On Tue, Feb 25, 2014 at 11:45:11AM +0000, Nick Thomas wrote: > On 25/02/14 10:09, Stefan Hajnoczi wrote: > > + def send(self, buf, event): > > + self.check(event, 'write', 'before') > > + self.sock.sendall(buf) > > + self.check(event, 'write', 'after') > > + > > + def recv(self, bufsize, event): > > + self.check(event, 'read', 'before') > > + data = recvall(self.sock, bufsize) > > + self.check(event, 'read', 'after') > > + return data > > There's a class of error I recently encountered in our out-of-tree proxy > component that only shows up if a read or write is interrupted partway > through. Perhaps you could have a "during" event here that happens after > bufsize/2 bytes is written? > > I've not looked at qemu's block/nbd code recently, so I don't know if > that exercises a particular failure path.
Yes, it can involve different code paths in the client. For example, the client may receive fields in separate recv(2) syscalls so there may be a unique path for each field. I think the easiest approach is to turn the 'when' option into a byte count. The connection will be terminated after the given number of bytes. Then 'before' becomes an alias for 0. 'after' becomes an alias for -1, the magic value for the entire data length. > > + def close(self): > > + self.sock.close() > > + > > +def negotiate(conn): > > + '''Negotiate export with client''' > > + # Send negotiation part 1 > > + buf = neg1_struct.pack(NBD_PASSWD, NBD_OPTS_MAGIC, 0) > > + conn.send(buf, event='neg1') > > + > > + # Receive export option > > + buf = conn.recv(export_struct.size, event='export') > > + export = export_tuple._make(export_struct.unpack(buf)) > > + assert export.magic == NBD_OPTS_MAGIC > > + assert export.opt == NBD_OPT_EXPORT_NAME > > + name = conn.recv(export.len, event='export-name') > > + > > + # Send negotiation part 2 > > + buf = neg2_struct.pack(8 * 1024 * 1024 * 1024, 0) # 8 GB capacity > > + conn.send(buf, event='neg2') > > Is it worth exercising the old-style negotiation too? Yes, probably a good idea.
