* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
*/
+#define _LGPL_SOURCE
#include <sys/types.h>
#include <sys/socket.h>
#include <sys/prctl.h>
#include <sys/mman.h>
#include <sys/stat.h>
+#include <sys/types.h>
+#include <sys/wait.h>
#include <fcntl.h>
#include <unistd.h>
#include <errno.h>
#include <assert.h>
#include <signal.h>
#include <urcu/uatomic.h>
+#include <urcu/futex.h>
#include <lttng-ust-comm.h>
#include <ust/usterr-signal-safe.h>
int root_handle;
int constructor_sem_posted;
int allowed;
+ int global;
char sock_path[PATH_MAX];
int socket;
/* Socket from app (connect) to session daemon (listen) for communication */
struct sock_info global_apps = {
.name = "global",
+ .global = 1,
.root_handle = -1,
.allowed = 1,
struct sock_info local_apps = {
.name = "local",
+ .global = 0,
.root_handle = -1,
.allowed = 0, /* Check setuid bit first */
.socket = -1,
};
+static int wait_poll_fallback;
+
extern void ltt_ring_buffer_client_overwrite_init(void);
extern void ltt_ring_buffer_client_discard_init(void);
extern void ltt_ring_buffer_metadata_client_init(void);
if (sock_info->constructor_sem_posted)
return 0;
sock_info->constructor_sem_posted = 1;
+ if (uatomic_read(&sem_count) <= 0) {
+ return 0;
+ }
ret = uatomic_add_return(&sem_count, -1);
if (ret == 0) {
ret = sem_post(&constructor_wait);
} else {
lur.ret_code = LTTCOMM_SESSION_FAIL;
}
+ if (lum->cmd == LTTNG_UST_STREAM) {
+ /*
+ * Special-case reply to send stream info.
+ * Use lum.u output.
+ */
+ lur.u.stream.memory_map_size = lum->u.stream.memory_map_size;
+ }
ret = send_reply(sock, &lur);
+ if (lum->cmd == LTTNG_UST_STREAM && ret >= 0) {
+ /* we also need to send the file descriptors. */
+ ret = lttcomm_send_fds_unix_sock(sock,
+ &lum->u.stream.shm_fd, &lum->u.stream.shm_fd,
+ 1, sizeof(int));
+ if (ret < 0) {
+ perror("send shm_fd");
+ goto error;
+ }
+ ret = lttcomm_send_fds_unix_sock(sock,
+ &lum->u.stream.wait_fd, &lum->u.stream.wait_fd,
+ 1, sizeof(int));
+ if (ret < 0) {
+ perror("send wait_fd");
+ goto error;
+ }
+ }
+error:
ust_unlock();
return ret;
}
}
}
+/*
+ * Using fork to set umask in the child process (not multi-thread safe).
+ * We deal with the shm_open vs ftruncate race (happening when the
+ * sessiond owns the shm and does not let everybody modify it, to ensure
+ * safety against shm_unlink) by simply letting the mmap fail and
+ * retrying after a few seconds.
+ * For global shm, everybody has rw access to it until the sessiond
+ * starts.
+ */
static
-char *get_map_shm(struct sock_info *sock_info)
+int get_wait_shm(struct sock_info *sock_info, size_t mmap_size)
{
- size_t mmap_size = sysconf(_SC_PAGE_SIZE);
int wait_shm_fd, ret;
- char *wait_shm_mmap;
+ pid_t pid;
/*
- * Get existing (read-only) shm, or open new shm.
- * First try to open read-only.
+ * Try to open read-only.
*/
- wait_shm_fd = shm_open(sock_info->wait_shm_path,
- O_RDONLY, 0700);
- if (wait_shm_fd >= 0)
- goto got_shm;
- /*
- * Real-only open did not work. If it is because it did
- * not exist, try creating it. Else it's a failure that
- * prohibits using shm.
- */
- if (errno != ENOENT) {
- ERR("Error opening shm %s", sock_info->wait_shm_path);
- goto error;
- }
- wait_shm_fd = shm_open(sock_info->wait_shm_path,
- O_RDWR | O_CREAT | O_EXCL, 0700);
- if (wait_shm_fd >= 0)
- goto created_shm;
- if (errno != EEXIST) {
+ wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
+ if (wait_shm_fd >= 0) {
+ goto end;
+ } else if (wait_shm_fd < 0 && errno != ENOENT) {
+ /*
+ * Real-only open did not work, and it's not because the
+ * entry was not present. It's a failure that prohibits
+ * using shm.
+ */
ERR("Error opening shm %s", sock_info->wait_shm_path);
- goto error;
+ goto end;
}
/*
- * If another process beat us to create the shm, we are
- * pretty certain the shm is available for us in
- * read-only mode.
+ * If the open failed because the file did not exist, try
+ * creating it ourself.
*/
- wait_shm_fd = shm_open(sock_info->wait_shm_path,
- O_RDWR | O_CREAT | O_EXCL, 0700);
- if (wait_shm_fd >= 0)
- goto got_shm;
- else
- goto error;
+ pid = fork();
+ if (pid > 0) {
+ int status;
-created_shm:
- ret = ftruncate(wait_shm_fd, mmap_size);
- if (ret) {
- PERROR("ftruncate");
- ret = close(wait_shm_fd);
+ /*
+ * Parent: wait for child to return, in which case the
+ * shared memory map will have been created.
+ */
+ pid = wait(&status);
+ if (!WIFEXITED(status) || WEXITSTATUS(status) != 0) {
+ wait_shm_fd = -1;
+ goto end;
+ }
+ /*
+ * Try to open read-only again after creation.
+ */
+ wait_shm_fd = shm_open(sock_info->wait_shm_path, O_RDONLY, 0);
+ if (wait_shm_fd < 0) {
+ /*
+ * Real-only open did not work. It's a failure
+ * that prohibits using shm.
+ */
+ ERR("Error opening shm %s", sock_info->wait_shm_path);
+ goto end;
+ }
+ goto end;
+ } else if (pid == 0) {
+ int create_mode;
+
+ /* Child */
+ create_mode = S_IRUSR | S_IWUSR | S_IRGRP;
+ if (sock_info->global)
+ create_mode |= S_IROTH | S_IWGRP | S_IWOTH;
+ /*
+ * We're alone in a child process, so we can modify the
+ * process-wide umask.
+ */
+ umask(~create_mode);
+ /*
+ * Try creating shm (or get rw access).
+ * We don't do an exclusive open, because we allow other
+ * processes to create+ftruncate it concurrently.
+ */
+ wait_shm_fd = shm_open(sock_info->wait_shm_path,
+ O_RDWR | O_CREAT, create_mode);
+ if (wait_shm_fd >= 0) {
+ ret = ftruncate(wait_shm_fd, mmap_size);
+ if (ret) {
+ PERROR("ftruncate");
+ exit(EXIT_FAILURE);
+ }
+ exit(EXIT_SUCCESS);
+ }
+ /*
+ * For local shm, we need to have rw access to accept
+ * opening it: this means the local sessiond will be
+ * able to wake us up. For global shm, we open it even
+ * if rw access is not granted, because the root.root
+ * sessiond will be able to override all rights and wake
+ * us up.
+ */
+ if (!sock_info->global && errno != EACCES) {
+ ERR("Error opening shm %s", sock_info->wait_shm_path);
+ exit(EXIT_FAILURE);
+ }
+ /*
+ * The shm exists, but we cannot open it RW. Report
+ * success.
+ */
+ exit(EXIT_SUCCESS);
+ } else {
+ return -1;
+ }
+end:
+ if (wait_shm_fd >= 0 && !sock_info->global) {
+ struct stat statbuf;
+
+ /*
+ * Ensure that our user is the owner of the shm file for
+ * local shm. If we do not own the file, it means our
+ * sessiond will not have access to wake us up (there is
+ * probably a rogue process trying to fake our
+ * sessiond). Fallback to polling method in this case.
+ */
+ ret = fstat(wait_shm_fd, &statbuf);
if (ret) {
- ERR("Error closing fd");
+ PERROR("fstat");
+ goto error_close;
}
- wait_shm_fd = -1;
+ if (statbuf.st_uid != getuid())
+ goto error_close;
+ }
+ return wait_shm_fd;
+
+error_close:
+ ret = close(wait_shm_fd);
+ if (ret) {
+ PERROR("Error closing fd");
+ }
+ return -1;
+}
+
+static
+char *get_map_shm(struct sock_info *sock_info)
+{
+ size_t mmap_size = sysconf(_SC_PAGE_SIZE);
+ int wait_shm_fd, ret;
+ char *wait_shm_mmap;
+
+ wait_shm_fd = get_wait_shm(sock_info, mmap_size);
+ if (wait_shm_fd < 0) {
goto error;
}
-got_shm:
wait_shm_mmap = mmap(NULL, mmap_size, PROT_READ,
MAP_SHARED, wait_shm_fd, 0);
- if (wait_shm_mmap == MAP_FAILED) {
- PERROR("mmap");
- goto error;
- }
/* close shm fd immediately after taking the mmap reference */
ret = close(wait_shm_fd);
if (ret) {
- ERR("Error closing fd");
+ PERROR("Error closing fd");
+ }
+ if (wait_shm_mmap == MAP_FAILED) {
+ DBG("mmap error (can be caused by race with sessiond). Fallback to poll mode.");
+ goto error;
}
return wait_shm_mmap;
static
void wait_for_sessiond(struct sock_info *sock_info)
{
+ int ret;
+
ust_lock();
if (lttng_ust_comm_should_quit) {
goto quit;
}
+ if (wait_poll_fallback) {
+ goto error;
+ }
if (!sock_info->wait_shm_mmap) {
sock_info->wait_shm_mmap = get_map_shm(sock_info);
if (!sock_info->wait_shm_mmap)
ust_unlock();
DBG("Waiting for %s apps sessiond", sock_info->name);
- /* Wait for futex wakeup TODO */
- sleep(5);
-
+ /* Wait for futex wakeup */
+ if (uatomic_read((int32_t *) sock_info->wait_shm_mmap) == 0) {
+ ret = futex_async((int32_t *) sock_info->wait_shm_mmap,
+ FUTEX_WAIT, 0, NULL, NULL, 0);
+ if (ret < 0) {
+ if (errno == EFAULT) {
+ wait_poll_fallback = 1;
+ ERR(
+"Linux kernels 2.6.33 to 3.0 (with the exception of stable versions) "
+"do not support FUTEX_WAKE on read-only memory mappings correctly. "
+"Please upgrade your kernel "
+"(fix is commit 9ea71503a8ed9184d2d0b8ccc4d269d05f7940ae in Linux kernel "
+"mainline). LTTng-UST will use polling mode fallback.");
+ }
+ PERROR("futex");
+ }
+ }
return;
quit:
error:
ust_unlock();
- /* Error handling: fallback on a 5 seconds sleep. */
- sleep(5);
return;
}
void *ust_listener_thread(void *arg)
{
struct sock_info *sock_info = arg;
- int sock, ret;
+ int sock, ret, prev_connect_failed = 0, has_waited = 0;
/* Restart trying to connect to the session daemon */
restart:
+ if (prev_connect_failed) {
+ /* Wait for sessiond availability with pipe */
+ wait_for_sessiond(sock_info);
+ if (has_waited) {
+ has_waited = 0;
+ /*
+ * Sleep for 5 seconds before retrying after a
+ * sequence of failure / wait / failure. This
+ * deals with a killed or broken session daemon.
+ */
+ sleep(5);
+ }
+ has_waited = 1;
+ prev_connect_failed = 0;
+ }
ust_lock();
if (lttng_ust_comm_should_quit) {
ret = lttcomm_connect_unix_sock(sock_info->sock_path);
if (ret < 0) {
ERR("Error connecting to %s apps socket", sock_info->name);
+ prev_connect_failed = 1;
/*
* If we cannot find the sessiond daemon, don't delay
* constructor execution.
ret = handle_register_done(sock_info);
assert(!ret);
ust_unlock();
-
- /* Wait for sessiond availability with pipe */
- wait_for_sessiond(sock_info);
goto restart;
}
ret = register_app_to_sessiond(sock);
if (ret < 0) {
ERR("Error registering to %s apps socket", sock_info->name);
+ prev_connect_failed = 1;
/*
* If we cannot register to the sessiond daemon, don't
* delay constructor execution.
ret = handle_register_done(sock_info);
assert(!ret);
ust_unlock();
- wait_for_sessiond(sock_info);
goto restart;
}
ust_unlock();