Regression: commit "ide: constify struct ide_port_info" causes breakage

Regression: commit "ide: constify struct ide_port_info" causes breakage

[Russell King]

commit 8562043606430185cad26d085d46adcc7ad67fd1 is broken, causing:

  CC      drivers/ide/pci/cmd64x.o
  CC      drivers/ide/pci/hpt366.o
drivers/ide/pci/hpt366.c:1428: error: hpt366_chipsets causes a section type conflict

and therefore should be reverted.

The problem arises because hpt366 has other data marked with __devinitdata,
so the compiler tries to define the initdata section with read-write
attributes at one point, and read-only attributes when encountering the
const-but-devinitdata declaration:

static struct hpt_info hpt36x __devinitdata = {
static struct hpt_info hpt370 __devinitdata = {
static struct hpt_info hpt370a __devinitdata = {
...
static const struct ide_port_info hpt366_chipsets[] __devinitdata = {

If we want ide_port_info to be read-only initdata, then we need
__devroinitdata and __roinitdata tags.

Note: it doesn't show up in cmd64x.c because that file contains no other
data marked as initdata or devinitdata.

-- 
Russell King
 Linux kernel    2.6 ARM Linux   - http://www.arm.linux.org.uk/
 maintainer of:

Re: Regression: commit "ide: constify struct ide_port_info" causes breakage

[Sergei Shtylyov]

Hello.

Russell King wrote:

> commit 8562043606430185cad26d085d46adcc7ad67fd1 is broken, causing:

>   CC      drivers/ide/pci/cmd64x.o
>   CC      drivers/ide/pci/hpt366.o
> drivers/ide/pci/hpt366.c:1428: error: hpt366_chipsets causes a section type conflict

> and therefore should be reverted.
> 
> The problem arises because hpt366 has other data marked with __devinitdata,
> so the compiler tries to define the initdata section with read-write
> attributes at one point, and read-only attributes when encountering the
> const-but-devinitdata declaration:
> 
> static struct hpt_info hpt36x __devinitdata = {
> static struct hpt_info hpt370 __devinitdata = {
> static struct hpt_info hpt370a __devinitdata = {
> ...
> static const struct ide_port_info hpt366_chipsets[] __devinitdata = {

> If we want ide_port_info to be read-only initdata, then we need
> __devroinitdata and __roinitdata tags.

    We can just convert cariables declared as 'struct hpt_info' to const too 
-- they are read-only templates.

MBR, Sergei

Re: Regression: commit "ide: constify struct ide_port_info" causes breakage

[Linus Torvalds]

On Fri, 26 Oct 2007, Russell King wrote:
>
> commit 8562043606430185cad26d085d46adcc7ad67fd1 is broken, causing:
> 
>   CC      drivers/ide/pci/cmd64x.o
>   CC      drivers/ide/pci/hpt366.o
> drivers/ide/pci/hpt366.c:1428: error: hpt366_chipsets causes a section type conflict
> 
> and therefore should be reverted.
> 
> The problem arises because hpt366 has other data marked with __devinitdata,
> so the compiler tries to define the initdata section with read-write
> attributes at one point, and read-only attributes when encountering the
> const-but-devinitdata declaration:

Sad. This is not too uncommon, but yeah, it's really sad that we cannot 
use "const" (as a compiler type safety marker) together with section 
definitions (to get it put in special init sections).

Bartlomiej - I don't think we need to revert the whole commit, but all 
those "static const xyz __devinitdata = { .." things probably do need to 
have the const removed, even if it would be otherwise correct ;(

		Linus

Re: Regression: commit "ide: constify struct ide_port_info" causes breakage

[Linus Torvalds]

On Fri, 26 Oct 2007, Sergei Shtylyov wrote:
> 
>    We can just convert cariables declared as 'struct hpt_info' to const too -- 
> they are read-only templates.

Generally "__devinitdata" is better than "const", since it means that the 
data may be thrown away.

So if you drop one over the other, it's generally the "const" that should 
be dropped ;/

		Linus

Re: Regression: commit "ide: constify struct ide_port_info" causes breakage

[Sergei Shtylyov]

Hello.

Linus Torvalds wrote:

>>   We can just convert cariables declared as 'struct hpt_info' to const too -- 
>>they are read-only templates.

> Generally "__devinitdata" is better than "const", since it means that the 
> data may be thrown away.

    Agreed.

> So if you drop one over the other,

    No, just add 'const' to already __devinitdata data to make gcc happy again.

> it's generally the "const" that should be dropped ;/

    Well, this may break when/if we need to add __devinitdata that is actually 
r/w -- not likely with this driver but still...

> 		Linus

WBR, Sergei

Re: Regression: commit "ide: constify struct ide_port_info" causes breakage

[Arjan van de Ven]

On Fri, 26 Oct 2007 09:30:20 -0700 (PDT)
Linus Torvalds <torvalds@linux-foundation.org> wrote:

> 
> 
> On Fri, 26 Oct 2007, Sergei Shtylyov wrote:
> > 
> >    We can just convert cariables declared as 'struct hpt_info' to
> > const too -- they are read-only templates.
> 
> Generally "__devinitdata" is better than "const", since it means that
> the data may be thrown away.
> 
> So if you drop one over the other, it's generally the "const" that
> should be dropped ;/
> 

we could do a
__constdevinitdata
that translate to "const" for the "not throw away" cases, and to the
throw-away secontion for the other cases...

Re: Regression: commit "ide: constify struct ide_port_info" causes breakage

[Bartlomiej Zolnierkiewicz]

On Friday 26 October 2007, Russell King wrote:
> commit 8562043606430185cad26d085d46adcc7ad67fd1 is broken, causing:
> 
>   CC      drivers/ide/pci/cmd64x.o
>   CC      drivers/ide/pci/hpt366.o
> drivers/ide/pci/hpt366.c:1428: error: hpt366_chipsets causes a section type conflict
> 
> and therefore should be reverted.

Fixed 3 days ago, patch was posted to LKML.

http://lkml.org/lkml/2007/10/23/411

[ Will be pushed to Linus together with few other fixes soon. ]

> The problem arises because hpt366 has other data marked with __devinitdata,
> so the compiler tries to define the initdata section with read-write
> attributes at one point, and read-only attributes when encountering the
> const-but-devinitdata declaration:
> 
> static struct hpt_info hpt36x __devinitdata = {
> static struct hpt_info hpt370 __devinitdata = {
> static struct hpt_info hpt370a __devinitdata = {
> ...
> static const struct ide_port_info hpt366_chipsets[] __devinitdata = {
> 
> If we want ide_port_info to be read-only initdata, then we need
> __devroinitdata and __roinitdata tags.

http://lkml.org/lkml/2007/10/23/412

Thanks,
Bart