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

Tue Oct 25 17:57:33 UTC 2016

Hello Shuah,

Shuah Khan wrote:
> On 10/19/2016 04:27 PM, Shuah Khan wrote:
>> 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
> 
> Okay more on this. I fixed xf86-video-armso to ask for EXYNOS_BO_CONTIG
> for ARMSOC_BO_SCANOUT and EXYNOS_BO_NONCONTIG in all other cases.
> 
> Revert of the commit: 3be1f6273441fe95dd442f44064387322e16b7e9
> 
> With this change, now display manager starts just fine. However, it turns
> out xf86-video-armsoc is obsoleted in favor of xf86-video-modesetting. The
> last update to xf86-video-armsoc git was 3 years ago.
IIRC xf86-video-armsoc was created to facilitate integration with the
proprietary Mali userspace blob. I don't think that can be done with the
modesetting DDX.

> I am not sure where to send the fix and doesn't look like it is being
> maintained actively. I can pursue it further and try to get this into
> xf86-video-armsoc provided I can find the maintainer for this seemingly
> inactive project.
I was wondering if anyone has actually complained about this issue?

> I brought in the latest xf86-video-modesetting bits from
> 
> https://cgit.freedesktop.org/xorg/driver/xf86-video-modesetting
> 
> I removed xf86-video-armsoc and installed xf86-video-modesetting and that
> worked just fine. xf86-video-modesetting uses dumb_create interface instead
> of DRM_IOCTL_EXYNOS_GEM_CREATE.
> 
> dumb_create interface eliminates the need for DRM_IOCTL_EXYNOS_GEM_CREATE
> in userspace. libdrm-2.4.71 does call DRM_IOCTL_EXYNOS_GEM_CREATE.
> 
> The question is do we still need to continue to support the custom gem
> create interface DRM_IOCTL_EXYNOS_GEM_CREATE?
I'd say yes. With just the dumb_create interface it is not possible to
change the type of buffer you get. And if you want to support any kind
of hw acceleration in the DDX, you probably want to control at least the
caching behaviour of the buffer.

> Some drm drivers seem to support it and some don't.
Not sure I understand this. I don't think any other DRM driver except
for exynos supports this ioctl. But all drivers have their own ioctls to
request memory (DRM_IOCTL_ETNAVIV_GEM_NEW, DRM_IOCTL_VC4_CREATE_BO, etc.)

>  Unfortunately, if userspace requests noncontig
> for scanout, we will continue to see problems with display manager when
> iommu is disabled. dumb create hides all of that, are there good reasons
> to continue to support it in exynos drm?
In addition to the stuff I said above, you can use it for render nodes.
That doesn't work for dumb_create.

> Exposing CONTIG and NONCONTIG to userspace appears to be causing problems
> when exynos drm determines it can't support non-contig GEM buffers in fb
> init after userspace allocates them.
Might be missing something here, but this whole thing just looks like a
bug in xf86-video-armsoc. No matter from which perspective you look, the
commit that changed all allocations to noncontig was plain wrong.

My guess: This was done to paper over some bug in some vendor kernel for
a product using an Exynos SoC.

With best wishes,
Tobias

> 
>>> 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.
>>>
> 
> exynos_drm_gem_dumb_create() doesn't take any flags, there is no need
> to change the above.
> 
> thanks,
> -- Shuah
> 