Re: [PATCH v9 0/4] fuse: mounts from non-init user namespaces

[ebiederm-aS9lmoZGLiVWk0Htik3J/w@public.gmane.org (Eric W. Biederman)]

Miklos Szeredi <mszeredi-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org> writes:

> On Thu, Mar 8, 2018 at 10:23 PM, Eric W. Biederman
> <ebiederm-aS9lmoZGLiVWk0Htik3J/w@public.gmane.org> wrote:
>>
>> This patchset builds on the work by Donsu Park and Seth Forshee and is
>> reduced to the set of patches that just affect fuse.  The non-fuse
>> vfs patches are far enough along we can ignore them except possibly for the
>> question of when does FS_USERNS_MOUNT get set in fuse_fs_type.
>>
>> Fuse with a block device has been left as an exercise for a later time.
>>
>> Since v5 I changed the core of this patchset around as the previous
>> patches were showing signs of bitrot.  Some important explanations were
>> missing, some important functionality was missing, and xattr handling
>> was completely absent.
>>
>> Since v6 I have:
>> - Removed the failure case from fuse_get_req_nofail_nopages that I
>>   added.
>> - Updated fuse to always to use posix_acl_access_xattr_handler, and
>>   posix_acl_default_xattr_handler, by teaching fuse to set
>>   ACL_DONT_CACHE when FUSE_POSIX_ACL is not set.
>>
>> Since v7 I have:
>> - Rethought and reworked how I am unifying the cached and the non-cached
>>   posix acl case so the code is cleaner and simpler.
>>   - I have dropped enhancements to caching negative acls when
>>     fc->no_getxattr is set.
>>   - Removed the need to wrap forget_all_cached_acls in fuse.
>> - Reorder the patches so the posix acl work comes first
>>
>> Since v8 I have:
>> - Dropped and postponed the unification of the uncached and the cached
>>   posix acls case.  The code is not hard but tricky enough it needs
>>   to be considered on it's own on it's own merits.
>>
>> Miklos can you take a look and see what you think?
>>
>> Miklos if you could pick these up I would appreciate it.  If not I can
>> merge these through the userns tree.
>
> Thank you Eric for moving this along.  Patches pushed to:
>
>   git://git.kernel.org/pub/scm/linux/kernel/git/mszeredi/fuse.git for-next
>
> I did just one modification to "fuse: Fail all requests with invalid
> uids or gids": instead of zeroing out the context for the nofail case,
> continue to use the "_munged" variants. I don't think this hurts and
> is better for backward compatibility (I guess the only relevant use
> would be for debugging output, but we don't want to regress even for
> that if not necessary)

Hmm...

The thing is the failure doesn't come in the difference between the
_munged and the normal variants.  The difference between
munged and non-munged variants is how they handled failure ((uid16_t)-2)
aka 0xfffe for munged and -1 for the non-munged case.

The failures are introduced by changing &init_user_ns to fc->user_ns.

The operations in question are iop->flush and fuse_force_forget (on an
error).   I don't know what value having ids on those paths will do
they are operations that must succeed, and they should not change the
on-disk ids.  I was thinking saying the most privileged id was asking
for the oepration would seem to make sense.

With the munged variants we will get (uid16_t)-2 aka 0xfffe aka
nobody asking for the operation if things don't map.  In practice
the don't map case is new.

Since the id's should not be looked at anyway I don't see it makes
much difference which ids we use so the munged case seems at least
plausible.

It might be better to use the non-munghed variant and do:
	if (req->in.h.uid == (uid_t)-1)
		req.in.h.uid = 0;
	if (req->in.h.gid == (gid_t)-1)
        	req.in.h.gid = 0;

That might be less surprising to userspace.  As I don't think the
unmapped case has ever occurred in practice yet.  The vfs will work hard
to keep the unmapped case from happening but only in the context of
i_uid and i_gid not current_fsuid and current_fsgid.

Eric