[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Emil Velikov]

On 16 February 2017 at 12:43, Tobias Jakobi
<tjakobi@math.uni-bielefeld.de> wrote:
> Hello,
>
> I was wondering about the following. Wasn't there some strict
> requirement about code going upstream, which also included that there
> was a full open-source driver stack for it?
>
> I don't see how this is the case for Mali, neither in the kernel, nor in
> userspace. I'm aware that the Mali kernel driver is open-source. But it
> is not upstream, maintained out of tree, and won't land upstream in its
> current form (no resemblence to a DRM driver at all). And let's not talk
> about the userspace part.
>
> So, why should this be here?
>
Have to agree with Tobias, here.

I can see the annoyance that Maxime and others have to go through to
their systems working.
At the same time, changing upstream kernel to suit out of tree
module(s) is not how things work. Right ?

Not to mention that the series adds stable ABI exclusively(?) used by
a module which does not seem to be in the process of getting merged.

Maxime, you're a great guy but I don't think this is suitable for
upstream... yet.

Regards,
Emil

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Maxime Ripard]

On Thu, Feb 16, 2017 at 04:54:45PM +0000, Emil Velikov wrote:
> On 16 February 2017 at 12:43, Tobias Jakobi
> <tjakobi@math.uni-bielefeld.de> wrote:
> > Hello,
> >
> > I was wondering about the following. Wasn't there some strict
> > requirement about code going upstream, which also included that there
> > was a full open-source driver stack for it?
> >
> > I don't see how this is the case for Mali, neither in the kernel, nor in
> > userspace. I'm aware that the Mali kernel driver is open-source. But it
> > is not upstream, maintained out of tree, and won't land upstream in its
> > current form (no resemblence to a DRM driver at all). And let's not talk
> > about the userspace part.
> >
> > So, why should this be here?
> >
> Have to agree with Tobias, here.
> 
> I can see the annoyance that Maxime and others have to go through to
> their systems working.
> At the same time, changing upstream kernel to suit out of tree
> module(s) is not how things work. Right ?
> 
> Not to mention that the series adds stable ABI exclusively(?) used by
> a module which does not seem to be in the process of getting merged.

It really doesn't have any relation to whether a particular component
is supported in Linux. Our git repo just happens to be the canonical
source of DT, but those DTs are also used in other systems and
projects that have *no* relation with Linux, and might have a
different view on things than we do.

There's been a long-running discussion about moving the DTs out of the
kernel and in a separate repo. Would you still be opposed to it if I
happened to contribute that binding to that repo, even if Linux didn't
have any in-tree support for it? I'm pretty sure you wouldn't, yet
this is the exact same case.

And taking the ACPI example once again, this doesn't seem to bother
you at all that ACPI reports that it has a device that is not
supported in-tree in Linux. Why is it any different in DT.

We already have DT bindings for out of tree drivers, there's really
nothing new here.

Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 801 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20170217/63ac141b/attachment.sig>

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Emil Velikov]

Hi Maxime,

As I feared things have taken a turn for the bitter end :-]

It seems that this is a heated topic, so I'l kindly ask that we try
the following:

 - For people such as myself/Tobias/others who feel that driver and DT
bindings should go hand in hand, prove them wrong.
But please, do so by pointing to the documentation (conclusion of a
previous discussion). This way you don't have to repeat yourself and
get [too] annoyed over silly suggestions.

 - The series has code changes which [seemingly] cater for out of tree
module(s).
Clearly state in the commit message who is the user, why it's save to
do so and get an Ack from more prominent [DRM] developers.

Please try to understand that I do not want to annoy/agitate you, I'm
merely pointing what seems [to me] as incorrect.
Nobody is perfect, so if I/others are wrong do point me/us to a
reading to educate ourselves.

Thanks
Emil

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Maxime Ripard]

Hi,

On Fri, Feb 17, 2017 at 08:39:33PM +0000, Emil Velikov wrote:
> As I feared things have taken a turn for the bitter end :-]
> 
> It seems that this is a heated topic, so I'l kindly ask that we try
> the following:
> 
>  - For people such as myself/Tobias/others who feel that driver and DT
> bindings should go hand in hand, prove them wrong.
> But please, do so by pointing to the documentation (conclusion of a
> previous discussion). This way you don't have to repeat yourself and
> get [too] annoyed over silly suggestions.

http://lxr.free-electrons.com/source/Documentation/devicetree/usage-model.txt#L13

"The "Open Firmware Device Tree", or simply Device Tree (DT), is a
data structure and language for describing hardware. More
specifically, it is a description of hardware that is readable by an
operating system so that the operating system doesn't need to hard
code details of the machine"

http://lxr.free-electrons.com/source/Documentation/devicetree/usage-model.txt#L79

"What it does do is provide a language for decoupling the hardware
configuration from the board and device driver support in the Linux
kernel (or any other operating system for that matter)."

And like I said, we already had bindings for out of tree bindings,
like this one:
https://patchwork.kernel.org/patch/9275707/

Which triggered no discussion at the time (but the technical one,
hence a v2, that should always be done).

> - The series has code changes which [seemingly] cater for out of tree
> module(s).

That patch was dropped, only DT changes remains now, and do not depend
of that missing patch anyway.

> Clearly state in the commit message who is the user, why it's save to
> do so and get an Ack from more prominent [DRM] developers.

DRM is really not important here. We could implement a driver using
i2c as far as the DT is concerned.

FreeBSD for example uses a different, !DRM framework to support our
display stack, and still uses the DT.

Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 801 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20170223/db231799/attachment.sig>

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Emil Velikov]

Hi Maxime,

Thanks for the links.

On 24 February 2017 at 00:19, Maxime Ripard
<maxime.ripard@free-electrons.com> wrote:
> Hi,
>
> On Fri, Feb 17, 2017 at 08:39:33PM +0000, Emil Velikov wrote:
>> As I feared things have taken a turn for the bitter end :-]
>>
>> It seems that this is a heated topic, so I'l kindly ask that we try
>> the following:
>>
>>  - For people such as myself/Tobias/others who feel that driver and DT
>> bindings should go hand in hand, prove them wrong.
>> But please, do so by pointing to the documentation (conclusion of a
>> previous discussion). This way you don't have to repeat yourself and
>> get [too] annoyed over silly suggestions.
>
> http://lxr.free-electrons.com/source/Documentation/devicetree/usage-model.txt#L13
>
> "The "Open Firmware Device Tree", or simply Device Tree (DT), is a
> data structure and language for describing hardware. More
> specifically, it is a description of hardware that is readable by an
> operating system so that the operating system doesn't need to hard
> code details of the machine"
>
> http://lxr.free-electrons.com/source/Documentation/devicetree/usage-model.txt#L79
>
> "What it does do is provide a language for decoupling the hardware
> configuration from the board and device driver support in the Linux
> kernel (or any other operating system for that matter)."
>
The above seems to imply that there is (merged) device driver support
in the Linux kernel (or other) that uses the bindings.

It's not my call to make any of the policy, so I'll just kindly
suggest improving the existing documentation:
 - Reword/elaborate if out of tree [Linux or in general?] drivers are
suitable counterpart.
 - Patches could/should reference the "other OS" driver, or the "other
OS" name at least ?

Rather than clumping the above in 2.1 a separate section would be better ?

> And like I said, we already had bindings for out of tree bindings,
> like this one:
> https://patchwork.kernel.org/patch/9275707/
>
> Which triggered no discussion at the time (but the technical one,
> hence a v2, that should always be done).
>
Needless to say, there's many of us waiting to see a Mali driver land
- hence the noise. It's not meant to belittle/sway the work you and
others do.

>> - The series has code changes which [seemingly] cater for out of tree
>> module(s).
>
> That patch was dropped, only DT changes remains now, and do not depend
> of that missing patch anyway.
>
>> Clearly state in the commit message who is the user, why it's save to
>> do so and get an Ack from more prominent [DRM] developers.
>
> DRM is really not important here. We could implement a driver using
> i2c as far as the DT is concerned.
>
What I meant to say is:

Please, provide clear expectations from the start - "Linux driver is
OOT with no ETA on landing" or "driver for $FOO OS is at $LINK".
Afaict Hans did the former in the patch mentioned. Perhaps you already
did - in which case pardon for missing it.

> FreeBSD for example uses a different, !DRM framework to support our
> display stack, and still uses the DT.
>
Interesting - do you have a link handy ? Does it use open-source usespace ?

Thanks
Emil

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Rask Ingemann Lambertsen]

On Fri, Feb 17, 2017 at 04:44:19PM +0100, Maxime Ripard wrote:
[...]
> We already have DT bindings for out of tree drivers, there's really
> nothing new here.

We have DT bindings for *hardware*, not for drivers. As stated in
Documentation/devicetree/usage-model.txt:

"The "Open Firmware Device Tree", or simply Device Tree (DT), is a data
structure and language for describing hardware.  More specifically, it
is a description of hardware that is readable by an operating system
so that the operating system doesn't need to hard code details of the
machine."

"2.1 High Level View
-------------------
The most important thing to understand is that the DT is simply a data
structure that describes the hardware."

-- 
Rask Ingemann Lambertsen

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Maxime Ripard]

Hi,

On Thu, Feb 16, 2017 at 01:43:06PM +0100, Tobias Jakobi wrote:
> I was wondering about the following. Wasn't there some strict
> requirement about code going upstream, which also included that there
> was a full open-source driver stack for it?
> 
> I don't see how this is the case for Mali, neither in the kernel, nor in
> userspace. I'm aware that the Mali kernel driver is open-source. But it
> is not upstream, maintained out of tree, and won't land upstream in its
> current form (no resemblence to a DRM driver at all). And let's not talk
> about the userspace part.
> 
> So, why should this be here?

The device tree is a representation of the hardware itself. The state
of the driver support doesn't change the hardware you're running on,
just like your BIOS/UEFI on x86 won't change the device it reports to
Linux based on whether it has a driver for it.

So yes, unfortunately, we don't have a driver upstream at the
moment. But that doesn't prevent us from describing the hardware
accurately.

Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 801 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20170216/0f4319a5/attachment-0001.sig>

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Tobias Jakobi]

Hello Maxime,

Maxime Ripard wrote:
> Hi,
> 
> On Thu, Feb 16, 2017 at 01:43:06PM +0100, Tobias Jakobi wrote:
>> I was wondering about the following. Wasn't there some strict
>> requirement about code going upstream, which also included that there
>> was a full open-source driver stack for it?
>>
>> I don't see how this is the case for Mali, neither in the kernel, nor in
>> userspace. I'm aware that the Mali kernel driver is open-source. But it
>> is not upstream, maintained out of tree, and won't land upstream in its
>> current form (no resemblence to a DRM driver at all). And let's not talk
>> about the userspace part.
>>
>> So, why should this be here?
> 
> The device tree is a representation of the hardware itself. The state
> of the driver support doesn't change the hardware you're running on,
> just like your BIOS/UEFI on x86 won't change the device it reports to
> Linux based on whether it has a driver for it.
Like Emil already said, the new bindings and the DT entries are solely
introduced to support a proprietary out-of-tree module.

The current workflow when introducing new DT entries is the following:
- upstream a driver that uses the entries
- THEN add the new entries

I'm against adding such entries without having any upstream "consumer".


With best wishes,
Tobias


> So yes, unfortunately, we don't have a driver upstream at the
> moment. But that doesn't prevent us from describing the hardware
> accurately.
> 
> Maxime
> 

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Emil Velikov]

On 17 February 2017 at 12:45, Tobias Jakobi
<tjakobi@math.uni-bielefeld.de> wrote:
> Hello Maxime,
>
> Maxime Ripard wrote:
>> Hi,
>>
>> On Thu, Feb 16, 2017 at 01:43:06PM +0100, Tobias Jakobi wrote:
>>> I was wondering about the following. Wasn't there some strict
>>> requirement about code going upstream, which also included that there
>>> was a full open-source driver stack for it?
>>>
>>> I don't see how this is the case for Mali, neither in the kernel, nor in
>>> userspace. I'm aware that the Mali kernel driver is open-source. But it
>>> is not upstream, maintained out of tree, and won't land upstream in its
>>> current form (no resemblence to a DRM driver at all). And let's not talk
>>> about the userspace part.
>>>
>>> So, why should this be here?
>>
>> The device tree is a representation of the hardware itself. The state
>> of the driver support doesn't change the hardware you're running on,
>> just like your BIOS/UEFI on x86 won't change the device it reports to
>> Linux based on whether it has a driver for it.
> Like Emil already said, the new bindings and the DT entries are solely
> introduced to support a proprietary out-of-tree module.
>
> The current workflow when introducing new DT entries is the following:
> - upstream a driver that uses the entries
> - THEN add the new entries
>
That's the ideal route that I was thinking of.

At the same time, if prominent DRM people believe that we can/should
turn a blind eye, so be it.
I'm not trying to make Maxime's life hard, but point out that things
feel iffy IMHO.

Thanks
Emil

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Alexandre Belloni]

On 17/02/2017 at 13:45:44 +0100, Tobias Jakobi wrote:
> > The device tree is a representation of the hardware itself. The state
> > of the driver support doesn't change the hardware you're running on,
> > just like your BIOS/UEFI on x86 won't change the device it reports to
> > Linux based on whether it has a driver for it.
> Like Emil already said, the new bindings and the DT entries are solely
> introduced to support a proprietary out-of-tree module.
> 

Because device tree describes the hardware, the added binding doesn't
support any particular module. The eventually upstreamed drvier will
share the same bindings.

> The current workflow when introducing new DT entries is the following:
> - upstream a driver that uses the entries
> - THEN add the new entries
> 

Exactly not, if you do that, checkpatch will complain loudly. Because
you must not add a driver using bindings that are not documented first.


-- 
Alexandre Belloni, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Tobias Jakobi]

Alexandre Belloni wrote:
> On 17/02/2017 at 13:45:44 +0100, Tobias Jakobi wrote:
>>> The device tree is a representation of the hardware itself. The state
>>> of the driver support doesn't change the hardware you're running on,
>>> just like your BIOS/UEFI on x86 won't change the device it reports to
>>> Linux based on whether it has a driver for it.
>> Like Emil already said, the new bindings and the DT entries are solely
>> introduced to support a proprietary out-of-tree module.
>>
> 
> Because device tree describes the hardware, the added binding doesn't
> support any particular module. The eventually upstreamed drvier will
> share the same bindings.
OK, can we then agree that we _only_ merge the bindings and the entries,
once this driver is upstream?

Driver upstreaming and DT work go hand-in-hand. It's usually after a lot
of discussion that new bindings get finalised. And for that discussion
to happen we need to know how the driver uses the information from the
DT. Otherwise we have no way to evaluate if the description is in any
way "appropriate".

And no, I don't follow the "DT is a separate/independent thing" thought.
It maybe is in an ideal world, but we've seen it now often enough that
bindings turned out to be poorly designed, even though they looked fine
at first.

With best wishes,
Tobias


> 
>> The current workflow when introducing new DT entries is the following:
>> - upstream a driver that uses the entries
>> - THEN add the new entries
>>
> 
> Exactly not, if you do that, checkpatch will complain loudly. Because
> you must not add a driver using bindings that are not documented first.
> 
> 

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Rob Herring]

On Fri, Feb 17, 2017 at 9:56 AM, Tobias Jakobi
<tjakobi@math.uni-bielefeld.de> wrote:
> Alexandre Belloni wrote:
>> On 17/02/2017 at 13:45:44 +0100, Tobias Jakobi wrote:
>>>> The device tree is a representation of the hardware itself. The state
>>>> of the driver support doesn't change the hardware you're running on,
>>>> just like your BIOS/UEFI on x86 won't change the device it reports to
>>>> Linux based on whether it has a driver for it.
>>> Like Emil already said, the new bindings and the DT entries are solely
>>> introduced to support a proprietary out-of-tree module.
>>>
>>
>> Because device tree describes the hardware, the added binding doesn't
>> support any particular module. The eventually upstreamed drvier will
>> share the same bindings.
> OK, can we then agree that we _only_ merge the bindings and the entries,
> once this driver is upstream?

Absolutely not.

> Driver upstreaming and DT work go hand-in-hand. It's usually after a lot
> of discussion that new bindings get finalised. And for that discussion
> to happen we need to know how the driver uses the information from the
> DT. Otherwise we have no way to evaluate if the description is in any
> way "appropriate".
>
> And no, I don't follow the "DT is a separate/independent thing" thought.
> It maybe is in an ideal world, but we've seen it now often enough that
> bindings turned out to be poorly designed, even though they looked fine
> at first.

Certainly, that happens (though arguably that was more often from lack
of review). But this one is self contained, using standard, existing
properties. I'm not worried about us getting it right. If this was
something new or different, then certainly yes I would want to see the
code.

Rob

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Maxime Ripard]

On Fri, Feb 17, 2017 at 01:45:44PM +0100, Tobias Jakobi wrote:
> Hello Maxime,
> 
> Maxime Ripard wrote:
> > Hi,
> > 
> > On Thu, Feb 16, 2017 at 01:43:06PM +0100, Tobias Jakobi wrote:
> >> I was wondering about the following. Wasn't there some strict
> >> requirement about code going upstream, which also included that there
> >> was a full open-source driver stack for it?
> >>
> >> I don't see how this is the case for Mali, neither in the kernel, nor in
> >> userspace. I'm aware that the Mali kernel driver is open-source. But it
> >> is not upstream, maintained out of tree, and won't land upstream in its
> >> current form (no resemblence to a DRM driver at all). And let's not talk
> >> about the userspace part.
> >>
> >> So, why should this be here?
> > 
> > The device tree is a representation of the hardware itself. The state
> > of the driver support doesn't change the hardware you're running on,
> > just like your BIOS/UEFI on x86 won't change the device it reports to
> > Linux based on whether it has a driver for it.
>
> Like Emil already said, the new bindings and the DT entries are solely
> introduced to support a proprietary out-of-tree module.

No. This new binding and the DT entries are solely introduced to
describe a device found in a number of SoCs, just like any other DT
binding we have.

> The current workflow when introducing new DT entries is the following:
> - upstream a driver that uses the entries
> - THEN add the new entries

And that's never been the preferred workflow, for *any* patches.

Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 801 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20170217/8cdb7cbf/attachment.sig>

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Thierry Reding]

On Fri, Feb 17, 2017 at 04:43:41PM +0100, Maxime Ripard wrote:
> On Fri, Feb 17, 2017 at 01:45:44PM +0100, Tobias Jakobi wrote:
> > Hello Maxime,
> > 
> > Maxime Ripard wrote:
> > > Hi,
> > > 
> > > On Thu, Feb 16, 2017 at 01:43:06PM +0100, Tobias Jakobi wrote:
> > >> I was wondering about the following. Wasn't there some strict
> > >> requirement about code going upstream, which also included that there
> > >> was a full open-source driver stack for it?
> > >>
> > >> I don't see how this is the case for Mali, neither in the kernel, nor in
> > >> userspace. I'm aware that the Mali kernel driver is open-source. But it
> > >> is not upstream, maintained out of tree, and won't land upstream in its
> > >> current form (no resemblence to a DRM driver at all). And let's not talk
> > >> about the userspace part.
> > >>
> > >> So, why should this be here?
> > > 
> > > The device tree is a representation of the hardware itself. The state
> > > of the driver support doesn't change the hardware you're running on,
> > > just like your BIOS/UEFI on x86 won't change the device it reports to
> > > Linux based on whether it has a driver for it.
> >
> > Like Emil already said, the new bindings and the DT entries are solely
> > introduced to support a proprietary out-of-tree module.
> 
> No. This new binding and the DT entries are solely introduced to
> describe a device found in a number of SoCs, just like any other DT
> binding we have.
> 
> > The current workflow when introducing new DT entries is the following:
> > - upstream a driver that uses the entries
> > - THEN add the new entries
> 
> And that's never been the preferred workflow, for *any* patches.

Actually it has. How else are you going to test that your driver really
works? You've got to have both pieces before you can verify that they're
both adequate. So the typical workflow is to:

	1) define the bindings
	2) write a driver that implements the bindings
	3) add entries to device tree files

Usually it doesn't matter in which order you do the above because they
are all part of the same patch series. But that's not what you're doing
here. The more general problem here is that you're providing device tree
content (and therefore ABI) that's based on a binding which has no
upstream users. So you don't actually have a way of validating that what
you merge is going to be an adequate description.

You're probably going to respond: "but DT describes hardware, so it must
be known already, there won't be a need for changes". Unfortunately that
is only partially true. We've had a number of occasions where it later
turned out that a binding was in fact not an adequate description, and
then we've had to jump through hoops in order to preserve backwards
compatibility. That's already annoying enough if you've got in-tree
users, but it's going to be even more painful if you start out with an
out-of-tree user.

All of that said, you've got an Acked-by from Rob and that's about as
good as it's going to get. So I'm not going to NAK this. But I will
caution against this, because I don't think you're doing yourself any
favours with this.

So perhaps the question that we should ask is this: what do you gain by
merging this series? The fact remains that you don't have an upstream
driver that implements this binding, so ultimately you're going to be
carrying patches in some development tree anyway. Why not simply stash
these patches into the same tree? That should be about the same amount
of work for you and your users, but it has the advantage of not locking
you into something that you may regret.

Thierry
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20170220/9f6e67e4/attachment.sig>

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Maxime Ripard]

Hi Thierry,

On Mon, Feb 20, 2017 at 05:49:26PM +0100, Thierry Reding wrote:
> On Fri, Feb 17, 2017 at 04:43:41PM +0100, Maxime Ripard wrote:
> > On Fri, Feb 17, 2017 at 01:45:44PM +0100, Tobias Jakobi wrote:
> > > Hello Maxime,
> > > 
> > > Maxime Ripard wrote:
> > > > Hi,
> > > > 
> > > > On Thu, Feb 16, 2017 at 01:43:06PM +0100, Tobias Jakobi wrote:
> > > >> I was wondering about the following. Wasn't there some strict
> > > >> requirement about code going upstream, which also included that there
> > > >> was a full open-source driver stack for it?
> > > >>
> > > >> I don't see how this is the case for Mali, neither in the kernel, nor in
> > > >> userspace. I'm aware that the Mali kernel driver is open-source. But it
> > > >> is not upstream, maintained out of tree, and won't land upstream in its
> > > >> current form (no resemblence to a DRM driver at all). And let's not talk
> > > >> about the userspace part.
> > > >>
> > > >> So, why should this be here?
> > > > 
> > > > The device tree is a representation of the hardware itself. The state
> > > > of the driver support doesn't change the hardware you're running on,
> > > > just like your BIOS/UEFI on x86 won't change the device it reports to
> > > > Linux based on whether it has a driver for it.
> > >
> > > Like Emil already said, the new bindings and the DT entries are solely
> > > introduced to support a proprietary out-of-tree module.
> > 
> > No. This new binding and the DT entries are solely introduced to
> > describe a device found in a number of SoCs, just like any other DT
> > binding we have.
> > 
> > > The current workflow when introducing new DT entries is the following:
> > > - upstream a driver that uses the entries
> > > - THEN add the new entries
> > 
> > And that's never been the preferred workflow, for *any* patches.
> 
> Actually it has. How else are you going to test that your driver really
> works? You've got to have both pieces before you can verify that they're
> both adequate. So the typical workflow is to:
> 
> 	1) define the bindings
> 	2) write a driver that implements the bindings
> 	3) add entries to device tree files
> 
> Usually it doesn't matter in which order you do the above because they
> are all part of the same patch series. But that's not what you're doing
> here. The more general problem here is that you're providing device tree
> content (and therefore ABI) that's based on a binding which has no
> upstream users. So you don't actually have a way of validating that what
> you merge is going to be an adequate description.
> 
> You're probably going to respond: "but DT describes hardware, so it must
> be known already, there won't be a need for changes". Unfortunately that
> is only partially true. We've had a number of occasions where it later
> turned out that a binding was in fact not an adequate description, and
> then we've had to jump through hoops in order to preserve backwards
> compatibility. That's already annoying enough if you've got in-tree
> users, but it's going to be even more painful if you start out with an
> out-of-tree user.
> 
> All of that said, you've got an Acked-by from Rob and that's about as
> good as it's going to get. So I'm not going to NAK this. But I will
> caution against this, because I don't think you're doing yourself any
> favours with this.
> 
> So perhaps the question that we should ask is this: what do you gain by
> merging this series? The fact remains that you don't have an upstream
> driver that implements this binding, so ultimately you're going to be
> carrying patches in some development tree anyway. Why not simply stash
> these patches into the same tree? That should be about the same amount
> of work for you and your users, but it has the advantage of not locking
> you into something that you may regret.

This is really a usability issue. I don't want the average Jane / Joe
to have to patch and recompile the DT in order to get the GPU running
on her / his distro of choice. The only thing that should be needed
would be to install an (or a couple of) extra package in order to get
everything running. Just like it's done for any other GPU out there,
or wifi, disregarding whether it's supported upstream or not.

Another nice thing is also that the mali bindings have been a huge
mess so far. If we have a common bindings, everything will just work
the same for all the platforms, and we can use the same driver there,
allowing us at least to consolidate the source code.

I'm not too afraid about getting something wrong. We already have code
working for two different vendors, and all the other GPUs already
described in DT (nvidia's, the Adreno) all have the same kind of
binding that we have.

Maxime

-- 
Maxime Ripard, Free Electrons
Embedded Linux and Kernel engineering
http://free-electrons.com
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 801 bytes
Desc: not available
URL: <http://lists.infradead.org/pipermail/linux-arm-kernel/attachments/20170222/d94edd35/attachment.sig>

[PATCH 0/8] ARM: sun8i: a33: Mali improvements

[Maxime Ripard]

Hi,

This serie is building on the recently merged bindings for the ARM Mali
Utgard GPU.

The two features that are supported with this serie are DVFS and the fbdev
support. The first one uses devfreq and is pretty standard, the only
addition being the generic OPP mechanism we have, plus some DT and Kconfig
patches.

Running on framebuffer is a bit more tedious, since we need to access the
CMA memory region size and base. This is quite trivial to do as well
through the memory-region bindings, but require to export a few symbols
along the way to make sure our module builds properly.

Let me know what you think,
Maxime

Maxime Ripard (8):
  ARM: sun8i: Fix the mali clock rate
  dt-bindings: gpu: mali: Add optional memory-region
  mm: cma: Export a few symbols
  drm/sun4i: Grab reserved memory region
  ARM: sun8i: a33: Add shared display memory pool
  dt-bindings: gpu: mali: Add optional OPPs
  ARM: sunxi: Select PM_OPP
  ARM: sun8i: a33: Add the Mali OPPs

 Documentation/devicetree/bindings/gpu/arm,mali-utgard.txt |  8 ++-
 arch/arm/boot/dts/sun8i-a23-a33.dtsi                      |  2 +-
 arch/arm/boot/dts/sun8i-a33.dtsi                          | 34 ++++++++-
 arch/arm/mach-sunxi/Kconfig                               |  1 +-
 drivers/base/dma-contiguous.c                             |  1 +-
 drivers/gpu/drm/sun4i/sun4i_drv.c                         | 19 ++--
 mm/cma.c                                                  |  2 +-
 7 files changed, 61 insertions(+), 6 deletions(-)

base-commit: a2138ce584d59571dd18a6cf3417cb90be7625d8
-- 
git-series 0.8.11