On 11/21/2014 03:44 PM, Ciprian Barbu wrote:
On Fri, Nov 21, 2014 at 1:31 PM, Taras Kondratiuk
<[email protected]> wrote:
On 11/20/2014 09:02 PM, Ciprian Barbu wrote:
Signed-off-by: Ciprian Barbu <[email protected]>
---
The testcases are based almost entirely on the odp_example.
There are no alloc tests and I added a test case for
odp_schedule_wait_time.
The major differencs between the odp_example and this cunit is the
partition
into testcases, the odp_example calls every test case from one big
function.
I had to work some magic in order to be able to pass arguments to test
cases,
I hope is not too hard to follow.
configure.ac | 1 +
test/cunit/Makefile.am | 2 +
test/cunit/schedule/Makefile.am | 10 +
test/cunit/schedule/odp_schedule_test.c | 844
++++++++++++++++++++++++++
test/cunit/schedule/odp_schedule_testsuites.c | 35 ++
test/cunit/schedule/odp_schedule_testsuites.h | 21 +
6 files changed, 913 insertions(+)
create mode 100644 test/cunit/schedule/Makefile.am
create mode 100644 test/cunit/schedule/odp_schedule_test.c
create mode 100644 test/cunit/schedule/odp_schedule_testsuites.c
create mode 100644 test/cunit/schedule/odp_schedule_testsuites.h
diff --git a/configure.ac b/configure.ac
index fcd7279..a47db72 100644
--- a/configure.ac
+++ b/configure.ac
@@ -173,6 +173,7 @@ AC_CONFIG_FILES([Makefile
test/Makefile
test/api_test/Makefile
test/cunit/Makefile
+ test/cunit/schedule/Makefile
pkgconfig/libodp.pc])
AC_SEARCH_LIBS([timer_create],[rt posix4])
diff --git a/test/cunit/Makefile.am b/test/cunit/Makefile.am
index 439e134..b6033ee 100644
--- a/test/cunit/Makefile.am
+++ b/test/cunit/Makefile.am
@@ -3,6 +3,8 @@ include $(top_srcdir)/test/Makefile.inc
AM_CFLAGS += -I$(CUNIT_PATH)/include
AM_LDFLAGS += -L$(CUNIT_PATH)/lib -static -lcunit
+SUBDIRS = schedule
+
Don't create a separate makefile, so all test binaries will be the same
directory.
Did you get that feedback on private? I don't see it in the comments.
Anyway, I can drop the extra Makefile no problem.
Anders complained that 'make check' has some issues with this. And I've
noticed that Alex have changed Crypto tests in this way.
if ODP_CUNIT_ENABLED
TESTS = ${bin_PROGRAMS}
check_PROGRAMS = ${bin_PROGRAMS}
+
+struct CU_TestInfo schedule_tests[] = {
+ _CU_TEST_INFO(schedule_wait_time),
+ _CU_TEST_INFO(schedule_one_single_lo),
+ _CU_TEST_INFO(schedule_single_lo),
+ _CU_TEST_INFO(schedule_one_many_lo),
+ _CU_TEST_INFO(schedule_many_lo),
+ _CU_TEST_INFO(schedule_multi_lo),
+ _CU_TEST_INFO(schedule_one_single_hi),
+ _CU_TEST_INFO(schedule_single_hi),
+ _CU_TEST_INFO(schedule_one_many_hi),
+ _CU_TEST_INFO(schedule_many_hi),
+ _CU_TEST_INFO(schedule_multi_hi),
+ CU_TEST_INFO_NULL,
+};
schedule_one() will be dropped, so these tests also can be dropped.
Yes I know I had to drop them. I kept them in for this RFC for easy
comparison against odp_example
I think odp_example is not a good base for scheduler tests. It was
written as benchmark, but not as a verification test.
That's actually not the feedback I got from Mike, correct me if I'm wrong.
My main concerns about odp_example:
- it does millions of iterations which is not necessary for functional
verification (at least for basic).
- it makes wrong assumption about scheduler fairness.
- it doesn't check main functional features mentioned in a quote below.
It would be better to start from scratch, that try to modify
odp_example.
I imagine basic scheduler testing in following steps:
1. One queue, single thread. odp_schedule() return all enqueued buffers.
Test each queue type (PARALLEL, ATOMIC, ORDERED).
2. Many queues, single thread. odp_schedule() return all enqueued
buffers. Verify buffer source queue. Test each queue type.
3. Queues with different priorities, single thread. odp_schedule()
return all buffers according to queue priorities. Test each queue
type.
4. Same as 3 but multi-threaded.
5. One ATOMIC queue, several threads. Verify that only one thread at a
time can get a buffer from that queue.
6. Same as 5, but use odp_schedule_release_atomic() and check that one
more thread could get an event. That is an optional test, because
odp_schedule_release_atomic() is a hint and may be ignored by
platform.
7. Two queues one of them ORDERED, several threads. Verify that buffers
scheduled from ORDERED queue are enqueue into the second queue in
correct order.
8. Test scheduler timeout APIs.
Basic scheduler tests are missing:
- verify that priority works correctly.
- verify that atomicity works correctly for ATOMIC queues.
- verify order for ORDERED queues.
That's good input, thank you.
_______________________________________________
lng-odp mailing list
[email protected]
http://lists.linaro.org/mailman/listinfo/lng-odp
