Re: [PATCH 0/3] Implement /proc/built-in file similar to /proc/modules

[Kirill Tkhai <tkhai@yandex.ru>]

On 14.09.2014 22:13, Greg KH wrote:
> On Sun, Sep 14, 2014 at 10:05:46PM +0400, Kirill Tkhai wrote:
>> On 14.09.2014 21:39, Greg KH wrote:
>>> On Sun, Sep 14, 2014 at 09:31:58PM +0400, Kirill Tkhai wrote:
>>>> On 14.09.2014 19:38, Greg KH wrote:
>>>>> On Sun, Sep 14, 2014 at 02:18:13PM +0400, Kirill Tkhai wrote:
>>>>>> This series implements a possibility to show the list of built-in drivers
>>>>>> to userspace. The names of drivers will be the same as when they are modules.
>>>>>
>>>>> Have you looked at /sys/modules/ ?  Doesn't that show what you want
>>>>> here?
>>>>
>>>> There are only the drivers in "/sys/module" which have parameters.
>>>> Drivers without parameters do not appear there.
>>>
>>> Ah, didn't realize that.  Should be easy to fix though, if you really
>>> wanted to list the modules.  Much better than a random proc file that
>>> you have to parse :)
>>
>> But it looks like one file is better than many new directories.
> 
> Why?

It's just an unification with /proc/modules. Why should we do any
difference between external and built-in modules? It's the same,
it's similar, it's better to parse when they can be shown similar.

> 
>> Furthermore some utils already may consider /sys/module directory as
>> a directory where all drivers have parameters. Is it good if we add
>> new ones of different type there?
> 
> What would break if you add new directories there with no parameters?

I do not know no one, but it does not mean they do not exist. I just do
not want to break anything which already exists.

>>>>>> So, if your system has "loop" driver then it appears either in /proc/modules
>>>>>> or in /proc/built-in and userspace will be able to know about this.
>>>>>>
>>>>>> Now this is impossible. The only way to get kernel configuration is
>>>>>> /proc/config.gz, but CONFIG_* names can change from time to time. Module
>>>>>> names are more or less standardized.
>>>>>
>>>>> Module names aren't "standardized", we change them at times when needed,
>>>>> just like CONFIG_ names.
>>>>>
>>>>> What is your end goal here?  As you say, config.gz is the real kernel
>>>>> configuration, just having a list of modules built in isn't going to
>>>>> help much in getting a working kernel config without it.
>>>>
>>>> It looks like userspace applications oriented on modules names rather
>>>> than on CONFIG_XXX parameters. /proc/config.gz is optional and userspace
>>>> applications can't base on it.
>>>>
>>>> For example, when I compile "loop" module built-in and "loop" is in
>>>> /etc/modules, init script warns about this module is not present and
>>>> can't be autoloaded. The script does not store CONFIG_XXX <-> module_xxx
>>>> conformity. And nobody stores it.
>>>>
>>>> When iptables wants extra functionality, it requests a module. Etc.
>>>>
>>>> Nobody is oriented on CONFIG_XXX parameters. It would be simple for
>>>> userspace to add a support of /proc/built-in analysing. It's very
>>>> similar to /proc/modules.
>>>
>>> Shouldn't userspace focus on the functionality a module provides, not
>>> the module name itself?  Can't a test for the loop "module" just test to
>>> see if the loop control device is present?  Same for iptables (there's
>>> modprobe rules for iptable modules I think...)
>>>
>>> In other words, don't focus on the module names, focus on the userspace
>>> function a module provides, there should always be a way to check that
>>> at run time (if not, then the module doesn't actually do much...)
>>
>> Hm, I'm not sure that anybody stores CONFIG_XXX <-> module_xxx
>> conformity. Everybody bases on module name. If application is seeing
>> CONFIG_XXX=m, but the functionality, which it want's, is not available,
>> what it has to do? How should it convert CONFIG_XXX to module name?
> 
> Why would an application ever care about CONFIG_XXX at runtime?
> 
>> So, many applications want module name instead of CONFIG_XXX, I believe.
> 
> No, they want the functionality that a module provides, not the module
> name, or some random configuation option.
> 
> It seems like you are trying to solve a problem that isn't there.  What
> program is broken right now that this new proc file (or sysfs directory)
> would fix?

The initial reason was I'm building custom kernels for more than 10
years (not so long, I agree), and every boot I see a big list of modules
from distribution /etc/module, which can't be autoloaded. I prefer to
build drivers in kernel. I tried to find is there a way for userspace to
understand that a module are present, but there is no a way. So this is
a reason.