9pfs: local: fix fchmodat_nofollow() limitations

This function has to ensure it doesn't follow a symlink that could be used
to escape the virtfs directory. This could be easily achieved if fchmodat()
on linux honored the AT_SYMLINK_NOFOLLOW flag as described in POSIX, but
it doesn't. There was a tentative to implement a new fchmodat2() syscall
with the correct semantics:

https://patchwork.kernel.org/patch/9596301/



but it didn't gain much momentum. Also it was suggested to look at an O_PATH
based solution in the first place.

The current implementation covers most use-cases, but it notably fails if:
- the target path has access rights equal to 0000 (openat() returns EPERM),
  => once you've done chmod(0000) on a file, you can never chmod() again
- the target path is UNIX domain socket (openat() returns ENXIO)
  => bind() of UNIX domain sockets fails if the file is on 9pfs

The solution is to use O_PATH: openat() now succeeds in both cases, and we
can ensure the path isn't a symlink with fstat(). The associated entry in
"/proc/self/fd" can hence be safely passed to the regular chmod() syscall.

The previous behavior is kept for older systems that don't have O_PATH.

Signed-off-by: Greg Kurz <groug@kaod.org>
Reviewed-by: Eric Blake <eblake@redhat.com>
Tested-by: Zhi Yong Wu <zhiyong.wu@ucloud.cn>
Acked-by: Philippe Mathieu-Daudé <f4bug@amsat.org>
(cherry picked from commit 4751fd53)
Signed-off-by: Michael Roth <mdroth@linux.vnet.ibm.com>