Re: [PATCH 27/43] userns: Use uid_eq gid_eq helpers when comparing kuids and kgids in the vfs

["Serge E. Hallyn" <serge@hallyn.com>]

Quoting Eric W. Biederman (ebiederm@xmission.com):
> "Serge E. Hallyn" <serge@hallyn.com> writes:
> 
> > Quoting Eric W. Beiderman (ebiederm@xmission.com):
> >> From: Eric W. Biederman <ebiederm@xmission.com>
> >> 
> >> Signed-off-by: Eric W. Biederman <ebiederm@xmission.com>
> >> ---
> >>  fs/attr.c                |    8 ++++----
> >>  fs/exec.c                |   10 +++++-----
> >>  fs/fcntl.c               |    6 +++---
> >>  fs/ioprio.c              |    4 ++--
> >>  fs/locks.c               |    2 +-
> >>  fs/namei.c               |    8 ++++----
> >>  include/linux/quotaops.h |    4 ++--
> >>  7 files changed, 21 insertions(+), 21 deletions(-)
> >> 
> 
> >> @@ -2120,7 +2120,7 @@ void do_coredump(long signr, int exit_code, struct pt_regs *regs)
> >>  	if (__get_dumpable(cprm.mm_flags) == 2) {
> >>  		/* Setuid core dump mode */
> >>  		flag = O_EXCL;		/* Stop rewrite attacks */
> >> -		cred->fsuid = 0;	/* Dump root private */
> >> +		cred->fsuid = GLOBAL_ROOT_UID;	/* Dump root private */
> >
> > Sorry, one more - can this be the per-ns root uid?  The coredumps should
> > be ok to belong to privileged users in the namespace right?
> 
> I'm not certain it was clear when you were looking at this that
> this is about dumping core from suid applications, not normal
> applications. 

Right, that makes sense,  but I was thinking we could just find the
root uid for the namespace in which current_euid() is valid.

Waiting is fine :)

>  Looking at the code in commoncap and commit_creds it looks like it is a
> bug that we don't call set_dumpable(new, suid_dumpable) in common cap

That sounds like a bug, yes.

> when we use file capabilities.  I might be wrong but I think we escape
> the test in commit_creds in that case.
> 
> 
> Having thought about it we can make this per namespace but not in
> this patch.
> 
> Things that I see as missing.
> - We likely need to make the suid_dumpable sysctl per namespace.
>   There is a prctl so it is already per process.
> - We would need to capture the user_namespace at mm creation time,
>   during exec, so we know which root user we could use.
>
>   By it's nature we know an mm can't escape a user namespace so the
>   user namespace an mm is created in will have a root user we can
>   dump core as.
> 
> I was wondering if we could relax this to a uid captured at mm creation
> time (and certainly we can capture the root user), but there are enough
> weird cases I don't think it is possible to safely allow anything more
> relaxed that the root of the user_namespace that created the mm.
> 
> I don't believe we can't use the user_namespace of current because the
> application may have been suid and then cloned a new user namespace
> keeping the mm or perhaps just the uid/euid split.
> 
> So in short it is doable but a little tricky so it doesn't belong in
> a patch with a bunch of boring and safe conversions.
> 
> Eric

Thanks for elaborating on that.  Makes sense to wait.

-serge