Re: [Nbd] [RFC 4/4] nbd: Add support for nbd as root device

[Alex Bligh <alex@...872...>]

On 5 Feb 2015, at 08:18, Wouter Verhelst <w@...112...> wrote:

> I had been thinking it might make sense for NBD to have something like
> an "nbd master", in the style of /dev/pts/ptmx; a /dev node whom you ask
> for an NBD device assignment, and which replies with a name, or
> something along those lines.

/dev/tun is another similar (and from what I recall slightly simpler)
precedent.

> The problem, as ever, is backwards compatibility. If taken to its
> logical extreme, this nbdmaster devnode would exist from module
> initialization, but things like /dev/nbd0 etc wouldn't be, only when
> they are registered. That would not be backwards compatible...

OK, how about this for a route to backward compatibility.

Currently the /dev/nbdN devices only appear when the nbd module is
first loaded, via udev or whatever, when the relevant device nodes
are created. As standard, 16 of these are created. Access to these
devices is inherently racy for the reasons described above. Let's
call these 'static' nbd devices.

As a replacement, we make the nbd module create N (where N is
a configurable number on the module command line inititially
defaulting to 16, but may become zero later) such devices which
are static nbd devices. These work as normal. We also let it
create dynamic nbd devices (meaning those mediated by the
/dev/nbd control device). These could be numbered from N up,
though numbering them with a gap might be useful to avoid
naive software relying on the old scheme enumerating from
0 upwards.

Userland software, over time, can be converted to use /dev/nbd
if it exists to safely create nbd devices. If it finds /dev/nbd
is absent, it can fall back to the old route.

Eventually, the default value of N can be reduced to zero,
in which case only the new route will work. This default value
could be a kernel compile option, in which case a distro that
is confident it has refreshed all its nbd software could if
it wished reduce N to zero earlier than the mainline kernel.

-- 
Alex Bligh