Simplifying devfs: minor == unit

[Kostik Belousov]

On Tue, May 27, 2008 at 03:06:15PM +0200, Ed Schouten wrote:
> Hello everyone,
> 
> Right before I left to BSDCan I was looking at the devfs code. When I
> started hacking the TTY code, I discovered minor/unit numbers are still
> actively used within the FreeBSD kernel, even though they are never
> exposed to userspace. Devfs automatically generates an inode number for
> each device. Right now, st_rdev is always equal to st_ino, which still
> guarantees device numbers are unique throughout the system.
> 
> In an experimental branch in Perforce, I decided to see what would
> happen if I would completely remove minor numbers from device drivers.
> This means make_dev()'s minor argument is removed, but also the minor(),
> unit2minor(), etc. functions. Drivers could use si_drv0 directly, just
> like si_drv1 and si_drv2 are used right now.
> 
> This doesn't seem to be possible because of the design of the device
> cloner (not the eventhandler, just the clone_* routines), which
> preallocates an unnamed device with a specific unit numer, which can
> later be passed to make_dev().
> 
> This is why I want to do this in little steps right now. I was thinking
> about doing the following:
> 
> - si_drv0 currently contains the minor number. We could alter the
>   minor2unit(), etc. routines to make minor numbers equal to unit
>   numbers. This means most routines will now become a no-op. See
>   attachment.
> 
> - When that hits the tree, we could decide to run a big regexp on the
>   source code to make drivers use si_drv0 directly.
> 
> - I've seen most drivers only use the device cloner, because they need
>   descriptor local storage. It turns out more drivers need this than I
>   initially thought. kib@ has a patch for this, so I hope this gets
>   committed one of these {days,weeks,months}.
The patch was committed ~ a week ago.
> 
> - After we've got file descriptor local storage, I think we can live
>   without the cloner. This means we could consider removing the minor
>   number argument from make_dev(), removing the unique unit number
>   restriction we currently have inside devfs, which causes many drivers
>   to use number pools for no obvious reason.
I think we cannot live without clones regardless of devfs_cdevpriv.
The model assumed for the pty, snp and probably several other devices
actually requires new cdev instead of the priv data.

> 
> I was thinking about discussing this patch with my mentor + committing
> it somewhere in the nearby future. Any comments?

Making minor == unit number looks to be not a bad idea, please, look at
the saga of the tty_pty.c revs. 1.153, 156, 1.157. Making the devices use
si_drv0 directly probably is not so good since we remove the indirection
layer that is already present and allows for some (minor) freedom in the
devfs/kern_conf implementation.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 195 bytes
Desc: not available
Url : http://lists.freebsd.org/pipermail/freebsd-arch/attachments/20080527/7065b6b5/attachment.pgp