Re: [PATCH] uio/pdrv_genirq: Add OF support

[Hans J. Koch]

On Thu, Mar 31, 2011 at 01:48:32PM -0600, Grant Likely wrote:
> On Thu, Mar 31, 2011 at 1:23 PM, Hans J. Koch <hjk@xxxxxxxxxxxx> wrote:
> > On Thu, Mar 31, 2011 at 07:57:47PM +0200, Michal Simek wrote:
> >> Hans J. Koch wrote:
> >> >On Thu, Mar 31, 2011 at 03:28:41PM +0200, Michal Simek wrote:
> >> >>>>+ Â Â Â Â uioinfo->name = pdev->dev.of_node->name;
> >> >>>>+ Â Â Â Â /* Use version for storing full IP name for identification */
> >> >>>>+ Â Â Â Â uioinfo->version = pdev->dev.of_node->full_name;
> >> >>>I don't think this is apropriate, but will leave that to Hans.
> >> >>I was thinking what to add and I choose full_name because I can read
> >> >>this value and identify which UIO is this device.
> >> >>I know that there should be version but there is no version string in DTS.
> >> >
> >> >The purpose of uio_info->version is to give the userspace part of the driver
> >> >additional information. Kernel part and userspace part might be developed
> >> >independently, and there should be a chance for the userspace part to find
> >> >out if a certain feature is already supported by the kernel part without
> >> >having to do dirty kernel version checks.
> >> >
> >> >So, uio_info->version is an information about the driver, not the hardware.
> >> >
> >> >Example: You write a UIO driver for a chip you use in a project. You don't
> >> >need all the functionality of that chip. One year later you need additional
> >> >chip functionality, and it turns out that you have to do certain
> >> >initializations in the kernel part. Your new userspace will need the new
> >> >kernel driver, but there are lots of older kernels around in your customers
> >> >devices. In that case, your userspace part can simply check the version
> >> >string in sysfs and require at least your new version.
> >>
> >> I understand reasons but this information is not in device tree and
> >> it must be setup.
> >> Grant suggested compatible string but it is not the best option too.
> >
> > In uio_pdrv_genirq, uio_info->version is hardcoded in platform data. Hardware
> > initialization can also take place in the same platform specific file, which
> > is common practice on archs like ARM. Therefore, a driver specific versioning
> > can make sense for UIO, even if the driver code itself doesn't change.
> >
> > If you have no equivalent for that in device tree, you should create a new
> > generic driver (uio_of_genirq?) that simply doesn't support this kind of
> > versioning.
> 
> I'd avoid that.  Current trend is to move away from separate
> of-specific data because the driver is essentially identical other
> than the data source being different (platform_data vs. a device node
> pointer).

The point here his that UIO drivers are _not_ like normal drivers, because
they're split into a kernel and a userspace part. If there are changes in the
hardware initialization the kernel performs, and the userspace part of the
driver needs to know about it, we use the "version" attribute in sysfs to
communicate that.

OK, userspace could check the kernel version. But UIO drivers are mostly used
in environments where custom non-mainline kernels reporting arbitrary version
numbers are running (less than 1% of the UIO drivers in use are ever posted
on LKML, I guess). That makes it at least difficult for these people to write
a clean UIO driver that can deal with different kernels.

Killing the "version" attribute means breaking some out-of-tree UIO drivers.
Personally, I'm fine with that. But completely throwing away the possibility
of that kind of versioning is one step too far IMHO.

For UIO, it is not enough to just know "we have this chip using that driver",
at least not for a generic driver like the one proposed in this patch.

Thanks,
Hans
--
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/