Re: [PATCH 0/16 v6] PCI: Linux kernel SR-IOV support

[Anthony Liguori <anthony@codemonkey.ws>]

Andi Kleen wrote:
> Anthony Liguori <anthony@codemonkey.ws> writes:
>> What we would rather do in KVM, is have the VFs appear in the host as
>> standard network devices.  We would then like to back our existing PV
>> driver to this VF directly bypassing the host networking stack.  A key
>> feature here is being able to fill the VF's receive queue with guest
>> memory instead of host kernel memory so that you can get zero-copy
>> receive traffic.  This will perform just as well as doing passthrough
>> (at least) and avoid all that ugliness of dealing with SR-IOV in the
>> guest.
> 
> But you shift a lot of ugliness into the host network stack again.
> Not sure that is a good trade off.
> 
> Also it would always require context switches and I believe one
> of the reasons for the PV/VF model is very low latency IO and having
> heavyweight switches to the host and back would be against that.

I don't think it's established that PV/VF will have less latency than 
using virtio-net.  virtio-net requires a world switch to send a group of 
packets.  The cost of this (if it stays in kernel) is only a few 
thousand cycles on the most modern processors.

Using VT-d means that for every DMA fetch that misses in the IOTLB, you 
potentially have to do four memory fetches to main memory.  There will 
be additional packet latency using VT-d compared to native, it's just 
not known how much at this time.

Regards,

Anthony Liguori


> -Andi
>