Re: SCSI device numbering (was: Re: Ideas for v2.1

Linus Torvalds (
Tue, 2 Jul 1996 10:33:47 +0300 (EET DST)

On Mon, 1 Jul 1996, Leonard N. Zubkoff wrote:
> So something like this:
> - one major number for each SCSI bus (= usually one controller)
> - 4 (8?) bits SCSI ID
> - 7 bits SCSI LUN
> - 4 bits partition information
> - ???
> Yuk. Why should the dev_t need to encode these subsystem specific details?

This wouldn't be a "dev_t" issue: as far as the _rest_ of the kernel is
concerned, dev_t is just a 32-bit number. That goes without saying: I
will totally ignore any patches that try to force any "regular" stuff on
the device numbers.

What I mean above is just what the SCSI drivers would use _internally_.
External to the SCSI drivers, nothing even knows or cares..

(We do want some kind of SCSI rules for the minor numbers, because we
want the device numbers to at least look _similar_ even if we change
controllers. But this is really internal to SCSI, not a "kernel" issue
per se).

> I'd much rather see:
> - 12 bits for the major number of each device type pretty much as we
> have now.
> (device types are the things that have a common set of properties)
> - 16 bits for an index into a dynamically allocated dense array of
> pointers to objects of that type
> - 4 bits for partition information or other specialization

NO! You seem to expect the kernel to actually care. The kernel should
_not_ care at all about the actual layout of the device numbers. If some
laser-guided-missile driver wants to use the 20 bit minor number for
targeting information, it can do so. The kernel doesn't care.

We're discussing only how the _internals_ of the SCSI layer decode the
number it gets passed. For the rest of the kernel, there just needs to be
some simple interface to get some specific information, for example
something like:

unsigned long dev_size(kdev_t); /* 512-byte sectors */
int dev_readonly(kdev_t); /* readonly: != 0 */
int check_disk_change(kdev_t); /* disk change info */

and then how the actual device lookup is done should be done in the
driver (just add a few fields to the "struct device_struct", and let the
drivers fill them in at register time, the same way it fills in the
"struct file_operations *" fields).

> I think that it's the *devices* that are the central objects here, not their
> implementation on specific SCSI channels. I don't see any reason that a dev_t
> needs to map directly to any SCSI details. Each SCSI Device can refer to the
> specific bus, the driver that handles that bus, and whatever details are
> necessary to address that device.

There is no need at all to even discuss internals of "dev_t" without
specifying a driver: "dev_t" _has_ no internals (apart from the number of
bits for major/minor numbers, and even that is mainly just a interface
issue) on it's own. Only when talking about a specific driver does it
make sense to try to specify what the numbers mean.

And the numbers _will_ differ from driver to driver: IDE disks use 6 bits
for partition information, while the floppy driver uses no bits at all,
and instead uses 2 bits for specifying which floppy and 6 bits for
specifying floppy type. I'm not going to accept any patches that try to
"regularize" this - making up rules just for the fun of having rules
makes no sense.

HOWEVER, within the SCSI subsystem we _do_ need to have some rules,
because a user that changes his SCSI controller from a NCR controller to
a BusLogic controller does _not_ want his /dev setup to change: his disks
didn't change. So for SCSI it makes sense to make some rules for how the
drivers should interpret the minor numbers.

> I think 65536 devices of each type is probably sufficient for the moment. If
> not, we can shrink the major number a bit more. The reason 32 bits aren't
> enough in most of the suggestions I've seen is that we're trying to encode
> irrelevant and unnecessary information into the minor numbers. dev_t does not
> need to be the way of naming devices from the user's perspective.

dev_t _IS_ the way of naming devices when it comes to the kernel. There
simply _is_ no other naming. The fact that you can then do stuff like

mknod b 256 0x010101 /dev/scsi/id1lun1part1

or whatever is irrelevant (although that also needs to be standardized,
but that is no longer a kernel issue). The same goes for the names you
suggested (/dev/scsi/DEC_RRD4 or whatever) - those can be generated
on-the-fly at boot time. But when it comes to the kernel, all of this is
irrelevant, and the only thing that means anything is the device number.

The current scheme works fine, where we give device numbers _completely_
dynamically. There is nothing inherently wrong with this scheme together
with the scripts that Eric has written to create the user-level names for
the devices. The reason people want to change that scheme is that it's a
bit _too_ dynamic: the device numbers tend to change during the lifetime
of the machine, even if the hardware doesn't change (just unload the
driver, turn off one device, re-load the driver and voila, all the
device numbers have changed ;-). And some other trivial setup changes
also tend to result in more problems than they really should.

But if anybody thought I have been advocating some kind of "global device
numbers" for everything (SCSI, laser pointers, network cards, robot arms,
keyboards..), think again. device numbers are supposed to be just ID's
for the hardware, and they shouldn't have any _inherent_ properties.