit_mutex
);
1672 * Do NOT join threads: use of sys_futex makes it impossible to
1673 * join the threads without using async-cancel, but async-cancel
1674 * is delivered by a signal, which could hit the target thread
1675 * anywhere in its code path, including while the ust_lock() is
1676 * held, causing a deadlock for the other thread. Let the OS
1677 * cleanup the threads if there are stalled in a syscall.
1679 lttng_ust_cleanup(1);
1683 * We exclude the worker threads across fork and clone (except
1684 * CLONE_VM), because these system calls only keep the forking thread
1685 * running in the child. Therefore, we don't want to call fork or clone
1686 * in the middle of an tracepoint or ust tracing state modification.
1687 * Holding this mutex protects these structures across fork and clone.
1689 void ust_before_fork(sigset_t
*save_sigset
)
1692 * Disable signals. This is to avoid that the child intervenes
1693 * before it is properly setup for tracing. It is safer to
1694 * disable all signals, because then we know we are not breaking
1695 * anything by restoring the original mask.
1700 if (URCU_TLS(lttng_ust_nest_count
))
1702 /* Disable signals */
1703 sigfillset(&all_sigs
);
1704 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1706 PERROR("sigprocmask");
1709 pthread_mutex_lock(&ust_fork_mutex
);
1712 rcu_bp_before_fork();
1715 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1719 DBG("process %d", getpid());
1722 pthread_mutex_unlock(&ust_fork_mutex
);
1724 /* Restore signals */
1725 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1727 PERROR("sigprocmask");
1731 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1733 if (URCU_TLS(lttng_ust_nest_count
))
1735 DBG("process %d", getpid());
1736 rcu_bp_after_fork_parent();
1737 /* Release mutexes and reenable signals */
1738 ust_after_fork_common(restore_sigset
);
1742 * After fork, in the child, we need to cleanup all the leftover state,
1743 * except the worker thread which already magically disappeared thanks
1744 * to the weird Linux fork semantics. After tyding up, we call
1745 * lttng_ust_init() again to start over as a new PID.
1747 * This is meant for forks() that have tracing in the child between the
1748 * fork and following exec call (if there is any).
1750 void ust_after_fork_child(sigset_t
*restore_sigset
)
1752 if (URCU_TLS(lttng_ust_nest_count
))
1754 DBG("process %d", getpid());
1755 /* Release urcu mutexes */
1756 rcu_bp_after_fork_child();
1757 lttng_ust_cleanup(0);
1758 lttng_context_vtid_reset();
1759 /* Release mutexes and reenable signals */
1760 ust_after_fork_common(restore_sigset
);
1764 void lttng_ust_sockinfo_session_enabled(void *owner
)
1766 struct sock_info
*sock_info
= owner
;
1767 sock_info
->statedump_pending
= 1;