Bug#661627: Avoid /tmp ?

[vladz <vladz@devzero.fr>]

On Fri, Mar 02, 2012 at 02:29:33PM +0100, Julien Cristau wrote:
> I'm not convinced the chown can be removed.  And 'mkdir -m 1777 foo' is
> not any more atomic than 'mkdir foo && chmod 1777 foo'.  

The command "mkdir -m" calls the mkdir() syscall, and its second
argument seems to be the mode.  

  $ man 2 mkdir
  [...]
  int mkdir(const char *pathname, mode_t mode);

Maybe I'm wrong, but this is what I call an atomic way to create and set
permissions (ie. two operations in a unique syscall).  For example:

  $ strace mkdir -m 222 /tmp/foo
  [...]
  mkdir("/tmp/foo", 0222)                 = 0

> The problem is
> that I'd still like to be able to chown and chmod /tmp/.X11-unix if it
> already exists as a directory when the script runs.  I can do that in C
> with something like this:
> 
>   ret = mkdir("/tmp/.X11-unix", 0700);
>   if (ret == 0 || errno == EEXIST) {
>     fd = open("/tmp/.X11-unix", O_RDONLY | O_NOFOLLOW);
>     if (fd < 0)
>       fail();
>     fstat(fd, &st);
>     if (!S_ISDIR(st.st_mode))
>       fail();
>     if (fchown(fd, 0, 0)) fail();
>     if (fchmod(fd, 01777)) fail();

Yes, fchown & fchmod use file descriptors, those functions are safer
than commands chmod & chown that directly use filenames...  This is why
I consider those commands dangerous.

> hmm, how about this:
>
>mkdir -p /tmp/.X11-unix
>chown -h root:root /tmp/.X11-unix
>stat=$(LC_ALL=C stat -c '%u %g %F' /tmp/.X11-unix)
>if [ "$stat" != '0 0 directory' ]; then
>  exit 1
>fi
>chmod 1777 /tmp/.X11-unix

This would work (even if it uses chmod), but wasn't the Bash approach (test
with "-O", "-G" and "-d") simpler than using "stat"?

> Agreed.  Or drop those two functions and call set_up_dir "$SOCKET_DIR &&
> set_up_dir $ICE_DIR" directly.

Agreed. :)