int vz_env_create(vps_handler *h, envid_t veid, vps_res *res,
int wait_p[2], int old_wait_p[2], int err_p[2],
env_create_FN fn, void *data)
{
int ret, pid, errcode;
int old_wait_fd;
int status_p[2];
struct sigaction act, actold;
if (check_var(res->fs.root, "VE_ROOT is not set"))
return VZ_VE_ROOT_NOTSET;
if (pipe(status_p) < 0) {
logger(-1, errno, "Can not create pipe");
return VZ_RESOURCE_ERROR;
}
sigaction(SIGCHLD, NULL, &actold);
sigemptyset(&act.sa_mask);
act.sa_handler = SIG_IGN;
act.sa_flags = SA_NOCLDSTOP;
sigaction(SIGCHLD, &act, NULL);
get_osrelease(res);
if ((pid = fork()) < 0) {
logger(-1, errno, "Can not fork");
ret = VZ_RESOURCE_ERROR;
goto err;
} else if (pid == 0) {
dup2(status_p[1], STDIN_FILENO);
close(status_p[0]);
close(status_p[1]);
fcntl(STDIN_FILENO, F_SETFD, FD_CLOEXEC);
fcntl(err_p[1], F_SETFD, FD_CLOEXEC);
close(err_p[0]);
fcntl(wait_p[0], F_SETFD, FD_CLOEXEC);
close(wait_p[1]);
if (old_wait_p) {
fcntl(old_wait_p[0], F_SETFD, FD_CLOEXEC);
close(old_wait_p[1]);
old_wait_fd = old_wait_p[0];
} else
old_wait_fd = -1;
ret = vz_real_env_create(h, veid, res, wait_p[0],
old_wait_fd, err_p[1], fn, data);
if (ret)
write(STDIN_FILENO, &ret, sizeof(ret));
exit(ret);
}
/* Wait for environment created */
close(status_p[1]);
close(wait_p[0]);
if (old_wait_p)
close(old_wait_p[0]);
close(err_p[1]);
ret = read(status_p[0], &errcode, sizeof(errcode));
if (ret > 0) {
ret = errcode;
switch(ret) {
case VZ_NO_ACCES:
logger(-1, 0, "Permission denied");
break;
case VZ_BAD_KERNEL:
logger(-1, 0, "Invalid kernel, or some kernel"
" modules are not loaded");
break;
case VZ_SET_CAP:
logger(-1, 0, "Unable to set capability");
break;
case VZ_RESOURCE_ERROR:
logger(-1, 0, "Not enough resources"
" to start environment");
break;
case VZ_WAIT_FAILED:
logger(0, 0, "Unable to set"
" wait functionality");
break;
case VZ_SET_OSRELEASE:
logger(-1, 0, "Unable to set osrelease to %s",
res->env.osrelease);
break;
}
}
err:
close(status_p[1]);
close(status_p[0]);
sigaction(SIGCHLD, &actold, NULL);
return ret;
}
static int _env_create(vps_handler *h, envid_t veid, int wait_p, int err_p,
void *data)
{
struct vzctl_env_create_data env_create_data;
struct env_create_param3 create_param;
int fd, ret;
vps_res *res;
char *argv[] = {"init", "-z", " ", NULL};
char *envp[] = {"HOME=/", "TERM=linux", NULL};
res = (vps_res *) data;
memset(&create_param, 0, sizeof(create_param));
create_param.iptables_mask = get_ipt_mask(res->env.ipt_mask);
logger(3, 0, "Set iptables mask %#10.8llx",
(unsigned long long) create_param.iptables_mask);
clean_hardlink_dir("/");
if (res->cpu.vcpus != NULL)
create_param.total_vcpus = *res->cpu.vcpus;
env_create_data.veid = veid;
env_create_data.class_id = 0;
env_create_data.flags = VE_CREATE | VE_EXCLUSIVE;
env_create_data.data = &create_param;
env_create_data.datalen = sizeof(create_param);
create_param.feature_mask = res->env.features_mask;
create_param.known_features = res->env.features_known;
/* sysfs enabled by default, unless explicitly disabled */
if (! (res->env.features_known & VE_FEATURE_SYSFS)) {
create_param.feature_mask |= VE_FEATURE_SYSFS;
create_param.known_features |= VE_FEATURE_SYSFS;
}
logger(3, 0, "Set features mask %016llx/%016llx",
create_param.feature_mask,
create_param.known_features);
/* Close all fds except stdin. stdin is status pipe */
close(STDERR_FILENO);
close(STDOUT_FILENO);
close_fds(0, wait_p, err_p, h->vzfd, -1);
try:
ret = vz_env_create_data_ioctl(h, &env_create_data);
if (ret < 0) {
switch(errno) {
case EINVAL:
ret = VZ_ENVCREATE_ERROR;
/* Run-time kernel did not understand the
* latest create_param -- so retry with
* the old env_create_param structs.
*/
switch (env_create_data.datalen) {
case sizeof(struct env_create_param3):
env_create_data.datalen =
sizeof(struct env_create_param2);
goto try;
case sizeof(struct env_create_param2):
env_create_data.datalen =
sizeof(struct env_create_param);
goto try;
}
break;
case EACCES:
/* License is not loaded */
ret = VZ_NO_ACCES;
break;
case ENOTTY:
/* Some vz modules are not present */
ret = VZ_BAD_KERNEL;
break;
default:
logger(-1, errno, "env_create error");
ret = VZ_ENVCREATE_ERROR;
break;
}
goto env_err;
}
if (res->env.osrelease != NULL) {
ret = vz_env_configure(h->vzfd, veid,
res->env.osrelease);
if (ret != 0)
goto env_err;
}
close(h->vzfd);
/* Create /fastboot to skip run fsck */
fd = open("/fastboot", O_CREAT | O_RDONLY, 0644);
close(fd);
if (res->misc.wait == YES) {
if (add_reach_runlevel_mark()) {
ret = VZ_WAIT_FAILED;
goto env_err;
}
}
mount("proc", "/proc", "proc", 0, 0);
if (stat_file("/sys"))
mount("sysfs", "/sys", "sysfs", 0, 0);
if (create_param.feature_mask & VE_FEATURE_NFSD) {
mount("nfsd", "/proc/fs/nfsd", "nfsd", 0, 0);
make_dir("/var/lib/nfs/rpc_pipefs", 1);
mount("sunrpc", "/var/lib/nfs/rpc_pipefs", "rpc_pipefs", 0, 0);
}
configure_sysctl();
if (res->dq.ugidlimit != NULL)
mk_quota_link();
/* Close status descriptor to report that
* environment is created.
*/
close(STDIN_FILENO);
/* Now we wait until CT setup will be done
If no error, then start init, otherwise exit.
*/
if (read(wait_p, &ret, sizeof(ret)) == 0)
return 0;
if ((fd = open("/dev/null", O_RDWR)) != -1) {
dup2(fd, 0);
dup2(fd, 1);
dup2(fd, 2);
}
logger(10, 0, "Starting init");
execve("/sbin/init", argv, envp);
execve("/etc/init", argv, envp);
execve("/bin/init", argv, envp);
ret = VZ_FS_BAD_TMPL;
write(err_p, &ret, sizeof(ret));
env_err:
return ret;
}
static int vz_real_env_create(vps_handler *h, envid_t veid, vps_res *res,
int wait_p, int old_wait_p, int err_p, env_create_FN fn, void *data)
{
int ret, pid;
if ((ret = vz_chroot(res->fs.root)))
return ret;
if ((ret = vz_setluid(veid)))
return ret;
if ((ret = setup_resource_management(h, veid, res)))
return ret;
/* Create another process for proper resource accounting */
if ((pid = fork()) < 0) {
logger(-1, errno, "Unable to fork");
return VZ_RESOURCE_ERROR;
} else if (pid == 0) {
if ((ret = vps_set_cap(veid, &res->env, &res->cap)))
goto env_err;
if (fn == NULL) {
ret = _env_create(h, veid, wait_p, err_p, (void *)res);
} else {
ret = fn(h, veid, wait_p, old_wait_p, err_p, data);
}
env_err:
if (ret)
write(STDIN_FILENO, &ret, sizeof(ret));
exit(ret);
}
return 0;
}
int vz_env_create(vps_handler *h, envid_t veid, vps_res *res,
int wait_p[2], int old_wait_p[2], int err_p[2],
env_create_FN fn, void *data)
{
int ret, pid, errcode;
int old_wait_fd;
int status_p[2];
struct sigaction act, actold;
if (check_var(res->fs.root, "VE_ROOT is not set"))
return VZ_VE_ROOT_NOTSET;
if (pipe(status_p) < 0) {
logger(-1, errno, "Can not create pipe");
return VZ_RESOURCE_ERROR;
}
sigaction(SIGCHLD, NULL, &actold);
sigemptyset(&act.sa_mask);
act.sa_handler = SIG_IGN;
act.sa_flags = SA_NOCLDSTOP;
sigaction(SIGCHLD, &act, NULL);
get_osrelease(res);
if ((pid = fork()) < 0) {
logger(-1, errno, "Can not fork");
ret = VZ_RESOURCE_ERROR;
goto err;
} else if (pid == 0) {
dup2(status_p[1], STDIN_FILENO);
close(status_p[0]);
close(status_p[1]);
fcntl(STDIN_FILENO, F_SETFD, FD_CLOEXEC);
fcntl(err_p[1], F_SETFD, FD_CLOEXEC);
close(err_p[0]);
fcntl(wait_p[0], F_SETFD, FD_CLOEXEC);
close(wait_p[1]);
if (old_wait_p) {
fcntl(old_wait_p[0], F_SETFD, FD_CLOEXEC);
close(old_wait_p[1]);
old_wait_fd = old_wait_p[0];
} else
old_wait_fd = -1;
ret = vz_real_env_create(h, veid, res, wait_p[0],
old_wait_fd, err_p[1], fn, data);
if (ret)
write(STDIN_FILENO, &ret, sizeof(ret));
exit(ret);
}
/* Wait for environment created */
close(status_p[1]);
close(wait_p[0]);
if (old_wait_p)
close(old_wait_p[0]);
close(err_p[1]);
ret = read(status_p[0], &errcode, sizeof(errcode));
if (ret > 0) {
ret = errcode;
switch(ret) {
case VZ_NO_ACCES:
logger(-1, 0, "Permission denied");
break;
case VZ_BAD_KERNEL:
logger(-1, 0, "Invalid kernel, or some kernel"
" modules are not loaded");
break;
case VZ_SET_CAP:
logger(-1, 0, "Unable to set capability");
break;
case VZ_RESOURCE_ERROR:
logger(-1, 0, "Not enough resources"
" to start environment");
break;
case VZ_WAIT_FAILED:
logger(0, 0, "Unable to set"
" wait functionality");
break;
case VZ_SET_OSRELEASE:
logger(-1, 0, "Unable to set osrelease to %s",
res->env.osrelease);
break;
}
}
err:
close(status_p[1]);
close(status_p[0]);
sigaction(SIGCHLD, &actold, NULL);
return ret;
}