Bug 67016

Perhaps mesa needs to be updated after this commit. I have tried to change mesa to align to 32768 instead of 4096. With some changes I can make the test pass a few times before it hangs. It does not hard lock everytime with the changes, so I managed to get a dmesg. Don't know how useful it is though because of my, probably incorrect, changes.

Does the test work reliably with that kernel commit reverted?  You don't need to adjust the alignment of anything in mesa.  I don't see how that commit would cause any regressions.  It just adjusts the alignment of VM page table blocks in the kernel driver.  Even if cayman doesn't need 32K VM page table alignment, over-aligning shouldn't hurt.

(In reply to comment #2)
> Does the test work reliably with that kernel commit reverted?  You don't
> need to adjust the alignment of anything in mesa.  I don't see how that
> commit would cause any regressions.  It just adjusts the alignment of VM
> page table blocks in the kernel driver.  Even if cayman doesn't need 32K VM
> page table alignment, over-aligning shouldn't hurt.

I ran the test 500 times without fail with RADEON_VA=0 with that kernel commit.

I also ran the test 500 times without fail with that kernel commit reverted and RADEON_VA=1.

Created attachment 82599 [details]
dmesg

I managed to get a dmesg without my patches.

Do you still get the issue with dpm disabled?

That commit causes problems (VM faults, lockups) on my Cape Verde card as well.

I suspect other kernel code needs to be adjusted for the increased buffer sizes, or something like that.

(In reply to comment #5)
> Do you still get the issue with dpm disabled?

yes

Ok.  I'll go ahead and revert it for now.

Created attachment 82616 [details] [review]
possible fix

This patch fixes the issue for me.

I think that patch is correct.  We have to align the PTE block size as well.

At least it works for me, I have run a complete piglit test(quick.tests) with that patch without issues, with dpm enabled.

(In reply to comment #9)
> Created attachment 82616 [details] [review] [review]
> possible fix
> 
> This patch fixes the issue for me.

The patch itself is correct, but the problem is it shouldn't be necessary!

When the increment is incorrect we should just make allot of small page directory updates instead of one big update, not as efficient but should work also...

Could you try to setting the increment to something like 0xffffffff and so disable the accumulation of updates and see if that still doesn't work?

Created attachment 82622 [details] [review]
only align the pt base to 32k

Does this patch help?  Only the page table base address should need the 32k alignment.

(In reply to comment #13)
> Created attachment 82622 [details] [review] [review]
> only align the pt base to 32k
> 
> Does this patch help?  Only the page table base address should need the 32k
> alignment.

Yes that patch fixes the problem.

(In reply to comment #14)
> (In reply to comment #13)
> > Created attachment 82622 [details] [review] [review] [review]
> > only align the pt base to 32k
> > 
> > Does this patch help?  Only the page table base address should need the 32k
> > alignment.
> 
> Yes that patch fixes the problem.

Even if the problem is fixed for now can you please make the test I suggested?

We are still having some problems with the virtual memory support for NI and it would be nice if  we can narrow thos down a bit more.

(In reply to comment #15)
> (In reply to comment #14)
> > (In reply to comment #13)
> > > Created attachment 82622 [details] [review] [review] [review] [review]
> > > only align the pt base to 32k
> > > 
> > > Does this patch help?  Only the page table base address should need the 32k
> > > alignment.
> > 
> > Yes that patch fixes the problem.
> 
> Even if the problem is fixed for now can you please make the test I
> suggested?
> 
> We are still having some problems with the virtual memory support for NI and
> it would be nice if  we can narrow thos down a bit more.

I reset the branch to drm-fixes-3.11 and set incr to 0xffffffff. The computer booted fine but when I ran vs-textureSize-compare the computer locked up.

*** Bug 67102 has been marked as a duplicate of this bug. ***

-- GitLab Migration Automatic Message --

This bug has been migrated to freedesktop.org's GitLab instance and has been closed from further activity.

You can subscribe and participate further through the new bug through this link to our GitLab instance: https://gitlab.freedesktop.org/mesa/mesa/issues/454.