[PATCH] exynos-drm: Fix display manager failing to start without IOMMU problem

Shuah Khan shuahkh at osg.samsung.com
Wed Oct 19 22:27:23 UTC 2016


On 10/19/2016 08:16 AM, Inki Dae wrote:
> Hi Shuah,
> 
> 2016-10-13 8:11 GMT+09:00 Shuah Khan <shuahkh at osg.samsung.com>:
>> Hi Inki,
>>
>> On 08/15/2016 10:40 PM, Inki Dae wrote:
>>
>>>>
>>>> okay the very first commit that added IOMMU support
>>>> introduced the code that rejects non-contig gem memory
>>>> type without IOMMU.
>>>>
>>>> commit 0519f9a12d0113caab78980c48a7902d2bd40c2c
>>>> Author: Inki Dae <inki.dae at samsung.com>
>>>> Date:   Sat Oct 20 07:53:42 2012 -0700
>>>>
>>>>     drm/exynos: add iommu support for exynos drm framework
>>>>
>>
>> I haven't given up on this yet. I am still seeing the following failure:
>>
>> Additional debug messages I added:
>> [   15.287403] exynos_drm_gem_create_ioctl() 1
>> [   15.287419] exynos_drm_gem_create() flags 1
>>
>> [   15.311511] [drm:exynos_drm_framebuffer_init] *ERROR* Non-contiguous GEM memory is not supported.
>>
>> Additional debug message I added:
>> [   15.318981] [drm:exynos_user_fb_create] *ERROR* failed to initialize framebuffer
>>
>> This is what happens:
>>
>> 1. exynos_drm_gem_create_ioctl() gets called with EXYNOS_BO_NONCONTIG request
>> 2. exynos_drm_gem_create(0 goes ahead and creates the GEM buffers
>> 3. exynos_user_fb_create() tries to associate GEM to fb and fails during
>>    check_fb_gem_memory_type()
>>
>> At this point, there is no recovery and lightdm fails
>>
>> xf86-video-armsoc/src/drmmode_exynos/drmmode_exynos.c assumes contiguous
>> allocations are not supported in some exynos drm versions: The following
>> commit introduced this change:
>>
>> https://git.linaro.org/arm/xorg/driver/xf86-video-armsoc.git/commitdiff/3be1f6273441fe95dd442f44064387322e16b7e9
>>
>> excerpts from the diff:-       if (create_gem->buf_type == ARMSOC_BO_SCANOUT)
>> -               create_exynos.flags = EXYNOS_BO_CONTIG;
>> -       else
>> -               create_exynos.flags = EXYNOS_BO_NONCONTIG;
>> +
>> +       /* Contiguous allocations are not supported in some exynos drm versions.
>> +        * When they are supported all allocations are effectively contiguous
>> +        * anyway, so for simplicity we always request non contiguous buffers.
>> +        */
>> +       create_exynos.flags = EXYNOS_BO_NONCONTIG;
>>
> 
> Above comment, "Contiguous allocations are not supported in some
> exynos drm versions.", seems wrong assumption.
> The root cause, contiguous allocation is not supported, would be that
> they used Linux kernel which didn't have CMA region enough - as
> default 16MB, or didn't declare CMA region enough for the DMA device.
> So I think they should not force to flag EXYNOS_BO_NONCONTIG and they
> should manage the error case if the allocation failed.

This assumption doesn't sound correct and forcing NONCONTIG isn't right
either. 

> 
>> There might have been logic on exynos_drm that forced Contig when it coudn't
>> support NONCONTIG. At least, that is what this comment suggests. This assumption
>> doesn't appear to be a good one and not sure if this change was made to fix a bug.
>>
>> After the IOMMU support, this assumption is no longer true. Hence, with IOMMU
>> support, latest kernels have a mismatch with the installed xf86-video-armsoc
>>
>> This is what I am running into. This leads to the following question:
>>
>> 1. How do we ensure exynos_drm kernel changes don't break user-space
>>    specifically xf86-video-armsoc
>> 2. This seems to have gone undetected for a while. I see a change in
>>    exynos_drm_gem_dumb_create() that is probably addressing this type
>>    of breakage. Commit 122beea84bb90236b1ae545f08267af58591c21b adds
>>    handling for IOMMU NONCONTIG case.
> 
> Seems this patch has a problem. This patch forces to flag NONCONTIG if
> iommu is enabled. The flag should be depend on the argument from
> user-space.
> I.e., if user-space requested a gem allocation with CONTIG flag, then
> Exynos drm driver should allocate contiguous memory even though iommu
> is enabled.
> 
>>
>> Anyway, I am interested in getting the exynos_drm kernel side code
>> and xf86-video-armsoc in sync to resolve the issue.
>>
>> Could you recommend a going forward plan?
> 
> My opinion are,
> 
> 1. Do not force to flag EXYNOS_BO_NONCONTIG at xf86-video-armsoc
> 2. Do not force to flag NONCONTIG at Exynos drm driver even though
> iommu is enabled and flag allocation type with the argument from
> user-space.
> 
> I think you could try to post above patches.
> 

Sounds good. I will work on the above two patches.

thanks,
-- Shuah




More information about the dri-devel mailing list