Dynamic major/minor numbers (or dropping them completely)

Dynamic major/minor numbers (or dropping them completely)

[Eduard-Gabriel Munteanu]

*This message was transferred with a trial version of CommuniGate(r) Pro*
Currently, the kernel has the following properties:
1) initramfs can be used to boot the system. We don't need any 
predefined /dev entries.
2) udev can be started from the initramfs to create the required entries 
in /dev. udev doesn't care about major/minor numbers.
3) Most distros already use udev and maybe initramfs. If there are 
exceptions, they can be easily converted.

For the first part, I'm asking: is there any reason why new char/block 
drivers shouldn't use dynamic major/minor numbers? Is there any reason 
against converting the whole kernel to dynamic major/minor numbers?

Okay, maybe the previous questions looked useless from a pragmatic POV. 
But why shouldn't the whole major/minor numbering system be dropped 
completely? sysfs already maintains a hierachy of device drivers and 
kernel subsystems, one which is better than the major/minor system. The 
current system could be replaced by a single-numbered, 
dynamically-allocated scheme.

Device files could be stored on a tmpfs filesystem, so that we don't 
make any changes to current filesystems. Apps won't need to be modified, 
since they access /dev entries by name, provided udev maintains the 
current naming scheme.

Any thoughts on this?

Re: Dynamic major/minor numbers (or dropping them completely)

[Chris Snook]

Eduard-Gabriel Munteanu wrote:
> *This message was transferred with a trial version of CommuniGate(r) Pro*
> Currently, the kernel has the following properties:
> 1) initramfs can be used to boot the system. We don't need any 
> predefined /dev entries.
> 2) udev can be started from the initramfs to create the required entries 
> in /dev. udev doesn't care about major/minor numbers.
> 3) Most distros already use udev and maybe initramfs. If there are 
> exceptions, they can be easily converted.
> 
> For the first part, I'm asking: is there any reason why new char/block 
> drivers shouldn't use dynamic major/minor numbers? Is there any reason 
> against converting the whole kernel to dynamic major/minor numbers?
> 
> Okay, maybe the previous questions looked useless from a pragmatic POV. 
> But why shouldn't the whole major/minor numbering system be dropped 
> completely? sysfs already maintains a hierachy of device drivers and 
> kernel subsystems, one which is better than the major/minor system. The 
> current system could be replaced by a single-numbered, 
> dynamically-allocated scheme.
> 
> Device files could be stored on a tmpfs filesystem, so that we don't 
> make any changes to current filesystems. Apps won't need to be modified, 
> since they access /dev entries by name, provided udev maintains the 
> current naming scheme.
> 
> Any thoughts on this?

You're correct that dynamic major/minor numbers are sufficient for most 
purposes, but embedded users really need their static numbers.  As for ripping 
out major/minor numberings, that's a non-starter.  Too much of our device 
management infrastructure is based around this numbering scheme, and there isn't 
really anything wrong with it to justify breaking everything in the change.

As a rule of thumb, if you ever find yourself wondering why we still support 
doing statically something we can now do dynamically, the answer is generally 
that doing it dynamically sucks for embedded.

	-- Chris

Re: Dynamic major/minor numbers (or dropping them completely)

[Al Viro]

On Fri, Aug 03, 2007 at 05:13:51PM -0400, Chris Snook wrote:
> You're correct that dynamic major/minor numbers are sufficient for most 
> purposes, but embedded users really need their static numbers.  As for 
> ripping out major/minor numberings, that's a non-starter.  Too much of our 
> device management infrastructure is based around this numbering scheme, and 
> there isn't really anything wrong with it to justify breaking everything in 
> the change.
> 
> As a rule of thumb, if you ever find yourself wondering why we still 
> support doing statically something we can now do dynamically, the answer is 
> generally that doing it dynamically sucks for embedded.

And not only embedded.  I'm quite happy _not_ running udev on anything
I have root on, except for one test box set exactly to make sure that
patches do not break things for udev-infested boxen.