virtio_net occasionally stops sending packets

virtio_net occasionally stops sending packets

[Brian Rak]

We've been running into a fairly persistent issue where virtio_net 
adapters will suddenly stop sending packets when running under KVM.  
This has persisted through several qemu versions, and a large number of 
guest kernel upgrades.

What we end up seeing is the guest continuing to receive packets, but 
refusing to transmit anything.

If we leave ping running for a minute or two, it will eventually start 
printing "ping: sendmsg: No buffer space available" messages.  tcpdump 
will show nothing being sent from the guest. `ip -s link` will show RX 
bytes/packets incrementing, but not TX. `tc -s qdisc show dev eth1` 
shows the 'dropped' counter incrementing with every new ping attempt.

I attempted to run 'dropwatch -l kas', which showed me a bunch of lines 
that looked like '4 drops at pfifo_fast_enqueue+85'.

So far, we haven't been able to consistently reproduce this.  We have a 
few guests that will hit this issue roughly once every two weeks, but we 
haven't been able to reproduce it on demand.  This seems to happen with 
guests that have more then one network adapter attached.  I do not think 
we've seen it on guests that only have one NIC.

We've tried guest kernels as new as 4.14.13, and qemu versions as new as 
2.11.0.  This doesn't appear to be related to the physical network at 
all, we've seen this happen with a variety of network backends:
* qemu 'multicast' networks
* macvtap attached to a vxlan interface
* bridged interface

We've tried disabling a variety of offloads (gso, tso4, tso6, ecn) from 
both the host and guest sides.  This didn't really have any effect.

The only way to fix this once it breaks is to restart the guest OS.  
`ifdown eth1; ifup eth1` doesn't seem to help.

How can I determine if this is a qemu issue, or an issue with the 
virtio_net driver?  We have not tried this with other virtual nic types 
yet.  I'm not sure if that would provide any useful information or not.  
We're still working on figuring out how to reproduce this, but I'm not 
terribly hopeful about coming up with a simple set of reproduction steps.

This was my post about it a few years ago: 
https://lists.nongnu.org/archive/html/qemu-devel/2015-01/msg03907.html

[Qemu-devel] virtio_net occasionally stops sending packets

[Brian Rak]

We've been running into a fairly persistent issue where virtio_net 
adapters will suddenly stop sending packets when running under KVM.  
This has persisted through several qemu versions, and a large number of 
guest kernel upgrades.

What we end up seeing is the guest continuing to receive packets, but 
refusing to transmit anything.

If we leave ping running for a minute or two, it will eventually start 
printing "ping: sendmsg: No buffer space available" messages.  tcpdump 
will show nothing being sent from the guest. `ip -s link` will show RX 
bytes/packets incrementing, but not TX. `tc -s qdisc show dev eth1` 
shows the 'dropped' counter incrementing with every new ping attempt.

I attempted to run 'dropwatch -l kas', which showed me a bunch of lines 
that looked like '4 drops at pfifo_fast_enqueue+85'.

So far, we haven't been able to consistently reproduce this.  We have a 
few guests that will hit this issue roughly once every two weeks, but we 
haven't been able to reproduce it on demand.  This seems to happen with 
guests that have more then one network adapter attached.  I do not think 
we've seen it on guests that only have one NIC.

We've tried guest kernels as new as 4.14.13, and qemu versions as new as 
2.11.0.  This doesn't appear to be related to the physical network at 
all, we've seen this happen with a variety of network backends:
* qemu 'multicast' networks
* macvtap attached to a vxlan interface
* bridged interface

We've tried disabling a variety of offloads (gso, tso4, tso6, ecn) from 
both the host and guest sides.  This didn't really have any effect.

The only way to fix this once it breaks is to restart the guest OS.  
`ifdown eth1; ifup eth1` doesn't seem to help.

How can I determine if this is a qemu issue, or an issue with the 
virtio_net driver?  We have not tried this with other virtual nic types 
yet.  I'm not sure if that would provide any useful information or not.  
We're still working on figuring out how to reproduce this, but I'm not 
terribly hopeful about coming up with a simple set of reproduction steps.

This was my post about it a few years ago: 
https://lists.nongnu.org/archive/html/qemu-devel/2015-01/msg03907.html

Re: virtio_net occasionally stops sending packets

[Brian Rak]

Interestingly, we thought we had just been seeing this on the second 
nic.  However, we've went back through some older logs and realized that 
it's been happening on the first nic as well, though it's presenting 
itself slightly differently there (ping just says "Destination Host 
Unreachable").  I think this is because the VM is unable to resolve it's 
default-route via ARP.

Unfortunately, this machine doesn't have drop monitoring enabled, so I 
can't tell if it's getting dropped in the same spot (it probably is 
though.. tc is still showing the drop counter incrementing)


On 1/19/2018 3:28 PM, Brian Rak wrote:
> We've been running into a fairly persistent issue where virtio_net 
> adapters will suddenly stop sending packets when running under KVM.  
> This has persisted through several qemu versions, and a large number 
> of guest kernel upgrades.
>
> What we end up seeing is the guest continuing to receive packets, but 
> refusing to transmit anything.
>
> If we leave ping running for a minute or two, it will eventually start 
> printing "ping: sendmsg: No buffer space available" messages.  tcpdump 
> will show nothing being sent from the guest. `ip -s link` will show RX 
> bytes/packets incrementing, but not TX. `tc -s qdisc show dev eth1` 
> shows the 'dropped' counter incrementing with every new ping attempt.
>
> I attempted to run 'dropwatch -l kas', which showed me a bunch of 
> lines that looked like '4 drops at pfifo_fast_enqueue+85'.
>
> So far, we haven't been able to consistently reproduce this.  We have 
> a few guests that will hit this issue roughly once every two weeks, 
> but we haven't been able to reproduce it on demand.  This seems to 
> happen with guests that have more then one network adapter attached.  
> I do not think we've seen it on guests that only have one NIC.
>
> We've tried guest kernels as new as 4.14.13, and qemu versions as new 
> as 2.11.0.  This doesn't appear to be related to the physical network 
> at all, we've seen this happen with a variety of network backends:
> * qemu 'multicast' networks
> * macvtap attached to a vxlan interface
> * bridged interface
>
> We've tried disabling a variety of offloads (gso, tso4, tso6, ecn) 
> from both the host and guest sides.  This didn't really have any effect.
>
> The only way to fix this once it breaks is to restart the guest OS.  
> `ifdown eth1; ifup eth1` doesn't seem to help.
>
> How can I determine if this is a qemu issue, or an issue with the 
> virtio_net driver?  We have not tried this with other virtual nic 
> types yet.  I'm not sure if that would provide any useful information 
> or not.  We're still working on figuring out how to reproduce this, 
> but I'm not terribly hopeful about coming up with a simple set of 
> reproduction steps.
>
> This was my post about it a few years ago: 
> https://lists.nongnu.org/archive/html/qemu-devel/2015-01/msg03907.html
>

Re: [Qemu-devel] virtio_net occasionally stops sending packets

[Brian Rak]

Interestingly, we thought we had just been seeing this on the second 
nic.  However, we've went back through some older logs and realized that 
it's been happening on the first nic as well, though it's presenting 
itself slightly differently there (ping just says "Destination Host 
Unreachable").  I think this is because the VM is unable to resolve it's 
default-route via ARP.

Unfortunately, this machine doesn't have drop monitoring enabled, so I 
can't tell if it's getting dropped in the same spot (it probably is 
though.. tc is still showing the drop counter incrementing)


On 1/19/2018 3:28 PM, Brian Rak wrote:
> We've been running into a fairly persistent issue where virtio_net 
> adapters will suddenly stop sending packets when running under KVM.  
> This has persisted through several qemu versions, and a large number 
> of guest kernel upgrades.
>
> What we end up seeing is the guest continuing to receive packets, but 
> refusing to transmit anything.
>
> If we leave ping running for a minute or two, it will eventually start 
> printing "ping: sendmsg: No buffer space available" messages.  tcpdump 
> will show nothing being sent from the guest. `ip -s link` will show RX 
> bytes/packets incrementing, but not TX. `tc -s qdisc show dev eth1` 
> shows the 'dropped' counter incrementing with every new ping attempt.
>
> I attempted to run 'dropwatch -l kas', which showed me a bunch of 
> lines that looked like '4 drops at pfifo_fast_enqueue+85'.
>
> So far, we haven't been able to consistently reproduce this.  We have 
> a few guests that will hit this issue roughly once every two weeks, 
> but we haven't been able to reproduce it on demand.  This seems to 
> happen with guests that have more then one network adapter attached.  
> I do not think we've seen it on guests that only have one NIC.
>
> We've tried guest kernels as new as 4.14.13, and qemu versions as new 
> as 2.11.0.  This doesn't appear to be related to the physical network 
> at all, we've seen this happen with a variety of network backends:
> * qemu 'multicast' networks
> * macvtap attached to a vxlan interface
> * bridged interface
>
> We've tried disabling a variety of offloads (gso, tso4, tso6, ecn) 
> from both the host and guest sides.  This didn't really have any effect.
>
> The only way to fix this once it breaks is to restart the guest OS.  
> `ifdown eth1; ifup eth1` doesn't seem to help.
>
> How can I determine if this is a qemu issue, or an issue with the 
> virtio_net driver?  We have not tried this with other virtual nic 
> types yet.  I'm not sure if that would provide any useful information 
> or not.  We're still working on figuring out how to reproduce this, 
> but I'm not terribly hopeful about coming up with a simple set of 
> reproduction steps.
>
> This was my post about it a few years ago: 
> https://lists.nongnu.org/archive/html/qemu-devel/2015-01/msg03907.html
>