Re: Multiple handlers per marker

[Mathieu Desnoyers <compudj@krystal.dyndns.org>]

* Mike Mason (mmlnx@us.ibm.com) wrote:
> Mathieu Desnoyers wrote:
> >* Mathieu Desnoyers (mathieu.desnoyers@polymtl.ca) wrote:
> >>* Mike Mason (mmlnx@us.ibm.com) wrote:
> >>>Hi Mathieu,
> >>>
> >>>Are you aware of any working being done to allow multiple handlers to be 
> >>>attached to a marker?  Something like what kprobes allows.  I've started 
> >>>to look into this and don't want to duplicate efforts.
> >>>
> >>Nope, but I know we will have to address this.
> >>
> >>Something along the lines of walking an RCU list of function pointers,
> >>calling them.
> >>
> >>The only downside I see is that we will have to pass a va_list * instead
> >>of real va args. The could make the marker site a little bit bigger and
> >>will change the probe callback arguments.
> >>
> >>What do you think about these ideas ?
> >>
> >>If we can find a way to make the common case (only one probe connected)
> >>_ultra_ fast, and yet architecture independent, that would be awesome. A
> >>simple call is kind of hard to beat though.. So we may have to think
> >>about a design with :
> >>
> >>- One call at the marker site
> >>- if 1 probe is installed :
> >>  - If the format string is empty, connect a probe without va args.
> >>  - If the format string is not empty, connect a "stage 1" probe that 
> >>  takes
> >>    the va args, starts/ends the va_list and calls _one_ function (let's
> >>    call it "stage 2" probe), that takes va_list as parameter.
> >>- if more than 1 probe is installed :
> >>  - The stage 1 probe creates the va_list and passes it to each function
> >>    connected, iterated with an RCU list.
> >>
> >>What do you think ?
> 
> Your proposal sounds reasonable to me.  How would marker_arm(), 
> marker_disarm() and marker_probe_unregister() change?  They'll need to work 
> on a per probe handler basis, rather than per marker.  They'll need to know 
> the marker name *and* the specific handler, or perhaps just the handler if 
> you keep a master list of all handlers.  In any case, the interface will 
> need to change.
> 

marker_arm/disarm: They already use a refcount, so no change is needed
:)

I just want to make sure that we can register/unregister probes when a
marker is armed.. I'll have to be careful about this.

My constraints are kind of difficult, but I think I can manage to
implement a good solution :
- Minimal memory footprint when disabled.
- Fast standard case (1 probe). Small memory footprint in that case.
- Use RCU-style updates to the markers structures (this one is hard).
  Requires atomic updates, quiescent states...
- Don't create va_list when a marker has no arguments.

I'm half-way there.. I'll keep you posted when I get to an interesting
solution.


> >>
> >>Mathieu
> >>
> >
> >I'm working on an implementation.
> 
> Great!  When you need reviewers/testers let me know.
> 
This is always appreciated.

Thanks,

Mathieu

> I'm copying this to the SystemTap list.  Multiple markers is a "must have" 
> feature for SystemTap.
> 
> Thanks,
> Mike
> 
> >
> >>>Thanks,
> >>>Mike
> >>-- 
> >>Mathieu Desnoyers
> >>Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
> >>OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 
> >>9A68
> >
> 

-- 
Mathieu Desnoyers
Computer Engineering Ph.D. Student, Ecole Polytechnique de Montreal
OpenPGP key fingerprint: 8CD5 52C3 8E3C 4140 715F  BA06 3F25 A8FE 3BAE 9A68