Re: [PATCH v4 7/8] drm/fourcc: amlogic: Add modifier definitions for the Scatter layout

[Pekka Paalanen <ppaalanen@gmail.com>]

[-- Attachment #1: Type: text/plain, Size: 4323 bytes --]

On Wed, 25 Mar 2020 17:18:15 +0100
Neil Armstrong <narmstrong@baylibre.com> wrote:

> Hi,
> 
> On 25/03/2020 14:49, Pekka Paalanen wrote:
> > On Wed, 25 Mar 2020 11:24:15 +0100
> > Neil Armstrong <narmstrong@baylibre.com> wrote:
> >   
> >> Hi,
> >>
> >> On 25/03/2020 10:04, Simon Ser wrote:  
> >>> On Wednesday, March 25, 2020 9:50 AM, Neil Armstrong <narmstrong@baylibre.com> wrote:
> >>>     
> >>>> Amlogic uses a proprietary lossless image compression protocol and format
> >>>> for their hardware video codec accelerators, either video decoders or
> >>>> video input encoders.
> >>>>
> >>>> This introduces the Scatter Memory layout, means the header contains IOMMU
> >>>> references to the compressed frames content to optimize memory access
> >>>> and layout.
> >>>>
> >>>> In this mode, only the header memory address is needed, thus the content
> >>>> memory organization is tied to the current producer execution and cannot
> >>>> be saved/dumped neither transferrable between Amlogic SoCs supporting this
> >>>> modifier.    
> >>>
> >>> I don't think this is suitable for modifiers. User-space relies on
> >>> being able to copy a buffer from one machine to another over the
> >>> network. It would be pretty annoying for user-space to have a blacklist
> >>> of modifiers that don't work this way.
> >>>
> >>> Example of such user-space:
> >>> https://gitlab.freedesktop.org/mstoeckl/waypipe/
> >>>     
> >>
> >> I really understand your point, but this is one of the use-cases we need solve.
> >> This is why I split the fourcc patch and added an explicit comment.
> >>
> >> Please point me a way to display such buffer, the HW exists, works like that and
> >> it's a fact and can't change.
> >>
> >> It will be the same for secure zero-copy buffers we can't map from userspace, but
> >> only the HW decoder can read/write and HW display can read.  
> > 
> > The comparison to secure buffers is a good one.
> > 
> > Are buffers with the DRM_FORMAT_MOD_AMLOGIC_FBC_LAYOUT_SCATTER modifier
> > meaningfully mmappable to CPU always / sometimes / never /
> > varies-and-cannot-know?  
> 
> mmappable, yes in our WIP V4L2 driver in non-secure path, meaningful, absolutely never.
> 
> So yeah, these should not be mmappable since not meaningful.

Ok. So we have a modifier that means there is no point in even trying to
mmap the buffer.

Not being able to mmap automatically makes things like waypipe not be
able to work on the buffer, so the buffer cannot be replicated over a
network, hence there is no compatibility issue. However, it still
leaves the problem that, since waypipe is "just" a message relay that
does not participate in the protocol really, the two end points might
still negotiate to use a modifier that waypipe cannot handle.

Secure buffers have the same problem: by definition, one must not be
able to replicate the buffer elsewhere.

To me it seems there needs to be a way to identify buffers that cannot
be mmapped. mmap() failing is obvious, but in waypipe's case it is too
late - the end points have already negotiated the formats and modifiers
and they cannot handle failures afterwards.

> > 
> > Maybe this type should be handled similar to secure buffers, with the
> > exception that they are not actually secured but only mostly
> > inaccessible. Then again, I haven't looked at any of the secure buffer
> > proposals.  
> 
> Actually, the Amlogic platforms offers secure video path using these exact
> modifiers, AFAIK it doesn't support the NV12 dual-write output in secure.
> 
> AFAIK last submission is from AMD, and it doesn't talk at all about mmapability
> of the secure BOs.

To me, a secure buffer concept automatically implies that there cannot
be CPU access to it. The CPU is not trusted, right? Not even the kernel.
I would assume secure implies no mmap. So I wonder, how does the secure
buffers proposal manage userspace like waypipe?

Or, is the secure buffer proposal allowing mmap, but the content is
indecipherable? Maybe they shouldn't allow mmap?

I think much of the criticism against this modifier should also be
presented to a secure buffers proposal and see how that turns out. If
they have the same problem, maybe you could use their solution?


Thanks,
pq

[-- Attachment #2: OpenPGP digital signature --]
[-- Type: application/pgp-signature, Size: 833 bytes --]