Re: [...] kprobes [minutes] portion 20060406

Re: [...] kprobes [minutes] portion 20060406

[Frank Ch. Eigler]

Hi -

On Thu, Apr 06, 2006 at 02:14:46PM -0700, Hien Q Nguyen wrote:
> User space probe
> Prasanna had prototype that execute the handler in the user space [...]

Beyond merely dispatching, it is important to consider what such a
handler can do.  Would it still be useful if it cannot call into the
kernel-side runtime, if it cannot share variables with kernel-space
probes?


- FChE

Re: [...] kprobes [minutes] portion 20060406

[Richard J Moore]

Unfortunately I missed the discussion. Am I right in deducing that
user-space probes as being discussed here are designed or intended as
probes for use with an application's perspective? If so, I believe that is
wrong or at least only attending to part of the need. For me the real
benefit of user-space probes is to provide an  extension to a kernel probe
that gives it full access to the system.  The arguments on lkml put forward
by Christoph and Arjan have value when one thinks of a user-probe only
serving user-space code and a user-space problem.  But we have tools
already in place to deal with that.  Have we fully appreciated the need
when addressing certain system or kernel problems that one needs access to
full system resources, which will include user-space code and data.  Anyone
who works on performance problems (hangs being an extreme example)
involving the networking stack will corroborate this need. An more
generally anyone that works with a kernel-based server component of a
client-server model will also corroborate this need.

Richard

- -
Richard J Moore
IBM Advanced Linux Response Team - Linux Technology Centre
MOBEX: 264807; Mobile (+44) (0)7739-875237
Office: (+44) (0)1962-817072


                                                                           
             "Frank Ch.                                                    
             Eigler"                                                       
             <fche@redhat.co                                            To 
             m>                       Hien Q Nguyen <nguyhien@us.ibm.com>  
             Sent by:                                                   cc 
             systemtap-owner          systemtap@sources.redhat.com         
             @sourceware.org                                           bcc 
                                                                           
                                                                   Subject 
             06/04/2006               Re: [...] kprobes [minutes] portion  
             22:28                    20060406                             
                                                                           
                                                                           




Hi -

On Thu, Apr 06, 2006 at 02:14:46PM -0700, Hien Q Nguyen wrote:
> User space probe
> Prasanna had prototype that execute the handler in the user space [...]

Beyond merely dispatching, it is important to consider what such a
handler can do.  Would it still be useful if it cannot call into the
kernel-side runtime, if it cannot share variables with kernel-space
probes?


- FChE

Re: [...] kprobes [minutes] portion 20060406

[Prasanna S Panchamukhi]

On Thu, Apr 06, 2006 at 05:28:38PM -0400, Frank Ch. Eigler wrote:
> Hi -
> 
> On Thu, Apr 06, 2006 at 02:14:46PM -0700, Hien Q Nguyen wrote:
> > User space probe
> > Prasanna had prototype that execute the handler in the user space [...]
> 
> Beyond merely dispatching, it is important to consider what such a
> handler can do.  Would it still be useful if it cannot call into the
> kernel-side runtime, if it cannot share variables with kernel-space
> probes?

No. The purpose of prototype was to see how it can be implemented.
To my opinion having handler run in user-space does not suit well for
system-wide tracing. But that might be one of the alternative approaches
to per-process tracing, since ptrace is found to have more performance
related problems. 


Thanks
Prasanna
-- 
Thanks & Regards
Prasanna S Panchamukhi
Linux Technology Center
India Software Labs, IBM Bangalore
Email: prasanna@in.ibm.com
Ph: 91-80-51776329

Re: [...] kprobes [minutes] portion 20060406

[Richard J Moore]

systemtap-owner@sourceware.org wrote on 07/04/2006 09:25:17:

>
> On Thu, Apr 06, 2006 at 05:28:38PM -0400, Frank Ch. Eigler wrote:
> > Hi -
> >
> > On Thu, Apr 06, 2006 at 02:14:46PM -0700, Hien Q Nguyen wrote:
> > > User space probe
> > > Prasanna had prototype that execute the handler in the user space
[...]
> >
> > Beyond merely dispatching, it is important to consider what such a
> > handler can do.  Would it still be useful if it cannot call into the
> > kernel-side runtime, if it cannot share variables with kernel-space
> > probes?
>
> No. The purpose of prototype was to see how it can be implemented.
> To my opinion having handler run in user-space does not suit well for
> system-wide tracing. But that might be one of the alternative approaches
> to per-process tracing, since ptrace is found to have more performance
> related problems.
>
>
 Agreed, but don't let the creation of per-process probe handlers detract
from the argument for a system-wide capabiltiy; that's the real
battle-ground.

Richard


> Thanks
> Prasanna
> --
> Thanks & Regards
> Prasanna S Panchamukhi
> Linux Technology Center
> India Software Labs, IBM Bangalore
> Email: prasanna@in.ibm.com
> Ph: 91-80-51776329