On Tue, Oct 27, 2009 at 05:46:51PM +0200, Pasi Kärkkäinen wrote:

I am not really comfortable with it. It has the same drawbacks as the fix for the drm_scatter, where we blindly assume phys_to_bus(virt_to_phys(X)) will give us the same value as what dma_alloc_coherent provides. We should save that bus address somewhere...

Saving it somewhere (perhaps in some of the structs the drm_ttm allocates) could do it. But we should probably differentiate between memory that is being allocated for DMA transfers vs other things so that we don't over-exercise the dma_alloc_coherent. Thought maybe the memory returned via drm_tt calls are only used for DMA transfers.

We can figure this out. Pasi, I don't have a modesetting working machine, but you do. Can you compile your pv_ops with the fix I provided earlier, along with enabling CONFIG_DMA_API_DEBUG=y. Once mode-setting is turned on and your machine is humming along (maybe even run glxgears), compile the attached module and load it. You should get a kernel dump off all devices that are using the DMA buffers. Can you e-mail me that back please?