[U-Boot] [PATCH v4 6/6] common: Generic loader for file system

[Simon Glass]

Hi Michal,

On 30 July 2018 at 07:30, Michal Simek <michal.simek at xilinx.com> wrote:
> On 30.7.2018 15:26, Simon Glass wrote:
>> Hi,
>>
>> On 27 July 2018 at 02:40, Chee, Tien Fong <tien.fong.chee at intel.com> wrote:
>>>
>>> On Thu, 2018-07-26 at 11:03 +0200, Michal Simek wrote:
>>>> On 25.7.2018 18:03, Tom Rini wrote:
>>>>>
>>>>> On Wed, Jul 25, 2018 at 09:47:17AM -0600, Simon Glass wrote:
>>>>>>
>>>>>> Hi,
>>>>>>
>>>>>> On 25 July 2018 at 03:48, Michal Simek <michal.simek at xilinx.com>
>>>>>> wrote:
>>>>>>>
>>>>>>>
>>>>>>> On 25.7.2018 08:31, Chee, Tien Fong wrote:
>>>>>>>>
>>>>>>>> On Wed, 2018-07-18 at 16:48 +0200, Michal Simek wrote:
>>>>>>>>>
>>>>>>>>> On 6.7.2018 10:28, tien.fong.chee at intel.com wrote:
>>>>>>>>>>
>>>>>>>>>>
>>>>>>>>>> From: Tien Fong Chee <tien.fong.chee at intel.com>
>>>>>>>>>>
>>>>>> [...]
>>>>>>
>>>>>>>
>>>>>>>>
>>>>>>>>>
>>>>>>>>>
>>>>>>>>> Also that DT binding is quite weird and I don't think you
>>>>>>>>> will get
>>>>>>>>> ACK
>>>>>>>>> for this from device tree community at all. I think that
>>>>>>>>> calling via
>>>>>>>>> platdata and avoid DT nodes would be better way to go.
>>>>>>>> Why do you think DT binding is weird? The DT is designed
>>>>>>>> based on Simon
>>>>>>>> proposal, and i believe following the rules in DTS spec.
>>>>>>>> There are some DT benefits with current design, i think
>>>>>>>> someone may be
>>>>>>>> maintainer need to made the final decision on the design.
>>>>>>> It is software configuration in file which should mostly
>>>>>>> describe
>>>>>>> hardware and state for hardware configuration.
>>>>>>>
>>>>>>> Your fs_loader node is purely describe sw configuration which
>>>>>>> shouldn't
>>>>>>> be here.
>>>>>>> You have there run time configuration via variables. I think
>>>>>>> using only
>>>>>>> this way is enough. Default variables will match what you would
>>>>>>> want to
>>>>>>> add to DT.
>>>>>> I think DT makes sense in the U-Boot context.
>>>>>>
>>>>>> We don't have a user space to handle policy decisions, and the
>>>>>> 'chosen' node is a good place to configure these common features.
>>>>>>
>>>>>> While you can argue that the partition or filesystem where an
>>>>>> image
>>>>>> comes from is a software config, it is something that has to be
>>>>>> configured. It has impact on hardware too, since the FPGA has to
>>>>>> get
>>>>>> its firmware from somewhere. We use the chosen node to specify
>>>>>> the
>>>>>> UART to use, and this is no different. Again, we don't have user-
>>>>>> space
>>>>>> config files in U-Boot.
>>>>>>
>>>>>> This argument comes up from time to time and I'd really like to
>>>>>> put it
>>>>>> to bed for U-Boot. I understand that Linux has its own approach
>>>>>> and
>>>>>> rules, but in some cases they serve U-Boot poorly.
>>>>> I want to second this as well.  So long as we're using our prefix
>>>>> and
>>>>> we've thought through and discussed what we're trying to do here,
>>>>> it's
>>>>> OK to do things that might not be accepted for Linux.
>>>>>
>>>> I have not a problem with using chosen node with u-boot prefix
>>>> properties and my colleague hopefully with finish work about moving
>>>> u-boot,dm-pre-reloc; to chosen node where it should be (because
>>>> current
>>>> solution has also problem with ordering).
>>>>
>>>> In this loader case doc is saying that you can rewrite it with
>>>> variables
>>>> on the prompt (or via script).
>>>> For cases that you want to autodetect platform and pass/load correct
>>>> dtb
>>>> which setup u-boot this can be problematic and using DT is could be
>>>> considered as easier for use.
>>>>
>>>> In this case this is what was proposed:
>>>>
>>>> +     fs_loader0: fs-loader at 0 {
>>>> +             u-boot,dm-pre-reloc;
>>>> +             compatible = "u-boot,fs-loader";
>>>> +             phandlepart = <&mmc 1>;
>>>> +     };
>>>>
>>>> +     fs_loader1: fs-loader at 1 {
>>>> +             u-boot,dm-pre-reloc;
>>>> +             compatible = "u-boot,fs-loader";
>>>> +             mtdpart = "UBI",
>>>> +             ubivol = "ubi0";
>>>> +     };
>>>>
>>>> u-boot,dm-pre-reloc; requires DM_FLAG_PRE_RELOC which is not setup
>>>> for
>>>> this driver - it means this should be here.
>>> You are right, i missed this one. The intention of design enables user
>>> to call any loader with default storage through the sequence number  if
>>> fs loader is not defined in chosen. For example, there is a case where
>>> system loading the file from SDMMC, NAND and QSPI.
>>>>
>>>> compatible = "u-boot,fs-loader"; - bind and probe are empty that's
>>>> why
>>>> this is only used for filling platdata but driver has no user that's
>>>> why
>>>> this is unused till someone calls that functions.
>>>>
>>>> phandlepart/mtdpart/ubivol is just for setup.
>>> There are some benefits with driver model:
>>> 1. Saving space, calling when need.
>>> 2. Handle memory allocation and deallocation automatically.
>>>>
>>>> For the first case you can just use in chosen node:
>>>> u-boot,fs-loader = <&mmc 1>;
>>>>
>>>> And for UBIfs. I have never played with that but I expect it
>>>> shouldn't
>>>> be big problem to describe it differently too (something like)
>>>> u-boot,fs-loader = <0 ubi0>;
>>> Need consider description for UBIFS, using fs-loader seems not working
>>> for UBIFS, since more arguments such as mtdpartition and mtd volume
>>> need passing into driver. In order to avoid messing, fs_loader can act
>>> the pointer to the chosen.
>>>
>>> Anyway, i have no strong opinion with driver designed via platdata or
>>> driver model if we can resolve the problem for UBIFS and maintainers
>>> agree with it.
>>>>
>>>> Then this driver/interface can stay in DT where it should stay. The
>>>> only
>>>> thing is how this should be initialized because there is no
>>>> compatible
>>>> string. But you can do that via platdata for platforms which want to
>>>> use
>>>> this.
>>
>> We should add a compatible string then :-)
>
> Isn't driver name used in case of platdata initialization?

If the node is in /chosen and has a compatible string, it will be
bound automatically. Manually binding a device is really just a
fallback for particular situations (e.g. buses like PCI where we often
rely on probing to find out what is on the bus).

Regards,
Simon