2018-04-23 21:40:21

by David Lechner

[permalink] [raw]
Subject: Re: [RFC work-in-progress 0/7] of: platform: use early platform routines instead of OF_DECLARE

FYI: It looks like the CC for Stephen and Arnd was messed up, so I
fixed.

On 04/23/2018 01:38 PM, Bartosz Golaszewski wrote:
> From: Bartosz Golaszewski <[email protected]>
>
> Hi David, Sekhar,
>
> since platform devices are generally considered more desirable than
> CLK_OF_DECLARE, TIMER_OF_DECLARE etc. and we need to figure out how to
> handle the clocks that need to be initialized early in the boot
> process on DaVinci, I thought that I could give the early_platform
> mechanism a try.
>
> This API is only used on one architecture (sh) but seems to work just
> fine on ARM. It allows to register early platform drivers and then
> probe them early in the boot process. So far only machine code is
> supported but with a bit of hacking I was able to probe a DT device.
>
> This is a very dirty and far-from-upstream proof of concept that allows
> to probe the (so far dummy) davinci timer platform device during the
> call to init_time (from machine_desc).
>
> The idea is to have a special compatible fallback string: "earlydev"
> that similarily to "syscon" would be added to device nodes that need
> early probing. Then we'd call the of_early_platform_populate()
> function that would find all compatible nodes and populate them
> long before all the "normal" nodes.

FWIW, "earlydev" sounds like a driver implementation detail, so not
something that should be included in the device tree. We only need
this because Linux needs a clocksource early on, but that doesn't
mean that all device tree users need to do the same.

I'm sure it makes things easier for a proof of concept though. :-)

>
> This would allow us to make the davinci timer a normal platform device
> and possibly also probe the psc and pll drivers earlier than we do now.
>
> The early platform API even allows us to check if we're being probed
> early in probe() so we can possibly probe the driver twice if needed:
> only doing the critical stuff first and then completing the process
> later.
>
> If you think this is a good idea, I would like to continue on that
> and eventually make it an alternative to OF_DECLARE macros.
>
> For a quick conversion of the davinci timer to a platform driver
> I image we'd need to use platform data lookup that would be passed
> to of_early_platform_populate().

On the surface, it certainly sounds like a good idea to me. Do we have
access to struct device of the platform device when using this early
platform device? I remember when I was working on the clock drivers, I
tried registering a platform device in the init_time callback but the
kernel crashed because kobj stuff was not initialized yet. I'm guessing
that the early platform device somehow works around this.



2018-04-24 07:28:18

by Bartosz Golaszewski

[permalink] [raw]
Subject: Re: [RFC work-in-progress 0/7] of: platform: use early platform routines instead of OF_DECLARE

2018-04-23 23:38 GMT+02:00 David Lechner <[email protected]>:
> FYI: It looks like the CC for Stephen and Arnd was messed up, so I
> fixed.
>

Thanks!

> On 04/23/2018 01:38 PM, Bartosz Golaszewski wrote:
>>
>> From: Bartosz Golaszewski <[email protected]>
>>
>> Hi David, Sekhar,
>>
>> since platform devices are generally considered more desirable than
>> CLK_OF_DECLARE, TIMER_OF_DECLARE etc. and we need to figure out how to
>> handle the clocks that need to be initialized early in the boot
>> process on DaVinci, I thought that I could give the early_platform
>> mechanism a try.
>>
>> This API is only used on one architecture (sh) but seems to work just
>> fine on ARM. It allows to register early platform drivers and then
>> probe them early in the boot process. So far only machine code is
>> supported but with a bit of hacking I was able to probe a DT device.
>>
>> This is a very dirty and far-from-upstream proof of concept that allows
>> to probe the (so far dummy) davinci timer platform device during the
>> call to init_time (from machine_desc).
>>
>> The idea is to have a special compatible fallback string: "earlydev"
>> that similarily to "syscon" would be added to device nodes that need
>> early probing. Then we'd call the of_early_platform_populate()
>> function that would find all compatible nodes and populate them
>> long before all the "normal" nodes.
>
>
> FWIW, "earlydev" sounds like a driver implementation detail, so not
> something that should be included in the device tree. We only need
> this because Linux needs a clocksource early on, but that doesn't
> mean that all device tree users need to do the same.
>
> I'm sure it makes things easier for a proof of concept though. :-)
>

We already have "syscon" which too is more an implementation detail
than HW description. I should have probably Cc'ed Rob Herring. I'll do
it with a more polished version I should have today.

>>
>> This would allow us to make the davinci timer a normal platform device
>> and possibly also probe the psc and pll drivers earlier than we do now.
>>
>> The early platform API even allows us to check if we're being probed
>> early in probe() so we can possibly probe the driver twice if needed:
>> only doing the critical stuff first and then completing the process
>> later.
>>
>> If you think this is a good idea, I would like to continue on that
>> and eventually make it an alternative to OF_DECLARE macros.
>>
>> For a quick conversion of the davinci timer to a platform driver
>> I image we'd need to use platform data lookup that would be passed
>> to of_early_platform_populate().
>
>
> On the surface, it certainly sounds like a good idea to me. Do we have
> access to struct device of the platform device when using this early
> platform device? I remember when I was working on the clock drivers, I
> tried registering a platform device in the init_time callback but the
> kernel crashed because kobj stuff was not initialized yet. I'm guessing
> that the early platform device somehow works around this.
>

Yes, it seems we do. I was getting kobj stack dumps too when trying to
register a device using just platform_device_register() and it went
away as soon as I switched to early platform.

Best regards,
Bartosz Golaszewski

2018-04-24 09:04:31

by Sekhar Nori

[permalink] [raw]
Subject: Re: [RFC work-in-progress 0/7] of: platform: use early platform routines instead of OF_DECLARE

On Tuesday 24 April 2018 12:56 PM, Bartosz Golaszewski wrote:
> 2018-04-23 23:38 GMT+02:00 David Lechner <[email protected]>:
>> FYI: It looks like the CC for Stephen and Arnd was messed up, so I
>> fixed.
>>
>
> Thanks!
>
>> On 04/23/2018 01:38 PM, Bartosz Golaszewski wrote:
>>>
>>> From: Bartosz Golaszewski <[email protected]>
>>>
>>> Hi David, Sekhar,
>>>
>>> since platform devices are generally considered more desirable than
>>> CLK_OF_DECLARE, TIMER_OF_DECLARE etc. and we need to figure out how to
>>> handle the clocks that need to be initialized early in the boot
>>> process on DaVinci, I thought that I could give the early_platform
>>> mechanism a try.
>>>
>>> This API is only used on one architecture (sh) but seems to work just
>>> fine on ARM. It allows to register early platform drivers and then
>>> probe them early in the boot process. So far only machine code is
>>> supported but with a bit of hacking I was able to probe a DT device.
>>>
>>> This is a very dirty and far-from-upstream proof of concept that allows
>>> to probe the (so far dummy) davinci timer platform device during the
>>> call to init_time (from machine_desc).
>>>
>>> The idea is to have a special compatible fallback string: "earlydev"
>>> that similarily to "syscon" would be added to device nodes that need
>>> early probing. Then we'd call the of_early_platform_populate()
>>> function that would find all compatible nodes and populate them
>>> long before all the "normal" nodes.
>>
>>
>> FWIW, "earlydev" sounds like a driver implementation detail, so not
>> something that should be included in the device tree. We only need
>> this because Linux needs a clocksource early on, but that doesn't
>> mean that all device tree users need to do the same.
>>
>> I'm sure it makes things easier for a proof of concept though. :-)
>>
>
> We already have "syscon" which too is more an implementation detail
> than HW description. I should have probably Cc'ed Rob Herring. I'll do
> it with a more polished version I should have today.

Yeah, we should check with DT maintainers here. Even if there is push
back on this, I suppose we can make of_early_platform_populate() take a
list of compatible strings whose DT nodes should be considered early
platform devices?

>>> This would allow us to make the davinci timer a normal platform device
>>> and possibly also probe the psc and pll drivers earlier than we do now.
>>>
>>> The early platform API even allows us to check if we're being probed
>>> early in probe() so we can possibly probe the driver twice if needed:
>>> only doing the critical stuff first and then completing the process
>>> later.
>>>
>>> If you think this is a good idea, I would like to continue on that
>>> and eventually make it an alternative to OF_DECLARE macros.
>>>
>>> For a quick conversion of the davinci timer to a platform driver
>>> I image we'd need to use platform data lookup that would be passed
>>> to of_early_platform_populate().
>>
>>
>> On the surface, it certainly sounds like a good idea to me. Do we have
>> access to struct device of the platform device when using this early
>> platform device? I remember when I was working on the clock drivers, I
>> tried registering a platform device in the init_time callback but the
>> kernel crashed because kobj stuff was not initialized yet. I'm guessing
>> that the early platform device somehow works around this.
>>
>
> Yes, it seems we do. I was getting kobj stack dumps too when trying to
> register a device using just platform_device_register() and it went
> away as soon as I switched to early platform.

I agree, it sounds like a good idea to use for clock and timer devices.
Thanks for looking into this. Looking forward to the more polished version.

Thanks,
Sekhar