cifs.upcall: fix regression in kerberos mount

cifs.upcall: fix regression in kerberos mount

The fix for CVE-2021-20208 in commit e461afd ("cifs.upcall: try to use
container ipc/uts/net/pid/mnt/user namespaces") introduced a
regression for kerberos mounts when cifs-utils is built with
libcap-ng. It makes mount fail with ENOKEY "Required key not
available".

Current state:

mount.cifs
 '---> mount() ---> kernel
       	           negprot, session setup (need security blob for krb)
                   request_key("cifs.spnego",  payload="pid=%d;username=...")
                               upcall
  /sbin/request-key <--------------'
   reads /etc/request-keys.conf
   dispatch cifs.spnego request
   calls /usr/sbin/cifs.upcall <key id>
   - drop privileges (capabilities)
   - fetch keyid
   - parse payload
   - switch to mount.cifs namespaces
   - call krb5_xxx() funcs
   - generate security blob
   - set key value to security blob
      '-----------------------------------> kernel
                                         put blob in session setup packet
  				       continue auth
  				       open tcon
  				       get share root
  				       setup superblock
mount.cifs mount() returns    <-----------'

By the time cifs.upcall tries to switch to namespaces, enough
capabilities have dropped in trim_capabilities() that it makes setns()
fail with EPERM.

setns() requires CAP_SYS_ADMIN.

With libcap trim_capabilities() is a no-op.

This fix:

- moves the namespace switch earlier so that operations like
  setgroups(), setgid(), scanning of pid environment, ... happens in the
  contained namespaces.
- moves trim_capabilities() after the namespace switch
- moves the string processing to decode the key request payload in a
  child process with minimum capabilities. the decoded data is shared
  with the parent process via shared memory obtained with mmap().

Fixes: e461afd ("cifs.upcall: try to use container ipc/uts/net/pid/mnt/user namespaces")
Signed-off-by: Aurelien Aptel <aaptel@suse.com>