On Wed 29-04-20 12:42:01, Dave Chinner wrote:
> On Tue, Apr 28, 2020 at 06:48:24PM +0200, Jan Kara wrote:
> > On Tue 28-04-20 08:53:51, Darrick J. Wong wrote:
> > > On Tue, Apr 28, 2020 at 05:32:28PM +0200, Jan Kara wrote:
> > > > > dd if=/dev/zero of=someoutput oflag=append
> > > > > dd: failed to open 'someoutput': Invalid argument
> > > >
> > > > Yes, that's a bug that should be fixed. Thanks for reporting this! -1 means
> > > > 'this id is not expressible in current user namespace' and some code gets
> > > > confused along the way. We should refuse to set project -1 for a file...
> > >
> > > Awkward part: projid 4294967295 is allowed on XFS (at least by the
> > > kernel), though the xfs quota tools do not permit that.
> >
> > Are you OK with just refusing to set projid 4294967295 for everybody? Or
> > should we just not try to translate project IDs through user namespaces?
> > Because XFS does not seem to translate them while ext4 does... What a mess.
>
> We do not translate project IDs through user names space because
> they are not usable as a mappable id. Project IDs are only used for
> customised aggregation of space accounting, unlike UIDs and GIDS
> that are used primarily for access control. IOWs, PRIDs are
> fundamentally different to UIDs and GIDs.
>
> Project IDs were already being used in the init namespace for
> directory quotas to limit containers using bind mounts on a host
> filesystem to an amount of disk space less than the entire hosting
> filesystem. And once you use PRIDs in the init namespace, they
> cannot be used by users in other user namespaces, regardless of
> whether they are mappable or not.
OK, understood.
> Essentially, the project ID mapping stuff was implemented by someone
> who didn't understand what project IDs were or how project IDs were
> being used, and then refused to listen to the people who knew these
> things and wanted them to drop the PRID mapping stuff. And then
> Linus pulled their tree containing all the uid/gid/prid mapping code
> without warning and we've been stuck with this shit ever since.
>
> Hence in XFS we simply do not allow project IDs to be manipulated
> outside of the init user namespace, and so mapping them is
> irrelevant because users in confined namespaces cannot usefully
> interact with them in any way.
So in ext4 we also don't currently allow anybody outside init user
namespace to change project IDs. Also as I'm now checking the projid
handling in ext4 more closely, we always transform project ID only to/from
init_user_ns (even in FSGETXATTR ioctl) so it's more or less pointless and
equivalent to XFS not transforming anything AFAIU.
So the only problem is really with VFS quota code. There we do mapping of
passed project ID from current_user_ns() in fs/quota/quota.c before passing
the ID further to the core quota code. Practically, this is only relevant
for GETQUOTA quotactl calls because all the others are restricted to
init_user_ns capable CAP_SYS_ADMIN so they can get called only from
init_user_ns.
Now we also have a check like:
/* Filesystems outside of init_user_ns not yet supported */
if (sb->s_user_ns != &init_user_ns) {
error = -EINVAL;
goto out_fmt;
}
in dquot_load_quota_sb() which is the quota enabling function. So we don't
allow any quotas for filesystems outside of init_user_ns. So the
qid_has_mapping() checks are mostly pointless as sb->s_user_ns is always
init_user_ns. But this is except for id -1, which doesn't have mapping even
in init_user_ns...
So I'm pondering what's the best way out of this mess. Currently, the
mapping of project IDs in quota code has rather limited impact and we may
be able to get away with just removing it (i.e. without causing a
regression for any real user). So that's certainly one option. But then we
should probably also remove the capability to specify (non-trivial) project
ID maps for user namespaces because having maps that are not actually
applied is pretty confusing.
Then there's a second option: Is there a reason *not* to map project IDs
in user namespaces? I understand it's pointless with how project ids are
currently used but it does not harm either AFAIU. The only real harm is
with id -1 not being usable. Also when people create fs mount option where
project ID is changeable by CAP_SYS_ADMIN (or maybe CAP_SYS_RESOURCE)
capable user - and there are several people asking for a functionality like
this - then fully mapping project IDs would IMHO make more sence.
Honza
--
Jan Kara <jack@xxxxxxxx>
SUSE Labs, CR
