On Thu, Sep 20, 2012 at 05:55:14PM +0300, Orit Wasserman wrote: > On 09/20/2012 04:19 PM, Markus Armbruster wrote: > > Orit Wasserman <owass...@redhat.com> writes: > > > >> Changes from v2: > >> - remove the use of getnameinfo > >> - remove errp for inet_connect_addr > >> - remove QemuOpt "block" > >> - fix errors in wait_for_connect > >> - pass ConnectState as a parameter to allow concurrent connect ops > >> > >> getaddrinfo can give us a list of addresses, but we only try to > >> connect to the first one. If that fails we never proceed to > >> the next one. This is common on desktop setups that often have ipv6 > >> configured but not actually working. > >> A simple way to reproduce the problem is migration: > >> for the destination use -incoming tcp:0:4444, run migrate -d > >> tcp:localhost:4444 > >> migration will fail on hosts that have both IPv4 and IPV6 address for > >> localhost. > >> > >> To fix this, refactor address resolution code and make > >> inet_nonblocking_connect > >> retry connection with a different address. > > > > Almost there for connect. > > > > I'm afraid we have a similar problem with listen: we bind only on the > > first address that works. Shouldn't we bind all of them? > > > > http://www.akkadia.org/drepper/userapi-ipv6.html > > > yes listen should be fixed but lets do it in a separate patch set. > > Orit
One useful hack is to set AI_ADDRCONFIG hint, this way you don't get ipv6 addresses if your system does not have any configured, this results in faster connections on some boxes. http://linux.die.net/man/3/getaddrinfo This probably should be a separate patch too, if for no other reason than that it makes failures in connect much harder to trigger and test. -- MST
