[PATCH] dmabuf: allow exporter to define customs ioctls

Daniel Vetter daniel at ffwll.ch
Wed Mar 9 10:08:30 UTC 2016


On Wed, Mar 9, 2016 at 10:40 AM, Benjamin Gaignard
<benjamin.gaignard at linaro.org> wrote:
> dmabuf have just accept one ioctl for cache management but I think
> that some exporter may need also need custom iotcl.
>
> For example I'm working since a while on a way to secure buffer (SMAF)
> [1] which is a central allocator + 2 ioctl for set/get secure status
> of the buffer.
> This not depend on one device but on the buffer itself and so need to
> be handled by the exporter.
> To implement this I see 3 solutions:
> - create an allocator with ioctl like SMAF but until now I don't get
> lot of feedback on this so I guess it is a wrong way.

Not much feedback doesn't mean wrong way, just that not a lot of
people care about this problem. Iirc I commented on some very early
draft with "makes sense, but not a problem for i915". I think adding
special ioctls to a SMAF allocator to set/control security status is
the right approach. At least assuming what I've seen thus far - it's
also really hard to assess SMAF if you don't have the entire driver
stack to look at (kernel gpu driver + userspace). Lack of the full
stack is probably another reason why you don't get comments, not
knowing those details makes review pretty much impossible.

> - add ioctl and ops in dmabuf but if we do that for buffer securing
> how many ioctls will have to be handled by dmabuf for other feature?

Atm you're the only one, and I can't think of anything else really we
might want ot expose on dma-buf. Maybe some way to get at the implicit
fences, for interop with explicit fencing. But that's something that
doesn't even need a new callback to dma-buf exporters. Generic ioctl
interfaces are seriously considered bad ABI style ;-)

> - let exporters define their own ioctl (like done in ION), it could
> allow to create per features helpers for the exporters.

ION todo says that all those ioctl should go (or well most of them)
since with the cache control ioctl they duplicate core functionality.
Or stuff we don't really want to support in upstream. It's all
captured.

> What is the best option when you have to do something relative to the
> buffer and to the device ?

Like Christian said, add an ioctl on the device driver fd that takes
the dma-buf as parameter.
-Daniel

>
> Regards,
> Benjamin
>
> [1] https://lwn.net/Articles/676270/
>
> 2016-03-09 10:18 GMT+01:00 Christian König <deathsimple at vodafone.de>:
>> Am 09.03.2016 um 10:03 schrieb Daniel Vetter:
>>>
>>> On Wed, Mar 9, 2016 at 9:54 AM, Benjamin Gaignard
>>> <benjamin.gaignard at linaro.org> wrote:
>>>>
>>>> In addition of the already existing operations allow exporter
>>>> to use it own custom ioctls.
>>>>
>>>> Signed-off-by: Benjamin Gaignard <benjamin.gaignard at linaro.org>
>>>
>>> First reaction: No way ever! More seriously, please start by
>>> explaining why you need this.
>>
>>
>> I was about to complain as well. Device specific IOCTLs should probably
>> better defined as operation on the device which takes the DMA-Buf file
>> descriptor as a parameter.
>>
>> If you really need this (which I doubt) then please define a range of IOCTL
>> numbers for which it should apply.
>>
>> Regards,
>> Christian.
>>
>>
>>> -Daniel
>>>
>>>> ---
>>>>   drivers/dma-buf/dma-buf.c | 3 +++
>>>>   include/linux/dma-buf.h   | 5 +++++
>>>>   2 files changed, 8 insertions(+)
>>>>
>>>> diff --git a/drivers/dma-buf/dma-buf.c b/drivers/dma-buf/dma-buf.c
>>>> index 9810d1d..6abd129 100644
>>>> --- a/drivers/dma-buf/dma-buf.c
>>>> +++ b/drivers/dma-buf/dma-buf.c
>>>> @@ -291,6 +291,9 @@ static long dma_buf_ioctl(struct file *file,
>>>>
>>>>                  return 0;
>>>>          default:
>>>> +               if (dmabuf->ops->ioctl)
>>>> +                       return dmabuf->ops->ioctl(dmabuf, cmd, arg);
>>>> +
>>>>                  return -ENOTTY;
>>>>          }
>>>>   }
>>>> diff --git a/include/linux/dma-buf.h b/include/linux/dma-buf.h
>>>> index 532108e..b6f9837 100644
>>>> --- a/include/linux/dma-buf.h
>>>> +++ b/include/linux/dma-buf.h
>>>> @@ -70,6 +70,9 @@ struct dma_buf_attachment;
>>>>    * @vmap: [optional] creates a virtual mapping for the buffer into
>>>> kernel
>>>>    *       address space. Same restrictions as for vmap and friends
>>>> apply.
>>>>    * @vunmap: [optional] unmaps a vmap from the buffer
>>>> + * @ioctl: [optional] ioctls supported by the exporter.
>>>> + *        It is up to the exporter to do the proper copy_{from/to}_user
>>>> + *        calls. Should return -EINVAL in case of error.
>>>>    */
>>>>   struct dma_buf_ops {
>>>>          int (*attach)(struct dma_buf *, struct device *,
>>>> @@ -104,6 +107,8 @@ struct dma_buf_ops {
>>>>
>>>>          void *(*vmap)(struct dma_buf *);
>>>>          void (*vunmap)(struct dma_buf *, void *vaddr);
>>>> +
>>>> +       int (*ioctl)(struct dma_buf *, unsigned int cmd, unsigned long
>>>> arg);
>>>>   };
>>>>
>>>>   /**
>>>> --
>>>> 1.9.1
>>>>
>>>> _______________________________________________
>>>> dri-devel mailing list
>>>> dri-devel at lists.freedesktop.org
>>>> https://lists.freedesktop.org/mailman/listinfo/dri-devel
>>>
>>>
>>>
>>
>
>
>
> --
> Benjamin Gaignard
>
> Graphic Working Group
>
> Linaro.org │ Open source software for ARM SoCs
>
> Follow Linaro: Facebook | Twitter | Blog



-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch


More information about the dri-devel mailing list