On 8/24/20 6:28 PM, Daniel Henrique Barboza wrote:
On 8/23/20 1:52 AM, Laine Stump wrote:
Back when macvtap support was added in commit 315baab9443 in Feb. 2010
(libvirt-0.7.7), it was setup to autogenerate a name for the device if
one wasn't supplied, in the pattern "macvtap%d" (or "macvlan%d"),
similar to the way an unspecified standard tap device name will lead
to an autogenerated "vnet%d".
As a matter of fact, in commit ca1b7cc8e45 added in May 2010, the code
was changed to *always* ignore a supplied device name for macvtap
interfaces by deleting any name immediately during the <interface>
parsing (this was intended to prevent one domain which had failed to
completely start from deleting the macvtap device of another domain
which had subsequently been provided the same device name (this will
seem mildly ironic later). This was later fixed to only clear the
device name when inactive XML was being parsed. HOWEVER - this was
only done if the xml was <interface type='direct'> - autogenerated
names were not cleared for <interface type='network'> (which could
also result in a macvtap device).
Although the names of "vnetX" tap devices had always been
automatically cleared when parsing <interface> (see commit d1304583d
from July 2008 (!)), at the time macvtap support was added, both vnetX
and macvtapX device names were always included when formatting the
XML.
Then in commit a8be259d0cc (July 2011, libvirt-0.9.4), <interface>
formatting was changed to also clear out "vnetX" device names during
XML formatting as well. However the same treatment wasn't given to
"macvtapX".
The result of all this (among other things) was that when a running guest
I guess this is a stray sentence ^
You are guessing correctly.
My ferret-like brain is easily distracted in the middle of a thought,
and sometimes when I pop my stack back to where I was before, I forget
to clean up some loose ends. In this case, the above fragment just needs
to be removed.
Now in 2020, there has been a report that a failed migration leads to
the macvtap device of some other unrelated guest on the destination
host losing its network connectivity. It was determined that this was
due to the domain XML in the migration containing a macvtap device
name, e.g. "macvtap0", that was already in use on the
destination. Normally this wouldn't be a problem, because libvirt
would see that the device was already in use, and then find a
different unused name. But in this case, other external problems were
causing the migration to fail prior to selecting a macvtap device and
successfully opening it, and during error recovery, qemuProcessStop()
was called, which went through all def->nets objects and (if they were
macvtap) deleted the device specified in net->ifname; since libvirt
hadn't gotten to the point of replacing the incoming "macvtap0" with
the name of a device it actually created for this guest, that meant
that "macvtap0" was deleted, *even though it was currently in use by a
different guest*!
Whew!
So, it turns out that when formatting "migratable" XML, "vnetX"
devices are omitted, just as when formatting "inactive" XML (I'm sure
there's a bit of code somewhere that says "if (migratable) then set
inactive too", but found it was easier to just try it out with "virsh
dumpxml blah --migratable"). By making the code in both interface
parsing and formatting consistent for "vnetX", "macvtapX", and
"macvlanX", we can thus make sure that the autogenerated (and unneeded
/ completely *not* wanted) macvtap device name will not be sent with
the migration XML. This way when a migration fails, net->ifname will
be NULL, and libvirt won't have any device to try and delete.
Signed-off-by: Laine Stump <la...@redhat.com>
---
Cool story and good patch.
Well, I live for cool stories, after all :-) (seriously, I don't know if
anyone else so frequently has commit log messages longer than the actual
patch. My wife would say I'm confusing the situation by giving too much
information; I say I'm being thorough :-))
