Patchwork [U-Boot] Function prototype conflicts with standalone apps

login
register
mail settings
Submitter Chris Packham
Date Jan. 16, 2013, 4:23 a.m.
Message ID <CAFOYHZBRZHH-g1Us5CcUTwj9wi+giw9RPnZ6SV+cY4jVy=p-5w@mail.gmail.com>
Download mbox | patch
Permalink /patch/212386/
State RFC
Delegated to: Albert ARIBAUD
Headers show

Comments

Chris Packham - Jan. 16, 2013, 4:23 a.m.
Hi,

I've just run into something porting an existing out of tree board to
u-boot 2012.10 but I think it points to a generic issue for standalone
applications.

Consider the following change

Because I'm not linking with the u-boot object file, I should be able to
use any function name I like in my application as long as it isn't one of
the functions in exports.h (at least in theory). Unfortunately I end up
with the following compiler error

  hello_world.c:27: error: conflicting types for ‘net_init’
  uboot/include/net.h:489: error: previous declaration of ‘net_init’ was
here
  make[1]: *** [hello_world.o] Error 1

If I replace #include <common.h> in my app with the first hunk of includes
from the top of common.h  then I can compile just fine.

I was wondering if it made sense to people to have standalone applications
define something like __STANDALONE__ either via CPPFLAGS or in the source
itself and use the presence of that to exclude the majority of common.h
when used in standalone applications. Or alternatively move the required
bits to exports.h.

Thanks,
Chris
Albert ARIBAUD - Jan. 16, 2013, 7:25 a.m.
Hi Chris,

On Wed, 16 Jan 2013 17:23:58 +1300, Chris Packham
<judge.packham@gmail.com> wrote:

> Hi,
> 
> I've just run into something porting an existing out of tree board to
> u-boot 2012.10 but I think it points to a generic issue for standalone
> applications.
> 
> Consider the following change
> 
> diff --git a/examples/standalone/hello_world.c
> b/examples/standalone/hello_world.c
> index 067c390..d2e6a77 100644
> --- a/examples/standalone/hello_world.c
> +++ b/examples/standalone/hello_world.c
> @@ -24,7 +24,7 @@
>  #include <common.h>
>  #include <exports.h>
> 
> -int hello_world (int argc, char * const argv[])
> +int net_init (int argc, char * const argv[])
>  {
>         int i;
> 
> Because I'm not linking with the u-boot object file, I should be able to
> use any function name I like in my application as long as it isn't one of
> the functions in exports.h (at least in theory). Unfortunately I end up
> with the following compiler error
> 
>   hello_world.c:27: error: conflicting types for ‘net_init’
>   uboot/include/net.h:489: error: previous declaration of ‘net_init’ was
> here
>   make[1]: *** [hello_world.o] Error 1
> 
> If I replace #include <common.h> in my app with the first hunk of includes
> from the top of common.h  then I can compile just fine.
> 
> I was wondering if it made sense to people to have standalone applications
> define something like __STANDALONE__ either via CPPFLAGS or in the source
> itself and use the presence of that to exclude the majority of common.h
> when used in standalone applications. Or alternatively move the required
> bits to exports.h.

(long rant ahead. Short answer after end of rant)

[RANT]

Code writers indeed have a right to name any function or other object
any way they choose... within the constraints of the situation.

Some of these constraints stem from the tools -- you just cannot put an
ampersand in a C object name, for instance -- and some stem from the
'agreement' entered into when using a library -- precisely, the
agreement on the name and semantics of such and such object names.

Here, by including exports.h, you enter an agreement in which
the object name 'net_init' receives a specific meaning. What you want
is to benefit from the agreement without abiding by it.

Now this can be changed, technically, as most things are, and a new
kind of agreement could be devised with fine-grain control on which
object names would or would not be defined. The question is, *should*
this be done?

Would you, analogously, suggest that Linux app developers be able to
opt out of defining fopen() when they #include <stdio.h> because they
feel they have a right to define 'char* fopen(float F)' in their code if
they so please? And that it should be done so for just about any
kernel-exported symbol? I suspect not.

So why ask this from U-Boot?

[/RANT]

I personally will NAK such a suggestion. I don't see the point in
adding complexity just to solve a naming conflict between a framework,
de facto standard, name and a freely-modifiable application name. Just
rename the application function -- that'll be all the better since that
will also remove potential misunderstanding for readers of your code.

> Thanks,
> Chris

Amicalement,
Albert ARIBAUD - Jan. 16, 2013, 7:28 a.m.
Small fix:

> Here, by including exports.h

Actually common.h, but the point stands.

Amicalement,
Chris Packham - Jan. 16, 2013, 10:16 a.m.
Hi Albert,

On 01/16/2013 08:25 PM, Albert ARIBAUD (U-Boot) wrote:
> Hi Chris,
> 
> On Wed, 16 Jan 2013 17:23:58 +1300, Chris Packham
> <judge.packham@gmail.com> wrote:
> 
>> Hi,
>>
>> I've just run into something porting an existing out of tree board to
>> u-boot 2012.10 but I think it points to a generic issue for standalone
>> applications.
>>
>> Consider the following change
>>
>> diff --git a/examples/standalone/hello_world.c
>> b/examples/standalone/hello_world.c
>> index 067c390..d2e6a77 100644
>> --- a/examples/standalone/hello_world.c
>> +++ b/examples/standalone/hello_world.c
>> @@ -24,7 +24,7 @@
>>  #include <common.h>
>>  #include <exports.h>
>>
>> -int hello_world (int argc, char * const argv[])
>> +int net_init (int argc, char * const argv[])
>>  {
>>         int i;
>>
>> Because I'm not linking with the u-boot object file, I should be able to
>> use any function name I like in my application as long as it isn't one of
>> the functions in exports.h (at least in theory). Unfortunately I end up
>> with the following compiler error
>>
>>   hello_world.c:27: error: conflicting types for ‘net_init’
>>   uboot/include/net.h:489: error: previous declaration of ‘net_init’ was
>> here
>>   make[1]: *** [hello_world.o] Error 1
>>
>> If I replace #include <common.h> in my app with the first hunk of includes
>> from the top of common.h  then I can compile just fine.
>>
>> I was wondering if it made sense to people to have standalone applications
>> define something like __STANDALONE__ either via CPPFLAGS or in the source
>> itself and use the presence of that to exclude the majority of common.h
>> when used in standalone applications. Or alternatively move the required
>> bits to exports.h.
> 
> (long rant ahead. Short answer after end of rant)

Short response: Yep I can live with that by making some changes to my
standalone application. I just thought it might be cleaner if a minimal
set of definitions were provided.

> [RANT]
> 
> Code writers indeed have a right to name any function or other object
> any way they choose... within the constraints of the situation.
> 
> Some of these constraints stem from the tools -- you just cannot put an
> ampersand in a C object name, for instance -- and some stem from the
> 'agreement' entered into when using a library -- precisely, the
> agreement on the name and semantics of such and such object names.
> 
> Here, by including exports.h, you enter an agreement in which
> the object name 'net_init' receives a specific meaning. What you want
> is to benefit from the agreement without abiding by it.
> 
> Now this can be changed, technically, as most things are, and a new
> kind of agreement could be devised with fine-grain control on which
> object names would or would not be defined. The question is, *should*
> this be done?
> 
> Would you, analogously, suggest that Linux app developers be able to
> opt out of defining fopen() when they #include <stdio.h> because they
> feel they have a right to define 'char* fopen(float F)' in their code if
> they so please? And that it should be done so for just about any
> kernel-exported symbol? I suspect not.

Actually this is my point. The symbols aren't exported. They're just in
the header file. The kernel solution for this is using __KERNEL__ and
filtering the exported headers to remove the kernel internals not needed
by userland. If for some reason I did define a different fopen I'd get a
link error whether I included stdio or not.

> So why ask this from U-Boot?
> 
> [/RANT]
> 
> I personally will NAK such a suggestion. I don't see the point in
> adding complexity just to solve a naming conflict between a framework,
> de facto standard, name and a freely-modifiable application name. Just
> rename the application function -- that'll be all the better since that
> will also remove potential misunderstanding for readers of your code.
> 
>> Thanks,
>> Chris
> 
> Amicalement,
>
Albert ARIBAUD - Jan. 16, 2013, 12:57 p.m.
Hi Chris,

On Wed, 16 Jan 2013 23:16:07 +1300, Chris Packham
<judge.packham@gmail.com> wrote:

> Hi Albert,
> 
> On 01/16/2013 08:25 PM, Albert ARIBAUD (U-Boot) wrote:
> > Hi Chris,
> > 
> > On Wed, 16 Jan 2013 17:23:58 +1300, Chris Packham
> > <judge.packham@gmail.com> wrote:
> > 
> >> Hi,
> >>
> >> I've just run into something porting an existing out of tree board to
> >> u-boot 2012.10 but I think it points to a generic issue for standalone
> >> applications.
> >>
> >> Consider the following change
> >>
> >> diff --git a/examples/standalone/hello_world.c
> >> b/examples/standalone/hello_world.c
> >> index 067c390..d2e6a77 100644
> >> --- a/examples/standalone/hello_world.c
> >> +++ b/examples/standalone/hello_world.c
> >> @@ -24,7 +24,7 @@
> >>  #include <common.h>
> >>  #include <exports.h>
> >>
> >> -int hello_world (int argc, char * const argv[])
> >> +int net_init (int argc, char * const argv[])
> >>  {
> >>         int i;
> >>
> >> Because I'm not linking with the u-boot object file, I should be able to
> >> use any function name I like in my application as long as it isn't one of
> >> the functions in exports.h (at least in theory). Unfortunately I end up
> >> with the following compiler error
> >>
> >>   hello_world.c:27: error: conflicting types for ‘net_init’
> >>   uboot/include/net.h:489: error: previous declaration of ‘net_init’ was
> >> here
> >>   make[1]: *** [hello_world.o] Error 1
> >>
> >> If I replace #include <common.h> in my app with the first hunk of includes
> >> from the top of common.h  then I can compile just fine.
> >>
> >> I was wondering if it made sense to people to have standalone applications
> >> define something like __STANDALONE__ either via CPPFLAGS or in the source
> >> itself and use the presence of that to exclude the majority of common.h
> >> when used in standalone applications. Or alternatively move the required
> >> bits to exports.h.
> > 
> > (long rant ahead. Short answer after end of rant)
> 
> Short response: Yep I can live with that by making some changes to my
> standalone application. I just thought it might be cleaner if a minimal
> set of definitions were provided.
> 
> > [RANT]
> > 
> > Code writers indeed have a right to name any function or other object
> > any way they choose... within the constraints of the situation.
> > 
> > Some of these constraints stem from the tools -- you just cannot put an
> > ampersand in a C object name, for instance -- and some stem from the
> > 'agreement' entered into when using a library -- precisely, the
> > agreement on the name and semantics of such and such object names.
> > 
> > Here, by including exports.h, you enter an agreement in which
> > the object name 'net_init' receives a specific meaning. What you want
> > is to benefit from the agreement without abiding by it.
> > 
> > Now this can be changed, technically, as most things are, and a new
> > kind of agreement could be devised with fine-grain control on which
> > object names would or would not be defined. The question is, *should*
> > this be done?
> > 
> > Would you, analogously, suggest that Linux app developers be able to
> > opt out of defining fopen() when they #include <stdio.h> because they
> > feel they have a right to define 'char* fopen(float F)' in their code if
> > they so please? And that it should be done so for just about any
> > kernel-exported symbol? I suspect not.
> 
> Actually this is my point. The symbols aren't exported. They're just in
> the header file. The kernel solution for this is using __KERNEL__ and
> filtering the exported headers to remove the kernel internals not needed
> by userland. If for some reason I did define a different fopen I'd get a
> link error whether I included stdio or not.

Ok, so I misunderstood your objective -- my bad. It is not a problem of
wanting to use a function which is part of the API; it's a problem of
having to pull in non-API function declarations along with the API ones.

I still think it's a bad idea to use a homonym of an existing U-Boot
function, but at least you should not be prevented from doing so if
the function is not actually in the API.

But also I still dislike the solution you are suggestiong, because I
think that the actual problem is not properly stated. Re-reading the
common.h and exports.h header files, IIUC the actual problem is that
exports.h depends on some types from common.h but common.h also defines
some internal U-Boot functions. The problem is that common.h is needed
for different and incompatible purposes: (public) API on the one hand,
internal declarations on the other hand.

Thus, I'd rather have common.h simply split into "a header file
with basic types needed by both public API and internal code" and "a
header file defining some internal U-Boot functions, possibly
#including the shared basic types header file", with no conditional
involved. New header file names to be chosen wisely.

Amicalement,

Patch

diff --git a/examples/standalone/hello_world.c
b/examples/standalone/hello_world.c
index 067c390..d2e6a77 100644
--- a/examples/standalone/hello_world.c
+++ b/examples/standalone/hello_world.c
@@ -24,7 +24,7 @@ 
 #include <common.h>
 #include <exports.h>

-int hello_world (int argc, char * const argv[])
+int net_init (int argc, char * const argv[])
 {
        int i;