Limit "resources.cores" does not work after nomad client restart

[TheSpbra1n]

Hello.
We test resources.cores for nomad jobs and found a bug: after nomad client restart, cores limit does not work.

Nomad version

Nomad v1.4.2 (039d70e)

Operating system and Environment details

Codename: focal

Reproduction steps

1. Run job:

# job name
job "cpu-loader" {
  region      = "global"
  datacenters = ["dc1"]

  type = "service"

  meta {
      ...
  }

  group "cpu-loader-group" {
    count = 1

    # task
    task "cpu-loader-task" {
      kill_signal = "SIGTERM"
      driver = "docker"

      config {
        image        = "docker-image"
      }

      env {
          ...
      }

      resources {
       memory = 1000
       cores  = 4 
      }
    }
  }
}

2. Stop job
3. Restart nomad client
4. Run job again

(for loader we used: https://github.com/vikyd/go-cpu-load )

Expected Result

Job used only 4 cores

Actual Result

Job used more than 4 core

[RobloxMatt]

We've seen something that sounds like this problem as well. Do you see something like this in your logs?

[WARN] client.cpuset.v1: failed to ensure reserved cpuset.cpus interface exists; disable cpuset management: error="mkdir /sys/fs/cgroup/cpuset/nomad/reserved: file exists"

[shoenig]

Thanks for the report @TheSpbra1n and for the logs @RobloxMatt, it seems this got broken during a refactoring in 0a3d57f where we (I) missed the edge case where the cgroup structure is already set on Client initialization. Should only affect machines using cgroups v1.