On Thu, Sep 01, 2022 at 05:17:14PM +0100, Richard W.M. Jones wrote:
[*]It's easier to skip on server failure than to try and write an nbdkit patch to add yet another --config feature probe just to depend on new-enough nbdkit to gracefully probe in advance if the server should succeed.
- /* FIXME: For now, we reject this client-side, but it is overly strict. */ + /* Older servers don't permit this, but there is no reliable indicator + * of whether nbdkit is new enough, so just skip the rest of the test + * if the attempt fails (then read the logs to see that the skip was + * indeed caused by the server, and not an accidental client-side bug). + */
In theory you could parse nbdkit --dump-config, although I agree this approach is fine too.
It only helps to parse --dump-config if --dump-config has an entry that tells us that nbdkit accepts LIST_META_CONTEXT without STRUCTURED_REPLY first (and that still wouldn't help the window of releases that had the feature but not a --dump-config entry, if we decide to add such an entry). But I did think of another way to test it: If we had new APIs: int64_t nbd_stats_opt_packets_sent(handle); int64_t nbd_stats_opt_bytes_sent(handle); int64_t nbd_stats_opt_packets_received(handle); int64_t nbd_stats_opt_bytes_received(handle); int64_t nbd_stats_transmission_packets_sent(handle); int64_t nbd_stats_transmission_bytes_sent(handle); int64_t nbd_stats_transmission_packets_received(handle); int64_t nbd_stats_transmission_bytes_received(handle); that basically count every outgoing packet and byte of NBD_OPT, NBD_REP, NBD_CMD, and response header in either direction, it becomes easy to track when something is squelched client-side by whether the transmission counts increment. And it may be interesting to know how many bytes/packets were involved in the NBD protocol over the life of a connection. -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3266 Virtualization: qemu.org | libvirt.org