Wrapping io.Copy silently disables Go zero-copy transfers

Category: Engineering post
Status: discussion
Sources: segflow.github.io
Summary: Assel Meher (Grafana Labs) walks through how Go's standard library dispatches io.Copy to sendfile(2) when copying an *os.File to a *net.TCPConn and to splice(2) for socket-to-socket forwarding, keeping data in kernel buffers. A 512 MiB benchmark shows the direct path using about 2,958 sendfile syscalls, while wrapping the source in a plain reader falls back to userspace copies with roughly 131,000 read and write calls and 3.4 times the CPU. Only *io.LimitedReader is recognized by the runtime, so middleware such as a logging reader disables the fast path.
Why it matters: Proxies and file servers written in Go lose kernel zero-copy transparently when a reader wraps the source, and the fix is to expose optional interfaces or avoid wrapping.