Re: [PATCH RFC] vhost: basic device IOTLB support

[Jason Wang <jasowang@redhat.com>]

On 01/04/2016 09:39 AM, Yang Zhang wrote:
> On 2015/12/31 15:13, Jason Wang wrote:
>> This patch tries to implement an device IOTLB for vhost. This could be
>> used with for co-operation with userspace(qemu) implementation of
>> iommu for a secure DMA environment in guest.
>>
>> The idea is simple. When vhost meets an IOTLB miss, it will request
>> the assistance of userspace to do the translation, this is done
>> through:
>>
>> - Fill the translation request in a preset userspace address (This
>>    address is set through ioctl VHOST_SET_IOTLB_REQUEST_ENTRY).
>> - Notify userspace through eventfd (This eventfd was set through ioctl
>>    VHOST_SET_IOTLB_FD).
>>
>> When userspace finishes the translation, it will update the vhost
>> IOTLB through VHOST_UPDATE_IOTLB ioctl. Userspace is also in charge of
>> snooping the IOTLB invalidation of IOMMU IOTLB and use
>> VHOST_UPDATE_IOTLB to invalidate the possible entry in vhost.
>
> Is there any performance data shows the difference with IOTLB supporting?

Basic testing show it was slower than without IOTLB.

> I doubt we may see performance decrease since the flush code path is
> longer than before.
>

Yes, it also depend on the TLB hit rate.

If lots of dynamic mappings and unmappings are used in guest (e.g normal
Linux driver). This method should be much more slower since:

- lots of invalidation and its path is slow.
- the hit rate is low and the high price of userspace assisted address
translation.
- limitation of userspace IOMMU/IOTLB implementation (qemu's vtd
emulation simply empty all entries when it's full).

Another method is to implement kernel IOMMU (e.g vtd). But I'm not sure
vhost is the best place to do this, since vhost should be architecture
independent. Maybe we'd better to do it in kvm or have a pv IOMMU
implementation in vhost.

Another side, if fixed mappings were used in guest, (e.g dpdk in guest).
We have the possibility to have 100% hit rate with almost no
invalidation, the performance penalty should be ignorable, this should
be the main use case for this patch.

The patch is just a prototype for discussion. Any other ideas are welcomed.

Thanks