Re: [PATCH 1/5] vring: Replace mmap() interface with ioctl()

[Anthony Liguori <aliguori@us.ibm.com>]

Avi Kivity wrote:
> Anthony Liguori wrote:
>>>
>>> And also, because memory hotplug and 64-bit PCI BARs require 
>>> reserving an infinite virtual address space range.  Not to mention 
>>> that someone needs to update the dirty bitmap in case we're live 
>>> migrating.
>>
>> You can certainly hotplug to the next RAM address so it doesn't 
>> require infinite space.  
>
> But you need to reserve that space to prevent mallocs from going 
> there.  How much space do you reserve?

It's deterministic at least.  Any physical system has a maximum amount 
of physical memory that it supports so hotplug cannot exceed that amount.

>> You wouldn't send a packet from/to a PCI IO region so I don't think 
>> that practically speaking that's a problem.
>
> If we implement interguest shared memory as a pci device, then it 
> becomes a problem.

We can set an explicit BAR to keep the physical address reasonable.

I'm not arguing that we should enforce base/limit, just that it is 
possible.  I think the burden is to prove that enforcing base/limit 
provides a significant performance boost to warrant the complexity.

The nastiest bit of manipulating the ring in-kernel is that it requires 
TX mitigation to be within the kernel.  This means that adjusting the 
heuristics for adaptive TX mitigation will require host kernel 
modifications which stinks IMHO.

Regards,

Anthony Liguori