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. Maybe qio_channel_readv/writev_full could have "offset" and "bytes" arguments. Most code in iov_send_recv could be cut-and-pasted. (When sheepdog is converted to QIOChannel, iov_send_recv can go away). Paolo