Screen corruption using radeon kernel driver

[Robin Murphy]

On 2022-11-30 19:59, Mikhail Krylov wrote:
> On Wed, Nov 30, 2022 at 11:07:32AM -0500, Alex Deucher wrote:
>> On Wed, Nov 30, 2022 at 10:42 AM Robin Murphy <robin.murphy at arm.com> wrote:
>>>
>>> On 2022-11-30 14:28, Alex Deucher wrote:
>>>> On Wed, Nov 30, 2022 at 7:54 AM Robin Murphy <robin.murphy at arm.com> wrote:
>>>>>
>>>>> On 2022-11-29 17:11, Mikhail Krylov wrote:
>>>>>> On Tue, Nov 29, 2022 at 11:05:28AM -0500, Alex Deucher wrote:
>>>>>>> On Tue, Nov 29, 2022 at 10:59 AM Mikhail Krylov <sqarert at gmail.com> wrote:
>>>>>>>>
>>>>>>>> On Tue, Nov 29, 2022 at 09:44:19AM -0500, Alex Deucher wrote:
>>>>>>>>> On Mon, Nov 28, 2022 at 3:48 PM Mikhail Krylov <sqarert at gmail.com> wrote:
>>>>>>>>>>
>>>>>>>>>> On Mon, Nov 28, 2022 at 09:50:50AM -0500, Alex Deucher wrote:
>>>>>>>>>>
>>>>>>>>>>>>> [excessive quoting removed]
>>>>>>>>>>
>>>>>>>>>>>> So, is there any progress on this issue? I do understand it's not a high
>>>>>>>>>>>> priority one, and today I've checked it on 6.0 kernel, and
>>>>>>>>>>>> unfortunately, it still persists...
>>>>>>>>>>>>
>>>>>>>>>>>> I'm considering writing a patch that will allow user to override
>>>>>>>>>>>> need_dma32/dma_bits setting with a module parameter. I'll have some time
>>>>>>>>>>>> after the New Year for that.
>>>>>>>>>>>>
>>>>>>>>>>>> Is it at all possible that such a patch will be merged into kernel?
>>>>>>>>>>>>
>>>>>>>>>>> On Mon, Nov 28, 2022 at 9:31 AM Mikhail Krylov <sqarert at gmail.com> wrote:
>>>>>>>>>>> Unless someone familiar with HIMEM can figure out what is going wrong
>>>>>>>>>>> we should just revert the patch.
>>>>>>>>>>>
>>>>>>>>>>> Alex
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> Okay, I was suggesting that mostly because
>>>>>>>>>>
>>>>>>>>>> a) it works for me with dma_bits = 40 (I understand that's what it is
>>>>>>>>>> without the original patch applied);
>>>>>>>>>>
>>>>>>>>>> b) there's a hint of uncertainity on this line
>>>>>>>>>> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/gpu/drm/radeon/radeon_device.c#n1359
>>>>>>>>>> saying that for AGP dma_bits = 32 is the safest option, so apparently there are
>>>>>>>>>> setups, unlike mine, where dma_bits = 32 is better than 40.
>>>>>>>>>>
>>>>>>>>>> But I'm in no position to argue, just wanted to make myself clear.
>>>>>>>>>> I'm okay with rebuilding the kernel for my machine until the original
>>>>>>>>>> patch is reverted or any other fix is applied.
>>>>>>>>>
>>>>>>>>> What GPU do you have and is it AGP?  If it is AGP, does setting
>>>>>>>>> radeon.agpmode=-1 also fix it?
>>>>>>>>>
>>>>>>>>> Alex
>>>>>>>>
>>>>>>>> That is ATI Radeon X1950, and, unfortunately, radeon.agpmode=-1 doesn't
>>>>>>>> help, it just makes 3D acceleration in games such as OpenArena stop
>>>>>>>> working.
>>>>>>>
>>>>>>> Just to confirm, is the board AGP or PCIe?
>>>>>>>
>>>>>>> Alex
>>>>>>
>>>>>> It is AGP. That's an old machine.
>>>>>
>>>>> Can you check whether dma_addressing_limited() is actually returning the
>>>>> expected result at the point of radeon_ttm_init()? Disabling highmem is
>>>>> presumably just hiding whatever problem exists, by throwing away all
>>>>>    >32-bit RAM such that use_dma32 doesn't matter.
>>>>
>>>> The device in question only supports a 32 bit DMA mask so
>>>> dma_addressing_limited() should return true.  Bounce buffers are not
>>>> really usable on GPUs because they map so much memory.  If
>>>> dma_addressing_limited() returns false, that would explain it.
>>>
>>> Right, it appears to be the only part of the offending commit that
>>> *could* reasonably make any difference, so I'm primarily wondering if
>>> dma_get_required_mask() somehow gets confused.
>>
>> Mikhail,
>>
>> Can you see that dma_addressing_limited() and dma_get_required_mask()
>> return in this case?
>>
>> Alex
>>
>>
>>>
>>> Thanks,
>>> Robin.
> 
> Unfortunately, right now I don't have enough time for kernel
> modifications and rebuilds (I will later!), so I did a quick-and-dirty
> research with kprobe.
> 
> The problem is that dma_addressing_limited() seems to be inlined and
> kprobe fails to intercept it.
> 
> But I managed to get the result of dma_get_required_mask(). It returns
> 0x7fffffff (!) on the vanilla (with the patch, buggy) kernel:
>   
> $ sudo kprobe-perf 'r:dma_get_required_mask $retval'
> Tracing kprobe dma_get_required_mask. Ctrl-C to end.
>          modprobe-1244    [000] d...   105.582816: dma_get_required_mask: (radeon_ttm_init+0x61/0x240 [radeon] <- dma_get_required_mask) arg1=0x7fffffff
> 
> This function does not even get called in the kernel without the patch
> that I built myself. I believe that's because ttm_bo_device_init()
> doesn't call it without the patch.
> 
> Hope that helps at least a bit. If not, I'll be able to do more thorough
> research in a couple of weeks, probably.

Hmm, just to clarify, what's your actual RAM layout? I've been assuming
that the issue must be caused by unexpected DMA address truncation, but
double-checking the older threads it seems that might not be the case.
I just did a quick sanity-check of both HIGHMEM4G and HIGHMEM64G configs
in a VM with either 2GB or 4GB of RAM assigned, and the
dma_direct_get_required_mask() calculation seemed to return the
appropriate result for all combinations.

Otherwise, the only significant difference of use_dma32 seems to be to
switch TTM's allocation flags from GFP_HIGHUSER to GFP_DMA32. Could it
just be that the highmem support somewhere between TTM and radeon has
bitrotted, and it hasn't been noticed until this change because everyone
still using a 32-bit system with highmem also happens not to be using a
newer 40-bit-capable GPU? Or perhaps it never worked for AGP at all, in
which case an explicit special case might be clearer?

diff --git a/drivers/gpu/drm/radeon/radeon_ttm.c b/drivers/gpu/drm/radeon/radeon_ttm.c
index d33fec488713..acb2d534bff5 100644
--- a/drivers/gpu/drm/radeon/radeon_ttm.c
+++ b/drivers/gpu/drm/radeon/radeon_ttm.c
@@ -696,6 +696,7 @@ int radeon_ttm_init(struct radeon_device *rdev)
  			       rdev->ddev->anon_inode->i_mapping,
  			       rdev->ddev->vma_offset_manager,
  			       rdev->need_swiotlb,
+			       rdev->flags & RADEON_IS_AGP ||
  			       dma_addressing_limited(&rdev->pdev->dev));
  	if (r) {
  		DRM_ERROR("failed initializing buffer object driver(%d).\n", r);
  

Robin.