@amritk/nish-x86_64-linux 0.16.0 → 0.17.0

package/std/net/quic-recovery.ts

@@ -0,0 +1,849 @@
+/**
+ * `nish/net/quic-recovery` — QUIC loss detection and congestion control
+ * (RFC 9002): the packets each packet number space has in flight, the
+ * round-trip estimate, loss by packet and by time threshold, the probe
+ * timeout with its backoff, NewReno, and a pacer. Sans-IO: it never sees a
+ * byte of a packet, only packet numbers, sizes and the caller's clock.
+ *
+ *     import { QuicRecovery, QUIC_RECOVERY_TIMEOUT_PTO } from "nish/net/quic-recovery";
+ *
+ *     const recovery = new QuicRecovery();
+ *     const slot: i32 = recovery.onPacketSent(space, pn, size, now);     // every ack-eliciting packet
+ *     recovery.onAck(space, frame.ackRanges, frame.ackRangeCount, delay, now);
+ *     … recovery.spaces[space].acked / .lost name the slots it settled …
+ *     if (now >= recovery.deadline(blocked)) { recovery.onTimeout(now, blocked); }
+ *
+ * **What it tracks.** Each space keeps its ack-eliciting packets in a fixed
+ * ring (`QuicSentPackets`), in the order they were sent, so the memory is
+ * fixed when the connection is made and nothing is allocated per packet. A
+ * packet that elicits no acknowledgement (one that carries only ACK, PADDING
+ * or CONNECTION_CLOSE) is not recorded: it is never in flight, and neither
+ * an RTT sample nor a loss is ever taken from it (RFC 9002 §2, §5.1). The
+ * caller keeps what each packet carried in its own arrays, indexed by the
+ * slot `onPacketSent` answers, and reads `acked` and `lost` after each
+ * `onAck` and `onTimeout` to learn which slots were settled. A slot named
+ * there keeps its packet number, time and size until the next call.
+ *
+ * **Time and units.** Every time and every duration is the caller's
+ * monotonic clock in milliseconds, which is also the timer granularity
+ * (`QUIC_RECOVERY_GRANULARITY`, §6.1.2), and the arithmetic is Appendix A's
+ * and Appendix B's in integers: `smoothed_rtt = (7 × smoothed_rtt +
+ * adjusted_rtt) / 8`, `rttvar = (3 × rttvar + |smoothed_rtt −
+ * adjusted_rtt|) / 4`, each division truncating. Sizes are bytes.
+ *
+ * **What the server leaves out.** It is written for the server side of a
+ * connection: there is no client anti-deadlock probe (§6.2.2.1), since a
+ * server never has to send to unblock its peer, and nothing reads ECN
+ * (§7.1, B.7). A congestion window that the sender does not use is not
+ * grown (§7.8): an acknowledgement grows it only if, when it arrived, the
+ * bytes in flight were at least half the window.
+ *
+ * **What a peer cannot do.** Every input is checked: the ranges of an
+ * acknowledgement are read only up to the count given and the array's
+ * length, a packet acknowledged twice counts once, and
+ * the backoff stops doubling at `QUIC_RECOVERY_MAX_BACKOFF`, so no duration
+ * overflows. A peer that never acknowledges anything costs a probe each
+ * timeout, at a rate that halves each time, until the idle timeout ends the
+ * connection.
+ *
+ * Written from RFC 9002 §5 to §7 and its Appendices A and B, in this
+ * module's own structure; nothing here is ported from another
+ * implementation. Private names carry the `quicRecovery` prefix
+ * (`docs/wp26-stdlib.md` §3e).
+ */
+import { QUIC_MAX_VARINT } from "nish/net/quic-packet"
+
+/** The Initial space's index, which is `TLS_LEVEL_INITIAL`. */
+export const QUIC_RECOVERY_INITIAL: i32 = 0
+/** The Handshake space's index. */
+export const QUIC_RECOVERY_HANDSHAKE: i32 = 1
+/** The Application Data space's index. */
+export const QUIC_RECOVERY_APPLICATION: i32 = 2
+/**
+ * How many packets the Initial and the Handshake space each keep in flight.
+ * A server's handshake flight is a few datagrams, and the anti-amplification
+ * limit holds it to three times what the client sent until its address is
+ * validated, so 16 leaves room for several probes on top.
+ */
+export const QUIC_RECOVERY_HANDSHAKE_CAPACITY: i32 = 16
+/**
+ * How many packets the Application Data space keeps in flight: 128
+ * full-sized datagrams is 150 KiB in flight, a 12 Mbit/s path at a 100 ms
+ * round trip, against what the record costs every connection from the
+ * start. When the ring is full the space sends nothing that would elicit an
+ * acknowledgement until one arrives, as if the congestion window were full.
+ */
+export const QUIC_RECOVERY_APPLICATION_CAPACITY: i32 = 128
+/** kPacketThreshold (§6.1.1): a packet three numbers below one acknowledged is lost. */
+export const QUIC_RECOVERY_PACKET_THRESHOLD: i64 = 3
+/** kGranularity (§6.1.2), in milliseconds: the timer resolution, and the least any loss delay or variance term is. */
+export const QUIC_RECOVERY_GRANULARITY: i64 = 1
+/** kInitialRtt (§6.2.2), in milliseconds: the round trip assumed before the first sample. */
+export const QUIC_RECOVERY_INITIAL_RTT: i64 = 333
+/** max_datagram_size (B.1): the 1200 bytes every datagram this server sends is at most. */
+export const QUIC_RECOVERY_DATAGRAM_SIZE: i64 = 1200
+/** kInitialWindow (§7.2): min(10 × 1200, max(14720, 2 × 1200)), so 12000 bytes. */
+export const QUIC_RECOVERY_INITIAL_WINDOW: i64 = 12000
+/** kMinimumWindow (§7.2): two datagrams, 2400 bytes. */
+export const QUIC_RECOVERY_MINIMUM_WINDOW: i64 = 2400
+/** kPersistentCongestionThreshold (§7.6.1): lost packets spanning three probe timeouts collapse the window. */
+export const QUIC_RECOVERY_PERSISTENT_THRESHOLD: i64 = 3
+/** The default max_ack_delay (RFC 9000 §18.2), in milliseconds, until the peer's transport parameters say otherwise. */
+export const QUIC_RECOVERY_DEFAULT_MAX_ACK_DELAY: i64 = 25
+/**
+ * The most times the probe timeout doubles (§6.2.1). Sixteen doublings of
+ * the longest probe timeout a peer can cause still fit an `i64`, and by then
+ * the idle timeout has ended the connection.
+ */
+export const QUIC_RECOVERY_MAX_BACKOFF: i32 = 16
+
+/** `onTimeout` ran nothing: no timer was due. */
+export const QUIC_RECOVERY_TIMEOUT_NONE: i32 = 0
+/** `onTimeout` declared packets lost by the time threshold; `lost` in `timeoutSpace` names them. */
+export const QUIC_RECOVERY_TIMEOUT_LOSS: i32 = 1
+/** `onTimeout` was a probe timeout: the caller sends a probe in `timeoutSpace` (§6.2.4). */
+export const QUIC_RECOVERY_TIMEOUT_PTO: i32 = 2
+
+/** A slot's states: free, a packet in flight, one acknowledged, one declared lost, and one this call declared lost. */
+const QUIC_RECOVERY_FREE: i32 = 0
+const QUIC_RECOVERY_SENT: i32 = 1
+const QUIC_RECOVERY_ACKED: i32 = 2
+const QUIC_RECOVERY_LOST: i32 = 3
+const QUIC_RECOVERY_LOST_NOW: i32 = 4
+
+/** The larger of two times or durations. */
+const quicRecoveryMax = (a: i64, b: i64): i64 => (a > b ? a : b)
+
+/**
+ * The ack-eliciting packets one space sent, in a ring of `capacity` slots
+ * in the order they went out: slot `slot(k)` is the `k`-th from the oldest
+ * still kept, for `k` below `count`. A packet stays in its slot after it is
+ * acknowledged or lost until every older one is settled too, which is what
+ * lets persistent congestion see that nothing between two lost packets was
+ * acknowledged (§7.6.2).
+ */
+export class QuicSentPackets {
+  pn: i64[]
+  timeSent: i64[]
+  size: i32[]
+  state: u8[]
+  /** The slots the last `onAck` acknowledged, `ackedCount` of them, oldest first. */
+  acked: i32[]
+  /** The slots the last `onAck` or `onTimeout` declared lost, `lostCount` of them, oldest first. */
+  lost: i32[]
+  /** The largest packet number the peer acknowledged in this space, or -1. */
+  largestAcked: i64 = -1
+  /** The packet number of the last packet recorded, or -1: each must be larger (RFC 9000 §12.3). */
+  lastPn: i64 = -1
+  /** loss_time (A.3): when the oldest packet not yet lost by the packet threshold becomes lost by time, or -1. */
+  lossTime: i64 = -1
+  /** time_of_last_ack_eliciting_packet (A.3), or -1 before the first. */
+  lastAckEliciting: i64 = -1
+  /** The bytes of this space's packets in flight. */
+  bytesInFlight: i64 = 0
+  capacity: i32 = 0
+  head: i32 = 0
+  count: i32 = 0
+  /** How many recorded packets are still in flight: neither acknowledged nor lost. */
+  inFlight: i32 = 0
+  ackedCount: i32 = 0
+  lostCount: i32 = 0
+
+  constructor(capacity: i32) {
+    this.capacity = capacity
+    this.pn = new Array<i64>(capacity)
+    this.timeSent = new Array<i64>(capacity)
+    this.size = new Array<i32>(capacity)
+    this.state = new Array<u8>(capacity)
+    this.acked = new Array<i32>(capacity)
+    this.lost = new Array<i32>(capacity)
+  }
+
+  /** The ring slot of the `k`-th packet kept, counting from the oldest. */
+  slot(k: i32): i32 {
+    return (this.head + k) % this.capacity
+  }
+
+  /** The state of slot `s`, or free for a slot outside the ring. */
+  stateAt(s: i32): i32 {
+    if (s < 0 || s >= toI32(this.state.length)) {
+      return QUIC_RECOVERY_FREE
+    }
+    return toI32(this.state[s])
+  }
+
+  /** Sets slot `s`'s state. */
+  setState(s: i32, value: i32): void {
+    if (s >= 0 && s < toI32(this.state.length)) {
+      this.state[s] = toU8(value)
+    }
+  }
+
+  /** The packet number in slot `s`, or -1. */
+  pnAt(s: i32): i64 {
+    return s >= 0 && s < toI32(this.pn.length) ? this.pn[s] : -1
+  }
+
+  /** When the packet in slot `s` was sent, or -1. */
+  timeAt(s: i32): i64 {
+    return s >= 0 && s < toI32(this.timeSent.length) ? this.timeSent[s] : -1
+  }
+
+  /** The size of the packet in slot `s`, or 0. */
+  sizeAt(s: i32): i32 {
+    return s >= 0 && s < toI32(this.size.length) ? this.size[s] : 0
+  }
+
+  /** The `k`-th slot the last call acknowledged, or -1. */
+  ackedSlot(k: i32): i32 {
+    return k >= 0 && k < this.ackedCount && k < toI32(this.acked.length) ? this.acked[k] : -1
+  }
+
+  /** The `k`-th slot the last call declared lost, or -1. */
+  lostSlot(k: i32): i32 {
+    return k >= 0 && k < this.lostCount && k < toI32(this.lost.length) ? this.lost[k] : -1
+  }
+
+  /** Whether slot `s` holds a packet still in flight. */
+  inFlightAt(s: i32): boolean {
+    return this.stateAt(s) === QUIC_RECOVERY_SENT
+  }
+
+  /** Whether every slot is taken, so nothing more can be recorded until a packet is settled. */
+  full(): boolean {
+    this.compact()
+    return this.count >= this.capacity
+  }
+
+  /** Frees the oldest slots that are settled, acknowledged or lost. */
+  compact(): void {
+    while (this.count > 0 && this.stateAt(this.head) !== QUIC_RECOVERY_SENT) {
+      this.setState(this.head, QUIC_RECOVERY_FREE)
+      this.head = (this.head + 1) % this.capacity
+      this.count = this.count - 1
+    }
+    if (this.count === 0) {
+      this.head = 0
+    }
+  }
+
+  /** Forgets the results of the last call. */
+  clearResults(): void {
+    this.ackedCount = 0
+    this.lostCount = 0
+  }
+
+  /** Takes packet `s` out of flight. */
+  settle(s: i32, state: i32): void {
+    this.setState(s, state)
+    this.inFlight = this.inFlight - 1
+    this.bytesInFlight = this.bytesInFlight - toI64(this.sizeAt(s))
+  }
+
+  /** Settles packet `s` as lost and lists it in `lost`. */
+  markLost(s: i32): void {
+    this.settle(s, QUIC_RECOVERY_LOST_NOW)
+    if (this.lostCount < toI32(this.lost.length)) {
+      this.lost[this.lostCount] = s
+      this.lostCount = this.lostCount + 1
+    }
+  }
+
+  /** Drops every packet, as when the space's keys are discarded (A.11). */
+  clear(): void {
+    for (let s: i32 = 0; s < toI32(this.state.length); s += 1) {
+      this.state[s] = toU8(QUIC_RECOVERY_FREE)
+    }
+    this.head = 0
+    this.count = 0
+    this.inFlight = 0
+    this.bytesInFlight = 0
+    this.lossTime = -1
+    this.lastAckEliciting = -1
+    this.clearResults()
+  }
+
+  /** Empties the ring and forgets every number, for a connection slot reused for another peer. */
+  reset(): void {
+    this.clear()
+    this.largestAcked = -1
+    this.lastPn = -1
+  }
+}
+
+/** Whether `pn` falls in one of the first `count` `[smallest, largest]` pairs of `ranges`. */
+const quicRecoveryInRanges = (ranges: i64[], count: i32, pn: i64): boolean => {
+  for (let k: i32 = 0; k < count; k += 1) {
+    const low: i32 = k * 2
+    const high: i32 = low + 1
+    if (low < 0 || low >= toI32(ranges.length) || high < 0 || high >= toI32(ranges.length)) {
+      return false
+    }
+    if (pn >= ranges[low] && pn <= ranges[high]) {
+      return true
+    }
+  }
+  return false
+}
+
+/**
+ * One connection's loss recovery and congestion control, across its three
+ * packet number spaces. The fields are readable: the RTT estimate
+ * (`latestRtt`, `smoothedRtt`, `rttVar`, `minRtt`), the window
+ * (`congestionWindow`, `ssthresh`, `recoveryStart`), `ptoCount`, and the
+ * counts a test or a log reads (`congestionEvents`, `persistentCongestions`).
+ */
+export class QuicRecovery {
+  spaces: QuicSentPackets[]
+  /** latest_rtt (§5.1): the last sample, in milliseconds. */
+  latestRtt: i64 = 0
+  /** smoothed_rtt (§5.3): kInitialRtt until the first sample. */
+  smoothedRtt: i64 = 0
+  /** rttvar (§5.3): kInitialRtt / 2 until the first sample. */
+  rttVar: i64 = 0
+  /** min_rtt (§5.2): 0 until the first sample. */
+  minRtt: i64 = 0
+  /** When the first RTT sample was taken, or -1 before it: persistent congestion counts only packets sent since (§7.6.2). */
+  firstSampleTime: i64 = -1
+  /** The peer's max_ack_delay in milliseconds (RFC 9000 §18.2), which the caller sets once it has read it. */
+  maxAckDelay: i64 = 0
+  /** congestion_window (B.2), in bytes. */
+  congestionWindow: i64 = 0
+  /** ssthresh (B.2): unbounded (2^62 − 1) until the first congestion event. */
+  ssthresh: i64 = 0
+  /** congestion_recovery_start_time (B.2), or -1 outside recovery. */
+  recoveryStart: i64 = -1
+  /** The pacer's credit in bytes (§7.7), and the time it was last topped up, or -1. */
+  pacerBudget: i64 = 0
+  pacerTime: i64 = -1
+  /** pto_count (A.3): how many probe timeouts fired since the last acknowledgement. */
+  ptoCount: i32 = 0
+  /** The space the last `onTimeout` acted on, or -1. */
+  timeoutSpace: i32 = -1
+  /** How many congestion events reduced the window, and how many of them were persistent congestion. */
+  congestionEvents: i32 = 0
+  persistentCongestions: i32 = 0
+  /** Whether the handshake is confirmed (RFC 9001 §4.1.2), which the caller sets: the peer's max_ack_delay then bounds its ACK delays, and the Application Data space has a probe timeout. */
+  handshakeConfirmed: boolean = false
+
+  constructor() {
+    this.spaces = [
+      new QuicSentPackets(QUIC_RECOVERY_HANDSHAKE_CAPACITY),
+      new QuicSentPackets(QUIC_RECOVERY_HANDSHAKE_CAPACITY),
+      new QuicSentPackets(QUIC_RECOVERY_APPLICATION_CAPACITY),
+    ]
+    this.reset()
+  }
+
+  /**
+   * Puts everything back as a new connection has it: every space empty, the
+   * RTT estimate and the window at their initial values. A connection slot
+   * reused for another peer calls it; the rings are kept.
+   */
+  reset(): void {
+    for (const sp of this.spaces) {
+      sp.reset()
+    }
+    this.latestRtt = 0
+    this.smoothedRtt = QUIC_RECOVERY_INITIAL_RTT
+    this.rttVar = QUIC_RECOVERY_INITIAL_RTT / 2
+    this.minRtt = 0
+    this.firstSampleTime = -1
+    this.maxAckDelay = QUIC_RECOVERY_DEFAULT_MAX_ACK_DELAY
+    this.congestionWindow = QUIC_RECOVERY_INITIAL_WINDOW
+    this.ssthresh = QUIC_MAX_VARINT
+    this.recoveryStart = -1
+    this.pacerBudget = QUIC_RECOVERY_INITIAL_WINDOW
+    this.pacerTime = -1
+    this.ptoCount = 0
+    this.timeoutSpace = -1
+    this.congestionEvents = 0
+    this.persistentCongestions = 0
+    this.handshakeConfirmed = false
+  }
+
+  /** Space `space`'s packets, or `null` for an index that names no space. */
+  space(space: i32): QuicSentPackets | null {
+    if (space < 0 || space >= toI32(this.spaces.length)) {
+      return null
+    }
+    return this.spaces[space]
+  }
+
+  /** bytes_in_flight (B.2): every space's packets in flight. */
+  bytesInFlight(): i64 {
+    let total: i64 = 0
+    for (const sp of this.spaces) {
+      total = total + sp.bytesInFlight
+    }
+    return total
+  }
+
+  /**
+   * Whether the window has room for another full-sized datagram (B.4). The
+   * caller sends nothing ack-eliciting while it has not, except a probe.
+   */
+  canSend(): boolean {
+    return this.bytesInFlight() + QUIC_RECOVERY_DATAGRAM_SIZE <= this.congestionWindow
+  }
+
+  /**
+   * Records an ack-eliciting packet (A.5): number `pn` of space `space`,
+   * `size` bytes on the wire, sent at `now`. Answers its slot, under which
+   * the caller keeps what it carried, or -1, recording nothing, for a space
+   * that does not exist, a size that is not positive, a packet number that is
+   * not above the last one recorded, or a ring that is full.
+   */
+  onPacketSent(space: i32, pn: i64, size: i32, now: i64): i32 {
+    const sp: QuicSentPackets | null = this.space(space)
+    if (sp === null || size <= 0 || pn <= sp.lastPn || pn > QUIC_MAX_VARINT || sp.full()) {
+      return -1
+    }
+    const s: i32 = sp.slot(sp.count)
+    if (s < 0 || s >= toI32(sp.pn.length) || s >= toI32(sp.timeSent.length) || s >= toI32(sp.size.length)) {
+      return -1
+    }
+    sp.pn[s] = pn
+    sp.timeSent[s] = now
+    sp.size[s] = size
+    sp.setState(s, QUIC_RECOVERY_SENT)
+    sp.count = sp.count + 1
+    sp.inFlight = sp.inFlight + 1
+    sp.bytesInFlight = sp.bytesInFlight + toI64(size)
+    sp.lastPn = pn
+    sp.lastAckEliciting = now
+    return s
+  }
+
+  /**
+   * An ACK frame for space `space` (A.7): `ranges` holds `count`
+   * `[smallest, largest]` pairs, highest first, as `QuicFrame.ackRanges`
+   * does, and `ackDelay` is the peer's ACK Delay in milliseconds. The
+   * packets newly acknowledged are listed in the space's `acked`, those it
+   * shows lost in `lost`; an RTT sample is taken when the largest
+   * acknowledged is newly acknowledged (§5.1), the window grows or shrinks
+   * (B.5, B.8), and the probe backoff resets. Answers how many packets were
+   * newly acknowledged, or -1, changing nothing, when the space does not
+   * exist or there is no range. A packet number the space never sent is the
+   * caller's to refuse (RFC 9000 §13.1): only it knows the packets that
+   * elicit no acknowledgement, which are not recorded here.
+   */
+  onAck(space: i32, ranges: i64[], count: i32, ackDelay: i64, now: i64): i32 {
+    const sp: QuicSentPackets | null = this.space(space)
+    if (sp === null || count <= 0 || toI32(ranges.length) < 2) {
+      return -1
+    }
+    const largest: i64 = ranges[1]
+    sp.compact()
+    sp.clearResults()
+    if (largest > sp.largestAcked) {
+      sp.largestAcked = largest
+    }
+    // The bytes in flight before this ACK decide whether the window was in
+    // use, and so whether it may grow (§7.8).
+    const inUse: boolean = this.bytesInFlight() * 2 >= this.congestionWindow
+    let sampleAt: i32 = -1
+    for (let k: i32 = 0; k < sp.count; k += 1) {
+      const s: i32 = sp.slot(k)
+      // The ring is in packet-number order, so nothing past the largest acknowledged is.
+      if (sp.pnAt(s) > largest) {
+        break
+      }
+      if (sp.stateAt(s) === QUIC_RECOVERY_SENT && quicRecoveryInRanges(ranges, count, sp.pnAt(s))) {
+        sp.settle(s, QUIC_RECOVERY_ACKED)
+        if (sp.ackedCount < toI32(sp.acked.length)) {
+          sp.acked[sp.ackedCount] = s
+          sp.ackedCount = sp.ackedCount + 1
+        }
+        if (sp.pnAt(s) === largest) {
+          sampleAt = s
+        }
+      }
+    }
+    if (sp.ackedCount === 0) {
+      return 0
+    }
+    if (sampleAt >= 0) {
+      this.latestRtt = quicRecoveryMax(now - sp.timeAt(sampleAt), 0)
+      // §5.3: the Initial space's ACKs are not delayed by the peer, so their delay is ignored.
+      const none: i64 = 0
+      this.updateRtt(space === QUIC_RECOVERY_INITIAL ? none : quicRecoveryMax(ackDelay, none), now)
+    }
+    this.detectLoss(sp, now)
+    this.packetsLost(sp, now)
+    for (let k: i32 = 0; k < sp.ackedCount; k += 1) {
+      this.packetAcked(sp, sp.ackedSlot(k), inUse)
+    }
+    this.ptoCount = 0
+    return sp.ackedCount
+  }
+
+  /** UpdateRtt (§5.3, A.7) with a sample in `latestRtt`, `ackDelay` the peer's delay in milliseconds. */
+  updateRtt(ackDelay: i64, now: i64): void {
+    const latest: i64 = this.latestRtt
+    if (this.firstSampleTime < 0) {
+      this.minRtt = latest
+      this.smoothedRtt = latest
+      this.rttVar = latest / 2
+      this.firstSampleTime = now
+      return
+    }
+    if (latest < this.minRtt) {
+      this.minRtt = latest
+    }
+    let delay: i64 = ackDelay
+    if (this.handshakeConfirmed && delay > this.maxAckDelay) {
+      delay = this.maxAckDelay
+    }
+    // The delay is taken off only where that leaves the sample at or above min_rtt.
+    let adjusted: i64 = latest
+    if (latest >= this.minRtt + delay) {
+      adjusted = latest - delay
+    }
+    const gap: i64 = this.smoothedRtt > adjusted ? this.smoothedRtt - adjusted : adjusted - this.smoothedRtt
+    this.rttVar = (3 * this.rttVar + gap) / 4
+    this.smoothedRtt = (7 * this.smoothedRtt + adjusted) / 8
+  }
+
+  /** The loss delay (§6.1.2): 9/8 of the larger of the latest and the smoothed RTT, at least the granularity. */
+  lossDelay(): i64 {
+    const delay: i64 = (9 * quicRecoveryMax(this.latestRtt, this.smoothedRtt)) / 8
+    return quicRecoveryMax(delay, QUIC_RECOVERY_GRANULARITY)
+  }
+
+  /**
+   * DetectAndRemoveLostPackets (A.10): every packet in flight at or below
+   * the largest acknowledged is lost once it was sent a loss delay ago, or
+   * three numbers below the largest acknowledged; the first that is neither
+   * sets `lossTime`.
+   */
+  detectLoss(sp: QuicSentPackets, now: i64): void {
+    sp.lossTime = -1
+    const delay: i64 = this.lossDelay()
+    const lostSendTime: i64 = now - delay
+    for (let k: i32 = 0; k < sp.count; k += 1) {
+      const s: i32 = sp.slot(k)
+      const pn: i64 = sp.pnAt(s)
+      if (pn > sp.largestAcked) {
+        break
+      }
+      if (sp.stateAt(s) !== QUIC_RECOVERY_SENT) {
+        continue
+      }
+      if (sp.timeAt(s) <= lostSendTime || sp.largestAcked >= pn + QUIC_RECOVERY_PACKET_THRESHOLD) {
+        sp.markLost(s)
+      } else {
+        const at: i64 = sp.timeAt(s) + delay
+        if (sp.lossTime < 0 || at < sp.lossTime) {
+          sp.lossTime = at
+        }
+      }
+    }
+  }
+
+  /**
+   * OnPacketsLost (B.8) for what `detectLoss` just declared lost: one
+   * congestion event for the newest of them, and persistent congestion
+   * when they span long enough.
+   */
+  packetsLost(sp: QuicSentPackets, now: i64): void {
+    if (sp.lostCount === 0) {
+      return
+    }
+    let newest: i64 = -1
+    for (let k: i32 = 0; k < sp.lostCount; k += 1) {
+      newest = quicRecoveryMax(newest, sp.timeAt(sp.lostSlot(k)))
+    }
+    this.congestionEvent(newest, now)
+    if (this.persistentCongestion(sp)) {
+      this.congestionWindow = QUIC_RECOVERY_MINIMUM_WINDOW
+      this.recoveryStart = -1
+      this.persistentCongestions = this.persistentCongestions + 1
+    }
+    for (let k: i32 = 0; k < sp.lostCount; k += 1) {
+      sp.setState(sp.lostSlot(k), QUIC_RECOVERY_LOST)
+    }
+  }
+
+  /**
+   * Whether the packets just declared lost establish persistent congestion
+   * (§7.6.2): a run of lost ack-eliciting packets, all sent since the first
+   * RTT sample, with nothing acknowledged between the first and the last,
+   * spanning more than three probe timeouts with max_ack_delay, and holding
+   * at least one packet this call declared lost.
+   */
+  persistentCongestion(sp: QuicSentPackets): boolean {
+    if (this.firstSampleTime < 0) {
+      return false
+    }
+    const span: i64 = (this.ptoBase() + this.maxAckDelay) * QUIC_RECOVERY_PERSISTENT_THRESHOLD
+    let first: i64 = -1
+    let fresh: boolean = false
+    for (let k: i32 = 0; k < sp.count; k += 1) {
+      const s: i32 = sp.slot(k)
+      const state: i32 = sp.stateAt(s)
+      const lost: boolean = state === QUIC_RECOVERY_LOST || state === QUIC_RECOVERY_LOST_NOW
+      if (state === QUIC_RECOVERY_ACKED || (lost && sp.timeAt(s) < this.firstSampleTime)) {
+        first = -1
+        fresh = false
+        continue
+      }
+      if (!lost) {
+        continue
+      }
+      if (first < 0) {
+        first = sp.timeAt(s)
+      }
+      fresh = fresh || state === QUIC_RECOVERY_LOST_NOW
+      if (fresh && sp.timeAt(s) - first > span) {
+        return true
+      }
+    }
+    return false
+  }
+
+  /** Whether a packet sent at `sentTime` was sent during the current recovery period (B.4). */
+  inRecovery(sentTime: i64): boolean {
+    return this.recoveryStart >= 0 && sentTime <= this.recoveryStart
+  }
+
+  /** OnCongestionEvent (B.6): halve the window, once per recovery period. */
+  congestionEvent(sentTime: i64, now: i64): void {
+    if (this.inRecovery(sentTime)) {
+      return
+    }
+    this.recoveryStart = now
+    this.ssthresh = this.congestionWindow / 2
+    this.congestionWindow = quicRecoveryMax(this.ssthresh, QUIC_RECOVERY_MINIMUM_WINDOW)
+    this.congestionEvents = this.congestionEvents + 1
+  }
+
+  /**
+   * OnPacketAcked (B.5) for slot `s`: in slow start the window grows by the
+   * packet's size, in congestion avoidance by 1200 × size / window; not at
+   * all for a packet sent during recovery, or when `inUse` says the window
+   * was not being used (§7.8).
+   */
+  packetAcked(sp: QuicSentPackets, s: i32, inUse: boolean): void {
+    if (!inUse || this.inRecovery(sp.timeAt(s))) {
+      return
+    }
+    const size: i64 = toI64(sp.sizeAt(s))
+    if (this.congestionWindow < this.ssthresh) {
+      this.congestionWindow = this.congestionWindow + size
+    } else {
+      this.congestionWindow =
+        this.congestionWindow + (QUIC_RECOVERY_DATAGRAM_SIZE * size) / this.congestionWindow
+    }
+  }
+
+  /**
+   * The probe timeout's period without backoff (§6.2.1): smoothed_rtt +
+   * max(4 × rttvar, kGranularity), plus the peer's max_ack_delay once the
+   * handshake is confirmed. It is also what a key update's old keys are kept
+   * for and what the idle timeout is at least three of.
+   */
+  probeTimeout(): i64 {
+    const base: i64 = this.ptoBase()
+    return this.handshakeConfirmed ? base + this.maxAckDelay : base
+  }
+
+  /** smoothed_rtt + max(4 × rttvar, kGranularity): every probe timeout's period before max_ack_delay and the backoff. */
+  ptoBase(): i64 {
+    return this.smoothedRtt + quicRecoveryMax(4 * this.rttVar, QUIC_RECOVERY_GRANULARITY)
+  }
+
+  /**
+   * GetPtoTimeAndSpace (A.8): when the probe timeout fires, the earliest
+   * over the spaces with an ack-eliciting packet in flight, with the backoff
+   * applied; the Application Data space counts only once the handshake is
+   * confirmed. Answers -1 with nothing in flight. `ptoSpace()` names the
+   * space it chose.
+   */
+  ptoTime(): i64 {
+    const space: i32 = this.ptoSpace()
+    if (space < 0) {
+      return -1
+    }
+    return this.ptoTimeOf(space)
+  }
+
+  /** The probe timeout of space `space`, from its last ack-eliciting packet, with the backoff. */
+  ptoTimeOf(space: i32): i64 {
+    const sp: QuicSentPackets | null = this.space(space)
+    if (sp === null || sp.lastAckEliciting < 0) {
+      return -1
+    }
+    const backoff: i64 = toI64(1) << toI64(this.ptoCount)
+    let duration: i64 = this.ptoBase()
+    if (space === QUIC_RECOVERY_APPLICATION) {
+      duration = duration + this.maxAckDelay
+    }
+    return sp.lastAckEliciting + duration * backoff
+  }
+
+  /** The space whose probe timeout fires first, or -1 when no space may have one. */
+  ptoSpace(): i32 {
+    let best: i32 = -1
+    let bestTime: i64 = -1
+    for (let space: i32 = 0; space < toI32(this.spaces.length); space += 1) {
+      const sp: QuicSentPackets = this.spaces[space]
+      if (sp.inFlight === 0 || (space === QUIC_RECOVERY_APPLICATION && !this.handshakeConfirmed)) {
+        continue
+      }
+      const at: i64 = this.ptoTimeOf(space)
+      if (best < 0 || at < bestTime) {
+        best = space
+        bestTime = at
+      }
+    }
+    return best
+  }
+
+  /** The space with the earliest `lossTime`, or -1 when none is set. */
+  lossSpace(): i32 {
+    let best: i32 = -1
+    let bestTime: i64 = -1
+    for (let space: i32 = 0; space < toI32(this.spaces.length); space += 1) {
+      const at: i64 = this.spaces[space].lossTime
+      if (at >= 0 && (best < 0 || at < bestTime)) {
+        best = space
+        bestTime = at
+      }
+    }
+    return best
+  }
+
+  /**
+   * SetLossDetectionTimer (A.8): when `onTimeout` is next due — the earliest
+   * time-threshold loss, else the probe timeout — or -1 for none. With
+   * `amplificationBlocked`, when the server may send nothing until the client
+   * does (RFC 9000 §8.1), there is no probe timeout (§6.2.2.1).
+   */
+  deadline(amplificationBlocked: boolean): i64 {
+    const loss: i32 = this.lossSpace()
+    if (loss >= 0) {
+      return this.spaces[loss].lossTime
+    }
+    if (amplificationBlocked) {
+      return -1
+    }
+    return this.ptoTime()
+  }
+
+  /**
+   * OnLossDetectionTimeout (A.9), at `now`. When a time-threshold loss is
+   * due, the space's packets are checked again and the lost ones listed in
+   * its `lost`: `QUIC_RECOVERY_TIMEOUT_LOSS`. Otherwise, when the probe
+   * timeout is due, the backoff doubles and the caller is to send a probe in
+   * `timeoutSpace`: `QUIC_RECOVERY_TIMEOUT_PTO`. Before either is due, or
+   * when nothing is timed, `QUIC_RECOVERY_TIMEOUT_NONE`.
+   */
+  onTimeout(now: i64, amplificationBlocked: boolean): i32 {
+    for (const sp of this.spaces) {
+      sp.compact()
+      sp.clearResults()
+    }
+    this.timeoutSpace = -1
+    const due: i64 = this.deadline(amplificationBlocked)
+    if (due < 0 || now < due) {
+      return QUIC_RECOVERY_TIMEOUT_NONE
+    }
+    const loss: i32 = this.lossSpace()
+    if (loss >= 0) {
+      const sp: QuicSentPackets = this.spaces[loss]
+      this.detectLoss(sp, now)
+      this.packetsLost(sp, now)
+      this.timeoutSpace = loss
+      return QUIC_RECOVERY_TIMEOUT_LOSS
+    }
+    this.timeoutSpace = this.ptoSpace()
+    if (this.ptoCount < QUIC_RECOVERY_MAX_BACKOFF) {
+      this.ptoCount = this.ptoCount + 1
+    }
+    return QUIC_RECOVERY_TIMEOUT_PTO
+  }
+
+  /**
+   * Gives up the oldest packet still in flight in space `space`, taking it
+   * out of flight without a congestion event, so that a probe can be recorded
+   * when the ring is full. The caller has already queued what the space had
+   * in flight again, which is what a probe sends. Answers the slot, or -1
+   * when nothing is in flight there.
+   */
+  evictOldest(space: i32): i32 {
+    const sp: QuicSentPackets | null = this.space(space)
+    if (sp === null) {
+      return -1
+    }
+    sp.compact()
+    for (let k: i32 = 0; k < sp.count; k += 1) {
+      const s: i32 = sp.slot(k)
+      if (sp.stateAt(s) === QUIC_RECOVERY_SENT) {
+        sp.settle(s, QUIC_RECOVERY_LOST)
+        return s
+      }
+    }
+    return -1
+  }
+
+  /**
+   * OnPacketNumberSpaceDiscarded (A.11): the space's keys are gone, so its
+   * packets leave flight uncounted, its timers stop, and the backoff resets.
+   */
+  discardSpace(space: i32): void {
+    const sp: QuicSentPackets | null = this.space(space)
+    if (sp === null) {
+      return
+    }
+    sp.clear()
+    this.ptoCount = 0
+  }
+
+  /**
+   * Tops the pacer up to `now` (§7.7): it earns 5/4 of the window per
+   * smoothed RTT, and holds at most the initial window, which is the largest
+   * burst it allows.
+   */
+  pacerRefill(now: i64): void {
+    if (this.pacerTime < 0 || now < this.pacerTime) {
+      this.pacerTime = now
+      return
+    }
+    // A minute is far past what fills the budget, and keeps the product in range.
+    let elapsed: i64 = now - this.pacerTime
+    if (elapsed > 60000) {
+      elapsed = 60000
+    }
+    const rtt: i64 = quicRecoveryMax(this.smoothedRtt, QUIC_RECOVERY_GRANULARITY)
+    const earned: i64 = (elapsed * this.congestionWindow * 5) / (rtt * 4)
+    this.pacerBudget = this.pacerBudget + earned
+    if (this.pacerBudget > QUIC_RECOVERY_INITIAL_WINDOW) {
+      this.pacerBudget = QUIC_RECOVERY_INITIAL_WINDOW
+    }
+    this.pacerTime = now
+  }
+
+  /**
+   * How many milliseconds after `now` the pacer lets a datagram of `size`
+   * bytes go: 0 when it may go now, else at least 1. The rate is 5/4 of the
+   * window per smoothed RTT (§7.7's N of 1.25).
+   */
+  pacerDelay(now: i64, size: i32): i64 {
+    this.pacerRefill(now)
+    const need: i64 = toI64(size) - this.pacerBudget
+    if (need <= 0) {
+      return 0
+    }
+    const rtt: i64 = quicRecoveryMax(this.smoothedRtt, QUIC_RECOVERY_GRANULARITY)
+    const rate: i64 = quicRecoveryMax(this.congestionWindow * 5, 1)
+    const wait: i64 = (need * rtt * 4 + rate - 1) / rate
+    return quicRecoveryMax(wait, 1)
+  }
+
+  /** Charges the pacer for a datagram of `size` bytes sent at `now`. */
+  onPaced(now: i64, size: i32): void {
+    this.pacerRefill(now)
+    this.pacerBudget = this.pacerBudget - toI64(size)
+  }
+}