On Sat, Jan 26, 2008 at 08:17:17PM +0100, Sam Ravnborg wrote:
quoted text
>On Sat, Jan 26, 2008 at 04:55:58AM -0500, Jeff Garzik wrote:
>> WANG Cong wrote:
>> >diff --git a/drivers/video/kyro/fbdev.c b/drivers/video/kyro/fbdev.c
>> >index acb9370..437ebd0 100644
>> >--- a/drivers/video/kyro/fbdev.c
>> >+++ b/drivers/video/kyro/fbdev.c
>> >@@ -90,7 +90,9 @@ static int nomtrr __devinitdata = 0;
>> > 
>> > /* PCI driver prototypes */
>> > static int kyrofb_probe(struct pci_dev *pdev, const struct pci_device_id 
>> > *ent);
>> >+#if defined(MODULE) || defined(CONFIG_HOTPLUG)
>> > static void kyrofb_remove(struct pci_dev *pdev);
>> >+#endif
>> > 
>> > static struct fb_videomode kyro_modedb[] __devinitdata = {
>> > 	{
>> >@@ -754,6 +756,7 @@ out_unmap:
>> > 	return -EINVAL;
>> > }
>> > 
>> >+#if defined(MODULE) || defined(CONFIG_HOTPLUG)
>> > static void __devexit kyrofb_remove(struct pci_dev *pdev)
>> > {
>> > 	struct fb_info *info = pci_get_drvdata(pdev);
>> >@@ -783,6 +786,7 @@ static void __devexit kyrofb_remove(struct pci_dev 
>> >*pdev)
>> > 	pci_set_drvdata(pdev, NULL);
>> > 	framebuffer_release(info);
>> > }
>> >+#endif
>> 
>> 
>> Quite strange -- due to __devexit_p() and the __devexit marker, ifdefs 
>> should not be needed.
>> 
>> I would look into why that isn't working as designed in these cases...
>
>I checked up on the synclink.c warning.
>We have the following code:
>
>static void synclink_remove_one (struct pci_dev *dev);
>
>...
>
>static struct pci_driver synclink_pci_driver = {
>        .remove         = __devexit_p(synclink_remove_one),
>};
>
>...
>
>static void __devexit synclink_remove_one (struct pci_dev *dev)
>{
>}
>
>And I double checked the preprocessed source to check
>that we applied the __attribute__((__used__)) to the function.

Yes. The only macro we used on synclink_remove_one is __devexit.
It's defination is:

#ifdef CONFIG_HOTPLUG
#define __devexit
#else
#define __devexit __exit
#endif

And __exit is:
#ifdef MODULE
#define __exit          __attribute__ ((__section__(".exit.text"))) __cold
#else
#define __exit          __attribute_used__ __attribute__ ((__section__(".exit.text"))) __cold
#endif

So only when !defined(MODULE) && !defined(CONFIG_HOTPLUG),
__attribute_used__ is applied. And this case is just that
condition, thus this should be a bogus of gcc.


quoted text
>
>Investigating a bit more I realized that gcc looses the
>__used__ attribution due to the prototype.
>So there are two correct fixes:
>a) move the function up so we do not need the forward
>   declaration
>b) add a __devexit to the forward decalration too.
>
>I strongly prefer the first version and this is the
>correct fix for these cases.
>
>Do we have a gcc bug here - I did not see a definitive answer in gcc docs?
>

Agreed. I will try a).


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