question about drivers/infiniband/core/cm.c's kobject usage

question about drivers/infiniband/core/cm.c's kobject usage

[Greg KH]

Hi,

I was working on some changes to the driver core that is cleaning up the
struct class fields, when I ran accross the usage of cm.c and the
infiniband_cm class.

It looks like you are registering "raw" kobjects in this class, chaining
things off of it, as if they were devices.

If so, why not just use struct device in the first place?  You are
creating a tree, which on modern distros, userspace will never see as
they are expecting everything to be showing up in /sys/devices/

Entries in /sys/class/*/* now are symlinks into the /sys/devices tree,
showing the representation of everything in one tree, not lots of little
trees all over the place.

So I was curious, was this done on purpose?  If so, why?  If not, any
objection to me switching it over to be using struct device properly?

thanks,

greg k-h

RE: question about drivers/infiniband/core/cm.c's kobject usage

[Sean Hefty]

>So I was curious, was this done on purpose?  If so, why?  If not, any
>objection to me switching it over to be using struct device properly?

It's entirely possible I have this wrong, but the intent is to export some
infiniband communication management message counters and relate them to the
corresponding ib_device/port.  For example:

/sys/class/infiniband_cm/<device name>/<port number>/<counter group>/<counter>

(E.g. /sys/class/infiniband_cm/mthca0/1/cm_tx_msgs/req)

If there's a better way to handle this, I have no objection to changing it.

- Sean

Re: question about drivers/infiniband/core/cm.c's kobject usage

[Greg KH]

On Thu, May 22, 2008 at 02:47:51PM -0700, Sean Hefty wrote:
> >So I was curious, was this done on purpose?  If so, why?  If not, any
> >objection to me switching it over to be using struct device properly?
> 
> It's entirely possible I have this wrong, but the intent is to export some
> infiniband communication management message counters and relate them to the
> corresponding ib_device/port.  For example:
> 
> /sys/class/infiniband_cm/<device name>/<port number>/<counter group>/<counter>
> 
> (E.g. /sys/class/infiniband_cm/mthca0/1/cm_tx_msgs/req)
> 
> If there's a better way to handle this, I have no objection to changing it.

Yes, just hang all of the <port number> stuff off of the original
<device name> struct device.  That seems like it would be much simpler.

thanks,

greg k-h