Realtime PI tests fail on SMP PREEMPT_RT after commit a05298e (“Get rid of autotools”)

[TamNguyen1507]

Environment:

Board: Linux board

Kernel: 6.1.141 (PREEMPT_RT)

Distro: Yocto

glibc: 2.39 / NPTL 2.39

LTP revision: a05298e (“testcases: realtime: Get rid of autotools”)

Description of the problem:
After updating to commit a05298e, several real-time priority-inheritance tests started failing on our SMP PREEMPT_RT board. In particular, testpi-1, testpi-4 and testpi-5 now fail, whereas they previously passed. Other tests (like testpi-0, testpi-2, testpi-6) still pass.

From the logs, the low-priority thread holding the mutex stays at pthread pol 0 pri 0 for the entire run, while the SCHED_FIFO noise thread at prio 40 loops continuously without being preempted. This indicates that priority inheritance and/or CPU affinity setup is no longer working as expected after the autotools cleanup. These tests assume all threads are on one CPU to observe PI behavior.

Expected result:
On a PREEMPT_RT system with PI support, testpi-1, testpi-4 and testpi-5 should pass; the low-priority thread should be boosted and preempt the high-priority noise thread as the parser expects.

Actual result:

testpi-1 and testpi-4 end with “Result: FAIL”.

testpi-5 ends with “appears to have completed” but no “Result: PASS”.

Best Regards,
Tam

[metan-ucw]

Are you sure that you have the right commit?

I've been staring at the diff for a while and the only relevant change for pi tests is:

@@ -113,15 +112,9 @@ extern double pass_criteria;
  */
 static inline int atomic_add(int i, atomic_t *v)
 {
-       /* XXX (garrcoop): only available in later versions of gcc */
-#if HAVE___SYNC_ADD_AND_FETCH
        return __sync_add_and_fetch(&v->counter, i);
-#else
-       printf("%s: %s\n", __func__, strerror(ENOSYS));
-       exit(1);
-       return -1;
-#endif
 }

And that is no-op as long as you have compiler that is officially supported by LTP.

[TamNguyen1507]

Hi @metan-ucw,

Thank you for your response.

I have run the pi-tests on the same kernel using both LTP20240524 and LTP20250530. With LTP20240524, the results matched my expectations, but with LTP20250530 the tests failed. Therefore, I started investigating which commit might have caused this failure.

I found that commit a05298e is a likely cause. When I checked out a05298e and ran the pi-tests, the result was FAIL, whereas with the previous commit 998df1a, the result was PASS. This led me to believe that a05298e is responsible for the failing test case.

I also noticed that in this commit the line including realtime_config.h was removed. Upon checking, I found that this header defines some macros, such as #define HAS_PRIORITY_INHERIT 1, which is used in testpi-5.c. I manually added this file back (with the same content as in LTP20240524) and included it again in testcases/realtime/include/librttest.h and testcases/realtime/lib/libstats.c.

After rebuilding and rerunning the tests on my target board, the results passed as expected.
Best regards,
Tam

[metan-ucw]

I see the problem now, that commit should have removed the HAS_PRIORITY_INHERIT ifdefs as well. I will send a patch to LTP mailing list.

[metan-ucw]

Patch send for a review https://patchwork.ozlabs.org/project/ltp/patch/20250925135448.19818-1-metan@ucw.cz/

[TamNguyen1507]

Hi @metan-ucw
Thank you for your support.
The result was as I expected after applying your patch file.
Best regards,
Tam

[metan-ucw]

Thanks for the quick test, patch has been merged.