Re: [Qemu-devel] Re: [RFC 1/1] pci-dma-api-v2

[Anthony Liguori <anthony@codemonkey.ws>]

Blue Swirl wrote:
> On 11/30/08, Andrea Arcangeli <aarcange@redhat.com> wrote:
>   
>> On Sun, Nov 30, 2008 at 09:11:28PM +0200, Blue Swirl wrote:
>>  > Yes, this version (attached) works.
>>
>>
>> Ok great! ;)
>>     
>
> This looks promising. Maybe the previous attempts tried to solve
> several problems at once, like the complex address translation schemes
> I have in mind, in addition to the basic zero-copy DMA transfer.This
> version leaves the DMA address translation to the caller and operates
> using physical memory addresses. Another way would be to use directly
> host pointers.
>   

I'm happy to separate address translation from just zero-copy IO.  I 
agree that a lot of the complexity is introduced trying to solve everything.

For me, just extending the BlockDriver API is a problem.  I'd like to 
see the API properly structured around a single vector API.  This is 
independent of zero copy.

I also think that in the very least, a DMA must be encapsulated in a 
structure, that has functions to pin()/unpin() that generate bounce 
buffers as necessary.

I think it's probably a good idea to separate a basic DMA api from a PCI 
DMA api.

Regards,

Anthony Liguori