Re: [PATCH RFC 1/3] Drivers: hv: kvp: convert userspace/kernel communication to using char device

["Radim Krčmář" <rkrcmar-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org>]

2015-03-03 10:53+0100, Vitaly Kuznetsov:
> Radim Krčmář <rkrcmar-H+wXaHxf7aLQT0dZR+AlfA@public.gmane.org> writes:
> > 2015-02-27 17:14+0100, Vitaly Kuznetsov:
> >> Re-implement the communication using misc char device. Use ioctl to do
> >> kernel/userspace version negotiation (doesn't make much sense at this moment
> >> as we're breaking backwards compatibility but can be used in future).
> >
> > The ioctl is used too creatively for my liking: as an out-of-band
> > communication that is required after the main channel has been opened.
> > It would be simpler to inject the version into first x bytes of the
> > stream, making a read() after open() mandatory.
> 
> We need to perform a handshake - kernel part sends its version and
> receives daemon's version.

(We could also design a backward-compatible protocol, which should be
 possible for an application this simple, but keeping options is wise.)

>                            We can definitelly pack everything in the
> data stream but why do we need to avoid ioctls?

I think it is better if the kernel sends its version (set of features)
first, so it would be just a simple two-way handshake.

Kernel-initiated communication is not possible over ioctl and it doesn't
give extra options for handshake either.

>                                                 It seems to me the
> handshake we're performing here belongs to a 'control' stream, not
> 'data' stream.

Handshake makes sense after we open the device and before any 'data' can
appear on it, so multiplexing the same carrier also prevents a lot of
stupid cases, like ioctls from different applications.
Not to mention that asynchronicity itself has a fairly bad record.

(I don't really understand the difference between 'control' and 'data'.)