"Daniel P. Berrange" <berrange(a)redhat.com> wrote:
Some of the existing usage of fork/exec in libvirt is done such that
the
child process is daemonized. In particular the libvirt_proxy and the
auto-spawned libvirtd for qemu:///session. Since we want to switch these
to use virExec, we need to suport a daemon mode.
This patch removes the two alternate virExec and virExecNonBlock functions
and renames the internal __virExec to virExec. It then gains a 'flags'
parameter which can be used to specify non-blocking mode, or daemon mode.
We also add the ability to pass in a list of environment variables which
get passed on to execve(). We also now explicitly close all file handles.
Although libvirt code is careful to set O_CLOSEXEC on all its file handles,
in multi-threaded apps there is a race condition between opening a FD and
setting O_CLOSEXEC. Furthermore, we can't guarentee that all applications
using libvirt are careful to set O_CLOSEXEC. Leaking FDs to a child is a
potential security risk and often causes SELinux AVCs to be raised. Thus
explicitely closing all FDs is a important safety net.
How about closing those FDs in the child instead, right after the fork?
Then, if a virExec caller has an open socket or file descriptor,
it won't be closed behind its back.
diff -r 28ddf9f5791c src/util.c
...
+ for (i = 3; i < openmax; i++)
+ if (i != infd &&
+ i != null &&
+ i != childout &&
+ i != childerr)
+ close(i);
+
+ if (flags & VIR_EXEC_DAEMON) {
+ if (setsid() < 0) {
+ ReportError(conn, VIR_ERR_INTERNAL_ERROR,
+ _("cannot become session leader: %s"),
+ strerror(errno));
+ _exit(1);
+ }
+
+ if (chdir("/") < 0) {
+ ReportError(conn, VIR_ERR_INTERNAL_ERROR,
+ _("cannot change to root directory: %s"),
+ strerror(errno));
+ _exit(1);
+ }
+
+ pid = fork();
...