TTM and AGP conflicts

[Jerome Glisse]

On Mon, Jan 09, 2012 at 10:07:02AM +0100, Thomas Hellstrom wrote:
> On 01/06/2012 04:51 PM, James Simmons wrote:
> >>>>You can achieve what you want by either adding a new domain so you would have
> >>>>system, vram, agp, pcidma and object can be bound to one and only one. Or you
> >>>>can hack your own agp ttm backend that could bind bo to agp or pci or
> >>>>both at the same time depending on what you want to achieve.
> >>>The question is how does one know which domain you want in tt_create.
> >>>Currenty drivers are using there dev_priv but if you have have more than
> >>>one option available how does one choose? Would you be okay with passing
> >>>in a domain flag?
> >>>
> >>Well i agree that something would be usefull there so the driver know
> >>which bind/unbind function it should use. Thomas i would prefer
> >>passing the bo to the tt_create callback but a flag is the minimum we
> >>need.
> >We can discuss this after the merge widow. Jerome your patch does fix a
> >regression whereas my proposal is a enhancement.
> 
> ..Back from parental leave.
> 
> I'm not 100% sure I understand the problem correctly, but I assume
> the problem is that
> you receive a "bind" request for pages allocated out of the wrong
> DMA pool? Is that correct?
> 
> In that case we need to make both the following happen:
> 1) The backend should be required to supply a fallback that can bind
> anyway by allocating the correct page type and copy.
> 2) I'm OK with providing a hint to tt_create. IIRC we're passing
> bo::mem to tt_bind. Would it be ok to pass the same to tt_create?
> 
> /Thomas
> 

Issue here is that different backend (AGP or PCI DMA or someother
communication way btw GPU and system memory) needs to use different
page allocator/backend. Thus tt_create need either to know about the
bo or at least about bo:mem.

Cheers,
Jerome