On Thu, Oct 01, 2015 at 09:41:37AM -0500, Seth Forshee wrote:
> On Thu, Oct 01, 2015 at 09:40:52AM -0400, Mike Snitzer wrote:
> > On Thu, Oct 01 2015 at 8:55am -0400,
> > Seth Forshee <seth.forshee@xxxxxxxxxxxxx> wrote:
> >
> > > On Wed, Sep 30, 2015 at 07:42:15PM -0400, Mike Snitzer wrote:
> > > > On Wed, Sep 30 2015 at 4:15pm -0400,
> > > > Seth Forshee <seth.forshee@xxxxxxxxxxxxx> wrote:
> > > >
> > > > > When mounting a filesystem on a block device there is currently
> > > > > no verification that the user has appropriate access to the
> > > > > device file passed to mount. This has not been an issue so far
> > > > > since the user in question has always been root, but this must
> > > > > be changed before allowing unprivileged users to mount in user
> > > > > namespaces.
> > > > >
> > > > > To fix this, add an argument to lookup_bdev() to specify the
> > > > > required permissions. If the mask of permissions is zero, or
> > > > > if the user has CAP_SYS_ADMIN, the permission check is skipped,
> > > > > otherwise the lookup fails if the user does not have the
> > > > > specified access rights for the inode at the supplied path.
> > > > >
> > > > > Callers associated with mounting are updated to pass permission
> > > > > masks to lookup_bdev() so that these mounts will fail for an
> > > > > unprivileged user who lacks permissions for the block device
> > > > > inode. All other callers pass 0 to maintain their current
> > > > > behaviors.
> > > > >
> > > > > Signed-off-by: Seth Forshee <seth.forshee@xxxxxxxxxxxxx>
> > > > > ---
> > > > > drivers/md/bcache/super.c | 2 +-
> > > > > drivers/md/dm-table.c | 2 +-
> > > > > drivers/mtd/mtdsuper.c | 6 +++++-
> > > > > fs/block_dev.c | 18 +++++++++++++++---
> > > > > fs/quota/quota.c | 2 +-
> > > > > include/linux/fs.h | 2 +-
> > > > > 6 files changed, 24 insertions(+), 8 deletions(-)
> > > > >
> > > > > diff --git a/drivers/md/dm-table.c b/drivers/md/dm-table.c
> > > > > index e76ed003769e..35bb3ea4cbe2 100644
> > > > > --- a/drivers/md/dm-table.c
> > > > > +++ b/drivers/md/dm-table.c
> > > > > @@ -380,7 +380,7 @@ int dm_get_device(struct dm_target *ti, const char *path, fmode_t mode,
> > > > > BUG_ON(!t);
> > > > >
> > > > > /* convert the path to a device */
> > > > > - bdev = lookup_bdev(path);
> > > > > + bdev = lookup_bdev(path, 0);
> > > > > if (IS_ERR(bdev)) {
> > > > > dev = name_to_dev_t(path);
> > > > > if (!dev)
> > > >
> > > > Given dm_get_device() is passed @mode why not have it do something like
> > > > you did in blkdev_get_by_path()? e.g.:
> > >
> > > I only dealt with code related to mounting in this patch since that's
> > > what I'm working on. I have it on my TODO list to consider converting
> > > other callers of lookup_bdev. But if you're sure doing so makes sense
> > > for dm_get_device and that it won't cause regressions then I could add a
> > > patch for it.
> >
> > OK, dm_get_device() is called in DM device activation path (by tools
> > like lvm2).
> >
> > After lookup_bdev() it goes on to call blkdev_get_by_dev() with this
> > call chain:
> > dm_get_device -> dm_get_table_device -> open_table_device -> blkdev_get_by_dev
> >
> > Not immediately clear to me why we'd need to augment blkdev_get_by_dev()
> > to do this checking also.
> >
> > However, thinking further: In a device stack (e.g. dm/lvm2, md, etc)
> > new virtual block devices are created that layer ontop of the
> > traditional block devices. This level of indirection may cause your
> > lookup_bdev() check to go on to succeed (if access constraints were not
> > established on the upper level dm or md device?). I'm just thinking
> > outloud here: but have you verified your changes work as intended on
> > devices created with either lvm2 or mdadm?
> >
> > What layer establishes access rights to historically root-only
> > priviledged block devices? Is it user namespaces?
>
> I'm going to start with this last question and work my way backwards.
>
> Who determines access rights to the block devices varies to some degree.
> Any normal user could unshare the user/mount namespaces and still see
> all the block devices in /dev. If we're going to allow that user to then
> mount block devices in their private namespaces, the kernel must verify
> that the user has appropriate permissions for the block device inode.
> That's the point of this patch. In a container context the host process
> which sets up the container might share some block devices into the
> container via bind mounts, in which case it would be responsible for
> setting up access rights (both via inode permissions and potentially via
> devcgroups). I also have some patches I'm working on for a loop psuedo
> filesystem which would allow a container to allocate loop devices for
> its own use (via the loop-control device), in which case the kernel
> creates a device node in the loopfs filesystem from which the new loop
> device was allocated, owned by the root user in the user namespace
> associated with the loopfs superblock.
>
> Obviously a DM block device is more complicated than a traditional block
> device. This patch should ensure that if an unprivileged user tries to
> mount a DM blkdev that it fails if the user lacks permissions for the DM
> blkdev inode. There's also the blkdevs behind the DM blkdev, and I'm not
> sure whether we need to additionally verify that the user has access to
> those devices. Maybe it's enough that someone with access to them set up
> the DM device, and someone with sufficient privileges gave that user
> access to the DM device. Do you have any thoughts?
>
> As for activating a DM device, I assume only root is permitted to do
> this. In that case lookup_bdev will skip the inode permission check
> even if a non-zero permission mask is passed.
I've been looking into this more and doing some testing.
As far as mounting goes, it ends up behaving as I expected. As long as
the user has permission to access the inode for the DM block device, the
device can be mounted. Access to the block devices behind the DM block
device is not required. I think this makes sense - it's consistent with
how read/write access to the DM block device works already independent
of user namespaces.
One thing I noticed that might be a concern (I'm not sure) are ioctls on
DM block devices. It looks to me like often these will be passed through
to the underlying block device.
It appears that all code paths which call dm_get_device originate with
ioctls on /dev/mapper/control. Doing this requires CAP_SYS_ADMIN, in
which case the permission check in lookup_bdev is skipped. So there's no
benefit to having dm_get_device pass a permission mask to
blkdev_get_by_path, nor is there any harm. It will only matter if at
some point userns-root is allowed to set up DM devices.
Seth
--
To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
