Re: Solving section mismatches

[Sam Ravnborg]

On Sun, Sep 27, 2009 at 05:41:16PM +0100, Russell King - ARM Linux wrote:
> Sam,
> 
> Any idea how to solve this:
> 
> WARNING: arch/arm/kernel/built-in.o(.text+0x1ebc): Section mismatch in reference from the function cpu_idle() to the function .cpuexit.text:cpu_die()
> The function cpu_idle() references a function in an exit section.
> Often the function cpu_die() has valid usage outside the exit section
> and the fix is to remove the __cpuexit annotation of cpu_die.
> 
> WARNING: arch/arm/kernel/built-in.o(.cpuexit.text+0x3c): Section mismatch in reference from the function cpu_die() to the function .cpuinit.text:secondary_start_kernel()
> The function __cpuexit cpu_die() references
> a function __cpuinit secondary_start_kernel().
> This is often seen when error handling in the exit function
> uses functionality in the init path.
> The fix is often to remove the __cpuinit annotation of
> secondary_start_kernel() so it may be used outside an init section.
> 
> Logically, the annotations are correct - in the first case, cpu_die()
> will only ever be called if hotplug CPU is enabled, since you can't
> offline a CPU without hotplug CPU enabled.  In that case, the __cpuexit.*
> sections are not discarded.

The annotation of cpu_die() is wrong.
To be annotated __cpuexit the function shall:
- be used in exit context and only in exit context with HOTPLUG_CPU=n
- be used outside exit context with HOTPLUG_CPU=y

cpu_die() fails on the first condition because it is only used
if HOTPLUG_CPU=y.
The annotation is wrongly used as a replacement for an ifdef.
As cpu_die() is already inside ifdef CONFIG_HOTPLUG_CPU is should
be enough to just remove the annotation.

Like this (copy'n'paste so it does not apply)

diff --git a/arch/arm/kernel/smp.c b/arch/arm/kernel/smp.c
index e0d3277..de4ef1c 100644
--- a/arch/arm/kernel/smp.c
+++ b/arch/arm/kernel/smp.c
@@ -214,7 +214,7 @@ void __cpuexit __cpu_die(unsigned int cpu)
  * of the other hotplug-cpu capable cores, so presumably coming
  * out of idle fixes this.
  */
-void __cpuexit cpu_die(void)
+void cpu_die(void)
 {
        unsigned int cpu = smp_processor_id();
 


I did not see any section mismatch with defconfig so I failed to
test if it made a difference.

> 
> In the second case, this is platform code to restart the CPU, and
> again the anotations are entirely correct - cpuexit code only exists
> if hotplug CPU is enabled, in which case the cpuinit code must also
> be kept around.  Therefore, it is entirely legal for cpuexit code to
> call cpuinit code - unlike __init vs __exit or __devinit vs __devexit.
> 
> Therefore, I believe both of these warnings to be incorrect.

Same reasoning as above.
And the patch ougth to solve this case too.

Please note that throughout the kernel there is a lot of places where
the __cpu* annotations has been misused.
I once suggested to drop the chek from modpost as I could not see
this being fixed - but alas no replies it never happened.

	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/