[libvirt-users] VF MAC not reverted to all zero MAC/domain xml MAC on VM restart

[Laine Stump]

(I should have said this initially - please don't top-post your replies
on this list. Put the answers/responses inline in context with the
questions. It makes it easier to follow the conversation).

On 05/06/2018 05:50 PM, Thirunavukarasu Sengalvarayan -X (tsengalv - HCL
TECHNOLOGIES LIMITED at Cisco) wrote:
> Hi Laine,
> 
>  
> 
> Yes we are setting the names as GE0-0 manually.

Okay, that's explained.

> We have turned trust ON for host IGB driver.

Interesting. I hadn't known that the igb driver supported trust=on - the
last time I tried it, it didn't.

> [root at nfvis libvirt]# ip link show GE0-0
> 
> 3: GE0-0: <BROADCAST,MULTICAST,ALLMULTI,UP,LOWER_UP> mtu 9216 <tel:9216>
> qdisc mq master ovs-system state UP mode DEFAULT qlen 1000 <tel:1000>
> 
> link/ether a0:23:9f:ce:b1:f8 brd ff:ff:ff:ff:ff:ff
> 
> vf 0 MAC *52:54:00:29:3c:be*, spoof checking on, link-state auto, *trust on*
> 
> vf 1 MAC 00:00:00:00:00:00, spoof checking on, link-state auto, *trust on*
> 
> That’s why we are able to change MAC on VF from Guest.

Okay, so now that part makes sense.

> 
> When we are using bonding on the guest VM, we need the host driver(IGB)
> to allow setting the guest mac on VF and to do that we need to enable VF
> trust on.(using the command “ip link set GE0-0 vf 0 trust on”)
> 
> So when VM reboots, as you mentioned guest VF driver is getting reloaded
> and we expect libvirt should revert VF’s MAC to the domain xml MAC and
> that’s not happening.


libvirt has no control over what happens when the guest OS reboots and
reloads the vf net driver. As a matter of fact, libvirt has no control
over what happens after the device is given to qemu for assignment until
the time that the qemu process is *terminated*, and it cleans up all the
devices that had been used by the guest.

When a guest OS reboots, this is just the guest OS reloading, so as far
as libvirt is concerned, the guest is just continuing to use the VF, so
no reason to reload any MAC address.

It appears that the problem is that when you have "trust=on" set for the
VF, then the PF driver in the host sees the new mac set for the VF and
sets the admin mac to that value. So the next time the VF driver is
reloaded, that new admin mac (aka the vf's mac set in the guest) is what
the vf mac gets set to.

In the past (when "trust=on" didn't exist) this wasn't a problem. Maybe
we need to explicitly support setting trust=on, as well as
re-initializing the MAC address when we get the RESET event from qemu;
it seems like this might be something that wouldn't always be desired,
so it may need to be configurable.

If you're using RHEL, can you open a support case asking for this
behavior - this will cause a properly prioritized bugzilla record to be
created and tracked. If you're not using RHEL, then you can either open
an upstream bugzilla (there is information on how to do this somewhere
on libvirt.org), or (even better) you could submit a patch to do it.

In the meantime, I'm fairly certain you can get the behavior you desire
by powering down the guest, then restarting it. This will cause libvirt
to reinitialize the admin mac with the original value set in libvirt's
config.

> 
>  
> 
> Thanks
> 
> Thiru.
> 
>  
> 
> *From:*sendmail <justsendmailnothingelse at gmail.com> *On Behalf Of *Laine
> Stump
> *Sent:* Friday, May 4, 2018 6:48 PM
> *To:* Chanda Mendon (cmendon) <cmendon at cisco.com>
> *Cc:* Thirunavukarasu Sengalvarayan -X (tsengalv - HCL TECHNOLOGIES
> LIMITED at Cisco) <tsengalv at cisco.com>; libvirt-users at redhat.com
> *Subject:* Re: VF MAC not reverted to all zero MAC/domain xml MAC on VM
> restart
> 
>  
> 
> On 05/04/2018 06:25 PM, Thirunavukarasu Sengalvarayan -X (tsengalv - HCL
> TECHNOLOGIES LIMITED at Cisco) wrote:
> 
>     Hi Laine,
> 
>      
> 
>     Thanks for taking the time to respond to my question. I think I have
>     not described my problem clearly.
> 
>      
> 
>     Let me explain my issue below with the information that you had
>     requested.
> 
>      
> 
>     My assumption according to the information you gave me is that the
>     admin MAC and VF MAC are the same in my case.
> 
>     I see a PF (GE0-0) interface but I don’t see a vfnetdev interface as
>     you mentioned in your email.
> 
> 
> If the VF has no separate netdev visible on the host, then it is not
> being bound to the VF net driver for some reason. This should cause no
> ill effect - it just means there is no "original" VF mac to restore.
> 
> 
>      
> 
>     * Given this assumption, when the host is booted, the admin MAC and
>     VF MAC are both*00:00:00:00:00:00. *
> 
> 
> Well, if there is no VF netdev on the host, there is no "vf mac" (and
> the igb driver wouldn't allow it to be 00:00:00:00:00:00). But again,
> not important
> 
> 
> 
>     [root at nfvis ~]# ip link show GE0-0
> 
> 
> "GEO-0"? That's not a normal name for an igb PF. Are you manually
> selecting this name?
> 
> 
>     3: GE0-0: <BROADCAST,MULTICAST,ALLMULTI,UP,LOWER_UP> mtu 9216
>     <tel:9216> qdisc mq master ovs-system state UP mode DEFAULT qlen
>     1000 <tel:1000>
> 
>     link/ether a0:23:9f:ce:b1:f8 brd ff:ff:ff:ff:ff:ff
> 
>     vf 0 MAC 00:00:00:00:00:00, spoof checking on, link-state auto,
>     trust off
> 
>     vf 1 MAC 00:00:00:00:00:00, spoof checking on, link-state auto,
>     trust off
> 
>     [root at nfvis ~]#
> 
>      
> 
>     * the VF is assigned to a guest, so the admin MAC is set to the
>     configured value *(52:54:00:29:3c:bf) *as shown below
> 
>     [root at nfvis ~]# ip link show GE0-0
> 
>     3: GE0-0: <BROADCAST,MULTICAST,ALLMULTI,UP,LOWER_UP> mtu 9216
>     <tel:9216> qdisc mq master ovs-system state UP mode DEFAULT qlen
>     1000 <tel:1000>
> 
>     link/ether a0:23:9f:ce:b1:f8 brd ff:ff:ff:ff:ff:ff
> 
>     vf 0 MAC *52:54:00:29:3c:bf*, spoof checking on, link-state auto,
>     trust on
> 
>     vf 1 MAC 00:00:00:00:00:00, spoof checking on, link-state auto, trust on
> 
>     [root at nfvis ~]#
> 
>      
> 
>     * Configure bond interface on the guest and added member interface
>     to the bond
> 
>        On doing the above step, bond interface mac*(52:54:00:29:3c:be)*
>     gets assigned to VF
> 
>     [root at nfvis libvirt]# ip link show GE0-0
> 
>     3: GE0-0: <BROADCAST,MULTICAST,ALLMULTI,UP,LOWER_UP> mtu 9216
>     <tel:9216> qdisc mq master ovs-system state UP mode DEFAULT qlen
>     1000 <tel:1000>
> 
>     link/ether a0:23:9f:ce:b1:f8 brd ff:ff:ff:ff:ff:ff
> 
>     vf 0 MAC *52:54:00:29:3c:be*, spoof checking on, link-state auto,
>     trust on
> 
>     vf 1 MAC 00:00:00:00:00:00, spoof checking on, link-state auto, trust on
> 
>     [root at nfvis libvirt]#
> 
> 
> Okay, this part makes no sense, for 2 reasons:
> 
> 1) once the admin MAC has been set by the host, a flag is set in the PF
> driver marking this VF as having an "administratively set" MAC (that's
> why I call it the "admin mac"), and after that point it should not be
> possible for a guest to modify the MAC address. If your guest is
> successfully setting the MAC of the VF, then either you don't have a
> device that uses the igb driver, or there is a bug in the driver.
> 
> 2) setting the MAC address of the VF shouldn't update the admin mac for
> that VF - they are separate entities, and only synched up when the VF
> driver is reloaded (and in that case it is the *admin mac* that is used
> as the master copy to set both of them)
> 
> 
> 
>      
> 
>     * The next step would be to restart the VM from within the VM
>     console (* we are not doing a shut down and start from the host *).
> 
> 
> Okay, so that reloads the VF driver in the guest, which would set it to
> the admin mac (which you have said showed to be ....:be just before the
> reset.
> 
> 
>        At this stage, the admin MAC is not reverted to the domain xml
>     mac*(52:54:00:29:3c:bf)*
> 
>        Instead it retains the same bond mac *(52:54:00:29:3c:be)* which
>     was set before the VM was restarted from the console.
> 
>      
> 
>     [root at nfvis libvirt]# ip link show GE0-0
> 
>     3: GE0-0: <BROADCAST,MULTICAST,ALLMULTI,UP,LOWER_UP> mtu 9216
>     <tel:9216> qdisc mq master ovs-system state UP mode DEFAULT qlen
>     1000 <tel:1000>
> 
>     link/ether a0:23:9f:ce:b1:f8 brd ff:ff:ff:ff:ff:ff
> 
>     vf 0 MAC *52:54:00:29:3c:be*, spoof checking on, link-state auto,
>     trust on
> 
>     vf 1 MAC 00:00:00:00:00:00, spoof checking on, link-state auto, trust on
> 
>      
> 
>     So the actual problem we face here, is after the VM is restarted and
>     after the members of bond interface are released, VF mac is not
>     getting reverted back to domain xml mac(*52:54:00:29:3c:bf*).
> 
>     Instead the VF mac and admin mac are set to the bond mac
>     “*52:54:00:29:3c:be*” and leading to a duplicate mac issue.
> 
>     My expectation was that after the VM is restarted from the console,
>     libvirt should revert the VF’s MAC to the original MAC,
> 
>     which in our case was the domain xml MAC, which is not happening in
>     my case.
> 
> 
> My expectation was that the guest would not be allowed to modify the MAC
> address of the VF at all, and certainly that even if a change to the
> guest MAC was allowed, that it wouldn't propagate to the PF's list of
> admin mac addresses for the VFs. So we've both been disappointed :-)
> 
> Your VF and PF drivers are behaving strangely. But if you completely
> shutdown the guest, then restart it, you should get the original
> "...:bf" back. Even if that works for you, I would not count on the
> ability of the guest to modify the VF's MAC address - that is not the
> defined behavior and thus is likely to change in the future.
> 
> 
>      
> 
>      
> 
>     Attached is the log file as requested.
> 
>      
> 
>     *_Interface related contents from domain xml: _*
> 
>     <interface type='hostdev' managed='yes'>
> 
>     <mac address='52:54:00:29:3c:bf'/>
> 
>     <driver name='vfio'/>
> 
>     <source>
> 
>     <address type='pci' domain='0x0000' bus='0x02' slot='0x10'
>     function='0x0'/>
> 
>     </source>
> 
>     <target dev='vnic1'/>
> 
>     <model type='virtio'/>
> 
>     <alias name='hostdev0'/>
> 
>     <address type='pci' domain='0x0000' bus='0x00' slot='0x04'
>     function='0x0'/>
> 
>     </interface>
> 
>      
> 
>     Thanks
> 
>     Thiru.
> 
>     *From:*sendmail <justsendmailnothingelse at gmail.com>
>     <mailto:justsendmailnothingelse at gmail.com> *On Behalf Of *Laine Stump
>     *Sent:* Wednesday, May 2, 2018 10:01 AM
>     *To:* Thirunavukarasu Sengalvarayan -X (tsengalv - HCL TECHNOLOGIES
>     LIMITED at Cisco) <tsengalv at cisco.com> <mailto:tsengalv at cisco.com>
>     *Cc:* Chanda Mendon (cmendon) <cmendon at cisco.com>
>     <mailto:cmendon at cisco.com>
>     *Subject:* Re: VF MAC not reverted to all zero MAC on VM restart
> 
>      
> 
>     On 04/27/2018 11:08 PM, Thirunavukarasu Sengalvarayan -X (tsengalv -
>     HCL TECHNOLOGIES LIMITED at Cisco) wrote:
> 
>         Hi Laine Stump,
> 
>          
> 
>         We are running linux based VM on KVM based Hypervisor and facing
>         issue with respect to VF MAC.
> 
>         Our Libvirt version: 3.2.0, host driver IGB version 5.2.16 and
>         VF driver IGBVF version 2.3.7.1
> 
>          
> 
>         *_Description of problem:_*
> 
>         When passing a VF to a guest, libvirt sets its MAC according to
>         the domain xml.
> 
>         On shutting down the VM or power off the VM, libvirt attempts to
>         restore its original MAC(all-zero mac). There is no issue with
>         this scenario.
> 
>         *When we restart the VM, there is no attempt to restore/revert
>         its original MAC(all zero mac). *
> 
>         This problem leads to duplicate MAC issue on the guest (if
>         configured with portchannel).
> 
>          
> 
>         Could you please help us in resolving the issue?
> 
> 
>     Questions like this should be sent to libvirt-users at redhat.com
>     <mailto:libvirt-users at redhat.com>rather than to an individual's
>     email address. This makes the likelyhood of getting an answer much
>     higher, and also creates an archive of the problem and eventual
>     solution, which may help others in the future. I'm Cc'ing this
>     response to that list so that any further communication will happen
>     there.
> 
>     Before reading any of the following, it's useful to know this - for
>     SRIOV network devices that *properly* support SRIOV (and I consider
>     those using the igb driver to be in this category) each VF has two
>     MAC addresses that need to be discussed:
> 
>     1) the VF netdev MAC address (which I will just call the "mac
>     address" - this is the MAC that is known to the VF net driver, and
>     displayed in the output of "ip link show dev $vfnetdev".
> 
>     2) the MAC address that is stored in the PF driver for each VF,
>     displayed in the "VF" lines immediately following the "ip link" info
>     of the *PF*, and which will be used to initialize the VF's own MAC
>     address when its driver is re-initialized (I will call this the
>     "admin mac")
> 
>     Some VF drivers initialize the mac to 00:00:00:00:00:00 (e.g. the
>     Cisco enic driver) and some initialize it to a random number (e.g.
>     igb and all the other Intel VF net drivers). Likewise, some PF
>     drivers initialize the admin macs to 00:00:00:00:00:00 and some to
>     random numbers (as a matter of fact, this behavior changes between
>     different versions of the same driver in at least one case!).
> 
>     Beyond this, some PF and VF drivers allow setting the mac/admin mac
>     to 00:00:00:00:00:00, and some prohibit one or the other. (in many
>     cases, a driver that itself initializes the mac/admin mac to
>     00:00:00:00:00:00 also prohibits setting it back to that same value
>     once it has been changed!)
> 
>     When libvirt sets up a VF to be used by a guest (either for vfio
>     device assignment, or for macvtap passthrough), it saves both the
>     mac and the admin mac with the intent of restoring them later.
> 
>     Once these values have been saved, libvirt will set the mac (in the
>     case of macvtap passthrough) or the admin mac (in the case of vfio
>     device assignment), and send the device on to qemu to be used by the
>     guest.
> 
>     After the guest is finished running (or the device is detached),
>     libvirt attempts to restore the settings it had saved at the
>     beginning. Trouble comes up in 2 ways though - 1) once the admin mac
>     has been set for a VF (via the PF), it is no longer possible to
>     directly set the mac (in the VF) (this is done for security reasons,
>     to prevent the guest from changing its mac address), and 2) as said
>     above, many drivers don't allow setting 00:00:00:00:00:00, but in
>     many cases that was the original setting.
> 
>     As of libvirt 3.2.0, libvirt uses the following strategies to work
>     around these problems:
> 
>     a) if setting a non-0 mac (to the VF) fails, then libvirt will try
>     setting the *admin MAC* to that value, then detaching/re-attaching
>     the VF net driver. This will cause the PF to reinitialize the mac in
>     the VF driver. (NB: if we do this, then we don't bother trying to
>     later re-set the admin MAC to its own original value; it's really
>     not important).
> 
>     b) if the failure was in setting the MAC to 00:00:00:00:00:00, then
>     libvirt will set the mac/admin mac to 02:00:00:00:00:00 (which
>     should be a legal value for any driver, but also should not conflict
>     with and "real" mac on the network).
> 
>     Okay, enough background. On to your problem.
> 
> 
>     To make sure I understand your situation correctly:
> 
>     * When the host is booted, the admin MAC is 00:00:00:00:00:00, and
>     VF mac is random (call it rr:rr:rr:rr:rr:rr).
> 
>     * the VF is assigned to a guest (with <interface type='hostdev'> I
>     guess?), so the admin MAC is set to the configured value (let's call
>     it gg:gg:gg:gg:gg:gg)
> 
>     * the guest is shutdown, which causes *both* mac and admin mac to be
>     set to rr:rr:rr:rr:rr:rr
> 
>     * the next time the guest is start *its mac is not changed*?
> 
>     There must be something I'm not getting. Can you maybe perform this
>     test and run "ip link show $PF; ip link show $VF" (substituting the
>     netdev names of the PF and VF for $PF and $VF) after each step to
>     illustrate the problem? Also, it would be helpful to add this to
>     /etc/libvirt/libvirtd.conf:
> 
>       log_filters="1:util.netdev"
>       log_outputs="1:file:/var/log/libvirt/libvirtd-netdev.log
>     <file://var/log/libvirt/libvirtd-netdev.log>"
> 
>     then restart libvirtd (systemctl restart libvirtd.service), and post
>     the contents of libvirtd-netdev.log somewhere and reference it in
>     your next reply. (don't forget to remove those lines and restart
>     libvirtd again after your testing is done!)
> 
>     Finally, you should include the contents of the <interface ...>
>     section of your config in your next response.
> 
>     (BTW, since you're using libvirt-3.2.0, I assume that you are using
>     either RHEL 7.4 or CentOS 7.4. If you are using RHEL, you should
>     open a customer case with Red Hat so that your problem is
>     appropriately prioritized).
> 
>  
> 
> 
> 
> _______________________________________________
> libvirt-users mailing list
> libvirt-users at redhat.com
> https://www.redhat.com/mailman/listinfo/libvirt-users
>