Eric Blake authored Mar 01, 2021 and Richard W.M. Jones committed Mar 02, 2021

While debugging some experimental nbdkit code that was triggering an
assertion failure in nbdkit, I noticed a secondary failure of nbdsh
also dying from an assertion:

libnbd: debug: nbdsh: nbd_opt_go: transition: NEWSTYLE.OPT_GO.SEND -> DEAD
libnbd: debug: nbdsh: nbd_opt_go: option queued, ignoring state machine failure
nbdsh: opt.c:86: nbd_unlocked_opt_go: Assertion `nbd_internal_is_state_negotiating (get_next_state (h))' failed.

Although my trigger was from non-production nbdkit code, libnbd should
never die from an assertion failure merely because a server
disappeared at the wrong moment during an incomplete reply to
NBD_OPT_GO or NBD_OPT_INFO.  If this is assigned a CVE, a followup
patch will add mention of it in docs/libnbd-security.pod.

Fixes: bbf1c513 (api: Give aio_opt_go a completion callback)
(cherry picked from commit fb4440de)