const __init

const __init

[Geert Uytterhoeven]

Since a while include/linux/init.h contains the line

    * Also note, that this data cannot be "const".

Why is this? Because const data will be put in a different section?

However, quite some code defines const __init variables (see list below).

So what should be done now?
  1. Remove const from __initdata variables
  2. Make __inidata work with const variables

FWIW, many sources still use __init for data, while it should be __initdata.

Gr{oetje,eeting}s,

						Geert

Appendix: here's the list of affected source files:

    arch/arm/kernel/setup.c
    arch/m68k/amiga/config.c
    arch/ppc/amiga/config.c
    arch/ppc/kernel/residual.c
    drivers/acorn/net/ether1.c
    drivers/acorn/net/ether3.c
    drivers/acorn/net/etherh.c
    drivers/atm/ambassador.c
    drivers/cdrom/cdu31a.c
    drivers/char/dsp56k.c
    drivers/char/pc110pad.c
    drivers/char/qpmouse.c
    drivers/char/softdog.c
    drivers/ide/buddha.c
    drivers/ide/rapide.c
    drivers/net/ewrk3.c
    drivers/net/hamradio/6pack.c
    drivers/net/hamradio/bpqether.c
    drivers/net/hamradio/mkiss.c
    drivers/net/hamradio/scc.c
    drivers/net/hamradio/yam.c
    drivers/net/strip.c
    drivers/net/tokenring/ibmtr.c
    drivers/net/wan/lapbether.c
    drivers/net/wan/z85230.c
    drivers/pci/names.c
    drivers/scsi/seagate.c
    drivers/sound/sonicvibes.c
    drivers/video/amifb.c
    drivers/video/aty128fb.c
    drivers/video/atyfb.c
    drivers/video/mdacon.c
    drivers/video/modedb.c
    drivers/video/newport_con.c
    drivers/video/promcon.c
    drivers/video/riva/fbdev.c
    drivers/video/sticon-bmode.c
    drivers/video/sticon.c
    drivers/video/tdfxfb.c
    drivers/video/vgacon.c
    drivers/zorro/names.c
    net/ax25/af_ax25.c
    net/ipv4/ipip.c
    net/ipx/af_ipx.c
    net/ipx/af_spx.c
    net/lapb/lapb_iface.c
    net/netrom/af_netrom.c
    net/unix/af_unix.c

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
							    -- Linus Torvalds

Re: const __init

[Jeff Garzik]

Geert Uytterhoeven wrote:
> 
> Since a while include/linux/init.h contains the line
> 
>     * Also note, that this data cannot be "const".
> 
> Why is this? Because const data will be put in a different section?

Causes a "section type conflict" build error, at least on x86.


> FWIW, many sources still use __init for data, while it should be __initdata.

*blink* wow

-- 
Jeff Garzik      | "Do you have to make light of everything?!"
Building 1024    | "I'm extremely serious about nailing your
MandrakeSoft     |  step-daughter, but other than that, yes."

Re: const __init

[Geert Uytterhoeven]

On Sun, 20 May 2001, Jeff Garzik wrote:
> Geert Uytterhoeven wrote:
> > Since a while include/linux/init.h contains the line
> > 
> >     * Also note, that this data cannot be "const".
> > 
> > Why is this? Because const data will be put in a different section?
> 
> Causes a "section type conflict" build error, at least on x86.

On m68k I only saw section type conflict errors when using __init while it
should have been __initdata.

Gr{oetje,eeting}s,

						Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
							    -- Linus Torvalds

Re: const __init

[Jeff Garzik]

Geert Uytterhoeven wrote:
> 
> On Sun, 20 May 2001, Jeff Garzik wrote:
> > Geert Uytterhoeven wrote:
> > > Since a while include/linux/init.h contains the line
> > >
> > >     * Also note, that this data cannot be "const".
> > >
> > > Why is this? Because const data will be put in a different section?
> >
> > Causes a "section type conflict" build error, at least on x86.
> 
> On m68k I only saw section type conflict errors when using __init while it
> should have been __initdata.

This might be a very valid point...

(let me know if the following test is flawed)

> [jgarzik@rum tmp]$ cat > sectest.c
> #include <linux/module.h>
> #include <linux/init.h>
> static const char version[] __initdata = "foo";
> [jgarzik@rum tmp]$ gcc -D__KERNEL__ -I/spare/cvs/linux_2_4/include -Wall -Wstrict-prototypes -O2 -fomit-frame-pointer -fno-strict-aliasing -pipe -mpreferred-stack-boundary=2 -march=i686    -c -o sectest.o sectest.c
> [jgarzik@rum tmp]$ 

No section type conflict appears.

-- 
Jeff Garzik      | "Do you have to make light of everything?!"
Building 1024    | "I'm extremely serious about nailing your
MandrakeSoft     |  step-daughter, but other than that, yes."

Re: const __init

[Ingo Oeser]

On Sun, May 20, 2001 at 05:34:48PM -0400, Jeff Garzik wrote:
> This might be a very valid point...
> 
> (let me know if the following test is flawed)
 
It is imho.

> > [jgarzik@rum tmp]$ cat > sectest.c
> > #include <linux/module.h>
> > #include <linux/init.h>
> > static const char version[] __initdata = "foo";
    static char version2[] __initdata = "bar";
> > [jgarzik@rum tmp]$ gcc -D__KERNEL__ -I/spare/cvs/linux_2_4/include -Wall -Wstrict-prototypes -O2 -fomit-frame-pointer -fno-strict-aliasing -pipe -mpreferred-stack-boundary=2 -march=i686    -c -o sectest.o sectest.c
> > [jgarzik@rum tmp]$ 
> 
> No section type conflict appears.

Now it SHOULD conflict on these binutils, but doesn't on mine (2.9.5) ;-)

It is decided to put it into .data.init as expected.

AFAIK "const" is only a promise to the compiler, that we write
this data ONCE and read only after this initial write. So the
decision on the section is implementation defined.

What I don't understand is, why GCC overrides our explicit
override (done by setting the "section attribute" explicitly).

I would consider this a BUG in GCC. I don't understand, why we
support this BUG...

Maybe some GCC people can enlighten me, why GCC ignores such
overrides, that are for the cases where we DO KNOW BETTER than
GCC, what section is correct.


Regards

Ingo Oeser
-- 
To the systems programmer,
users and applications serve only to provide a test load.

Re: const __init

[Richard Henderson]

On Mon, May 21, 2001 at 12:19:49AM +0200, Ingo Oeser wrote:
> AFAIK "const" is only a promise to the compiler, that we write
> this data ONCE and read only after this initial write. So the
> decision on the section is implementation defined.

No, the problem is not with which section, but what flags that
section should have.  If you put only "const" data in a section,
then the section should have SHF_WRITE clear.  Conversely, if
you put writable data in a section then SHF_WRITE should be set.

Now, one could argue that gcc should scan the entire file to
see if there are any non-constant data members added to a 
particular section, and set SHF_WRITE if any such exist.

My answer is: you would not like gcc's memory usage under these
conditions.


r~

Re: const __init

[Jamie Lokier]

Richard Henderson wrote:
> No, the problem is not with which section, but what flags that
> section should have.  If you put only "const" data in a section,
> then the section should have SHF_WRITE clear.  Conversely, if
> you put writable data in a section then SHF_WRITE should be set.
> 
> Now, one could argue that gcc should scan the entire file to
> see if there are any non-constant data members added to a 
> particular section, and set SHF_WRITE if any such exist.
> 
> My answer is: you would not like gcc's memory usage under these
> conditions.

No, I'd argue that gcc should offer a `section_flags' attribute to be
used along with the `section' attribute.  Then we change the definition
of __init.

-- Jamie.

Re: const __init

[Keith Owens]

On Sun, 20 May 2001 17:34:48 -0400, 
Jeff Garzik <jgarzik@mandrakesoft.com> wrote:
>(let me know if the following test is flawed)
>
> [jgarzik@rum tmp]$ cat > sectest.c
> #include <linux/module.h>
> #include <linux/init.h>
> static const char version[] __initdata = "foo";
> [jgarzik@rum tmp]$ gcc -D__KERNEL__ -I/spare/cvs/linux_2_4/include -Wall -Wstrict-prototypes -O2 -fomit-frame-pointer -fno-strict-aliasing -pipe -mpreferred-stack-boundary=2 -march=i686    -c -o sectest.o sectest.c
> [jgarzik@rum tmp]$ 
>
>No section type conflict appears.

With just one variable in initdata there is no conflict, it takes two
to conflict.

static const char var1[] __attribute__ ((__section__ (".data.init"))) = "foo";
int main(void) { return(0); }
static       int  var2[] __attribute__ ((__section__ (".data.init"))) = {0,1};

does cause a section conflict, egcs 1.1.2.

Interestingly enough, if var[12] are together, without the intervening
text, then gcc does not flag an error, instead it puts both variables
in section .data.init and marks it as read only.  This looks like a bug
in gcc.

Re: const __init

[Richard Henderson]

On Mon, May 21, 2001 at 01:07:50PM +1000, Keith Owens wrote:
> does cause a section conflict, egcs 1.1.2.
> 
> Interestingly enough, if var[12] are together, without the intervening
> text, then gcc does not flag an error, instead it puts both variables
> in section .data.init and marks it as read only.  This looks like a bug
> in gcc.

Fixed in compilers newer than 2 years old.


r~

Re: const __init

[J . A . Magallon]

On 05.21 Richard Henderson wrote:
> On Mon, May 21, 2001 at 01:07:50PM +1000, Keith Owens wrote:
> > does cause a section conflict, egcs 1.1.2.
> > 
> > Interestingly enough, if var[12] are together, without the intervening
> > text, then gcc does not flag an error, instead it puts both variables
> > in section .data.init and marks it as read only.  This looks like a bug
> > in gcc.
> 
> Fixed in compilers newer than 2 years old.
> 

Somebody should officially certify 2.95.2 (or better 2.95.3, if all the
kernel related bugs are solved), and change:

"The recommended compiler for the kernel is egcs 1.1.2 (gcc 2.91.66), and it
should be used when you need absolute stability. You may use gcc 2.95.x
instead if you wish, although it may cause problems."

-- 
J.A. Magallon                           #  Let the source be with you...        
mailto:jamagallon@able.es
Linux Mandrake release 8.1 (Cooker) for i586
Linux werewolf 2.4.4-ac11 #2 SMP Fri May 18 12:27:06 CEST 2001 i686

Re: const __init

[Franz Sirl]

On Sunday 20 May 2001 21:51, Geert Uytterhoeven wrote:
> Since a while include/linux/init.h contains the line
>
>     * Also note, that this data cannot be "const".
>
> Why is this? Because const data will be put in a different section?

Yes, and gcc3 errors on these constructs,  cause it cannot decide if the data 
should be put into a .data or .rodata section.
Dunno if it's worth to create a __initconstdata/__initrodata though, but it 
would be easy implement I guess.

>     drivers/video/aty128fb.c

Fixed in bk 2_4.

Franz.

Re: const __init

[Keith Owens]

On Sun, 20 May 2001 22:16:11 +0200, 
Franz Sirl <Franz.Sirl-kernel@lauterbach.com> wrote:
>Yes, and gcc3 errors on these constructs,  cause it cannot decide if the data 
>should be put into a .data or .rodata section.
>Dunno if it's worth to create a __initconstdata/__initrodata though, but it 
>would be easy implement I guess.

Not worth it.  Adding finer grained init.data sections requires changes
to every architecture's vmlinux.lds script but it gains you nothing,
the sections are discarded after the kernel has booted.

Re: const __init

[Russell King]

On Sun, May 20, 2001 at 09:51:04PM +0200, Geert Uytterhoeven wrote:
> Appendix: here's the list of affected source files:
> 
>     arch/arm/kernel/setup.c

Thanks for pointing it out.

--
Russell King (rmk@arm.linux.org.uk)                The developer of ARM Linux
             http://www.arm.linux.org.uk/personal/aboutme.html