[PATCH v2 16/50] image: Add Kconfig options for FIT in the host build

[Alex G.]

On 5/13/21 6:56 PM, Simon Glass wrote:
> Hi Alex,
> 
> On Thu, 13 May 2021 at 10:21, Alex G. <mr.nuke.me at gmail.com> wrote:
>>
>>
>>
>> On 5/12/21 12:30 PM, Simon Glass wrote:
>>> Hi Alex,
>>>
>>> On Wed, 12 May 2021 at 10:18, Alex G. <mr.nuke.me at gmail.com> wrote:
>>>>
>>>>
>>>>
>>>> On 5/12/21 10:54 AM, Simon Glass wrote:
>>>>> Hi Alex,
>>>>>
>>>>> On Wed, 12 May 2021 at 09:48, Alex G. <mr.nuke.me at gmail.com> wrote:
>>>>>>
>>>>>>
>>>>>>
>>>>>> On 5/12/21 9:51 AM, Simon Glass wrote:
>>>>>>> Hi Alex,
>>>>>>>
>>>>>>> On Tue, 11 May 2021 at 13:57, Alex G. <mr.nuke.me at gmail.com> wrote:
>>>>>>>>
>>>>>>>> On 5/6/21 9:24 AM, Simon Glass wrote:
>>>>>>
>>>>>> [snip]
>>>>>>
>>>>>>>>
>>>>>>>>> +
>>>>>>>>> +config HOST_FIT_PRINT
>>>>>>>>> +     def_bool y
>>>>>>>>> +     help
>>>>>>>>> +       Print the content of the FIT verbosely in the host build
>>>>>>>>
>>>>>>>> This option also doesn't make sense.This seems to do what 'mkimage -l'
>>>>>>>> already supports.
>>>>>>>
>>>>>>> Are you sure you have looked at the goal of the CONFIG_IS_ENABLED()
>>>>>>> changes? This is here purely to avoid #ifdefs in the share code.
>>>>>>
>>>>>> On the one hand, we have the cosmetic inconvenience caused by #ifdefs.
>>>>>> On the other hand we have the config system. To most users, the config
>>>>>> system is likely more visible, while it's mostly developers who will
>>>>>> ever see the ifdefs.
>>>>>>
>>>>>> Therefore, in order to get the developer convenience of less ifdefs, we
>>>>>> have to sacrifice user convenience by cloberring the Kconfig options. I
>>>>>> think this is back-to-front.
>>>>>
>>>>> These Kconfig options are not visible to users. They cannot be updated
>>>>> in defconfig, nor in 'make menuconfig', etc. They are purely there for
>>>>> the build system.
>>>>>
>>>>>>
>>>>>> Can we reduce the host config count to just SLL/NOSSL?
>>>>>
>>>>> The point here is that the code has a special case for host builds,
>>>>> and this is a means to remove that special case and make the code
>>>>> easier to maintain and follow.
>>>>
>>>> I understand where you're coming from. Without these changes, the code
>>>> knows what it should and should not do, correct? My argument is that if
>>>> the code has the logic to do the correct thing, that logic should not be
>>>> split with the config system.
>>>>
>>>> I agree with the goal of reducing clutter in the source code. I disagree
>>>> with this specific course of fixing it. Instead, I propose a single
>>>> kconfig for host tools for SSL on/off.
>>>>
>>>> The disadvantage of my proposal is that we have to refactor the common
>>>> code in a way consistent with the goals, instead of just changing some
>>>> #ifdefs to if(CONFIG_IS_ENABLED()). I admit that, off the top of my
>>>> head, I don't have a detailed plan on how to achieve this.
>>>
>>> You are mostly describing the status quo, so far as I understand it.
>>> The problem is with the code that is built for both boards and tools.
>>> For boards, we want this code to be compiled conditionally, depending
>>> on what options are enabled. For tools, we want the code to be
>>> compiled unconditionally.
>>>
>>> I can think of only three ways to do this:
>>>
>>> - status quo (add #ifdefs USE_HOSTCC wherever we need to)
>>> - my series (make use of hidden Kconfig options to avoid that)
>>> - put every single feature and associated lines of code in separate
>>> files and compile them conditionally for boards, but always for tools
>>>
>>> I believe the last option is actually impossible, or at least
>>> impractical. It would cause an explosion of source files to deal with
>>> all the various combinations, and would be quite painful to maintain
>>> also.
>>
>> I don't think the status quo is such a terrible solution, so I am
>> looking at the aspects that can benefit from improvement. Hence why it
>> may appear I am talking about the status quo.
>>
>> Let's assume CONFIG_IS_ENABLED() returns true on host builds, and for
>> those cases where you need to know if IS_HOST_BUILD(), there's a macro
>> for that. You have the oddball case that uses a CONFIG_ value that's
>> only defined on the target, and those you would have to split according
>> to your option #3.
>>
>> I don't think the above is as dire an explosion in source files as it seems.
>>
>> Another point of contention is checksum_algos and crypto_algos arrays
>> image-sig.c. On the target side, these should be in .u-boot-list. Status
>> quo is the definition of rsa_verify is hidden behind #if macros, which
>> just pushes the complexity out into the rsa.h headers.
>>
>> The two ideas here are CONFIG_IS_ENABLED() returns true for host code,
>> and image-sig.c is split bwtween host and target versions, the target
>> version making use of .uboot-list. Using these as the starting points, I
>> think we can get to a much better solution
> 
> I did consider simply defining CONFIG_IS_ENABLED() to true for the
> host, but rejected that because I worried it would break down at some
> point. Examples I can think of at the moment:
> 
> - conflicting or alternative options (OF_LIVE, OF_HOST_FILE, OF_SEPARATE)
> - code that is not actually wanted on the host (WATCHDOG,
> NEEDS_MANUAL_RELOC, FPGA)
> - code that we want on the host but not the board (so we end up with a
> dummy CONFIG for the boards?)
> - all the SPL / TPL / VPL options which would always end up being
> true, when in fact we probably want none of them
> 
> I think you should more clearly explain your objection to the hidden
> Kconfig options, since your original reason ("clobbering the Kconfig
> space") doesn't seem to have survived further analysis.

I thought it to have been already explained and settled. It objectively 
increase the complexity of the logic. Instead of the logic being defined 
in one place (code), it is now defined in two places (code and Kconfig).

And subjectively, when adding ECDSA support, I found the ifs/ifdefs that 
were derived from kconfig code to be extremely confusing. It would have 
helped trememdously if the host code would flow with less 'kconfig' 
decision points. I don't find this series to improve on that.

> Having said that, I can imagine with a lot of refactoring it might be
> possible to address some of those. I just don't see it as a feasible
> option from here. The effort would be better spent improving testing,
> I think. But if you'd like to code something up for it, I'd be
> interested to see it.

We are in agreement that refactoring will improve the situation. I don't 
think it's so dire that it's unfeasible. However, if you'd like, we can 
start in smaller chunks.

> Re the linker-list stuff, yes we should push more things to those
> instead of #ifdefs and weak functions. Hashing and crypto are prime
> examples. In fact host tools can use linker lists too if we need that.

Pretty low hanging fruit here. Let me try to code something up.

> Regards,
> Simon
>