Re: [libvirt] [PATCH] loopback is not always just lo

On Monday, 7 November 2016 13:10:06 CET Laine Stump wrote: > On 11/07/2016 04:15 AM, Daniel P. Berrange wrote: > > On Mon, Nov 07, 2016 at 01:52:05AM -0500, Laine Stump wrote: > >> On 11/06/2016 01:54 AM, Roman Bogorodskiy wrote: > >>> Daniel P. Berrange wrote: > >>> > >>>> On Tue, Nov 01, 2016 at 06:15:59PM +0300, Pavel Timofeev wrote: > >>>>> On BSD family OSes (Free/Net/Open/DragonFlyBSD, Mac OS) and > >>>>> Solaris loopback interface is called 'lo0' instead of just 'lo'. > >>>>> > >>>>> --- > >>>>> src/network/bridge_driver.c | 5 +++++ > >>>>> 1 file changed, 5 insertions(+) > >>>>> > >>>>> diff --git a/src/network/bridge_driver.c b/src/network/bridge_driver.c > >>>>> index a3ee3f3..9d94d65 100644 > >>>>> --- a/src/network/bridge_driver.c > >>>>> +++ b/src/network/bridge_driver.c > >>>>> @@ -1002,7 +1002,12 @@ networkDnsmasqConfContents(virNetworkObjPtr network, > >>>>> virBufferAsprintf(&configbuf, "pid-file=%s\n", pidfile); > >>>>> /* dnsmasq will *always* listen on localhost unless told otherwise */ > >>>>> +#ifdef __linux__ > >>>>> virBufferAddLit(&configbuf, "except-interface=lo\n"); > >>>>> +#else > >>>>> + /* BSD family OSes and Solaris call loopback interface as lo0 */ > >>>>> + virBufferAddLit(&configbuf, "except-interface=lo0\n"); > >>>>> +#endif > >>>>> if (dnsmasqCapsGet(caps, DNSMASQ_CAPS_BIND_DYNAMIC)) { > >>>>> /* using --bind-dynamic with only --interface (no > >>>> ACK will push shortly. > >>> This breaks all the networkxml2conftest tests, for example: > >>> > >>> $ VIR_TEST_DEBUG=1 VIR_TEST_RANGE=1 ./tests/networkxml2conftest > >>> TEST: networkxml2conftest > >>> 1) Network XML-2-Conf isolated-network ... > >>> In '/home/novel/code/libvirt/tests/networkxml2confdata/isolated-network.conf': > >>> Offset 303 > >>> Expect [ > >>> ] > >>> Actual [0 > >>> ] > >>> $ > >> On IRC the other day Roman mentioned the idea of just replacing "lo" with > >> "lo*" on all platforms. I checked the dnsmasq documentation and this simple > >> wildcarding is supported for the except-interface option, so I tried out > >> such a patch on my system and it seems to work properly. Maybe we should do > >> that instead of having yet another chunk of #ifdef __linux__ code (the test > >> case datafiles will still need to be changed, but at least they'll be > >> identical for all platforms). > > Any idea how far back wildcards work. If they still work on the RHEL-6 > > vintage dnsmasq, then I'd agree with you and it'll make the test case > > fix easier to avoid the conditional. > > Time for the "Monday Sigh". Sigh. > > I looked this up in dnsmasq's git repo, and it seems recognition of > wildcards in interface names was added in > > commit 49333cbdbef3bf86198b31a42bffad002de305d5 > Author: Simon Kelley <simon(a)thekelleys.org.uk&gt; > Date: Fri Mar 15 20:30:51 2013 +0000 > > Allow trailing '*' wildcard in interface names. > > That commit was first in dnsmasq 2.66, but RHEL6 is using dnsmasq 2.48. > > The strange thing is that I tried running dnsmasq on RHEL6 with > "--except-interface=lo*" and was able to run multiple instances. I guess > my test was flawed. (NetworkManager doesn't support use of dnsmasq for > DNS services, as Cole suggested as a test in: > > https://bugzilla.redhat.com/show_bug.cgi?id=886663#c2 > > But the source doesn't lie.) > > So I guess we need to parameterize the tests somehow (this seems > familiar, but I can't remember why). Another option could be to create two except-interface items in the configuration, "lo" and "lo0", regardless the OS. It seems like dnsmasq supports multiple arguments for --interface and --except-interface for a decade (I went back to v2.32 cca 2006, feel free to double check). If the lo* solution proposed above was deemed safe, just matching two specific interface names instead of everything starting with "lo" should be even more safe. What do you think? Should I go ahead and send the patch implementing my above proposal?