Summary

• When running in containers (Kubernetes, Docker), swoole_cpu_num() reports the host node's CPU count instead of the container's cgroup CPU limit
• This causes --workers=auto and --task-workers=auto to spawn far more processes than the container can handle, leading to OOMKills
• The fix reads the container's cgroup CPU quota (v2 and v1) before falling back to swoole_cpu_num()

Example

A pod with 1 CPU limit on a 96-core node:

• Before: 96 workers + 96 task workers (OOMKilled)
• After: 1 worker + 1 task worker

Test plan

• Existing SwooleExtensionTest::test_cpu_count still passes
• Added tests for container CPU detection and fallback behavior
• Verified cgroup v2 values on a real Kubernetes pod (cpu.max: 100000 100000 → 1 CPU)
• Full test suite passes (174 tests)