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>
45 #include <usterr-signal-safe.h>
46 #include "tracepoint-internal.h"
47 #include "ltt-tracer-core.h"
49 #include "../libringbuffer/tlsfixup.h"
52 * Has lttng ust comm constructor been called ?
54 static int initialized
;
57 * The ust_lock/ust_unlock lock is used as a communication thread mutex.
58 * Held when handling a command, also held by fork() to deal with
59 * removal of threads, and by exit path.
62 /* Should the ust comm thread quit ? */
63 static int lttng_ust_comm_should_quit
;
66 * Wait for either of these before continuing to the main
68 * - the register_done message from sessiond daemon
69 * (will let the sessiond daemon enable sessions before main
71 * - sessiond daemon is not reachable.
72 * - timeout (ensuring applications are resilient to session
75 static sem_t constructor_wait
;
77 * Doing this for both the global and local sessiond.
79 static int sem_count
= { 2 };
82 * Info about socket and associated listener thread.
86 pthread_t ust_listener
; /* listener thread */
88 int constructor_sem_posted
;
92 char sock_path
[PATH_MAX
];
95 char wait_shm_path
[PATH_MAX
];
99 /* Socket from app (connect) to session daemon (listen) for communication */
100 struct sock_info global_apps
= {
107 .sock_path
= DEFAULT_GLOBAL_APPS_UNIX_SOCK
,
110 .wait_shm_path
= DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH
,
113 /* TODO: allow global_apps_sock_path override */
115 struct sock_info local_apps
= {
119 .allowed
= 0, /* Check setuid bit first */
124 static int wait_poll_fallback
;
126 extern void ltt_ring_buffer_client_overwrite_init(void);
127 extern void ltt_ring_buffer_client_discard_init(void);
128 extern void ltt_ring_buffer_metadata_client_init(void);
129 extern void ltt_ring_buffer_client_overwrite_exit(void);
130 extern void ltt_ring_buffer_client_discard_exit(void);
131 extern void ltt_ring_buffer_metadata_client_exit(void);
134 int setup_local_apps(void)
136 const char *home_dir
;
141 * Disallow per-user tracing for setuid binaries.
143 if (uid
!= geteuid()) {
144 local_apps
.allowed
= 0;
147 local_apps
.allowed
= 1;
149 home_dir
= (const char *) getenv("HOME");
152 snprintf(local_apps
.sock_path
, PATH_MAX
,
153 DEFAULT_HOME_APPS_UNIX_SOCK
, home_dir
);
154 snprintf(local_apps
.wait_shm_path
, PATH_MAX
,
155 DEFAULT_HOME_APPS_WAIT_SHM_PATH
, uid
);
160 int register_app_to_sessiond(int socket
)
170 uint32_t bits_per_long
;
171 char name
[16]; /* process name */
174 reg_msg
.major
= LTTNG_UST_COMM_VERSION_MAJOR
;
175 reg_msg
.minor
= LTTNG_UST_COMM_VERSION_MINOR
;
176 reg_msg
.pid
= getpid();
177 reg_msg
.ppid
= getppid();
178 reg_msg
.uid
= getuid();
179 reg_msg
.gid
= getgid();
180 reg_msg
.bits_per_long
= CAA_BITS_PER_LONG
;
181 lttng_ust_getprocname(reg_msg
.name
);
183 ret
= ustcomm_send_unix_sock(socket
, ®_msg
, sizeof(reg_msg
));
184 if (ret
>= 0 && ret
!= sizeof(reg_msg
))
190 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
194 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
197 DBG("message successfully sent");
200 if (errno
== ECONNRESET
) {
201 printf("remote end closed connection\n");
206 printf("incorrect message size: %zd\n", len
);
212 int handle_register_done(struct sock_info
*sock_info
)
216 if (sock_info
->constructor_sem_posted
)
218 sock_info
->constructor_sem_posted
= 1;
219 if (uatomic_read(&sem_count
) <= 0) {
222 ret
= uatomic_add_return(&sem_count
, -1);
224 ret
= sem_post(&constructor_wait
);
231 int handle_message(struct sock_info
*sock_info
,
232 int sock
, struct ustcomm_ust_msg
*lum
)
235 const struct lttng_ust_objd_ops
*ops
;
236 struct ustcomm_ust_reply lur
;
242 memset(&lur
, 0, sizeof(lur
));
244 if (lttng_ust_comm_should_quit
) {
249 ops
= objd_ops(lum
->handle
);
256 case LTTNG_UST_REGISTER_DONE
:
257 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
258 ret
= handle_register_done(sock_info
);
262 case LTTNG_UST_RELEASE
:
263 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
266 ret
= lttng_ust_objd_unref(lum
->handle
);
270 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
271 (unsigned long) &lum
->u
,
279 lur
.handle
= lum
->handle
;
283 lur
.ret_code
= USTCOMM_OK
;
285 //lur.ret_code = USTCOMM_SESSION_FAIL;
290 case LTTNG_UST_STREAM
:
292 * Special-case reply to send stream info.
295 lur
.u
.stream
.memory_map_size
= *args
.stream
.memory_map_size
;
296 shm_fd
= *args
.stream
.shm_fd
;
297 wait_fd
= *args
.stream
.wait_fd
;
299 case LTTNG_UST_METADATA
:
300 case LTTNG_UST_CHANNEL
:
301 lur
.u
.channel
.memory_map_size
= *args
.channel
.memory_map_size
;
302 shm_fd
= *args
.channel
.shm_fd
;
303 wait_fd
= *args
.channel
.wait_fd
;
305 case LTTNG_UST_TRACER_VERSION
:
306 lur
.u
.version
= lum
->u
.version
;
308 case LTTNG_UST_TRACEPOINT_LIST_GET
:
309 memcpy(&lur
.u
.tracepoint
, &lum
->u
.tracepoint
, sizeof(lur
.u
.tracepoint
));
313 ret
= send_reply(sock
, &lur
);
315 perror("error sending reply");
319 if ((lum
->cmd
== LTTNG_UST_STREAM
320 || lum
->cmd
== LTTNG_UST_CHANNEL
321 || lum
->cmd
== LTTNG_UST_METADATA
)
322 && lur
.ret_code
== USTCOMM_OK
) {
323 /* we also need to send the file descriptors. */
324 ret
= ustcomm_send_fds_unix_sock(sock
,
328 perror("send shm_fd");
331 ret
= ustcomm_send_fds_unix_sock(sock
,
335 perror("send wait_fd");
340 * We still have the memory map reference, and the fds have been
341 * sent to the sessiond. We can therefore close those fds. Note
342 * that we keep the write side of the wait_fd open, but close
345 if (lur
.ret_code
== USTCOMM_OK
) {
347 case LTTNG_UST_STREAM
:
351 PERROR("Error closing stream shm_fd");
353 *args
.stream
.shm_fd
= -1;
356 ret
= close(wait_fd
);
358 PERROR("Error closing stream wait_fd");
360 *args
.stream
.wait_fd
= -1;
363 case LTTNG_UST_METADATA
:
364 case LTTNG_UST_CHANNEL
:
368 PERROR("Error closing channel shm_fd");
370 *args
.channel
.shm_fd
= -1;
373 ret
= close(wait_fd
);
375 PERROR("Error closing channel wait_fd");
377 *args
.channel
.wait_fd
= -1;
389 void cleanup_sock_info(struct sock_info
*sock_info
, int exiting
)
393 if (sock_info
->socket
!= -1) {
394 ret
= close(sock_info
->socket
);
396 ERR("Error closing apps socket");
398 sock_info
->socket
= -1;
400 if (sock_info
->root_handle
!= -1) {
401 ret
= lttng_ust_objd_unref(sock_info
->root_handle
);
403 ERR("Error unref root handle");
405 sock_info
->root_handle
= -1;
407 sock_info
->constructor_sem_posted
= 0;
409 * wait_shm_mmap is used by listener threads outside of the
410 * ust lock, so we cannot tear it down ourselves, because we
411 * cannot join on these threads. Leave this task to the OS
414 if (!exiting
&& sock_info
->wait_shm_mmap
) {
415 ret
= munmap(sock_info
->wait_shm_mmap
, sysconf(_SC_PAGE_SIZE
));
417 ERR("Error unmapping wait shm");
419 sock_info
->wait_shm_mmap
= NULL
;
424 * Using fork to set umask in the child process (not multi-thread safe).
425 * We deal with the shm_open vs ftruncate race (happening when the
426 * sessiond owns the shm and does not let everybody modify it, to ensure
427 * safety against shm_unlink) by simply letting the mmap fail and
428 * retrying after a few seconds.
429 * For global shm, everybody has rw access to it until the sessiond
433 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
435 int wait_shm_fd
, ret
;
439 * Try to open read-only.
441 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
442 if (wait_shm_fd
>= 0) {
444 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
446 * Real-only open did not work, and it's not because the
447 * entry was not present. It's a failure that prohibits
450 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
454 * If the open failed because the file did not exist, try
455 * creating it ourself.
462 * Parent: wait for child to return, in which case the
463 * shared memory map will have been created.
466 if (pid
< 0 || !WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
471 * Try to open read-only again after creation.
473 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
474 if (wait_shm_fd
< 0) {
476 * Real-only open did not work. It's a failure
477 * that prohibits using shm.
479 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
483 } else if (pid
== 0) {
487 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
488 if (sock_info
->global
)
489 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
491 * We're alone in a child process, so we can modify the
492 * process-wide umask.
496 * Try creating shm (or get rw access).
497 * We don't do an exclusive open, because we allow other
498 * processes to create+ftruncate it concurrently.
500 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
501 O_RDWR
| O_CREAT
, create_mode
);
502 if (wait_shm_fd
>= 0) {
503 ret
= ftruncate(wait_shm_fd
, mmap_size
);
511 * For local shm, we need to have rw access to accept
512 * opening it: this means the local sessiond will be
513 * able to wake us up. For global shm, we open it even
514 * if rw access is not granted, because the root.root
515 * sessiond will be able to override all rights and wake
518 if (!sock_info
->global
&& errno
!= EACCES
) {
519 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
523 * The shm exists, but we cannot open it RW. Report
531 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
535 * Ensure that our user is the owner of the shm file for
536 * local shm. If we do not own the file, it means our
537 * sessiond will not have access to wake us up (there is
538 * probably a rogue process trying to fake our
539 * sessiond). Fallback to polling method in this case.
541 ret
= fstat(wait_shm_fd
, &statbuf
);
546 if (statbuf
.st_uid
!= getuid())
552 ret
= close(wait_shm_fd
);
554 PERROR("Error closing fd");
560 char *get_map_shm(struct sock_info
*sock_info
)
562 size_t mmap_size
= sysconf(_SC_PAGE_SIZE
);
563 int wait_shm_fd
, ret
;
566 wait_shm_fd
= get_wait_shm(sock_info
, mmap_size
);
567 if (wait_shm_fd
< 0) {
570 wait_shm_mmap
= mmap(NULL
, mmap_size
, PROT_READ
,
571 MAP_SHARED
, wait_shm_fd
, 0);
572 /* close shm fd immediately after taking the mmap reference */
573 ret
= close(wait_shm_fd
);
575 PERROR("Error closing fd");
577 if (wait_shm_mmap
== MAP_FAILED
) {
578 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
581 return wait_shm_mmap
;
588 void wait_for_sessiond(struct sock_info
*sock_info
)
593 if (lttng_ust_comm_should_quit
) {
596 if (wait_poll_fallback
) {
599 if (!sock_info
->wait_shm_mmap
) {
600 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
601 if (!sock_info
->wait_shm_mmap
)
606 DBG("Waiting for %s apps sessiond", sock_info
->name
);
607 /* Wait for futex wakeup */
608 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
609 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
610 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
612 if (errno
== EFAULT
) {
613 wait_poll_fallback
= 1;
615 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
616 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
617 "Please upgrade your kernel "
618 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
619 "mainline). LTTng-UST will use polling mode fallback.");
637 * This thread does not allocate any resource, except within
638 * handle_message, within mutex protection. This mutex protects against
640 * The other moment it allocates resources is at socket connexion, which
641 * is also protected by the mutex.
644 void *ust_listener_thread(void *arg
)
646 struct sock_info
*sock_info
= arg
;
647 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
649 /* Restart trying to connect to the session daemon */
651 if (prev_connect_failed
) {
652 /* Wait for sessiond availability with pipe */
653 wait_for_sessiond(sock_info
);
657 * Sleep for 5 seconds before retrying after a
658 * sequence of failure / wait / failure. This
659 * deals with a killed or broken session daemon.
664 prev_connect_failed
= 0;
668 if (lttng_ust_comm_should_quit
) {
673 if (sock_info
->socket
!= -1) {
674 ret
= close(sock_info
->socket
);
676 ERR("Error closing %s apps socket", sock_info
->name
);
678 sock_info
->socket
= -1;
682 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
684 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
685 prev_connect_failed
= 1;
687 * If we cannot find the sessiond daemon, don't delay
688 * constructor execution.
690 ret
= handle_register_done(sock_info
);
696 sock_info
->socket
= sock
= ret
;
699 * Create only one root handle per listener thread for the whole
702 if (sock_info
->root_handle
== -1) {
703 ret
= lttng_abi_create_root_handle();
705 ERR("Error creating root handle");
709 sock_info
->root_handle
= ret
;
712 ret
= register_app_to_sessiond(sock
);
714 ERR("Error registering to %s apps socket", sock_info
->name
);
715 prev_connect_failed
= 1;
717 * If we cannot register to the sessiond daemon, don't
718 * delay constructor execution.
720 ret
= handle_register_done(sock_info
);
729 struct ustcomm_ust_msg lum
;
731 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
733 case 0: /* orderly shutdown */
734 DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info
->name
);
737 * Either sessiond has shutdown or refused us by closing the socket.
738 * In either case, we don't want to delay construction execution,
739 * and we need to wait before retry.
741 prev_connect_failed
= 1;
743 * If we cannot register to the sessiond daemon, don't
744 * delay constructor execution.
746 ret
= handle_register_done(sock_info
);
751 DBG("message received\n");
752 ret
= handle_message(sock_info
, sock
, &lum
);
754 ERR("Error handling message for %s socket", sock_info
->name
);
758 DBG("Receive failed from lttng-sessiond with errno %d", errno
);
759 if (errno
== ECONNRESET
) {
760 ERR("%s remote end closed connection\n", sock_info
->name
);
765 ERR("incorrect message size (%s socket): %zd\n", sock_info
->name
, len
);
771 goto restart
; /* try to reconnect */
777 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
780 int get_timeout(struct timespec
*constructor_timeout
)
782 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
786 str_delay
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
788 constructor_delay_ms
= strtol(str_delay
, NULL
, 10);
791 switch (constructor_delay_ms
) {
792 case -1:/* fall-through */
794 return constructor_delay_ms
;
800 * If we are unable to find the current time, don't wait.
802 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
806 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
807 constructor_timeout
->tv_nsec
+=
808 (constructor_delay_ms
% 1000UL) * 1000000UL;
809 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
810 constructor_timeout
->tv_sec
++;
811 constructor_timeout
->tv_nsec
-= 1000000000UL;
817 * sessiond monitoring thread: monitor presence of global and per-user
818 * sessiond by polling the application common named pipe.
822 void __attribute__((constructor
)) lttng_ust_init(void)
824 struct timespec constructor_timeout
;
828 if (uatomic_xchg(&initialized
, 1) == 1)
832 * Fixup interdependency between TLS fixup mutex (which happens
833 * to be the dynamic linker mutex) and ust_lock, taken within
836 lttng_fixup_event_tls();
837 lttng_fixup_ringbuffer_tls();
838 lttng_fixup_vtid_tls();
841 * We want precise control over the order in which we construct
842 * our sub-libraries vs starting to receive commands from
843 * sessiond (otherwise leading to errors when trying to create
844 * sessiond before the init functions are completed).
848 ltt_ring_buffer_metadata_client_init();
849 ltt_ring_buffer_client_overwrite_init();
850 ltt_ring_buffer_client_discard_init();
852 timeout_mode
= get_timeout(&constructor_timeout
);
854 ret
= sem_init(&constructor_wait
, 0, 0);
857 ret
= setup_local_apps();
859 ERR("Error setting up to local apps");
861 ret
= pthread_create(&local_apps
.ust_listener
, NULL
,
862 ust_listener_thread
, &local_apps
);
864 if (local_apps
.allowed
) {
865 ret
= pthread_create(&global_apps
.ust_listener
, NULL
,
866 ust_listener_thread
, &global_apps
);
868 handle_register_done(&local_apps
);
871 switch (timeout_mode
) {
872 case 1: /* timeout wait */
874 ret
= sem_timedwait(&constructor_wait
,
875 &constructor_timeout
);
876 } while (ret
< 0 && errno
== EINTR
);
877 if (ret
< 0 && errno
== ETIMEDOUT
) {
878 ERR("Timed out waiting for ltt-sessiond");
883 case -1:/* wait forever */
885 ret
= sem_wait(&constructor_wait
);
886 } while (ret
< 0 && errno
== EINTR
);
889 case 0: /* no timeout */
895 void lttng_ust_cleanup(int exiting
)
897 cleanup_sock_info(&global_apps
, exiting
);
898 if (local_apps
.allowed
) {
899 cleanup_sock_info(&local_apps
, exiting
);
902 * The teardown in this function all affect data structures
903 * accessed under the UST lock by the listener thread. This
904 * lock, along with the lttng_ust_comm_should_quit flag, ensure
905 * that none of these threads are accessing this data at this
908 lttng_ust_abi_exit();
909 lttng_ust_events_exit();
910 ltt_ring_buffer_client_discard_exit();
911 ltt_ring_buffer_client_overwrite_exit();
912 ltt_ring_buffer_metadata_client_exit();
915 /* Reinitialize values for fork */
917 lttng_ust_comm_should_quit
= 0;
922 void __attribute__((destructor
)) lttng_ust_exit(void)
927 * Using pthread_cancel here because:
928 * A) we don't want to hang application teardown.
929 * B) the thread is not allocating any resource.
933 * Require the communication thread to quit. Synchronize with
934 * mutexes to ensure it is not in a mutex critical section when
935 * pthread_cancel is later called.
938 lttng_ust_comm_should_quit
= 1;
942 ret
= pthread_cancel(global_apps
.ust_listener
);
944 ERR("Error cancelling global ust listener thread");
946 if (local_apps
.allowed
) {
947 ret
= pthread_cancel(local_apps
.ust_listener
);
949 ERR("Error cancelling local ust listener thread");
953 * Do NOT join threads: use of sys_futex makes it impossible to
954 * join the threads without using async-cancel, but async-cancel
955 * is delivered by a signal, which could hit the target thread
956 * anywhere in its code path, including while the ust_lock() is
957 * held, causing a deadlock for the other thread. Let the OS
958 * cleanup the threads if there are stalled in a syscall.
960 lttng_ust_cleanup(1);
964 * We exclude the worker threads across fork and clone (except
965 * CLONE_VM), because these system calls only keep the forking thread
966 * running in the child. Therefore, we don't want to call fork or clone
967 * in the middle of an tracepoint or ust tracing state modification.
968 * Holding this mutex protects these structures across fork and clone.
970 void ust_before_fork(sigset_t
*save_sigset
)
973 * Disable signals. This is to avoid that the child intervenes
974 * before it is properly setup for tracing. It is safer to
975 * disable all signals, because then we know we are not breaking
976 * anything by restoring the original mask.
981 /* Disable signals */
982 sigfillset(&all_sigs
);
983 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
985 PERROR("sigprocmask");
988 rcu_bp_before_fork();
991 static void ust_after_fork_common(sigset_t
*restore_sigset
)
995 DBG("process %d", getpid());
997 /* Restore signals */
998 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
1000 PERROR("sigprocmask");
1004 void ust_after_fork_parent(sigset_t
*restore_sigset
)
1006 DBG("process %d", getpid());
1007 rcu_bp_after_fork_parent();
1008 /* Release mutexes and reenable signals */
1009 ust_after_fork_common(restore_sigset
);
1013 * After fork, in the child, we need to cleanup all the leftover state,
1014 * except the worker thread which already magically disappeared thanks
1015 * to the weird Linux fork semantics. After tyding up, we call
1016 * lttng_ust_init() again to start over as a new PID.
1018 * This is meant for forks() that have tracing in the child between the
1019 * fork and following exec call (if there is any).
1021 void ust_after_fork_child(sigset_t
*restore_sigset
)
1023 DBG("process %d", getpid());
1024 /* Release urcu mutexes */
1025 rcu_bp_after_fork_child();
1026 lttng_ust_cleanup(0);
1027 lttng_context_vtid_reset();
1028 /* Release mutexes and reenable signals */
1029 ust_after_fork_common(restore_sigset
);