Re: [Dprobes] Re: When Kprobes is fired ,how can we do performance measurement?

[Roberto J. D. <dr...@in...>]

If the probe is inactive, there is no overhead, since the original 
unmodified code is executed.

When the probe is registered, the instruction is replaced by an 
interrupt instruction, which singlesteps the original instruction and 
then/after the probe handler in its interrupt handler. The overhead of 
this interruption is really low (few instructions) but your probe 
handler can be quite heavy (pathological, as Richard calls). I think 
the probe is run in the interrupt handler and so cannot be preempted 
inside the kernel (I still don't understand how kernel preemption works 
very well).

To measure how long your probe takes, you could use the get_jiffies() 
kernel function in the begining and end of your probe, or modify the 
kprobes patch.



On 11/10/2004, at 06:01, lin shang wrote:

> Hi,Richard
>  
> Thank you for your feedback quickly.
>  
> What you said means the active probe's overhead related to the 
> effect which probes induced when system is running. In spite of the 
> overhead is very high,if probe and system execute normally we can 
> ignore it.  There is no need to be concerned about performance 
> overhead and there aren't  measure standard,too.
> Do I comprehend the meaning right?
>  
> Regards,
> shanglin

-- 
Roberto Jung Drebes <dr...@in...>
Porto Alegre, RS - Brasil
http://www.inf.ufrgs.br/~drebes/