On 26/09/2017 00:19, Eric Blake wrote:
+ /* here we deal with successful structured reply */
+ switch (s->reply.type) {
+ QEMUIOVector sub_qiov;
+ case NBD_SREP_TYPE_OFFSET_DATA:
This is putting a LOT of smarts directly into the receive routine.
Here's where I was previously wondering (and I think Paolo as well)
whether it might be better to split the efforts: the generic function
reads off the chunk information and any payload, but a per-command
callback function then parses the chunks. Especially since the
definition of the chunks differs on a per-command basis (yes, the NBD
spec will try to not reuse an SREP chunk type across multiple commands
unless the semantics are similar, but that's a bit more fragile). This
particularly matters given my statement above that you want a
discriminated union, rather than a struct that contains unused fields,
for handling different SREP chunk types.
I think there should be two kinds of replies: 1) read directly into a
QEMUIOVector, using structured replies only as an encapsulation of the
payload; 2) read a chunk at a time into malloc-ed memory, yielding back
to the calling coroutine after receiving one complete chunk.
In the end this probably means that you have a read_chunk_header
function and a read_chunk function. READ has a loop that calls
read_chunk_header followed by direct reading into the QEMUIOVector,
while everyone else calls read_chunk.