[BUG] AMD family 25 model 7 wrong regwidth

[brinkcoder]

We use AMD EPYC 9254 and 9454 in our cluster (both are Family 25, Model 15, Zen4). In power.c, for all ZEN4_EPYC processors the code sets

power_info.statusRegWidth = 64;

However, at least these two models only have a register width of 32. As a result, we observe a counter wrap every few hours, and the wrap value calculated with 64 bits yields incorrect power results.

I’m not sure how to best distinguish these processors in the code. A patch to correctly set the status register width to 32 for these models would be appreciated.

[TomTheBear]

Can you provide links to the official docs for these chips?

Please attach the output of /proc/cpuinfo of one CPU core per system so we can see how to differentiate the different types of Zen4.

[brinkcoder]

Ok, this is super weird. The official document says it is a 64 bit register (page 253):
55901_B2_pub_1.pdf
The cpuinfos:
cpuinfo_9254.txt
cpuinfo_9454.txt

But I added a function to perfmon.c so that I can see the raw values in likwidMetric.go:
diff_perfmon.c.txt
diff_likwidMetric.go.txt

You can see a wrap on thread 6 at 32 bit:
raw_counters.txt

[root@cpu001 cc-metric-collector]# grep "thread 6" raw_counters.txt

Raw counter value for event PWR0 on thread 6: 4289021270.000000

Raw counter value for event PWR0 on thread 6: 1696099.000000

I am clueless how to this can happen.

[TomTheBear]

It might be a documentation issue. Unfortunately, there is no other setting visible in the Linux kernel sources.

[brinkcoder]

I found the issue. statusRegWidth is correctly set to 64 for our ZEN4_EPYC.

The underlying issue is in likwid.h, lines 1694-1698, the typedef for PowerData. How could the 64bit counter be stored in a uint32_t?
That's why I observed a wrap-around at 32 bits every few hours. I changed it to uint64_t, and the problem is solved.

I'm not sure if it should generally be set to uint64_t or if you'd prefer using an ifdef (if that's even possible, as I'm not very familiar with C).

[TomTheBear]

Good find. It shouldn't be a big deal to update the PowerData to 64 bit.