irtt-rs/crates/irtt-client/AGENTS.md

irtt-client guidance

Repository-wide guidance in the root AGENTS.md also applies here.

Linux kernel TX timestamp capture

On Linux, with the ancillary feature, the client best-effort upgrades its socket from RX-only SO_TIMESTAMPING to RX+TX immediately after a successful Open, before the first Echo probe can be sent:

socket created
    -> RX timestamping configured (RX_SOFTWARE | SOFTWARE)
    -> Open attempts/retries (untimestamped)
    -> successful Open
    -> best-effort enable TX timestamping (+ TX_SOFTWARE | OPT_ID | OPT_TSONLY)
    -> first Echo probe

TX_SOFTWARE asks the kernel for a software send timestamp; OPT_ID asks it to tag each notification with an automatically incrementing per-socket ID; OPT_TSONLY means the notification never needs to carry the original payload, so a zero-length receive buffer is enough. The upgrade is entirely best-effort: if it fails, the client keeps operating on the original RX-only configuration (try_enable_tx_timestamping restores it explicitly) and timing falls back to userspace sent_at as it always has. TX capability failure is never a public error and is never logged from the library.

Correlation invariant: kernel ID == wire sequence

The client uses the kernel's automatic OPT_ID directly as the probe's wire sequence number — no separate correlation map. This is safe because:

• A failed/nonblocking send never consumes an ID. Linux increments the per-socket ID counter (sk->sk_tskey) speculatively while building a datagram and rolls the increment back (atomic_dec) on every error path out of __ip_append_data/__ip6_append_data (net/ipv4/ip_output.c, net/ipv6/ip6_output.c), guarded by a local hold_tskey flag. A WouldBlock try_send therefore can never leave a gap in, or otherwise advance, the ID the kernel will assign to the next successful submission. (include/net/sock.h's _sock_tx_timestamp performs the same increment for the non-corked path.) Client-side, the existing send transaction only calls commit_probe_sent — which is the only thing that advances next_wire_seq — after a successful socket send; a WouldBlock result never reaches it.
• The two counters start together and advance together. wire_seq starts at 0 on every successful Open (SessionMachine::commit_open builds a fresh ActiveSession with next_wire_seq: 0) and advances by exactly 1 only via commit_probe_sent, which only runs after a confirmed successful send — the same condition under which the kernel's counter advances by 1. TX timestamping is enabled only after that same successful Open and before any probe is sent, so both counters' zero points line up exactly.
• One socket, one session, no reopen. SessionMachine's state machine only reaches Open from Connected, and commit_open/commit_local_close never route back to Connected. A Client/AsyncClient cannot reopen a second session on the same socket, so there is no second epoch of wire sequence numbers to collide with a stale kernel ID.
• The one extra send (Close) is harmless. Close sends one datagram after TX timestamping is already enabled, generating one TX timestamp record with no corresponding wire_seq (Close does not advance next_wire_seq). Because the session is terminal immediately after, nothing will ever look up that ID; record_kernel_tx_timestamp's "unknown ID -> ignore" rule discards it safely.

MSG_ERRQUEUE drains

receive::linux::error_queue decodes one MSG_ERRQUEUE record at a time (try_recv_error_queue_record) and classifies it: a genuine send-timestamp completion (TxTimestamp), a timestamp-origin record that is not a usable SND completion (MalformedOrUnsupportedTimestamp — never treated as fatal, since this client only ever requests TX_SOFTWARE/SND, so anything else on that origin is unexpected metadata, not a network error), a genuine non-timestamp socket/network error (SocketError), or nothing (Ignored). Origin (SO_EE_ORIGIN_TIMESTAMPING or not) is the only thing that decides between the last two categories.

receive::linux::drain_tx_timestamps reads up to MAX_TX_TIMESTAMP_RECORDS_PER_DRAIN (32) records per call: nonblocking, one record at a time, no heap allocation, stopping early once the queue is empty. It is a starvation guard, not a queue-capacity promise — one TX timestamp is generated per successful probe, and drains happen at several natural choke points, so a modest constant comfortably covers realistic backlog. Client/AsyncClient call it after a successful probe send, before processing an inbound reply, and before evaluating timeouts; it never waits for a timestamp; an absent timestamp after all of those opportunities simply stays absent. A genuine SocketError record stops the drain and is surfaced through the ordinary ClientError::Socket path.

The Tokio adapter uses the same raw nonblocking recvmsg(MSG_ERRQUEUE | MSG_DONTWAIT) rather than try_io/poll_recv_ready: error-queue readiness is not ordinary read readiness, and driving it through the Tokio reactor risks clearing readiness state that the normal Echo receive path depends on. There is no background task or second AsyncFd around the Tokio-owned socket.

Retained state

PendingProbe carries an optional kernel_tx_timestamp: Option<SystemTime> alongside the userspace sent_at. It survives pending -> timed-out retention (SessionMachine::record_kernel_tx_timestamp updates whichever of PendingMap/TimedOutMap still holds the probe) and is otherwise bounded by the same limits those maps already enforce — no new unbounded state. First valid observation wins on duplicates.

Pre-send anchor vs. post-send sent_at (F-04)

The probe send transaction splits into two timestamps around the socket send, not one:

• A private pre-send anchor, sampled immediately before the fallible commit preflight (SessionMachine::finalize_probe_commit) that must happen before the socket send: timeout deadline checked_add arithmetic (so DurationOverflow still guarantees nothing was transmitted) and the kernel TX lower-plausibility bound. Its monotonic half becomes PendingProbe::timeout_at's basis; its wall half becomes PendingProbe::tx_not_before_wall. It is never exposed publicly.
• The public sent_at, a paired wall/monotonic ClientTimestamp captured immediately after the socket send returns successfully (SessionMachine::commit_probe_sent, which is infallible — all fallible work already happened against the pre-send anchor). This is the RTT endpoint, the userspace upstream-OWD fallback endpoint, timer_error's basis, and the value surfaced on EchoSent/EchoReply/EchoLoss/ LateReply.

So timeout_at is deliberately not derived from sent_at.mono — it remains anchored to before the send, on purpose, so the existing pre-send timeout/overflow/no-send contract is unchanged. Only the measurement sent_at moved to after the send. A WouldBlock try_send (Tokio) never samples a sent_at, never calls finalize_probe_commit's matching commit_probe_sent, and never advances the machine or schedule; each retry gets its own fresh pre-send anchor, and only a successful attempt's anchor and sent_at are ever retained.

The kernel TX timestamp is eligible for exactly one measurement: the upstream (client_to_server) one-way delay send endpoint. compute_one_way's private preferred_send_wall helper prefers it over the userspace sent_at.wall sample when

tx_not_before_wall <= kernel_tx_timestamp <= reply_received_at.wall

(inclusive, compared only against the client's own wall clock — never against the remote server's wall timestamp). The lower bound is the pre-send tx_not_before_wall, not the post-send sent_at.wall: a legitimate TX_SOFTWARE timestamp can be generated while the send syscall is still in flight, so it can legitimately precede the post-send sent_at sample without being implausible. Otherwise, including when no kernel timestamp was observed, the upstream endpoint falls back to sent_at.wall (the post-send sample); a rejected observation is never erased from PendingProbe, only skipped for that measurement. There is no maximum-lag bound: unlike the kernel receive side, legitimate client send-to-reply time can exceed a second.

sent_at remains the sole input to RTT and the downstream (server_to_client) one-way delay, exactly as before this capability existed; TX_SOFTWARE is still a software kernel timestamp near the driver handoff, not a NIC hardware transmit time. IPDV is likewise unaffected — it is computed from the existing userspace timestamp fields only. A normal reply and a measurable late reply (one whose PendingProbe is still retained in TimedOutMap) apply the identical selection policy; an untracked late reply has no retained probe and stays unmeasurable, and a kernel timestamp that arrives after a reply has already been processed is simply never observed by that reply.