Re: [libvirt] [PATCH] qemu: enable direct migration over IPv6
Sorry, I only noticed your reply today.
On 02/19/13 17:36, John Ferlan wrote:
On 02/15/2013 05:13 AM, Peter Krempa wrote:
> On 02/15/13 11:00, Ján Tomko wrote:
>> Use virURIParse in qemuMigrationPrepareDirect to allow parsing
>> IPv6 addresses, which would cause an 'incorrect :port' error message
>> before.
>>
>> To be able to migrate over IPv6, QEMU needs to listen on [::] instead
>> of 0.0.0.0. This patch adds a call to getaddrinfo and sets the listen
>> address based on the result.
>>
>> This will break migration if a hostname that can only be resolved on the
>> source machine is passed in the migration URI, or if it does not resolve
>> to the same address family on both sides.
Yuck... "feature-wise" you can pass hints to getaddrinfo() to tell it
what type of address family you want... You can walk the info->ai_next
list looking for the family you want/prefer. Secondary feature-wise
qemu could say which it prefers to use on the "-incoming" value...
In this case, I assume the first family returned by getaddrinfo is the
preferred one. I'm not sure how qemu would have more clue than libvirt
about which address to use.
>> }
>> }
>>
>> + if (getaddrinfo(hostname, NULL, NULL, &info)) {
>> + virReportError(VIR_ERR_INTERNAL_ERROR,
>> + _("unable to get address info for %s"),
>> hostname);
>> + goto cleanup;
>
> Isn't there a possibility to fall back on IPv4 if this fails to minimize
> the chance of regressing?
>
Yeah - this is tricky for some of the reasons I listed above. It seems
we need a way to tell or force the target qemu to use one family style
over the other because that's what the source resolved to using.
Qemu does support ipv4 or ipv6 flags for hostnames. From a quick glance
at the git history it seems it always has. Maybe we could always add the
address family option to the migration URI to force this?
Assuming I'm reading things correctly we are about to tell the target
qemu how to start and to listen over the "first" style of address we
found in the source hosts' list. The source connect will use that
uri_out to perform the migration. The issue I can see becomes what if
the target (for some reason) doesn't support/have/use the family that
the host found first? When it goes to start up a localhost port, it
will fail right? Then what?
No, the perform phase happens on the destination, but the result is
still the same. If the families do not match (or they do match but they
can't connect to each other via that family), migration will fail. Then
the user either has to change the network configuration or specify the
IP address directly.
>> if (*uri_out)
>> VIR_DEBUG("Generated uri_out=%s", *uri_out);
>>
>> - /* QEMU will be started with -incoming tcp:0.0.0.0:port */
>> - snprintf(migrateFrom, sizeof(migrateFrom), "tcp:0.0.0.0:%d",
>> this_port);
>> + /* QEMU will be started with -incoming tcp:0.0.0.0:port
>> + * or -incoming tcp:[::]:port for IPv6 */
>> + if (ipv6) {
>> + snprintf(migrateFrom, sizeof(migrateFrom),
>> + "tcp:[::]:%d", this_port);
I thought IPv6 localhost was "::1"... Or is "::" a synonym? It's
been a
while since I had to think about this and just took a quick google...
It's the IPv6 equivalent of 0.0.0.0, meaning any address.
We don't allow migration on localhost.
>> + } else {
>> + snprintf(migrateFrom, sizeof(migrateFrom),
>> + "tcp:0.0.0.0:%d", this_port);
>
> I thing this would be doable. Just do IPv4 by default if the resolution
> fails.
>
Hmm... which resolution fails do you mean? Perhaps the "feature" needs
to be set/check some field that says the target qemu wants/desires IPv6;
otherwise, always fall back to use IPv4.
If the hostname specified by the user can only be resolved on the
source, migration still works, but erroring out on getaddrinfo failure
would break it.
We can definitely tell qemu to listen on v4/v6 if we received an IP
address. As for hostnames, we can either guess it from how it resolves
on the source, destination or get the input from the user. Maybe we
could add a migration flag for this and add ipv4 or ipv6 option to the
migration URI for qemu based on the presence/absence of this flag. Or
only do IPv6 migration if a URI with a v6 address or tcp6: scheme was
present?
Jan