#![forbid(unsafe_code)] use std::sync::{Mutex, OnceLock}; use std::time::{Duration, Instant}; static CAPTURE_ORIGIN: OnceLock = OnceLock::new(); const DEFAULT_SOURCE_LAG_CAP_MS: u64 = 250; const DEFAULT_SOURCE_LEAD_CAP_MS: u64 = 80; fn origin() -> Instant { *CAPTURE_ORIGIN.get_or_init(Instant::now) } /// Return the shared live-capture timestamp for upstream camera/mic packets. /// /// Inputs: none. /// Outputs: microseconds elapsed since the relay child first stamped live media. /// Why: camera and microphone capture pipelines run independently, so they need /// one explicit common origin before the server can keep them on the same live /// call timeline. #[must_use] pub fn capture_pts_us() -> u64 { origin().elapsed().as_micros().min(u64::MAX as u128) as u64 } /// Measure how old one shared capture timestamp is right now. /// /// Inputs: a packet timestamp previously produced by `capture_pts_us`. /// Outputs: the elapsed age as a `Duration`. /// Why: upstream freshness telemetry should use the same shared live clock as /// packet timestamps so queue-age calculations stay honest. #[must_use] #[allow(dead_code)] pub fn packet_age(pts_us: u64) -> Duration { Duration::from_micros(capture_pts_us().saturating_sub(pts_us)) } /// Decide whether extra upstream timing instrumentation should be emitted. /// /// Inputs: none. /// Outputs: `true` when detailed capture/rebase timing logs are enabled. /// Why: A/V sync work needs bursts of deep timing visibility without leaving /// noisy logs on during normal live operation. #[must_use] pub fn upstream_timing_trace_enabled() -> bool { std::env::var("LESAVKA_UPSTREAM_TIMING_TRACE") .ok() .map(|value| { let trimmed = value.trim(); !(trimmed.eq_ignore_ascii_case("0") || trimmed.eq_ignore_ascii_case("false") || trimmed.eq_ignore_ascii_case("no") || trimmed.eq_ignore_ascii_case("off")) }) .unwrap_or(false) } /// Cap how far source-derived packet timestamps may trail the live capture clock. /// /// Inputs: none. /// Outputs: the maximum tolerated lag between a rebased source PTS and the /// current capture clock. /// Why: encoded appsink buffers can emerge well after the raw capture moment, /// and trusting those delayed buffer PTS values without a guard can make the /// server believe fresh audio/video packets are already hopelessly stale. #[must_use] pub fn upstream_source_lag_cap() -> Duration { std::env::var("LESAVKA_UPSTREAM_SOURCE_LAG_CAP_MS") .ok() .and_then(|raw| raw.trim().parse::().ok()) .filter(|value| *value > 0) .map(Duration::from_millis) .unwrap_or_else(|| Duration::from_millis(DEFAULT_SOURCE_LAG_CAP_MS)) } /// Cap how far source-derived packet timestamps may lead the live capture clock. /// /// Inputs: none. /// Outputs: the maximum tolerated future lead for source-based packet PTS. /// Why: live sources can flush a burst of future-stamped buffers; if those /// future timestamps escape, the server freezes media waiting for local backlog. #[must_use] pub fn upstream_source_lead_cap() -> Duration { std::env::var("LESAVKA_UPSTREAM_SOURCE_LEAD_CAP_MS") .ok() .and_then(|raw| raw.trim().parse::().ok()) .filter(|value| *value > 0) .map(Duration::from_millis) .unwrap_or_else(|| Duration::from_millis(DEFAULT_SOURCE_LEAD_CAP_MS)) } #[derive(Debug, Default)] struct SourcePtsRebaserState { source_base_us: Option, capture_base_us: Option, last_packet_pts_us: Option, } /// Rebase source-buffer timestamps onto the shared client capture clock. /// /// Inputs: optional source PTS values from one live capture pipeline. /// Outputs: packet timestamps that share the same client clock origin across /// camera and microphone while still advancing based on source timing rather /// than late appsink pull time. /// Why: camera and microphone encode paths can add different queue/encode /// delays before appsink pull, so stamping at pull time bakes skew into the /// packets before the server ever sees them. #[derive(Debug, Default)] pub struct SourcePtsRebaser { state: Mutex, } /// Snapshot of one client-side timestamp rebasing decision. #[derive(Clone, Copy, Debug, Eq, PartialEq)] pub struct RebasedSourcePts { pub packet_pts_us: u64, pub capture_now_us: u64, pub source_pts_us: Option, pub source_base_us: Option, pub capture_base_us: Option, pub used_source_pts: bool, pub lag_clamped: bool, pub lead_clamped: bool, } #[derive(Debug, Default)] struct DurationPacedSourcePtsState { next_packet_pts_us: Option, last_packet_pts_us: Option, } /// Rebase encoded packet timing by anchoring once, then pacing by duration. /// /// Inputs: optional source PTS from one encoded packet stream plus the packet's /// declared duration and a freshness lag cap. /// Outputs: packet timestamps on the shared client capture clock that advance /// by actual media duration instead of trusting potentially stretched parser /// PTS on every packet. /// Why: encoded audio parsers can emit packet PTS values that do not track /// real packet duration faithfully, which can make the server pace audio far /// too slowly or quickly even when the underlying capture stream is healthy. #[derive(Debug, Default)] pub struct DurationPacedSourcePtsRebaser { anchor_rebaser: SourcePtsRebaser, state: Mutex, } impl SourcePtsRebaser { /// Translate one source-buffer timestamp onto the shared capture clock. /// /// Inputs: the buffer PTS if available plus the minimum monotonic step. /// Outputs: a rebased packet timestamp and the values used to derive it. /// Why: source PTS should drive packet timing when available, but packets /// must still remain monotonic even if buffers repeat or arrive oddly. #[allow(dead_code)] #[must_use] pub fn rebase_or_now(&self, source_pts_us: Option, min_step_us: u64) -> RebasedSourcePts { self.rebase_with_lag_cap(source_pts_us, min_step_us, None) } /// Translate one source-buffer timestamp onto the shared capture clock /// while bounding how stale that source-derived timestamp may become. /// /// Inputs: optional source PTS, minimum monotonic step, and an optional /// maximum lag behind the current capture clock. /// Outputs: a rebased packet timestamp plus details about any lag clamp. /// Why: encoder/parser queues can batch source buffers, so a pure /// source-PTS timeline may fall far behind real packet availability and /// poison server-side freshness calculations. #[must_use] pub fn rebase_with_lag_cap( &self, source_pts_us: Option, min_step_us: u64, max_lag: Option, ) -> RebasedSourcePts { let capture_now_us = capture_pts_us(); let mut state = self .state .lock() .expect("source pts rebaser mutex poisoned"); let mut packet_pts_us = capture_now_us; let mut used_source_pts = false; let mut lag_clamped = false; let mut lead_clamped = false; if let Some(source_pts_us) = source_pts_us { let source_base_us = *state.source_base_us.get_or_insert(source_pts_us); let capture_base_us = *state.capture_base_us.get_or_insert(capture_now_us); state.capture_base_us = Some(capture_base_us); packet_pts_us = capture_base_us.saturating_add(source_pts_us.saturating_sub(source_base_us)); used_source_pts = true; } if used_source_pts && let Some(max_lag) = max_lag { let lag_floor_us = capture_now_us.saturating_sub(max_lag.as_micros().min(u64::MAX as u128) as u64); if packet_pts_us < lag_floor_us { packet_pts_us = lag_floor_us; lag_clamped = true; } let lead_ceiling_us = capture_now_us.saturating_add( upstream_source_lead_cap().as_micros().min(u64::MAX as u128) as u64, ); if packet_pts_us > lead_ceiling_us { packet_pts_us = lead_ceiling_us; lead_clamped = true; } } if let Some(last_packet_pts_us) = state.last_packet_pts_us && packet_pts_us <= last_packet_pts_us { packet_pts_us = last_packet_pts_us.saturating_add(min_step_us.max(1)); } state.last_packet_pts_us = Some(packet_pts_us); RebasedSourcePts { packet_pts_us, capture_now_us, source_pts_us, source_base_us: state.source_base_us, capture_base_us: state.capture_base_us, used_source_pts, lag_clamped, lead_clamped, } } } impl DurationPacedSourcePtsRebaser { /// Rebase one encoded packet onto the shared capture clock. /// /// Inputs: optional packet PTS, the packet media duration in microseconds, /// and a freshness lag cap behind the live capture clock. /// Outputs: a rebased packet timestamp plus the values used to derive it. /// Why: once the first encoded packet is anchored, the safest pacing signal /// for compressed audio is its actual packet duration, with a live lag /// clamp to keep delayed batches from resurrecting stale timing. #[must_use] pub fn rebase_with_packet_duration( &self, source_pts_us: Option, packet_duration_us: u64, max_lag: Duration, ) -> RebasedSourcePts { let step_us = packet_duration_us.max(1); let mut rebased = self.anchor_rebaser .rebase_with_lag_cap(source_pts_us, step_us, Some(max_lag)); let lag_floor_us = rebased .capture_now_us .saturating_sub(max_lag.as_micros().min(u64::MAX as u128) as u64); let mut state = self .state .lock() .expect("duration paced source pts rebaser mutex poisoned"); let mut packet_pts_us = state.next_packet_pts_us.unwrap_or(rebased.packet_pts_us); if packet_pts_us < lag_floor_us { packet_pts_us = lag_floor_us; rebased.lag_clamped = true; } let lead_ceiling_us = rebased .capture_now_us .saturating_add(upstream_source_lead_cap().as_micros().min(u64::MAX as u128) as u64); if packet_pts_us > lead_ceiling_us { packet_pts_us = lead_ceiling_us; rebased.lead_clamped = true; } if let Some(last_packet_pts_us) = state.last_packet_pts_us && packet_pts_us <= last_packet_pts_us { packet_pts_us = last_packet_pts_us.saturating_add(1); } state.last_packet_pts_us = Some(packet_pts_us); state.next_packet_pts_us = Some(packet_pts_us.saturating_add(step_us)); rebased.packet_pts_us = packet_pts_us; rebased } } #[cfg(test)] #[path = "live_capture_clock/tests.rs"] mod tests;