[PATCH v2 1/7] drm/exynos: add super device support

[PATCH v2 1/7] drm/exynos: add super device support

[Tomasz Figa]

[adding more people and MLs on Cc for further discussion]

On 04.04.2014 17:44, Inki Dae wrote:
> 2014-04-04 22:55 GMT+09:00 Tomasz Figa <t.figa@samsung.com>:
>> Hi Inki,
>>
>>
>> On 01.04.2014 14:37, Inki Dae wrote:
>>>
>>> This patch adds super device support to bind sub drivers
>>> using device tree.
>>>
>>> For this, you should add a super device node to each machine dt files
>>> like belows,
>>>
>>> In case of using MIPI-DSI,
>>>          display-subsystem {
>>>                  compatible = "samsung,exynos-display-subsystem";
>>>                  ports = <&fimd>, <&dsi>;
>>>          };
>>>
>>> In case of using DisplayPort,
>>>          display-subsystem {
>>>                  compatible = "samsung,exynos-display-subsystem";
>>>                  ports = <&fimd>, <&dp>;
>>>          };
>>>
>>> In case of using Parallel panel,
>>>          display-subsystem {
>>>                  compatible = "samsung,exynos-display-subsystem";
>>>                  ports = <&fimd>;
>>>          };
>>>
>>> And if you don't add connector device node to ports property,
>>> default parallel panel driver, exynos_drm_dpi module, will be used.
>>>
>>> ports property can have the following device nodes,
>>>          fimd, mixer, Image Enhancer, MIPI-DSI, eDP, LVDS Bridge, or HDMI
>>>
>>> With this patch, we can resolve the probing order issue without
>>> some global lists. So this patch also removes the unnecessary lists and
>>> stuff related to these lists.
>>
>>
>> I can see several problems with this approach:
>>
>> 1) It breaks compatibility with existing DT. After this patch it is no
>> longer possible to use old device trees and get a working DRM. However, in
>> my opinion, this requirement can be relaxed if we make sure that any users
>> are properly converted.
>>
>> 2) What happens if in Kconfig you disable a driver for a component that is
>
> I'm not sure what you meant but there wouldn't be no way that users
> *cannot disable* the driver for a component. The driver would be
> exynos_drm_drv, not separated module, and would always be built as
> long as users want to use *exynos drm driver*.

I think you don't understand what I mean. Let me show you an example:

You have a board with a DSI panel and also a HDMI output. So you have a 
supernode pointing to FIMD, DSI and HDMI.

Now, an user finds that he doesn't need HDMI in his system, so he turns 
off CONFIG_DRM_EXYNOS_HDMI. The supernode still points at hdmi node and 
Exynos DRM core will register it as a component, but HDMI driver is not 
available and will never probe, leading the whole Exynos DRM to never 
initialize. Is this a desired behavior?

>
>> listed in supernode? If I'm reading the code correctly, Exynos DRM will not
>
> And the only case a component isn't added to the list is when users
> disabled sub driver.

See above.

The code creating the list of components to wait for 
(exynos_drm_add_components()) doesn't seem to consider which sub-drivers 
are actually enabled in kernel config.

>> register, which is completely wrong. Users should be able to select which
>> drivers should be compiled into their kernels.
>
> So users are be able to select drivers they want to use, and will be
> compiled correctly. So no, the only thing users can disable is each
> sub driver, not core module.
>
>>
>> 3) Such approach leads to complete integration of all Exynos DRM drivers,
>> without possibility of loading some sub-drivers as modules. I know that
>> current driver design doesn't support it either, but if this series is
>
> No, current drm driver *must also be built* as one integrated single
> drm driver without super device approach.

As I wrote, I know that current design before this patch isn't modular 
either, but this is not my concern here. See below.

> So the super device approach
> *has no any effect on existing design*,  and what the super device
> approch tries to do is to resolve the probe order issue to sub drivers
> *without some codes specific to Exynos drm*.

My concern is that the supernode design is actually carving such broken 
non-modular design in stone. Remember that we are currently heading 
towards a fully multi-platform kernel where it is critical for such 
subsystems to be modular, because the same zImage is going to be running 
on completely different machines.

>
>> claimed to improve things, it should really do so.
>>
>> 4) Exactly the same can be achieved without changing the DT bindings at all.
>> In fact even without adding any new single property or node to DT. We
>> discussed this with Andrzej and Marek today and came to a solution in which
>> just by adding a little bit of code to Exynos DRM subdrivers, you could
>> guarantee correct registration of Exynos DRM platform and also get rid of
>> #ifdeffery in exynos_drm_drv.c. Andrzej will send an RFC after the weekend.
>
> I'm not sure but I had implemented below prototype codes for that, see
> the below link,
> https://git.kernel.org/cgit/linux/kernel/git/daeinki/drm-exynos.git/commit/?h=exynos-bridge-test&id=2860ad02d736e300034ddeabce3da92614922b4e
>
> I guess what you said would be similler approach.

Not exactly. The approach we found does mostly the same as componentized 
subsystem framework but without _any_ extra data in Device Tree. Just 
based on the list of subsystem sub-drivers that is already available to 
the master driver.

>
> And I still think the use of the component framework would be the best
> solution and *Linux generic way* for resolving the probe order issue
> without any specific codes. So I'm not advocating the compoent
> framework but I tend not to want to use specific codes.
>

I understand your concern. I also believe that generic frameworks should 
be reused wherever possible. However the componentized subsystem 
framework is a bit of overkill for this simple problem. Moreover, Device 
Tree is not a trash can where any data that can be thought of can be 
thrown as you go, but rather a hardware description that is supposed to 
be a stable ABI and needs to be well-thought. So, if something can be 
done in a way that doesn't require additional data, it's better to do it 
that way.

>>
>> 5) This series seems to break DPI display support with runtime PM enabled.
>> Universal C210 just hangs on second FIMD probe, after first one fails with
>> probe deferral. This needs more investigation, though.
>
> For -next, I never expect that pm operations would be operated
> perfactly. They are really not for product. Just adding new feature
> and drivers to -next, and then fixing them while in RC. And RC process
> would also be for it. Actually, I see pm interfaces of exynos_drm_dsi
> driver leads to break a single driver model because pm operations
> should be done at top level of exynos drm, and some codes Sean didn't
> fix. So such things are in my fix-todo-list. I tend to accept new
> features and drivers for -next as long as they have no big problem,
> And that is my style I maintain.

Unless it breaks so much that the system is unable to boot at all, which 
is the case. As I said, the system just hangs, like something would 
access hardware that is not properly initialized, e.g. without power 
domain or clocks enabled.

Anyway, I don't agree that regressions are allowed in linux-next, 
especially if found before applying a patch. Linux-next is a tree in 
which patches that are supposed to be ready to be pulled by Linus should 
be pushed. Of course nobody can spot all the regressions before the 
patches hitting -next and that's fine, as -next is an integration 
_testing_ tree. However even if already in -next, regressions should be 
fixed ASAP to minimize the number of fix patches needed during -rc period.

Best regards,
Tomasz

[PATCH v2 1/7] drm/exynos: add super device support

[Russell King - ARM Linux]

On Sat, Apr 05, 2014 at 07:32:50PM +0200, Tomasz Figa wrote:
> Not exactly. The approach we found does mostly the same as componentized  
> subsystem framework but without _any_ extra data in Device Tree. Just  
> based on the list of subsystem sub-drivers that is already available to  
> the master driver.

The existing approach is fundamentally broken.  Yes, your solution may
work for the probing case, but have you tried unbinding any of your
sub-drivers?

>From what I can see, that causes a kernel oops for one very simple reason -
you destroy stuff while it's still in use.  Let's look at an example:

struct platform_driver ipp_driver = {
        .probe          = ipp_probe,
        .remove         = ipp_remove,
        .driver         = {
                .name   = "exynos-drm-ipp",
                .owner  = THIS_MODULE,
                .pm     = &ipp_pm_ops,
        },
};

static int ipp_remove(struct platform_device *pdev)
{
        struct ipp_context *ctx = platform_get_drvdata(pdev);

        /* unregister sub driver */
        exynos_drm_subdrv_unregister(&ctx->subdrv);

        /* remove,destroy ipp idr */
        idr_destroy(&ctx->ipp_idr);
        idr_destroy(&ctx->prop_idr);

        mutex_destroy(&ctx->ipp_lock);
        mutex_destroy(&ctx->prop_lock);

        /* destroy command, event work queue */
        destroy_workqueue(ctx->cmd_workq);
        destroy_workqueue(ctx->event_workq);

        return 0;
}

int exynos_drm_subdrv_unregister(struct exynos_drm_subdrv *subdrv)
{
        if (!subdrv)
                return -EINVAL;

        list_del(&subdrv->list);

        return 0;
}

Oh dear, that destroys a whole pile of resources which could already
be in use without telling anything that it's about to do that.

I'm sure if I continue looking at the exynos stuff, it'll show similar
crap all over the place.

What you have now in mainline is not a solution.  It's a crappy bodge.

-- 
FTTC broadband for 0.8mile line: now at 9.7Mbps down 460kbps up... slowly
improving, and getting towards what was expected from it.

[PATCH v2 1/7] drm/exynos: add super device support

[Tomasz Figa]

On 05.04.2014 20:24, Russell King - ARM Linux wrote:
> On Sat, Apr 05, 2014 at 07:32:50PM +0200, Tomasz Figa wrote:
>> Not exactly. The approach we found does mostly the same as componentized
>> subsystem framework but without _any_ extra data in Device Tree. Just
>> based on the list of subsystem sub-drivers that is already available to
>> the master driver.
>
> The existing approach is fundamentally broken.  Yes, your solution may
> work for the probing case, but have you tried unbinding any of your
> sub-drivers?
>
>  From what I can see, that causes a kernel oops for one very simple reason -
> you destroy stuff while it's still in use.  Let's look at an example:
>
> struct platform_driver ipp_driver = {
>          .probe          = ipp_probe,
>          .remove         = ipp_remove,
>          .driver         = {
>                  .name   = "exynos-drm-ipp",
>                  .owner  = THIS_MODULE,
>                  .pm     = &ipp_pm_ops,
>          },
> };
>
> static int ipp_remove(struct platform_device *pdev)
> {
>          struct ipp_context *ctx = platform_get_drvdata(pdev);
>
>          /* unregister sub driver */
>          exynos_drm_subdrv_unregister(&ctx->subdrv);
>
>          /* remove,destroy ipp idr */
>          idr_destroy(&ctx->ipp_idr);
>          idr_destroy(&ctx->prop_idr);
>
>          mutex_destroy(&ctx->ipp_lock);
>          mutex_destroy(&ctx->prop_lock);
>
>          /* destroy command, event work queue */
>          destroy_workqueue(ctx->cmd_workq);
>          destroy_workqueue(ctx->event_workq);
>
>          return 0;
> }
>
> int exynos_drm_subdrv_unregister(struct exynos_drm_subdrv *subdrv)
> {
>          if (!subdrv)
>                  return -EINVAL;
>
>          list_del(&subdrv->list);
>
>          return 0;
> }
>
> Oh dear, that destroys a whole pile of resources which could already
> be in use without telling anything that it's about to do that.
>
> I'm sure if I continue looking at the exynos stuff, it'll show similar
> crap all over the place.
>
> What you have now in mainline is not a solution.  It's a crappy bodge.
>

Undoubtedly. Nobody here is trying to state the opposite.

Maybe my words have been misinterpreted, but all I'm suggesting here is 
that there is no need to add any new data to DT to solve the same issue 
to the same extent as componentized subsystem framework, at least in 
Exynos case.

Best regards,
Tomasz

[PATCH v2 1/7] drm/exynos: add super device support

[Russell King - ARM Linux]

On Sat, Apr 05, 2014 at 08:31:15PM +0200, Tomasz Figa wrote:
> Maybe my words have been misinterpreted, but all I'm suggesting here is  
> that there is no need to add any new data to DT to solve the same issue  
> to the same extent as componentized subsystem framework, at least in  
> Exynos case.

Right, so we seem to have agreement that what exynos is currently doing
is fundamentally broken.

So the next question is, how is it going to get fixed?

-- 
FTTC broadband for 0.8mile line: now at 9.7Mbps down 460kbps up... slowly
improving, and getting towards what was expected from it.

[PATCH v2 1/7] drm/exynos: add super device support

[Tomasz Figa]

On 05.04.2014 20:52, Russell King - ARM Linux wrote:
> On Sat, Apr 05, 2014 at 08:31:15PM +0200, Tomasz Figa wrote:
>> Maybe my words have been misinterpreted, but all I'm suggesting here is
>> that there is no need to add any new data to DT to solve the same issue
>> to the same extent as componentized subsystem framework, at least in
>> Exynos case.
>
> Right, so we seem to have agreement that what exynos is currently doing
> is fundamentally broken.
>
> So the next question is, how is it going to get fixed?
>

I believe this is a separate issue from what is being discussed in this 
thread, but anyway, I'd wait with this until Monday, so we could discuss 
this with Inki and Andrzej, who is the person responsible for Exynos DRM 
in my team. I was just passing by. ;)

Best regards,
Tomasz

[PATCH v2 1/7] drm/exynos: add super device support

[Inki Dae]

I'd be really happy if you gave me such opinions before pull request.

Anyway, below is my comments.

2014-04-06 2:32 GMT+09:00 Tomasz Figa <tomasz.figa@gmail.com>:
> [adding more people and MLs on Cc for further discussion]
>
>
> On 04.04.2014 17:44, Inki Dae wrote:
>>
>> 2014-04-04 22:55 GMT+09:00 Tomasz Figa <t.figa@samsung.com>:
>>>
>>> Hi Inki,
>>>
>>>
>>> On 01.04.2014 14:37, Inki Dae wrote:
>>>>
>>>>
>>>> This patch adds super device support to bind sub drivers
>>>> using device tree.
>>>>
>>>> For this, you should add a super device node to each machine dt files
>>>> like belows,
>>>>
>>>> In case of using MIPI-DSI,
>>>>          display-subsystem {
>>>>                  compatible = "samsung,exynos-display-subsystem";
>>>>                  ports = <&fimd>, <&dsi>;
>>>>          };
>>>>
>>>> In case of using DisplayPort,
>>>>          display-subsystem {
>>>>                  compatible = "samsung,exynos-display-subsystem";
>>>>                  ports = <&fimd>, <&dp>;
>>>>          };
>>>>
>>>> In case of using Parallel panel,
>>>>          display-subsystem {
>>>>                  compatible = "samsung,exynos-display-subsystem";
>>>>                  ports = <&fimd>;
>>>>          };
>>>>
>>>> And if you don't add connector device node to ports property,
>>>> default parallel panel driver, exynos_drm_dpi module, will be used.
>>>>
>>>> ports property can have the following device nodes,
>>>>          fimd, mixer, Image Enhancer, MIPI-DSI, eDP, LVDS Bridge, or
>>>> HDMI
>>>>
>>>> With this patch, we can resolve the probing order issue without
>>>> some global lists. So this patch also removes the unnecessary lists and
>>>> stuff related to these lists.
>>>
>>>
>>>
>>> I can see several problems with this approach:
>>>
>>> 1) It breaks compatibility with existing DT. After this patch it is no
>>> longer possible to use old device trees and get a working DRM. However,
>>> in
>>> my opinion, this requirement can be relaxed if we make sure that any
>>> users
>>> are properly converted.
>>>
>>> 2) What happens if in Kconfig you disable a driver for a component that
>>> is
>>
>>
>> I'm not sure what you meant but there wouldn't be no way that users
>> *cannot disable* the driver for a component. The driver would be
>> exynos_drm_drv, not separated module, and would always be built as
>> long as users want to use *exynos drm driver*.
>
>
> I think you don't understand what I mean. Let me show you an example:
>
> You have a board with a DSI panel and also a HDMI output. So you have a
> supernode pointing to FIMD, DSI and HDMI.
>
> Now, an user finds that he doesn't need HDMI in his system, so he turns off
> CONFIG_DRM_EXYNOS_HDMI. The supernode still points at hdmi node and Exynos
> DRM core will register it as a component, but HDMI driver is not available
> and will never probe, leading the whole Exynos DRM to never initialize. Is
> this a desired behavior?
>

Ok, now I understood. Your comment was not enough to me.

First of all, if I see component codes correctly then it seems your
misunderstanding. See the below comments,

>
>>
>>> listed in supernode? If I'm reading the code correctly, Exynos DRM will
>>> not
>>
>>
>> And the only case a component isn't added to the list is when users
>> disabled sub driver.
>
>
> See above.
>
> The code creating the list of components to wait for
> (exynos_drm_add_components()) doesn't seem to consider which sub-drivers are
> actually enabled in kernel config.
>

Are you sure?

exynos_drm_add_components() will try to attach components *added to
component_lists. And these components will be added by only
corresponding sub drivers to the component_lists and
master->components.

So in this case, if users disabled HDMI support then a commponent
object for HDMI sub driver doesn't exists in the component_lists and
master->components. This means that a component object for HDMI sub
driver *cannot be attached to master object*.

As a result, component_bind_add() will ignor component_bind call for
HDMI sub driver so HDMI driver will not be bounded. The only
components added by sub drivers will be bound, component->ops->bind().

For more understanding, it seems like you need to look into below codes,

static int exynos_drm_add_components(...)
{
        ...
        for (i == 0;; i++) {
                ...
                node = of_parse_phandle(np, "ports", i);
                ...
                ret = component_master_add_child(m, compare_of, node);
                ...
        }
}


And below codes,

int component_master_add_child(...)
{
        list_for_each_entry(c, &component_list, node) {
                if (c->master)
                        continue;

                if (compare(...)) {
                         component_attach_master(master, c);
                         ...
                }
        }
}

And below codes,

static void component_attach_master(master, c)
{
        c->master = master;
        list_add_tail(&c->master_node, &master->comonents);
}


As you can see above, the only components added to component_list can
be attached to master. And the important thing is that components can
be added by sub drivers to the component_list.

And below codes that actually tries to bind each sub drivers,

int component_bind_add(...)
{
        ....
        list_for_each_entry(c, &master->components, master_node) {
                   ret = component_bind(c, master, data);
                   ...
        }
        ...
}

The hdmi driver users disabled doesn't exist to master->components list.
How Exynos DRM cannot be initialized?

Of course, there may be my missing point, but I think I see them
correctly. Anyway I will test them tomorrow.

Thanks,
Inki Dae

>
>>> register, which is completely wrong. Users should be able to select which
>>> drivers should be compiled into their kernels.
>>
>>
>> So users are be able to select drivers they want to use, and will be
>> compiled correctly. So no, the only thing users can disable is each
>> sub driver, not core module.
>>
>>>
>>> 3) Such approach leads to complete integration of all Exynos DRM drivers,
>>> without possibility of loading some sub-drivers as modules. I know that
>>> current driver design doesn't support it either, but if this series is
>>
>>
>> No, current drm driver *must also be built* as one integrated single
>> drm driver without super device approach.
>
>
> As I wrote, I know that current design before this patch isn't modular
> either, but this is not my concern here. See below.
>
>
>> So the super device approach
>> *has no any effect on existing design*,  and what the super device
>> approch tries to do is to resolve the probe order issue to sub drivers
>> *without some codes specific to Exynos drm*.
>
>
> My concern is that the supernode design is actually carving such broken
> non-modular design in stone. Remember that we are currently heading towards
> a fully multi-platform kernel where it is critical for such subsystems to be
> modular, because the same zImage is going to be running on completely
> different machines.
>
>
>>
>>> claimed to improve things, it should really do so.
>>>
>>> 4) Exactly the same can be achieved without changing the DT bindings at
>>> all.
>>> In fact even without adding any new single property or node to DT. We
>>> discussed this with Andrzej and Marek today and came to a solution in
>>> which
>>> just by adding a little bit of code to Exynos DRM subdrivers, you could
>>> guarantee correct registration of Exynos DRM platform and also get rid of
>>> #ifdeffery in exynos_drm_drv.c. Andrzej will send an RFC after the
>>> weekend.
>>
>>
>> I'm not sure but I had implemented below prototype codes for that, see
>> the below link,
>>
>> https://git.kernel.org/cgit/linux/kernel/git/daeinki/drm-exynos.git/commit/?h=exynos-bridge-test&id=2860ad02d736e300034ddeabce3da92614922b4e
>>
>> I guess what you said would be similler approach.
>
>
> Not exactly. The approach we found does mostly the same as componentized
> subsystem framework but without _any_ extra data in Device Tree. Just based
> on the list of subsystem sub-drivers that is already available to the master
> driver.
>
>
>>
>> And I still think the use of the component framework would be the best
>> solution and *Linux generic way* for resolving the probe order issue
>> without any specific codes. So I'm not advocating the compoent
>> framework but I tend not to want to use specific codes.
>>
>
> I understand your concern. I also believe that generic frameworks should be
> reused wherever possible. However the componentized subsystem framework is a
> bit of overkill for this simple problem. Moreover, Device Tree is not a
> trash can where any data that can be thought of can be thrown as you go, but
> rather a hardware description that is supposed to be a stable ABI and needs
> to be well-thought. So, if something can be done in a way that doesn't
> require additional data, it's better to do it that way.
>
>
>>>
>>> 5) This series seems to break DPI display support with runtime PM
>>> enabled.
>>> Universal C210 just hangs on second FIMD probe, after first one fails
>>> with
>>> probe deferral. This needs more investigation, though.
>>
>>
>> For -next, I never expect that pm operations would be operated
>> perfactly. They are really not for product. Just adding new feature
>> and drivers to -next, and then fixing them while in RC. And RC process
>> would also be for it. Actually, I see pm interfaces of exynos_drm_dsi
>> driver leads to break a single driver model because pm operations
>> should be done at top level of exynos drm, and some codes Sean didn't
>> fix. So such things are in my fix-todo-list. I tend to accept new
>> features and drivers for -next as long as they have no big problem,
>> And that is my style I maintain.
>
>
> Unless it breaks so much that the system is unable to boot at all, which is
> the case. As I said, the system just hangs, like something would access
> hardware that is not properly initialized, e.g. without power domain or
> clocks enabled.
>
> Anyway, I don't agree that regressions are allowed in linux-next, especially
> if found before applying a patch. Linux-next is a tree in which patches that
> are supposed to be ready to be pulled by Linus should be pushed. Of course
> nobody can spot all the regressions before the patches hitting -next and
> that's fine, as -next is an integration _testing_ tree. However even if
> already in -next, regressions should be fixed ASAP to minimize the number of
> fix patches needed during -rc period.
>
>
> Best regards,
> Tomasz
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH v2 1/7] drm/exynos: add super device support

[Inki Dae]

2014-04-06 3:31 GMT+09:00, Tomasz Figa <tomasz.figa@gmail.com>:
>
>
> On 05.04.2014 20:24, Russell King - ARM Linux wrote:
>> On Sat, Apr 05, 2014 at 07:32:50PM +0200, Tomasz Figa wrote:
>>> Not exactly. The approach we found does mostly the same as componentized
>>> subsystem framework but without _any_ extra data in Device Tree. Just
>>> based on the list of subsystem sub-drivers that is already available to
>>> the master driver.
>>
>> The existing approach is fundamentally broken.  Yes, your solution may
>> work for the probing case, but have you tried unbinding any of your
>> sub-drivers?
>>
>>  From what I can see, that causes a kernel oops for one very simple reason
>> -
>> you destroy stuff while it's still in use.  Let's look at an example:
>>
>> struct platform_driver ipp_driver = {
>>          .probe          = ipp_probe,
>>          .remove         = ipp_remove,
>>          .driver         = {
>>                  .name   = "exynos-drm-ipp",
>>                  .owner  = THIS_MODULE,
>>                  .pm     = &ipp_pm_ops,
>>          },
>> };
>>
>> static int ipp_remove(struct platform_device *pdev)
>> {
>>          struct ipp_context *ctx = platform_get_drvdata(pdev);
>>
>>          /* unregister sub driver */
>>          exynos_drm_subdrv_unregister(&ctx->subdrv);
>>
>>          /* remove,destroy ipp idr */
>>          idr_destroy(&ctx->ipp_idr);
>>          idr_destroy(&ctx->prop_idr);
>>
>>          mutex_destroy(&ctx->ipp_lock);
>>          mutex_destroy(&ctx->prop_lock);
>>
>>          /* destroy command, event work queue */
>>          destroy_workqueue(ctx->cmd_workq);
>>          destroy_workqueue(ctx->event_workq);
>>
>>          return 0;
>> }
>>
>> int exynos_drm_subdrv_unregister(struct exynos_drm_subdrv *subdrv)
>> {
>>          if (!subdrv)
>>                  return -EINVAL;
>>
>>          list_del(&subdrv->list);
>>
>>          return 0;
>> }
>>
>> Oh dear, that destroys a whole pile of resources which could already
>> be in use without telling anything that it's about to do that.
>>
>> I'm sure if I continue looking at the exynos stuff, it'll show similar
>> crap all over the place.
>>
>> What you have now in mainline is not a solution.  It's a crappy bodge.
>>
>
> Undoubtedly. Nobody here is trying to state the opposite.
>
> Maybe my words have been misinterpreted, but all I'm suggesting here is
> that there is no need to add any new data to DT to solve the same issue
> to the same extent as componentized subsystem framework, at least in
> Exynos case.

As you may know, there is exynos chip that has two display
controllers. So it is possible to compose display pipe lines at same
time like below,

fimd0-----bridges..------Panel
fimd1-----maybe bridges..-----CPU Panel

How we can care such pipelines without component? Of course, it would
be possible somehow but I'm not sure there could be better and more
generic way than super device.

And super device will make existing dtb broken but this time it might
be good opportunity to more generic way before more users use existing
dt.

Thanks,
Inki Dae
>
> Best regards,
> Tomasz
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel
>

[PATCH v2 1/7] drm/exynos: add super device support

[Russell King - ARM Linux]

On Sun, Apr 06, 2014 at 01:21:24PM +0900, Inki Dae wrote:
> As you may know, there is exynos chip that has two display
> controllers. So it is possible to compose display pipe lines at same
> time like below,
> 
> fimd0-----bridges..------Panel
> fimd1-----maybe bridges..-----CPU Panel
> 
> How we can care such pipelines without component? Of course, it would
> be possible somehow but I'm not sure there could be better and more
> generic way than super device.
> 
> And super device will make existing dtb broken but this time it might
> be good opportunity to more generic way before more users use existing
> dt.

There has been a discussion ongoing about how to best represent display
(and capture) stuff in DT using graphs, which you may wish to consider
looking at.  That discussion never reached any kind of conclusion before
the merge window, but it does need to reach a conclusion for the next
window.

-- 
FTTC broadband for 0.8mile line: now at 9.7Mbps down 460kbps up... slowly
improving, and getting towards what was expected from it.

[PATCH v2 1/7] drm/exynos: add super device support

[Inki Dae]

Hi Russell,

2014-04-06 17:47 GMT+09:00 Russell King - ARM Linux <linux@arm.linux.org.uk>:
> On Sun, Apr 06, 2014 at 01:21:24PM +0900, Inki Dae wrote:
>> As you may know, there is exynos chip that has two display
>> controllers. So it is possible to compose display pipe lines at same
>> time like below,
>>
>> fimd0-----bridges..------Panel
>> fimd1-----maybe bridges..-----CPU Panel
>>
>> How we can care such pipelines without component? Of course, it would
>> be possible somehow but I'm not sure there could be better and more
>> generic way than super device.
>>
>> And super device will make existing dtb broken but this time it might
>> be good opportunity to more generic way before more users use existing
>> dt.
>
> There has been a discussion ongoing about how to best represent display
> (and capture) stuff in DT using graphs, which you may wish to consider

Right, there was my wrong comments. Actually, I had made a discussion
about it. For this, you can refer to below link,
http://www.spinics.net/lists/dri-devel/msg55555.html

My intension there was that let's use super device approach to resolve
the probe order issue of drm drivers for embedded system, and the
graphs to compose their display pipelines. And for this, we need
generic common structure to handle various bridge devices that could
be controlled by crtc and encoder/connector drivers.

I'd be happy to get your opinions about it.

> looking at.  That discussion never reached any kind of conclusion before
> the merge window, but it does need to reach a conclusion for the next
> window.

Yes. I still think the super device approach is a best solution, and I
think the only problem here is backward compatibility.  I have a feel
that we might resolve that issue easier then we think. I think all we
have to do for it is to implement some codes for backward
compatibility on top of super device approach.

This way, we could have a better design for the future drm drivers
without being tied to the backward compatibility.

Thanks,
Inki Dae

>
> --
> FTTC broadband for 0.8mile line: now at 9.7Mbps down 460kbps up... slowly
> improving, and getting towards what was expected from it.
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH v2 1/7] drm/exynos: add super device support

[Andrzej Hajda]

Hi Inki and Tomasz,

On 04/06/2014 05:15 AM, Inki Dae wrote:

(...)
> The code creating the list of components to wait for
> (exynos_drm_add_components()) doesn't seem to consider which sub-drivers are
> actually enabled in kernel config.
>
> Are you sure?
>
> exynos_drm_add_components() will try to attach components *added to
> component_lists. And these components will be added by only
> corresponding sub drivers to the component_lists and
> master->components.
>
> So in this case, if users disabled HDMI support then a commponent
> object for HDMI sub driver doesn't exists in the component_lists and
> master->components. This means that a component object for HDMI sub
> driver *cannot be attached to master object*.
>
> As a result, component_bind_add() will ignor component_bind call for
> HDMI sub driver so HDMI driver will not be bounded. The only
> components added by sub drivers will be bound, component->ops->bind().
>
> For more understanding, it seems like you need to look into below codes,
>
> static int exynos_drm_add_components(...)
> {
>         ...
>         for (i == 0;; i++) {
>                 ...
>                 node = of_parse_phandle(np, "ports", i);
>                 ...
>                 ret = component_master_add_child(m, compare_of, node);
>                 ...
>         }
> }

Hmm, In case HDMI driver is not present, HDMI device is not probed and
HDMI component is not added, so component_master_add_child returns
an error when it tries to add hdmi component and master is never brought
up, am I correct?

>
>
> And below codes,
>
> int component_master_add_child(...)
> {
>         list_for_each_entry(c, &component_list, node) {
>                 if (c->master)
>                         continue;
>
>                 if (compare(...)) {
>                          component_attach_master(master, c);
>                          ...
>                 }
>         }
> }
>
> And below codes,
>
> static void component_attach_master(master, c)
> {
>         c->master = master;
>         list_add_tail(&c->master_node, &master->comonents);
> }
>
>
> As you can see above, the only components added to component_list can
> be attached to master. And the important thing is that components can
> be added by sub drivers to the component_list.
>
> And below codes that actually tries to bind each sub drivers,
>
> int component_bind_add(...)
> {
>         ....
>         list_for_each_entry(c, &master->components, master_node) {
>                    ret = component_bind(c, master, data);
>                    ...
>         }
>         ...
> }
>
> The hdmi driver users disabled doesn't exist to master->components list.
> How Exynos DRM cannot be initialized?
>
> Of course, there may be my missing point, but I think I see them
> correctly. Anyway I will test them tomorrow.
>
> Thanks,
> Inki Dae
>
>>>> register, which is completely wrong. Users should be able to select which
>>>> drivers should be compiled into their kernels.
>>>
>>> So users are be able to select drivers they want to use, and will be
>>> compiled correctly. So no, the only thing users can disable is each
>>> sub driver, not core module.
>>>
>>>> 3) Such approach leads to complete integration of all Exynos DRM drivers,
>>>> without possibility of loading some sub-drivers as modules. I know that
>>>> current driver design doesn't support it either, but if this series is
>>>
>>> No, current drm driver *must also be built* as one integrated single
>>> drm driver without super device approach.
>>
>> As I wrote, I know that current design before this patch isn't modular
>> either, but this is not my concern here. See below.
>>
>>
>>> So the super device approach
>>> *has no any effect on existing design*,  and what the super device
>>> approch tries to do is to resolve the probe order issue to sub drivers
>>> *without some codes specific to Exynos drm*.
>>
>> My concern is that the supernode design is actually carving such broken
>> non-modular design in stone. Remember that we are currently heading towards
>> a fully multi-platform kernel where it is critical for such subsystems to be
>> modular, because the same zImage is going to be running on completely
>> different machines.
>>
>>
>>>> claimed to improve things, it should really do so.
>>>>
>>>> 4) Exactly the same can be achieved without changing the DT bindings at
>>>> all.
>>>> In fact even without adding any new single property or node to DT. We
>>>> discussed this with Andrzej and Marek today and came to a solution in
>>>> which
>>>> just by adding a little bit of code to Exynos DRM subdrivers, you could
>>>> guarantee correct registration of Exynos DRM platform and also get rid of
>>>> #ifdeffery in exynos_drm_drv.c. Andrzej will send an RFC after the
>>>> weekend.
>>>
>>> I'm not sure but I had implemented below prototype codes for that, see
>>> the below link,
>>>
>>> https://git.kernel.org/cgit/linux/kernel/git/daeinki/drm-exynos.git/commit/?h=exynos-bridge-test&id=2860ad02d736e300034ddeabce3da92614922b4e
>>>
>>> I guess what you said would be similler approach.
>>
>> Not exactly. The approach we found does mostly the same as componentized
>> subsystem framework but without _any_ extra data in Device Tree. Just based
>> on the list of subsystem sub-drivers that is already available to the master
>> driver.

I am not so sure which approach is better, anyway I hope to post RFC soon,
as some material for discussion.

>>
>>
>>> And I still think the use of the component framework would be the best
>>> solution and *Linux generic way* for resolving the probe order issue
>>> without any specific codes. So I'm not advocating the compoent
>>> framework but I tend not to want to use specific codes.
>>>
>> I understand your concern. I also believe that generic frameworks should be
>> reused wherever possible. However the componentized subsystem framework is a
>> bit of overkill for this simple problem. Moreover, Device Tree is not a
>> trash can where any data that can be thought of can be thrown as you go, but
>> rather a hardware description that is supposed to be a stable ABI and needs
>> to be well-thought. So, if something can be done in a way that doesn't
>> require additional data, it's better to do it that way.
>>
>>
>>>> 5) This series seems to break DPI display support with runtime PM
>>>> enabled.
>>>> Universal C210 just hangs on second FIMD probe, after first one fails
>>>> with
>>>> probe deferral. This needs more investigation, though.

If we are talking about the same issue, it is a problem of panel
initialization sequence on some
panels. I will post fix for it.


Regards
Andrzej

[PATCH v2 1/7] drm/exynos: add super device support

[Tomasz Figa]

Hi Andrzej,

On 07.04.2014 16:18, Andrzej Hajda wrote:
> Hi Inki and Tomasz,
>
> On 04/06/2014 05:15 AM, Inki Dae wrote:
>
> (...)
>> The code creating the list of components to wait for
>> (exynos_drm_add_components()) doesn't seem to consider which sub-drivers are
>> actually enabled in kernel config.
>>
>> Are you sure?
>>
>> exynos_drm_add_components() will try to attach components *added to
>> component_lists. And these components will be added by only
>> corresponding sub drivers to the component_lists and
>> master->components.
>>
>> So in this case, if users disabled HDMI support then a commponent
>> object for HDMI sub driver doesn't exists in the component_lists and
>> master->components. This means that a component object for HDMI sub
>> driver *cannot be attached to master object*.
>>
>> As a result, component_bind_add() will ignor component_bind call for
>> HDMI sub driver so HDMI driver will not be bounded. The only
>> components added by sub drivers will be bound, component->ops->bind().
>>
>> For more understanding, it seems like you need to look into below codes,
>>
>> static int exynos_drm_add_components(...)
>> {
>>          ...
>>          for (i == 0;; i++) {
>>                  ...
>>                  node = of_parse_phandle(np, "ports", i);
>>                  ...
>>                  ret = component_master_add_child(m, compare_of, node);
>>                  ...
>>          }
>> }
>
> Hmm, In case HDMI driver is not present, HDMI device is not probed and
> HDMI component is not added, so component_master_add_child returns
> an error when it tries to add hdmi component and master is never brought
> up, am I correct?
>

This is my thought here as well.

>>
>>
>> And below codes,
>>
>> int component_master_add_child(...)
>> {
>>          list_for_each_entry(c, &component_list, node) {
>>                  if (c->master)
>>                          continue;
>>
>>                  if (compare(...)) {
>>                           component_attach_master(master, c);
>>                           ...
>>                  }
>>          }
>> }
>>
>> And below codes,
>>
>> static void component_attach_master(master, c)
>> {
>>          c->master = master;
>>          list_add_tail(&c->master_node, &master->comonents);
>> }
>>
>>
>> As you can see above, the only components added to component_list can
>> be attached to master. And the important thing is that components can
>> be added by sub drivers to the component_list.
>>
>> And below codes that actually tries to bind each sub drivers,
>>
>> int component_bind_add(...)
>> {
>>          ....
>>          list_for_each_entry(c, &master->components, master_node) {
>>                     ret = component_bind(c, master, data);
>>                     ...
>>          }
>>          ...
>> }
>>
>> The hdmi driver users disabled doesn't exist to master->components list.
>> How Exynos DRM cannot be initialized?
>>
>> Of course, there may be my missing point, but I think I see them
>> correctly. Anyway I will test them tomorrow.
>>
>> Thanks,
>> Inki Dae
>>
>>>>> register, which is completely wrong. Users should be able to select which
>>>>> drivers should be compiled into their kernels.
>>>>
>>>> So users are be able to select drivers they want to use, and will be
>>>> compiled correctly. So no, the only thing users can disable is each
>>>> sub driver, not core module.
>>>>
>>>>> 3) Such approach leads to complete integration of all Exynos DRM drivers,
>>>>> without possibility of loading some sub-drivers as modules. I know that
>>>>> current driver design doesn't support it either, but if this series is
>>>>
>>>> No, current drm driver *must also be built* as one integrated single
>>>> drm driver without super device approach.
>>>
>>> As I wrote, I know that current design before this patch isn't modular
>>> either, but this is not my concern here. See below.
>>>
>>>
>>>> So the super device approach
>>>> *has no any effect on existing design*,  and what the super device
>>>> approch tries to do is to resolve the probe order issue to sub drivers
>>>> *without some codes specific to Exynos drm*.
>>>
>>> My concern is that the supernode design is actually carving such broken
>>> non-modular design in stone. Remember that we are currently heading towards
>>> a fully multi-platform kernel where it is critical for such subsystems to be
>>> modular, because the same zImage is going to be running on completely
>>> different machines.
>>>
>>>
>>>>> claimed to improve things, it should really do so.
>>>>>
>>>>> 4) Exactly the same can be achieved without changing the DT bindings at
>>>>> all.
>>>>> In fact even without adding any new single property or node to DT. We
>>>>> discussed this with Andrzej and Marek today and came to a solution in
>>>>> which
>>>>> just by adding a little bit of code to Exynos DRM subdrivers, you could
>>>>> guarantee correct registration of Exynos DRM platform and also get rid of
>>>>> #ifdeffery in exynos_drm_drv.c. Andrzej will send an RFC after the
>>>>> weekend.
>>>>
>>>> I'm not sure but I had implemented below prototype codes for that, see
>>>> the below link,
>>>>
>>>> https://git.kernel.org/cgit/linux/kernel/git/daeinki/drm-exynos.git/commit/?h=exynos-bridge-test&id=2860ad02d736e300034ddeabce3da92614922b4e
>>>>
>>>> I guess what you said would be similler approach.
>>>
>>> Not exactly. The approach we found does mostly the same as componentized
>>> subsystem framework but without _any_ extra data in Device Tree. Just based
>>> on the list of subsystem sub-drivers that is already available to the master
>>> driver.
>
> I am not so sure which approach is better, anyway I hope to post RFC soon,
> as some material for discussion.
>

Great. Looking forward for it.

>>>
>>>
>>>> And I still think the use of the component framework would be the best
>>>> solution and *Linux generic way* for resolving the probe order issue
>>>> without any specific codes. So I'm not advocating the compoent
>>>> framework but I tend not to want to use specific codes.
>>>>
>>> I understand your concern. I also believe that generic frameworks should be
>>> reused wherever possible. However the componentized subsystem framework is a
>>> bit of overkill for this simple problem. Moreover, Device Tree is not a
>>> trash can where any data that can be thought of can be thrown as you go, but
>>> rather a hardware description that is supposed to be a stable ABI and needs
>>> to be well-thought. So, if something can be done in a way that doesn't
>>> require additional data, it's better to do it that way.
>>>
>>>
>>>>> 5) This series seems to break DPI display support with runtime PM
>>>>> enabled.
>>>>> Universal C210 just hangs on second FIMD probe, after first one fails
>>>>> with
>>>>> probe deferral. This needs more investigation, though.
>
> If we are talking about the same issue, it is a problem of panel
> initialization sequence on some
> panels. I will post fix for it.

That's possible.

I'm not saying that it's a problem of this series itself, but it has 
been introduced/triggered by this series. My intention was to point that 
such issues need to be investigated before pushing the patches upstream.

Best regards,
Tomasz

[PATCH v2 1/7] drm/exynos: add super device support

[Inki Dae]

Hi Andrzej,

2014-04-07 23:18 GMT+09:00 Andrzej Hajda <a.hajda@samsung.com>:
> Hi Inki and Tomasz,
>
> On 04/06/2014 05:15 AM, Inki Dae wrote:
>
> (...)
>> The code creating the list of components to wait for
>> (exynos_drm_add_components()) doesn't seem to consider which sub-drivers are
>> actually enabled in kernel config.
>>
>> Are you sure?
>>
>> exynos_drm_add_components() will try to attach components *added to
>> component_lists. And these components will be added by only
>> corresponding sub drivers to the component_lists and
>> master->components.
>>
>> So in this case, if users disabled HDMI support then a commponent
>> object for HDMI sub driver doesn't exists in the component_lists and
>> master->components. This means that a component object for HDMI sub
>> driver *cannot be attached to master object*.
>>
>> As a result, component_bind_add() will ignor component_bind call for
>> HDMI sub driver so HDMI driver will not be bounded. The only
>> components added by sub drivers will be bound, component->ops->bind().
>>
>> For more understanding, it seems like you need to look into below codes,
>>
>> static int exynos_drm_add_components(...)
>> {
>>         ...
>>         for (i == 0;; i++) {
>>                 ...
>>                 node = of_parse_phandle(np, "ports", i);
>>                 ...
>>                 ret = component_master_add_child(m, compare_of, node);
>>                 ...
>>         }
>> }
>
> Hmm, In case HDMI driver is not present, HDMI device is not probed and
> HDMI component is not added, so component_master_add_child returns
> an error when it tries to add hdmi component and master is never brought
> up, am I correct?
>

Ok, after that,

ret = component_master_add(,..)
if (ret < 0)
         DRM_DEBUG_KMS("re-tried by last sub driver probed later.\n");

As you can see above, if component_master_add() is failed, return 0,
*not error*. And then component_add() called by other kms drivers,
late probing of hdmi or fimd probing - if there is any kms driver
enabled - will try to bring up master again by calling
try_to_bring_up_master() from beginning.

And if component_list is empty then it means that there is no any kms
driver enabled.

Do you still think in that case, exynos drm initialization will be failed?

Thanks,
Inki Dae

>>
>>
>> And below codes,
>>
>> int component_master_add_child(...)
>> {
>>         list_for_each_entry(c, &component_list, node) {
>>                 if (c->master)
>>                         continue;
>>
>>                 if (compare(...)) {
>>                          component_attach_master(master, c);
>>                          ...
>>                 }
>>         }
>> }
>>
>> And below codes,
>>
>> static void component_attach_master(master, c)
>> {
>>         c->master = master;
>>         list_add_tail(&c->master_node, &master->comonents);
>> }
>>
>>
>> As you can see above, the only components added to component_list can
>> be attached to master. And the important thing is that components can
>> be added by sub drivers to the component_list.
>>
>> And below codes that actually tries to bind each sub drivers,
>>
>> int component_bind_add(...)
>> {
>>         ....
>>         list_for_each_entry(c, &master->components, master_node) {
>>                    ret = component_bind(c, master, data);
>>                    ...
>>         }
>>         ...
>> }
>>
>> The hdmi driver users disabled doesn't exist to master->components list.
>> How Exynos DRM cannot be initialized?
>>
>> Of course, there may be my missing point, but I think I see them
>> correctly. Anyway I will test them tomorrow.
>>
>> Thanks,
>> Inki Dae
>>
>>>>> register, which is completely wrong. Users should be able to select which
>>>>> drivers should be compiled into their kernels.
>>>>
>>>> So users are be able to select drivers they want to use, and will be
>>>> compiled correctly. So no, the only thing users can disable is each
>>>> sub driver, not core module.
>>>>
>>>>> 3) Such approach leads to complete integration of all Exynos DRM drivers,
>>>>> without possibility of loading some sub-drivers as modules. I know that
>>>>> current driver design doesn't support it either, but if this series is
>>>>
>>>> No, current drm driver *must also be built* as one integrated single
>>>> drm driver without super device approach.
>>>
>>> As I wrote, I know that current design before this patch isn't modular
>>> either, but this is not my concern here. See below.
>>>
>>>
>>>> So the super device approach
>>>> *has no any effect on existing design*,  and what the super device
>>>> approch tries to do is to resolve the probe order issue to sub drivers
>>>> *without some codes specific to Exynos drm*.
>>>
>>> My concern is that the supernode design is actually carving such broken
>>> non-modular design in stone. Remember that we are currently heading towards
>>> a fully multi-platform kernel where it is critical for such subsystems to be
>>> modular, because the same zImage is going to be running on completely
>>> different machines.
>>>
>>>
>>>>> claimed to improve things, it should really do so.
>>>>>
>>>>> 4) Exactly the same can be achieved without changing the DT bindings at
>>>>> all.
>>>>> In fact even without adding any new single property or node to DT. We
>>>>> discussed this with Andrzej and Marek today and came to a solution in
>>>>> which
>>>>> just by adding a little bit of code to Exynos DRM subdrivers, you could
>>>>> guarantee correct registration of Exynos DRM platform and also get rid of
>>>>> #ifdeffery in exynos_drm_drv.c. Andrzej will send an RFC after the
>>>>> weekend.
>>>>
>>>> I'm not sure but I had implemented below prototype codes for that, see
>>>> the below link,
>>>>
>>>> https://git.kernel.org/cgit/linux/kernel/git/daeinki/drm-exynos.git/commit/?h=exynos-bridge-test&id=2860ad02d736e300034ddeabce3da92614922b4e
>>>>
>>>> I guess what you said would be similler approach.
>>>
>>> Not exactly. The approach we found does mostly the same as componentized
>>> subsystem framework but without _any_ extra data in Device Tree. Just based
>>> on the list of subsystem sub-drivers that is already available to the master
>>> driver.
>
> I am not so sure which approach is better, anyway I hope to post RFC soon,
> as some material for discussion.
>

Feel free to post RFC but as I mentioned eailer, it'd be good way to
use existing infrastructure, not specific one.

Thanks,
Inki Dae

>>>
>>>
>>>> And I still think the use of the component framework would be the best
>>>> solution and *Linux generic way* for resolving the probe order issue
>>>> without any specific codes. So I'm not advocating the compoent
>>>> framework but I tend not to want to use specific codes.
>>>>
>>> I understand your concern. I also believe that generic frameworks should be
>>> reused wherever possible. However the componentized subsystem framework is a
>>> bit of overkill for this simple problem. Moreover, Device Tree is not a
>>> trash can where any data that can be thought of can be thrown as you go, but
>>> rather a hardware description that is supposed to be a stable ABI and needs
>>> to be well-thought. So, if something can be done in a way that doesn't
>>> require additional data, it's better to do it that way.
>>>
>>>
>>>>> 5) This series seems to break DPI display support with runtime PM
>>>>> enabled.
>>>>> Universal C210 just hangs on second FIMD probe, after first one fails
>>>>> with
>>>>> probe deferral. This needs more investigation, though.
>
> If we are talking about the same issue, it is a problem of panel
> initialization sequence on some
> panels. I will post fix for it.
>
>
> Regards
> Andrzej
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel

[PATCH v2 1/7] drm/exynos: add super device support

[Tomasz Figa]

On 07.04.2014 17:16, Inki Dae wrote:
> Hi Andrzej,
>
> 2014-04-07 23:18 GMT+09:00 Andrzej Hajda <a.hajda@samsung.com>:
>> Hi Inki and Tomasz,
>>
>> On 04/06/2014 05:15 AM, Inki Dae wrote:
>>
>> (...)
>>> The code creating the list of components to wait for
>>> (exynos_drm_add_components()) doesn't seem to consider which sub-drivers are
>>> actually enabled in kernel config.
>>>
>>> Are you sure?
>>>
>>> exynos_drm_add_components() will try to attach components *added to
>>> component_lists. And these components will be added by only
>>> corresponding sub drivers to the component_lists and
>>> master->components.
>>>
>>> So in this case, if users disabled HDMI support then a commponent
>>> object for HDMI sub driver doesn't exists in the component_lists and
>>> master->components. This means that a component object for HDMI sub
>>> driver *cannot be attached to master object*.
>>>
>>> As a result, component_bind_add() will ignor component_bind call for
>>> HDMI sub driver so HDMI driver will not be bounded. The only
>>> components added by sub drivers will be bound, component->ops->bind().
>>>
>>> For more understanding, it seems like you need to look into below codes,
>>>
>>> static int exynos_drm_add_components(...)
>>> {
>>>          ...
>>>          for (i == 0;; i++) {
>>>                  ...
>>>                  node = of_parse_phandle(np, "ports", i);
>>>                  ...
>>>                  ret = component_master_add_child(m, compare_of, node);
>>>                  ...
>>>          }
>>> }
>>
>> Hmm, In case HDMI driver is not present, HDMI device is not probed and
>> HDMI component is not added, so component_master_add_child returns
>> an error when it tries to add hdmi component and master is never brought
>> up, am I correct?
>>
>
> Ok, after that,
>
> ret = component_master_add(,..)
> if (ret < 0)
>           DRM_DEBUG_KMS("re-tried by last sub driver probed later.\n");
>
> As you can see above, if component_master_add() is failed, return 0,
> *not error*. And then component_add() called by other kms drivers,
> late probing of hdmi or fimd probing - if there is any kms driver
> enabled - will try to bring up master again by calling
> try_to_bring_up_master() from beginning.
>
> And if component_list is empty then it means that there is no any kms
> driver enabled.
>
> Do you still think in that case, exynos drm initialization will be failed?

There will be no HDMI driver to call component_add(), because HDMI sub 
driver will be disabled in kernel config and any previous 
component_master_add_child() for the phandle pointing to HDMI node will 
wail, because such component will never be registered.

The complete failure path is as follows:

exynos_drm_platform_probe()
	component_master_add()
		try_to_bring_up_master()
			exynos_drm_add_components()
				// phandle to HDMI node
				component_master_add_child()
				= -ENXIO
			= -ENXIO
		= 0 // but no call to master->ops->bind(master->dev);
	= 0
= 0 // but Exynos DRM platform is not registered yet

Now if any other late-probed sub-driver comes, the sequence will be as 
follows (let's take FIMD as an example):

fimd_probe()
	component_add()
		try_to_bring_up_masters()
			try_to_bring_up_master()
				exynos_drm_add_components()
					// phandle to HDMI node
					component_master_add_child()
					= -ENXIO
				= -ENXIO
			= 0 // but no call to 	
			    // master->ops->bind(master->dev);
		= 0
	= 0
= 0 // but Exynos DRM platform still not registered

And the same for any late-probed driver, unless HDMI drivers loads, but 
there is no HDMI driver, because it is disabled in kernel config and not 
compiled in.

Best regards,
Tomasz

[PATCH v2 1/7] drm/exynos: add super device support

[Inki Dae]

2014-04-08 0:40 GMT+09:00, Tomasz Figa <tomasz.figa@gmail.com>:
> On 07.04.2014 17:16, Inki Dae wrote:
>> Hi Andrzej,
>>
>> 2014-04-07 23:18 GMT+09:00 Andrzej Hajda <a.hajda@samsung.com>:
>>> Hi Inki and Tomasz,
>>>
>>> On 04/06/2014 05:15 AM, Inki Dae wrote:
>>>
>>> (...)
>>>> The code creating the list of components to wait for
>>>> (exynos_drm_add_components()) doesn't seem to consider which sub-drivers
>>>> are
>>>> actually enabled in kernel config.
>>>>
>>>> Are you sure?
>>>>
>>>> exynos_drm_add_components() will try to attach components *added to
>>>> component_lists. And these components will be added by only
>>>> corresponding sub drivers to the component_lists and
>>>> master->components.
>>>>
>>>> So in this case, if users disabled HDMI support then a commponent
>>>> object for HDMI sub driver doesn't exists in the component_lists and
>>>> master->components. This means that a component object for HDMI sub
>>>> driver *cannot be attached to master object*.
>>>>
>>>> As a result, component_bind_add() will ignor component_bind call for
>>>> HDMI sub driver so HDMI driver will not be bounded. The only
>>>> components added by sub drivers will be bound, component->ops->bind().
>>>>
>>>> For more understanding, it seems like you need to look into below
>>>> codes,
>>>>
>>>> static int exynos_drm_add_components(...)
>>>> {
>>>>          ...
>>>>          for (i == 0;; i++) {
>>>>                  ...
>>>>                  node = of_parse_phandle(np, "ports", i);
>>>>                  ...
>>>>                  ret = component_master_add_child(m, compare_of, node);
>>>>                  ...
>>>>          }
>>>> }
>>>
>>> Hmm, In case HDMI driver is not present, HDMI device is not probed and
>>> HDMI component is not added, so component_master_add_child returns
>>> an error when it tries to add hdmi component and master is never brought
>>> up, am I correct?
>>>
>>
>> Ok, after that,
>>
>> ret = component_master_add(,..)
>> if (ret < 0)
>>           DRM_DEBUG_KMS("re-tried by last sub driver probed later.\n");
>>
>> As you can see above, if component_master_add() is failed, return 0,
>> *not error*. And then component_add() called by other kms drivers,
>> late probing of hdmi or fimd probing - if there is any kms driver
>> enabled - will try to bring up master again by calling
>> try_to_bring_up_master() from beginning.
>>
>> And if component_list is empty then it means that there is no any kms
>> driver enabled.
>>
>> Do you still think in that case, exynos drm initialization will be
>> failed?
>
> There will be no HDMI driver to call component_add(), because HDMI sub
> driver will be disabled in kernel config and any previous
> component_master_add_child() for the phandle pointing to HDMI node will
> wail, because such component will never be registered.
>
> The complete failure path is as follows:
>
> exynos_drm_platform_probe()
> 	component_master_add()
> 		try_to_bring_up_master()
> 			exynos_drm_add_components()
> 				// phandle to HDMI node
> 				component_master_add_child()
> 				= -ENXIO
> 			= -ENXIO
> 		= 0 // but no call to master->ops->bind(master->dev);
> 	= 0
> = 0 // but Exynos DRM platform is not registered yet
>
> Now if any other late-probed sub-driver comes, the sequence will be as
> follows (let's take FIMD as an example):
>
> fimd_probe()
> 	component_add()
> 		try_to_bring_up_masters()
> 			try_to_bring_up_master()
> 				exynos_drm_add_components()
> 					// phandle to HDMI node
> 					component_master_add_child()
> 					= -ENXIO
> 				= -ENXIO
> 			= 0 // but no call to 	
> 			    // master->ops->bind(master->dev);
> 		= 0
> 	= 0
> = 0 // but Exynos DRM platform still not registered
>
> And the same for any late-probed driver, unless HDMI drivers loads, but
> there is no HDMI driver, because it is disabled in kernel config and not
> compiled in.
>

Ah, right. I missed that it gets phandle to hdmi, and compares the
phandle to one in component_list.
Yes, in this case, exynos drm driver will be failed because
component_list never has a component object for hdmi.
So this shouldn't return as long as there is any component - in
component_list - that can be attached to master.

Hm.. the foot of the candle is dark.

For this, I fixed and tested it like below,
for (i = 0;; i++) {
        node = of_parse_phandle(np, "ports", i);
        ...
        ret = component_master_add_child(...);
        ...
        if (!ret)
                attached_cnt++;
}

if (!attached_cnt)
        return -ENXIO;
...

And one more thing, it seems that we could  resolve dt broken issue
easily by getting phandle to existing device node directly.

Thanks,
Inki Dae

> Best regards,
> Tomasz
> _______________________________________________
> dri-devel mailing list
> dri-devel at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/dri-devel
>