Re: [Xen-devel] [PATCH RFC 06/12] libxl: set correct permissions for the full backend path

[Ian Campbell <Ian.Campbell@xxxxxxxxxx>]

On Thu, 2013-09-26 at 11:20 +0200, Roger Pau Monnà wrote:
> On 25/09/13 17:18, Ian Campbell wrote:
> > On Wed, 2013-09-25 at 17:10 +0200, Roger Pau Monnà wrote:
> >> On 25/09/13 17:00, Ian Campbell wrote:
> >>> On Wed, 2013-09-25 at 16:02 +0200, Roger Pau Monnà wrote:
> >>>> On 25/09/13 15:57, Ian Campbell wrote:
> >>>>> On Mon, 2013-09-23 at 12:30 +0200, Roger Pau Monne wrote:
> >>>>>> The backend path should be fully owned by the domain where it resides.
> >>>>>
> >>>>> I can see why /local/domain/<domid>backends/<type>/<id> would need to be
> >>>>> owned by the backend dom, but why
> >>>>> do /local/domain/<domid>backends/<type>/, 
> >>>>> /local/domain/<domid>backends/, etc need to be?
> >>>>
> >>>> The path /local/domain/<domid>backends/<type>/<guest_domid>/<id> is
> >>>> already owned by the driver domain, the problem is that if the driver
> >>>> domain has to perform the clean up of this path it should be able to
> >>>> fully remove it, otherwise we are leaving empty directories around in
> >>>> xenstore (backend/<type>/<guest_domid> and so).
> >>>>
> >>>> And performing the clean up from the toolstack domain is not that easy,
> >>>> we will have to add a way to signal the toolstack domain that the driver
> >>>> domain has finished the disconnection and the directory can be cleaned.
> >>>
> >>> Don't we need that anyway so that the toolstack domain can synchronise
> >>> the end of the asynchronous op which is the remove?
> >>
> >> Not really IMHO, the toolstack waits for the device frontend to
> >> disconnect and then it removes the frontend xenstore entries. Then (or
> >> in parallel) the domain that's handling the backend should deal with the
> >> disconnection of the backend and perform any necessary actions and clean
> >> up the backend xenstore entries.
> > 
> > So e.g. xl destroy can potentially return while the backend is still
> > closing down? That would mean it might fail to reboot the guest, because
> > the block device in the driver domain is still busy.
> 
> True, since we don't synchronize with the toolstack there's a small
> window were a new domain could be created before the driver domain
> hasn't finished unplugging. This is a problem already with current udev
> implementation, but I think we can solve this by adding a watch in the
> toolstack that waits for the backend path to be destroyed (when the
> backend is on a driver domain, of course).
> 
> I will expand this patch to include the above mentioned check.

Please do.

Can you also enumerate the intended permissions for each bit
of /local/domain/<domid>/backends/<type>/<guest_domid>/<id>

I think you are changing the perms of "backends" onwards, but is that
really necessary? The backend should be cleaning up <id>, which needs
write perms for <guest_domid> but could the toolstack not do the rest?

Ian.


_______________________________________________
Xen-devel mailing list
Xen-devel@xxxxxxxxxxxxx
http://lists.xen.org/xen-devel