[cgl_discussion] Announcing Non-intrusive Kernel Monitor Project

[Rusty Lynch]

On Thu, Jun 10, 2004 at 03:28:26PM -0700, Steven Dake wrote:
> patch looks acceptable..  Although character drivers that implement
> ioctls are poorly received...  Have you thought of a sysfs mechanism to
> plug this into the kernel?
> 
> Thanks
> -steve

I kind of expected to get push back on the char device with ioctl mechanism for
registering events.  The way I was looking at this, this patch was more of 
an experiment to see how much of a system impact we would see for delivering
the key events via a signal.  Using a char device was just the quickest way
to get something working.

I'm not so sure sysfs is the right answer.  The registration/deregistration
stuff doesn't seem to map very well to sysfs.

Other ideas are:
  - add new system calls
  - add a proc entry
  - add a netlink interface 

One thing I like about the char device is that as soon as the device is
closed, we can nuke all the registration request the monitor made.  I guess
the same thing could be done with the other methods, but it wouldn't be
as simple.

So... I don't know.  I was hoping to hear some strong arguments here on 
how to [de]register (ie. char device, system call, etc.)

    --rusty   

> 
> On Thu, 2004-06-10 at 15:19, Rusty Lynch wrote:
> > On Wed, Jun 09, 2004 at 10:19:03AM -0700, Sabharwal, Atul wrote:
> > > I have been experimenting with the CGL 2.0 AVL 8.2 requirement on
> > > Non-intrusive kernel
> > > based monitoring of processes and threads.  For your reference, the
> > > requirement is
> > > quoted below :
> > 
> > Looks like outlook ate the patch.  Here is the patch again (without the
> > help of outlook)
> > 
> > BTW, is this how everyone is interpreting the requirement?  There isn't much
> > to this and the performance/scalability is basically what ever real-time
> > signals can do.
> > 
> > I would at least like to hear if this is on or off target for the requirement
> > before Atul takes this to LKML.  We have played with this on various ia32 and
> > IPF systems with hardly noticeable impact while monitoring a significant load,
> > and no measurable impact while not monitoring anything.
> > 
> >     --rusty