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"
58 * Has lttng ust comm constructor been called ?
60 static int initialized
;
63 * The ust_lock/ust_unlock lock is used as a communication thread mutex.
64 * Held when handling a command, also held by fork() to deal with
65 * removal of threads, and by exit path.
67 * The UST lock is the centralized mutex across UST tracing control and
70 * ust_exit_mutex must never nest in ust_mutex.
72 * ust_fork_mutex must never nest in ust_mutex.
74 * ust_mutex_nest is a per-thread nesting counter, allowing the perf
75 * counter lazy initialization called by events within the statedump,
76 * which traces while the ust_mutex is held.
78 * ust_lock nests within the dynamic loader lock (within glibc) because
79 * it is taken within the library constructor.
81 static pthread_mutex_t ust_mutex
= PTHREAD_MUTEX_INITIALIZER
;
83 /* Allow nesting the ust_mutex within the same thread. */
84 static DEFINE_URCU_TLS(int, ust_mutex_nest
);
87 * ust_exit_mutex protects thread_active variable wrt thread exit. It
88 * cannot be done by ust_mutex because pthread_cancel(), which takes an
89 * internal libc lock, cannot nest within ust_mutex.
91 * It never nests within a ust_mutex.
93 static pthread_mutex_t ust_exit_mutex
= PTHREAD_MUTEX_INITIALIZER
;
96 * ust_fork_mutex protects base address statedump tracing against forks. It
97 * prevents the dynamic loader lock to be taken (by base address statedump
98 * tracing) while a fork is happening, thus preventing deadlock issues with
99 * the dynamic loader lock.
101 static pthread_mutex_t ust_fork_mutex
= PTHREAD_MUTEX_INITIALIZER
;
103 /* Should the ust comm thread quit ? */
104 static int lttng_ust_comm_should_quit
;
107 * Return 0 on success, -1 if should quit.
108 * The lock is taken in both cases.
113 sigset_t sig_all_blocked
, orig_mask
;
116 sigfillset(&sig_all_blocked
);
117 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
119 ERR("pthread_sigmask: %s", strerror(ret
));
121 if (!URCU_TLS(ust_mutex_nest
)++)
122 pthread_mutex_lock(&ust_mutex
);
123 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
125 ERR("pthread_sigmask: %s", strerror(ret
));
127 if (lttng_ust_comm_should_quit
) {
135 * ust_lock_nocheck() can be used in constructors/destructors, because
136 * they are already nested within the dynamic loader lock, and therefore
137 * have exclusive access against execution of liblttng-ust destructor.
140 void ust_lock_nocheck(void)
142 sigset_t sig_all_blocked
, orig_mask
;
145 sigfillset(&sig_all_blocked
);
146 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
148 ERR("pthread_sigmask: %s", strerror(ret
));
150 if (!URCU_TLS(ust_mutex_nest
)++)
151 pthread_mutex_lock(&ust_mutex
);
152 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
154 ERR("pthread_sigmask: %s", strerror(ret
));
161 void ust_unlock(void)
163 sigset_t sig_all_blocked
, orig_mask
;
166 sigfillset(&sig_all_blocked
);
167 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_mask
);
169 ERR("pthread_sigmask: %s", strerror(ret
));
171 if (!--URCU_TLS(ust_mutex_nest
))
172 pthread_mutex_unlock(&ust_mutex
);
173 ret
= pthread_sigmask(SIG_SETMASK
, &orig_mask
, NULL
);
175 ERR("pthread_sigmask: %s", strerror(ret
));
180 * Wait for either of these before continuing to the main
182 * - the register_done message from sessiond daemon
183 * (will let the sessiond daemon enable sessions before main
185 * - sessiond daemon is not reachable.
186 * - timeout (ensuring applications are resilient to session
189 static sem_t constructor_wait
;
191 * Doing this for both the global and local sessiond.
193 static int sem_count
= { 2 };
196 * Counting nesting within lttng-ust. Used to ensure that calling fork()
197 * from liblttng-ust does not execute the pre/post fork handlers.
199 static DEFINE_URCU_TLS(int, lttng_ust_nest_count
);
202 * Info about socket and associated listener thread.
206 pthread_t ust_listener
; /* listener thread */
208 int constructor_sem_posted
;
213 char sock_path
[PATH_MAX
];
217 char wait_shm_path
[PATH_MAX
];
219 /* Keep track of lazy state dump not performed yet. */
220 int statedump_pending
;
223 /* Socket from app (connect) to session daemon (listen) for communication */
224 struct sock_info global_apps
= {
232 .sock_path
= LTTNG_DEFAULT_RUNDIR
"/" LTTNG_UST_SOCK_FILENAME
,
236 .wait_shm_path
= "/" LTTNG_UST_WAIT_FILENAME
,
238 .statedump_pending
= 0,
241 /* TODO: allow global_apps_sock_path override */
243 struct sock_info local_apps
= {
247 .allowed
= 0, /* Check setuid bit first */
253 .statedump_pending
= 0,
256 static int wait_poll_fallback
;
258 static const char *cmd_name_mapping
[] = {
259 [ LTTNG_UST_RELEASE
] = "Release",
260 [ LTTNG_UST_SESSION
] = "Create Session",
261 [ LTTNG_UST_TRACER_VERSION
] = "Get Tracer Version",
263 [ LTTNG_UST_TRACEPOINT_LIST
] = "Create Tracepoint List",
264 [ LTTNG_UST_WAIT_QUIESCENT
] = "Wait for Quiescent State",
265 [ LTTNG_UST_REGISTER_DONE
] = "Registration Done",
266 [ LTTNG_UST_TRACEPOINT_FIELD_LIST
] = "Create Tracepoint Field List",
268 /* Session FD commands */
269 [ LTTNG_UST_CHANNEL
] = "Create Channel",
270 [ LTTNG_UST_SESSION_START
] = "Start Session",
271 [ LTTNG_UST_SESSION_STOP
] = "Stop Session",
273 /* Channel FD commands */
274 [ LTTNG_UST_STREAM
] = "Create Stream",
275 [ LTTNG_UST_EVENT
] = "Create Event",
277 /* Event and Channel FD commands */
278 [ LTTNG_UST_CONTEXT
] = "Create Context",
279 [ LTTNG_UST_FLUSH_BUFFER
] = "Flush Buffer",
281 /* Event, Channel and Session commands */
282 [ LTTNG_UST_ENABLE
] = "Enable",
283 [ LTTNG_UST_DISABLE
] = "Disable",
285 /* Tracepoint list commands */
286 [ LTTNG_UST_TRACEPOINT_LIST_GET
] = "List Next Tracepoint",
287 [ LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
] = "List Next Tracepoint Field",
289 /* Event FD commands */
290 [ LTTNG_UST_FILTER
] = "Create Filter",
291 [ LTTNG_UST_EXCLUSION
] = "Add exclusions to event",
294 static const char *str_timeout
;
295 static int got_timeout_env
;
297 extern void lttng_ring_buffer_client_overwrite_init(void);
298 extern void lttng_ring_buffer_client_overwrite_rt_init(void);
299 extern void lttng_ring_buffer_client_discard_init(void);
300 extern void lttng_ring_buffer_client_discard_rt_init(void);
301 extern void lttng_ring_buffer_metadata_client_init(void);
302 extern void lttng_ring_buffer_client_overwrite_exit(void);
303 extern void lttng_ring_buffer_client_overwrite_rt_exit(void);
304 extern void lttng_ring_buffer_client_discard_exit(void);
305 extern void lttng_ring_buffer_client_discard_rt_exit(void);
306 extern void lttng_ring_buffer_metadata_client_exit(void);
309 * Returns the HOME directory path. Caller MUST NOT free(3) the returned
313 const char *get_lttng_home_dir(void)
317 val
= (const char *) getenv("LTTNG_HOME");
321 return (const char *) getenv("HOME");
325 * Force a read (imply TLS fixup for dlopen) of TLS variables.
328 void lttng_fixup_nest_count_tls(void)
330 asm volatile ("" : : "m" (URCU_TLS(lttng_ust_nest_count
)));
334 void lttng_fixup_ust_mutex_nest_tls(void)
336 asm volatile ("" : : "m" (URCU_TLS(ust_mutex_nest
)));
343 void lttng_fixup_urcu_bp_tls(void)
349 int lttng_get_notify_socket(void *owner
)
351 struct sock_info
*info
= owner
;
353 return info
->notify_socket
;
357 void print_cmd(int cmd
, int handle
)
359 const char *cmd_name
= "Unknown";
361 if (cmd
>= 0 && cmd
< LTTNG_ARRAY_SIZE(cmd_name_mapping
)
362 && cmd_name_mapping
[cmd
]) {
363 cmd_name
= cmd_name_mapping
[cmd
];
365 DBG("Message Received \"%s\" (%d), Handle \"%s\" (%d)",
367 lttng_ust_obj_get_name(handle
), handle
);
371 int setup_local_apps(void)
373 const char *home_dir
;
378 * Disallow per-user tracing for setuid binaries.
380 if (uid
!= geteuid()) {
381 assert(local_apps
.allowed
== 0);
384 home_dir
= get_lttng_home_dir();
386 WARN("HOME environment variable not set. Disabling LTTng-UST per-user tracing.");
387 assert(local_apps
.allowed
== 0);
390 local_apps
.allowed
= 1;
391 snprintf(local_apps
.sock_path
, PATH_MAX
, "%s/%s/%s",
393 LTTNG_DEFAULT_HOME_RUNDIR
,
394 LTTNG_UST_SOCK_FILENAME
);
395 snprintf(local_apps
.wait_shm_path
, PATH_MAX
, "/%s-%u",
396 LTTNG_UST_WAIT_FILENAME
,
402 * Get notify_sock timeout, in ms.
403 * -1: don't wait. 0: wait forever. >0: timeout, in ms.
406 long get_timeout(void)
408 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
410 if (!got_timeout_env
) {
411 str_timeout
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
415 constructor_delay_ms
= strtol(str_timeout
, NULL
, 10);
416 return constructor_delay_ms
;
420 long get_notify_sock_timeout(void)
422 return get_timeout();
426 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
429 int get_constructor_timeout(struct timespec
*constructor_timeout
)
431 long constructor_delay_ms
;
434 constructor_delay_ms
= get_timeout();
436 switch (constructor_delay_ms
) {
437 case -1:/* fall-through */
439 return constructor_delay_ms
;
445 * If we are unable to find the current time, don't wait.
447 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
451 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
452 constructor_timeout
->tv_nsec
+=
453 (constructor_delay_ms
% 1000UL) * 1000000UL;
454 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
455 constructor_timeout
->tv_sec
++;
456 constructor_timeout
->tv_nsec
-= 1000000000UL;
462 int register_to_sessiond(int socket
, enum ustctl_socket_type type
)
464 return ustcomm_send_reg_msg(socket
,
467 lttng_alignof(uint8_t) * CHAR_BIT
,
468 lttng_alignof(uint16_t) * CHAR_BIT
,
469 lttng_alignof(uint32_t) * CHAR_BIT
,
470 lttng_alignof(uint64_t) * CHAR_BIT
,
471 lttng_alignof(unsigned long) * CHAR_BIT
);
475 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
479 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
482 DBG("message successfully sent");
485 if (len
== -ECONNRESET
) {
486 DBG("remote end closed connection");
491 DBG("incorrect message size: %zd", len
);
497 int handle_register_done(struct sock_info
*sock_info
)
501 if (sock_info
->constructor_sem_posted
)
503 sock_info
->constructor_sem_posted
= 1;
504 if (uatomic_read(&sem_count
) <= 0) {
507 ret
= uatomic_add_return(&sem_count
, -1);
509 ret
= sem_post(&constructor_wait
);
516 * Only execute pending statedump after the constructor semaphore has
517 * been posted by each listener thread. This means statedump will only
518 * be performed after the "registration done" command is received from
519 * each session daemon the application is connected to.
521 * This ensures we don't run into deadlock issues with the dynamic
522 * loader mutex, which is held while the constructor is called and
523 * waiting on the constructor semaphore. All operations requiring this
524 * dynamic loader lock need to be postponed using this mechanism.
527 void handle_pending_statedump(struct sock_info
*sock_info
)
529 int ctor_passed
= sock_info
->constructor_sem_posted
;
531 if (ctor_passed
&& sock_info
->statedump_pending
) {
532 sock_info
->statedump_pending
= 0;
533 pthread_mutex_lock(&ust_fork_mutex
);
534 lttng_handle_pending_statedump(sock_info
);
535 pthread_mutex_unlock(&ust_fork_mutex
);
540 int handle_message(struct sock_info
*sock_info
,
541 int sock
, struct ustcomm_ust_msg
*lum
)
544 const struct lttng_ust_objd_ops
*ops
;
545 struct ustcomm_ust_reply lur
;
549 memset(&lur
, 0, sizeof(lur
));
552 ret
= -LTTNG_UST_ERR_EXITING
;
556 ops
= objd_ops(lum
->handle
);
563 case LTTNG_UST_REGISTER_DONE
:
564 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
565 ret
= handle_register_done(sock_info
);
569 case LTTNG_UST_RELEASE
:
570 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
573 ret
= lttng_ust_objd_unref(lum
->handle
, 1);
575 case LTTNG_UST_FILTER
:
577 /* Receive filter data */
578 struct lttng_ust_filter_bytecode_node
*bytecode
;
580 if (lum
->u
.filter
.data_size
> FILTER_BYTECODE_MAX_LEN
) {
581 ERR("Filter data size is too large: %u bytes",
582 lum
->u
.filter
.data_size
);
587 if (lum
->u
.filter
.reloc_offset
> lum
->u
.filter
.data_size
) {
588 ERR("Filter reloc offset %u is not within data",
589 lum
->u
.filter
.reloc_offset
);
594 bytecode
= zmalloc(sizeof(*bytecode
) + lum
->u
.filter
.data_size
);
599 len
= ustcomm_recv_unix_sock(sock
, bytecode
->bc
.data
,
600 lum
->u
.filter
.data_size
);
602 case 0: /* orderly shutdown */
607 if (len
== lum
->u
.filter
.data_size
) {
608 DBG("filter data received");
610 } else if (len
< 0) {
611 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
612 if (len
== -ECONNRESET
) {
613 ERR("%s remote end closed connection", sock_info
->name
);
622 DBG("incorrect filter data message size: %zd", len
);
628 bytecode
->bc
.len
= lum
->u
.filter
.data_size
;
629 bytecode
->bc
.reloc_offset
= lum
->u
.filter
.reloc_offset
;
630 bytecode
->bc
.seqnum
= lum
->u
.filter
.seqnum
;
632 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
633 (unsigned long) bytecode
,
638 /* don't free bytecode if everything went fine. */
645 case LTTNG_UST_EXCLUSION
:
647 /* Receive exclusion names */
648 struct lttng_ust_excluder_node
*node
;
651 count
= lum
->u
.exclusion
.count
;
653 /* There are no names to read */
657 node
= zmalloc(sizeof(*node
) +
658 count
* LTTNG_UST_SYM_NAME_LEN
);
663 node
->excluder
.count
= count
;
664 len
= ustcomm_recv_unix_sock(sock
, node
->excluder
.names
,
665 count
* LTTNG_UST_SYM_NAME_LEN
);
667 case 0: /* orderly shutdown */
672 if (len
== count
* LTTNG_UST_SYM_NAME_LEN
) {
673 DBG("Exclusion data received");
675 } else if (len
< 0) {
676 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
677 if (len
== -ECONNRESET
) {
678 ERR("%s remote end closed connection", sock_info
->name
);
687 DBG("Incorrect exclusion data message size: %zd", len
);
694 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
695 (unsigned long) node
,
700 /* Don't free exclusion data if everything went fine. */
707 case LTTNG_UST_CHANNEL
:
712 len
= ustcomm_recv_channel_from_sessiond(sock
,
713 &chan_data
, lum
->u
.channel
.len
,
716 case 0: /* orderly shutdown */
720 if (len
== lum
->u
.channel
.len
) {
721 DBG("channel data received");
723 } else if (len
< 0) {
724 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
725 if (len
== -ECONNRESET
) {
726 ERR("%s remote end closed connection", sock_info
->name
);
733 DBG("incorrect channel data message size: %zd", len
);
738 args
.channel
.chan_data
= chan_data
;
739 args
.channel
.wakeup_fd
= wakeup_fd
;
741 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
742 (unsigned long) &lum
->u
,
748 case LTTNG_UST_STREAM
:
750 /* Receive shm_fd, wakeup_fd */
751 ret
= ustcomm_recv_stream_from_sessiond(sock
,
754 &args
.stream
.wakeup_fd
);
759 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
760 (unsigned long) &lum
->u
,
768 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
769 (unsigned long) &lum
->u
,
777 lur
.handle
= lum
->handle
;
781 lur
.ret_code
= LTTNG_UST_OK
;
784 * Use -LTTNG_UST_ERR as wildcard for UST internal
785 * error that are not caused by the transport, except if
786 * we already have a more precise error message to
789 if (ret
> -LTTNG_UST_ERR
) {
790 /* Translate code to UST error. */
793 lur
.ret_code
= -LTTNG_UST_ERR_EXIST
;
796 lur
.ret_code
= -LTTNG_UST_ERR_INVAL
;
799 lur
.ret_code
= -LTTNG_UST_ERR_NOENT
;
802 lur
.ret_code
= -LTTNG_UST_ERR_PERM
;
805 lur
.ret_code
= -LTTNG_UST_ERR_NOSYS
;
808 lur
.ret_code
= -LTTNG_UST_ERR
;
817 case LTTNG_UST_TRACER_VERSION
:
818 lur
.u
.version
= lum
->u
.version
;
820 case LTTNG_UST_TRACEPOINT_LIST_GET
:
821 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
825 DBG("Return value: %d", lur
.ret_val
);
826 ret
= send_reply(sock
, &lur
);
828 DBG("error sending reply");
833 * LTTNG_UST_TRACEPOINT_FIELD_LIST_GET needs to send the field
836 if (lur
.ret_code
== LTTNG_UST_OK
) {
838 case LTTNG_UST_TRACEPOINT_FIELD_LIST_GET
:
839 len
= ustcomm_send_unix_sock(sock
,
840 &args
.field_list
.entry
,
841 sizeof(args
.field_list
.entry
));
846 if (len
!= sizeof(args
.field_list
.entry
)) {
857 * Performed delayed statedump operations outside of the UST
858 * lock. We need to take the dynamic loader lock before we take
859 * the UST lock internally within handle_pending_statedump().
861 handle_pending_statedump(sock_info
);
867 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
871 if (sock_info
->root_handle
!= -1) {
872 ret
= lttng_ust_objd_unref(sock_info
->root_handle
, 1);
874 ERR("Error unref root handle");
876 sock_info
->root_handle
= -1;
878 sock_info
->constructor_sem_posted
= 0;
881 * wait_shm_mmap, socket and notify socket are used by listener
882 * threads outside of the ust lock, so we cannot tear them down
883 * ourselves, because we cannot join on these threads. Leave
884 * responsibility of cleaning up these resources to the OS
890 if (sock_info
->socket
!= -1) {
891 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
893 ERR("Error closing ust cmd socket");
895 sock_info
->socket
= -1;
897 if (sock_info
->notify_socket
!= -1) {
898 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
900 ERR("Error closing ust notify socket");
902 sock_info
->notify_socket
= -1;
904 if (sock_info
->wait_shm_mmap
) {
907 page_size
= sysconf(_SC_PAGE_SIZE
);
909 ret
= munmap(sock_info
->wait_shm_mmap
, page_size
);
911 ERR("Error unmapping wait shm");
914 sock_info
->wait_shm_mmap
= NULL
;
919 * Using fork to set umask in the child process (not multi-thread safe).
920 * We deal with the shm_open vs ftruncate race (happening when the
921 * sessiond owns the shm and does not let everybody modify it, to ensure
922 * safety against shm_unlink) by simply letting the mmap fail and
923 * retrying after a few seconds.
924 * For global shm, everybody has rw access to it until the sessiond
928 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
930 int wait_shm_fd
, ret
;
934 * Try to open read-only.
936 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
937 if (wait_shm_fd
>= 0) {
940 size_t bytes_read
= 0;
943 * Try to read the fd. If unable to do so, try opening
947 len
= read(wait_shm_fd
,
948 &((char *) &tmp_read
)[bytes_read
],
949 sizeof(tmp_read
) - bytes_read
);
953 } while ((len
< 0 && errno
== EINTR
)
954 || (len
> 0 && bytes_read
< sizeof(tmp_read
)));
955 if (bytes_read
!= sizeof(tmp_read
)) {
956 ret
= close(wait_shm_fd
);
958 ERR("close wait_shm_fd");
963 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
965 * Real-only open did not work, and it's not because the
966 * entry was not present. It's a failure that prohibits
969 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
975 * If the open failed because the file did not exist, or because
976 * the file was not truncated yet, try creating it ourself.
978 URCU_TLS(lttng_ust_nest_count
)++;
980 URCU_TLS(lttng_ust_nest_count
)--;
985 * Parent: wait for child to return, in which case the
986 * shared memory map will have been created.
989 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
994 * Try to open read-only again after creation.
996 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
997 if (wait_shm_fd
< 0) {
999 * Real-only open did not work. It's a failure
1000 * that prohibits using shm.
1002 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
1006 } else if (pid
== 0) {
1010 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
1011 if (sock_info
->global
)
1012 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
1014 * We're alone in a child process, so we can modify the
1015 * process-wide umask.
1017 umask(~create_mode
);
1019 * Try creating shm (or get rw access).
1020 * We don't do an exclusive open, because we allow other
1021 * processes to create+ftruncate it concurrently.
1023 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
1024 O_RDWR
| O_CREAT
, create_mode
);
1025 if (wait_shm_fd
>= 0) {
1026 ret
= ftruncate(wait_shm_fd
, mmap_size
);
1028 PERROR("ftruncate");
1029 _exit(EXIT_FAILURE
);
1031 _exit(EXIT_SUCCESS
);
1034 * For local shm, we need to have rw access to accept
1035 * opening it: this means the local sessiond will be
1036 * able to wake us up. For global shm, we open it even
1037 * if rw access is not granted, because the root.root
1038 * sessiond will be able to override all rights and wake
1041 if (!sock_info
->global
&& errno
!= EACCES
) {
1042 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
1043 _exit(EXIT_FAILURE
);
1046 * The shm exists, but we cannot open it RW. Report
1049 _exit(EXIT_SUCCESS
);
1054 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
1055 struct stat statbuf
;
1058 * Ensure that our user is the owner of the shm file for
1059 * local shm. If we do not own the file, it means our
1060 * sessiond will not have access to wake us up (there is
1061 * probably a rogue process trying to fake our
1062 * sessiond). Fallback to polling method in this case.
1064 ret
= fstat(wait_shm_fd
, &statbuf
);
1069 if (statbuf
.st_uid
!= getuid())
1075 ret
= close(wait_shm_fd
);
1077 PERROR("Error closing fd");
1083 char *get_map_shm(struct sock_info
*sock_info
)
1086 int wait_shm_fd
, ret
;
1087 char *wait_shm_mmap
;
1089 page_size
= sysconf(_SC_PAGE_SIZE
);
1090 if (page_size
< 0) {
1094 wait_shm_fd
= get_wait_shm(sock_info
, page_size
);
1095 if (wait_shm_fd
< 0) {
1098 wait_shm_mmap
= mmap(NULL
, page_size
, PROT_READ
,
1099 MAP_SHARED
, wait_shm_fd
, 0);
1100 /* close shm fd immediately after taking the mmap reference */
1101 ret
= close(wait_shm_fd
);
1103 PERROR("Error closing fd");
1105 if (wait_shm_mmap
== MAP_FAILED
) {
1106 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
1109 return wait_shm_mmap
;
1116 void wait_for_sessiond(struct sock_info
*sock_info
)
1123 if (wait_poll_fallback
) {
1126 if (!sock_info
->wait_shm_mmap
) {
1127 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
1128 if (!sock_info
->wait_shm_mmap
)
1133 DBG("Waiting for %s apps sessiond", sock_info
->name
);
1134 /* Wait for futex wakeup */
1135 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
1136 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
1137 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
1139 if (errno
== EFAULT
) {
1140 wait_poll_fallback
= 1;
1142 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
1143 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
1144 "Please upgrade your kernel "
1145 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
1146 "mainline). LTTng-UST will use polling mode fallback.");
1164 * This thread does not allocate any resource, except within
1165 * handle_message, within mutex protection. This mutex protects against
1167 * The other moment it allocates resources is at socket connection, which
1168 * is also protected by the mutex.
1171 void *ust_listener_thread(void *arg
)
1173 struct sock_info
*sock_info
= arg
;
1174 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
1177 /* Restart trying to connect to the session daemon */
1179 if (prev_connect_failed
) {
1180 /* Wait for sessiond availability with pipe */
1181 wait_for_sessiond(sock_info
);
1185 * Sleep for 5 seconds before retrying after a
1186 * sequence of failure / wait / failure. This
1187 * deals with a killed or broken session daemon.
1192 prev_connect_failed
= 0;
1195 if (sock_info
->socket
!= -1) {
1196 ret
= ustcomm_close_unix_sock(sock_info
->socket
);
1198 ERR("Error closing %s ust cmd socket",
1201 sock_info
->socket
= -1;
1203 if (sock_info
->notify_socket
!= -1) {
1204 ret
= ustcomm_close_unix_sock(sock_info
->notify_socket
);
1206 ERR("Error closing %s ust notify socket",
1209 sock_info
->notify_socket
= -1;
1213 * Register. We need to perform both connect and sending
1214 * registration message before doing the next connect otherwise
1215 * we may reach unix socket connect queue max limits and block
1216 * on the 2nd connect while the session daemon is awaiting the
1217 * first connect registration message.
1219 /* Connect cmd socket */
1220 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1222 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1223 prev_connect_failed
= 1;
1230 * If we cannot find the sessiond daemon, don't delay
1231 * constructor execution.
1233 ret
= handle_register_done(sock_info
);
1238 sock_info
->socket
= ret
;
1245 * Create only one root handle per listener thread for the whole
1246 * process lifetime, so we ensure we get ID which is statically
1247 * assigned to the root handle.
1249 if (sock_info
->root_handle
== -1) {
1250 ret
= lttng_abi_create_root_handle();
1252 ERR("Error creating root handle");
1255 sock_info
->root_handle
= ret
;
1258 ret
= register_to_sessiond(sock_info
->socket
, USTCTL_SOCKET_CMD
);
1260 ERR("Error registering to %s ust cmd socket",
1262 prev_connect_failed
= 1;
1264 * If we cannot register to the sessiond daemon, don't
1265 * delay constructor execution.
1267 ret
= handle_register_done(sock_info
);
1275 /* Connect notify socket */
1276 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
1278 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
1279 prev_connect_failed
= 1;
1286 * If we cannot find the sessiond daemon, don't delay
1287 * constructor execution.
1289 ret
= handle_register_done(sock_info
);
1294 sock_info
->notify_socket
= ret
;
1296 timeout
= get_notify_sock_timeout();
1299 * Give at least 10ms to sessiond to reply to
1304 ret
= ustcomm_setsockopt_rcv_timeout(sock_info
->notify_socket
,
1307 WARN("Error setting socket receive timeout");
1309 ret
= ustcomm_setsockopt_snd_timeout(sock_info
->notify_socket
,
1312 WARN("Error setting socket send timeout");
1314 } else if (timeout
< -1) {
1315 WARN("Unsupported timeout value %ld", timeout
);
1322 ret
= register_to_sessiond(sock_info
->notify_socket
,
1323 USTCTL_SOCKET_NOTIFY
);
1325 ERR("Error registering to %s ust notify socket",
1327 prev_connect_failed
= 1;
1329 * If we cannot register to the sessiond daemon, don't
1330 * delay constructor execution.
1332 ret
= handle_register_done(sock_info
);
1337 sock
= sock_info
->socket
;
1343 struct ustcomm_ust_msg lum
;
1345 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
1347 case 0: /* orderly shutdown */
1348 DBG("%s lttng-sessiond has performed an orderly shutdown", sock_info
->name
);
1353 * Either sessiond has shutdown or refused us by closing the socket.
1354 * In either case, we don't want to delay construction execution,
1355 * and we need to wait before retry.
1357 prev_connect_failed
= 1;
1359 * If we cannot register to the sessiond daemon, don't
1360 * delay constructor execution.
1362 ret
= handle_register_done(sock_info
);
1367 print_cmd(lum
.cmd
, lum
.handle
);
1368 ret
= handle_message(sock_info
, sock
, &lum
);
1370 ERR("Error handling message for %s socket", sock_info
->name
);
1375 DBG("Receive failed from lttng-sessiond with errno %d", (int) -len
);
1377 DBG("incorrect message size (%s socket): %zd", sock_info
->name
, len
);
1379 if (len
== -ECONNRESET
) {
1380 DBG("%s remote end closed connection", sock_info
->name
);
1391 /* Cleanup socket handles before trying to reconnect */
1392 lttng_ust_objd_table_owner_cleanup(sock_info
);
1394 goto restart
; /* try to reconnect */
1399 pthread_mutex_lock(&ust_exit_mutex
);
1400 sock_info
->thread_active
= 0;
1401 pthread_mutex_unlock(&ust_exit_mutex
);
1406 * Weak symbol to call when the ust malloc wrapper is not loaded.
1408 __attribute__((weak
))
1409 void lttng_ust_malloc_wrapper_init(void)
1414 * sessiond monitoring thread: monitor presence of global and per-user
1415 * sessiond by polling the application common named pipe.
1417 void __attribute__((constructor
)) lttng_ust_init(void)
1419 struct timespec constructor_timeout
;
1420 sigset_t sig_all_blocked
, orig_parent_mask
;
1421 pthread_attr_t thread_attr
;
1425 if (uatomic_xchg(&initialized
, 1) == 1)
1429 * Fixup interdependency between TLS fixup mutex (which happens
1430 * to be the dynamic linker mutex) and ust_lock, taken within
1433 lttng_fixup_urcu_bp_tls();
1434 lttng_fixup_ringbuffer_tls();
1435 lttng_fixup_vtid_tls();
1436 lttng_fixup_nest_count_tls();
1437 lttng_fixup_procname_tls();
1438 lttng_fixup_ust_mutex_nest_tls();
1441 * We want precise control over the order in which we construct
1442 * our sub-libraries vs starting to receive commands from
1443 * sessiond (otherwise leading to errors when trying to create
1444 * sessiond before the init functions are completed).
1448 lttng_ust_clock_init();
1449 lttng_ust_baddr_statedump_init();
1450 lttng_ring_buffer_metadata_client_init();
1451 lttng_ring_buffer_client_overwrite_init();
1452 lttng_ring_buffer_client_overwrite_rt_init();
1453 lttng_ring_buffer_client_discard_init();
1454 lttng_ring_buffer_client_discard_rt_init();
1455 lttng_perf_counter_init();
1456 lttng_context_init();
1458 * Invoke ust malloc wrapper init before starting other threads.
1460 lttng_ust_malloc_wrapper_init();
1462 timeout_mode
= get_constructor_timeout(&constructor_timeout
);
1464 ret
= sem_init(&constructor_wait
, 0, 0);
1467 ret
= setup_local_apps();
1469 DBG("local apps setup returned %d", ret
);
1472 /* A new thread created by pthread_create inherits the signal mask
1473 * from the parent. To avoid any signal being received by the
1474 * listener thread, we block all signals temporarily in the parent,
1475 * while we create the listener thread.
1477 sigfillset(&sig_all_blocked
);
1478 ret
= pthread_sigmask(SIG_SETMASK
, &sig_all_blocked
, &orig_parent_mask
);
1480 ERR("pthread_sigmask: %s", strerror(ret
));
1483 ret
= pthread_attr_init(&thread_attr
);
1485 ERR("pthread_attr_init: %s", strerror(ret
));
1487 ret
= pthread_attr_setdetachstate(&thread_attr
, PTHREAD_CREATE_DETACHED
);
1489 ERR("pthread_attr_setdetachstate: %s", strerror(ret
));
1492 pthread_mutex_lock(&ust_exit_mutex
);
1493 ret
= pthread_create(&global_apps
.ust_listener
, &thread_attr
,
1494 ust_listener_thread
, &global_apps
);
1496 ERR("pthread_create global: %s", strerror(ret
));
1498 global_apps
.thread_active
= 1;
1499 pthread_mutex_unlock(&ust_exit_mutex
);
1501 if (local_apps
.allowed
) {
1502 pthread_mutex_lock(&ust_exit_mutex
);
1503 ret
= pthread_create(&local_apps
.ust_listener
, &thread_attr
,
1504 ust_listener_thread
, &local_apps
);
1506 ERR("pthread_create local: %s", strerror(ret
));
1508 local_apps
.thread_active
= 1;
1509 pthread_mutex_unlock(&ust_exit_mutex
);
1511 handle_register_done(&local_apps
);
1513 ret
= pthread_attr_destroy(&thread_attr
);
1515 ERR("pthread_attr_destroy: %s", strerror(ret
));
1518 /* Restore original signal mask in parent */
1519 ret
= pthread_sigmask(SIG_SETMASK
, &orig_parent_mask
, NULL
);
1521 ERR("pthread_sigmask: %s", strerror(ret
));
1524 switch (timeout_mode
) {
1525 case 1: /* timeout wait */
1527 ret
= sem_timedwait(&constructor_wait
,
1528 &constructor_timeout
);
1529 } while (ret
< 0 && errno
== EINTR
);
1530 if (ret
< 0 && errno
== ETIMEDOUT
) {
1531 ERR("Timed out waiting for lttng-sessiond");
1536 case -1:/* wait forever */
1538 ret
= sem_wait(&constructor_wait
);
1539 } while (ret
< 0 && errno
== EINTR
);
1542 case 0: /* no timeout */
1548 void lttng_ust_cleanup(int exiting
)
1550 cleanup_sock_info(&global_apps
, exiting
);
1551 if (local_apps
.allowed
) {
1552 cleanup_sock_info(&local_apps
, exiting
);
1555 * The teardown in this function all affect data structures
1556 * accessed under the UST lock by the listener thread. This
1557 * lock, along with the lttng_ust_comm_should_quit flag, ensure
1558 * that none of these threads are accessing this data at this
1561 lttng_ust_abi_exit();
1562 lttng_ust_events_exit();
1563 lttng_context_exit();
1564 lttng_perf_counter_exit();
1565 lttng_ring_buffer_client_discard_rt_exit();
1566 lttng_ring_buffer_client_discard_exit();
1567 lttng_ring_buffer_client_overwrite_rt_exit();
1568 lttng_ring_buffer_client_overwrite_exit();
1569 lttng_ring_buffer_metadata_client_exit();
1570 lttng_ust_baddr_statedump_destroy();
1573 /* Reinitialize values for fork */
1575 lttng_ust_comm_should_quit
= 0;
1580 void __attribute__((destructor
)) lttng_ust_exit(void)
1585 * Using pthread_cancel here because:
1586 * A) we don't want to hang application teardown.
1587 * B) the thread is not allocating any resource.
1591 * Require the communication thread to quit. Synchronize with
1592 * mutexes to ensure it is not in a mutex critical section when
1593 * pthread_cancel is later called.
1596 lttng_ust_comm_should_quit
= 1;
1599 pthread_mutex_lock(&ust_exit_mutex
);
1600 /* cancel threads */
1601 if (global_apps
.thread_active
) {
1602 ret
= pthread_cancel(global_apps
.ust_listener
);
1604 ERR("Error cancelling global ust listener thread: %s",
1607 global_apps
.thread_active
= 0;
1610 if (local_apps
.thread_active
) {
1611 ret
= pthread_cancel(local_apps
.ust_listener
);
1613 ERR("Error cancelling local ust listener thread: %s",
1616 local_apps
.thread_active
= 0;
1619 pthread_mutex_unlock(&ust_exit_mutex
);
1622 * Do NOT join threads: use of sys_futex makes it impossible to
1623 * join the threads without using async-cancel, but async-cancel
1624 * is delivered by a signal, which could hit the target thread
1625 * anywhere in its code path, including while the ust_lock() is
1626 * held, causing a deadlock for the other thread. Let the OS
1627 * cleanup the threads if there are stalled in a syscall.
1629 lttng_ust_cleanup(1);
1633 * We exclude the worker threads across fork and clone (except
1634 * CLONE_VM), because these system calls only keep the forking thread
1635 * running in the child. Therefore, we don't want to call fork or clone
1636 * in the middle of an tracepoint or ust tracing state modification.
1637 * Holding this mutex protects these structures across fork and clone.
1639 void ust_before_fork(sigset_t
*save_sigset
)
1642 * Disable signals. This is to avoid that the child intervenes
1643 * before it is properly setup for tracing. It is safer to
1644 * disable all signals, because then we know we are not breaking
1645 * anything by restoring the original mask.
1650 if (URCU_TLS(lttng_ust_nest_count
))
1652 /* Disable signals */
1653 sigfillset(&all_sigs
);
1654 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
1656 PERROR("sigprocmask");
1659 pthread_mutex_lock(&ust_fork_mutex
);
1662 rcu_bp_before_fork();
1665 static void ust_after_fork_common(sigset_t
*restore_sigset
)
1669 DBG("process %d", getpid());
1672 pthread_mutex_unlock(&ust_fork_mutex
);
1674 /* Restore signals */
1675 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1677 PERROR("sigprocmask");
1681 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1683 if (URCU_TLS(lttng_ust_nest_count
))
1685 DBG("process %d", getpid());
1686 rcu_bp_after_fork_parent();
1687 /* Release mutexes and reenable signals */
1688 ust_after_fork_common(restore_sigset
);
1692 * After fork, in the child, we need to cleanup all the leftover state,
1693 * except the worker thread which already magically disappeared thanks
1694 * to the weird Linux fork semantics. After tyding up, we call
1695 * lttng_ust_init() again to start over as a new PID.
1697 * This is meant for forks() that have tracing in the child between the
1698 * fork and following exec call (if there is any).
1700 void ust_after_fork_child(sigset_t
*restore_sigset
)
1702 if (URCU_TLS(lttng_ust_nest_count
))
1704 DBG("process %d", getpid());
1705 /* Release urcu mutexes */
1706 rcu_bp_after_fork_child();
1707 lttng_ust_cleanup(0);
1708 lttng_context_vtid_reset();
1709 /* Release mutexes and reenable signals */
1710 ust_after_fork_common(restore_sigset
);
1714 void lttng_ust_sockinfo_session_enabled(void *owner
)
1716 struct sock_info
*sock_info
= owner
;
1717 sock_info
->statedump_pending
= 1;