On Thu, 18 Oct 2018 at 16:11, Gerd Hoffmann [off-list ref] wrote:

On Thu, Oct 18, 2018 at 11:41:52AM +1000, Dave Airlie wrote:

quoted

On Mon, 1 Oct 2018 at 20:33, Gerd Hoffmann [off-list ref] wrote:

quoted

Remove the virtio_gpu_object_{attach,detach} calls from move_notify()
callback.  Add them to the ttm_tt_{populate,unpopulate} callbacks, which
is the correct place to handle this.

The new ttm_tt_{populate,unpopulate} callbacks call the
ttm_pool_populate()/unpopulate() functions (which are the default
implementation in case the callbacks not present) for the actual ttm
work.  Additionally virtio_gpu_object_{attach,detach} is called to
update the state on the host.

This to me feels more like a bind/unbind operation rather than a
populate/unpopulate operation,

bind is " Bind the backend pages into the aperture in the location"

whereas populate is

allocate pages for a ttm.

I ran into that trap too ;)

My first attempt was to map this to bind/unbind.  But this is not
correct and therefore didn't work very well.

virtio_gpu_object_attach() will send a scatter list of the pages
allocated for the object to the host (so the host knows where to
copy from/to when processing the transfer_from/to calls).  So IMO
it should be done on population not when binding.

Well bind on AGP is the same thing, we'd fill the AGP GART table on
bind, so that the AGP GPU could access the pages.

So I'm interested in why using bind/unbind failed if you have some more info?

Dave.