arch: x86_64: handle npot CPU topology

This PR addresses a bug in which the cpu topology of a guest
with non power-of-two number of cores is incorrect. For example,
in some contexts, a virtual machine with 2-sockets and 12-cores
will incorrectly believe that 16 cores are on socket 1 and 8
cores are on socket 2. In other cases, common topology enumeration
software such as hwloc will crash.

The root of the problem was the way that cloud-hypervisor generates
apic_id. On x86_64, the (x2) apic_id embeds information about cpu
topology. The cpuid instruction is primarily used to discover the
number of sockets, dies, cores, threads, etc. Using this information,
the (x2) apic_id is masked to determine which {core, die, socket} the
cpu is on. When the cpu topology is not a power of two
(e.g. a 12-core machine), this requires non-contiguous (x2) apic_id.

Signed-off-by: Thomas Barrett <tbarrett@crusoeenergy.com>
This commit is contained in:
Thomas Barrett
2023-12-22 03:38:13 +00:00
committed by Rob Bradford
parent a2ceb00fbc
commit 5c0b66529a
6 changed files with 164 additions and 34 deletions

View File

@@ -1035,6 +1035,40 @@ fn test_cpu_topology(threads_per_core: u8, cores_per_package: u8, packages: u8,
.unwrap_or(0),
packages
);
#[cfg(target_arch = "x86_64")]
{
let mut cpu_id = 0;
for package_id in 0..packages {
for core_id in 0..cores_per_package {
for _ in 0..threads_per_core {
assert_eq!(
guest
.ssh_command(&format!("cat /sys/devices/system/cpu/cpu{cpu_id}/topology/physical_package_id"))
.unwrap()
.trim()
.parse::<u8>()
.unwrap_or(0),
package_id
);
assert_eq!(
guest
.ssh_command(&format!(
"cat /sys/devices/system/cpu/cpu{cpu_id}/topology/core_id"
))
.unwrap()
.trim()
.parse::<u8>()
.unwrap_or(0),
core_id
);
cpu_id += 1;
}
}
}
}
});
let _ = child.kill();