Re: devfs

Perry Harrington (
Sun, 18 Jan 1998 00:05:50 -0800 (PST)

The primary argument (beside what single digit terminology to use) seems
to be "what if we extend this, or want to support this". The goal of the
canonical naming scheme is to hierarchly address devices in the system, in
a logical, consistent manner. I propose this solution:

Obviously Solaris left some things out. We want support for Host->channel,
etc, not a problem. However the notation of partitions, slices, etc is
going to be a hairball. My proposal is that the device name be created
all the way up to the lun, eg: h0c0t0d1 and then make that a kernel maintained
(device driver) directory. To access the device directly, ideally you'd
just access the directory, but that may not be possible. So, simply putting
a default file called '0' in the directory would make accessing that SCSI
device: h0c0t0d1/0

Then, when all of the modules for supporting partitions, slices, BSD disklabels,
etc are implemented, they just specify to add their own logical subnaming to
the directory. That way it unclutters the directory where the device names
are stored, it assures a basic logical consistency for device access, and it
provides for developer level kernel modules which do whatever disk formatting
scheme they see fit.

This solution eliminates the endless holy wars over 's'lice, 'p'artition, and
any other, possibly commercial vendor created, disk organization scheme.

Just think, if we could say to (Veritas?): "yeah, you don't have to dink with
partitions, slices, etc, you can simply create your own (braindamaged?) disk
organizational scheme, and it works like this on all systems and doesn't
require any non-standard kernel hacks". Linux would be better than any
commercial OS, and we could say that the meritocracy method of debate and
thought brought them an OS with the dainbramage optional! :)


> On Thu, Jan 15, 1998 at 05:47:45PM -0500, James Bottomley wrote:
> >
> > sdc: sdc1 <solaris: [s0] sdc5 [s1] sdc6 [s3] sdc7 [s5] sdc8 [s6] sdc9 >
> >
> > showing I have one MS-DOS partition on sdc1 containing my solaris slices.
> >
> > I haven't actually got around to looking at the devfs patches yet. When I do
> > I'll try to hammer the solaris x86 slice recognition code into a form which
> > can be used by devfs. You can see that if one disc has both solaris and linux
> > extended partitions, we rapidly run out of minor numbers.
> This is where we should use p1s0 to refer to sdc5.
> I still think that the device names should reflect the nature of the disk.
> So the sdc1 partition should be xxxxxxp1, it's slices are xxxxxxp1s0,
> xxxxxxp1s0, etc. And if there weren't any slices then xxxxxxp1 is it.
> We should NOT come up with a naming scheme that can name every device and
> just fill in 0's. It just doesn't make sense. This is the perfect example
> of why.
> I suppose that somebody will retort with something about logical drives in
> extended partitions needing a naming scheme so that what we now call hda5
> becomes xxxxxxp1e1. I think that's silly. It just doesn't reflect how DOS
> style partitions are used. But p1 should refer to the whole extended
> partition and p5, p6, etc should be the logicals, exactly as we have
> envisaged it so far.
> --
> [ Kevin Lentin Email: ]

Perry Harrington       Linux rules all OSes.    APSoft      ()
email: 			Think Blue. /\