fix(client): use kernel receive time for downstream OWD
This commit is contained in:
4 files changed
+555
-10
No files matched your search
@@ -348,7 +348,11 @@ pub struct OneWayDelaySample {
|
||||
/// Server-to-client delay.
|
||||
///
|
||||
/// This is computed from the server send wall time and the client receive
|
||||
/// wall time when both are available.
|
||||
/// wall time when both are available. The client receive endpoint is the
|
||||
/// kernel receive timestamp reported with the datagram when one is
|
||||
/// available and plausible for this receive, and the userspace receive wall
|
||||
/// time otherwise. RTT is unaffected and remains a userspace monotonic
|
||||
/// measurement.
|
||||
pub server_to_client: Option<SignedDuration>,
|
||||
}
|
||||
|
||||
@@ -377,5 +381,11 @@ pub struct PacketMeta {
|
||||
pub traffic_class: Option<u8>,
|
||||
pub dscp: Option<u8>,
|
||||
pub ecn: Option<u8>,
|
||||
/// Kernel software receive timestamp reported with the datagram, where the
|
||||
/// platform provides one.
|
||||
///
|
||||
/// This is the raw observed value, reported whether or not it was accepted
|
||||
/// as the downstream one-way delay endpoint. It is a kernel software
|
||||
/// timestamp, not a NIC hardware ingress time.
|
||||
pub kernel_rx_timestamp: Option<std::time::SystemTime>,
|
||||
}
|
||||
Reference in new issue
Block a user