Re: drm_gem_dumb_map_offset patch

Re: drm_gem_dumb_map_offset patch

[Noralf Trønnes]

Hi,

[add Daniel Vetter]
[cc dri-devel]

Den 29.05.2019 15.09, skrev Pierre Yves MORDRET:
> Hello Noralf,
> 
> Sorry for bothering you with question but I need to better understand the
> rational about a patch you did in DRM/GEM.
> 
> First of all, let me introduce myself.
> I'm currently employee to STMicroelectronics company and in charge of GPU
> integration within STM32 MPU (Cortex A7 + Cortex M4)
> 
> On Cortex A7 is running a Linux Kernel 4.19 as for today.
> 
> We came across some trouble when we switch from Kernel 4.14 to 4.19 for GPU
> stack. On august you submit this commit :
> 
> 	90378e58919285637aa0f063c04ba0c6449d98b1
>     	    drm/gem: drm_gem_dumb_map_offset(): reject dma-buf
> 
>     	    Reject mapping an imported dma-buf since is's an invalid use-case.
> 
> In Userland GPU stack we have such statements :
>    bo_map_fd
>        DRM_IOCTL_MODE_MAP_DUMB
>        mmap (offset)
> 
> With the patch above, ioctl returns an error EINVAL and prevents the nmap.
> As for today we revert this patch and it works fine in our end.
> 
> Thus the questions are :
>  - what is the rational behind this fix ?
>  - What we doing wrong this situation ?
>  - What do we need in such situation ?
> 

I need to pass those on to Daniel Vetter (DRM maintainer) since he
wanted the change. The details were never clear to me.
Some of the discussion is here:
https://patchwork.freedesktop.org/patch/172242/

Noralf.

> Many thanks in advance.
> Best Regards
> 
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: drm_gem_dumb_map_offset patch

[Daniel Vetter]

On Fri, May 31, 2019 at 1:28 PM Noralf Trønnes <noralf@tronnes.org> wrote:
>
> Hi,
>
> [add Daniel Vetter]
> [cc dri-devel]
>
> Den 29.05.2019 15.09, skrev Pierre Yves MORDRET:
> > Hello Noralf,
> >
> > Sorry for bothering you with question but I need to better understand the
> > rational about a patch you did in DRM/GEM.
> >
> > First of all, let me introduce myself.
> > I'm currently employee to STMicroelectronics company and in charge of GPU
> > integration within STM32 MPU (Cortex A7 + Cortex M4)
> >
> > On Cortex A7 is running a Linux Kernel 4.19 as for today.
> >
> > We came across some trouble when we switch from Kernel 4.14 to 4.19 for GPU
> > stack. On august you submit this commit :
> >
> >       90378e58919285637aa0f063c04ba0c6449d98b1
> >           drm/gem: drm_gem_dumb_map_offset(): reject dma-buf
> >
> >           Reject mapping an imported dma-buf since is's an invalid use-case.
> >
> > In Userland GPU stack we have such statements :
> >    bo_map_fd
> >        DRM_IOCTL_MODE_MAP_DUMB
> >        mmap (offset)
> >
> > With the patch above, ioctl returns an error EINVAL and prevents the nmap.
> > As for today we revert this patch and it works fine in our end.

Link to source code would be good.

> > Thus the questions are :
> >  - what is the rational behind this fix ?
> >  - What we doing wrong this situation ?
> >  - What do we need in such situation ?
> >
>
> I need to pass those on to Daniel Vetter (DRM maintainer) since he
> wanted the change. The details were never clear to me.
> Some of the discussion is here:
> https://patchwork.freedesktop.org/patch/172242/

Dumb mmap is for buffer created using dumb_create ioctl, and nothing
else. Anything else really has at best undefined behaviour. E.g. for
dma-buf you really need to call the begin/end_cpu_access ioctl, and
dumb buffers simply have no provision for that. Hence why we can't
support this in general.

Thanks, Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: drm_gem_dumb_map_offset patch

[Pierre Yves MORDRET]

Hi all,

Many thanks for your valuable support and answers.

Since Dumb mmap is for buffer created using dumb_create ioctl we won't use it
anymore. In place a mmap/ummap is called with DMA Buf FD.
After some tests it seems working in our side.

Many thanks again.
Regards.

On 5/31/19 1:36 PM, Daniel Vetter wrote:
> On Fri, May 31, 2019 at 1:28 PM Noralf Trønnes <noralf@tronnes.org> wrote:
>>
>> Hi,
>>
>> [add Daniel Vetter]
>> [cc dri-devel]
>>
>> Den 29.05.2019 15.09, skrev Pierre Yves MORDRET:
>>> Hello Noralf,
>>>
>>> Sorry for bothering you with question but I need to better understand the
>>> rational about a patch you did in DRM/GEM.
>>>
>>> First of all, let me introduce myself.
>>> I'm currently employee to STMicroelectronics company and in charge of GPU
>>> integration within STM32 MPU (Cortex A7 + Cortex M4)
>>>
>>> On Cortex A7 is running a Linux Kernel 4.19 as for today.
>>>
>>> We came across some trouble when we switch from Kernel 4.14 to 4.19 for GPU
>>> stack. On august you submit this commit :
>>>
>>>       90378e58919285637aa0f063c04ba0c6449d98b1
>>>           drm/gem: drm_gem_dumb_map_offset(): reject dma-buf
>>>
>>>           Reject mapping an imported dma-buf since is's an invalid use-case.
>>>
>>> In Userland GPU stack we have such statements :
>>>    bo_map_fd
>>>        DRM_IOCTL_MODE_MAP_DUMB
>>>        mmap (offset)
>>>
>>> With the patch above, ioctl returns an error EINVAL and prevents the nmap.
>>> As for today we revert this patch and it works fine in our end.
> 
> Link to source code would be good.
> 
>>> Thus the questions are :
>>>  - what is the rational behind this fix ?
>>>  - What we doing wrong this situation ?
>>>  - What do we need in such situation ?
>>>
>>
>> I need to pass those on to Daniel Vetter (DRM maintainer) since he
>> wanted the change. The details were never clear to me.
>> Some of the discussion is here:
>> https://patchwork.freedesktop.org/patch/172242/
> 
> Dumb mmap is for buffer created using dumb_create ioctl, and nothing
> else. Anything else really has at best undefined behaviour. E.g. for
> dma-buf you really need to call the begin/end_cpu_access ioctl, and
> dumb buffers simply have no provision for that. Hence why we can't
> support this in general.
> 
> Thanks, Daniel
> 
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: drm_gem_dumb_map_offset patch

[Daniel Vetter]

On Mon, Jun 3, 2019 at 2:25 PM Pierre Yves MORDRET
<pierre-yves.mordret@st.com> wrote:
>
> Hi all,
>
> Many thanks for your valuable support and answers.
>
> Since Dumb mmap is for buffer created using dumb_create ioctl we won't use it
> anymore. In place a mmap/ummap is called with DMA Buf FD.
> After some tests it seems working in our side.

Please note that with dma-buf you need to call the begin/end cpu
access ioctl, or depending upon driver/platform the mmap access might
be incoherent.
-Daniel

>
> Many thanks again.
> Regards.
>
> On 5/31/19 1:36 PM, Daniel Vetter wrote:
> > On Fri, May 31, 2019 at 1:28 PM Noralf Trønnes <noralf@tronnes.org> wrote:
> >>
> >> Hi,
> >>
> >> [add Daniel Vetter]
> >> [cc dri-devel]
> >>
> >> Den 29.05.2019 15.09, skrev Pierre Yves MORDRET:
> >>> Hello Noralf,
> >>>
> >>> Sorry for bothering you with question but I need to better understand the
> >>> rational about a patch you did in DRM/GEM.
> >>>
> >>> First of all, let me introduce myself.
> >>> I'm currently employee to STMicroelectronics company and in charge of GPU
> >>> integration within STM32 MPU (Cortex A7 + Cortex M4)
> >>>
> >>> On Cortex A7 is running a Linux Kernel 4.19 as for today.
> >>>
> >>> We came across some trouble when we switch from Kernel 4.14 to 4.19 for GPU
> >>> stack. On august you submit this commit :
> >>>
> >>>       90378e58919285637aa0f063c04ba0c6449d98b1
> >>>           drm/gem: drm_gem_dumb_map_offset(): reject dma-buf
> >>>
> >>>           Reject mapping an imported dma-buf since is's an invalid use-case.
> >>>
> >>> In Userland GPU stack we have such statements :
> >>>    bo_map_fd
> >>>        DRM_IOCTL_MODE_MAP_DUMB
> >>>        mmap (offset)
> >>>
> >>> With the patch above, ioctl returns an error EINVAL and prevents the nmap.
> >>> As for today we revert this patch and it works fine in our end.
> >
> > Link to source code would be good.
> >
> >>> Thus the questions are :
> >>>  - what is the rational behind this fix ?
> >>>  - What we doing wrong this situation ?
> >>>  - What do we need in such situation ?
> >>>
> >>
> >> I need to pass those on to Daniel Vetter (DRM maintainer) since he
> >> wanted the change. The details were never clear to me.
> >> Some of the discussion is here:
> >> https://patchwork.freedesktop.org/patch/172242/
> >
> > Dumb mmap is for buffer created using dumb_create ioctl, and nothing
> > else. Anything else really has at best undefined behaviour. E.g. for
> > dma-buf you really need to call the begin/end_cpu_access ioctl, and
> > dumb buffers simply have no provision for that. Hence why we can't
> > support this in general.
> >
> > Thanks, Daniel
> >



-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch
_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel