v3.1.0

Added

• Memory working set for containers. ContainerLimits::$memoryWorkingSetBytes
  reports cgroup memory usage minus inactive file cache, the way cAdvisor and
  Kubernetes define the working set: memory.current - inactive_file on
  cgroup v2, memory.usage_in_bytes - total_inactive_file on v1, floored at 0.
  Raw usage counts reclaimable page cache, so a container that writes files
  reads as close to its limit even though the kernel would hand that memory
  back. memoryWorkingSetUtilizationPercentage() and
  availableMemoryBytesFromWorkingSet() go with it. All three are null when
  memory.stat cannot be read. memoryUsageBytes and its helpers are
  unchanged, and so is limits().

Fixed

• limits() in a cgroup without a memory limit pairs host total with host
  usage. It used to pair the host's total with the usage of only the
  processes inside the cgroup — every systemd service, and every container
  whose limit sits on a parent cgroup (a Kubernetes pod or ECS task limit).
  availableMemoryBytes() and memoryUtilization() described neither scope,
  and a capacity planner built on them over-packed the machine. Cgroups with a
  limit are unchanged. (#10)
• cgroup v1 oomKillCount counts OOM kills. It read the under_oom flag,
  which is 1 only while the cgroup is stalled with the OOM killer disabled, so
  it stayed 0 however many processes the kernel killed. It now reads the
  oom_kill counter in memory.oom_control (Linux 4.13+), and is null on
  older kernels. cgroup v2 is unchanged. (#11)

Thanks to @Orrison for #10 and #11.