TTM and AGP conflicts

[Konrad Rzeszutek Wilk]

On Tue, Jan 03, 2012 at 05:59:33PM -0500, Konrad Rzeszutek Wilk wrote:
> On Tue, Jan 03, 2012 at 05:43:40PM -0500, Jerome Glisse wrote:
> > On Fri, Dec 23, 2011 at 01:19:43AM +0000, James Simmons wrote:
> > > 
> > > > > Hi!
> > > > >
> > > > >        I updated the openchrome tree and while testing on the AGP system
> > > > > discovered some interesting problems with the new TTM changes. The
> > > > > problems center around the ttm_tt_[un]populate which I modeled after the
> > > > > radeon and nouveau driver.
> > > > >        First problem I noticed was on a AGP system that my ttm_tt_populate
> > > > > function would oops. Tracking it down I found the problem was the
> > > > > ttm_agp_tt_create calls ttm_tt_init instead of ttm_dma_tt_init so once my
> > > > > ttm_tt_populate function would attempt to touch the dma_address it would
> > > > > oops. The second issue is the assumption of the cast for struct ttm_tt in
> > > > > both the populate and unpopulate function. For the AGP case the proper
> > > > > case would be to ttm_agp_backend.
> > > > >        At this point my assumption is the ttm_bo_move function has to be
> > > > > rewritten to handle the AGP case to avoid calling ttm_tt_bind and in all
> > > > > cases ttm_tt_bind needs to be avoided. Looking at the radeon and nouveau
> > > > > drivers I don't see that testing happening. Am I just missing something?
> > > > 
> > > > Happens on AGP radeons as well:
> > > > https://bugs.freedesktop.org/show_bug.cgi?id=43719
> > > 
> > > 	So I'm not crazy, so this needs to be fixed. Here is what my 
> > > understanding of the TTM layer is. My impression is that struct ttm_bo_driver
> > > handles multiple domains, AGP, pcie etc and in each method you have to 
> > > handle each specific domain you support. Also *move gives the impression of
> > > moving between these different domains. 
> > > 	Where as for struct ttm_backend_func was more for allocating from
> > > a specific domain. Also I never saw a clear way to handle multiple backends. 
> > > For example my AGP systems can also do pci dma as well. 
> > 
> > > _______________________________________________
> > > dri-devel mailing list
> > > dri-devel at lists.freedesktop.org
> > > http://lists.freedesktop.org/mailman/listinfo/dri-devel
> > 
> > Attached is patch to fix this, so sorry about that, i must have lost my
> > agp change along the way when working on the patchset. This patch is not
> > extensively tested, i will do more testing tomorrow on more gpu, might
> > even found an nvidia agp i can try. Again sorry for breaking this.
> 
> Hey Jerome,
> 
> Was going to look at this week and see how it performs before (and after)
> the squash ttm bind+populate operation. Any thoughts of what benchmarks I
> should run?

OK, Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk at oracle.com>


I ran it on my radeon AGP cards and they performed nicely. I do need to
run it with the nouveau AGP card tomorrow (barring any move_notify issues).