On 10/10/2011 03:00 PM, Folkert van Heusden wrote:
In practice, I'd try to avoid reads that are more than 1 MB minus the reply header size, and writes that are more than 1 MB minus the request header size. qemu-nbd used to choke on those, though the next release will only choke on more than 1 MB. A good way to see what kind of request the kernel might make is to use dd with the O_DIRECT flag.That is too tricky: a new kernel version might show differnt behaviour. Imho we should specify a limit or make it configurable or put it in the protocol. We have 124 bytes of handshake left, we can use 8 of them for a maximum blocksize.
O_DIRECT bypasses everything clever that the kernel might do. Paolo