This is the mail archive of the systemtap@sourceware.org mailing list for the systemtap project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Backward compatibility for insn probe point


Roland McGrath wrote:
>> I poked around the kernel source some more, but couldn't see what was
>> going wrong.  
> 
> I figured you'd use some stap probes to follow the code paths!

Oh I did, but so much of it is inlined that it was incredibly difficult
to follow.

>> I've changed the itrace code to stop the task after each step trap (so
>> that it acts more like ptrace).  I've tested this on several kernels
>> (2.6.18-141.el5/ppc, 2.6.18-128.1.10.el5/x86_64/i686, and
>> 2.6.25-14.fc9.ppc64) and it seems to work correctly.
>>
>> Does this seem like a reasonable work-around?  Could there be problems
>> with this approach?
> 
> I presume it kills performance.  But what works works, that's a what a
> work-around is.  I'd hope that you don't make it use this "not really
> right" mode for kernels with the modern utrace interface that doesn't have
> this bug.

Yes, this is only for old utrace.  As far as performance goes, I've
benchmarked single-stepping '/bin/ls' on x86_64 with both approaches.
Here's what I saw ('time' output, averages of 5 runs):

- no stopping on each step trap:
real   0m3.735s
user   0m0.328s
sys    0m3.359s

- stopping on each step trap:
real   0m4.101s
user   0m0.336s
sys    0m3.692s

I could also limit this work-around to ppc-only, to not penalize other
architectures.

One last thing.  I thought I'd try block stepping, so I got access to an
ia64 machine.  Unfortunately, using systemtap insn probes (either single
or block step) lock up the system with a spinlock lockup.  Sigh.

-- 
David Smith
dsmith@redhat.com
Red Hat
http://www.redhat.com
256.217.0141 (direct)
256.837.0057 (fax)


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]