Hi,

 

I’m trying power saving feature in my server with intel sapphire rapid CPU. With “tsx=on” kernel parameter, I have both rtm and waitpkg available in CPU flag.

I tried rte_power_monitor_multi API to monitor on multiple address, but it never enters C0.2/C0.1 state.

To find the reason I patched the code to return the transaction abort code as below:

        /* transaction abort, possible write to one of wait addresses */

        if (rc != RTE_XBEGIN_STARTED)

-               return 0;

+               return rc;

and it turn out that RTM transaction always aborted with 0 (a few times with 4 or 6 also, which I think maybe expected due to memory change of monitored addresses).

I found some explanation of the abort state here:

https://www.intel.com/content/www/us/en/develop/documentation/cpp-compiler-developer-guide-and-reference/top/compiler-reference/intrinsics/intrinsics-for-avx2/intrinsics-for-tsx/intrinsics-for-restrict-transactional-mem-ops/restricted-transactional-memory-overview.html

For abort code 0, it just say one example (CPUID used within transcation), but it seems CPUID is not used within rte_power_monitor_multi.

I also tried rte_power_pause and rte_power_monitor, both works as expected, CPU enters C0.2 state (checked by perf stat --e r20ec -C xx)

 

Now, I don’t know why the RTM transaction aborted with 0 and don’t know how to debug it.

 

Can anyone throw some light on this?

 

Thank you.

 

Br, Xiaoping