ARM: amba: defining module aliases for AMBA driver autoloading

There's no special reason why AMBA device drivers should not be auto-loadable via udev, but udev currently has no way to map AMBA device IDs to drivers.

As part of the effort to help enable the building of multiple platforms into a single kernel image in the future, it's desirable to be able to build any non-critical platform-specific drivers as modules.

A straightforward solution is to use modaliases to allow udev to identify the correct driver module to load.

This series adds the needed MODALIAS uevent property when sending events concerning amba bus devices, and adds appropriate MODULE_ALIAS() declarations to a couple of modules.

Briefly tested on Versatile Express.

Any comments and feedback are welcome.

Issues:

* Do new module alises need to be globally agreed/registered somewhere?

* Because a driver's ID table is in match-and-mask format whereas udev uses string pattern matching, we effectively have to maintain two ID tables per driver, containing the same information in different formats. The patch to mmci.c gives an example.

I predict that maintenance of those duplicated tables will be somewhat painful and error-prone. However, the necessary transformations, while simple, are beyond the scope of the C preprocessor.

In order to avoid this duplication of information, an extra (but simple) bit of build-time infrastructure would be needed.

I think this effort would be worth it -- does anyone have strong views on this?

Dave Martin (3):
ARM: amba: pass a suitable modalias to udev
ARM: sound/arm/aaci.c: Define amba module alias
ARM: drivers/mmc/host/mmci.c: Define amba module alias

 drivers/amba/bus.c      |    9 ++++++++-
 drivers/mmc/host/mmci.c |    8 ++++++++
 sound/arm/aaci.c        |    2 ++
 3 files changed, 18 insertions(+), 1 deletions(-)

--
1.7.4.1

This patch defines and imlpements a new modalias for AMBA devices:

MODALIAS=amba:xxxxxxxx

where 'x' is a hex digit in lower case.

This enables udev to auto-load modules transparently for AMBA devices, provided that each driver makes suitable MODULE_ALIAS() declarations based on the supported AMBA device IDs.

Without this, drivers built as modules never get loaded, unless they are loaded manually.

---
 drivers/amba/bus.c |    9 ++++++++-
 1 files changed, 8 insertions(+), 1 deletions(-)

index d74926e..7d2ba0b 100644
@@ -52,7 +52,14 @@ static int amba_uevent(struct device *dev, struct kobj_uevent_env *env)
 	int retval = 0;
 
 	retval = add_uevent_var(env, "AMBA_ID=%08x", pcdev->periphid);
-	return retval;
+	if (retval)
+		return retval;
+
+	retval = add_uevent_var(env, "MODALIAS=amba:%08x", pcdev->periphid);
+	if (retval)
+		return retval;
+
+	return 0;
 }
 #else
 #define amba_uevent NULL

1.7.4.1

---
 sound/arm/aaci.c |    2 ++
 1 files changed, 2 insertions(+), 0 deletions(-)

index d0cead3..2e68bfd 100644
@@ -1097,6 +1097,8 @@ static struct amba_id aaci_ids[] = {
 	{ 0, 0 },
 };
 
+MODULE_ALIAS("amba:???41041");
+
 static struct amba_driver aaci_driver = {
 	.drv		= {
 		.name	= DRIVER_NAME,

1.7.4.1

---
 drivers/mmc/host/mmci.c |    8 ++++++++
 1 files changed, 8 insertions(+), 0 deletions(-)

index 56e9a41..ea2a72c 100644
@@ -1477,6 +1477,14 @@ static struct amba_id mmci_ids[] = {
 	{ 0, 0 },
 };
 
+MODULE_ALIAS("amba:00?41180");
+MODULE_ALIAS("amba:01?41180");
+MODULE_ALIAS("amba:???41181");
+MODULE_ALIAS("amba:??180180");
+MODULE_ALIAS("amba:??280180");
+MODULE_ALIAS("amba:0?480180");
+MODULE_ALIAS("amba:1?480180");
+
 static struct amba_driver mmci_driver = {
 	.drv		= {
 		.name	= DRIVER_NAME,

1.7.4.1

I don't think there is. Having the bus prefix in there is definitely a good thing.

I think there is some support for udev to use PCI ID tables directly. I don't know how that works - I suspect it requires module tools support though. Might be worth investigating.

OK, I'll stick with "amba:" for now, then.

Hmmm, it looks like there's some magic for other bus types like PCI and USB, where some scripting generates appropriate module aliases from drivers' device tables (see scripts/mod/file2alias.c)

I'll see whether we can hook into that. This would allow us to avoid having to describe the aliases explicitly.

Thanks for the feedback.

Cheers
---Dave