Re: per-cpu statistics (Re: CVS commit: src)

[yamt%mwd.biglobe.ne.jp@localhost (YAMAMOTO Takashi)]

> Jason Thorpe wrote:
> 
> > On Apr 10, 2008, at 8:59 PM, YAMAMOTO Takashi wrote:
> > 
> > > while converting existing sysctls is fine, we need to introduce a
> > > kernel-userland interface for currently kmem-based statistics
> > > (including evcnt) at some point.  i'd like to see something like
> > > solaris kstat rather than sprinkling random nodes in the sysctl tree.
> > 
> > That's certainly fine with me.
> 
> I've got a simple proof-of-concept sysctl for fetching evcnts, which is
> really the all-or-nothing approach.
> 
> I've just start looking at the Solaris kstat interface.  Is there really
> much of a difference between a sysctl node and a named kstat?  My
> first thought after a bit of reading is that we could add an interface
> to query evcnts by name, but then we're just back at sysctl nodes.
> At which stage, do we just do away with evcnts and use sysctl nodes
> anyways?
> 
> I'll keep reading up on kstat, but can someone easily summarise what
> kstat gives us over sysctl?
> 
> Cheers,
> Simon.

sysctl nodes for this purpose is abuse in the first place.
why do you prefer it?  do you know linux procfs mess? :)

YAMAMOTO Takashi