On 3/25/19 11:21 AM, Eric Blake wrote: > When you fix the first bug (the client setting iter.err on a simple > error reply from the server because the server's reply wasn't > structured, to now the client setting just iter.request_ret because it > successfully parsed an error out of the server's reply), this code is > now reachable with iter.err == NULL. Since the server replied with an > error, extent->length is 0. So this code (added in 7f86068d) says that > since iter.err is not set, we give up on the server and hang up (but > only for simple errors, not for structured errors). It is a regression, > because prior to 7f86068d, we did not have iter.request_ret, and instead > went by the presence or absence of iter.fatal, and iter.fatal was not > set when dealing with a simple error when a structured error was supplied. > > So, the full setup of things to test: > > server that always fails with structured error: > - client before 7f86068d (qemu 3.1): connection stays alive, regardless > of this patch > - client after 7f86068d: connection stays alive, regardless of this patch > > server that always fails with simple error: > - client before 7f86068d: (second half of patch does not apply, so only > the first half is needed): > - without first half, connection dies because "Protocol error: simple > reply when structured reply chunk was expected"
What's really annoying is that iter.err does not show up anywhere except maybe in trace messages; on stderr, the symptoms are merely: Failed to get allocation status: Invalid argument because nbd_co_do_receive_one_chunk returns -EINVAL for iter.ret. > - with first half, connection stays alive as it does for structured error > - client after 7f86068d, with various stages of this patch: > - without either half (or with second half only), connection dies > because "Protocol error: simple reply when structured reply chunk was > expected" again, shows up only as Failed to get allocation status: Invalid argument > - with first half, connection dies because "server did not reply with > any status extents" Here, shows up only as Failed to get allocation status: Input/output error because without the second hunk, iter.ret is set to -EIO. > - with both halves, connection stays alive > -- Eric Blake, Principal Software Engineer Red Hat, Inc. +1-919-301-3226 Virtualization: qemu.org | libvirt.org
signature.asc
Description: OpenPGP digital signature
