Re: [PATCH] misc/cs5535: Fix section mismatch derived fromcs5535_mfgpt_drv variable

[Sam Ravnborg]

On Mon, Jan 03, 2011 at 01:14:03PM -0800, Andres Salomon wrote:
> On Mon, 3 Jan 2011 21:50:19 +0100
> Sam Ravnborg <sam@xxxxxxxxxxxx> wrote:
> 
> > On Mon, Jan 03, 2011 at 12:34:31PM -0800, Andres Salomon wrote:
> > > On Mon, 3 Jan 2011 20:51:31 +0100
> > > Sam Ravnborg <sam@xxxxxxxxxxxx> wrote:
> > > 
> > > > On Mon, Jan 03, 2011 at 11:43:10AM -0800, Andrew Morton wrote:
> > > > > On Mon,  3 Jan 2011 03:51:28 +0100
> > > > > Sedat Dilek <sedat.dilek@xxxxxxxxxxxxxx> wrote:
> > > > > 
> > > > > > >From my build.log:
> > > > > > 
> > > > > > WARNING: drivers/misc/cs5535-mfgpt.o(.data+0x0): Section
> > > > > > mismatch in reference from the variable cs5535_mfgpt_drv to
> > > > > > the function .devinit.text:cs5535_mfgpt_probe() The variable
> > > > > > cs5535_mfgpt_drv references the function __devinit
> > > > > > cs5535_mfgpt_probe() If the reference is valid then annotate
> > > > > > the variable with __init* or __refdata (see linux/init.h) or
> > > > > > name the variable: *driver, *_template, *_timer, *_sht,
> > > > > > *_ops, *_probe, *_probe_one, *_console,
> > > > > 
> > > > > Confused.  cs5535_mfgpt_probe() _does_ have a name of the form
> > > > > "*_probe", so why did it warn?
> > > > 
> > > > The varibale need to follow the naming scheme.
> > > > In this case the function was named *_probe - which has no
> > > > significance in the section mismatch checks.
> > > > 
> > > > > > variable with __init* or __refdata (see linux/init.h) or name
> > > > > > the variable:
> > > >     ^^^^^^^^
> > > > ^^^^^^^^^ 
> > > > 
> > > > 	Sam
> > > 
> > > I see.  So it should really be renamed cs5535_mfgpt_driver, or _drv
> > > added to the whitelist.  Or the whitelist dropped, and all drivers
> > > changed to explicitly mark the structs with __devinitdata.
> > 
> > It is preferred that the variables are annotated __refdata.
> 
> Should they not be devinitdata, since they're structures referencing
> devinit (ie, hotplug init) hooks?
> 
> The two places where I traditionally went for documentation on such
> things was linux/init.h and Documentation/PCI/pci.txt.  The latter
> states:
> 
>         __devinit       Device initialization code.
>                         Identical to __init if the kernel is not
>         compiled with CONFIG_HOTPLUG, normal function otherwise.
> 
>         o Do not mark the struct pci_driver.
> 
> linux/init.h says:
> 
>  * modpost not to issue a warning.  Intended semantics is that a code or
>  * data tagged __ref* can reference code or data from init section
>    without
>  * producing a warning (of course, no warning does not mean code is
>  * correct, so optimally document why the __ref is needed and why it's
>    OK).
> 
> 
> It would be nice to know why refdata is correct here, and devinitdata
> is not.

The annotation really depends on the lifetime of the variable.

A variable that is used when the kernel is fully operational shall be
annotated with __refdata.
This will relocate the data to a special section that modpost
recognize and thus do not warn about references to functions
/data that are discarded during the startup phase.

__init is used for data that are uncnditionally discarded after
the init phase.
__devinitdata is used for data that is discarded after the initphase
is hotplug is not enabled - if hotplug is enabled then the
data are not discared.

So __init and __initdata have impact on the lifetime of the
data  + how modpost check references - where __refdata is
only used to shut up warnings from modpost.

__refdata was introduced late in the process where we
fixed a lot of section mismatch warnings.
So most of the fixes that hit the kernel was renaming
variables so references to init data/functions did not
cause warnings.

I looked into a more precise way to do the checkss
back then so we could do the check down on member
level in for example a "struct driver" - but I never
came up with anything good.
So therefore we are stuck with the less optimal check
algorithm today.

	Sam
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/