[Lguest] [kvm-devel] [RFC PATCH 0/4] Inter-guest virtio I/O example with lguest

[Avi Kivity]

Anthony Liguori wrote:
> Avi Kivity wrote:
>> Anthony Liguori wrote:
>>> Avi Kivity wrote:
>>>
>>> Each guest's host userspace mmaps the other guest's address space.  
>>> The userspace then does a copy on both the tx and rx paths.
>>>
>>
>> Well, that's better security-wise (I'd still prefer to avoid it, so 
>> we can run each guest under a separate uid), but then we lose 
>> performance wise.
>
> What performance win?  I'm not sure the copies can be eliminated in 
> the case of interguest IO.
>

I guess not.  But at least you can dma instead of busy-copying.

> Fast interguest IO means mmap()'ing the other guest's address space 
> read-only.  

This implies trusting the other userspace, which is not a good thing.  
Let the kernel copy, we already trust it, and it has more resources to 
do the copy.

> If you had a pv dma registration api you could conceivably only allow 
> the active dma entries to be mapped but my fear would be that the 
> zap'ing on unregister would hurt performance.
>

Yes, mmu games are costly.  They also only work on page granularity 
which isn't always possible to guarantee.


>>> Conceivably, this could be done as a read-only mapping so that each 
>>> guest userspace copies only the rx packets.  That's about as secure 
>>> as you're going to get with this approach I think.
>>>
>>
>> Maybe we can terminate the virtio queue in the host kernel as a pipe, 
>> and splice pipes together.
>>
>> That gives us guest-guest and guest-process communications, and if 
>> you use aio the kernel can use a dma engine for the copy.
>
> Ah, so you're looking to use a DMA engine for accelerated copy.  
> Perhaps the answer is to expose the DMA engine via a userspace API?

That's one option, but it still involves sharing all of memory.  
Splicing pipes might be better.

-- 
error compiling committee.c: too many arguments to function