[PATCH v10 00/17] drm/exynos: atomic modesetting support

[Joonyoung Shim]

On 06/10/2015 10:36 PM, Gustavo Padovan wrote:
> Hi Marek,
> 
> 2015-06-10 Marek Szyprowski <m.szyprowski at samsung.com>:
> 
>> Hello,
>>
>> On 2015-06-01 17:04, Gustavo Padovan wrote:
>>> From: Gustavo Padovan <gustavo.padovan at collabora.co.uk>
>>>
>>> Hi,
>>>
>>> Here goes the full support for atomic modesetting on exynos. I've
>>> split the patches in the various phases of atomic support.
>>
>> Thanks for this patchses, however I've noticed a problem after applying
>> them.
>> The issue gets revealed when support for IOMMU is enabled. I've did my tests
>> with Exynos HDMI driver on Odroid U3 board.
>>
>> To demonstrated the issue I've added following additional debug in the
>> exynos
>> mixer driver in mixer_graph_buffer() function:
>> pr_info("dma addr %pad plane->src_width %d plane->src_height %d\n",
>> &plane->dma_addr[0], plane->src_width, plane->src_height);
>>
>> Before applying patches setting 640x480 mode and getting back to 1920x1080
>> console generates following log:
>>
>> # modetest -M exynos -s 23:640x480
>> setting mode 640x480-60Hz at XR24 on connectors 23, crtc 21
>> [ 3860.617151] dma 0xbc500000 plane->src_width 640 plane->src_height 480
>> ^C
>> [ 3870.555232] dma 0xbbd00000 plane->src_width 1920 plane->src_height 1080
>> [ 3870.565696] dma 0xbbd00000 plane->src_width 1920 plane->src_height 1080
>>
>> After applying atomic modesetting patchset:
>> # modetest -M exynos -s 24:640x480
>> [  142.540122] dma 0xbbd00000 plane->src_width 1920 plane->src_height 1080
>> [  142.550726] dma 0xbbd00000 plane->src_width 1920 plane->src_height 1080
>> setting mode 640x480-60Hz at XR24 on connectors 24, crtc 22
>> [  142.643672] dma 0xbc500000 plane->src_width 1920 plane->src_height 1080
>> [  142.759982] dma 0xbc500000 plane->src_width 640 plane->src_height 480
>> ^C
>> [  154.986040] dma 0xbbd00000 plane->src_width 1920 plane->src_height 1080
>>
>> As you can see from the above log, mixer_graph_buffer function is called
>> several times. 0xbbd00000 is the DMA address of the 1920x1080 framebuffer
>> and 0xbc500000 is the DMA address of the allocated 640x480 buffer.
>> mixer_graph_buffer() is first called with the new DMA address of the
>> framebuffer, but with the old mode parameters (1920x1080 size) and then
>> in the next call it updates the plane parameters to the correct values
>> (size changed to 640x480). When IOMMU is not used, this can be easily
>> missed, but after enabling IOMMU support, any DMA access to unallocated
>> address causes IOMMU PAGE FAULT. Here it will happen after changing DMA
>> address of the buffer without changing the size.
>>
>> A quick workaround to resolve this multiple calls to mixer_graph_buffer()
>> with partially updated mode values is to remove calls to
>> mixer_window_suspend/mixer_window_resume from mixer_disable and
>> mixer_disable functions, but I expect that this is not the right
>> approach.
>>
>> Probably the same problem can be observed with Exynos FIMD driver.
>>
>> Gustavo: could you check if mixer_enable functions should really
>> call mixer_window_resume function, which in turn calls mixer_win_commit,
>> which calls mixer_graph_buffer with partially updated display buffer
>> data?
> 
> It should not, you are right. This is actually the correct fix. Atomic
> modesetting should not do chained calls, e.g., crtc_disable calling
> plane_disable. This change was already in my plan actually, but as I had
> IOMMU disabled I didn't see it here, and I didn't create this patch yet.
> 

Right, but it needs that crtc_disable calls plane_disable in exynos
driver internally. Because crtc is disabled before plane is disabled,
it means plane_disable just returns without any register changes,
then we cannot be sure setting register to disable plane when crtc is
disable.

I think a solution is enough only to eliminate calling xxx_resume
function in exynos driver function when enable plane, we can remove it
because it's called to enable plane from drm atomic modeset framework.