Ryzen 3950X - Unknown S3 issue within fixed performance counters

[cyring]

AMD/Matisse

This is an unsolved issue about counters corruption after resuming processor from S3 (Suspend To Ram).

Looking into the UCC column, delta value of 18446744073709551615, the positive value of -1, let me think that the counter is stuck to reset.

UCC is bound to the x86 register MSR_IA32_APERF

CoreFreq/corefreqk.c

Line 10548 in 07069d5

MSR_CORE_PERF_UCC, Core->Counter[T].C0.UCC, \

The same test on a 5300U does not produce such issue!

Searching among the AMD specification updates has not revealed an errata.

Before S3

After S3

Unexpected workaround

Strangely, the effect tends to disappear over several periods and the MSR_IA32_APERF per CPU starts to count again, with occasional corruptions.

[cyring]

When resuming the higher part of counters APERF and/or MPERF is overflow with 0xFFFFFFFF

rdmsr -aX 0x000000e7

FFFFFFFFC0E01292
FFFFFFFFD3BFDE0D
FFFFFFFFC28CD6A6
FFFFFFFFC111C53A
FFFFFFFFC4B5D257
FFFFFFFFC1D4DC30
FFFFFFFFCA59D1F6
FFFFFFFFDF2B868A
...

So the fix is just to reset registers with a zero write.

wrmsr -a 0x000000e7 0x0