Re: [PATCH 1/3] switch_creds: Syscall to switch creds for file server ops

[ebiederm@xmission.com (Eric W. Biederman)]

Al Viro <viro@ZenIV.linux.org.uk> writes:

> On Wed, Oct 16, 2013 at 03:01:57PM -0700, Jim Lieb wrote:
>> File servers must do some operations with the credentials of
>> their client.  This syscall switches the key credentials similar
>> to nfsd_setuser() in fs/nfsd/auth.c  with the capability of retaining a
>> handle to the credentials by way of an fd for an open anonymous file.
>> This makes switching for subsequent operations for that client more efficient.
>
> 	Yet Another Untyped Multiplexor.  Inna bun.  Onna stick.
> CMOT Dibbler special...
>
> 	Switching creds to those of opener of given file descriptor
> is fine, but in any realistic situation you'll get all the real win
> from that - you should cache those fds (which you seem to do), and
> then setuid/etc. is done once per cache miss.  Making the magical
> "set them all at once" mess (complete with non-trivial structure,
> 32/64bit compat, etc.) pointless.  Moreover, you don't need any magic
> files at all - just set the creds and open /dev/null and there's your fd.
> With proper creds associated with it.  While we are at it, just _start_
> with opening /dev/null.  With your initial creds.  Voila - revert is
> simply switch to that fd's creds.
>
> 	IOW, you really need only one syscall:

That doesn't look bad but it does need capable(CAP_SETUID) &&
capable(CAP_SETGID) or possibly something a little more refined.

I don't think we want file descriptor passing to all of a sudden become
a grant of privilege, beyond what the passed fd can do.

> SYSCALL_DEFINE1(switch_cred, int, fd)
> {
> 	struct fd f = fdget(fd);
> 	if (!f.file)
> 		return -EBADF;
> 	put_cred(override_creds(f.file->f_cred);
> 	fdput(f);
> 	return 0;
> }
>
> and that's all there is to it.

Eric