mirror of
https://github.com/slackhq/nebula.git
synced 2026-08-15 04:56:59 +02:00
JackDoan
10e9514e44
Each listenIn goroutine locks its OS thread and pins it to one CPU (sched_setaffinity), so every UDP send from that goroutine leaves through the same XPS-selected NIC TX ring instead of being sprayed across rings and reordered. On by default via tun.pin_threads; queue i pins to the i-th entry of the process's allowed CPU set (respecting cpuset/taskset masks, whose IDs are often not 0..NumCPU-1), or to an explicit tun.cpu_affinity list, validated against that same allowed set. Linux only; pinning is a no-op elsewhere. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
19 lines
610 B
Go
19 lines
610 B
Go
//go:build !linux || android || e2e_testing
|
|
|
|
package util
|
|
|
|
// PinThreadToCPU is a no-op outside Linux: only Linux exposes a stable
|
|
// per-thread CPU affinity API and only Linux has XPS-driven TX ring
|
|
// selection in the first place. On every other platform there's nothing
|
|
// to fix here.
|
|
func PinThreadToCPU(_ int) error {
|
|
return nil
|
|
}
|
|
|
|
// AllowedCPUs has no meaningful answer off Linux (no sched_getaffinity), so it
|
|
// reports "unknown" by returning a nil slice and nil error. Callers treat an
|
|
// empty result as "fall back to the default CPU choice".
|
|
func AllowedCPUs() ([]int, error) {
|
|
return nil, nil
|
|
}
|