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>
25 #include <sys/prctl.h>
28 #include <sys/types.h>
34 #include <semaphore.h>
38 #include <urcu/uatomic.h>
39 #include <urcu/futex.h>
41 #include <lttng-ust-comm.h>
42 #include <ust/lttng-events.h>
43 #include <ust/usterr-signal-safe.h>
44 #include <ust/lttng-ust-abi.h>
45 #include <ust/tracepoint.h>
46 #include <ust/tracepoint-internal.h>
48 #include "ltt-tracer-core.h"
51 * Has lttng ust comm constructor been called ?
53 static int initialized
;
56 * The ust_lock/ust_unlock lock is used as a communication thread mutex.
57 * Held when handling a command, also held by fork() to deal with
58 * removal of threads, and by exit path.
61 /* Should the ust comm thread quit ? */
62 static int lttng_ust_comm_should_quit
;
65 * Wait for either of these before continuing to the main
67 * - the register_done message from sessiond daemon
68 * (will let the sessiond daemon enable sessions before main
70 * - sessiond daemon is not reachable.
71 * - timeout (ensuring applications are resilient to session
74 static sem_t constructor_wait
;
76 * Doing this for both the global and local sessiond.
78 static int sem_count
= { 2 };
81 * Info about socket and associated listener thread.
85 pthread_t ust_listener
; /* listener thread */
87 int constructor_sem_posted
;
91 char sock_path
[PATH_MAX
];
94 char wait_shm_path
[PATH_MAX
];
98 /* Socket from app (connect) to session daemon (listen) for communication */
99 struct sock_info global_apps
= {
106 .sock_path
= DEFAULT_GLOBAL_APPS_UNIX_SOCK
,
109 .wait_shm_path
= DEFAULT_GLOBAL_APPS_WAIT_SHM_PATH
,
112 /* TODO: allow global_apps_sock_path override */
114 struct sock_info local_apps
= {
118 .allowed
= 0, /* Check setuid bit first */
123 static int wait_poll_fallback
;
125 extern void ltt_ring_buffer_client_overwrite_init(void);
126 extern void ltt_ring_buffer_client_discard_init(void);
127 extern void ltt_ring_buffer_metadata_client_init(void);
128 extern void ltt_ring_buffer_client_overwrite_exit(void);
129 extern void ltt_ring_buffer_client_discard_exit(void);
130 extern void ltt_ring_buffer_metadata_client_exit(void);
133 int setup_local_apps(void)
135 const char *home_dir
;
140 * Disallow per-user tracing for setuid binaries.
142 if (uid
!= geteuid()) {
143 local_apps
.allowed
= 0;
146 local_apps
.allowed
= 1;
148 home_dir
= (const char *) getenv("HOME");
151 snprintf(local_apps
.sock_path
, PATH_MAX
,
152 DEFAULT_HOME_APPS_UNIX_SOCK
, home_dir
);
153 snprintf(local_apps
.wait_shm_path
, PATH_MAX
,
154 DEFAULT_HOME_APPS_WAIT_SHM_PATH
, uid
);
159 int register_app_to_sessiond(int socket
)
170 char name
[16]; /* process name */
173 reg_msg
.major
= LTTNG_UST_COMM_VERSION_MAJOR
;
174 reg_msg
.minor
= LTTNG_UST_COMM_VERSION_MINOR
;
175 reg_msg
.pid
= getpid();
176 reg_msg
.ppid
= getppid();
177 reg_msg
.uid
= getuid();
178 reg_msg
.gid
= getgid();
179 prctl_ret
= prctl(PR_GET_NAME
, (unsigned long) reg_msg
.name
, 0, 0, 0);
181 ERR("Error executing prctl");
185 ret
= ustcomm_send_unix_sock(socket
, ®_msg
, sizeof(reg_msg
));
186 if (ret
>= 0 && ret
!= sizeof(reg_msg
))
192 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
196 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
199 DBG("message successfully sent");
202 if (errno
== ECONNRESET
) {
203 printf("remote end closed connection\n");
208 printf("incorrect message size: %zd\n", len
);
214 int handle_register_done(struct sock_info
*sock_info
)
218 if (sock_info
->constructor_sem_posted
)
220 sock_info
->constructor_sem_posted
= 1;
221 if (uatomic_read(&sem_count
) <= 0) {
224 ret
= uatomic_add_return(&sem_count
, -1);
226 ret
= sem_post(&constructor_wait
);
233 int handle_message(struct sock_info
*sock_info
,
234 int sock
, struct ustcomm_ust_msg
*lum
)
237 const struct objd_ops
*ops
;
238 struct ustcomm_ust_reply lur
;
243 memset(&lur
, 0, sizeof(lur
));
245 if (lttng_ust_comm_should_quit
) {
250 ops
= objd_ops(lum
->handle
);
257 case LTTNG_UST_REGISTER_DONE
:
258 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
259 ret
= handle_register_done(sock_info
);
263 case LTTNG_UST_RELEASE
:
264 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
267 ret
= objd_unref(lum
->handle
);
271 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
272 (unsigned long) &lum
->u
);
279 lur
.handle
= lum
->handle
;
283 lur
.ret_code
= USTCOMM_OK
;
285 //lur.ret_code = USTCOMM_SESSION_FAIL;
289 case LTTNG_UST_STREAM
:
291 * Special-case reply to send stream info.
294 lur
.u
.stream
.memory_map_size
= lum
->u
.stream
.memory_map_size
;
295 shm_fd
= lum
->u
.stream
.shm_fd
;
296 wait_fd
= lum
->u
.stream
.wait_fd
;
298 case LTTNG_UST_METADATA
:
299 case LTTNG_UST_CHANNEL
:
300 lur
.u
.channel
.memory_map_size
= lum
->u
.channel
.memory_map_size
;
301 shm_fd
= lum
->u
.channel
.shm_fd
;
302 wait_fd
= lum
->u
.channel
.wait_fd
;
304 case LTTNG_UST_VERSION
:
305 lur
.u
.version
= lum
->u
.version
;
308 ret
= send_reply(sock
, &lur
);
310 perror("error sending reply");
314 if ((lum
->cmd
== LTTNG_UST_STREAM
315 || lum
->cmd
== LTTNG_UST_CHANNEL
316 || lum
->cmd
== LTTNG_UST_METADATA
)
317 && lur
.ret_code
== USTCOMM_OK
) {
318 /* we also need to send the file descriptors. */
319 ret
= ustcomm_send_fds_unix_sock(sock
,
323 perror("send shm_fd");
326 ret
= ustcomm_send_fds_unix_sock(sock
,
330 perror("send wait_fd");
340 void cleanup_sock_info(struct sock_info
*sock_info
)
344 if (sock_info
->socket
!= -1) {
345 ret
= close(sock_info
->socket
);
347 ERR("Error closing apps socket");
349 sock_info
->socket
= -1;
351 if (sock_info
->root_handle
!= -1) {
352 ret
= objd_unref(sock_info
->root_handle
);
354 ERR("Error unref root handle");
356 sock_info
->root_handle
= -1;
358 sock_info
->constructor_sem_posted
= 0;
359 if (sock_info
->wait_shm_mmap
) {
360 ret
= munmap(sock_info
->wait_shm_mmap
, sysconf(_SC_PAGE_SIZE
));
362 ERR("Error unmapping wait shm");
364 sock_info
->wait_shm_mmap
= NULL
;
369 * Using fork to set umask in the child process (not multi-thread safe).
370 * We deal with the shm_open vs ftruncate race (happening when the
371 * sessiond owns the shm and does not let everybody modify it, to ensure
372 * safety against shm_unlink) by simply letting the mmap fail and
373 * retrying after a few seconds.
374 * For global shm, everybody has rw access to it until the sessiond
378 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
380 int wait_shm_fd
, ret
;
384 * Try to open read-only.
386 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
387 if (wait_shm_fd
>= 0) {
389 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
391 * Real-only open did not work, and it's not because the
392 * entry was not present. It's a failure that prohibits
395 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
399 * If the open failed because the file did not exist, try
400 * creating it ourself.
407 * Parent: wait for child to return, in which case the
408 * shared memory map will have been created.
411 if (!WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
416 * Try to open read-only again after creation.
418 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
419 if (wait_shm_fd
< 0) {
421 * Real-only open did not work. It's a failure
422 * that prohibits using shm.
424 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
428 } else if (pid
== 0) {
432 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
433 if (sock_info
->global
)
434 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
436 * We're alone in a child process, so we can modify the
437 * process-wide umask.
441 * Try creating shm (or get rw access).
442 * We don't do an exclusive open, because we allow other
443 * processes to create+ftruncate it concurrently.
445 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
446 O_RDWR
| O_CREAT
, create_mode
);
447 if (wait_shm_fd
>= 0) {
448 ret
= ftruncate(wait_shm_fd
, mmap_size
);
456 * For local shm, we need to have rw access to accept
457 * opening it: this means the local sessiond will be
458 * able to wake us up. For global shm, we open it even
459 * if rw access is not granted, because the root.root
460 * sessiond will be able to override all rights and wake
463 if (!sock_info
->global
&& errno
!= EACCES
) {
464 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
468 * The shm exists, but we cannot open it RW. Report
476 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
480 * Ensure that our user is the owner of the shm file for
481 * local shm. If we do not own the file, it means our
482 * sessiond will not have access to wake us up (there is
483 * probably a rogue process trying to fake our
484 * sessiond). Fallback to polling method in this case.
486 ret
= fstat(wait_shm_fd
, &statbuf
);
491 if (statbuf
.st_uid
!= getuid())
497 ret
= close(wait_shm_fd
);
499 PERROR("Error closing fd");
505 char *get_map_shm(struct sock_info
*sock_info
)
507 size_t mmap_size
= sysconf(_SC_PAGE_SIZE
);
508 int wait_shm_fd
, ret
;
511 wait_shm_fd
= get_wait_shm(sock_info
, mmap_size
);
512 if (wait_shm_fd
< 0) {
515 wait_shm_mmap
= mmap(NULL
, mmap_size
, PROT_READ
,
516 MAP_SHARED
, wait_shm_fd
, 0);
517 /* close shm fd immediately after taking the mmap reference */
518 ret
= close(wait_shm_fd
);
520 PERROR("Error closing fd");
522 if (wait_shm_mmap
== MAP_FAILED
) {
523 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
526 return wait_shm_mmap
;
533 void wait_for_sessiond(struct sock_info
*sock_info
)
538 if (lttng_ust_comm_should_quit
) {
541 if (wait_poll_fallback
) {
544 if (!sock_info
->wait_shm_mmap
) {
545 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
546 if (!sock_info
->wait_shm_mmap
)
551 DBG("Waiting for %s apps sessiond", sock_info
->name
);
552 /* Wait for futex wakeup */
553 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
554 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
555 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
557 if (errno
== EFAULT
) {
558 wait_poll_fallback
= 1;
560 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
561 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
562 "Please upgrade your kernel "
563 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
564 "mainline). LTTng-UST will use polling mode fallback.");
581 * This thread does not allocate any resource, except within
582 * handle_message, within mutex protection. This mutex protects against
584 * The other moment it allocates resources is at socket connexion, which
585 * is also protected by the mutex.
588 void *ust_listener_thread(void *arg
)
590 struct sock_info
*sock_info
= arg
;
591 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
593 /* Restart trying to connect to the session daemon */
595 if (prev_connect_failed
) {
596 /* Wait for sessiond availability with pipe */
597 wait_for_sessiond(sock_info
);
601 * Sleep for 5 seconds before retrying after a
602 * sequence of failure / wait / failure. This
603 * deals with a killed or broken session daemon.
608 prev_connect_failed
= 0;
612 if (lttng_ust_comm_should_quit
) {
617 if (sock_info
->socket
!= -1) {
618 ret
= close(sock_info
->socket
);
620 ERR("Error closing %s apps socket", sock_info
->name
);
622 sock_info
->socket
= -1;
626 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
628 ERR("Error connecting to %s apps socket", sock_info
->name
);
629 prev_connect_failed
= 1;
631 * If we cannot find the sessiond daemon, don't delay
632 * constructor execution.
634 ret
= handle_register_done(sock_info
);
640 sock_info
->socket
= sock
= ret
;
643 * Create only one root handle per listener thread for the whole
646 if (sock_info
->root_handle
== -1) {
647 ret
= lttng_abi_create_root_handle();
649 ERR("Error creating root handle");
653 sock_info
->root_handle
= ret
;
656 ret
= register_app_to_sessiond(sock
);
658 ERR("Error registering to %s apps socket", sock_info
->name
);
659 prev_connect_failed
= 1;
661 * If we cannot register to the sessiond daemon, don't
662 * delay constructor execution.
664 ret
= handle_register_done(sock_info
);
673 struct ustcomm_ust_msg lum
;
675 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
677 case 0: /* orderly shutdown */
678 DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info
->name
);
681 DBG("message received\n");
682 ret
= handle_message(sock_info
, sock
, &lum
);
684 ERR("Error handling message for %s socket", sock_info
->name
);
688 if (errno
== ECONNRESET
) {
689 ERR("%s remote end closed connection\n", sock_info
->name
);
694 ERR("incorrect message size (%s socket): %zd\n", sock_info
->name
, len
);
700 goto restart
; /* try to reconnect */
706 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
709 int get_timeout(struct timespec
*constructor_timeout
)
711 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
715 str_delay
= getenv("UST_REGISTER_TIMEOUT");
717 constructor_delay_ms
= strtol(str_delay
, NULL
, 10);
720 switch (constructor_delay_ms
) {
721 case -1:/* fall-through */
723 return constructor_delay_ms
;
729 * If we are unable to find the current time, don't wait.
731 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
735 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
736 constructor_timeout
->tv_nsec
+=
737 (constructor_delay_ms
% 1000UL) * 1000000UL;
738 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
739 constructor_timeout
->tv_sec
++;
740 constructor_timeout
->tv_nsec
-= 1000000000UL;
746 * sessiond monitoring thread: monitor presence of global and per-user
747 * sessiond by polling the application common named pipe.
751 void __attribute__((constructor
)) lttng_ust_init(void)
753 struct timespec constructor_timeout
;
757 if (uatomic_xchg(&initialized
, 1) == 1)
761 * We want precise control over the order in which we construct
762 * our sub-libraries vs starting to receive commands from
763 * sessiond (otherwise leading to errors when trying to create
764 * sessiond before the init functions are completed).
768 ltt_ring_buffer_metadata_client_init();
769 ltt_ring_buffer_client_overwrite_init();
770 ltt_ring_buffer_client_discard_init();
772 timeout_mode
= get_timeout(&constructor_timeout
);
774 ret
= sem_init(&constructor_wait
, 0, 0);
777 ret
= setup_local_apps();
779 ERR("Error setting up to local apps");
781 ret
= pthread_create(&local_apps
.ust_listener
, NULL
,
782 ust_listener_thread
, &local_apps
);
784 if (local_apps
.allowed
) {
785 ret
= pthread_create(&global_apps
.ust_listener
, NULL
,
786 ust_listener_thread
, &global_apps
);
788 handle_register_done(&local_apps
);
791 switch (timeout_mode
) {
792 case 1: /* timeout wait */
794 ret
= sem_timedwait(&constructor_wait
,
795 &constructor_timeout
);
796 } while (ret
< 0 && errno
== EINTR
);
797 if (ret
< 0 && errno
== ETIMEDOUT
) {
798 ERR("Timed out waiting for ltt-sessiond");
803 case -1:/* wait forever */
805 ret
= sem_wait(&constructor_wait
);
806 } while (ret
< 0 && errno
== EINTR
);
809 case 0: /* no timeout */
815 void lttng_ust_cleanup(int exiting
)
817 cleanup_sock_info(&global_apps
);
818 if (local_apps
.allowed
) {
819 cleanup_sock_info(&local_apps
);
821 lttng_ust_abi_exit();
823 ltt_ring_buffer_client_discard_exit();
824 ltt_ring_buffer_client_overwrite_exit();
825 ltt_ring_buffer_metadata_client_exit();
828 /* Reinitialize values for fork */
830 lttng_ust_comm_should_quit
= 0;
835 void __attribute__((destructor
)) lttng_ust_exit(void)
840 * Using pthread_cancel here because:
841 * A) we don't want to hang application teardown.
842 * B) the thread is not allocating any resource.
846 * Require the communication thread to quit. Synchronize with
847 * mutexes to ensure it is not in a mutex critical section when
848 * pthread_cancel is later called.
851 lttng_ust_comm_should_quit
= 1;
854 ret
= pthread_cancel(global_apps
.ust_listener
);
856 ERR("Error cancelling global ust listener thread");
858 if (local_apps
.allowed
) {
859 ret
= pthread_cancel(local_apps
.ust_listener
);
861 ERR("Error cancelling local ust listener thread");
864 lttng_ust_cleanup(1);
868 * We exclude the worker threads across fork and clone (except
869 * CLONE_VM), because these system calls only keep the forking thread
870 * running in the child. Therefore, we don't want to call fork or clone
871 * in the middle of an tracepoint or ust tracing state modification.
872 * Holding this mutex protects these structures across fork and clone.
874 void ust_before_fork(ust_fork_info_t
*fork_info
)
877 * Disable signals. This is to avoid that the child intervenes
878 * before it is properly setup for tracing. It is safer to
879 * disable all signals, because then we know we are not breaking
880 * anything by restoring the original mask.
885 /* Disable signals */
886 sigfillset(&all_sigs
);
887 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, &fork_info
->orig_sigs
);
889 PERROR("sigprocmask");
892 rcu_bp_before_fork();
895 static void ust_after_fork_common(ust_fork_info_t
*fork_info
)
899 DBG("process %d", getpid());
901 /* Restore signals */
902 ret
= sigprocmask(SIG_SETMASK
, &fork_info
->orig_sigs
, NULL
);
904 PERROR("sigprocmask");
908 void ust_after_fork_parent(ust_fork_info_t
*fork_info
)
910 DBG("process %d", getpid());
911 rcu_bp_after_fork_parent();
912 /* Release mutexes and reenable signals */
913 ust_after_fork_common(fork_info
);
917 * After fork, in the child, we need to cleanup all the leftover state,
918 * except the worker thread which already magically disappeared thanks
919 * to the weird Linux fork semantics. After tyding up, we call
920 * lttng_ust_init() again to start over as a new PID.
922 * This is meant for forks() that have tracing in the child between the
923 * fork and following exec call (if there is any).
925 void ust_after_fork_child(ust_fork_info_t
*fork_info
)
927 DBG("process %d", getpid());
928 /* Release urcu mutexes */
929 rcu_bp_after_fork_child();
930 lttng_ust_cleanup(0);
931 lttng_context_vtid_reset();
932 /* Release mutexes and reenable signals */
933 ust_after_fork_common(fork_info
);