Re: [Qemu-devel] [RFC] libvhost-user: implement VHOST_USER_PROTOCOL_F_KICK_CALL_MSGS

[Michael S. Tsirkin]

On Mon, Sep 09, 2019 at 05:34:13PM +0200, Johannes Berg wrote:
> On Mon, 2019-09-09 at 17:26 +0200, Johannes Berg wrote:
> > 
> > Maybe instead we should just add a "VHOST_USER_REPLY_ERROR" bit (e.g.
> > bit 4 after NEED_REPLY). Qemu in vhost_user_read_header() validates that
> > it received REPLY_MASK | VERSION, so it would reject the message at that
> > point.
> > 
> > Another possibility would be to define the highest bit of the 'request'
> > field to indicate an error, so for GET_FEATURES we'd return the value
> > 0x80000000 | GET_FEATURES.
> 
> However, one way or another, that basically leaves us with three
> different ways of indicating an error:
> 
>  1) already defined errors in existing messages - we can't change them
>     since those are handled at runtime now, e.g. VHOST_USER_POSTCOPY_END
>     returns a u64 value with an error status, and current code cannot
>     deal with an error flag in the 'request' or 'flags' field
>  2) F_REPLY_ACK errors to messages that do not specify a response at all
>  3) this new way of indicating an error back from messages that specify
>     a response, but the response has no inherent way of returning an
>     error
> 
> To me that really feels a bit too complex from the spec POV. But I don't
> see a way to generalize this without another extension, and again the
> device cannot choose which extensions it supports since the master
> chooses them and just sets them.
> 
> Perhaps I really should just stick a "g_assert()" into the code at that
> point,

There's the old way: close the socket.
This will make reads fail gracefully.
If we don't want complexity right now, I'd go with that.


> and have it crash, since it's likely that F_KICK_CALL_MSGS isn't
> even going to be implemented in qemu (unless it grows simulation support
> and then it'd all be conditional on some simulation command-line option)
> 
> 
> 
> And actually ... you got the order wrong:
> 
> > > Next command is GET_FEATURES. Return an error response from that
> > > and device init will fail.
> 
> That's not the case. We *start* with GET_FEATURES, if that includes
> protocol features then we do GET_PROTOCOL_FEATURES next, and then we get
> the # of queues next ...
> 
> Though the whole discussion pretty much applies equivalently to
> GET_QUEUES_NUM instead of GET_FEATURES.
> 
> johannes