Review of PR #918: "perf(sfu): send media from all goroutines at once on the shared UDP socket"

From mnema-talk v3.3.13, rendered on this site.

Independent adversarial review of the first commit (a direct sendto(2) path for the SFU's mux sockets and the built-in TURN relay's socket, and counters for Pion's silent buffer drops). Verdict on that commit: MERGE AFTER FIXES. No correctness or security defect was found. The findings were gaps in tests and docs, and the second commit addresses them; see the resolution column.

Findings

#SeverityFindingResolution
F1MediumThe EAGAIN fallback (internal/fastudp/fastudp_linux.go) was untested. A mutation that returned the error instead of falling back passed every test.sendto is a package variable that tests replace. TestFullSendBufferFallsBackToTheRuntime asserts that a would-block send still goes out through the runtime with the full length. TestDirectSendErrorsAreReturned and TestSendtoReportsWouldBlockAsNotDone cover the other outcomes.
F2MediumThe buffer counter test called Write, but srtp writes through WriteWithAttributes. Deleting the override would have silenced the counter without failing the test.The test now writes through WriteWithAttributes, the method Pion calls, and fails if the buffer loses it.
F3LowThe mux counter depends on Pion's exact log text, and nothing checked it.TestPionLogLinesStillMatch reads the Pion module source (go list -m) and asserts that both matched log lines still exist (the mux drop and the unhandled-RTCP line). It skips only when the source is unavailable.
F4LowThe mux socket test listened on udp4, while Pion listens on udp with an interface address. A wildcard bind would yield a dual-stack socket that is not wrapped.The test now opens the socket exactly the way Pion's mux does.
F5Lowdoc/benchmark.md had no entry for the result, and reason="mux" also counts RTCP and DTLS.The new section "The mixed Talk holds: the socket's write lock" has the before and after runs. doc/operations.md states that mux also counts RTCP and DTLS packets of any member.
F6InfoWrite deadlines are not applied on the direct path, a writer that falls back may wait behind direct writers, and direct-path errors carry no addresses.Accepted. Pion's mux uses deadlines only to unblock writes stuck for 5 s, and a direct send never blocks. The fallback fairness note is in the package doc. No caller reads the error's address fields.

What was verified

  • Descriptor lifetime:
    • Every direct send reads closed under a read lock, and Close sets it under the write lock before closing the socket.
    • Every close path of these sockets goes through Conn.Close: Pion's UDPMuxDefault.Close and its constructor's error path, turn.Server.Close, and the relay's closers.
    • Nothing closes the embedded conn directly.
    • A send therefore never reaches a released or reused descriptor number.
  • System call:
    • EINTR is retried, and UDP never sends partially.
    • A zero-length payload takes the runtime's path.
    • The port is written in network byte order; a mutation swapping it fails two tests.
    • IPv4-mapped destinations are unmapped. IPv6 and dual-stack sockets are not wrapped.
    • The unsafe.Pointer conversions sit in the Syscall6 argument list.
    • -gcflags=-m shows that neither the payload nor the address escapes: 0 allocations per send.
  • Pion integration:
    • asAddrPortReaderWriter picks the wrapper's WriteToAddrPort, and buffer sizes still reach the socket.
    • No ECN reader is lost, because mux candidates never had one.
    • pion/turn uses only ReadFrom, WriteTo and LocalAddr and asserts no concrete type.
    • The RTCP buffer is still a plain attrBuffer, so the existing type assertion holds.
  • Security:
    • A remote peer cannot influence the descriptor.
    • Destinations still come only from Pion's ICE and STUN addresses and from the TURN 5-tuple.
    • The packet count per input is unchanged, so there is no amplification and no new panic path.
  • Runs:
    • GOMAXPROCS=4 go test -race -count=3 ./internal/fastudp ./internal/turnrelay and the targeted ./internal/sfu tests pass on macOS.
    • The Linux test binaries (arm64) pass in an Alpine container.
    • go test -race passes in a Linux Go container, exercising the direct path.
    • BenchmarkParallelSend on Linux arm64 with 4 goroutines: 724 ns/op through the runtime's lock against 314 ns/op direct.
  • Repo limits:
    • Files: transport.go 377 lines, ingress.go 346, fastudp.go 117.
    • Test files, functions and cognitive complexity are within limits.

Something wrong on this page? Tell us.