[RFC][PATCH] Devices visibility container

[Dave Hansen]

On Wed, 2007-09-26 at 13:09 -0600, Eric W. Biederman wrote:
> Dave Hansen <haveblue at us.ibm.com> writes:
> 
> > On Tue, 2007-09-25 at 07:30 -0600, Eric W. Biederman wrote:
> >> Pavel Emelyanov <xemul at openvz.org> writes:
> >> >
> >> > Oh! Can you provide us an example when after the migration some
> >> > device's major+minor pair change on the same device?
> >> 
> >> SCSI disks on a SAN.  Network accessible block devices.
> >> All kinds of logical/virtual devices like ttys, the loop device, and
> >> ramdisks.
> >> 
> >> It isn't especially frequent that something cares, but fundamentally
> >> the same issues apply.
> >
> > To be clear, this just covers cases where an application has
> > _internalized_ the device number, right?
> 
> Also cases where you want to call mknod in the container.

mknod of device files only, yeah.

> > Most applications should be pretty happy with the devices having
> > persistent device names across a restart, and we can do that with udev
> > and no kernel patching.
> 
> Yes.  But the applications that do internalize stat data from files
> aren't that uncommon.  git, and backup software etc.
>
> There is also a fair bit of work that is needed to get sysfs
> and the hotplug events isolated, when we start allowing mknod etc.
> 
> Basically if I figure if we are going to deal with this we need to handle
> the entire problem because these pieces are user visible.  I don't
> think it is a great priority.

Exactly.  We have to allow mknod before any of this gets interesting in
the least.  

-- Dave