Adding libnbd list (libguestfs) in cc On Fri, Jan 28, 2022 at 11:56:19AM +0000, Richard W.M. Jones wrote: And only after I already replied to your other email, did I then see your followup recommending to read this one instead ;) The NBD spec says the client is non-complying when sending under-sized requests. If the server accepts it anyway (presumably with RMW performance pessimizations, as qemu-nbd does), that's a QoI bonus. Yeah, that might make sense - a command-line option for stating "I know the server has a nice QoI feature, and I don't mind the performance pessimization". Another thing to consider: the way the NBD spec is written, the rules about a client sending unaligned requests being non-compliant only applies to a client that requested block size information in the first place. If the client did not request block alignment information, the server should honor anything at alignment of 512 or above (even if it would prefer a larger minimum); performance may suffer, but this is needed to cater to older clients that don't know how to request alignments - and what's more, qemu-nbd specifically has code that changes what it advertises if the client did not query (that is, an advertisement of 64k is ONLY possible if the client requested alignment details). So maybe the question becomes whether libnbd needs a knob on whether to request alignment information. Libnbd commit 9e9c74755 (libnbd v1.3.10) is where I added the code to unconditionally query for alignment info. Given that we now know of a case where NOT querying causes qemu-nbd to behave differently in what it advertises, adding such a knob to the API makes total sense, at which point, 'nbdcopy --ignore-align' becomes a way to request that the client not request alignment in the first place, rather than as a way to call nbd_set_strict_mode() to turn off alignment checking. Looks like I have some API work do propose in libnbd... -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3266 Virtualization: qemu.org | libvirt.org