Re: adding OF_DYNAMIC proc interface

[Benjamin Herrenschmidt <benh-XVmvHMARGAS8U2dJNN8I7kB+6BGkLq7r@public.gmane.org>]

On Fri, 2012-09-28 at 11:46 -0500, Alan Tull wrote:
> Hello,
> 
> The following patch adds a /proc/ofdt interface to add or remove device tree
> nodes dynamically.
> 
> Based on earlier feedback, I've changed my driver to use /proc instead of
> creating a new ioctl (the old thread is at 
> http://www.mail-archive.com/devicetree-discuss-uLR06cmDAlY/bJ5BZ2RsiQ@public.gmane.org/msg17333.html)
> 
> 
> I was hoping to get some early feedback from others who might be interested
> who were discussing this on an earlier thread about OF_DYNAMIC usage.
> 
> This code doesn't do any notification for drivers yet. It can add multiple
> nodes and they will show up properly under /proc/device-tree.  It has an
> issue that shows up when removing nodes (it appears that the memory used by
> proc gets corrupted after the add).

(Adding Arnd here)

Have you guys considered whether a better approach would be a file
system ? IE, create a node by creating a directory, add files for
properties etc... ?

It might need some trick to make the node "active" (in order to not
internally in the kernel start exposing unfinished nodes), maybe a
special file, maybe a permission trick ...

It should be at least considered rather than a new interface that tries
to re-implement semantics (such as notification) that are pretty much
already provided by a fs.

Also, what about the existing /proc/device-tree ?

Finally, if we are doing something new and for some reason not a fs,
what about using sysfs rather than /proc ? Really /proc is not a great
idea here.

Cheers,
Ben.