LKET tests

LKET tests

[Frank Ch. Eigler]

Hi -

Now that LKET has had some time to grow, it appears to be time to step
back and extend the test suite to focus on LKET code.  For a first
step, this may need little more than a buildok test that instantiates
all the probes.  For serious testing though, there should be a proper
pass-5 test that runs some set of them for real.

- FChE

resolving probe point context_switch

[Jason Yeh]

Hi,

I have been trying to get Systemtap to place probe in the scheduler and 
hoping to experiment using Systemtap to emulate certain behavior of 
Perfmon2 similar to what Frank mentioned in the email titled 
"Model-Specific Register". Running the example sched_snoop.stp was 
successful only on Athlon XP 32-bit laptop running Suse 10.1 using 
2.6.16.13 kernel. Opteron machines running kernel 2.6.18 on Ubuntu 
(compiled using gcc 4.1), Fedora Core 6, and Suse 10.1 will end with 
subsets of the following errors:

semantic error: no match for probe point
while: resolving probe point kernel.inline("context_switch")
semantic error: no match for probe point
while: resolving probe point scheduler.cpu_off
semantic error: no match for probe point
while: resolving probe point kernel.inline("finish_task_switch")
semantic error: no match for probe point
while: resolving probe point scheduler.cpu_on
semantic error: no match for probe point
while: resolving probe point kernel.inline("pull_task")
semantic error: no match for probe point
while: resolving probe point scheduler.migrate
semantic error: no match for probe point
while: resolving probe point kernel.inline("idle_balance")
semantic error: no match for probe point
while: resolving probe point scheduler.balance

I have seem "pull_task" being problematic in previous emails, but could 
not find any workaround mentioned. Is there any workaround for these 
unresolved probes? Thanks.

Jason

Re: resolving probe point context_switch

[Frank Ch. Eigler]

"Jason Yeh" <jason.yeh@amd.com> writes:

> [...]  I have been trying to get Systemtap to place probe in the
> scheduler and hoping to experiment using Systemtap to emulate
> certain behavior of Perfmon2 similar to what Frank mentioned in the
> email titled "Model-Specific Register".

Thanks for giving that a try!

> [...]
> while: resolving probe point kernel.inline("context_switch")
> while: resolving probe point kernel.inline("finish_task_switch")
> while: resolving probe point kernel.inline("pull_task")
> while: resolving probe point kernel.inline("idle_balance")

We have some known weaknesses probing inline functions, which we may
be able to improve significantly.  In the mean time, something coarser
might have to do:

      probe kernel.statement("schedule") { /* leaving tid() on cpu() */ }
      probe kernel.function("schedule").return { /* resuming tid() on cpu() */ }

FWIW, the scheduler is one of the spots that would benefit most from
static instrumentation markers.


- FChE