Re: Feature Request: Ability to decode bus/dma address back into physical address

[Jerome Glisse <j.glisse-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>]

On Wed, Aug 02, 2017 at 10:26:40AM +0200, Christian König wrote:
> Hi Jerome,
> 
> sorry for being a bit late to the discussion and the top posting.
> 
> But I think you miss a very important point here, which makes the whole
> discussion on how to implement completely superfluous:
> 
> We already have a functionality to access the content of BOs in a process
> for debugging purpose which works similar as you described and just uses the
> BO housekeeping structures in the driver to access the pages and VRAM
> locations in question.
> 
> See here for the implementation:
> 1. drm/ttm: Implement vm_operations_struct.access v2
> (http://www.spinics.net/lists/dri-devel/msg147302.html)
> 2. drm/amdgpu: Implement ttm_bo_driver.access_memory callback v2
> (http://www.spinics.net/lists/dri-devel/msg147303.html)
> 
> Those patches allow you to just attach gdb to a process and access the
> content of every CPU mapped buffer, even when that buffer is in CPU
> invisible VRAM.
> 
> So the primary goal of that effort is *NOT* to make the BO content
> accessible to the debugger through the BO housekeeping, cause that's
> something we already have.
> 
> The goal here is to walk the different page tables and hardware mapping
> functionalities to access the data just in the same way the GPU would do to
> catch problems.
> 
> That works fine as long as IOMMU is disabled, but when it is enabled the
> additional mapping breaks our neck and we don't knew if the stuff the page
> table dumper is producing is correct or not.
> 
> So what we need is just a way to translate dma addresses back to struct
> pages to check their validity.
> 
> I've considered to just add this information to amdgpu_gem_info, but then we
> get page->dma-address mapping instead dma-address->page as we want.
> 

So to summarize you are saying you do not trust the value you get from
pci_map_page() ?

If not then i stress again that you have all the informations you need
inside the amdgpu driver. You can take the same scheme i propose to
dump ttm.dma_address[] and compare against content of GPU page table.

Jérôme