cpuallocator, topology-aware: handle weird die setups better#643
cpuallocator, topology-aware: handle weird die setups better#643askervin merged 2 commits intocontainers:mainfrom
Conversation
askervin
left a comment
askervin
left a comment
There was a problem hiding this comment.
...and after reading the second commit, too, now I think I recall this case. If we suspect there is an error in kernel die reporting, then Error in log is fine, too.
I'm already good taking this in as is.
LGTM
@askervin I'm not sure if this is due to the kernel or if it is really what the topology-enumeration using the CPUID leaf 0x1f really returns on those platforms. We should probably retest this once the 7.0 kernel is out, because that has some topology enumeration fixes. But TBH, I suspect that this is what leaf 0x1f really returns on that particular HW... |
Do not panic for weird die setups. If an LLC group spans multiple dies, sort cache groups only by socket, NUMA node and lowest CPU. Signed-off-by: Krisztian Litkey <krisztian.litkey@intel.com>
Omit die pools if dies are the same as clusters. Signed-off-by: Krisztian Litkey <krisztian.litkey@intel.com>
klihub
force-pushed
the
fixes/weird-die-setup
branch
from
a4522ad to
2e933df
Compare
That's right, sounds more of an ACPI issue. As this PR looks safe when die numbering works fine, and it writes nice warnings in the log in other cases, I'll merge it now and let's test with linux 7.0. (I don't want to keep this PR pending and our topology detection doing weird things on weirdly reported dies...) |
2 participants
This PR fixes the most immediate problems for weird die setups in the topology reported by the kernel. In particular the commits