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>
40 #include <urcu/compiler.h>
42 #include <lttng/ust-events.h>
43 #include <lttng/ust-abi.h>
44 #include <lttng/ust.h>
46 #include <usterr-signal-safe.h>
47 #include "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 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 prctl_ret
= prctl(PR_GET_NAME
, (unsigned long) reg_msg
.name
, 0, 0, 0);
183 ERR("Error executing prctl");
187 ret
= ustcomm_send_unix_sock(socket
, ®_msg
, sizeof(reg_msg
));
188 if (ret
>= 0 && ret
!= sizeof(reg_msg
))
194 int send_reply(int sock
, struct ustcomm_ust_reply
*lur
)
198 len
= ustcomm_send_unix_sock(sock
, lur
, sizeof(*lur
));
201 DBG("message successfully sent");
204 if (errno
== ECONNRESET
) {
205 printf("remote end closed connection\n");
210 printf("incorrect message size: %zd\n", len
);
216 int handle_register_done(struct sock_info
*sock_info
)
220 if (sock_info
->constructor_sem_posted
)
222 sock_info
->constructor_sem_posted
= 1;
223 if (uatomic_read(&sem_count
) <= 0) {
226 ret
= uatomic_add_return(&sem_count
, -1);
228 ret
= sem_post(&constructor_wait
);
235 int handle_message(struct sock_info
*sock_info
,
236 int sock
, struct ustcomm_ust_msg
*lum
)
239 const struct lttng_ust_objd_ops
*ops
;
240 struct ustcomm_ust_reply lur
;
245 memset(&lur
, 0, sizeof(lur
));
247 if (lttng_ust_comm_should_quit
) {
252 ops
= objd_ops(lum
->handle
);
259 case LTTNG_UST_REGISTER_DONE
:
260 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
261 ret
= handle_register_done(sock_info
);
265 case LTTNG_UST_RELEASE
:
266 if (lum
->handle
== LTTNG_UST_ROOT_HANDLE
)
269 ret
= lttng_ust_objd_unref(lum
->handle
);
273 ret
= ops
->cmd(lum
->handle
, lum
->cmd
,
274 (unsigned long) &lum
->u
);
281 lur
.handle
= lum
->handle
;
285 lur
.ret_code
= USTCOMM_OK
;
287 //lur.ret_code = USTCOMM_SESSION_FAIL;
291 case LTTNG_UST_STREAM
:
293 * Special-case reply to send stream info.
296 lur
.u
.stream
.memory_map_size
= lum
->u
.stream
.memory_map_size
;
297 shm_fd
= lum
->u
.stream
.shm_fd
;
298 wait_fd
= lum
->u
.stream
.wait_fd
;
300 case LTTNG_UST_METADATA
:
301 case LTTNG_UST_CHANNEL
:
302 lur
.u
.channel
.memory_map_size
= lum
->u
.channel
.memory_map_size
;
303 shm_fd
= lum
->u
.channel
.shm_fd
;
304 wait_fd
= lum
->u
.channel
.wait_fd
;
306 case LTTNG_UST_TRACER_VERSION
:
307 lur
.u
.version
= lum
->u
.version
;
309 case LTTNG_UST_TRACEPOINT_LIST_GET
:
310 memcpy(lur
.u
.tracepoint_list_entry
,
311 lum
->u
.tracepoint_list_entry
,
312 LTTNG_UST_SYM_NAME_LEN
);
315 ret
= send_reply(sock
, &lur
);
317 perror("error sending reply");
321 if ((lum
->cmd
== LTTNG_UST_STREAM
322 || lum
->cmd
== LTTNG_UST_CHANNEL
323 || lum
->cmd
== LTTNG_UST_METADATA
)
324 && lur
.ret_code
== USTCOMM_OK
) {
325 /* we also need to send the file descriptors. */
326 ret
= ustcomm_send_fds_unix_sock(sock
,
330 perror("send shm_fd");
333 ret
= ustcomm_send_fds_unix_sock(sock
,
337 perror("send wait_fd");
347 void cleanup_sock_info(struct sock_info
*sock_info
)
351 if (sock_info
->socket
!= -1) {
352 ret
= close(sock_info
->socket
);
354 ERR("Error closing apps socket");
356 sock_info
->socket
= -1;
358 if (sock_info
->root_handle
!= -1) {
359 ret
= lttng_ust_objd_unref(sock_info
->root_handle
);
361 ERR("Error unref root handle");
363 sock_info
->root_handle
= -1;
365 sock_info
->constructor_sem_posted
= 0;
366 if (sock_info
->wait_shm_mmap
) {
367 ret
= munmap(sock_info
->wait_shm_mmap
, sysconf(_SC_PAGE_SIZE
));
369 ERR("Error unmapping wait shm");
371 sock_info
->wait_shm_mmap
= NULL
;
376 * Using fork to set umask in the child process (not multi-thread safe).
377 * We deal with the shm_open vs ftruncate race (happening when the
378 * sessiond owns the shm and does not let everybody modify it, to ensure
379 * safety against shm_unlink) by simply letting the mmap fail and
380 * retrying after a few seconds.
381 * For global shm, everybody has rw access to it until the sessiond
385 int get_wait_shm(struct sock_info
*sock_info
, size_t mmap_size
)
387 int wait_shm_fd
, ret
;
391 * Try to open read-only.
393 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
394 if (wait_shm_fd
>= 0) {
396 } else if (wait_shm_fd
< 0 && errno
!= ENOENT
) {
398 * Real-only open did not work, and it's not because the
399 * entry was not present. It's a failure that prohibits
402 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
406 * If the open failed because the file did not exist, try
407 * creating it ourself.
414 * Parent: wait for child to return, in which case the
415 * shared memory map will have been created.
418 if (!WIFEXITED(status
) || WEXITSTATUS(status
) != 0) {
423 * Try to open read-only again after creation.
425 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
, O_RDONLY
, 0);
426 if (wait_shm_fd
< 0) {
428 * Real-only open did not work. It's a failure
429 * that prohibits using shm.
431 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
435 } else if (pid
== 0) {
439 create_mode
= S_IRUSR
| S_IWUSR
| S_IRGRP
;
440 if (sock_info
->global
)
441 create_mode
|= S_IROTH
| S_IWGRP
| S_IWOTH
;
443 * We're alone in a child process, so we can modify the
444 * process-wide umask.
448 * Try creating shm (or get rw access).
449 * We don't do an exclusive open, because we allow other
450 * processes to create+ftruncate it concurrently.
452 wait_shm_fd
= shm_open(sock_info
->wait_shm_path
,
453 O_RDWR
| O_CREAT
, create_mode
);
454 if (wait_shm_fd
>= 0) {
455 ret
= ftruncate(wait_shm_fd
, mmap_size
);
463 * For local shm, we need to have rw access to accept
464 * opening it: this means the local sessiond will be
465 * able to wake us up. For global shm, we open it even
466 * if rw access is not granted, because the root.root
467 * sessiond will be able to override all rights and wake
470 if (!sock_info
->global
&& errno
!= EACCES
) {
471 ERR("Error opening shm %s", sock_info
->wait_shm_path
);
475 * The shm exists, but we cannot open it RW. Report
483 if (wait_shm_fd
>= 0 && !sock_info
->global
) {
487 * Ensure that our user is the owner of the shm file for
488 * local shm. If we do not own the file, it means our
489 * sessiond will not have access to wake us up (there is
490 * probably a rogue process trying to fake our
491 * sessiond). Fallback to polling method in this case.
493 ret
= fstat(wait_shm_fd
, &statbuf
);
498 if (statbuf
.st_uid
!= getuid())
504 ret
= close(wait_shm_fd
);
506 PERROR("Error closing fd");
512 char *get_map_shm(struct sock_info
*sock_info
)
514 size_t mmap_size
= sysconf(_SC_PAGE_SIZE
);
515 int wait_shm_fd
, ret
;
518 wait_shm_fd
= get_wait_shm(sock_info
, mmap_size
);
519 if (wait_shm_fd
< 0) {
522 wait_shm_mmap
= mmap(NULL
, mmap_size
, PROT_READ
,
523 MAP_SHARED
, wait_shm_fd
, 0);
524 /* close shm fd immediately after taking the mmap reference */
525 ret
= close(wait_shm_fd
);
527 PERROR("Error closing fd");
529 if (wait_shm_mmap
== MAP_FAILED
) {
530 DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
533 return wait_shm_mmap
;
540 void wait_for_sessiond(struct sock_info
*sock_info
)
545 if (lttng_ust_comm_should_quit
) {
548 if (wait_poll_fallback
) {
551 if (!sock_info
->wait_shm_mmap
) {
552 sock_info
->wait_shm_mmap
= get_map_shm(sock_info
);
553 if (!sock_info
->wait_shm_mmap
)
558 DBG("Waiting for %s apps sessiond", sock_info
->name
);
559 /* Wait for futex wakeup */
560 if (uatomic_read((int32_t *) sock_info
->wait_shm_mmap
) == 0) {
561 ret
= futex_async((int32_t *) sock_info
->wait_shm_mmap
,
562 FUTEX_WAIT
, 0, NULL
, NULL
, 0);
564 if (errno
== EFAULT
) {
565 wait_poll_fallback
= 1;
567 "Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
568 "do not support FUTEX_WAKE on read-only memory mappings correctly. "
569 "Please upgrade your kernel "
570 "(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
571 "mainline). LTTng-UST will use polling mode fallback.");
588 * This thread does not allocate any resource, except within
589 * handle_message, within mutex protection. This mutex protects against
591 * The other moment it allocates resources is at socket connexion, which
592 * is also protected by the mutex.
595 void *ust_listener_thread(void *arg
)
597 struct sock_info
*sock_info
= arg
;
598 int sock
, ret
, prev_connect_failed
= 0, has_waited
= 0;
600 /* Restart trying to connect to the session daemon */
602 if (prev_connect_failed
) {
603 /* Wait for sessiond availability with pipe */
604 wait_for_sessiond(sock_info
);
608 * Sleep for 5 seconds before retrying after a
609 * sequence of failure / wait / failure. This
610 * deals with a killed or broken session daemon.
615 prev_connect_failed
= 0;
619 if (lttng_ust_comm_should_quit
) {
624 if (sock_info
->socket
!= -1) {
625 ret
= close(sock_info
->socket
);
627 ERR("Error closing %s apps socket", sock_info
->name
);
629 sock_info
->socket
= -1;
633 ret
= ustcomm_connect_unix_sock(sock_info
->sock_path
);
635 DBG("Info: sessiond not accepting connections to %s apps socket", sock_info
->name
);
636 prev_connect_failed
= 1;
638 * If we cannot find the sessiond daemon, don't delay
639 * constructor execution.
641 ret
= handle_register_done(sock_info
);
647 sock_info
->socket
= sock
= ret
;
650 * Create only one root handle per listener thread for the whole
653 if (sock_info
->root_handle
== -1) {
654 ret
= lttng_abi_create_root_handle();
656 ERR("Error creating root handle");
660 sock_info
->root_handle
= ret
;
663 ret
= register_app_to_sessiond(sock
);
665 ERR("Error registering to %s apps socket", sock_info
->name
);
666 prev_connect_failed
= 1;
668 * If we cannot register to the sessiond daemon, don't
669 * delay constructor execution.
671 ret
= handle_register_done(sock_info
);
680 struct ustcomm_ust_msg lum
;
682 len
= ustcomm_recv_unix_sock(sock
, &lum
, sizeof(lum
));
684 case 0: /* orderly shutdown */
685 DBG("%s ltt-sessiond has performed an orderly shutdown\n", sock_info
->name
);
688 * Either sessiond has shutdown or refused us by closing the socket.
689 * In either case, we don't want to delay construction execution,
690 * and we need to wait before retry.
692 prev_connect_failed
= 1;
694 * If we cannot register to the sessiond daemon, don't
695 * delay constructor execution.
697 ret
= handle_register_done(sock_info
);
702 DBG("message received\n");
703 ret
= handle_message(sock_info
, sock
, &lum
);
705 ERR("Error handling message for %s socket", sock_info
->name
);
709 DBG("Receive failed from lttng-sessiond with errno %d", errno
);
710 if (errno
== ECONNRESET
) {
711 ERR("%s remote end closed connection\n", sock_info
->name
);
716 ERR("incorrect message size (%s socket): %zd\n", sock_info
->name
, len
);
722 goto restart
; /* try to reconnect */
728 * Return values: -1: don't wait. 0: wait forever. 1: timeout wait.
731 int get_timeout(struct timespec
*constructor_timeout
)
733 long constructor_delay_ms
= LTTNG_UST_DEFAULT_CONSTRUCTOR_TIMEOUT_MS
;
737 str_delay
= getenv("LTTNG_UST_REGISTER_TIMEOUT");
739 constructor_delay_ms
= strtol(str_delay
, NULL
, 10);
742 switch (constructor_delay_ms
) {
743 case -1:/* fall-through */
745 return constructor_delay_ms
;
751 * If we are unable to find the current time, don't wait.
753 ret
= clock_gettime(CLOCK_REALTIME
, constructor_timeout
);
757 constructor_timeout
->tv_sec
+= constructor_delay_ms
/ 1000UL;
758 constructor_timeout
->tv_nsec
+=
759 (constructor_delay_ms
% 1000UL) * 1000000UL;
760 if (constructor_timeout
->tv_nsec
>= 1000000000UL) {
761 constructor_timeout
->tv_sec
++;
762 constructor_timeout
->tv_nsec
-= 1000000000UL;
768 * sessiond monitoring thread: monitor presence of global and per-user
769 * sessiond by polling the application common named pipe.
773 void __attribute__((constructor
)) lttng_ust_init(void)
775 struct timespec constructor_timeout
;
779 if (uatomic_xchg(&initialized
, 1) == 1)
783 * We want precise control over the order in which we construct
784 * our sub-libraries vs starting to receive commands from
785 * sessiond (otherwise leading to errors when trying to create
786 * sessiond before the init functions are completed).
790 ltt_ring_buffer_metadata_client_init();
791 ltt_ring_buffer_client_overwrite_init();
792 ltt_ring_buffer_client_discard_init();
794 timeout_mode
= get_timeout(&constructor_timeout
);
796 ret
= sem_init(&constructor_wait
, 0, 0);
799 ret
= setup_local_apps();
801 ERR("Error setting up to local apps");
803 ret
= pthread_create(&local_apps
.ust_listener
, NULL
,
804 ust_listener_thread
, &local_apps
);
806 if (local_apps
.allowed
) {
807 ret
= pthread_create(&global_apps
.ust_listener
, NULL
,
808 ust_listener_thread
, &global_apps
);
810 handle_register_done(&local_apps
);
813 switch (timeout_mode
) {
814 case 1: /* timeout wait */
816 ret
= sem_timedwait(&constructor_wait
,
817 &constructor_timeout
);
818 } while (ret
< 0 && errno
== EINTR
);
819 if (ret
< 0 && errno
== ETIMEDOUT
) {
820 ERR("Timed out waiting for ltt-sessiond");
825 case -1:/* wait forever */
827 ret
= sem_wait(&constructor_wait
);
828 } while (ret
< 0 && errno
== EINTR
);
831 case 0: /* no timeout */
837 void lttng_ust_cleanup(int exiting
)
839 cleanup_sock_info(&global_apps
);
840 if (local_apps
.allowed
) {
841 cleanup_sock_info(&local_apps
);
843 lttng_ust_abi_exit();
844 lttng_ust_events_exit();
845 ltt_ring_buffer_client_discard_exit();
846 ltt_ring_buffer_client_overwrite_exit();
847 ltt_ring_buffer_metadata_client_exit();
850 /* Reinitialize values for fork */
852 lttng_ust_comm_should_quit
= 0;
857 void __attribute__((destructor
)) lttng_ust_exit(void)
862 * Using pthread_cancel here because:
863 * A) we don't want to hang application teardown.
864 * B) the thread is not allocating any resource.
868 * Require the communication thread to quit. Synchronize with
869 * mutexes to ensure it is not in a mutex critical section when
870 * pthread_cancel is later called.
873 lttng_ust_comm_should_quit
= 1;
876 ret
= pthread_cancel(global_apps
.ust_listener
);
878 ERR("Error cancelling global ust listener thread");
880 if (local_apps
.allowed
) {
881 ret
= pthread_cancel(local_apps
.ust_listener
);
883 ERR("Error cancelling local ust listener thread");
886 lttng_ust_cleanup(1);
890 * We exclude the worker threads across fork and clone (except
891 * CLONE_VM), because these system calls only keep the forking thread
892 * running in the child. Therefore, we don't want to call fork or clone
893 * in the middle of an tracepoint or ust tracing state modification.
894 * Holding this mutex protects these structures across fork and clone.
896 void ust_before_fork(sigset_t
*save_sigset
)
899 * Disable signals. This is to avoid that the child intervenes
900 * before it is properly setup for tracing. It is safer to
901 * disable all signals, because then we know we are not breaking
902 * anything by restoring the original mask.
907 /* Disable signals */
908 sigfillset(&all_sigs
);
909 ret
= sigprocmask(SIG_BLOCK
, &all_sigs
, save_sigset
);
911 PERROR("sigprocmask");
914 rcu_bp_before_fork();
917 static void ust_after_fork_common(sigset_t
*restore_sigset
)
921 DBG("process %d", getpid());
923 /* Restore signals */
924 ret
= sigprocmask(SIG_SETMASK
, restore_sigset
, NULL
);
926 PERROR("sigprocmask");
930 void ust_after_fork_parent(sigset_t
*restore_sigset
)
932 DBG("process %d", getpid());
933 rcu_bp_after_fork_parent();
934 /* Release mutexes and reenable signals */
935 ust_after_fork_common(restore_sigset
);
939 * After fork, in the child, we need to cleanup all the leftover state,
940 * except the worker thread which already magically disappeared thanks
941 * to the weird Linux fork semantics. After tyding up, we call
942 * lttng_ust_init() again to start over as a new PID.
944 * This is meant for forks() that have tracing in the child between the
945 * fork and following exec call (if there is any).
947 void ust_after_fork_child(sigset_t
*restore_sigset
)
949 DBG("process %d", getpid());
950 /* Release urcu mutexes */
951 rcu_bp_after_fork_child();
952 lttng_ust_cleanup(0);
953 lttng_context_vtid_reset();
954 /* Release mutexes and reenable signals */
955 ust_after_fork_common(restore_sigset
);