Steven Rostedt wrote:
On Fri, 10 Jul 2009, Darren Hart wrote:
Subrata Modak wrote:
On Fri, 2009-07-10 at 22:00 +0530, Sripathi Kodi wrote:
On Wednesday 08 July 2009 23:43:53 Subrata Modak wrote:
Darren/Sri/Gowri,
Where do you want me to put this exactly inside the RT tree ?
Hi Subrata,
Going by how the tests are organized currently, I think this should go
into it's own directory under testcases/realtime/func. We will need to
add a makefile to it. Are you looking at us to help you with this?
Correct. Please send me a patch which integrates it into RT tests build,
install & run.
Just got back from a week vacation and am burning through mail as fast as I
can :-) Haven't had a look yet, but does this test use librttest.h? I suspect
not. We'll need to adapt it to run within the existing ltp real-time testing
framework, which includes things like buffered output as well as mlocking
support.
Lastly, I'm not sure this test does anything effectively different than
prio-wake, already in the tree. My other concerns with the test are its
explicit 1ms preemption criteria (as Steven described it anyway). We are
trying to move away from criteria being inherent in measurement tests, and 1
ms seems like an awfully long priority inversion to be an acceptable criteria
to many users.
Steven, am I missing something conceptually here?
Hmm, I missed this email, sorry for the late reply.
What does prio-wake do?
This test is what I used to develop the rt scheduler in mainline (as well
as in -rt). It wakes up N+1 tasks with lowering real time priorities.
Where N is the number of CPUs in the system. Then it makes sure that the
these tasks spread out across the CPUs. Most tests just test the highest
priority task in the system. But those tests usually miss the second
highest prio task in the system. If you have a second highest prio task in
the system and a CPU is available to run, then it should run on that CPU.
But what happens is that it can wait to be migrated and can take millisecs
to wake up.
This test makes sure that all the high prio tasks that are in the running
state are actually running on a CPU if it can.
Make sense?
Yup. This is different than prio-wake. Prio-wake creates some number
of threads of varying priorities and puts them to sleep on
pthread_cond_wait and then wakes them with pthread_cond_broadcast, then
tests that the threads were woken in priority order. Fails on
kernels/glibc without requeue pi support as a FUTEX_WAKE(all) was used
for PI mutexes.
So to include this in the ltp/restcases/realtime suite, the test should
make use of the librttest.c and libstats.c apis, standardized argument
parsing, buffered logging mechanism, and standardized output formatting.
I can't get to that for a little while myself, but can assist if you
have any questions. Alternatively we can wait a little while, and
perhaps someone on my team will be able to help merge it with the
realtime test infrastructure in LTP.
(BTW, current -rt and mainline now fail this test :-? )
Hrm, uh oh. Adding it to my lengthy list of things to try and look at :/
Thanks Steve,
--
Darren Hart
IBM Linux Technology Center
Real-Time Linux Team
--
To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
