Re: [PATCH] [PNP] 'modalias' sysfs export

[Adam Belay <ambx1@neo.rr.com>]

On Mon, Mar 13, 2006 at 05:14:58AM +0100, Kay Sievers wrote:
> 
> Macio solved the problem by adding all devices to a single string and
> let the device table match one of these id's in that single string:
>   http://www.kernel.org/git/?p=linux/kernel/git/torvalds/linux-2.6.git;a=blob;hb=HEAD;f=drivers/macintosh/macio_sysfs.c#l42
> 
> We should first check if that is possible for PnP too, or solve that
> problem in general at that level before we introduce such a hack.

I do have some concerns about merging every ID into a single string.  The
orginal design goal of having multiple IDs was to allow vendors to specify
a single high priority ID that a driver that supports the device's complete
feature set could match against.  If that driver is unavailable, it is
acceptable to search for other drivers that might match against a
compatibility ID and support a partial feature set.  Now if we just search
for the first driver that matches anything in a single ID string without
regard to the order IDs are presented, then we're not supporting the
specification.

More generally speaking, it seems to me there are four main options:

1.) We remove the modalias strings from all buses, and generate them in
userspace exclusively.  We may loose the ability to support new buses
without specialized userspace software, but we gain a great deal of
flexibility and can eventually implement more advanced hardware detection
schemes without depreciating existing kernel interfaces or parsing strings
that are limiting when compared to bus-specific data.  Also, at least we
have a uniform sysfs interface.

2.) We selectively export modalias strings on buses where this sort of
thing works and use hacks for other buses.

3.) We export multiline sysfs modalias attributes and tell userspace
hotplug developers that they're wrong and must change their assumptions.

4.) We export a single line modalias with each ID appended to the previous ID.
Userspace must pay careful attention to the order, but because the format is
bus-specific, it will have to be handled in a very specialized way. (e.g. PCI
has class codes, PnP has compatibility IDs, etc)

In the long run, I think option 1 is the best choice.  I'm more concerned with
flexibility than having a simplistic but limited hardware detection mechanism.
Also, I prefer to keep code out of the kernel when there isn't a clear
functionality advantage.  "file2alias" is not a kernel-level interface, but
rather implementation specific to modutils and various module scripts included
with the kernel source.  Therefore, I don't think that sysfs is obligated to be
specially tailored toward modprobe, even if it is convenient for some buses.

But I'm also interested in a practical short-term solution.  What are your
thoughts?  Would method #2 be acceptable?

Thanks,
Adam