[PATCH v2 12/26] drm/exynos: Split manager/display/subdrv

[Sean Paul]

On Tue, Oct 22, 2013 at 1:30 AM, Inki Dae <inki.dae at samsung.com> wrote:
>
>
>> -----Original Message-----
>> From: Sean Paul [mailto:seanpaul at chromium.org]
>> Sent: Tuesday, October 22, 2013 6:18 AM
>> To: Inki Dae
>> Cc: dri-devel; Dave Airlie; Tomasz Figa; Stéphane Marchesin
>> Subject: Re: [PATCH v2 12/26] drm/exynos: Split manager/display/subdrv
>>
>> On Mon, Oct 21, 2013 at 10:46 AM, Sean Paul <seanpaul at chromium.org> wrote:
>> > On Thu, Oct 17, 2013 at 10:31 PM, Inki Dae <inki.dae at samsung.com> wrote:
>> >>
>> >>
>> >>> -----Original Message-----
>> >>> From: Sean Paul [mailto:seanpaul at chromium.org]
>> >>> Sent: Thursday, October 17, 2013 11:37 PM
>> >>> To: Inki Dae
>> >>> Cc: dri-devel; Dave Airlie; Tomasz Figa; Stéphane Marchesin
>> >>> Subject: Re: [PATCH v2 12/26] drm/exynos: Split manager/display/subdrv
>> >>>
>> >>> On Thu, Oct 17, 2013 at 4:21 AM, Inki Dae <inki.dae at samsung.com>
> wrote:
>> >>> >
>> >>> >
>> >>> >> -----Original Message-----
>> >>> >> From: Sean Paul [mailto:seanpaul at chromium.org]
>> >>> >> Sent: Thursday, October 17, 2013 4:27 AM
>> >>> >> To: dri-devel at lists.freedesktop.org; inki.dae at samsung.com
>> >>> >> Cc: airlied at linux.ie; tomasz.figa at gmail.com; marcheu at chromium.org;
>> Sean
>> >>> >> Paul
>> >>> >> Subject: [PATCH v2 12/26] drm/exynos: Split manager/display/subdrv
>> >>> >>
>> >>> >> This patch splits display and manager from subdrv. The result is
>> that
>> >>> >> crtc functions can directly call into manager callbacks and encoder
>> >>> >> functions can directly call into display callbacks. This will allow
>> >>> >> us to remove the exynos_drm_hdmi shim and support mixer/hdmi &
>> fimd/dp
>> >>> >> with common code.
>> >>> >>
>> >>> >> Signed-off-by: Sean Paul <seanpaul at chromium.org>
>> >>> >> ---
>> >>> >>
>> >>> >> Changes in v2:
>> >>> >>       - Pass display into display_ops instead of context
>> >>> >
>> >>> > Sorry but it seems like more reasonable to pass device object into
>> >>> > display_ops and manager_ops.
>> >>> >
>> >>>
>> >>>
>> >>> So you've changed your mind from when you said the following?
>> >>>
>> >>> >>> manager->ops->xxx(manager, ...);
>> >>> >>> display->ops->xxx(display, ...);
>> >>> >>>
>> >>> >>> Agree.
>> >>>
>> >>
>> >>
>> >> True. Before that, My comment was to pass device object into
>> display_ops and
>> >> manager_ops, and then you said the good solution is to pass manager and
>> >> display to each driver. At that time, I thought no matter how the
>> callback
>> >> is called if the framework doesn't call callbacks of each driver with
>> ctx.
>> >> So I agreed.
>> >>
>> >>
>> >>> It would have been nice if you had changed your mind *before* I
>> >>> reworked everything. At any rate, I think it's still the right thing
>> >>> to do.
>> >>
>> >> Really sorry about that. And I will add new patch for it so you don't
>> need
>> >> to concern about that.
>> >>
>> >>>
>> >>>
>> >>> > I'm not sure but display_ops could be implemented in other framework
>> >>> based
>> >>> > driver such as CDF based lcd panel driver. So if you pass display -
>> it's
>> >>> > specific to exynos drm framework - into display_ops, the other
>> framework
>> >>> > based driver should include specific exynos drm header.
>> >>> >
>> >>>
>> >>> AFAIK, CDF will not land in its current separate-from-drm form, we
>> >>> don't need to worry about this. Furthermore, these ops should just go
>> >>> away and become drm_crtc/drm_encoder/drm_connector funcs anyways.
>> >>>
>> >>
>> >> Can you assure the display_ops never implemented in other framework
>> based
>> >> driver, not CDF? At any rate, I think all possibilities should be
>> opened.
>> >>
>> >
>> > I don't think we should let an RFC framework make the code more
>> > complicated for unclear benefit. By removing manager/display entirely,
>> > we can get rid of a *lot* of other code that is basically just
>> > plumbing drm hooks (exynos_drm_connector is a good example).
>> >
>>
>> I hacked this up today to prove it out. Check out the top 5 commits in
>> https://github.com/crseanpaul/exynos-drm-next/commits/linux-next-exynos-
>> staging.
>> It removes exynos_drm_connector in favor of just implementing
>> drm_connector directly. This same treatment should be done for
>> exynos_drm_encoder and exynos_drm_crtc, but I didn't get around to
>> doing this.
>>
>> As you can see, it cuts out a lot of code and removes an entire
>> abstraction layer. Much nicer :)
>>
>
> It seems that you implements connector in each device driver. Can't they be
> combined as common spot, exynos_connector, again to avoid codes from
> duplicated? :)

There's nothing of substance being duplicated. In fact, by getting rid
of the exynos_drm_connector layer, we deleted 150 lines. If you really
take a look at exynos_drm_connector, it's not doing anything useful.
All it does is translate the drm callbacks into display callbacks, so
I think it's much better to just implement the drm callbacks directly.

There are a bunch of real bugs that we've found as a result of having
these abstraction layers. Take, for example, dpms. Before this
patchset, dpms for fimd was being tracked separately in fimd driver,
exynos_drm_encoder, exynos_drm_crtc, and exynos_drm_connector.
Furthermore, during suspend, only fimd driver's dpms state was
updated, so the others were incorrect. There was also this weird
gymnastics that had to happen when dpms was changed in the encoder
since it had to walk up to the connector level to change its dpms
state. If fimd just directly implemented
drm_crtc/drm_encoder/drm_connector (before dp was moved in), this
problem wouldn't exist. The same goes for HDMI/mixer.

Take a look at exynos_drm_encoder.c  in my tree
(https://github.com/crseanpaul/exynos-drm-next/blob/linux-next-exynos-staging/drivers/gpu/drm/exynos/exynos_drm_encoder.c),
what does it do that's useful to abstract? All that it does is just
call display ops, it's completely useless. The same is true for
exynos_drm_connector, it's just dead weight. There is some useful
stuff in exynos_drm_crtc for page flipping, that would be better
served as a helper library, though.

> The abstraction layer you mentioned also means a common spot.
> Another one, you patch also makes each sub driver have strongly dependency
> of drm framework. So how we can support existing backlight and lcd class
> based lcd panel drivers if the connector is implemented in each device
> driver later?  the drm header files should be included in
> drivers/video/backlight/xxx_lcd.c?
>

drm_bridge or drm_panel seem like good candidates for this.


> And, I will introduce a new framework to support existing lcd panel drivers
> and display bus drivers soon; as of now for Exynos drm, and the framework is
> being tested internally. With this framework, encoder and connector will be
> created when lcd panel or display bus driver such as eDP is probed: it
> doesn’t really need to create encoder and connector in advance if lcd panel
> or display bus driver isn't probed yet. Regardless of crtc, and encoder and
> connector creation order, when last one is created, crtc and connector will
> be connected each other. And exynos_drm_display could be implemented in
> other frameworks if we have common structure for display device driver. And
> also the framework will support lvds driver according to Linux device driver
> model.
>

I don't really follow what you're trying to do here, but I think we
should be moving in the direction of fewer abstractions in the exynos
driver, not more :)

Sean



> Thanks,
> Inki Dae
>
>> Sean
>>
>> >>>
>> >>> > And another one, the patch 6 passes manager object to manager_ops,
>> and
>> >>> for
>> >>> > this, you made the manager object to be set to driver data;
>> >>> > platform_set_drvdata(pdev, &manager). That isn't reasonable.
>> Generally,
>> >>> > driver_data would point to device driver's context object.
>> >>> >
>> >>>
>> >>> I'm not sure why this isn't reasonable, but it's a moot point. The
>> >>> driver data is only used up until we get rid of the pm ops, it needn't
>> >>> be set at all once things go through dpms.
>> >>>
>> >>
>> >> Generally, device drivers can call its own internal functions, and they
>> will
>> >> call that functions with ctx. However, if you set manager to
>> driver_data
>> >> then that functions should be called with manager object and also
>> internally
>> >> that functions should get ctx from the manager object. What is the
>> purpose
>> >> of manager? Do you think it's reasonable?
>> >>
>> >
>> > So, to avoid setting the manager as the drvdata, we could implement
>> > something like fimd_dpms_ctx(ctx, mode) that takes ctx and the manager
>> > callback calls it fimd_dpms(mgr, mode) { ctx = mgr->ctx;
>> > fimd_dpms_ctx(ctx, mode); }. Alternatively, you can save a pointer to
>> > mgr in ctx, but that creates a circular link between the two. IMO,
>> > both of those solutions suck :)
>> >
>> > I'd much rather just set drvdata to the manager and call the hook
>> > directly. Like I said earlier, this is just a temporary step since we
>> > remove these pm ops later in the patch series.
>> >
>> > Sean
>> >
>> >
>> >> Anyway, I'd like to say really sorry about inconvenient again. So I
>> will fix
>> >> it.
>> >>
>> >> Thanks,
>> >> Inki Dae
>> >>
>> >>> Sean
>> >>>
>> >>>
>> >>> > Thanks,
>> >>> > Inki Dae
>> >>> >
>> >>
>